Wireless sensing using a basic model

Abstract

Examples of wireless sensing, tracking, imaging, and occupancy detection are described. [Solution] In one example, the method described comprises generating channel information (CI) data based on at least one radio channel, generating a training dataset based on the CI data, the training dataset comprising a plurality of CI pairs, original CI data, and a mask, training a base model using the training dataset based on an aggregation of a contrast loss function and a reconstruction loss function, training a plurality of task-specific models, and performing a plurality of radio sensing tasks based on the base model and the plurality of task-specific models. Each of the plurality of task-specific models is used together with the base model to perform a corresponding one of the plurality of radio sensing tasks.

Description

Technical Field

[0001] The present disclosure generally relates to wireless sensing, tracking, imaging, and occupancy detection. More specifically, the present disclosure relates to wireless sensing based on a base model, wireless sensing in a network of networks having coded addresses, measurement enhancement, wireless sensing using classifier probing and narrowing, wireless sensing for in-vehicle child presence detection, wireless human imaging, passive tracking using wireless signals, and wireless-based occupancy detection and activity monitoring.

Background Art

[0002] With the popularization of Internet of Things (IoT) devices, indoor intelligent applications such as security monitoring, intruder detection, occupancy monitoring, and activity recognition have attracted great attention. However, these applications frequently have the problem of a high false alarm rate because they cannot recognize humans as well as non-human objects such as pets, robotic vacuum cleaners, and electrical appliances. This discrimination ability is essential especially for applications related to security, health monitoring, automation, and energy management. False recognition can lead to user frustration, damage trust, and prevent the practical and widespread adoption of these technologies. Considering the popularization of pets, robotic vacuum cleaners, and electrical appliances in a residential environment, it is extremely important to develop a reliable system that can accurately recognize human and non-human objects.

[0003] Precisely distinguishing between human and non-human movements remains a challenge today. For example, camera-based methods and thermal sensor-based approaches can only detect moving objects within a line of sight (LOS). Furthermore, camera-based systems raise privacy concerns. Strategies using radar to distinguish pets from humans based on vital signs assume the pet is stationary, which is unrealistic. In addition, these methods have strict device placement requirements, are often limited to the LOS, and assume that the object is moving within a given area.

[0004] The development of IoT applications means that billions of home appliances, phones, smart devices, security systems, environmental sensors, vehicles and buildings, and other wirelessly connected devices will transmit data, communicate with each other or with people, and everything will be constantly measurable and trackable. Among the various approaches to measuring what is happening in the surrounding environment, wireless sensing has attracted increasing attention in recent years due to the ubiquitous deployment of wireless devices. Furthermore, since human activity affects the propagation of wireless signals, understanding and analyzing how wireless signals respond to human activity can reveal a wealth of information about that activity. As more bandwidth becomes available in next-generation wireless systems, wireless sensing will enable many more smart IoT applications in the near future than can be imagined today. This is because increased bandwidth allows for the detection of more multipaths in rich scattering environments such as indoors and in urban areas, and these can be treated as hundreds of virtual antennas / sensors. Given the potential for many IoT devices to be available for wireless sensing, there is a need for efficient and effective methods to utilize multiple devices for wireless sensing.

[0005] Radio frequency (RF) object imaging (of objects such as humans, animals, pets, furniture, machinery, and vehicles) is a new field of research spurred by the increasing availability of RF radar radio devices. While existing research has achieved object reconstruction (e.g., reconstruction of the human body for pose estimation), object identification through imaging (e.g., humans) has not been feasible due to low resolution.

[0006] The rapid growth of IoT devices has driven the development of indoor tracking systems. Existing indoor tracking systems primarily employ visual, acoustic, infrared, or radar technologies. However, these technologies have limitations that can hinder their effectiveness and scalability. Visual-based systems can infringe on privacy and fail in poor lighting conditions, while acoustic systems are often susceptible to background noise. Infrared systems require line of sight (LOS) to function properly, and radar-based systems require expensive, specialized hardware and have limited coverage.

[0007] The affordability and widespread availability of Wi-Fi, along with its cost-effectiveness and ubiquitous nature, have made it a popular choice for indoor tracking research and applications. Many Wi-Fi-based indoor tracking systems require specific device placements, and the complexity of these systems is quite high. Some systems also require extensive training. Furthermore, most systems require the tracked target to carry a dedicated device during tracking, which can be impractical or inconvenient in everyday use.

[0008] Occupancy detection, such as detecting human occupancy indoors, has become an essential technology in modern society, aimed at improving energy efficiency and occupant comfort. This technology enables control of heating, ventilation, and air conditioning (HVAC) systems, lighting, and other smart energy and resource optimization applications, thereby reducing energy consumption and costs. Furthermore, occupancy detection can enhance security and safety measures by identifying unauthorized presence. Traditional occupancy detection approaches, such as cameras, PIR sensors, RFID sensors, and motion sensors, are often cost-ineffective due to the need for additional installation. They also have problems with privacy infringement and low accuracy.

[0009] The ubiquitous availability of IoT smart devices has made using Wi-Fi for occupancy detection a popular solution due to its convenience, cost-effectiveness, privacy-preserving capabilities, and broad coverage. However, existing Wi-Fi sensing applications cannot identify stationary humans, such as people sleeping, because detection relies solely on motion characteristics. Even though breath detection in the absence of motion can be used to detect static human targets, detecting breath requires high-quality CSI, which is typically extracted from 5GHz devices with a bandwidth of at least 40MHz. However, many IoT devices do not provide the CSI quality necessary for accurate breath detection. Furthermore, model-based Wi-Fi sensing approaches that rely on interpretable statistics such as motion and breath may not work well in increasingly complex environments as other sources of motion introduce noise and interference. [Overview of the project]

[0010] This disclosure relates in general to wireless sensing, tracking, imaging, and occupancy detection. More specifically, this teaching relates to wireless sensing based on a basic model, wireless sensing in a network of networks with coded addresses, wireless sensing using measurement augmentation, probing and narrowing down of classifiers, wireless sensing for detecting the presence of children in vehicles, wireless human imaging, passive tracking using wireless signals, and wireless-based occupancy detection and activity monitoring.

[0011] In one embodiment, a method for wireless sensing is described. This method involves acquiring channel information (CI) data generated based on at least one wireless channel, A training dataset is generated based on the CI data, which includes multiple pairs of CI samples and associated masks. At least partially, the contrastive loss function is determined based on a first similarity metric between the CI data of each CI pair in the training dataset, The reconstruction loss function is determined based on a second similarity metric between the original CI data and the predicted CI data generated based on the mask. The total loss function is determined based on the aggregation of the contrastive loss function and the reconstruction loss function, Training the underlying model using the training dataset by determining the model parameters of the underlying model to minimize the overall loss function, Training multiple task-specific models, The process involves executing a plurality of wireless sensing tasks based on the aforementioned basic model and the plurality of task-specific models, wherein each of the plurality of task-specific models is used together with the basic model to execute one of the plurality of wireless sensing tasks. Includes.

[0012] In another embodiment, a device for wireless sensing is described. This device is At least one processor, At least one memory, Equipped with, The memory, when executed, obtains channel information (CI) data generated based on at least one radio channel from the at least one processor, The process involves generating a training dataset based on the aforementioned CI data, wherein the training dataset includes multiple CI pairs, the original CI data, and a mask. At least partially, the contrastive loss function is determined based on a first similarity metric between the CI data of each CI pair in the training dataset, Determining a reconstruction loss function based on a second similarity metric, wherein the second similarity metric is a metric between the original CI data and the predicted CI data generated based on the mask, The total loss function is determined based on the aggregation of the contrastive loss function and the reconstruction loss function, Training the underlying model using the training dataset by determining the model parameters of the underlying model to minimize the overall loss function, Training multiple task-specific models, The process involves executing a plurality of wireless sensing tasks based on the aforementioned basic model and the plurality of task-specific models, wherein each of the plurality of task-specific models is used together with the basic model to execute one of the plurality of wireless sensing tasks. Stores instructions that perform actions including those mentioned.

[0013] In yet another embodiment, a system for wireless sensing is described. The system comprises at least one local device and a cloud server. The at least one local device is configured to acquire channel information (CI) data generated based on at least one wireless channel and to generate a training dataset based on the CI data, the training dataset comprising a plurality of CI pairs, original CI data, and a mask. The cloud server is configured, at least in part, to train a base model using the training dataset and to train a plurality of task-specific models by determining a contrastive loss function based on a first similarity metric between the CI data of each CI pair in the training dataset, determining a reconstruction loss function based on a second similarity metric between the original CI data and predicted CI data generated based on the mask, determining a total loss function based on the aggregation of the contrastive loss function and the reconstruction loss function, and determining the model parameters of the base model to minimize the total loss function. The at least one local device and the cloud server are further configured to perform a plurality of wireless sensing tasks based on the base model and the plurality of task-specific models. Each of the aforementioned task-specific models is used, together with the underlying model, to perform one of the aforementioned wireless sensing tasks.

[0014] In a different embodiment, a method for radio sensing using coded addresses to identify transmitting devices is described. The method involves associating a plurality of Type 1 devices in a venue with Type 2 devices in the venue, wherein each of the plurality of Type 1 devices is a heterogeneous radio device having a first type, and the Type 2 devices are heterogeneous radio devices having a second type different from the first type, and each of the Type 2 devices has a plurality of coded addresses associated with each of the Type 1 devices in the plurality of Type 1 devices. The Type 2 device transmits a time series of multiple radio trigger signals to multiple Type 1 devices via a radio channel within the venue in order to trigger radio sounding, wherein each time series of radio trigger signals is transmitted to each Type 1 device along with the encoded address of the Type 2 device associated with each Type 1 device, and the encoded address is embedded as the source address in each radio trigger signal of each time series. Each Type 1 device transmits its respective radio sounding signal to the Type 2 device in response to its respective radio trigger signal, along with its respective associated coded address, wherein the respective associated coded address is embedded as the destination address in the respective radio sounding signal. The Type 2 device receives multiple radio sounding signals from the multiple Type 1 devices, which are addressed to one of the coded addresses of the Type 2 device. For each received wireless sounding signal, the processor of the Type 2 device acquires channel information (CI) for each of the wireless channels within the venue based on the received wireless sounding signal, The processor compares the received encoded address of the Type 2 device, which is embedded as a destination address in the received wireless sounding signal, with a plurality of encoded addresses stored in a database. The processor identifies each Type 1 device having an associated stored coded address of the Type 2 device that matches the received coded address as the transmitting device of the received radio sounding signal and each of the CIs, The processor assembles time series (TSCI) of multiple CIs associated with the multiple Type 1 devices, wherein each TSCI associated with each Type 1 device includes all CIs originating from each of the Type 1 devices. The processor performs a wireless sensing task based on the plurality of TSCIs associated with the plurality of Type 1 devices, Includes.

[0015] In another embodiment, a wireless device within a venue is described. This wireless device comprises a data storage for storing a plurality of coded addresses of the wireless device, a transmitter, a receiver, and a processor communicatively coupled to the data storage, the transmitter, and the receiver. The wireless device is associated with a plurality of Type 1 devices within the venue. Each of the plurality of Type 1 devices is a heterogeneous wireless device having a first type. The wireless device is a Type 2 device having a second type different from the first type. Each of the plurality of coded addresses is associated with each of the Type 1 devices within the plurality of Type 1 devices. The transmitter is configured to transmit a time series of a plurality of wireless trigger signals to the plurality of Type 1 devices via a wireless channel within the venue to trigger wireless sounding. Each time series of wireless trigger signals is transmitted to the respective Type 1 device along with the coded address of the Type 2 device associated with the respective Type 1 device. The coded address is embedded as a source address in each wireless trigger signal of each time series of the wireless trigger signals. The receiver is configured to receive a plurality of wireless sounding signals from the plurality of Type 1 devices, addressed to one of the coded addresses of the Type 2 device. Each radio sounding signal is received from each Type 1 device in response to each respective radio trigger signal, along with the associated coded address of the Type 2 device. The associated coded address is embedded as the destination address in each radio sounding signal.The processor, for each received radio sounding signal, obtains channel information (CI) for each of the radio channels in the venue based on the received radio sounding signal, compares the received encoded address of the Type 2 device embedded as a destination address in the received radio sounding signal with a plurality of encoded addresses stored in the data storage, and identifies each Type 1 device having an associated stored encoded address of the Type 2 device that matches the received encoded address as the transmitting device for the received radio sounding signal and each of the CIs. A time series (TSCI) of multiple CIs associated with the multiple Type 1 devices is assembled, where each TSCI associated with each Type 1 device includes all CIs originating from each of the Type 1 devices. It is configured to perform a wireless sensing task based on the plurality of TSCIs associated with the plurality of Type 1 devices.

[0016] In yet another embodiment, a system for wireless sensing is described. The system comprises a plurality of Type 1 devices in a venue, each being a heterogeneous wireless device having a first type, and a set of Type 2 devices in the venue. The Type 2 devices are heterogeneous wireless devices having a second type distinct from the first type. Each of the Type 2 devices has a plurality of coded addresses associated with each of the plurality of Type 1 devices. The Type 2 devices are configured to transmit a sequence of multiple wireless trigger signals to the plurality of Type 1 devices via a wireless channel in the venue in order to trigger wireless sounding. Each sequence of wireless trigger signals is transmitted to the respective Type 1 device along with the coded address of the Type 2 device associated with the respective Type 1 device. The coded address is embedded as a source address in each wireless trigger signal of each sequence of wireless trigger signals. Each Type 1 device is configured to transmit a respective wireless sounding signal to the Type 2 device in response to each respective wireless trigger signal, along with the respective associated coded address of the Type 2 device. The respective associated coded address is embedded as a destination address in the respective wireless sounding signal. The Type 2 device receives multiple radio sounding signals from the multiple Type 1 devices, which are addressed to any of the coded addresses of the Type 2 device. For each received wireless sounding signal, obtain respective channel information (CI) of the wireless channels in the venue based on the received wireless sounding signal, compare the received encoded address of the type 2 device embedded as a destination address in the received wireless sounding signal with a plurality of encoded addresses stored in a database, and identify each type 1 device having the associated stored encoded address that matches the received encoded address as the transmitting device of the received wireless sounding signal and the respective CI, Assemble a time series (TSCI) of a plurality of CIs associated with the plurality of type 1 devices, where each TSCI associated with each type 1 device includes all CIs transmitted from the respective type 1 device, It is configured to perform a wireless sensing task based on the plurality of TSCs associated with the plurality of type 1 devices.

[0017] In different embodiments, a method for wireless sensing involving conversion component selection is described. The method includes receiving, by a receiver, a wireless signal transmitted by a transmitter via a wireless multipath channel in a venue, where the received wireless signal is different from the transmitted wireless signal due to the wireless multipath channel and the movement of objects in the venue, Obtaining a time series (TSCI) of channel information (CI) of the wireless multipath channel based on the received wireless signal, where each channel information (CI) includes N1 CI components (CICs), For each of the respective CIs having N1 CICs of the TSCI, calculating a conversion from each of the respective CIs in the first domain of the N1 CICs to a respective converted CI (TCI) having N2 converted CICs (TCICs) in the second domain, where N1 <= N2, Selecting N3, which is the number of each of the TCICs of the TCI, such that for any N3, N3 <= N2, and N1, N2, and N3 are all positive integers, Calculating an enhanced TCI (ETCI) having N2 TCICs by highlighting the selected N3 TCICs of the TCI in the second domain, Calculating the inverse transformation of the ETCI from the second domain to the first domain to obtain each ECI having N2 CICs, wherein the inverse transformation is the inverse of the transformation, thereby calculating each enhanced CI (ECI) having N2 CICs, Calculating motion information (MI) based on the ECI of the TSCI, Executing a sensing task based on the MI, and including.

[0018] In another embodiment, a wireless device within a venue is described. This wireless device comprises a receiver configured to receive a wireless signal transmitted by a transmitting device via a wireless multipath channel of the venue, wherein the received wireless signal is different from the transmitted wireless signal due to the wireless multipath channel and the movement of objects within the venue, the receiver, a memory, and a processor communicatively coupled to the memory. The processor obtains a time series (TSCI) of channel information (CI) of the wireless multipath channel based on the received wireless signal, where each channel information (CI) includes N1 CI components (CIC), and for each CI having N1 CICs in the TSCI, calculates the conversion from each CI in a first domain of N1 CICs to each transformed CI (TCI) having N2 transformed CICs (TCICs) in a second domain, where N1 <= N2, and selects N3, which is the number of TCICs in each TCI, where N3 <= N2 for any N3, and N1, N The system is configured to calculate an enhanced TCI (ETCI) having N2 TCICs by selecting 2 and N3, highlighting each of the N3 selected TCICs in the TCI in the second domain, and to obtain each ECI having N2 CICs by calculating the inverse transformation of the ETCI from the second domain to the first domain, where the inverse transformation is the inverse of the transformation, and to calculate motion information (MI) based on the ECI of the TSCI and to perform a sensing task based on the MI.

[0019] In yet another embodiment, a system for wireless sensing is described. The system comprises a transmitter configured to transmit a wireless signal over a wireless multipath channel of a venue, and a receiver. The receiver is configured to receive the wireless signal over the wireless multipath channel of the venue, the received wireless signal being different from the transmitted wireless signal due to the wireless multipath channel and the movement of objects in the venue, and is configured to obtain a time series (TSCI) of channel information (CI) of the wireless multipath channel based on the received wireless signal, where each channel information (CI) includes N1 CI components (CIC), and for each CI having N1 CICs in the TSCI, the calculation is performed to convert each CI in a first domain of N1 CICs to each converted CI (TCI) having N2 converted CICs (TCIC) in a second domain, where N1 <= N2. The system is configured to select N3, which is the number of TCICs in the TCI, such that for any N3, N3 <= N2, and N1, N2, and N3 are all positive integers; to calculate an enhanced TCI (ETCI) having N2 TCICs by highlighting each of the N3 selected TCICs in the TCI in the second domain; to calculate the inverse transformation of the ETCI from the second domain to the first domain to obtain each ECI having N2 CICs, such that the inverse transformation is the inverse of the transformation; to calculate each enhanced CI (ECI) having N2 CICs; to calculate motion information (MI) based on the ECI of the TSCI; and to perform a sensing task based on the MI.

[0020] In different embodiments, a method performed by a system for wireless sensing, involving probing and narrowing down of classifiers, is described. This method involves, in the probing phase of the system, acquiring multiple raw measurement data by the sensing device of the system, processing the multiple raw measurement data by the processor of the system to construct multiple input data for a classifier, performing classification using the classifier by inputting each of the multiple input data to the classifier by the processor, and the classifier calculating multiple output analyses based on the multiple input data, each output analysis being calculated by the classifier based on its respective input data, and mapping the multiple output analyses to a plurality of mapped results, each output analysis being mapped to its respective mapped result, and identifying at least one reference input data associated with a reference output analysis and a reference mapping result, each reference input data being one of the multiple input data for the classifier, whose respective reference result is available and different from the reference mapping result, and each With respect to the reference input data, construct a plurality of perturbation input data for the classifier by perturbing the reference input data, each perturbation input data being constructed based on the respective perturbations of the reference input data; perform the classification using the classifier by inputting each of the plurality of perturbation input data into the classifier; have the classifier calculate a plurality of perturbation output analyses based on the plurality of perturbation input data, each perturbation output analysis being calculated by the classifier based on the respective perturbation input data; map the plurality of perturbation output analyses to a plurality of perturbation mapping results, each perturbation output analysis being mapped to the respective perturbation mapping result; compare each perturbation mapping result with the reference result associated with the reference input data, which is different from the reference mapping result; and if at least one perturbation mapping result deviates from the reference mapping result and is the same as the reference result, select the perturbation input data for retraining the classifier.The process includes selecting at least one selected perturbation input data, which is one of the multiple perturbation input data associated with one of the at least one perturbation mapping results, and retraining the classifier in the retraining phase of the system based on the reference results associated with each of the at least one selected perturbation input data.

[0021] In another embodiment, a device for wireless sensing with probing and filtering of a classifier is described. The device comprises a receiver configured to acquire a plurality of raw measurement data in the probing phase of the device, and a classifier having memory and a processor. The processor processes the plurality of raw measurement data in the probing phase of the device to construct a plurality of input data, uses each of the plurality of input data as input to the classifier to perform classification, calculates a plurality of output analyses based on the plurality of input data, where each output analysis is calculated based on its respective input data, maps the plurality of output analyses to a plurality of mapped results, where each output analysis is mapped to its respective mapped result, each identifying at least one reference input data associated with a reference output analysis and a reference mapping result, where each reference input data is one of the plurality of input data for the classifier, whose respective reference result is available and different from the reference mapping result, and for each reference input data for the classifier, constructs a plurality of perturbed input data for the classifier by perturbing the reference input data, where each perturbed input data is The system is constructed based on the perturbations of each of the reference input data, performs the classification using each of the plurality of perturbation input data as input to the classifier, calculates each of the plurality of perturbation output analyses based on the plurality of perturbation input data, where each perturbation output analysis is calculated based on the respective perturbation input data, maps the plurality of perturbation output analyses to a plurality of perturbation mapping results, where each perturbation output analysis is mapped to the respective perturbation mapping result, compares each perturbation mapping result to the reference result associated with the reference input data which is different from the reference mapping result, and is configured to select at least one selected perturbation input data, which is one of the plurality of perturbation input data associated with one of the at least one perturbation mapping results, as selected perturbation input data for retraining the classifier if at least one perturbation mapping result deviates from the reference mapping result and is the same as the reference result.The processor is configured to retrain the classifier during the retraining phase of the device based on the reference result associated with each of the at least one selected perturbation input data.

[0022] In yet another embodiment, a system for wireless sensing with probing and filtering by a classifier is described. The system comprises a sensing device configured to acquire multiple raw measurement data in the probing phase of the system, and a classifier having memory and a processor.The processor, in the probing phase of the system, processes the plurality of raw measurement data to construct a plurality of input data, uses each of the plurality of input data as input to the classifier to perform classification, calculates a plurality of output analyses based on the plurality of input data, where each output analysis is calculated based on its respective input data, maps the plurality of output analyses to a plurality of mapped results, where each output analysis is mapped to its respective mapped result, identifies at least one reference input data associated with a reference output analysis and a reference mapping result, where each reference input data is one of the plurality of input data for the classifier, whose respective reference result is available and different from the reference mapping result, and for each reference input data for the classifier, constructs a plurality of perturbed input data for the classifier by perturbing the reference input data, where each perturbed input data is The system is configured to construct a perturbation on each of the reference input data, perform the classification using each of the plurality of perturbation input data as input to the classifier, calculate each of the plurality of perturbation output analyses on the plurality of perturbation input data, where each perturbation output analysis is calculated on the respective perturbation input data, map the plurality of perturbation output analyses to a plurality of perturbation mapping results, where each perturbation output analysis is mapped to the respective perturbation mapping result, compare each perturbation mapping result to the reference result associated with the reference input data which is different from the reference mapping result, and if at least one perturbation mapping result deviates from the reference mapping result and is the same as the reference result, select at least one selected perturbation input data which is one of the plurality of perturbation input data associated with one of the at least one perturbation mapping results as selected perturbation input data for retraining the classifier. The processor is configured to retrain the classifier in the retraining phase of the system based on the reference result associated with each of the at least one selected perturbation input data.

[0023] In one embodiment, a method for detecting the presence of a child in a vehicle is described. This method involves transmitting a radio signal via a radio channel in the vehicle using a transmitting device, and receiving the radio signal using a receiving device in the vehicle, wherein the received radio signal is different from the transmitted radio signal due to the radio channel and the movement of the user in the vehicle; obtaining a time series (TSCI) of channel information (CI) of the radio channel based on the received radio signal; calculating motion statistics (MS) based on N5 channel information (CI) of the TSCI over a certain period; and detecting the user's movement over the period by comparing the motion statistics with a first threshold, wherein the user's movement is detected when the motion statistics exceed the first threshold; and associating each of the N9 sliding time windows in the period with its respective timestamp. The process involves calculating the one-dimensional (1D) transformation of each CI of the TSCI in the sliding time window, constructing the 2D transformation matrix by assembling and concatenating the N9 1D transformations as columns of a 2D transformation matrix in ascending order of the associated timestamps, detecting the number of horizontal streaks in the 2D transformation matrix, detecting the user's breathing behavior during the period based on the detected number of horizontal streaks in the 2D transformation matrix and associated criteria, calculating the vibration coefficient (OR) by counting zero crossings of the 2D transformation matrix in the transformation domain, calculating the vibration CS by mapping the OR to a vibration confidence score (CS) using a first mapping function, calculating the DTW cost by applying DTW to the 2D transformation matrix, and mapping the DTW cost to a DTW using a second mapping function. By mapping to CS, the dynamic time warping (DTW) CS is calculated, and the streak count of the number of horizontal streaks in the detected 2D transformation matrix is ​​calculated, and the peak CS is calculated by mapping the streak count to the peak CS using a third mapping function, and the vibration CS, the DTWThe method comprises: calculating a combined CS (CCS) as a first aggregate of CS and the peak CS; calculating a final confidence score (FCS) as a second moving aggregate of the CCS and a number of temporally adjacent CCS associated with adjacent periods of the period; rejecting any detection of the breathing motion in the period if the FCS is below a second threshold; detecting the presence of the user in the period if (a) the user's movement is detected, or (b) the user's breathing motion is detected and not rejected; and determining, when the presence of the user is detected, whether the user is a child or an adult based on the TSCI and a classifier, wherein the classifier comprises a convolutional neural network (CNN) followed by a fully connected network (FCN) having input data including (1) a plurality of first 2D input matrices constructed from the amplitudes of the TSCI without phase, and (2) a plurality of second 2D input matrices constructed from the phases of the TSCI without amplitude.

[0024] In another embodiment, a device for detecting the presence of a child in a vehicle is described. The device comprises a receiver configured to receive a radio signal transmitted by a transmitter over a radio channel in a vehicle, wherein the received radio signal is different from the transmitted radio signal due to the radio channel and the movement of a user in the vehicle, and a processor. The processor obtains a time series (TSCI) of channel information (CI) of the radio channel based on the received radio signal, calculates motion statistics (MS) based on N5 channel information (CI) of the TSCI over a period of time, and detects the user's movement over the period by comparing the motion statistics to a first threshold, where the user's movement is detected when the motion statistics exceed the first threshold, and calculates the respective one-dimensional (1D) transformation of the CI of the TSCI in the sliding time window associated with the respective timestamp for each of the N9 sliding time windows over the period, and the associated timestamp The process involves constructing a 2D transformation matrix by assembling and concatenating the N9 1D transformations in ascending order of the number of steps, detecting the number of horizontal streaks in the 2D transformation matrix, detecting the user's breathing behavior during the period based on the detected number of horizontal streaks in the 2D transformation matrix and associated criteria, calculating the vibration coefficient (OR) by counting zero crossings in the 2D transformation matrix in the transformation domain, calculating the vibration CS by mapping the OR to the vibration confidence score (CS) using a first mapping function, calculating the DTW cost by applying DTW to the 2D transformation matrix, calculating the dynamic time warping (DTW) CS by mapping the DTW cost to the DTW CS using a second mapping function, calculating the streak count of the detected number of horizontal streaks in the 2D transformation matrix, and calculating the peak CS by mapping the streak count to the peak CS using a third mapping function, and then calculating the vibration CS, the DTWThe system is configured to calculate a combined CS (CCS) as a first aggregate of the CS and the peak CS, calculate a final confidence score (FCS) as a second moving aggregate of the CCS and a number of temporally adjacent CCS associated with adjacent periods of the period, reject any detection of the breathing motion in the period if the FCS is below a second threshold, detect the presence of the user in the period if (a) the user's movement is detected, or (b) the user's breathing motion is detected and not rejected, and determine whether the user is a child or an adult based on the TSCI and classifier when the presence of the user is detected, wherein the classifier comprises a convolutional neural network (CNN) followed by a fully connected network (FCN) having input data including (1) a first plurality of 2D input matrices constructed only from the amplitudes of the TSCI without phase, and (2) a second plurality of 2D input matrices constructed only from the phases of the TSCI without amplitude.

[0025] In yet another embodiment, a system for detecting the presence of a child in a vehicle is described. The system comprises a transmitting device configured to transmit a radio signal over a radio channel in the vehicle, a receiving device configured to receive the radio signal, wherein the received radio signal is different from the transmitted radio signal due to the radio channel and the movement of a user in the vehicle, and a processor. The processor obtains a time series (TSCI) of channel information (CI) of the radio channel based on the received radio signal, calculates motion statistics (MS) based on N5 channel information (CI) of the TSCI over a period of time, and detects the user's movement over the period by comparing the motion statistics to a first threshold, where the user's movement is detected when the motion statistics exceed the first threshold, and calculates the respective one-dimensional (1D) transformation of the CI of the TSCI in the sliding time window associated with each timestamp for each of the N9 sliding time windows over the period, and the associated timestamp The process involves constructing a 2D transformation matrix by assembling and concatenating the N9 1D transformations in ascending order of the number of steps, detecting the number of horizontal streaks in the 2D transformation matrix, detecting the user's breathing behavior during the period based on the detected number of horizontal streaks in the 2D transformation matrix and associated criteria, calculating the vibration coefficient (OR) by counting zero crossings in the 2D transformation matrix in the transformation domain, calculating the vibration CS by mapping the OR to the vibration confidence score (CS) using a first mapping function, calculating the DTW cost by applying DTW to the 2D transformation matrix, calculating the dynamic time warping (DTW) CS by mapping the DTW cost to the DTW CS using a second mapping function, calculating the streak count of the detected number of horizontal streaks in the 2D transformation matrix, and calculating the peak CS by mapping the streak count to the peak CS using a third mapping function, and then calculating the vibration CS, the DTWThe system is configured to calculate a combined CS (CCS) as a first aggregate of the CS and the peak CS, calculate a final confidence score (FCS) as a second moving aggregate of the CCS and a number of temporally adjacent CCS associated with adjacent periods of the period, reject any detection of the breathing motion in the period if the FCS is below a second threshold, detect the presence of the user in the period if (a) the user's movement is detected, or (b) the user's breathing motion is detected and not rejected, and determine whether the user is a child or an adult based on the TSCI and classifier when the presence of the user is detected, wherein the classifier comprises a convolutional neural network (CNN) followed by a fully connected network (FCN) having input data including (1) a first plurality of 2D input matrices constructed only from the amplitudes of the TSCI without phase, and (2) a second plurality of 2D input matrices constructed only from the phases of the TSCI without amplitude.

[0026] In yet another embodiment, a vehicle is described. The vehicle comprises a transmitting device configured to transmit radio signals over a radio channel within the vehicle, a receiving device configured to receive the radio signals, wherein the received radio signals differ from the transmitted radio signals due to the radio channel and the movement of a user within the vehicle, and a processor. The processor obtains a time series (TSCI) of channel information (CI) of the radio channel based on the received radio signals, calculates motion statistics (MS) based on N5 channel information (CI) of the TSCI over a period of time, and detects the user's movement over the period by comparing the motion statistics to a first threshold, where the user's movement is detected when the motion statistics exceed the first threshold, and calculates each one-dimensional (1D) transformation of the CI of the TSCI in the sliding time window associated with each timestamp for each of N9 sliding time windows over the period, and the associated timestamp The process involves constructing a 2D transformation matrix by assembling and concatenating the N9 1D transformations in ascending order of the number of steps, detecting the number of horizontal streaks in the 2D transformation matrix, detecting the user's breathing behavior during the period based on the detected number of horizontal streaks in the 2D transformation matrix and associated criteria, calculating the vibration coefficient (OR) by counting zero crossings in the 2D transformation matrix in the transformation domain, calculating the vibration CS by mapping the OR to the vibration confidence score (CS) using a first mapping function, calculating the DTW cost by applying DTW to the 2D transformation matrix, calculating the dynamic time warping (DTW) CS by mapping the DTW cost to the DTW CS using a second mapping function, calculating the streak count of the detected number of horizontal streaks in the 2D transformation matrix, and calculating the peak CS by mapping the streak count to the peak CS using a third mapping function, and then calculating the vibration CS, the DTWThe system is configured to calculate a combined CS (CCS) as a first aggregate of the CS and the peak CS, calculate a final confidence score (FCS) as a second moving aggregate of the CCS and a number of temporally adjacent CCS associated with adjacent periods of the period, reject any detection of the breathing motion in the period if the FCS is below a second threshold, detect the presence of the user in the period if (a) the user's movement is detected, or (b) the user's breathing motion is detected and not rejected, and determine whether the user is a child or an adult based on the TSCI and classifier when the presence of the user is detected, wherein the classifier comprises a convolutional neural network (CNN) followed by a fully connected network (FCN) having input data including (1) a first plurality of 2D input matrices constructed only from the amplitudes of the TSCI without phase, and (2) a second plurality of 2D input matrices constructed only from the phases of the TSCI without amplitude.

[0027] In one embodiment, a method for encoder-decoder-based high-resolution imaging is described. The method includes: assembling a k1-dimensional (k1-D) imaging matrix by a processor based on arranging and concatenating a plurality of k2-dimensional (k2-D) input imaging matrices; encoding the k1-D imaging matrix by an encoder, which is a k1-D coding neural network, to generate (k1+k3)-D first intermediate matrices; generating (k2+k3)-D second intermediate matrices based on the (k1+k3)-D first intermediate matrices; and decoding the (k2+k3)-D second intermediate matrices by a decoder, which is a k2-D decoding neural network, to generate a k2-D output imaging matrix, wherein at least one skip connection exists between the encoder and the decoder, and the imaging resolution of the k2-D output imaging matrix is ​​greater than the imaging resolution of any of the plurality of k2-D input imaging matrices.

[0028] In another embodiment, a device for encoder-decoder-based high-resolution imaging is described. The device comprises a processor and a memory for storing instructions, which, when executed, cause the processor to perform operations including: assembling a k1-dimensional (k1-D) imaging matrix by arranging and concatenating a plurality of k2-dimensional (k2-D) input imaging matrices; encoding the k1-D imaging matrix using an encoder, which is a k1-D coding neural network, to produce a (k1+k3)-D first intermediate matrix; generating a (k2+k3)-D second intermediate matrix based on the (k1+k3)-D first intermediate matrix; and decoding the (k2+k3)-D second intermediate matrix using a decoder, which is a k2-D decoding neural network, to produce a k2-D output imaging matrix, wherein at least one skip connection exists between the encoder and the decoder, and the imaging resolution of the k2-D output imaging matrix is ​​greater than the imaging resolution of any of the plurality of k2-D input imaging matrices.

[0029] In yet another embodiment, a system for encoder-decoder-based high-resolution imaging is described. This system comprises a processor configured to assemble a k1-dimensional (k1-D) imaging matrix by arranging and concatenating a plurality of k2-dimensional (k2-D) input imaging matrices; an encoder which is a k1-D coding neural network configured to encode the k1-D imaging matrix to produce a (k1+k3)-D first intermediate matrix, wherein the processor is further configured to produce a (k2+k3)-D second intermediate matrix based on the (k1+k3)-D first intermediate matrix; and a decoder which is a k2-D decoding neural network configured to decode the (k2+k3)-D second intermediate matrix to produce a k2-D output imaging matrix, wherein at least one skip connection exists between the encoder and the decoder, and the imaging resolution of the k2-D output imaging matrix is ​​greater than the imaging resolution of any of the plurality of k2-D input imaging matrices.

[0030] In one embodiment, a method for wireless tracking is described. The method involves transmitting a wireless signal via a wireless channel of a venue by at least one transmitter, and receiving the wireless signal via the wireless channel of the venue by a receiver, wherein the received wireless signal is different from the transmitted wireless signal due to the movement of the wireless channel and objects in the venue, and neither the at least one transmitter nor the receiver moves with the object; obtaining a time series (TSCI) of channel information (CI) of the wireless channel in a time window, wherein the TSCI is generated based on the received wireless signal; and generating a plurality of features based on the TSCI over the time window. The method includes generating, where the plurality of features include at least proximity metric (PM) features, motion statistics (MS) features, and spatial features; obtaining a signature map of the venue from a database, where each location on the signature map is assigned one or more location signatures representing unique features associated with that location when some object is present at that location; and for each timestamp in the time window, determining the estimated location of the object within the venue based on the signature map and the plurality of features, and generating a trajectory of the object within the venue across the time window based on the estimated location of the object.

[0031] In another embodiment, a device for wireless tracking is described. This device comprises a receiver and a processor. The receiver is configured to receive a wireless signal from at least one transmitter via a wireless channel in a venue. The received wireless signal is different from the transmitted wireless signal due to the wireless channel and the movement of an object in the venue. Neither the at least one transmitter nor the receiver moves with the object. The processor is configured to acquire a time series (TSCI) of channel information (CI) of the radio channel in a time window, wherein the TSCI is generated based on the received radio signal and is configured to generate a plurality of features based on the TSCI across the time window, wherein the plurality of features include at least proximity metric (PM) features, motion statistics (MS) features, and spatial features; and is configured to acquire a signature map of the venue from a database, wherein each location on the signature map is assigned one or more location signatures representing unique features associated with that location when some object is present at that location; and for each timestamp in the time window, the processor is configured to determine the estimated location of the object within the venue based on the signature map and the plurality of features, and to generate a trajectory of the object within the venue across the time window based on the estimated location of the object.

[0032] In yet another embodiment, a system for wireless tracking is described. This system comprises at least one transmitter, a receiver, and a processor. The at least one transmitter is configured to transmit a wireless signal over a wireless channel in a venue. The receiver is configured to receive the wireless signal over the wireless channel in the venue. The received wireless signal is different from the transmitted wireless signal due to the wireless channel and the movement of an object in the venue. Neither the at least one transmitter nor the receiver moves with the object. The processor is configured to acquire a time series (TSCI) of channel information (CI) of the radio channel in a time window, wherein the TSCI is generated based on the received radio signal, and is configured to generate a plurality of features based on the TSCI across the time window, wherein the plurality of features include at least proximity metric (PM) features, motion statistics (MS) features, and spatial features, and is configured to acquire a signature map of the venue from a database, wherein each location on the signature map is assigned one or more location signatures representing unique features associated with that location when some object is present at that location, and for each timestamp in the time window, is configured to determine the estimated location of the object within the venue based on the signature map and the plurality of features, and to generate the trajectory of the object within the venue across the time window based on the estimated location of the object.

[0033] In one embodiment, a method for wireless-based occupancy detection is described. The method involves each of a plurality of transmitters transmitting their respective wireless signals over a wireless channel of a venue, and a receiver receiving the respective wireless signals from each of the plurality of transmitters over the wireless channel of the venue, wherein the received wireless signals are different from the transmitted wireless signals due to the wireless channel and the movement of the users when a user is present in the venue, and obtaining a time series (TSCI) of a plurality of channel information (CI) of the wireless channel, wherein each of the plurality of TSCIs is generated based on the received wireless signals from each of the plurality of transmitters, and calculating a plurality of autocorrelation function (ACF) segments based on the plurality of TSCIs. Each of the multiple ACF segments is calculated based on the CI of each of the multiple TSCIs, and generates a plurality of feature maps using a first deep learning model shared by the multiple ACF segments, wherein each of the multiple ACF segments is individually input into the first deep learning model to generate each of the plurality of feature maps, inputs the plurality of feature maps together into a second deep learning model to generate an aggregate representation that is independent of the number and location of the plurality of transmitters, calculates the probability of user presence based on the aggregate representation, and detects the presence of the user in the venue based on a threshold and the probability of user presence.

[0034] In another embodiment, a device for wireless-based occupancy detection is described. The device comprises a receiver and a processor. The receiver is configured to receive wireless signals from each of a plurality of transmitters via a wireless channel of the venue. The received wireless signals differ from the transmitted wireless signals due to the wireless channel and the user's movement when a user is present in the venue. The processor is configured to acquire a time series (TSCI) of multiple channel information (CI) of the wireless channel, where each of the multiple TSCIs is generated based on the received wireless signal from each of the multiple transmitters, and is configured to compute a plurality of autocorrelation function (ACF) segments based on the multiple TSCIs, where each of the multiple ACF segments is computed based on the CI of each of the multiple TSCIs, and is configured to generate a plurality of feature maps using a first deep learning model shared by the multiple ACF segments, where each of the multiple ACF segments is individually input into the first deep learning model to generate each of the multiple feature maps, and the multiple feature maps are input together into a second deep learning model to generate an aggregate representation independent of the number and location of the multiple transmitters, calculate the probability of user presence based on the aggregate representation, and detect the presence of the user in the venue based on a threshold and the probability of user presence.

[0035] In yet another embodiment, a system for wireless-based occupancy detection is described. This system comprises a plurality of transmitters, each configured to transmit its own wireless signal over a wireless channel of a venue, a receiver, and a processor. The receiver is configured to receive the respective wireless signals from each of the plurality of transmitters over the wireless channel of the venue. The received wireless signals differ from the transmitted wireless signals due to the wireless channel and the user's movement when a user is present in the venue. The processor is configured to acquire a time series (TSCI) of multiple channel information (CI) of the wireless channel, where each of the multiple TSCIs is generated based on the received wireless signal from each of the multiple transmitters, and is configured to compute multiple autocorrelation function (ACF) segments based on the multiple TSCIs, where each of the multiple ACF segments is computed based on the CI of each of the multiple TSCIs, and is configured to generate multiple feature maps using a first deep learning model shared by the multiple ACF segments, where each of the multiple ACF segments is individually input into the first deep learning model to generate each of the multiple feature maps, and the multiple feature maps are input together into a second deep learning model to generate an aggregate representation independent of the number and location of the multiple transmitters, calculate the probability of user presence based on the aggregate representation, and detect the presence of the user in the venue based on a threshold and the probability of user presence.

[0036] Other concepts relate to software for implementing this disclosure relating to wireless sensing, tracking, imaging, and occupancy detection. Additional novel features are described in part in the following description, and some will become apparent to those skilled in the art by examining the following and accompanying drawings, or can be learned by manufacturing or operating the embodiments. Novel features of this disclosure can be realized and achieved by practicing or using various aspects of the methods, means, and combinations described in the detailed embodiments discussed below. [Brief explanation of the drawing]

[0037] The methods, systems, and / or devices described herein will be further described in relation to exemplary embodiments. These exemplary embodiments will be described in detail with reference to the drawings. These embodiments are non-limiting exemplary embodiments, and similar reference numerals represent similar structures through some of the drawings.

[0038] [Figure 1] Figure 1 shows an exemplary framework of a system for wireless sensing using a basic model, according to several embodiments of the present disclosure.

[0039] [Figure 2] Figure 2 illustrates an exemplary process for training and running a foundational model for wireless sensing according to several embodiments of the present disclosure.

[0040] [Figure 3] Figure 3 illustrates an exemplary method for combining the determination of a wireless sensing task based on multiple links, according to several embodiments of the present disclosure.

[0041] [Figure 4] Figure 4 illustrates another exemplary method for combining the determination of a wireless sensing task based on multiple links, according to some embodiments of the present disclosure.

[0042] [Figure 5]Figure 5 illustrates an exemplary process for multi-task learning based on a foundational model, according to several embodiments of the present disclosure.

[0043] [Figure 6] Figure 6 shows exemplary masks used for contrast loss in some embodiments of the present disclosure.

[0044] [Figure 7] Figure 7 shows exemplary architectures of autoencoders according to several embodiments of the present disclosure.

[0045] [Figure 8] Figure 8 shows an exemplary process for applying a mask according to several embodiments of the present disclosure.

[0046] [Figure 9] Figure 9 shows a flowchart illustrating an exemplary method for wireless sensing using a basic model, according to several embodiments of the present disclosure.

[0047] [Figure 10] Figure 10 shows exemplary scenarios of wireless sensing in a venue according to several embodiments of the present disclosure.

[0048] [Figure 11] Figure 11 shows an exemplary floor plan and arrangement of wireless devices for wireless sensing according to some embodiments of the present disclosure.

[0049] [Figure 12] Figure 12 shows a flowchart illustrating an exemplary method for hybrid wireless plus-assisted fall detection based on wireless sensing, according to several embodiments of the present disclosure.

[0050] [Figure 13]Figure 13 shows an exemplary system for performing wireless sensing in a venue using multiple device groups with room-by-room deployments, according to some embodiments of the present disclosure.

[0051] [Figure 14] Figure 14 shows an exemplary system for performing wireless sensing in a venue using multiple device groups in a house-wide deployment, according to several embodiments of the present disclosure.

[0052] [Figure 15] Figure 15 shows a flowchart illustrating an exemplary method for performing wireless sensing in a venue using multiple device groups, according to some embodiments of the present disclosure.

[0053] [Figure 16] Figure 16 shows an exemplary floor plan for wireless sensing to display sensing motion statistics and analysis according to some embodiments of the present disclosure.

[0054] [Figure 17] Figure 17 shows a flowchart illustrating an exemplary method of a wireless sensing presentation system according to several embodiments of the present disclosure.

[0055] [Figure 18] Figure 18 shows a flowchart illustrating an exemplary method for performing a sensing procedure with a selective proxy, according to several embodiments of the present disclosure.

[0056] [Figure 19] Figure 19 shows an exemplary system for performing wireless sensing in a network of networks with inter-network sounding, according to several embodiments of the present disclosure.

[0057] [Figure 20]Figure 20 shows a flowchart illustrating an exemplary method for performing inter-network sounding according to several embodiments of the present disclosure.

[0058] [Figure 21] Figure 21 shows a flowchart illustrating an exemplary method for performing wireless sensing in a network of networks with inter-network sounding, according to some embodiments of the present disclosure.

[0059] [Figure 22A] A flowchart illustrates an exemplary method for performing wireless sensing using encoded addresses, according to some embodiments of this disclosure. [Figure 22B] A flowchart illustrates an exemplary method for performing wireless sensing using encoded addresses, according to some embodiments of this disclosure.

[0060] [Figure 23] Figure 23 shows exemplary scenarios of wireless sensing in a venue according to several embodiments of the present disclosure.

[0061] [Figure 24] Figure 24 shows an exemplary framework of a system for wireless positioning according to several embodiments of the present disclosure.

[0062] [Figure 25A] This disclosure illustrates exemplary multipath selection filtering according to several embodiments. [Figure 25B] This disclosure illustrates exemplary multipath selection filtering according to several embodiments.

[0063] [Figure 26A] This disclosure demonstrates exemplary performance of moving and distributing PCA components across CSI packets according to several embodiments of this disclosure. [Figure 26B]This disclosure demonstrates exemplary performance of moving and distributing PCA components across CSI packets according to several embodiments of this disclosure.

[0064] [Figure 27] Figure 27 shows a flowchart illustrating an exemplary method for performing wireless sensing with conversion component selection according to some embodiments of the present disclosure.

[0065] [Figure 28] Figure 28 shows a flowchart illustrating an exemplary method for converting CI of component N1 to TCI of component N2, according to several embodiments of the present disclosure.

[0066] [Figure 29] Figure 29 shows a flowchart of an exemplary method for selecting N3 TCICs from N2 components of a TCI, according to some embodiments of the present disclosure.

[0067] [Figure 30] Figure 30 shows a flowchart of exemplary methods, according to several embodiments of the present disclosure, for highlighting selected N3 TCICs and de-highlighting unselected N2-N3 TCICs.

[0068] [Figure 31A] A flowchart illustrates an exemplary method for calculating task results using enhanced CI, according to several embodiments of this disclosure. [Figure 31B] A flowchart illustrates an exemplary method for calculating task results using enhanced CI, according to several embodiments of this disclosure.

[0069] [Figure 32] Figure 32 shows a flowchart of an exemplary method for calculating task results based on multiple domain-based analyses, according to some embodiments of the present disclosure.

[0070] [Figure 33]Figure 33 shows a flowchart illustrating an exemplary method for constructing spatial diversity vectors to generate spatial domain analysis, according to several embodiments of the present disclosure.

[0071] [Figure 34] Figure 34 shows exemplary scenarios of wireless sensing in a venue according to several embodiments of the present disclosure.

[0072] [Figure 35] Figure 35 shows an exemplary floor plan and arrangement of wireless devices for wireless sensing according to some embodiments of the present disclosure.

[0073] [Figure 36] Figure 36 shows an exemplary framework of a system for fall detection according to several embodiments of the present disclosure.

[0074] [Figure 37] Figure 37 shows exemplary methods for iterative strain-class activation maps (ID-CAM) according to several embodiments of the present disclosure.

[0075] [Figure 38] Figure 38 shows a flowchart illustrating an exemplary method for performing wireless sensing with classifier probing and refinement according to some embodiments of the present disclosure.

[0076] [Figure 39] Figure 39 shows a flowchart illustrating an exemplary method for generating at least one selected perturbation input data according to some embodiments of the present disclosure.

[0077] [Figure 40] Figure 40 shows a flowchart illustrating an exemplary method for selecting specific selected perturbation input data according to some embodiments of the present disclosure.

[0078] [Figure 41] Figure 41 shows a flowchart of an exemplary method for retraining a classifier based on selected supplemental perturbation input data, according to some embodiments of the present disclosure.

[0079] [Figure 42] Figure 42 shows a flowchart illustrating an exemplary method for constructing input data according to some embodiments of the present disclosure.

[0080] [Figure 43] Figure 43 shows a flowchart illustrating an exemplary method for constructing perturbation input data based on reference input data, according to some embodiments of the present disclosure.

[0081] [Figure 44] Figure 44 shows a flowchart illustrating exemplary methods for correcting each marked component of reference input data according to some embodiments of the present disclosure.

[0082] [Figure 45] Figure 45 shows a flowchart of another exemplary method for correcting each marked component of reference input data according to some embodiments of the present disclosure.

[0083] [Figure 46] Figure 46 shows a flowchart illustrating an exemplary method for generating a subset of marked components according to several embodiments of the present disclosure.

[0084] [Figure 47] Figure 47 shows an exemplary framework of a wireless sensing system for detecting the presence of a child in a vehicle, according to several embodiments of the present disclosure.

[0085] [Figure 48] Figure 48 shows exemplary configurations and arrangements of wireless devices for in-vehicle wireless sensing according to some embodiments of the present disclosure.

[0086] [Figure 49] Figure 49 shows exemplary classification networks based on convolutional neural networks according to several embodiments of the present disclosure.

[0087] [Figure 50A] The present disclosure illustrates exemplary antenna configurations for in-vehicle wireless sensing according to several embodiments. [Figure 50B] The present disclosure illustrates exemplary antenna configurations for in-vehicle wireless sensing according to several embodiments. [Figure 50C] The present disclosure illustrates exemplary antenna configurations for in-vehicle wireless sensing according to several embodiments.

[0088] [Figure 51A] A flowchart illustrates an exemplary method for performing wireless sensing for detecting the presence of a child inside a vehicle, according to some embodiments of the present disclosure. [Figure 51B] A flowchart illustrates an exemplary method for performing wireless sensing for detecting the presence of a child inside a vehicle, according to some embodiments of the present disclosure.

[0089] [Figure 52] Figure 52 shows flowcharts of exemplary methods for calculating vibration reliability score (CS), dynamic time warping (DTW) CS, and peak CS according to some embodiments of the present disclosure.

[0090] [Figure 53] Figure 53 shows a flowchart of an exemplary method for detecting the number of horizontal streaks in a 2D transformation matrix, according to some embodiments of the present disclosure.

[0091] [Figure 54A] A flowchart illustrating an exemplary method for detecting a user's breathing behavior over a period of time, according to several embodiments of this disclosure, is shown. [Figure 54B] A flowchart illustrating an exemplary method for detecting a user's breathing behavior over a period of time, according to several embodiments of this disclosure, is shown.

[0092] [Figure 55] Figure 55 shows a flowchart illustrating an exemplary method for calculating the vibration coefficient in a transformation domain according to some embodiments of the present disclosure.

[0093] [Figure 56] Figure 56 shows a flowchart illustrating an exemplary method for calculating DTW costs according to several embodiments of the present disclosure.

[0094] [Figure 57] Figure 57 shows a flowchart illustrating an exemplary method for calculating a 2D matrix according to some embodiments of the present disclosure.

[0095] [Figure 58A] A flowchart illustrating exemplary methods for generating output probability scores for a fully connected network, according to some embodiments of this disclosure, is shown. [Figure 58B] A flowchart illustrating exemplary methods for generating output probability scores for a fully connected network, according to some embodiments of this disclosure, is shown.

[0096] [Figure 59A] A flowchart illustrates exemplary methods of wireless sensing for intruder detection with human-non-human detection, according to several embodiments of the present disclosure. [Figure 59B] A flowchart illustrates exemplary methods of wireless sensing for intruder detection with human-non-human detection, according to several embodiments of the present disclosure.

[0097] [Figure 60] Figure 60 shows a flowchart illustrating an exemplary method for running a neural network according to several embodiments of the present disclosure.

[0098] [Figure 61] Figure 61 shows a flowchart illustrating exemplary methods for detecting human and non-human motion according to several embodiments of the present disclosure.

[0099] [Figure 62] Figure 62 shows an exemplary framework of an encoder-decoder-based system for high-resolution imaging according to several embodiments of the present disclosure.

[0100] [Figure 63] Figure 63 shows an exemplary architecture of a generator in an encoder-decoder-based system for high-resolution imaging, according to several embodiments of the present disclosure.

[0101] [Figure 64] Figure 64 shows an exemplary architecture of a discriminator in an encoder-decoder-based system for high-resolution imaging, according to several embodiments of the present disclosure.

[0102] [Figure 65] Figure 65 shows an exemplary diagram of a system for CSI evaluation according to several embodiments of the present disclosure.

[0103] [Figure 66] Figure 66 shows a flowchart of an exemplary method for encoder-decoder-based high-resolution imaging according to several embodiments of the present disclosure.

[0104] [Figure 67] Figure 67 shows a flowchart of an exemplary method for generating output matrices for different layers of an encoder, according to some embodiments of the present disclosure.

[0105] [Figure 68]Figure 68 shows a flowchart of an exemplary method for generating an output imaging matrix according to some embodiments of the present disclosure.

[0106] [Figure 69] Figure 69 shows a flowchart illustrating an exemplary method for handling skipped connections according to some embodiments of the present disclosure.

[0107] [Figure 70] Figure 70 shows a flowchart of an exemplary method for quality assessment of channel measurements for wireless sensing, according to some embodiments of the present disclosure.

[0108] [Figure 71A] A flowchart illustrating exemplary methods for task-specific qualification according to several embodiments of this disclosure is shown. [Figure 71B] A flowchart illustrating exemplary methods for task-specific qualification according to several embodiments of this disclosure is shown.

[0109] [Figure 72] Figure 72 shows a flowchart illustrating an exemplary method for calculating a task-specific CI timing rate score according to several embodiments of the present disclosure.

[0110] [Figure 73] Figure 73 shows a flowchart of an exemplary method for calculating a CI timing regularity score according to some embodiments of the present disclosure.

[0111] [Figure 74] Figure 74 shows a flowchart illustrating an exemplary method for calculating the summation of multiple CI timing variations according to some embodiments of the present disclosure.

[0112] [Figure 75]Figure 75 shows a flowchart of an exemplary method for calculating the CI amplitude score according to some embodiments of the present disclosure.

[0113] [Figure 76] Figure 76 shows an exemplary framework of a system for wireless tracking according to several embodiments of the present disclosure.

[0114] [Figure 77] Figure 77 shows an exemplary position estimator in a wireless tracking system according to several embodiments of the present disclosure.

[0115] [Figure 78] Figure 78 shows an exemplary trajectory refinement module in a wireless tracking system according to several embodiments of the present disclosure.

[0116] [Figure 79] Figure 79 illustrates some embodiments of the present disclosure of indoor tracking based on wireless signals.

[0117] [Figure 80] Figure 80 shows exemplary walking scenarios and corresponding proximity metrics according to several embodiments of the present disclosure.

[0118] [Figure 81] Figure 81 shows exemplary floor plans and corresponding location signature matching over time according to several embodiments of the present disclosure.

[0119] [Figure 82] Figure 82 shows an overview of a system for wireless tracking according to several embodiments of the present disclosure.

[0120] [Figure 83] Figure 83 shows exemplary floor plans for various scenarios according to several embodiments of the present disclosure.

[0121] [Figure 84] Figure 84 shows a flowchart illustrating an exemplary method for wireless tracking according to several embodiments of the present disclosure.

[0122] [Figure 85] Figure 85 shows an exemplary framework of a wireless-based occupancy detection system according to several embodiments of the present disclosure.

[0123] [Figure 86] Figure 86 shows an exemplary system for detecting the presence of a child, according to several embodiments of the present disclosure.

[0124] [Figure 87] Figure 87 shows exemplary antenna configurations for detecting the presence of children according to several embodiments of the present disclosure.

[0125] [Figure 88] Figure 88 shows exemplary antenna configurations and data acquisition setups according to several embodiments of the present disclosure.

[0126] [Figure 89] Figure 89 shows exemplary confusion matrices for classifying children, adults, and empty (absent) according to some embodiments of the present disclosure.

[0127] [Figure 90] Figure 90 shows exemplary ROC curves generated from different baseline network architectures according to several embodiments of the present disclosure.

[0128] [Figure 91] Figure 91 shows an exemplary system for deep learning-based occupation detection according to several embodiments of the present disclosure.

[0129] [Figure 92]Figure 92 shows exemplary processes for training and retraining a deep learning model according to some embodiments of the present disclosure.

[0130] [Figure 93] Figure 93 shows an exemplary block diagram of a first wireless device of a system for wireless sensing, tracking, imaging, and occupancy detection according to some embodiments of the present disclosure.

[0131] [Figure 94] Figure 94 shows an exemplary block diagram of a second wireless device of a system for wireless sensing, tracking, imaging, and occupancy detection according to some embodiments of the present disclosure.

[0132] [Figure 95] Figure 95 shows a flowchart illustrating an exemplary method for wireless-based occupation detection according to several embodiments of the present disclosure.

[0133] [Figure 96] Figure 96 shows a flowchart of another exemplary method for wireless-based occupation detection according to some embodiments of the present disclosure.

[0134] [Figure 97] Figure 97 shows an exemplary process for intrusion detection according to several embodiments of the present disclosure.

[0135] [Figure 98] Figure 98 shows an exemplary long-short-term memory (LSTM) neural network for intrusion detection according to several embodiments of the present disclosure.

[0136] [Figure 99] Figure 99 shows an exemplary fusion shot learning process according to several embodiments of the present disclosure.

[0137] [Figure 100]Figure 100 shows an exemplary data input preparation process for motion source recognition according to several embodiments of the present disclosure. [Modes for carrying out the invention]

[0138] The symbol " / " disclosed herein means "and / or". For example, "A / B" means "A and / or B". In some embodiments, methods / devices / systems / software for a wireless monitoring system are disclosed. A time series of channel information (CI) for a wireless multipath channel is obtained using a processor, memory communicably coupled to the processor, and an instruction set stored in memory. The time series of CI (TSCI) can be extracted from radio (sounding) signals transmitted through a channel in a venue from a first type heterogeneous wireless device (e.g., a wireless transmitter (TX), a "bot" device) to a second type heterogeneous wireless device (e.g., a wireless receiver (RX), an "origin" device). The channel is affected by the representation / movement of objects in the venue. Object / representation / movement characteristics / spatiotemporal information (STI) / motion information (MI) can be calculated / monitored based on the TSCI. A task (e.g., sensing target) can be performed based on the characteristics / STI / MI. A task-related presentation can be generated in a user interface (UI) on the user's device. Characteristics / STI / MI can be Motion Index Value (MIV) / Activity Data (AD) / Motion Data (MD) / Motion Sensing Data (MSD) / Motion Detection Data (MDD) / Motion Score (MS) / Motion Statistics (MS2).

[0139] Expressions may include arrangement, arrangement of movable parts, position / velocity / acceleration / attitude / orientation / direction / identifiable location / area / presence / spatial coordinates, static expression / presentation / state / size / length / width / height / angle / scale / curve / surface / area / volume / pose / posture / sign / body language, dynamic expression / movement / sequence / movement / activity / behavior / gesture / gait / extension / contraction / distortion / deformation, bodily expression (e.g., head / face / eyes / mouth / tongue / hair / voice / neck / limbs / arms / hands / legs / feet / muscles / movable parts), surface expression / shape / texture / material / color / electromagnetic (EM) properties / visual pattern / moisture level / reflectivity / transparency / flexibility, material properties (e.g., biological tissue / hair / fabric / metal / wood / leather / plastic / artificial material / solid / liquid / gas / temperature), changes in expression, and / or any combination thereof.

[0140] A wireless multipath channel may include communication channels, analog frequency channels (e.g., carrier frequencies around 700 / 800 / 900 MHz or around 1.8 / 1.9 / 2.4 / 3 / 5 / 6 / 27 / 60 / 70+ GHz), coded channels (e.g., those in CDMA), and / or wireless / cellular network / system channels (e.g., WLAN, WiFi, mesh, 4G / LTE / 5G / 6G / 7G / 8G / 9G / 10G, Bluetooth, Zigbee, UWB, RFID, microwave). It may include multiple channels, which may be continuous (e.g., adjacent / overlapping bands) or discontinuous (e.g., non-overlapping bands, 2.4 GHz / 5 GHz). While channels are used to transmit wireless signals and perform sensing measurements, data (e.g., TSCI / features / components / characteristics / STI / MI / analysis / task output, auxiliary / non-sensing data / network traffic) may be communicated / transmitted over the channels.

[0141] A wireless signal may include a series of probe signals. It could be any of the following: EM radiation, radio frequency (RF) / optical / band-limited / baseband signals, signals in licensed / unlicensed / ISM bands, MIMO transmissions, sensing transmissions, or wireless / mobile / cellular / optical communications / network / mesh / downlink / uplink / unicast / multicast / broadcast signals. It may comply with standards / protocols (e.g., WLAN, WWAN, WPAN, WBAN, international / national / industry / de facto, IEEE / 802 / 802.11 / 15 / 16, WiFi, 802.11n / ac / ax / be / bf, 3G / 4G / LTE / 5G / 6G / 7G / 8G / 9G / 10G, 3GPP® / Bluetooth® / BLE / Zigbee / NFC / RFID / UWB / WiMax). A probe signal may include any of the following: protocol / standard / beacon / pilot / sounding / excitation / illumination / handshake / synchronization / reference / source / motion probe / detection / sensing / management / control / data / null data / beacon / pilot / request / response / association / re-association / dis-association / authentication / action / report / pole / announcement / extension / query / acknowledgment frame / packet / signal, and / or null data frame (NDP) / RTS / CTS / QoS / CF-Poll / CF-Ack / block acknowledgment / reference / training / synchronization. It may include line-of-sight (LOS) / non-LOS components (or path / link). It may have embedded data. A probe signal may be replaced by (or embedded within) a data signal. Each frame / packet / signal may include a preamble / header / payload. It may include a training sequence, short (STF) / long (LTF) training fields, L-STF / L-LTF / L-SIG / HE-STF / HE-LTF / HE-SIG-A / HE-SIG-B, and a channel estimation field (CEF). It can be used to wirelessly transmit power from a first-type device to a second-type device. The sounding rate of the signal can be adjusted to control the amount of power transmitted. The probe signal may be transmitted in bursts.

[0142] TSCI can be extracted / acquired from radio signals (e.g., by ICs / chips) at the layers of Type 2 devices (e.g., OSI reference model layers, PHY / MAC / Data Link / Logical Link Control / Network / Transport / Session / Presentation / Application Layer, TCP / IP / Internet / Link Layer). It can be extracted from received radio / derived signals. It may include radio sensing measurements obtained in communication protocols (e.g., radio / cellular communication standards / networks, 4G / LTE / 5G / 6G / 7G / 8G / 9G / 10G, WiFi, IEEE 802.11 / 11bf / 15 / 16). Each CI can be extracted from a probe / sounding signal and associated with a timestamp. TSCI can be associated with start / stop time / duration / amount of CI / sampling / sounding frequency / period. Motion detection / sensing signals can be recognized / identified based on probe signals. TSCI can be stored / retrieved / accessed / preprocessed / processed / postprocessed / adjusted / analyzed / monitored. TSCI / features / components / characteristics / STI / MI / analysis / task results can be communicated to edge / cloud servers / Type 1 / Type 2 / hubs / data aggregators / other devices / systems / networks.

[0143] Type 1 / Type 2 devices may include components (hardware / software) such as electronic equipment / chips / integrated circuits (ICs) / RF circuits / antennas / modems / TX / RX / transceivers / RF interfaces (e.g., 2.4 / 5 / 6 / 27 / 60 / 70+GHz radio / front / backhaul radio) / networks / interfaces / processors / memory / modules / circuits / boards / software / firmware / connectors / structures / enclosures / housings / structures. They may also include access points (APs) / base stations / mesh / routers / repeaters / hubs / wireless stations / clients / terminals / "origin satellites" / "tracker bots" and / or Internet of Things (IoT) / consumer electronics / wearables / accessories / peripherals / furniture / amenities / gadgets / vehicles / modules / wireless / unicast / multicast / broadcast / nodes / hubs / targets / sensors / portables / mobile / cellular / communication / motion detection / source / destination / standards-compliant devices. It may include additional attributes such as auxiliary functions / network connectivity / purpose / brand / model / appearance / form / shape / color / material / specifications. It may be heterogeneous because the above (e.g., components / device type / additional attributes) may differ for each different Type 1 (or Type 2) device.

[0144] Type 1 / Type 2 devices may or may not be authenticated / associated / colocated. They may be the same device. Type 1 / Type 2 / portable / neighboring / other devices, sensing / measurement sessions / links between them, and / or objects / representations / motions / characteristics / STI / MI / tasks may be associated with an identity / identification / identifier (ID) such as a UUID, associated / unassociated STA ID (ASID / USID / AID / UID). A Type 2 device may passively observe / monitor / receive radio signals from a Type 1 device without establishing a connection with it (e.g., association / authentication / handshake) or requesting service from it. Type 1 / Type 2 devices may move with the tracked object / other object.

[0145] A Type 1 (TX) device may function as a Type 2 (RX) device temporarily, sporadically, continuously, repeatedly, interchangeably, alternately, simultaneously, concurrently, or in parallel, and vice versa. A Type 1 device may be a Type 2 device. A device may function as a Type 1 or Type 2 device temporarily, sporadically, continuously, repeatedly, simultaneously, concurrently, or in parallel. Multiple radio nodes may exist, each being a Type 1 or Type 2 device. When two nodes transmit and receive / communicate radio signals, a TSCI may be obtained between them. Object properties / STI / MI may be monitored individually based on a single TSCI or collectively based on multiple TSCIs.

[0146] The movement / representation of an object can be monitored actively using Type 1 / Type 2 devices that move with the object (e.g., wearable devices / automated guided vehicles / AGVs), or passively using Type 1 / Type 2 devices that do not move with the object (e.g., both types of stationary devices).

[0147] Tasks may be performed with or without reference to the trained / collected / processed / computed / transmitted / stored reference / trained / initial database / profile / baseline during the training phase. The database may be retrained / updated / reset.

[0148] A presentation may include UI / GUI / text / messages / forms / web pages / visual / images / video / graphics / animations / geometric / symbols / emojis / codes / color / shading / sound / music / voice / audio / mechanical / gestures / vibrations / haptic presentations. Characteristics / STI / MI / task results / time series of another quantity may be displayed / presented in the presentation. Any calculation may be performed / shared by a processor (or logic unit / chip / IC) / Type 1 / Type 2 / user / neighbor / another device / local / edge / cloud server / hub / data / signal analysis subsystem / sensing initiator / responder / SBP initiator / responder / AP / non-AP. A presentation may include any of the following: monthly / weekly / daily / simplified / detailed / cross-sectional / small / large / form factor / color coding / comparison / summary / web view, animation / voice announcement / repeating motion / periodicity of representation / repetitive characteristics.

[0149] Multiple Type 1 (or Type 2) devices may interact with a Type 2 (or Type 1) device. Multiple Type 1 (or Type 2) devices may be synchronous / asynchronous and / or use the same / different channels / sensing parameters / settings (e.g., sounding frequency / bandwidth / antenna). A Type 2 device may receive another signal from a Type 1 / another Type 1 device. A Type 1 device may transmit another signal to a Type 2 / another Type 2 device. The radio signals transmitted (or received) by them may be sporadic / temporary / continuous / repeated / synchronous / simultaneous / parallel / contemporary. They may operate independently / cooperatively. Their data (e.g., TSCI / features / characteristics / STI / MI / intermediate task results) may be processed / monitored / analyzed independently or jointly / cooperatively.

[0150] Any device may operate based on some state / internal state / system state. Devices may communicate directly or via another / nearby / portable device / server / hub device / cloud server. A device / system may be associated with one or more users, along with its associated settings. Settings may be selected / selected / pre-programmed / modified / adjusted / corrected / changed over time. Methods may be executed / performed in the indicated / alternative order. Steps may be executed / repeated / reiterated in parallel. Users may include humans / adults / elderly / males / females / teenagers / children / babies / pets / animals / living organisms / machines / computer modules / software. Steps / operations / processes may differ for each different device (e.g., based on location / orientation / direction / role / user-related characteristics / settings / configuration / available resources / bandwidth / power / network connectivity / hardware / software / processor / coprocessor / memory / battery life / antenna / directional antenna / power settings / device parameters / characteristics / conditions / situations / states). Any / all devices may be controlled / regulated by a processor (e.g., Type 1 / Type 2 / Nearby / Portable / Another device / Server / Associated with a specified source). Some devices may be physically inside / part of / attached to a common device.

[0151] A Type 1 (or Type 2) device may be able to wirelessly couple with multiple Type 2 (or Type 1) devices. A Type 1 (or Type 2) device may be directed / controlled to switch / establish wireless coupling (e.g., association / authentication) from one Type 2 (or Type 1) device to another Type 2 (or another Type 1) device. Switching may be controlled by a server / hub device / processor / Type 1 device / Type 2 device. The wireless channel may differ before and after switching. A second wireless signal may be transmitted between a Type 1 (or Type 2) device and a second Type 2 (or second Type 1) device through a second channel. A second TSCI of the second channel may be extracted / acquired from the second signal. The first / second signals, first / second channels, first / second Type 1 devices, and / or first / second Type 2 devices may be identical / similar / collocated.

[0152] A Type 1 device may transmit / broadcast a radio signal to multiple Type 2 devices with or without establishing a connection (association / authentication) with each individual Type 2 device. It may transmit to a specific / common MAC address, which could be the MAC address of some device (e.g., a dummy receiver). Each Type 2 device may tune to a specific MAC address to receive the radio signal. A specific MAC address may be associated with a venue, which may be recorded in the association table of an association server (e.g., a hub device). A venue may be identified by a Type 1 / Type 2 device based on the radio signal received at a specific MAC address.

[0153] For example, a Type 2 device may be moved to a new venue. A Type 1 device may be newly set up at the venue so that the Type 1 and Type 2 devices do not recognize each other. During setup, the Type 1 device may be instructed / guided / directed / controlled (e.g., by a dummy receiver, hardware pin configuration / connection, stored configuration, local configuration, remote configuration, downloaded configuration, hub device, and / or server) to send radio signals (e.g., a series of probe signals) to a specific MAC address. Upon power-up, the Type 2 device may scan for probe signals according to a table of MAC addresses (e.g., different MAC addresses used for different venues such as residential / office / enclosure / floor / multi-story building / shop / airport / mall / stadium / hall / station / subway / site / area / zone / region / district / city / country / continent) that can be used for broadcasting in different locations (e.g., stored in a specified source, server, hub device, cloud server). When the Type 2 device detects a radio signal sent to a specific MAC address, it can use the table to identify the venue.

[0154] Channels may be selected from a set of candidate / selectable / acceptable channels. Candidate channels may be associated with different frequency bands / bandwidth / carrier frequencies / modulation / radio standards / coding / encryption / payload characteristics / network / ID / SSID / characteristics / configurations / parameters. A specific MAC address / selected channel may be changed / adjusted / modified / corrected over time (e.g., according to a timetable / rules / policies / modes / conditions / situations / changes). Selection / modification may be based on availability / conflict / traffic patterns / shared channels / inter-channel interference / effective bandwidth / random selection / pre-selected lists / plans. This may be done by a server (e.g., a hub device). They may be communicated (e.g., from or to a Type 1 / Type 2 / hub / another device / local / edge / cloud server).

[0155] A wireless connection (e.g., association / authentication) can be established between a Type 1 device and a nearby / portable / other device (e.g., using a signal handshake). The Type 1 device may transmit a first handshake signal (e.g., a sounding frame / probe signal / transmit request RTS) to the nearby / portable / other device. The nearby / portable / other device may respond to the first signal by transmitting a second handshake signal (e.g., command / transmittable / CTS) to the Type 1 device, triggering the Type 1 device to transmit / broadcast a wireless signal to multiple Type 2 devices without establishing a connection with the Type 2 devices. The second handshake signal may be a response / acknowledgment (e.g., ACK) to the first handshake signal. The second handshake signal may contain information about the venue / Type 1 device. The nearby / portable / other device may be a dummy device with the purpose (e.g., primary purpose, secondary purpose) of establishing a wireless connection with the Type 1 device, receiving the first signal, or transmitting the second signal. The nearby / portable / another device may be physically attached to the Type 1 device.

[0156] In another example, a nearby / portable / other device may send a third handshake signal to a Type 1 device, triggering the Type 1 device to broadcast the signal to multiple Type 2 devices without establishing a connection with them. The Type 1 device may respond to the third signal by sending a fourth handshake signal to another device.

[0157] A nearby / portable / another device can be used to trigger multiple Type 1 devices to broadcast. It may have multiple RF circuits to trigger multiple transmitters in parallel. Triggering can be sequential / partially sequential / partially / fully parallel. Parallel triggering can be achieved using additional devices to perform similar triggering in parallel with the nearby / portable / another device. After establishing a connection with a Type 1 device, the nearby / portable / another device can suspend / stop communication with the Type 1 device. It may enter an inactive / hibernation / sleep / standby / low power / off / power-down mode. Suspended communication can be resumed. The nearby / portable / another device may have a specific MAC address, and a Type 1 device may transmit signals to that specific MAC address.

[0158] A (first) radio signal may be transmitted at a first venue, through a first channel, by a first antenna of a first type device to some first second type device. A second radio signal may be transmitted at a second venue, through a second channel, by a second antenna of a first type device to some second second type device. The first / second signals may be transmitted at first / second (sounding) rates, respectively, and possibly to first / second MAC addresses, respectively. Any first / second channels / signals / rates / MAC addresses / antennas / second type devices may be identical / different / synchronous / asynchronous. The first / second venues may have identical / different sizes / shapes / multipath characteristics. The immediate areas around the first / second venues / first / second antennas may overlap. The first / second channel / signal may be WiFi+LTE (one WiFi, the other LTE), or WiFi+WiFi, or WiFi(2.4GHz)+WiFi(5GHz), or WiFi(5GHz, channel=a1, BW=a2)+WiFi(5GHz / channel=b1, BW=b2). Any first / second item (e.g., channel / signal / rate / MAC address / antenna / first type / second type device) may change / adjust / modify / change over time (e.g., based on a timetable / rules / policies / modes / conditions / situations / other changes).

[0159] Each Type 1 device can be a signal source for multiple Type 2 devices (i.e., it sends its respective probe signal to each Type 2 device). Each Type 2 device can asynchronously select a Type 1 device from among all Type 1 devices as its signal source. TSCI can be obtained by each Type 2 device from a series of probe signals from each Type 1 device. A Type 2 device can select a Type 1 device as its signal source (e.g., initially) from among all Type 1 devices based on information such as the identity / identification / identifier of the Type 1 / Type 2 device, task, past signal sources, history, characteristics, signal strength / quality, thresholds for switching signal sources, and / or user / account / profile / access information / parameters / inputs / requirements / criteria.

[0160] A database of available / candidate Type 1 (or Type 2) devices may be initialized / maintained / updated by Type 2 (or Type 1) devices. A Type 2 device may receive radio signals from multiple candidate Type 1 devices. It may select its Type 1 device (i.e., signal source) based on any of the following: signal quality / strength / regularity / channel / traffic / characteristics / attributes / state / task requirements / training task results / MAC address / identity / identifier / past signal sources / history / user instructions / other considerations.

[0161] Undesirable / bad / insufficient / problematic / unsatisfactory / unacceptable / unbearable / defective / harsh / undesirable / inappropriate / missing / inferior / inappropriate conditions may occur if (1) the timing between adjacent probe signals in the received radio signal becomes irregular and deviates from the agreed sounding rate (e.g., time perturbation exceeding the acceptable range), and / or (2) the processed / signal strength of the received signal is too weak (e.g., below the third threshold or below the fourth threshold for a significant proportion of time), where processing includes any low-pass / band-pass / high-pass / median / moving / weighted average / linear / nonlinear / smoothing filtering. Any thresholds / percentages / parameters may change over time. Such conditions may occur if the first / second type devices move further away or if the channel becomes congested.

[0162] Some settings (e.g., Type 1-Type 2 device pairing / signal source / network / association / probe signal / sounding rate / method / channel / bandwidth / system state / TSCI / TSMA / task / task parameters) may be changed / altered / adjusted / modified. Changes may follow timetables / rules / policies / modes / conditions (e.g., undesirable conditions) / other changes. For example, the sounding rate may normally be 100Hz, but may be changed to 1000Hz under harsh conditions or 1Hz under low power / standby conditions.

[0163] Settings may change based on task requirements (e.g., normally 100Hz, temporarily 1000Hz for 20 seconds). In a task, instantaneous systems may be adaptively / dynamically associated with classes / states / conditions (e.g., low / normal / high priority / urgent / critical / normal / privileged / non-subscription / subscription / paid / free). Settings (e.g., sounding rate) may be adjusted accordingly. Changes may be controlled by servers / hubs / Type 1 / Type 2 devices. Scheduled changes may be made according to a timetable. Changes may be made immediately when an emergency is detected, or incrementally when an ongoing condition is detected.

[0164] Characteristics / STI / MI can be monitored / analyzed individually based on TSCIs associated with a specific Type 1 / Type 2 device pair, jointly based on multiple TSCIs associated with multiple Type 1 / Type 2 device pairs, jointly based on any TSCI associated with a specific Type 2 device and any Type 1 device, jointly based on any TSCI associated with a specific Type 1 device and any Type 2 device, or comprehensively based on any TSCI associated with any Type 1 / Type 2 device.

[0165] Classifiers / classification / recognition / detection / estimation / projection / feature extraction / processing / filtering may be applied (e.g., to CI / CI features / characteristics / STI / MI) and / or training / retraining / updating. In the training stage, training may be performed based on multiple training TSCIs of some training radio multipath channel, or characteristics / STI / MI calculated from training TSCIs, which are obtained from training radio signals transmitted from a training type 1 device and received by a training type 2 device. Retraining / updating may be performed in the operation stage based on the training TSCI / current TSCI. Multiple classes (e.g., grouping / category / event / movement / activity / object / location) may exist associated with a venue / domain / zone / location / environment / residential / office / building / warehouse / facility / object / representation / movement / transportation / process / event / manufacturing / assembly line / maintenance / repair / navigation / object / emotional / mental / state / condition / stage / gesture / gait / action / movement / existence / transportation / daily / activity / history / event.

[0166] Classifiers can include linear / nonlinear / binary / multiclass / Bayesian classifiers / Fisher linear discriminant analysis / logistic regression / Markov chains / Monte Carlo / deep / neural networks / perceptrons / self-organizing maps / boosting / meta-algorithms / decision trees / random forests / genetic programming / kernel learning / KNN / support vector machines (SVM).

[0167] Feature extraction / projection may include subspace projection / principal component analysis (PCA) / independent component analysis (ICA) / vector quantization / singular value decomposition (SVD) / eigendecomposition / eigenvalues / time / frequency / orthogonal / nonorthogonal decomposition, processing / preprocessing / postprocessing. Each CI may contain multiple components (e.g., combinations of vectors / complex values). Each component may be preprocessed to give amplitude / phase or a function thereof.

[0168] Features include: feature extraction / projection output, amplitude / magnitude / phase / energy / power / intensity / luminance, presence / absence / proximity / likelihood / histogram, time / period / duration / frequency / component / decomposition / projection / bandwidth, local / global / maximum / minimum / zero crossing, repetitive / periodic / typical / habitual / one-time / atypical / sudden / exclusive / evolving / transient / changing / time / related / correlated features / patterns / trends / profiles / events / tendencies / slope / behavior, cause and effect / short-term / long-term / correlation / statistics / frequency / period / duration, motion / movement / location / map / coordinates / height / velocity / acceleration / angle / rotation / size / volume, suspicious / dangerous / warning event / warning / belief / proximity / collision, tracking / breathing / heart rate / Gait / Action / Event / Statistical / Hourly / Daily / Weekly / Monthly / Yearly Parameters / Statistics / Analysis, Happiness / Health / Disease / Medical Statistics / Analysis, State / Condition / Situation / Disease / Early / Instantaneous / Contemporary / Delayed Indicators / Suggestions / Signs / Indicators / Validators / Detectors / Symptoms, Babies / Patients / Machines / Devices / Temperature / Vehicles / Parking Lots / Venues / Lifts / Elevators / Spaces / Roads / Fluid Flow / Residential / Rooms / Offices / Homes / Buildings / Warehouses / Storage / Systems / Ventilation / Fans / Pipes / Ducts / People / Humans / Cars / Boats / Trucks / Airplanes / Drones / Downtown / Crowds / Impulsive Events / Cyclostationary / Environment / Vibration / Materials / Surfaces / 3D / 2D / Local / Global, and / or other measurable quantities / variables. Features may include a monotonic function of features, or a sliding aggregate of features in a sliding window.

[0169] Training may include AI / machine / deep / supervised / unsupervised / discriminative training / autoencoder / linear discriminant analysis / regression / clustering / tagging / labeling / Monte Carlo computation.

[0170] In the operational stage, the current events / movements / representations / objects in the venue at the current time can be classified by applying a classifier to the current TSCI / characteristics / STI / MI obtained from the current radio signal received by Type 2 devices in the venue from Type 1 devices in the venue. If there are multiple Type 1 / Type 2 devices, some / all of them (or their locations / antenna locations) may be substitutes for the corresponding training Type 1 / Type 2 devices (or locations / antenna locations). The Type 1 / Type 2 devices / signals / channels / venues / objects / movements may be identical / different from the corresponding training entities. The classifier may be applied to a sliding window. The current TSCI / characteristics / STI / MI may be augmented by the training TSCI / characteristics / STI / MI (or fragments / extracts) to bootstrap the classifier / classifier.

[0171] The first section / segment (with a first duration / start / end time) of the first TSCI (associated with the first type-to-second type device pair) may be aligned (e.g., using dynamic time stretching / DTW / matched filtering, possibly based on some mismatch / distance / similarity score / cost, or correlation / autocorrelation / cross-correlation) with the second section / segment (with a second duration / start / end time) of the second TSCI (associated with the second type-to-second type device pair), and each CI in the first section is mapped to a CI in the second section. The first / second TSCIs may be preprocessed. Some similarity score (on a component / item / link / segment basis) may be calculated. The similarity score may include any of mismatch / distance / similarity score / cost. Component-level similarity scores can be calculated between the components of the first item (CI / feature / characteristic / STI / MI) in the first section and the corresponding components of the corresponding mapped item (second item) in the second section. Item-level similarity scores can be calculated between the first and second items (e.g., based on the aggregation of the corresponding component-level similarity scores). Aggregations may include sum / weighted sum, product / weighted product, weighted mean / robust mean / trimmed mean / arithmetic / geometric / harmonic mean, median / mode, percentile, or any other aggregation of the above. Link-level similarity scores can be calculated between the first and second items associated with the links (TX-RX antenna pairs) of the first and second first-type-second-type device pairs (e.g., based on the aggregation of the corresponding item-level similarity scores). Segment-level similarity scores can be calculated between the first and second segments (e.g., based on the aggregation of the corresponding link-level similarity scores). The first and second segments may be sliding.

[0172] In DTW, the constraints can be satisfied by the first / second segment, the first / second item, another first (or second) item of the first (or second) segment, or any corresponding timestamp / duration / difference / derivative function. The time difference between the first / second items can be constrained (e.g., upper / lower bounds are imposed). The first (or second) section can be the entire first (or second) TSCI. The duration / start / end times of the first / second can be the same or different.

[0173] In one example, the first / second Type 1-Type 2 device pair may be identical, and the first / second TSCIs may be identical / different. If different, the first / second TSCIs may include pairs of current / reference, current / current, or reference / reference TSCIs. For "current / reference," the first TSCI may be the current TSCI obtained in the operational stage, and the second TSCI may be the reference TSCI obtained in the training stage. For "reference / reference," the first / second TSCIs may be two TSCIs obtained during the training stage (e.g., for two training events / states / classes). For "current / current," the first / second TSCIs may be two TSCIs obtained during the operational stage (e.g., associated with two different antennas or two measurement setups). In another example, the first / second Type 1-Type 2 device pair may be different, but may share a common device (either Type 1 or Type 2).

[0174] Aligned first / second segments (or their respective parts) can be represented as first / second vectors. A part may contain all items (for "segment units"), or all items associated with a TX-RX link (for "link units"), or one item (for "item units"), or one component (for "component units"). The similarity score may include any combination / aggregation / function of dot product / correlation / autocorrelation / correlation index / covariance / discriminant score / distance / Euclidean / absolute / L_k / weighted distance (between the first / second vectors). The similarity score can be normalized by vector length. Parameters derived from the similarity score can be modeled by a statistical distribution. The scale / location / other parameters of the statistical distribution can be estimated.

[0175] It should be noted that multiple sliding segments may exist. The classifier may be applied to a first / second pair of sliding segments to obtain a provisional classification result. It may associate the current event with a particular class based on one pair of segments / provisional classification results, or based on multiple pairs of segments / provisional classification results (e.g., if the similarity score is N times consecutively, or sufficiently high / low, or if it is the most frequent / least frequent, dominant (e.g., largest / smallest / dominant / unparalleled / most important / superior), or sufficiently important (e.g., above / below some threshold) among all candidate classes).

[0176] Channel information (CI) includes signal strength / amplitude / phase / timestamp / spectral power measurement / modem parameters / dynamic beamforming information / beamforming report / dynamic imaging information / channel representation information (CRI) / spatial map from dynamic beam processing / transfer function components / radio state / measurable variables / sensing data / measurements / coarse / fine layer information (e.g., PHY / MAC / data link layer) / digital gain / RF filter / front-end switch / DC offset / correction / IQ compensation settings / environmental influence on radio signal propagation / channel input to output conversion / environmental stability This may include any of the following: behavior / state profile / wireless channel measurement / received signal strength index (RSSI) / channel state information (CSI) / channel impulse response (CIR) / channel frequency response (CFR) / channel response (CR) / frequency component (e.g., subcarrier) characteristics / channel characteristics / channel filter response, auxiliary information, data / meta / user / account / access / security / session / status / monitoring / device / network / household / neighbor / environment / real-time / sensor / stored / encrypted / compressed / protected data, identity / identifier / identification.

[0177] Each CI can be associated with a timestamp / arrival time / frequency bandwidth / signature / phase / amplitude / trend / characteristics, frequency-like characteristics, time / frequency / time-frequency domain elements, and orthogonal / non-orthogonal decomposition characteristics of the signal through the channel. The timestamps of TSCIs can be irregular and can be corrected to be regular (e.g., by interpolation / resampling) at least over a sliding time window.

[0178] A TSCI may be, or include, a link-unit TSCI associated with the antennas of a Type 1 device and a Type 2 device. For a Type 1 device having M antennas and a Type 2 device having N antennas, there may be MN link-unit TSCIs.

[0179] CI / TSCI can be pre-processed / processed / post-processed / stored / retrieved / transmitted / received. Some modem / radio state parameters can be kept constant. Modem parameters can be applied to a radio subsystem and can represent radio states. Motion detection signals (e.g., baseband signals, and then decoded / demodulated packets) can be obtained by processing (e.g., down-converting) radio signals (e.g., RF / WiFi / LTE / 5G / 6G signals) by the radio subsystem using radio states represented by stored modem parameters. Modem parameters / radio states can be updated (e.g., using previous modem parameters / radio states). Both previous and updated modem parameters / radio states can be applied in the radio subsystem (e.g., to process signals / decode data). In the disclosed system, both can be acquired / compared / analyzed / processed / monitored.

[0180] Each CI may contain N1 CI components (CICs) (e.g., time / frequency domain components, decomposed components), each with a corresponding CIC index. Each CIC may contain real / imaginary / complex numbers, magnitude / phase / boolean values / flags, and / or any combination / subset. Each CI may contain a vector / matrix / set / collection of CICs. The CICs of a TSCI associated with a particular CIC index may form a CIC time series. A TSCI may be divided into a time series (TSCIC) of N1 CICs, each associated with its respective CIC index. Characteristics / STI / MI may be monitored based on the TSCIC. Some TSCIC may be selected for further processing based on some criterion / cost function / signal quality indicator (e.g., SNR, interference level).

[0181] Multiple multi-component properties / STI / MI of TSCICs (e.g., two components with indices 6 and 7, or three components indexed at 6, 7, and 10) can be calculated. In particular, a k-component property can be a function of k TSCICs, each having k corresponding CIC indices. When k=1, it is a single-component property and can constitute / form a one-dimensional (1D) function as the CIC index spans all possible values. When k=2, a two-component property can constitute / form a 2D function. In special cases, it may depend only on the difference between two indices. In such cases, it can constitute a 1D function. A sum property can be calculated based on one or more multi-component properties (e.g., weighted average / aggregate). The properties / STI / MI of an object / motion / representation can be monitored based on any multi-component property / sum property.

[0182] Characteristics / STI / MI are instantaneous / short-term / long-term / historical / repeated / repeat / reproducible / recursive / periodic / pseudoperiodic / regular / habitual / gradual / average / initial / final / current / past / future / predicted / changing / deviation / change / time / frequency / orthogonal / nonorthogonal / transform / decomposition / deterministic / stochastic / probable / dominant / major / prominent / representative / characteristic / important / unimportant / indicative / common / averaged / shared / typical / prototypical / persistent / abnormal / sudden / impulsive / abrupt / unusual / non Representative / Atypical / Suspicious / Dangerous / Warning / Evolving / Transient / One-time Quantity / Characteristics / Analysis / Features / Information, Cause and Effect, Correlation Indicator / Score, Auto / Cross-correlation / Covariance, Autocorrelation Function (ACF), Spectrum / Spectrogram / Power Spectral Density, Time / Frequency Function / Transformation / Projection, Initial / Final / Temporal / Change / Trend / Pattern / Predisposition / Slope / Behavior / Activity / History / Profile / Event, Location / Place / Localization / Spatial Coordinates / Change on Map / Path / Navigation / Tracking, Linear / Rotation / Horizontal / Vertical / Position / Distance / Displacement / Height Velocity / acceleration / change / angular velocity, direction / orientation, size / length / width / height / azimuth angle / area / volume / capacity, deformation / transformation, object / direction of motion / angle / shape / form / contraction / expansion, action / activity / movement, occurrence, fall / accident / security / event, period / frequency / rate / cycle / rhythm / count / quantity, timing / duration / interval, start / initiating / end / present / past / next time / quantity / information, type / grouping / classification / composition, presence / absence / proximity / approach / retreat / entry / exit, identity / identifier Head / Mouth / Eyes / Respiration / Heart / Hands / Handwriting / Arms / Body / Gestures / Legs / Gait / Organ Characteristics, Tidal Volume / Depth of Respiration / Airflow / Inspiratory / Expiratory Time / Ratio, Gait / Walking / Tools / Machines / Complex Movements, Signals / Motion Characteristics / Information / Features / Statistics / Parameters / Amplitude / Phase / Degree / Dynamics / Anomalies / Variability / Detection / Estimation / Recognition / Identification / Indicators, Gradient / Derivative / Higher-Order Derivative Function / Features / Mapping / Transformation of Another Feature, Discrepancy / Distance / Similarity Score / Cost / Indicators, Euclidean / Statistical / Weighted Distance, L1 / L2 / Lk Norm, Inner Product / Cross Product, Tags, Quantities that may include Tests,The amount consumed / not consumed, the state / physical / healthy / well-being / emotional / mental state, the output response, any configuration / combination, and / or any related characteristics / information / combination.

[0183] Test volume can be calculated. Characteristics / STI / MI can be calculated / monitored based on CI / TSCI / features / similarity scores / test volume. Static (or dynamic) segments / profiles can be identified / calculated / analyzed / monitored / extracted / acquired / marked / presented / indicated / highlighted / remembered / communicated by analyzing CI / TSCI / features / functions of features / test volume / characteristics / STI / MI (e.g., presence / detection / estimation / recognition / identification of target movement / movement). Test volume can be based on CI / TSCI / features / functions of features / characteristics / STI / MI. Test volume can be processed / tested / analyzed / compared.

[0184] The test quantity includes data / vectors / matrices / structure, characteristics / STI / MI, CI information (CII, e.g., CI / CIC / features / amplitude / phase), directional information (DI, e.g., directional CII), dominant / representative / characteristic / indicative / major / archetypal / exemplary / paradigmatic / prominent / common / shared / typical / prototypical / averaged / regular / persistent / normal / authentic / anomalous / anomalous / non-representative data / vectors / matrices / structure, similarity / discrepancy / distance score / cost / indicator, auto / cross-correlation / covariance, sum / mean / average / weighted / trim / arithmetic / geometric / harmonic mean, variance / deviation / absolute / squared deviation / averaged / median / sum / standard deviation / derivative / gradient / variance / sum / absolute / squared variation / spread / dispersion / variability. It may include any / any of the following functions: divergence / skewness / kurtosis / range / interquartile range / coefficient of variation / dispersion / L moment / interquartile variance coefficient / mean absolute / squared difference / Gini coefficient / relative mean difference / entropy / maximum / minimum / median / percentile / quartile, variance-to-mean ratio, maximum-to-minimum ratio, measure of variation / regularity / similarity, transient event / behavior, statistics / mode / likelihood / histogram / probability distribution function (pdf) / moment generating function / expectation function / expected value, behavior, repetition / periodicity / pseudoperiodicity, impulsivity / suddenness / occurrence / recurrence, temporal profile / characteristics, time / timing / duration / period / frequency / trend / history, initiation / start / end time / quantity / count, classification / type of movement, change, temporal / frequency / cycle change, etc.

[0185] Identifiers / identities / identifiers / IDs may include MAC addresses / ASID / USID / AID / UID / UUID, labels / tags / indexes, web links / addresses, numeric / alphanumeric IDs, names / passwords / accounts / account IDs, and / or other IDs. IDs may be assigned (e.g., by software / firmware / user / hardware, hardwired, via dongles). IDs may be stored / retrieved (e.g., locally / remotely / permanently / temporarily stored in databases / memory / cloud / edge / local / hub servers). IDs may be associated with any of the following: user / customer / household / information / data / address / telephone number / social security number, user / customer number / record / account, or timestamp / duration / timing. IDs may be made available to Type 1 / Type 2 devices / sensing / SBP initiators / responders. IDs may be used for registration / initialization / communication / identification / verification / discovery / recognition / authentication / access control / cloud access / networking / social networking / logging / recording / cataloging / classification / tagging / association / pairing / transactions / electronic transactions / intellectual property control (e.g., local / cloud / server / hub, first type / second type / neighbor / user / another device, by user).

[0186] Objects include people / pets / animals / plants / machines / users, babies / children / adults / elderly, professionals / specialists / leaders / commanders / managers / office staff / executives / doctors / nurses / workers / teachers / technicians / servicemen / repairmen / passengers / patients / customers / students / travelers / inmates / high-value individuals / , tracked objects, vehicles / cars / AGVs / drones / robots / wagons / transportation / remotely operated machines / carts / moving objects / goods / items / materials / parts / components / machines / lifts / elevators, goods / items / cargo / people / items / food / packages / luggage / equipment / cleaning supplies, telephones / computers / laptops / Tablets / dongles / plugins / companions / tools / peripherals / accessories / wearables / furniture / appliances / amenities / gadgets, IoT / networked / smart / portable devices, watches / glasses / speakers / toys / strollers / keys / wallets / purses / handbags / backpacks, goods / cargo / luggage / equipment / motors / machines / appliances / tables / chairs / air conditioners / doors / windows / heaters / fans, lighting / fixtures / stationary objects / televisions / cameras / audio / video / surveillance equipment / parts, tickets / parking / tolls / airline tickets, credit cards / plastic cards / access cards, fixed / changing / formless objects, mass / solid / liquid / gas / fluid / smoke / fire / flames, signage, electromagnetic (EM) sources / mediums, and / or other objects.

[0187] An object can have multiple parts, each with different movements (e.g., changes in position / location / direction). An object could be a person walking forward. While walking, their left / right hand may move in different directions with different instantaneous movements / velocities / accelerations.

[0188] The object may or may not be coupled to any network capable of communicating with any network such as WiFi, MiFi, 4G / LTE / 5G / 6G / 7G / 8G, Bluetooth / NFC / BLE / WiMAX / Zigbee / mesh / ad-hoc networks. The object may be a bulky machine with AC power that is moved during installation / cleaning / maintenance / renovation. It may be placed on / inside a movable platform such as an elevator / conveyor / lift / pad / belt / robot / drone / forklift / car / boat / vehicle. The first / second type device may be attached to / moved with the object. The first / second type device may be portable / part of another device (e.g., a module / device with a module, which may be large / quite large / small / heavy / bulky / light, e.g., coin-sized / cigarette-pack-sized) or embedded within it. Type 1 / Type 2 / Portable / Another device may or may not be attached to an object, may or may not move with the object, and may have wireless (e.g., via Bluetooth / BLE / Zigbee / NFC / WiFi) or wired (e.g., USB / Micro USB / Firewire / HDMI®) connections to nearby devices for network access (e.g., via WiFi / cellular network). Nearby devices may be objects / telephones / APs / IoT / devices / appliances / peripherals / amenities / furniture / vehicles / gadgets / wearables / networked / computing devices. Nearby devices may be connected to some server (e.g., a cloud server via a network / Internet). It may or may not be portable / movable, and may or may not move with the object. Type 1 / Type 2 / Portable / Nearby / Another device may be powered by batteries / solar / DC / AC / other power sources, and they may be replaceable / non-replaceable and rechargeable / non-rechargeable. It may be charged wirelessly.

[0189] Type 1 / Type 2 / Portable / Nearby / Another device includes: Computers / Laptops / Tablets / Pads / Telephones / Printers / Monitors / Batteries / Antennas, Peripherals / Accessories / Sockets / Plugs / Chargers / Switches / Adapters / Dongles, Internet of Things (IoT), TVs / Soundbars / HiFi / Speakers / Set-top Boxes / Remote Controls / Panels / Gaming Devices, APs / Cables / Broadband / Routers / Repeaters / Extenders, Home Appliances / Utilities / Fans / Refrigerators / Washing Machines / Dryers / Microwaves / Ovens / Stoves / Ranges / Lighting / Lamps / Tubes / Pipes / Faucets / Light Fixtures / Air Conditioners / Heaters / Smoke Detectors, Wearables / Watches / Glasses / Goggles / Buttons / Bracelets / Chains / Jewelry / Rings / Belts / Clothing / Clothes / Fabrics / Shirts / Pants / Dresses / Gloves / Shoes / Footwear / Hats / Headwear / Bags / Wallets / Cosmetics / Accessories / Books / Magazines / Paper / Stationery / Signage / Posters / Displays / Printed Materials, Furniture / Equipment This may include any of the following: goods / tables / desks / chairs / sofas / beds / cabinets / shelves / racks / storage / boxes / buckets / baskets / packaging / carts / tiles / strips / bricks / blocks / mats / panels / curtains / cushions / pads / carpets / materials / building materials / glass, amenities / sensors / clocks / pots / pans / wares / containers / bottles / cans / equipment / plates / cups / bowls / toys / balls / tools / pens / rackets / locks / bells / cameras / microphones / paintings / frames / mirrors / coffee makers / doors / windows, food / tablets / pharmaceuticals, implantable / transplantable / gadgets / equipment / devices / machines / controllers / mechanical tools, garage openers, keys / plastic cards / payment cards / credit cards / tickets, solar panels, key trackers, fire extinguishers, trash cans / bins, WiFi-enabled devices, smart devices / machines / systems / residential / office / building / warehouse / facility / vehicles / cars / bicycles / motorcycles / boats / ships / airplanes / carts / wagons.

[0190] One, two, or more Type 1 / Type 2 / Portable / Nearby / Another Device / Server may determine the initial properties / STI / MI of an object and / or share intermediate information. One of the Type 1 / Type 2 devices may move with the object (e.g., a "tracker bot"). The other of the Type 1 / Type 2 devices may not move with the object (e.g., an "origin satellite," an "origin register"). Both may have known properties / STI / MI. The initial properties / STI / MI may be calculated based on the known properties / STI / MI.

[0191] The venue includes sensing domains, rooms / houses / homes / offices / workplaces / buildings / facilities / warehouses / factories / shops / vehicles / sites, indoor / outdoor / enclosed / semi-enclosed / open / semi-open / closed / aerial / floating / underground spaces / areas / structures / enclosures, wood / glass / metal / materials / structures / frames / beams / panels / columns / walls / floors / doors / ceilings / windows / cavities / gaps / openings / reflective / refracting media / fluids / construction materials / spaces / areas with fixed or adjustable layouts or shapes, human / animal / plant bodies / cavities / organs / bones / blood / blood vessels / tracheas / teeth / soft or hard or rigid or non-rigid tissues, manufacturing / repair / maintenance / mining / parking lots / storage / transportation / delivery / logistics / sports / entertainment / amusement / public / recreation / government / community / elderly / elderly care / senior care facilities / space facilities / terminals / hubs, distribution centers / shops, machinery / engines / devices / assembly lines / workflows, cities / rural areas / suburbs / metropolitan areas, stairs / escalators / elevators / corridors / passages / tunnels / caves / caves / channels / ducts / pipes / tubes / lifts / wells / paths / roofs / basements / studios / alleys / roads / routes / highways / sewers / ventilation systems / networks, cars / trucks / buses / vans / containers / ships / boats / submarines / trains / trams / airplanes / mobile homes, stadiums / cities / playgrounds / parks / fields / tracks / courts / gymnasiums / halls / marts / markets / supermarkets / plazas / squares / construction sites / hotels / museums / schools / hospitals / universities / garages / malls / airports / train or bus stations / terminals / hubs / platforms, valleys / forests / groves / terrain / landscapes / gardens / parks / patios / land, and / or any space such as gas / oil / water pipes / lines. The venue may include the interior / exterior of a building / facility. The building / facility may have one or more floors, including an underground section.

[0192] Events may be monitored based on TSCI. Events may relate to objects / movements / gestures / gaits, such as falls, spins / hesitations / pauses, impacts (e.g., a person hitting a punching bag / door / bed / window / chair / table / desk / cabinet / box / another person / animal / bird / fly / ball / bowling / tennis / soccer / volleyball / football / baseball / basketball), two-body actions (e.g., a person releasing a balloon / catching a fish / shaping clay / writing on paper / typing on a computer), cars moving around a garage, people carrying smartphones / walking around a venue, and moving autonomous / mobile objects / machines (e.g., vacuum cleaners / utilities / autonomous vehicles / cars / drones).

[0193] Tasks may include: (a) Sensing tasks, any of the following: object / multiple objects / vehicle / machine / tool / human / baby / elderly / patient / intruder / pet presence / proximity / activity / daily activities / wellness / respiration / vital signs / heart rate / health status / sleep / sleep stage / walking / position / distance / speed / acceleration / navigation / tracking / movement / safety / danger / fall / intrusion / security / threat to life / emotion / movement / degree / pattern / periodic / repetitive / cyclostationary / steady / regular / transient / sudden / suspicious movement / irregularity / (b) Monitoring / sensing / detection / recognition / estimation / verification / identification / audience informatics / gait / gestures / in / room / area / zone / venue / of / in / out Wavenumber / time / function decomposition / neural network / map-based / model-based processing / correction / geometric estimation / analytical calculation, (c) IoT tasks, any of the following: smart tasks for venues / users / objects / people / pets / residential / home / office / workplace / building / facility / warehouse / factory / store / vehicle / site / structure / assembly line / IoT / device / system, energy / power management / transmission, wireless power transmission, interaction / engagement with users / objects / intruders / people / animals (e.g., presence / movement / gestures / gait / activity / behavior / voice / commands / instructions / queries / sounds) Music / sound / images / video / location / movement / danger / threat detection / recognition / monitoring / analysis / response / execution / synthesis, interaction / exchange / response / presentation / experience / media / multimedia / representation / sound / voice / music / images / imaging / video / animation / web pages / text / messages / notifications / reminders / queries / alert generation / search / play / display / rendering / synthesis, user / intruder / object input / movement / gestures / location / activity detection / recognition / monitoring / interpretation / analysis / recording / storage), device / system (e.g.,(d) Activation / control / configuration (e.g., turn on / off / control / lock / unlock / open / close / adjust / configure) vehicles / drones / electrical / mechanical / air conditioning / heating / lighting / ventilation / cleaning / entertainment / IoT / security / sirens / access systems / devices / doors / windows / garages / lifts / elevators / escalators / speakers / televisions / lighting / peripherals / accessories / wearables / furniture / appliances / amenities / gadgets / alarms / cameras / games / coffee / cooking / heaters / fans / housework / home / office equipment / devices / robots / vacuum cleaners / assembly lines), (d) miscellaneous tasks, any of the following: data / parameters / analysis / transmission / coding / encryption / storage / analysis, upgrade / management / configuration / adjustment / broadcast / synchronization / networking / encryption / communication / protection / compression / storage / databases / archives / queries / cloud computing / presentations / augmentation / virtual reality / other processing / tasks. The task may be performed by a first-type / second-type / neighborhood / portable / another device, and / or by a hub / local / edge / cloud server.

[0194] Tasks may also include: detection / recognition / monitoring / location / interpretation / analysis / recording / memory of users / visitors / intruders / objects / pets; interaction / engagement / conversation / dialogue / exchange with users / objects / visitors / intruders / humans / babies / pets; health / wellness / daily life / activities / behavior / patterns / exercise / eating / toilet visits / work / play / rest / sleep / relaxation / danger / routine / timing / habit / trend / normality / state / abnormality / regularity / irregularity / change / existence / movement / gestures / gait / expression / emotions / State / Stage / Voice / Command / Instruction / Question / Query / Music / Sound / Location / Movement / Fall / Threat / Discomfort / Illness / Environmental Detection / Location Identification / Localization / Recognition / Monitoring / Analysis / Interpretation / Learning / Training / Response / Execution / Synthesis / Generation / Recording / Memory / Summarization, Dialogue / Exchange / Response / Presentation / Report / Experience / Media / Multimedia / Representation / Sound / Voice / Music / Image / Imaging / Video / Animation / Webpage / Text / Message / Notification / Reminder / Inquiry / Warning Generation / Detection Detecting / playing / displaying / rendering / compositing, detecting / recognizing / monitoring / interpreting / analyzing / recording / memorizing user / intruder / object input / movement / gestures / location / activity), detecting / checking / monitoring / locating / managing / controlling / controlling / configuring / locking / unlocking / alert / disarming / opening / closing / interpreting / analyzing / recording / memorizing / controllingDetecting / monitoring / location of the user (sitting / sleeping on the sofa / in the bedroom / running on the treadmill / cooking / watching TV / eating in the kitchen or dining room / going up / down the stairs / outside / inside / using the toilet), automatically doing something when detected (e.g., generating a message / response / warning / explanation / notification / report), automatically doing something for the user when the user's presence is detected, turning on / off / waking / controlling / adjusting / dimming lights / music / radio / TV / HiFi / STB / computer / speaker / smart device / air conditioning / ventilation / heating system / curtains / lightshades, turning on / off / preheating / controlling coffee makers / electric kettles / cookers / ovens / microwaves / other cooking devices, checking / managing temperature / settings / weather forecasts / phone calls / messages / emails / system checks, presenting / interacting / engaging / dialogue / conversing (e.g., via smart speakers / displays / screens; via web pages / emails / messaging systems / notification systems).

[0195] When the user returns home by car, the task may automatically detect the user / car's approach, open the garage / door upon detection, turn on the driveway / garage lights and / or turn on the air conditioner / heater / fan as the user approaches the garage. When the user enters the house, the task may automatically turn on the porch lights / turn off the driveway or garage lights, play a greeting message to welcome the user, turn on the user's favorite music / radio / news / channels, open the curtains / blinds, monitor the user's mood, adjust the lighting / sound environment according to the mood / current or imminent events in the user's calendar (for example, if the user has dinner plans with their girlfriend soon, turn on romantic lighting / music), heat up food in the microwave that the user prepared in the morning, perform diagnostic checks on all systems in the house, check tomorrow's weather forecast / news of interest to the user, check the calendar / to-do list, play reminders, check answering machine / messaging system / email, and use a dialogue system / speech synthesis to speak. The task may involve mentally preparing / providing a report and / or (using audible tools such as speakers / HiFi / speech synthesis / sound / field / voice / music / song / dialogue systems, visual tools such as televisions / entertainment systems / computers / notebooks / tablets / displays / lights / colors / brightness / pattern symbols, haptics / virtual reality / gestures / tools, smart devices / appliances / materials / furniture / fixtures, servers / hub devices / cloud / fog / edge servers / home / mesh networks, messaging / notification / communication / scheduling / email tools, UI / GUI, scents / odors / fragrances / tastes, nerves / nervous system / tools, or any combination thereof) reminding the user of someone's birthday / calling him and preparing / providing a report. The task may also involve pre-turning on an air conditioner / heater / ventilation system and / or pre-adjusting the temperature setting of a smart thermostat.When a user moves from the entrance to the living room, the tasks may include turning on the living room lights, opening the living room curtains, opening the windows, turning off the entrance lights behind the user, turning on the TV / set-top box, setting the TV to the user's favorite channel, and / or adjusting appliances according to the user's preferences / conditions / state (e.g., adjusting the lighting, selecting / playing music to create a romantic atmosphere).

[0196] When the user wakes up in the morning, the task may detect the user moving around in the bedroom, open the blinds / curtains / windows, turn off the alarm clock, adjust the temperature from a night profile to a day profile, turn on the bedroom lights, turn on the bathroom lights as the user approaches the bathroom, check the radio / streaming channels and play the morning news, turn on the coffee maker, preheat the water, and / or turn off the security system. When the user walks from the bedroom to the kitchen, the task may turn on the kitchen / hallway lights, turn off the bedroom / bathroom lights, move music / messages / reminders from the bedroom to the kitchen, turn on the kitchen TV, change the TV to a morning news channel, lower the kitchen blinds, open the kitchen window, unlock the back door so the user can check the backyard, and / or adjust the kitchen temperature setting.

[0197] When a user leaves home for work, the task may detect the user's departure, play a farewell / "Have a good day" message, open and close the garage door, turn the garage / driveway lights on / off, close / lock all windows / doors (if the user forgets), turn off appliances (e.g., stove / microwave / oven), turn on / alarm the security system, adjust the lighting / air conditioning / heating / ventilation system to an "absent" profile to save energy, and / or send an alert / report / update to the user's smartphone.

[0198] Movement includes: stillness, sequences of movement, resting / immobile movement, movement / change of location, daily / weekly / monthly / yearly / repetitive / activity / behavior / action / routine, transient / time-varying / falling / repetitive / repeated / periodic / pseudoperiodic movement / breathing / heartbeat, deterministic / non-deterministic / stochastic / chaotic / random movement, complex / combined movement, non / pseudo / cyclo / stationary random movement, changes in electromagnetic properties, movement of humans / animals / plants / bodies / machines / mechanical / vehicles / drones, air / wind / weather / water / fluid / This may include ground / subsurface / earthquake movements, man-machine interactions, normal / abnormal / dangerous / warning / suspicious movements, imminent / rain / fire / flood / tsunami / explosion / collision, head / face / eyes / mouth / tongue / neck / fingers / hands / arms / shoulders / upper / lower / body / chest / abdomen / waist / legs / feet / joints / knees / elbows / skin / subcutaneous / subcutaneous tissue / blood vessels / intravenous / organs / heart / lungs / stomach / intestines / intestinal tract / eating / breathing / talking / singing / dancing / coordinated movements, facial / eye / mouth expressions, and / or hand / arm / gestures / gait / UI / keystrokes / typing strokes.

[0199] Type 1 / Type 2 devices may include heterogeneous ICs, low-noise amplifiers (LNAs), power amplifiers, transmit / receive switches, media access controllers, baseband radios, and / or 2.4 / 3.65 / 4.9 / 5 / 6 / sub-7 / over-7 / 28 / 60 / 76GHz / other radios. Heterogeneous ICs may include processors / memory / software / firmware / instructions. It may support broadband / wireless / mobile / mesh / cellular networks, WLAN / WAN / MAN, standards / IEEE / 3GPP(registered trademark) / WiFi / 4G / LTE / 5G / 6G / 7G / 8G, IEEE 802.11 / a / b / g / n / ac / ad / af / ah / ax / ay / az / be / bf / 15 / 16, and / or Bluetooth(registered trademark) / BLE / NFC / Zigbee / WiMax.

[0200] The processor may include any of the following: general-purpose / dedicated / purpose / embedded / multicore processors, microprocessors / microcontrollers, multi / parallel / CISC / RISC processors, CPUs / GPUs / DSPs / ASICs / FPGAs, and / or logic circuits. Memory may include non-volatile / volatile RAM / ROMs / EPROMs / EEPROMs, hard disks / SSDs, flash memory, CD- / DVD-ROMs, magnetic / optical / organic / storage systems / networks, network / cloud / edge / local / external / internal storage, and / or any non-temporary storage media. The instruction set may include machine executable code in hardware / ICs / software / firmware, and may be embedded / pre-loaded / loaded at boot time / on the fly / on-demand / pre-installed / installed / downloaded.

[0201] Processing / pre-processing / post-processing may be applied to data (e.g., TSCI / features / characteristics / STI / MI / test quantity / intermediate / data / analysis) and may have multiple steps. Steps / pre / post / processing may include any of the following: calculation of operand / LOS / non-LOS / single-link / multi-link / component / item / quantity functions, calculation / extraction / correction / cleaning of magnitude / norm / phase / features / energy / timebase / similarity / distance / characteristic score / scale, linear / non-linear / FIR / IIR / MA / AR / ARMA / Kalman / particle filtering, low-pass / band-pass / high-pass / median / rank / quartile / percentile / mode / selection Adaptive filtering, interpolation / extrapolation / decimation / subsampling / upsampling / resampling, matched filtering / enhancement / reconstruction / denoising / smoothing / conditioning / spectral analysis / average subtraction / removal, linear / nonlinear / inverse / frequency / time transformation, Fourier transform (FT) / DTFT / DFT / FFT / wavelets / Laplace / Hilbert / Hadamard / trigonometric functions / sine / cosine / DCT / powers of 2 / sparse / fast / frequency transformation, zero / cyclic / padding, gradual Rough-based transformation / processing, decomposition / orthogonal / non-orthogonal / hyperperfect projection / eigendecomposition / SVD / PCA / ICA / compressed sensing, grouping / folding / sorting / comparison / soft / hard / thresholding / clipping, first-order / quadratic / higher-order derivatives / integration / convolution / multiplication / division / addition / subtraction, local / global / maximization / minimization, recursive / iterative / constrained / batch processing, least squares mean / absolute error / deviation, cost function optimization, neural networks / detection / recognition / classification / identification / estimation / labeling / association / tagging / Mapping / Remapping / Training / Clustering / Machine / Supervised / Unsupervised / Semi-supervised Learning / Networks, Vectors / Quantization / Encryption / Compression / Matching Pursuit / Scrambling / Coding / Memory / Retrieval / Send / Receive / Time Domain / Frequency Domain / Normalization / Scaling / Expanding / Representation / Merge / Combine / Partition / Tracking / Monitoring / Shape / Silhouette / Motion / Activity / Analysis, PDF / Histogram Estimation / Importance / Monte Carlo Sampling, Error Detection / Protection / Correction, Do Nothing, Time-Variable / Adaptive Processing,Conditioning / weighting / averaging / selected components / links; arithmetic / geometric / harmonic / trimmed mean / centroid / medoid calculations; morphological / logical operations / substitution / combination / sorting / AND / OR / XOR / union / intersection; vector operations / addition / subtraction / multiplication / division; and / or other operations. Processing may be applied individually / jointly. Acceleration using GPU / DSP / coprocessor / multicore / multiprocessing may be applied.

[0202] Functions may include: characteristics / features / magnitude / phase / energy, scalar / vector / discrete / continuous / polynomial / exponential / logarithmic / trigonometric / transcendental / logical / piecewise / linear / algebraic / nonlinear / circular / piecewise linear / real / complex / vector values / inverse / absolute / indicator / restriction / floor / rounding / sign / composition / sliding / moving functions, derivatives / integrals, functions of functions, one-to-one / one-to-many / many-to-one / many-to-many functions, mean / mode / median / percentile / maximum / minimum / range / statistics / histogram, local / global maximum / minimum / zero crossing, variance / variability / spread / variance / Deviation / Standard deviation / Divergence / Range / Interquartile range / Total variation / Absolute / Total deviation, Arithmetic / Geometric / Harmonic / Trimmed mean / Squaring / Cube / Root / Power, Thresholding / Clipping / Rounding / Truncation / Quantization / Approximation, Time function processed by operations (e.g., filtering), Sine / Cosine / Tangent / Cotangent / Sector / Cosecant / Elliptic / Parabola / Hyperbola / Game / Zeta function, Probabilistic / Random / Ergodic / Stationary / Deterministic / Periodic / Iterative function, Inverse / Transform / Frequency / Discrete time / Laplace / Hilbert / Sine / Cosine / Triangular / Wavelet / Integer / Power of 2 / Sparse transformation, Intersection / Non-orthogonal / Eigenprojection / Decomposition / Eigenvalues / Singular values / PCA / ICA / SVD / Compressed sensing, Neural networks, Feature extraction, Functions of moving windows of adjacent terms in time series, Filtering functions / Convolution, Short-time / Discrete transforms / Fourier / Cosine / Sine / Hadamard / Wavelet / Sparse transform, Matching pursuit, Approximation, Graph-based processing / Transformation / Graph signal processing, Classification / Discrimination / Class / Group / Category / Labeling, Processing / Preprocessing / Postprocessing, Machine / Learning / Detection / Estimation / Feature extraction / Learning networks / Feature extraction / Denoising / Signal enhancement / Co Reading / encryption / mapping / vector quantization / remapping / lowpass / highpass / bandpass / matching / Kalman / particle / FIR / IIR / MA / AR / ARMA / median / mode / adaptive filtering, first-order / second-order / higher-order derivatives / integral / zero crossing / smoothing, up / down / random / importance / Monte Carlo sampling / resampling / transformation, interpolation / extrapolation, short / long-term statistics / auto / cross-correlation / moment generating function / time mean / weighted mean, special / Bessel / beta / gamma / Gauss / Poisson / integral complementary error function.

[0203] The sliding time window may have a time-varying width / size. It may be small / large at the start to allow for fast / accurate acquisition, and may increase / decrease over time to a steady-state size comparable to the frequency / period / duration / characteristics / STI / MI of the motion to be monitored. The window size / time shift between adjacent windows may be constant, or may be adaptive / dynamic / automatically changed / adjusted / modified (based on, for example, battery life / power consumption / available computing power / changes in the amount of target / nature of the motion to be monitored / user requirements / selections / instructions / commands).

[0204] The characteristic / STI / MI may be determined based on the characteristic values / characteristic points of the function and / or related function arguments (e.g., time / frequency). The function may be the result of a regression. The characteristic values / characteristic points may include the local / global / constrained / significant / first / second / ith maximum / minimum / extremum / zero crosses (e.g., with positive / negative time / frequency / arguments) of the function. The local signal-to-noise ratio (SNR) or SNR-like parameter may be calculated for each pair of adjacent local maximums (peaks) / local minimums (valleys) of the function, which may be a function (e.g., linear / logarithmic / exponential / monotonical / power / polynomial) of the fraction or difference between the amount of the local maximum (e.g., power / magnitude) and the amount of the local minimum. A local maximum (or minimum) may be significant if its SNR is greater than a threshold and / or its amplitude is greater than (or less than) another threshold. Local maximums / minims may be selected / identified / calculated using a persistence-based approach. Several significant local maximums / minims may be selected based on selection criteria (e.g., quality criteria / conditions, the strongest / consistent significant peak within the range). Significant peaks that are not selected may be saved / monitored as “reserved” peaks for use in future selections in future sliding time windows. For example, a particular peak (e.g., at a particular argument / time / frequency) may appear consistently over time. Initially, it may not be selected even if significant (because other peaks may be stronger). Subsequently, it may become consistently stronger / dominant. When selected, it may be tracked retrospectively and selected at an earlier time to replace previously selected peaks (which may be temporarily strong / dominant but not persistent / consistent). Peak consistency may be measured by tracing or the duration for which it is significant. Alternatively, local maximums / minims may be selected based on a finite state machine (FSM). The decision threshold may be time-varying and may be adjusted adaptively / dynamically (e.g., based on backtrace timing / FSM, or data distribution / statistics).

[0205] The Similarity Score (SS) / Component SS may be calculated based on two temporally adjacent CIs / CICs of one TSCI or two different TSCIs. The pairs may be obtained from the same / different sliding windows. The SS or Component SS may include time-reverse resonance intensity (TRRS), auto / cross-correlation / covariance, dot product of two vectors, L1 / L2 / Lk / Euclidean / statistical / weighted / distance score / norm / metric / quality metric, signal quality condition, statistical characteristics, discriminant score, neural network / deep learning network / machine learning / training / discrimination / weighted averaging / preprocessing / denoising / signal conditioning / filtering / temporal correction / timing compensation / phase offset compensation / transformation / component-by-component calculation / feature extraction / FSM, and / or other scores.

[0206] Any threshold may be fixed (e.g., 0, 0.5, 1, 1.5, 2), predetermined, and / or adaptively / dynamically determined (e.g., by FSM, or based on time / space / location / antenna / path / link / state / battery life / remaining battery life / available resources / power / computational power / network bandwidth). The threshold may be applied to a test quantifier to distinguish between two events / conditions / situations / states, A and B. Data (e.g., CI / TSCI / feature / similarity score / test quantifier / characteristic / STI / MI) may be collected under A / B in a training situation. A test quantifier calculated based on the data (e.g., its distribution) may be compared under A / B, and a threshold may be selected based on several criteria (e.g., maximum likelihood (ML), maximum posterior probability (MAP), discrimination training, minimum type I (or type II) error for a given type II (or type I) error, quality criteria, signal quality conditions). The threshold may be adjusted automatically / semi-automatically / manually / adaptively / dynamically (for example, to achieve different sensitivities), once / sometimes / frequently / periodically / repeatedly / occasionally / sporadically / on demand (based on, for example, objects / movement / position / orientation / action / characteristics / STI / MI / size / characteristics / traits / habits / behavior / venue / features / fixtures / furniture / barriers / materials / machines / living things / objects / boundaries / surfaces / mediums / maps / constraints / models / events / states / situations / conditions / time / timing / duration / state / history / user / preferences). The iterative algorithm may stop after N iterations, after a timeout period, or after the test quantity satisfies a condition (for example, the updated quantity is greater than the threshold), and that condition may be fixed or adjusted adaptively / dynamically.

[0207] Searching for local extrema may include constrained / minimizing / maximizing, statistical / dual / constrained / convex / global / local / combinatorial / infinite-dimensional / multiobjective / multimodal / nondifferentiable / particle swarm / simulation-based optimization, linear / nonlinear / quadratic / higher-order regression, linear / nonlinear / stochastic / constrained / dynamic / mathematical / discrete / convex semidefinite / conical / conical / interior point / fractional / integer / sequential / quadratic programming, conjugate / gradient / subgradient / coordinate / shrinking descent, Newtonian / simplex / iterative / point / ellipsoid / quasi-Newtonian / interpolation / mimetic / genetic / evolutionary / pattern / gravity search methods / algorithms, constraint satisfaction, variational methods, optimal control, spatial mapping, heuristics / metahouristics, numerical analysis, simultaneous perturbation stochastic approximation, stochastic tunneling, dynamic relaxation, hill climbing, simulated annealing, differential evolution, robust / line / taboo / reactive search / optimization, curve fitting, least squares, variational methods, and / or variants. It may be associated with the purpose / loss / cost / utility / fitness / energy function.

[0208] Regression may be performed using a regression function to fit data, or a function of data (e.g., ACF / converted / mapped), in a regression window. During iteration, the length / position of the regression window may be changed. The regression function may be linear / quadratic / cubic / polynomial / another function. Regression may minimize any of the following: mean / weighted / absolute / squared deviation, error, aggregated / component / weighted / mean / sum / absolute / squared / higher order / another error / cost (e.g., in a projection area / selected axis / orthogonal axis), robust error (e.g., a first error (e.g., squared) for smaller error magnitudes and a second error (e.g., absolute) for larger error magnitudes), and / or weighted sum / mean of multiple errors (e.g., absolute / squared error). Errors associated with different links / paths may have different weights (e.g., a link with less noise may have a higher weight). Regression parameters (e.g., time offset associated with the maximum / minimum regression error of the regression function in the regression window, window position / width) may be initialized and / or updated during iteration (e.g., based on target value / range / profile, characteristic / STI / MI / test amount, object movement / amount / count / position / state, past / current trend, position / amount / distribution of local extrema in previous window, carrier / sub-carrier frequency / bandwidth of signal, amount of antenna associated with channel, noise characteristic, histogram / distribution / center / F distribution, and / or threshold). When converged, the current time offset may be at the center / left / right (or fixed relative position) of the regression window.

[0209] In presentation, information may be displayed / presented (e.g., together with a venue map / environmental model). Information may include: current / past / corrected / approximate maps / location / velocity / acceleration / zones / regions / areas / segmentation / coverage area, direction / path / trace / history / traffic / summary, frequently visited areas, customer / crowd events / distribution / behavior, crowd management information, acceleration / velocity / vital signs / respiration / heart rate / activity / emotion / sleep / state / rest information, motion statistics / MI / STI, presence or absence of movement / person / pet / object / vital signs, gestures (e.g., hand / arm / foot / leg / body / head / face / mouth / eye) / meaning / control (control of devices using gestures), location-based gestures Interpretation of control / motion, identity / identifier (ID) (e.g., object / person / user / pet / zone / area, device / machine / vehicle / drone / car / boat / bicycle / television / air conditioner / fan, autonomous driving machine / device / vehicle), environment / weather information, gesture / gesture control / motion tracing, earthquake / explosion / storm / rain / fire / temperature, collision / impact / vibration, event / door / window / open / close / fall / accident / burn / freeze / water / wind / air movement event, repetitive / pseudoperiodic event (e.g., running on a treadmill, jumping up and down, jump rope, somersault), and / or vehicle events. Position may be one / two / three-dimensional (e.g., represented / displayed as 1D / 2D / 3D orthogonal / polar coordinates), relative (e.g., with respect to a map / environment model), or relational (e.g., distance from / near / from a point, midpoint between two points, around a corner, on an upper floor, on a table, on the ceiling, on the floor, on a sofa).

[0210] Information (e.g., location) may be marked / displayed with some symbol. The symbol may be time-varying / flashing / pulsating and may change in color / intensity / size / orientation. The symbol may be a number reflecting an instantaneous quantity (e.g., analysis / gesture / status / status / operation / movement / breathing / heart rate, temperature / network traffic / connectivity / remaining power). The characteristics of the symbol / size / orientation / color / intensity / rate / change may reflect respective movements. The information may be text or presented visually / orally (e.g., using recorded voice / text-to-speech) / mechanically (e.g., animated gadget, movement of a movable part).

[0211] The user device may include a smartphone / tablet / speaker / camera / display / TV / gadget / vehicle / home appliance / device / IoT, a device with UI / GUI / voice / audio / recording / capture / sensor / playback / display / animation / VR / AR (augmented reality) / voice (assistance / recognition / synthesis) functions, and / or a tablet / laptop / PC.

[0212] The map / floor plan / environment model (e.g., of a home / office / building / store / warehouse / facility) may be two / three / higher dimensional. It may change / evolve over time (e.g., rotate / zoom / move / jump on the screen). Walls / windows / doors / entrances / exits / no-go areas may be marked. It may include multiple layers (overlay). It may include a maintenance map / model including water pipes / gas pipes / wiring / cable wiring / air ducts / crawl spaces / ceilings / underground layouts.

[0213] The venue may be segmented / subdivided / zoned / grouped into multiple zones / domains / sectors / sections / territory / district / region / neighborhood / area / extension / space, such as bedroom / living room / dining room / lounge / storage room / utility / warehouse / meeting room / workroom / corridor / kitchen / entrance hall / garage / first floor / second floor / office / reception room / area / domain. The segments / domains / areas may be presented in a map / floor plan / model along with their presentation characteristics (e.g., brightness / intensity / luminance / color / chromaticity / texture / animation / flashing / rate).

[0214] An example of a disclosed system / device / method. Stephen and his family want to install the disclosed wireless motion detection system to detect movement in their 2,000 square foot, two-story townhouse in Seattle, Washington. Because his house is two stories, Stephen decides to use one Type 2 device (named A) and two Type 1 devices (named B and C) on the first floor. His first floor has three rooms in a straight line: the kitchen, dining room, and living room, with the dining room in the center. He divides the first floor into three zones (dining room, living room, kitchen) by placing A in the dining room, B in the kitchen, and C in the living room. When motion is detected by the AB pair and / or AC pair, the system analyzes the TSCI / feature / characteristic / STI / MI to associate the motion with one of the three zones.

[0215] When Steven and his family go camping on holidays, he uses a mobile phone app (e.g., an Android phone app or an iPhone® app) to turn on a motion detection system. If the system detects motion, an alert signal is sent to Steven (e.g., via SMS, email, or push message to the mobile phone app). If Steven pays a monthly fee (e.g., $10 per month), a service company (e.g., a security company) receives the alert signal via a wired (e.g., broadband) / wireless (e.g., WiFi / LTE / 5G) network and takes security actions (e.g., calling Steven to check if there is a problem, sending someone to the house to check, contacting the police on Steven's behalf).

[0216] Stephen loves his elderly mother and cares about her well-being when she's home alone. When his mother is home alone while other family members are out (e.g., at work / shopping / on vacation), Stephen uses a mobile app to turn on a motion detection system to make sure she's okay. He uses the mobile app to monitor his mother's movements around the house. Stephen knows his mother is okay when he uses the mobile app to see her moving through three areas of the house according to her daily routine. Stephen appreciates that the motion detection system helps him monitor his mother's well-being even while he's away from home.

[0217] On a typical day, his mother wakes up at 7 a.m., spends 20 minutes in the kitchen preparing breakfast, and eats breakfast in the dining room for 30 minutes. After that, she does her routine exercises in the living room before sitting on the sofa and watching her favorite TV shows. The motion detection system allows Stephen to see the timing of movements in three areas of the house. When the movements match the routine, Stephen has a rough idea that his mother is probably fine. However, if the movement pattern seems unusual (for example, no movement until 10 a.m., or staying still in the kitchen for a long time), Stephen suspects something is wrong and calls his mother to check on her. Stephen may even ask someone else (for example, a family member, neighbor, paid staff, friend, social worker, or service provider) to check on his mother.

[0218] One day, Steven feels the urge to rearrange the device. He simply unplugs it from its original AC power outlet and plugs it into a different one. He is pleased that the motion detection system is plug-and-play and that rearranging it does not affect its operation. He plugs it in, and it works immediately.

[0219] After a while, Stephen decides to set up a similar setup on the second floor (i.e., one Type 2 and two Type 1 devices) to monitor his upstairs bedroom. Again, he finds the system setup to be very simple, as all he has to do is plug the Type 2 and Type 1 devices into the AC power outlets on the second floor. No special installation is required. He can monitor activity on both the first and second floors using the same mobile app. Each Type 2 device on the first / second floor can interact with all the Type 1 devices on both floors. Stephen finds that the combined system gives him more than double the capabilities.

[0220] The disclosed system can be applied to many uses. Type 1 / Type 2 devices may be any WiFi-enabled device (e.g., smart IoT / appliances / TV / STB / speaker / refrigerator / cooker / oven / microwave / fan / heater / air conditioner / router / telephone / computer / tablet / accessories / plug / pipe / lamp / smoke detector / furniture / fixtures / shelf / cabinet / door / window / lock / sofa / tablet / piano / cooking utensils / wearable / watch / tag / key / ticket / belt / wallet / pen / hat / necklace / implanted / telephone / glasses / glass panel / game console) in a home / office / facility, on a table, ceiling, floor, or wall. They may be placed in a conference room to count people. They may form a well-being monitoring system to monitor the daily activities of the elderly and detect signs of symptoms (e.g., dementia, Alzheimer's disease). They may be used in a baby monitor to monitor a baby's vital signs (respiration). They may be placed in a bedroom to monitor sleep quality and detect sleep apnea. They may be placed inside vehicles to monitor the health of passengers and drivers and to detect drowsy drivers or babies left unattended in hot vehicles. They may be used in logistics to prevent human trafficking by monitoring people hiding in trucks / containers. They may be deployed by emergency services at disaster sites to search for victims trapped in rubble. They may be deployed in security systems to detect intruders.

[0221] In some embodiments, this disclosure discloses deep learning-based wireless sensing using a foundational model (FM). In some embodiments, a deep neural network (DNN; or deep learning network, or network of neural networks) / model / FM / LLM may be used as a classifier / detector to classify / detect an input into several resulting classes. The classifier / detector may include a neural network-based feature extraction module (or stage 1 network) and a subsequent neural network-based classifier / detector module (or stage 2 network). The feature extraction module may be a convolutional neural network (CNN) having N2 layers, each layer having multiple associated convolutional filters. The convolutional filters may be applied (individually / independently) to each of the multiple input matrices. Downsampling may be applied. The classifier / detector / DNN / model / FM / LLM module may include any of the following: feedforward neural networks (FNNs), fully connected networks (FCNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTMs), or transformers. The input to the DNN / model / FM / LLM may be computed / generated / prepared based on multiple 1D transformations (at one or more granularities).

[0222] In some examples, some of the variables used in this disclosure may include: N1 TSCIs (where N1 may vary in different situations / setups / venues, but DL may expect M1 TSCIs, where M1 may be greater than / less than or equal to N1), N2 layers of CNN (M2), N3 layers of FCN (M3), N4 CICs for each CI (M4), N5 CIs over a period of time (M5), N6 pairs from N1 TSCIs (M6), and N7 transmitted ANs. Tenor, N8 receiving antennas, N9 sliding time windows over a period of time (M9), N10 device pairs of Type 1 / Type 2 devices (M10), N11 points of 1D transformation (M11), N12 CICs selected from N4 (or component-based 1D transformations for each TSCI), N13 points of resampling (from N11 points), and N14 TSCI-based 1D transformations selected from N1 for each device pair.

[0223] In some examples, a wireless sensing system may include N10 TX-RX device pairs, N1 TSCIs per pair, and N4 CICs per CI. For example, a wireless sensing system may have N10 TX-RX device pairs in the venue, and each TX-RX device pair in the system comprises each type 1 heterogeneous wireless device (TX device) and each type 2 heterogeneous wireless device (RX device) in the system, working together to perform at least one wireless sensing task (e.g., occupancy / presence detection / learning / supervised learning / unsupervised learning / self-supervised learning / generative pretext task based on reconstruction / feature learning as a pretext task / feature similarity / task transfer / contrast learning, or multiple tasks / subtasks). The wireless sensing task may be performed in a robust and environment-independent manner (e.g., for any N10, for any TX-RX device pair with any bandwidth / modulation, for any placement / orientation of TX / RX devices, for TX / RX devices with any number of antennas, for any venue, etc.).

[0224] For each TX-RX device pair, each radio signal (e.g., WiFi signal, cellular signal, UWB, Bluetooth, radar signal, etc.) may be transmitted from each Type 1 device (TX), which has N7 transmitting antennas, to each Type 2 device (RX), which has N8 receiving antennas, via each radio channel in the venue. From each received radio signal, each Type 2 device may acquire / extract time series (TSCI, e.g., CSI, CIR, CFR, RSSI) of channel information for each radio channel, N1 = N7 * N8, where each TSCI is associated with one of the N7 transmitting antennas and one of the N8 receiving antennas.

[0225] In some cases, the transmitting antennas are close together, and the receiving antennas are also close together, resulting in all N1 TSCIs monitoring the same object / event within the same area / zone of the venue. Therefore, they may be analyzed together for the task (for example, by calculating the autocorrelation function (ACF) or short-time Fourier transform (STTFT) based on all N1 TSCIs).

[0226] In other cases, transmitting antennas may be strategically positioned in different locations / corners of the venue (e.g., different / adjacent / opposite corners of a car). Similarly, receiving antennas may be positioned in different locations / corners of the venue (e.g., different / adjacent / opposite corners inside a car). The corresponding N1 TSCIs may be used to monitor different areas / zones of the venue and / or the same / different objects / movements / events. Thus, each of the N1 TSCIs may be analyzed independently for the task (e.g., by each calculating N1 ACFs based on its respective TSCI).

[0227] TSCI may be preprocessed (e.g., denoising, phase cleaning, correction, compensation). Each channel information (CI) of N1 TSCIs may have N4 components (referred to as "CI components" or CIC, for example, components may be multiple subcarriers in the CFR, multiple taps in the CIR). Each CI may be represented as a vector of N4 tuples (each of the N4 vector components may be the respective CI component, or a function of the respective CI component (e.g., magnitude / magnitude squared / phase, or a function thereof)). N7, N8, N1, and / or N4 may be the same or different for different device pairs (e.g., different N1 due to different transmit / receive antenna amounts for Type 1 / Type 2 devices, different N4 due to different bandwidths of radio signals: 20 / 40 / 80 / 160 / 320 / 640 MHz, etc.). The TSCI may be decomposed into a time series of N4 CICs (TSCIC), where each TSCIC is associated with one of the N4 CICs of each CI in the TSCI. The characteristics of each CIC (such as magnitude, phase, or a function of such) may be calculated.

[0228] Multiple venues may exist. Each venue may have its own N10 device pairs. Different venues may have different numbers of N10. Each device pair may have its own N1 TSCIs. Different device pairs may have different N1s. For each venue, multiple states may exist (e.g., absent, present, no movement, human movement in the kitchen, non-human movement in the kitchen, human movement in the living room, non-human movement in the living room, etc.). For each venue, multiple sets of each N1 TSCI may be captured, each set corresponding to a different state.

[0229] In some embodiments, the wireless sensing system may include a deep learning-based system block. A deep neural network (DNN) (e.g., deep learning (DL), FNN, FCN, CNN, RNN, LSTM, transformer, network, deep network, network of networks, mixed expert (MoE), model, base model (FM), and / or large language model (LLM)) may be used to perform at least one wireless sensing task based on each N1 TSCI generated by N10 TX-RX device pairs. The wireless sensing system may comprise several main system blocks: (1) A TSCI generation block for each N10 TX-RX device pair, in which each RX generates N1 TSCIs based on the respective radio signals transmitted from each transmitter (TX) to each receiver (RX). N1 may be different for different device pairs. (Different device pairs may have the same / different N1, the same / different sounding frequencies, and / or the same / different radio signal bandwidths, resulting in the same / different amounts of CICs); (2) A data augmentation block that generates an augmented TSCI (including an augmented CI and / or an augmented CIC) based on the TSCI obtained in the CSI generation block and / or CSI preprocessing block. (The augmented TSCI may be used, for example, in controlled learning or predictive learning to generate an augmented DNN input for training / running a DNN); (3) A TSCI preprocessing block in which, for each TX-RX device pair, each N1 TSCI is converted / preprocessed into each M1 normalized input TSCI having a standardized / unified / normalized format / characteristics for the next block (embedding block). (M1 may be the same for all device pairs. The M1 normalized input TSCI having a normalized format may be used by the embedding block to generate DNN inputs for training / running the DNN); (4) For each TX-RX device pair, each M1 normalized input TSCI is used / processed to generate an embedding block that supplies a DNN input (embedding of the preprocessed input) to the next block (DNN). (TSCIs from different device pairs may have significantly different sampling characteristics / formats (e.g., sounding rate, amount of CIC) that are unsuitable for the embedding block, but M1 normalized input TSCIs with normalized characteristics / formats are readily used by the embedding block. The DNN input is in a format suitable for use by the DNN); (5) A deep neural network (DNN) block that takes DNN input as input and outputs task results. A DNN may include any of the following: neural network (NN), feedforward NN (FNN), convolutional NN (CNN), recurrent NN (RNN), long short-term memory (LSTM), transformer, autoencoder, fully connected (FCN), mixture expert (MoE), generative adversarial network (GAN), network of networks, and associated models (AI models), foundational models (FM), and / or large language models (LLM).

[0230] In the training phase of the DNN model, machine learning (e.g., supervised / unsupervised / semi-supervised / meta-learning, reinforcement training, few-shot training, transfer learning, etc.) may be used to train the DNN model (e.g., having a large set of model parameters) using a set of training data (e.g., a large set). The set of training data may include a set of TSCI obtained by various TX-RX device pairs that transmit wireless signals with various characteristics (sounding rate, carrier frequency, bandwidth) at various venues under various events / conditions / states and measure TSCI based on the received wireless signals. The device pairs may include a variety of wireless devices including access points (APs), client devices, IoT devices, home appliances, computing devices, smart devices, vehicle devices, etc. The training data may or may not be labeled. In unsupervised / semi-supervised training, the training data may not contain labels. In supervised / reinforcement / few-shot / meta / transfer training, some or all of the training data may be labeled. In machine learning, a loss function may be defined / computed, and the parameters of the DNN model may be adjusted to optimize (e.g., minimize) the loss function.

[0231] In some embodiments, the DNN of a wireless sensing system may be trained using a combination of contrast learning and predictive learning. The DNN may map each CI data to a latent / embedding point in the latent / embedding space. In contrast learning, a contrastive loss function is defined / calculated based on distance / similarity scores in the latent space (embedding space) and may be minimized for the set of training data by tuning the DNN model parameters. The DNN model parameters may be trained so that pairs of similar inputs (positive pairs, e.g., two augmented inputs, or augmented TSCI associated with a TSCI, or augmented DNN input associated with a DNN input) have small distance scores (or high similarity scores) in the latent / embedding space and therefore have a lower contrast loss. Pairs of dissimilar inputs (negative pairs, e.g., two different TSCIs associated with different events or different venues) should have large distances (or small similarity scores) and therefore a larger contrast loss.

[0232] In some embodiments, the DNN may generate predicted CI data. During predictive training, a reconstruction / predictive loss function (e.g., mean squared error, MSE) may be defined / calculated between the predicted CI data and the actual CI data and minimized over the set of training data by tuning the DNN model parameters.

[0233] In a wireless sensing system, the total loss function may be calculated / defined as an aggregate (e.g., weighted sum or weighted product) of the contrast loss function and the reconstruction loss function, or it may be minimized. In the weighted sum / product of the two loss functions, the weights may be fixed predefined values ​​(e.g., 0.1 / 0.3 / 0.5 / 0.7 / 0.9) throughout the training. Alternatively, during training, the weights may start from initial values ​​and gradually change as training progresses (e.g., according to a schedule or controlled by a rate parameter) (e.g., the weights gradually change from 0.1 to 0.5).

[0234] In some embodiments, the system may have a cloud edge architecture, where some system blocks (e.g., TSCI generation blocks, TSCI preprocessing blocks, data augmentation blocks, and / or embedding blocks) may reside at the edge (e.g., the processor of the RX in a TX-RX device pair, a local server, or device), and other blocks (e.g., DNN blocks, embedding blocks, data augmentation blocks) may reside in the cloud (e.g., within a cloud server).

[0235] In some cases, the transmission / storage of TSCIs is cumbersome and time-consuming, placing a very heavy load on transmission resources / bandwidth and storage resources / memory. Therefore, transmission of TSCIs from the edge to the cloud may be avoided by placing the TSCI generation block, TSCI preprocessing block, data augmentation block, and embedding block at the edge, and the DNN block at the cloud. All immediate processing of the TSCI may be performed at the edge (e.g., a processor or edge server associated with the RX) to generate the DNN input and augmented DNN input. The DNN input and augmented DNN input may be sent to the cloud (cloud server). DNN models with many parameters may be stored in the cloud, trained / tuned in the cloud, and executed in the cloud.

[0236] In some embodiments, the transmission of DNN inputs can still be substantial and critical. To avoid transmitting DNN inputs from the edge to the cloud, the DNN may be further broken down into two stages: (1) A CNN for extracting features from DNN input, and (2) Another neural network (e.g., transformer, autoencoder, FCN, RNN, LSTM) that takes the extracted features as input and outputs the task result. The Stage 1 CNN used to extract features from the DNN input may be further placed at the edge, while the Stage 2 neural network may be placed in the cloud. Features extracted by the Stage 1 CNN may be transmitted from the edge to the cloud. Typically, transmitting extracted features is less demanding than transmitting the DNN input (e.g., in terms of bandwidth, time delay, and memory storage).

[0237] In some cases, for each TX-RX device pair, the transmission of DNN inputs / extended DNN inputs or extracted features from the edge (e.g., the RX of the device pair) to the cloud may be further reduced by performing local motion detection at the edge (e.g., by a motion detection module on the RX of the device pair) based on N1 TSCIs (without using a DNN). The DNN inputs or extracted features may only be transmitted if motion is detected by the motion detection module. In the motion detection task, a similarity score may be calculated between two temporally adjacent CIs of the TSCI. Motion may be detected when the similarity score is greater than a threshold.

[0238] In some embodiments, the “DNN input” (or embedding) may be generated from M1 “normalized” “input” TSCIs for a deep neural network (DNN). A deep neural network (DNN) (e.g., deep learning (DL), FNN, FCN, CNN, RNN, LSTM, transformer, network, deep network, network of networks, mixed expert (MoE), model, foundational model (FM), and / or large language model (LLM)) may be used to perform at least one wireless sensing task based on a formatted input called the “DNN input” (having a specific format for the DNN, e.g., a kD matrix, a kD transformation matrix) generated / formatted / constructed from multiple CI / TSCIs. In deep learning (e.g., training / running / retraining / fine-tuning a DNN), to construct the DNN input, M1 normalized "input" TSCIs (with an associated normalized "input" sampling frequency of MF1, i.e., the normalized "input" CI of each normalized "input" TSCI is sampled as MF1 samples per second) may be used / needed / requested / expected / intended to be used in each input instance. Each CI of the M1 normalized "input" TSCIs may have M4 normalized "input" CICs. (The M1 normalized input TSCIs may be obtained / constructed from N1 TSCIs associated with one TX-RX device pair, or from each N1 TSCI associated with multiple TX-RX device pairs in the venue and / or additional venues.) A time series of the DNN input may exist.For each input instance, at least one "DNN input" to the DNN may be generated / built / assembled using M1 normalized input TSCIs (e.g., all normalized input CIs of M1 normalized input TSCIs within a sliding time window) (e.g., to be processed / analyzed by the DNN to train the DNN in the training phase or to perform at least one task).

[0239] In some embodiments, the DNN may be a mixed expert (MoE), or an expert within a MoE. Different experts within a MoE may have or use the same DNN input requirements / format / definition. Alternatively, different experts within a MoE may use or have different DNN input formats / requirements / definitions. With respect to a MoE, a device (e.g., AP) may select / determine / switch which one (or more) experts from the MoE are used / executed / invoke / selected to generate / calculate the respective DNN inputs associated with the selected expert in order to process M1 normalized input TSCIs (or the sliding time window of M1 normalized input TSCIs).

[0240] The DNN input to the DNN may be generated by applying one of the following operations to M1 normalized input TSCIs: preprocessing / processing / postprocessing / filtering, resampling, feature extraction, transformation, 1D / 2D / kD transformation at some granularity level, concatenation / joining / organization / assembly to form a matrix, and / or another operation. The DNN input may have a fixed size (e.g., a 1D / 2D / kD matrix with a fixed dimension and a fixed size in each dimension). The fixed dimension and / or fixed size may be initialized / determined before the task. Alternatively, the DNN input may have a flexible size that can be adaptively changed (e.g., changed during retraining / adaptation / learning / evolution in a task, task changes, and / or restriction / enhancement to a specific subtask of the task).

[0241] In some embodiments, one or more dimensions / elements of the DNN input may be masked (e.g., masked / forced to zero, replaced with zero, or made into some predefined pattern). A DNN input mask may be associated with the DNN input to indicate which elements are valid or invalid. The DNN input mask may have the same dimensions and size as the DNN input. Any DNN input mask element may be either "0" or "1". A mask element of "1" may mean that the corresponding element in the DNN input may be valid and may be used by the DNN. A mask element of "0" may mean that the corresponding element in the DNN input is invalid and may be ignored by the DNN (or should be ignored during DNN training / running / retraining / refinement). A logical operation (e.g., logical AND) may be performed between the mask element and the corresponding DNN input element to force the DNN input element to zero (or another default value). Alternatively, the DNN may choose to synthesize / generate padded / substituted / composite values ​​in place of invalid DNN input elements.

[0242] In some cases, the mask element may take additional values. For example, a mask element of "2" may mean that the corresponding element in the DNN input may be a generated element produced using some padding / composition / interpolation / generation algorithm and should be used with care by the DNN. In another example, a mask element of "3" may mean that the corresponding element in the DNN input may have been preprocessed / filtered / cleaned / denoised / conditioned in some way. Masks may be input to the DNN (e.g., as metadata). Masks may be included as part of the DNN input. For any DNN input, the associated DNN input mask may be used in the execution / run of the DNN. When a DNN input element is associated with a DNN input mask element of "1" (meaning "valid"), the DNN input element may be "valid" and may be used (in the usual way) in all associated DNN links (connections) within the DNN block.

[0243] In some embodiments, DNN input elements associated with a DNN input mask element of "0" (meaning "invalid") may be ignored / skipped / unused in the associated DNN links within the DNN. Alternatively, a DNN input element may be "invalid" and forced to have a value (e.g., zero) so that it has no effect in the DNN. For any DNN link directly connected to an "invalid" DNN input element, the model parameters associated with that "invalid" DNN input element may be considered temporarily "invalid" for the DNN input. Invalid DNN links may be excluded from the calculation of loss functions (e.g., contrast loss function, reconstruction loss function, total loss function). During the training phase, backpropagation may be used to propagate updated DNN model parameters backward. DNN model parameters for invalid links may not be updated for the DNN input. Calculations associated with invalid DNN links may not be performed (or may simply be forced to zero). Any node where all input links are invalid may be considered temporarily "invalid" for the DNN input.

[0244] In some embodiments, masking may be performed on a TSCI set, a TSCI, or a CI. There may be N1 TSCIs associated with a TX-RX device pair, each TSCI having N5 CIs within a certain period, and each CI having N4 CICs. From the N1 TSCIs, M1 normalized input TSCIs may be constructed, each normalized input TSCI having M5 normalized CIs within that period, and each normalized CI having M4 CICs. Any of these items may be valid / invalid / missing / noisy. A mask may be constructed / calculated / defined for each to mark the corresponding state.

[0245] Regarding TSCI set masks, some TSCIs in a set of N1 TSCIs and / or a set of M1 normalized input TSCIs may be invalid / missing / noisy / bad. Therefore, for N1 TSCIs / M1 normalized input TSCIs, a TSCI set mask with N1 / M1 mask elements may be constructed / calculated, where each mask element is associated with a respective TSCI. Similar to DNN input mask elements, any TSCI set mask element may be one of the following logical flags / values: "0" (to indicate an invalid TSCI), "1" (to indicate a valid TSCI), or "2" (e.g., to indicate a "padding," "composite," "denoised," or "processed" TSCI). Invalid TSCIs may not be used, but valid TSCIs may be used. Composite / padding / denoised / processed / borrowed TSCIs may be used / generated for invalid TSCIs and / or used to replace invalid TSCIs.

[0246] Regarding TSCI masks, in any of the N1 TSCIs / M1 normalized input TSCIs, some of the N5 / M5 CIs within that period may be missing / invalid. Therefore, for each N5 / M5 CI within each TSCI, a TSCI mask with N5 / M5 mask elements may be constructed / calculated, with each TSCI mask element associated with each CI within that period. Similar to DNN input mask elements, any TSCI mask element may be one of the following logical flags / values: "0" (to indicate an invalid CI), "1" (to indicate a valid CI), or "2" (e.g., to indicate a "padding," "composite," "denoised," or "processed" CI). Each padded / generated / replaced CI may be generated / composite for any invalid CI, and / or used to replace any invalid CI.

[0247] Regarding CI masks, in any CI of N1 TSCIs / M1 normalized input TSCIs, some of the N4 / M4 CICs within that period may be missing / invalid. Therefore, for the N4 / M4 CICs in each CI of each TSCI, a CI mask with N4 / M4 mask elements may be constructed / calculated, with each CI mask element associated with each CIC in the CI. Similar to DNN input mask elements, any CI mask element may be one of the following logical flags / values: "0" (to indicate an invalid CIC), "1" (to indicate a valid CIC), or "2" (e.g., to indicate a "padding," "composite," "denoised," or "processed" CIC). Each padded / generated / replaced CIC may be generated / composite for any invalid CIC, and / or used to replace any invalid CIC.

[0248] In some embodiments, data augmentation is applied to wireless sensing. Data augmentation may involve techniques for artificially augmenting / diversifying (i.e., increasing diversity and volume) training data by creating variations / modifications of existing data (e.g., N1 TSCIs, M1 normalized "input" TSCIs, "DNN inputs") without actually collecting new data. Data augmentation may be applied to each DNN input to generate one or more "augmented" DNN inputs for training / retraining / refining the DNN. Data augmentation may also be applied to CICs (whether normalized or not), CIs (whether normalized or not), TSCIs (whether normalized or not), and / or sets of TSCIs (whether normalized or not) to generate one or more "augmented" CICs, "augmented" CIs, "augmented" TSCIs, and / or sets of "augmented" TSCIs. Sets of augmented denormalized CICs / CIs / TSCIs / TSCIs may be used to generate / construct / form sets of one or more "augmented" normalized CICs / CIs / TSCIs / TSCIs. Any set of extended normalized CIC / CI / TSCI / TSCI may be used to generate extended DNN inputs for training / retraining / refining the DNN.

[0249] For example, data augmentation may be applied to TSCIs (whether normalized or not, or whether they are augmented or not) to generate augmented input TSCIs that can be used to construct an augmented set of M1 normalized input TSCIs to generate each augmented DNN input. In particular, k of the M1 normalized input TSCIs may be augmented to generate an augmented set of TSCIs (containing k augmented TSCIs and (M1-k) unaugmented TSCIs) to generate one or more DNN inputs, where k may be 0, 1, 2, ..., or M1. Different augmentation steps / techniques / procedures may be applied to different input TSCIs. Even if none of the M1 normalized input TSCIs are augmented, an augmented set of M1 TSCIs may be generated from the set of M1 normalized input TSCIs (for example, by shuffling the order of the set of M1 normalized input TSCIs, or by applying the same / common augmentation to all of the set of M1 normalized input TSCIs).

[0250] In some embodiments, there may be an imbalance in the training data set (training TSCI) for training a DNN across different classes. To generate more augmented training data, more augmentation may be applied to classes with insufficient / less / imbalanced training data. Even for classes with more / sufficient training data, augmentation may still be applied to a lesser extent to generate more diversity / diversification of the training data.

[0251] For matrix data, data augmentation may include geometric transformations, rotations, flips, cropping, scaling, translation, changes / adjustments to hue / luminance / contrast / saturation / dynamic range, edge enhancement, image filtering / blurring / sharpening, noise addition / injection, occlusion, masking of random regions, label preservation perturbations, etc. Tabular data can be represented as matrices with rows as samples (e.g., TSCI) and columns as features (e.g., links associated with TX and RX antennas). Data augmentation may include feature noise, feature shuffling (e.g., substitution / shuffling of non-critical columns), interpolation between samples (e.g., interpolation between rows), GANs for synthetic data, or synthetic sample generation (SMOTE). Time series data (e.g., TSCI) can be represented as matrices. Augmentation may include time warping, substitution, shuffling, magnitude scaling, window slicing, and temporal scaling / resampling / subsampling / interpolation.

[0252] In some embodiments, M1 normalized input TSCIs may be generated based on N1 TSCIs. The M1 normalized input TSCIs used to generate the DNN input may be obtained / constructed from N1 TSCIs of a device pair. (Alternatively, M1 normalized input TSCIs may be obtained / constructed from TSCIs of multiple device pairs. Multiple device pairs may be located / positioned / installed / present in similar / same sub-regions or sub-areas of the venue and therefore can monitor the same motion / state / activity in the sub-regions / sub-areas.) Since M1 may be fixed while N1 may differ for different device pairs, there are three cases: (1) M1=N1, (2) M1<N1、(3)M1> N1 is possible.

[0253] Case (1): When M1=N1, the M1 normalized input TSCIs do not have to be ordered (i.e., they are simply a set / collection of M1 TSCIs), and the M1 normalized input TSCIs can simply contain N1 TSCIs (i.e., only one combination). Alternatively, the M1 normalized input TSCIs may be ordered. There are (M1)! (i.e., M1 factorial, where (M1)!=(M1)*(M1-1)*(M1-2)*…*3*2*1) ways (permutations / shuffles) to generate M1 normalized input TSCIs from N1 TSCIs, resulting in (M1)! DNN inputs. In other words, there can be (M1)! input instances of M1 ordered normalized input TSCIs to generate (M1)! DNN inputs to train / run a DNN based on the same N1 TSCIs.

[0254] Case (2): When M1 < N1, the M1 normalized input TSCIs may include a subset of the N1 TSCIs. If the M1 normalized input TSCIs are not ordered, there can be [C_(N1)^(M1)] = [(N1)! / (N1 - M1)! / (M1)!] = (N1)*(N1 - 1)*(N1 - 2)*..*(N1 - M1 + 1) / M1 / (M1 - 1) / (M1 - 2) / … / 3 / 2 / 1 ways (or combinations) to generate the M1 normalized input TSCIs from the N1 TSCIs. In other words, there can be [(N1)! / (N1 - M1)! / (M1)!] input instances or DNN inputs (each containing M1 unordered input TSCIs) for training / executing the DNN based on the N1 TSCIs. If the M1 normalized input TSCIs are ordered, there can be [P_(N1)^(M1)] = [(N1)! / (N1 - M1)!] = (N1)*(N1 - 1)*(N1 - 2)*..*(N1 - M1 + 1) ways (permutations / shuffles) to generate the M1 normalized input TSCIs from the N1 TSCIs. In other words, there can be [(N1)! / (N1 - M1)!] input instances of the M1 ordered input TSCIs to generate [(N1)! / (N1 - M1)!] DNN inputs for training the DNN based on the N1 TSCIs.

[0255] Case (3): When M1 > N1, the N1 TSCIs can be used to construct a total of M1 composed TSCIs. Each set of the M1 composed TSCIs can be used to generate one (e.g., when the M1 TSCIs are not ordered) or multiple (e.g., when the M1 TSCIs are ordered) DNN inputs. Since M1 > N1, some of the N1 TSCIs can be "extended" by repeating / copying / cloning them one or more times (i.e., the repeated TSCIs can appear more than once among the M1 TSCIs).

[0256] Alternatively, one (or more) of the M1 configured TSCIs may be copied / borrowed / generated from one or more second TSCIs associated with a second TX-RX device pair (e.g., the original TX-RX device pair and the second TX-RX device pair may be in the same / similar sub-region of the venue, and therefore both can capture the same motion / state / activity). The second TSCI may be captured / acquired in the second TX-RX device pair with the same / different sampling parameters (e.g., different timing, different sampling / sounding rates, different amounts of CIC). In the copy / borrowing / generation process, the second TSCI may be fitted / resampled / interpolated / synchronized to match the rest of the M1 configured TSCIs.

[0257] Alternatively, one (or more) of the M1 configured TSCIs may be obtained by combining / merging / mixing two or more TSCIs (e.g., of N1 TSCIs and / or another N1 second TSCI associated with a second TX-RX device pair) to form a "combined" TSCI. For example, some of the N4 / M4 "combined" CICs of the "combined" CI of the combined TSCI may be the corresponding CICs of the CIs of the first TSCI (e.g., copied from there), while other parts of the N4 / M4 combined CICs of the combined CI may be the corresponding CICs of the CIs of the second TSCI (e.g., copied from there), and so on. Some of the N4 / M4 CICs of the CI of the combined TSCI may be an aggregation (e.g., a weighted sum / average) of each CIC of each CI of the first TSCI and each CIC of each CI of the second TSCI. The weights in any weighted aggregation may be the same / different for different CICs.

[0258] In some embodiments, the M1 normalized input TSCIs can be generated based on P1 TSCIs selected from N1 TSCIs. Alternatively, in any of cases (1), (2) or (3), P1 TSCIs (P1 < M1) can be selected from the N1 TSCIs (e.g., using maximum ratio combining / MRC), and the P1 selected TSCIs can be used to construct a total of M1 configured TSCIs for generating the DNN input (similar to case (3) above). In addition to including the P1 selected TSCIs among the M1 configured TSCIs, M1 - P1 additional TSCIs can be configured. Similar to case (3), some of the P1 selected TSCIs can be "expanded" by repeating / replicating / cloning them one or more times. One of the M1 configured TSCIs can be borrowed from another TX - RX device pair. Two or more of the P1 selected TSCIs (and / or any borrowed TSCIs from another TX - RX device pair) can be "combined" to form a combined TSCI that is used as one of the M1 configured TSCIs.

[0259] In some embodiments, for DNN inputs used to train / run / fine-tune a DNN, the standard / target / nominal / reference / required / unified / normalized sounding frequency (or the temporal rate of CI in the normalized input TSCI) may be assumed / required / expected / used / shared / common in M1 regularized input TSCIs. Associated with a device pair, the first sounding frequency may be the same / common among N1 TSCIs, and the second sounding frequency may be the same / common among M1 normalized input TSCIs. The second sounding frequency may be the standard sounding frequency for generating the DNN input and may be the same for all device pairs. However, the first sounding frequency may be the same / different from the second sounding frequency. If they are different, resampling may be performed to ensure that the M1 normalized input TSCIs have the second sounding frequency. Furthermore, the first sounding frequency can be different / the same for different device pairs (e.g., one could be 1 Hz (i.e., 1 sounding / second), another 10 Hz, and yet another 1000 Hz). Resampling (e.g., interpolation, 0th / 1st / 2nd / ... / higher-order interpolation, spline interpolation, decimation, subsampling, or any combination) can be performed on N1 TSCIs to generate M1 normalized input TSCIs such that the sounding frequencies of the resampled CI / CICs of some / all TSCIs associated with any device pair can have assumed / required / expected / used / shared / common sounding frequencies. Assume that a first TSCI with a first sounding frequency F1 is resampled to become a second TSCI with a second sounding frequency F2. Consider a sliding period T1. During the sliding period, there may be F1*T1 CI / CICs of the first TSCI, while during the sliding period, there may be F2*T1 CI / CICs of the second TSCI. Resampling may resample F1*T1 CI / CICs to generate F2*T1 CI / CICs based on some resampling filter.Resampling can occur before or after generating M1 normalized input TSCIs from N1 TSCIs.

[0260] In some embodiments, a time mask can be constructed / calculated / defined for each TSCI (whether normalized or not) within a given period. The time mask may include multiple time mask elements within a fine time standard (e.g., having a time resolution of 10ms for 100Hz, 1ms for 1000Hz, 0.2ms for 5000Hz, and 0.1ms for 10000Hz) that can represent samples at high sampling frequencies (e.g., 100 / 1000 / 5000 / 10000Hz), and each time mask element is associated with its respective relative timestamp (expressed with respect to the time standard) within that period. The time interval between consecutive time mask elements (the time difference between their timestamps) may be / reflect the time standard / resolution for representing the sampling timestamp. A time mask element of "1" may mean that the CI is valid / sampled at its respective timestamp, while a time mask element of "0" may mean that the CI is invalid / not sampled / not captured / not generated at its respective timestamp. Furthermore, there may be additional time mask element labels such as "2," which could indicate that each CI has been generated / combined using some method (e.g., zero-order retention / interpolation, first-order / linear interpolation, etc.). A time mask element label "3" could indicate that each CI has been preprocessed / cleaned / denoised / conditioned using some method. The time mask can be metadata for each TSCI. A time mask element label "4" could indicate that each has been resampled.

[0261] In some embodiments, data augmentation can be performed using time scaling and warping. Time scaling and / or time warping can be applied to any TSCI (e.g., any of N1 TSCIs or any of M1 normalized input TSCIs) to generate an augmented TSCI. Time scaling (e.g., time compression, or time extension from 7 seconds to 10 seconds) can be a data augmentation technique primarily used for time series data or continuous signals (e.g., TSCIs). It can change the temporal length (duration) of a signal (TSCI, or period / section of a TSCI) while preserving its essential characteristics. Time compression can shorten the duration of a signal (e.g., from 10 seconds to 7 seconds) by deleting samples (e.g., CIs of a TSCI) or resampling at a higher rate. Time extension can extend the duration of a signal (e.g., from 7 seconds to 10 seconds) by inserting samples or resampling at a lower rate. For example, a TSCI containing object motion that lasts for 10 seconds. Time compression can be applied to the TSCI to generate an augmented TSCI with object motion that lasts for 7 seconds. Time scaling can be linear or nonlinear. Dynamic time warping (DTW) can stretch or shrink segments nonlinearly to match a target length. Neural time scaling can use an autoencoder or generative adversarial network (GAN) to learn the optimal time warping function. Extended time masks can be associated with extended TSCI. Time scaling and / or time warping can be applied to the time mask of a TSCI to generate an extended time mask for an extended TSCI.

[0262] In some embodiments, data augmentation may be performed based on scaling / noise injection. By adding noise to each TSCI, CI, and CIC, augmented TSCI, augmented CI, and augmented CIC can be generated. To generate each augmented CIC, noise (e.g., Gaussian / Laplace noise, impulse noise, salt and pepper noise) may be added to one of the N4 / M4 CICs of the TSCI's CI. The resulting CI and TSCI become augmented CI and augmented TSCI by the augmented CIC. Noise with the same / different statistical distribution / behavior may be added to different CICs. To generate each augmented TSCI / CI / CIC, the features of the TSCI / CI / CIC (e.g., magnitude, magnitude squared, power) may be scaled / normalized.

[0263] In some embodiments, for each DNN input (i.e., each "to be augmented" DNN input) generated from M1 normalized input TSCIs to train / run the DNN, a number of related augmented DNN inputs can be generated (whether ordered or not, before or after resampling) to train / retrain / tune the DNN. Augmented DNN inputs can be generated by augmenting one or more of the M1 normalized input TSCIs, or by augmenting one or more of the N1 TSCIs. In generating / constructing augmented DNN inputs, the TSCI to be augmented (also called the TSCI to be replaced) can be replaced by augmented TSCIs generated from the TSCI to be augmented and / or the remaining TSCIs (i.e., the remainder of the M1 normalized input TSCIs and / or the remainder of the N1 TSCIs) using any data augmentation method.

[0264] In some embodiments, data augmentation may be performed based on the substitution of multiple TSCIs. In generating the augmented DNN input, some / all of the M1 normalized input TSCIs (or N1 TSCIs) may be augmented by applying substitution / shuffle / reordering to some / all of the M1 normalized input TSCIs (or N1 TSCIs). Individual TSCIs may or may not be modified. This substitution / shuffle / reordering is at the TSCI level.

[0265] In some embodiments, one or more CIs / matrices may be represented as vectors. Each CI may be represented as a vector. For example, a CI with N CICs may be an (N) tuple vector. Multiple vectors (e.g., CI vectors) may be represented jointly as a combined vector by concatenating (or combining / grouping / merging) each vector in some order. Numerous timestamped CIs of a TSCI may be concatenated / grouped / combined together (e.g., chronologically) to form a combined vector. For example, there may be N5 CIs of a TSCI within a certain period, and they may be represented as an (N*N5) tuple vector. Alternatively, a CI may be concatenated with a denoising version and / or transformed version of that CI to form a combined vector. Alternatively, multiple CIs / CICs from multiple TSCIs may be concatenated to form a combined vector. A two-dimensional matrix may be represented as a row-ordered vector or a column-ordered vector. Similarly, a k-dimensional matrix may be scanned in a scan order (e.g., a raster or zigzag scan order), and the scanned elements may then be represented as vectors.

[0266] In some embodiments, data augmentation may be performed using vector / subvector transformations / substitutions. Transformations and / or substitutions may be applied to any vector X (in vector space) or a subvector of X (in a subspace of vector space). A substitution is the arrangement of objects in a specific order. Substitutions (e.g., rearrangement / reorganization / reordering / shuffle) may be applied to elements of a vector / subvector or to a set of timestamped quantities (such as a set of CIs in TSCI that are naturally ordered based on timestamps). Transformations may be applied to vectors or subvectors. Transformations may include any of the following algebraic operations: addition (e.g., adding noise / Gaussian noise / impulse noise), subtraction, multiplication (e.g., multiplication by noise), division, scalar multiplication, substitution, shuffle, resampling, subsampling, truncation, quantization, extraction of subsets or subvectors, noise injection, and one-dimensional transformations. There are three possible ways to generate an extended input using extension / substitution / shuffle / transformation for (1) multiple TSCIs, (2) multiple CIs within a TSCI, and (3) multiple CICs within a CI.

[0267] (1) Expansion / Substitution / Transformation of Multiple TSCIs. First, expansion / transformation / substitution / shuffle can be applied to multiple TSCIs (e.g., N1 TSCIs, M1 normalized input TSCIs, or a subset thereof). Regardless of whether any / part / some / any / all of the individual input TSCIs are expanded, the multiple TSCIs can be ordered, and some / all of them can be substituted / reordered / rearranged / reorganized / shuffled (e.g., after substitution, the "i-th" TSCI becomes the "j-th" TSCI, where i and j are different).

[0268] (2) Extension / replacement / transformation of multiple CIs in a TSCI. Secondly, extension / transformation / replacement / shuffle can be applied to multiple CIs in a single TSCI. Multiple CIs may include CIs of a TSCI in one or more sliding time windows. Multiple sliding time windows may have the same / different window lengths. Each replacement / transformation / shuffle can be applied to each individual time window of a CI. Another replacement / transformation / shuffle can be applied to multiple time windows (for example, the "i-th" window may become the "j-th" window after replacement).

[0269] (3) Expansion / substitution / transformation of multiple CICs of a TSCI's CI. Thirdly, expansion / transformation / substitution / shuffle may be applied to multiple CICs of a CI of a single TSCI in order to generate an expanded CI, or to the elements of a vector in order to generate an expanded vector. Transformation / substitution / shuffle may include rotation and / or shift. In particular, to generate an expanded CI, the CIC of each CI may be rotated right / left and / or shifted right / left.

[0270] In some embodiments, data augmentation can be performed using matrix multiplication. Let X1 be an N-tuple vector (one-dimensional array, or rank-1 / order-1 tensor) that can represent one (or more) CIs having N4 / M4 CICs. Let X1(i) be the i-th element / component of X1. Let X2 be the augmented N-tuple vector obtained / generated by transforming / substituting / shuffling the N elements / components of X1. The augmented vector X2 can be computed by multiplying the vector X1 by an NxN square matrix A1 (e.g., N4xN4 if N=N4, or M4xM4 if N=M4), i.e., X2=A1*X1 (e.g., a transformation of rank-1 tensors, or a multilinear mapping). Thus, the augmentation can be linear.

[0271] In some embodiments, data augmentation can be performed using permutations. A1 may be an NxN identity matrix such that X2=X1. An identity matrix is ​​a diagonal matrix where the diagonal elements are 1 and the other elements are 0. Alternatively, A1 may be a permutation matrix, which can be an identity matrix with a permuted / shuffled column or row. Each permutation matrix has exactly one 1 in each column / row and the remaining elements in that column / row are 0. For example, A1 may be an inverse diagonal matrix where the inverse diagonal elements are 1 and the other elements are 0, which corresponds to inverting the elements of X1 so that the elements are in reverse order. For example, X2(1)=X1(N), X2(2)=X1(N-1), ..., X2(i)=X1(N+1-i), ...X2(N-1)=X1(2), X2(N)=X1(1).

[0272] In some embodiments, data augmentation can be performed using rotation. A1 may be an NxN identity matrix with its columns rotated right by k or left by Nk. Alternatively, A1 may be an identity matrix with its rows rotated up by k or down by Nk. X2 may be X1 rotated left (or up) by k, i.e., X2(i) = X1((i+k)mod N), where mod is the remainder. In other words, X2(1) = X1(1+k), X2(2) = X1(2+k), ..., X2(Nk) = X1(N), X2(N-k+1) = X1(1), X2(N-k+2) = X1(2), ..., X2(N) = X1(k).

[0273] In some embodiments, data augmentation can be performed using shifts. A1 may be an NxN identity matrix with its columns shifted to the right by k, or its rows shifted up by k. X2 may be X1 shifted to the left (or up) by k, i.e., X2(1)=X1(1+k), X2(2)=X1(2+k), ..., X2(Nk)=X1(N), X2(N-k+1)=0, X2(N-k+2)=0, ...X2(N)=0. A1 may also be an NxN identity matrix with its columns shifted to the left by k, or its rows shifted down by k. X2 can be X1 shifted to the right by k, i.e., X2(1)=0, X2(2)=0, ..., X2(N-k1)=0, X2(N-k+1)=X1(1), X2(N-k+2)=X1(2), ..., X2(N)=X1(k).

[0274] In some embodiments, data augmentation can be performed using cropping. The augmented vector X2 can be generated by cropping. Some elements of X1 can be forced to be 0. In some cases, some elements of X1 can be forced to be irregular (e.g., artificially large, very large). For example, all elements in the i-th row of A1 can be 0, so that the i-th element of X2 = A1 * X1 can be 0.

[0275] In some embodiments, the system can perform section expansion / substitution / transformation. An N-tuple vector X1 can be subdivided / divided into a number of "sections" or subvectors (e.g., first half, second half, or first / second / third / fourth quarters). Each expansion / substitution / transformation / shuffle can be applied to each individual section / subvector. Another expansion / substitution / transformation / shuffle can be applied to a number of sections / subvectors. Different sections / subvectors can have the same / different amounts of elements. For example, CI may be a CFR for a 160MHz channel, and the CI vector X1 can be subdivided into eight sections or eight subvectors, each having a CIC corresponding to its respective 20MHz bandwidth. Or, X1 can be subdivided into four sections / subvectors, each having a CIC corresponding to its respective 40MHz bandwidth. Or, X1 can be subdivided into two sections / subvectors, each having a CIC corresponding to its respective 80MHz bandwidth.

[0276] In some embodiments, each section / subvector may contain a set / assembly of consecutive vector elements (e.g., a CIC with consecutive indices). Alternatively, the vector elements within a section may not be consecutive (i.e., the indices may not be consecutive). For example, four sections / subvectors may be obtained by subsampling X1 with a coefficient of 4. The first section may contain the subsampled elements in the first subsampling phase (e.g., X1(1), X1(5), X1(9), ..., or X(4n+1) for n=0,1,...), while the second section may contain the subsampled elements in another phase (e.g., X1(2), X1(6), X1(10), ..., or X1(4n+2) for n=0,1,...). The third section may contain X1(3), X1(7), X1(11), ..., or X1(4n+3) for n=0,1,..., the fourth section may contain X1(4), X1(8), X1(12), ..., or X1(4n+4) for n=0,1,..., and so on. Different sections may have the same / different amounts of (ordered) vector elements / CIC. The number of sections can be ordered / indexed. Any respective transformation / substitution / shuffle (e.g., rotation / shift / invert / random) can be performed / applied independently in each section. Different / same transformations / substitutions / shuffles can be applied to different sections.

[0277] In some embodiments, some transformation / substitution / shuffle (e.g., rotation / shift / inversion / randomization) may be applied / performed on the number of sections. For example, the expanded vector X2 may be obtained by shifting or rotating the sections by k section positions (or k section indices), i.e., by shifting / rotating the i-th section to the (i+k)-th section. For example, the first section of X1 may be shifted to the second section of X2 (e.g., k=1), and the second section of X1 may be shifted to the third section of X2.

[0278] In some embodiments, CI interpolation may be performed in punctured transmissions. Puncture is a method of increasing channel efficiency / utilization / throughput and reducing latency by reducing the effects of interference on a radio channel, and can occur during the transmission of radio signals in one or more TSCIs of N10 TX-RX device pairs. In punctured transmissions in affected device pairs (e.g., preamble puncturing in 802.11be, BSS coloring in 802.11ax / be), one or more busy (e.g., with severe interference) or unavailable narrowband channels may be excluded ("punctured") from the broadband channel. As a result, some subbands / subcarriers of the broadband channel's bandwidth (i.e., some CICs in the CI) may be skipped / unused / unavailable (e.g., due to severe interference). For example, a particular 20 MHz subband within a 160 MHz band may be busy / blocked and may be skipped so that the available bandwidth is effectively 140 MHz (which may be a 160 MHz band with a 20 MHz "hole" or missing bandwidth). Without puncturing, a 160MHz communication channel may need to fall back to 80MHz due to interference. During punctured communication with associated “holes” or missing bandwidth in an affected TX-RX device pair over a period of time, some corresponding CICs of each associated CI captured by the TX-RX device pair over that period may be missing / unavailable / invalid. For each CI of N1 TSCIs associated with the affected device pair over that period, a first CI mask (as described above) can be constructed / defined, containing N4 CI mask elements (or components), each CI mask element / component associated with the respective CIC of the CI. Associated with the first CI mask, a second CI mask can be constructed / defined, containing M4 CI mask elements, for each CI of M1 normalized input TSCIs over that period.Any CI mask element can be any of the following logical flags / values: "0" (to indicate an invalid CIC), "1" (to indicate a valid CIC), or "2" (e.g., to indicate a "padded", "composite", "denoised", or "processed" CIC). To replace an invalid CIC in a CI, a composite / generated / replaced / borrowed / interpolated CIC may be generated. The generated CIC of a CI may be generated based on interpolation based on "adjacent" CICs. Adjacent CICs may include (a) other CICs of CIs that have adjacent CIC indices, and / or (b) CICs of temporally adjacent CIs within the same TSCI.

[0279] In some embodiments, data augmentation may be performed for multilink operation (MLO). Multiple bands may be used simultaneously for data / control transmissions. For example, in the case of a TX-RX device pair operating in multilink operation (MLO) (e.g., IEEE 802.11be (Wi-Fi 7)), three bands, namely 2.4 GHz, 5 GHz, and 6 GHz, may be used simultaneously. Multiple sets of N1 TSCIs may be obtained from the TX-RX device pair, each set associated with its respective band. N1 may be the same or different for different bands. The amount of CIC (N4) in different bands may be the same or different.

[0280] In multilink aggregation (MLA), multiple bandwidths across multiple links can be combined to achieve greater bandwidth. Each link may be associated with its respective TX-RX device pair and its respective set of N1 TSCIs. Different links may have the same / different N1 and / or N4. In multilink switching (MLS), multiple bandwidths across multiple links can be switched (e.g., one bandwidth at a time). Each bandwidth on each link may be associated with its respective set of N1 TSCIs. In asynchronous multilink (AML), each of multiple links may perform independent transmission, resulting in their respective sets of N1 TSCIs. In cooperative MLO (e.g., in 802.11bn), multiple APs can jointly manage a punctured channel to optimize the overall network performance. Each AP may generate its own set of N1 TSCIs. In co-BF (e.g., in 802.11bn), multiple (e.g., two) APs can simultaneously transmit to multiple target STAs on the same channel by coordinating beamforming and null steering to prevent interference. This can improve throughput / reliability and reduce latency in a multi-AP environment. Each AP can generate its own set of N1 TSCIs. Independent data augmentation can be applied to each set of N1 TSCIs. Multiple sets of each N1 TSCI can be combined / mixed / merged / replaced. The time mask can be shared by multiple TSCIs.

[0281] In some embodiments, a DNN input can be generated from M1 normalized input TSCIs for a deep neural network. To generate the DNN input, each normalized CI of the M1 normalized input TSCIs may be required / required / expected / assumed to have M4 normalized CICs. M4 may be a fixed / predefined number. However, since N4 in particular may differ for different device pairs, a CI of N1 TSCIs (of a device pair) having N4 CICs may not be directly used as a normalized CI (having M4 normalized CICs) due to the mismatch in the number of CICs. Thus, the M4 normalized CICs of a normalized CI may be constructed / transformed from some N4 "initial" CICs of some "initial" CI. Alternatively, a (M4) tuple normalized CI vector Y having M4 CICs / components / elements may be constructed from an (N4) tuple initial CI vector X having N4 initial components / elements / CICs. The vector Y can be calculated by multiplying the vector X by the matrix A2 of size M4xN4, i.e., Y = A2 * X. Often, the i-th row of matrix A2 can be a unit vector with (N4-1) zeros and one one at its j-th element, and consequently, the i-th element of Y becomes the j-th element of X. Sometimes, the i-th row of A2 can contain multiple non-zero elements that sum to 1, and consequently, the i-th element of Y becomes the weighted sum of the corresponding elements of X, with the corresponding weights shown in the i-th row of A2. The transformation from N4 initial CICs to M4 normalized CICs may occur before the construction of the M1 normalized input TSCI from the N1 TSCIs, such that each CI of the M1 normalized input TSCI can have M4 CICs. Alternatively, the conversion from N4 initial CICs to M4 normalized CICs may occur after the construction of the M1 normalized input TSCI, such that each CI of the M1 normalized input TSCI has N4 CICs, which can then be converted to M4 normalized CICs.

[0282] In some embodiments, M4 may be fixed while N4 may vary for different device pairs, resulting in three situations: (1) M4 = N4, (2) M4<N4、(3)M4> N4 is a possibility.

[0283] Case (1): When M4 = N4, the M4-tuple normalized CI vector Y can simply be the N4-tuple initial CI vector X (i.e., Y = X, or equivalently, the matrix A2 can be the identity matrix). Alternatively, the normalized CI vector Y can be a permutation of the initial CI vector (i.e., the matrix A2 can be an M4xM4 permutation matrix). There can be (M4)! different permutation matrices such that (M4)! different normalized CI vectors Y can be generated from each initial CI vector X. Data augmentation can be applied to each normalized CI vector Y, or to each DNN input generated by each normalized CI vector Y.

[0284]

[0285] Case (2): When M4 < N4, the normalized CI vector Y can include a subset of the elements of the initial CI vector X. There can be [P_(N4)^(M4)] = [(N4)! / (N4 - M4)!] = (N4)*(N4 - 1)*(N4 - 2)*..*(N4 - M4 + 1) ways (permutations / shuffles) to generate the M4-tuple normalized CI vector Y from the N4-tuple initial CI vector X. In other words, there can be [(N4)! / (N4 - M4)!] input instances of the M4-tuple normalized CI vector Y to generate [(N4)! / (N4 - M4)!] DNN inputs for training / running the DNN based on the N4-tuple initial CI vector X. One element of the M4-tuple Y can be a "summary" (e.g., linear combination, aggregation, weighted average / sum) of multiple elements of the N4-tuple X. The M4 elements of the M4-tuple Y can be selected from the N4 elements of the N4-tuple X based on an algorithm / criterion (e.g., maximum ratio combining / MRC, maximum eigenvalue, etc.).Case (3): If M4 > N4, then an M4 tuple Y requires more elements than an N4 tuple X holds, so at least M4-N4 elements must be generated for Y based on X. The first way to generate one or more M4 tuples Y from an N4 tuple X is zero padding, where zeros are filled around the elements of N4 tuple X to form an M4 tuple vector. A total of M4-N4 zeros may be filled at the beginning, at the end, or some at the beginning and some at the end. In some cases, some of the M4-N4 zeros may be added / placed / inserted among the N4 elements of X. The second way to generate an M4 tuple Y from an N4 tuple X is to "extend" X by repeating / duplicating / cloning / mirroring at least one element of X. For example, a periodic repetition of the elements of N4 tuple X may be formed, and M4 tuple Y may be any interval / section of the M4 elements of the periodic repetition. In the second example, the elements of an N4 tuple X can be mirrored to form a mirrored vector X2 containing approximately 2*N4 elements. Then, a periodic repetition of the elements of X2 can be formed, and the M4 tuple Y can be any interval / section of the M4 elements of the periodic repetition. The mirrored vector X2 can be either: (1) a (2*N4) tuple vector [X(1), X(2), ..., X(N4-1), X(N4), X(N4), X(N4-1), ..., X(2), X(1)] (where X(i) is the i-th element of X), (2) a (2*N4-1) tuple vector [X(1), X(2), ..., X(N4-1), (3) A (2*N4-1) tuple vector [X(1), X(2), ..., X(N4-1), X(N4), X(N4), X(N4-1), ..., X(2)] where X(1) appears once and X(N4) appears twice. (4) A (2*N4-2) tuple vector [X(1), X(2), ..., X(N4-1), X(N4), X(N4-1), ..., X(2)] where both X(1) and X(N4) appear once.

[0286] In some embodiments, a third method for generating an M4 tuple Y from an N4 tuple X is to generate each of the M4-N4 elements to be generated in the M4 tuple Y as a linear combination of each of one or more elements of the N4 tuple X. The generated M4-N4 elements can be placed / added / inserted at the beginning or end of the elements of the N4 tuple X, or some at the beginning and some at the end. In some cases, the generated elements can be placed / added / inserted among the N4 elements of X.

[0287] In some embodiments, a one-dimensional transformation of N11 points (e.g., autocorrelation function or ACF, short-time Fourier transform or STTFT) can be performed at a granularity level (e.g., per CIC, per TSCI, per device pair, per device pair). In a sliding time window, a one-dimensional (1D) N11 point transformation (e.g., ACF, STFT) can be computed for the CIC and / or CI of any TSCI associated with any of the N10 device pairs in the sliding time window (e.g., any of the N1 TSCIs, or the M1 normalized input TSCIs for each device pair, or the M1 normalized input TSCIs used to generate the DNN input for the DNN / LLM), where each one-dimensional transformation has N11 transformation components. Let X be an N11 tuple vector of CIC / CI having N11 vector components / CIC / CI. Let Y be an N11 tuple vector containing the one-dimensional transformation of X. A one-dimensional transformation can be linear, and as a result, Y can be calculated as Y=A*X, where A is the N11xN11 transformation matrix. A one-dimensional transformation can also be nonlinear, or it may contain both linear and nonlinear parts. A one-dimensional transformation can include any of the following: autocorrelation function (ACF), short-time Fourier transform (STFT), wavelet transform, Hadamard transform, or some frequency transform. A one-dimensional transformation can also be a trivial unit transformation for which no transformation is effectively performed for N11 CIC / CIs. Even though different device pairs may have the same / different N4, N7, N8, and / or N1, the same / different TX and RX configurations / or orientations, as well as the same / different system configurations / antenna types / antenna quantities / signal transmission / modulation (e.g., WiFi 2.4GHz, WiFi 5GHz, WiFi 6GHz, WiFi 60GHz, 3G / 4G / 5G / 6G / 7G / 8G, UWB, millimeter wave (mmWave), Bluetooth, etc.), N11 can be the same / different for all TSCIs associated with any of the N10 TX-RX device pairs. In particular, if all TSCIs are associated with the same sounding frequency, N11 can be the same for all TSCIs.

[0288] In some embodiments, the DNN input may be generated based on a one-dimensional transformation of N11 points. N11 point transformation coefficients of several TSCIs (e.g., M1 normalized input TSCIs) may be used to generate the DNN input (or M1 normalized input TSCI) for the DNN. Deep learning (DL, e.g., supervised learning, unsupervised learning, self-supervised learning) may be used to train or fine-tune a DNN / model / FM / LLM (for a task / subtask) using training TSCIs acquired / generated from various device pairs having the same / different N4, N7, N8, N1, and the same / different training placement / direction / venue, sampling / sounding frequency, carrier frequency bandwidth, signal transmission / modulation, etc. The trained DNN / model / FM / LLM may be used for various / arbitrary / all device pairs with various N4, N7, N8, and / or N1. A trained DNN / model / FM / LLM can be adapted / adopted / prompted / fine-tuned to function for various sensing tasks (e.g., "downstream" tasks) (e.g., with various additional DNNs). Fine-tuning can modify the DNN / model / underlying model / LLM itself, while prompting can change how the DNN / model / FM / LLM is used. There can be N9 sliding time windows within a given period (e.g., 0.1 / 1 / 10 / 100 / 1000 / 10000 seconds), and many periods can exist along the time axis. Adjacent periods may or may not overlap.

[0289] In some embodiments, different N1s may be used for different sliding time windows of different durations. The one-dimensional transformation can be performed for sliding time windows of different durations. Assume that the sounding frequency associated with the TSCI is F1. There may be a first sequence of sliding time windows, each with a duration T1, and a second sequence of sliding time windows, each with a duration T2. ​​For each sliding time window in the two sequences of sliding time windows, the one-dimensional transformation can be calculated. For the first sliding time window with a duration T1, the first N11 point one-dimensional transformation can be calculated based on the CI of the TSCI in the first sliding time window, with the first N11 being F1*T1. For the second sliding time window with a duration T2, the second N11 point one-dimensional transformation can be calculated based on the TSCI in the second sliding time window, with the second N11 being F1*T2. A short sliding time window (e.g., T1 = 10 seconds) may reflect short-term behavior, while a long sliding time window (e.g., T2 = 1000 seconds) may reflect long-term / longer-term behavior.

[0290] In some embodiments, a one-dimensional transformation of N11-point or N13-point with resampling may be performed. N11 may differ for different TSCIs. For example, a first TSCI may be associated with a first sounding frequency F1 (e.g., 10 Hz), and a second TSCI may be associated with a second frequency F2 (e.g., 50 Hz). The one-dimensional transformation may be performed on both TSCIs in their respective time windows associated with a common sliding period T1 (e.g., 1 / 10 / 100 / 1000 seconds). N11 of the first TSCI in the common period may be T1*F1, while another N11 of the second TSCI in the common period may be T1*F2. If two N11s are different for different TSCIs, then N11 one-dimensional transformation coefficients of the first TSCI in a common period may be resampled (e.g., interpolated, decimated, and / or subsampled) to give N13 (which may be equal to another N11) resampled transformation coefficients, where N13 can be the same for all TSCIs (e.g., N13 may be associated with a required / expected / used / shared / common sounding frequency). Alternatively, N11 CIs / CICs in a sliding period of the first TSCI (e.g., a common period T1 in the first / second TSCIs) may be resampled to give N13 resampled CIs / CICs (where N13 is the same for all TSCIs), and an N13-point one-dimensional transformation may be applied to the N13 resampled CIs in the sliding period to give N13 (resampled) transformation coefficients. The (resampled) transformation coefficients of N13 points can be used to generate DNN inputs to DL / DNN / models / FM / LLM using a common format / presentation / representation.

[0291] Different granularity levels may include per TSCIC, per TSCI, per device pair, and for all device pairs. One-dimensional transformations may be calculated at different granularity levels (e.g., per TSCIC / per component level, per TSCI level, per device pair level, or for all device pairs). A hierarchy of granularity levels may be established / constructed / determined. In that hierarchy, the per component / per TSCIC granularity level may be lower than the per TSCI granularity level (i.e., the per component / per TSCIC granularity level may be "lower" than the per TSCI granularity level), which may be lower than the per device pair granularity level, which may be even lower than the all device pairs granularity level. In any case, a one-dimensional transformation at a given granularity level may be a transformation at that granularity level (e.g., of CI or CIC) from a first domain (e.g., time domain) to a second domain (e.g., transformation domain), or vice versa, associated with that one-dimensional transformation. The transformation domain may be a "time lag" domain if the one-dimensional transformation is an ACF, or it may be a frequency domain if the one-dimensional transformation is a Fourier / wavelet / trigonometric function transformation (e.g., FFT).

[0292] In some embodiments, higher-grained quantities may be calculated based on multiple lower-grained quantities. A higher-grained quantity in a time window (e.g., a one-dimensional transformation, or a two-dimensional transformation, or dynamic statistics) may be calculated based on multiple lower-grained quantities in that time window. In particular, a higher-grained one-dimensional / two-dimensional / k-dimensional transformation may be an aggregation (e.g., sum / product / average, possibly weighted) of multiple lower-grained one-dimensional / two-dimensional / k-dimensional transformations. A one-dimensional / two-dimensional / k-dimensional transformation per TSCI may be an aggregation of one-dimensional / two-dimensional / k-dimensional transformations per multiple components (or per TSCIC). If each CI has N4 CICs, then a one-dimensional / two-dimensional / k-dimensional transformation per TSCI may be an aggregation of one-dimensional / two-dimensional / k-dimensional transformations per N4 TSCICs. The 1D / 2D / kD transformation for each device pair may be a first aggregation of 1D / 2D / kD transformations for multiple components (per TSCIC), a second aggregation of 1D / 2D / kD transformations for multiple TSCIs, or a third aggregation of the first and second aggregations. The 1D / 2D / kD transformation for all device pairs may be a first aggregation of 1D / 2D / kD transformations for multiple components / per TSCIC, a second aggregation of 1D / 2D / kD transformations for multiple TSCIs, a third aggregation of 1D / 2D / kD transformations for multiple device pairs, or a fourth aggregation of the first, second, and third aggregations. A 2D transformation may be constructed based on multiple 1D transformations. Recursively, a kD transformation may be constructed based on multiple (k-1)D transformations.

[0293] In some embodiments, when more lower-grained-level quantities (among several lower-grained-level quantities) are used to compute a higher-grained-level quantity, the resulting higher-grained-level quantity may be more substantial / representative / enlightening / reliable / confidential. Thus, one or more granular-level scores (e.g., confidence / reliability / confidence / reliability scores) may be computed for or associated with each granular-level quantity. The granular-level scores may be non-negative. The granular-level scores may be real numbers (e.g., non-negative) between an upper and lower bound. For example, it may be between -1 and 1, or between 0 and 1, or between 0 and 10, or between 0 and 100. The granular-level scores (or associated with) a higher-grained-level quantity (e.g., a one-dimensional transformation of a higher granular-level) may be greater when more lower-grained-level quantities (e.g., one-dimensional transformations of lower granular-levels) are used to compute the higher-grained-level quantity. The higher-grained-level quantities may be monotonically increasing / non-decreasing with respect to each lower-grained-level quantity. A quantity at a higher granularity level (e.g., a one-dimensional transformation) may be calculated as a weighted sum / average / quantity of multiple quantities at lower granularity levels. In calculating a weighted sum / average / quantity, quantities at lower granularity levels that have a higher granularity level score may be given or assigned a greater weight (i.e., quantities that are more substantial / representative / enlightening / certain / reliable / confidential are given a greater weight as reflected by their associated granularity level score). The weights may be a function of the associated score, which may be monotonic (e.g., monotonically increasing / non-decreasing). The weights may also be a function of the associated transformation index. Alternatively, all weights may be equal for all quantities at lower granularity levels (i.e., no weights). Any weights used in any weighted sum / average / quantity may be normalized.

[0294] In some embodiments, a component-level / TSCIC-level one-dimensional transformation (N4 / M4 for each TSCI, and N1*N4 / N1*M4 for each device pair) may be performed. At the component-level / TSCIC-level granularity (i.e., the lowest level), a component-level (or TSCIC-level) one-dimensional N11 point transformation (e.g., ACF, STFT) associated with the CIC of the CI of the TSCI of the device pair may be calculated for each N10 device pair in the sliding time window, for each N1 TSCI, for each N4 CIC of the CI (e.g., the first CIC of all CIs of each TSCI in the sliding time window). In the sliding time window, there may be N4 TSCIC-level / component-level one-dimensional transformations corresponding to each N4 CIC of the CI of each TSCI, with each component-level one-dimensional transformation associated with each of the N4 CICs of that TSCI. For each N1 TSCI of a device pair (which is a TX-RX pair containing a type 1 device (TX) and a type 2 device (RX)), there may be N1*N4 component-based / TSCI-based one-dimensional transformations. For M1 "input" TSCIs, each of which has M4 CICs for each "input" CI used to generate DNN inputs for training a DNN / model / FM / LLM, there may be M4 TSCIC-based one-dimensional transformations for each of the M4 "input" TSCICs of each input TSCI in a sliding time window, and a total of M1*M4 TSCIC-based one-dimensional transformations may exist for the M4 input TSCIs.

[0295] In some embodiments, a one-dimensional transformation at the TSCI level (one for each TSCI, N1 for each device pair) may be performed as an aggregation of N4 component-level / TSCI-level one-dimensional transformations. At the granularity level of each TSCI (or radio link level associated with a particular TX antenna and a particular RX antenna), a one-dimensional transformation for each TSCI associated with a device pair's TSCI (associated with each TX antenna and each RX antenna) may be calculated as an aggregation of N4 component-level one-dimensional transformations for that TSCI of the device pair within a sliding time window. There may be N1 one-dimensional transformations for each TSCI for a device pair, each associated with one of the N1 TSCIs of the device pair. For M1 normalized input TSCIs, there may be M1 one-dimensional transformations for each TSCI within a sliding time window. Alternatively, a one-dimensional transformation for each TSCI, such as an ACF, may be calculated directly for the N11 CIs of each TSCI. The calculation of a one-dimensional transformation (e.g., ACF or other correlation / covariance-based one-dimensional transformation) may include the product of two CIs. To calculate such a product, each of the two CIs may be represented as an N4 tuple CI vector having each N4 vector element / CIC (or an M4 tuple vector having each M4 CIC). The product of two CIs may also be the dot product of each of the two CI vectors. Aggregations may include any of the following: sum, product, weighted sum, weighted product, mean, weighted mean, arithmetic mean, geometric mean, harmonic mean, percentile, median (50th percentile), minimum (0th percentile), maximum (100th percentile), trimmed mean (mean of percentiles from A1 to A2 percentile), mode, etc.

[0296] In some embodiments, each higher-grained level one-dimensional transformation (e.g., one-dimensional transformation per TSCI, ACF per TSCI, STFT per TSCI, etc.) may be a weighted aggregate (e.g., weighted average / product / average) of the corresponding lower-grained level one-dimensional transformations. For example, each one-dimensional transformation per TSCI (e.g., ACF / STFT per TSCI) may be a weighted average of N4 associated one-dimensional transformations per TSCI / component. The weights may be calculated based on the Maximum Ratio Synthesis (MRC). The weights may all be equal so that the weighted average is simply an unweighted average. Some of the lower-grained level one-dimensional transformations may be selected to calculate the weighted average, while the rest may not be selected. For example, N12 of the N4 CICs (referred to as "selected CICs") or TSCICs (referred to as "selected TSCICs") may be selected to compute a one-dimensional transformation for each TSCI, while the remaining N4-N12 CICs (referred to as "unselected CICs") or TSCICs (referred to as "unselected TSCICs") may not be selected. The weights for the selected lower-grained one-dimensional transformations (e.g., N12 selected CICs / TSCICs) may be non-zero, while the weights for the unselected lower-grained one-dimensional transformations (e.g., N4-N12 unselected CICs / TSCICs) may be zero. The weights for the lower-grained one-dimensional transformations may be associated with, a function of, or proportional to, that lower-grained one-dimensional transformation evaluated in a particular transformation index. For example, the weights for the selected CIC / TSCIC (i.e., the weights for the associated component-specific / TSCIC-specific one-dimensional transformations) may be the associated component-specific / TSCIC-specific one-dimensional transformations evaluated in a predefined non-zero transformation index (e.g., a first non-zero transformation index). Alternatively, the weighted aggregation (e.g., a weighted average) may be a simple unweighted aggregation (e.g., an unweighted average / product / average) where all weights are equal. The weights may be normalized.For example, the weights for N12 selected CIC / TSCICs may be normalized to unity such that their sum is unity (i.e., 1).

[0297] In some embodiments, features (or characteristic values) of each lower-grained-level one-dimensional transformation may be calculated. The selection of the chosen lower-grained-level one-dimensional transformations may be based on an analysis / comparison of those lower-grained-level one-dimensional transformations. The features may include (or be based on) one or more local maxima and / or local minima of the one-dimensional transformation (or one or more local maxima / local minima of the first / second / higher-order derivatives of the one-dimensional transformation). The features may be one-dimensional transformations evaluated at a particular transformation index. Several lower-grained-level one-dimensional transformations having the largest (or smallest) features may be selected as the chosen lower-grained-level one-dimensional transformations. If there are N12 selected lower-grained-level one-dimensional transformations, then N12 lower-grained-level one-dimensional transformations having the top N12 features (or bottom N12 features) may be selected as the N12 chosen lower-grained-level one-dimensional transformations. For example, in order to calculate a one-dimensional transformation for each TSCI, all N4 component-specific / TSCIC one-dimensional transformations may be analyzed and compared to select N12 selected CIC / TSCICs. For each of the N4 component-specific / TSCIC one-dimensional transformations, a characteristic value for each component-specific / TSCIC (e.g., a one-dimensional transformation feature) may be calculated. Then, the N12 selected CICs may be selected based on the characteristic values ​​for each of the N4 components. The N12 selected CICs may include, or may not include, CICs associated with any of the following: the maximum (100th percentile) characteristic value, or the top (or maximum) N12 characteristic values, or multiple maxima, or the minimum (0th percentile) characteristic value, or the bottom (or minimum) N12 characteristic values, or multiple maxima, or the median (50th percentile) characteristic value, or a group of N12 percentile characteristic values ​​around the median.

[0298] In some embodiments, the characteristic value of any one-dimensional transformation (e.g., a one-dimensional transformation per component / per TSCIC, or per TSCI, or per device pair, or per device pair) may be a first aggregate (e.g., minimum) of all or a subset of the N11 transformation values ​​of that one-dimensional transformation. The subset of transformation values ​​may be associated with a subset of transformation indices, such as a range of transformation indices from transformation index a1 to index a2. a1 may be N11*b1, and a2 may be N11*b2. b1 may be 0.1 / 0.2 / 0.3 / 0.4 or other values. b2 may be 0.6 / 0.7 / 0.8 / 0.9 or other values. The first aggregation may include, or may not include, percentiles, maximum value (100th percentile), minimum value (0th percentile), median (50th percentile), mean, weighted mean, arithmetic mean, geometric mean, harmonic mean, trimmed mean (sum of percentiles from A1 percentile to A2 percentile, e.g., (A1,A2)=(0,10), or (3,17), or (33,66), or (90-100), or (83,97)), the difference between two percentiles (A2 percentile minus A1 percentile, e.g., (A1,A2)=(0,100), (10,90), (25,75)), and / or the transformed value in a particular transformation index (e.g., a first non-zero transformation index). In some embodiments, the particular characteristic value is the smallest transformed value in a subset of the transformation index. In particular, the characteristic value may be the minimum value of a subset of the transformation values, and each of the subsets of the transformation values ​​has a transformation index in the range from a1=N11*b1 to a2=N11*b2, with b1=0.2 and b2=0.8. The N12 selected CICs may be the N12 CICs associated with the lower N12 characteristic values.

[0299] In some embodiments, a one-dimensional transformation at the level of N10 device pairs may be performed as an aggregation of one-dimensional transformations for N1*N4 components or for N1 TSCIs. At the granularity level of each device pair, a first one-dimensional transformation for a device pair of TX and RX devices may be calculated as an aggregation of all one-dimensional transformations for each N1*N4 component associated with each of the N1 TSCIs of the device pair in a sliding time window, and each of the N4 CICs of the CI. A second one-dimensional transformation for the device pair may be calculated as another aggregation of all one-dimensional transformations for each of the N1 TSCIs associated with each of the N1 TSCIs of the device pair in a sliding time window. There may be N10 first (or second) one-dimensional transformations for each device pair.

[0300] In some embodiments, to compute a one-dimensional transformation for each device pair, one-dimensional transformations for N13 selected components (out of N1*N4) may be performed. The first one-dimensional transformation for each device pair may be a weighted average of all or a subset of the one-dimensional transformations for N1*N4 components / TSCICs (e.g., using Maximum Ratio Synthesis (MRC)). Of the N1*N4 component / TSCIC one-dimensional transformations, N13 may be selected to compute the first one-dimensional transformation for each device pair, while N1*N4-N13 component one-dimensional transformations may not be selected, where N13 < (N1*N4). The weights for the N13 selected component one-dimensional transformations (in the weighted average) may be non-zero, while the weights for the N1*N4-N13 unselected ones may be zero. The weights for the one-dimensional transformations for each selected component may be those one-dimensional transformations for each selected component, evaluated using a predefined non-zero transformation index (e.g., a first non-zero transformation index). The weights for the N13 one-dimensional transformations for each selected component may be normalized such that their sum is unity (i.e., 1).

[0301] In some embodiments, all N1*N4 component-specific one-dimensional transformations may be analyzed and compared in order to select one-dimensional transformations for each of the N13 selected components. Then, one characteristic value for each of the N1*N4 component-specific / TSCIC may be calculated for each of the N1*N4 component-specific / TSCIC one-dimensional transformations. Then, the N13 selected component-specific one-dimensional transformations may be selected based on the characteristic values ​​for each of the N1*N4 components. The N13 selected component-specific one-dimensional transformations may include, or may not include, one-dimensional transformations for each component associated with the maximum (100th percentile) characteristic value, or the top (or maximum) N13 characteristic values, or multiple maxima, or the minimum (0th percentile) characteristic value, or the bottom (or minimum) N13 characteristic values, or multiple maxima, or the median (50th percentile) characteristic value, or a group of N13 percentile characteristic values ​​around the median. The one-dimensional transformation for each of the N13 selected components may be associated with the characteristic values ​​for each of the lower (minimum) N13 components / TSCIC.

[0302] In some embodiments, to calculate the one-dimensional transformation for each device pair, N14 (out of N1) may be selected as the one-dimensional transformation for each selected TSCI. The one-dimensional transformation for each second device pair may be a weighted average of the one-dimensional transformations for all or a subset of the N1 TSCI (e.g., using maximum ratio combining (MRC)). While the number N14 of the one-dimensional transformations for the N1 TSCI may be selected to calculate the one-dimensional transformation for each second device pair, the one-dimensional transformations for N1 - N14 TSCI may not be selected, where N14 < N1. The weights for the one-dimensional transformations for the N14 selected TSCI may be non-zero, while the weights for the N1 - N14 non-selected ones may be zero. The weight for the one-dimensional transformation for each selected TSCI may be that one-dimensional transformation for each selected TSCI evaluated at a predefined non-zero transformation index (e.g., the first non-zero transformation index). The weights for the one-dimensional transformations for the N14 selected TSCI may be normalized such that their sum is unity (i.e., 1).

[0303] In some embodiments, one-dimensional transformations for N14 TSCIs may be selected by analyzing the characteristic values ​​for all N1 TSCIs. To select one-dimensional transformations for each of the N14 selected components, all one-dimensional transformations for N1 TSCIs may be analyzed and compared. Then, the characteristic values ​​for N1 TSCIs may be calculated one by one for each of the one-dimensional transformations for N1 TSCIs. Then, one-dimensional transformations for N14 selected TSCIs may be selected based on the characteristic values ​​for N1 TSCIs. One-dimensional transformations for N14 selected TSCIs may include, or may not include, one-dimensional transformations for each TSCI associated with the maximum (100th percentile) characteristic value, or the top (or maximum) N14 characteristic values, or multiple local maxima, or the minimum (0th percentile) characteristic value, or the bottom (or minimum) N14 characteristic values, or multiple local maxima, or the median (50th percentile) characteristic value, or a group of N14 percentile characteristic values ​​around the median. The one-dimensional transformation for each of the N14 selected TSCIs may be associated with the characteristic values ​​for each of the lower (minimum) N14 TSCIs.

[0304] In some embodiments, the one-dimensional transformation at the level of all device pairs may be performed as an aggregation of one-dimensional transformations for each N10*N1*N4 component, or for each N10*N1 TSCI, or for each N10 device pair. At the granularity level of all device pairs, the first one-dimensional transformation of all device pairs associated with all device pairs may be calculated as an aggregation of one-dimensional transformations for each N1*N4 component associated with each N1 TSCI, each N4 CIC of each CI, for all N10 device pairs in a sliding time window. The second one-dimensional transformation of all device pairs associated with all device pairs may be calculated as another aggregation of one-dimensional transformations for each N1 TSCI, each N1 TSCI associated with each N10 device pair in a sliding time window. A third one-dimensional transformation of all device pairs associated with all device pairs may be calculated as yet another aggregation of the one-dimensional transformations for each of the N10 device pairs associated with the N10 device pairs in a sliding time window.

[0305] In some embodiments, the system may select lower-grained one-dimensional transformations to compute the one-dimensional transformations for all device pairs. The one-dimensional transformations for all first / second / third device pairs may be a weighted average of all or a subset of each lower-grained one-dimensional transformation. Some of the lower-grained one-dimensional transformations may be selected to compute the one-dimensional transformations for all device pairs, while the remaining lower-grained one-dimensional transformations may not be selected. The weights for the selected lower-grained one-dimensional transformations (in the weighted average) may be non-zero, while the weights for the unselected ones may be zero. The weights for the selected lower-grained one-dimensional transformations may be those selected lower-grained one-dimensional transformations evaluated at a predefined non-zero transformation index (e.g., the first non-zero transformation index). The weights of all selected lower-grained one-dimensional transformations may be normalized such that their sum is unity.

[0306] In some embodiments, the system may select a chosen lower-grained one-dimensional transformation by analyzing characteristic values ​​at lower granularity. All lower-grained one-dimensional transformations may be analyzed and compared in order to select a chosen lower-grained one-dimensional transformation. Characteristic values ​​may be calculated for each lower-grained one-dimensional transformation, and the selected lower-grained one-dimensional transformation may be selected based on those characteristic values. The selected lower-grained one-dimensional transformation may include, or may not include, one associated with the maximum (100th percentile) characteristic value, or multiple upper (or maximum) characteristic values, or multiple local maxima, or the minimum (0th percentile) characteristic value, or multiple lower (or minimum) characteristic values, or multiple local maxima, or the median (50th percentile) characteristic value, or multiple percentile characteristic values ​​around the median. The selected lower-grained one-dimensional transformation may be associated with multiple lower (minimum) characteristic values.

[0307] In some embodiments, the system may perform sensing using motion statistics (MS) in a sliding time window. Generally, a sensing task (e.g., monitoring / detecting / estimating / recognizing / counting the presence / motion / respiration / vital signs / falls / abnormalities / movement / position / activities of daily living of one or more objects) may be performed for each sliding time window by calculating at least one motion statistic (MS) based on each N1 TSCI of each N10 device pair in the sliding time window. For example, motion detection may be performed by comparing the MS to a threshold. If at least one MS is greater than a certain threshold, motion of an object may be detected. The motion statistics (MS) may be calculated at different granularities, i.e., for all device pairs, per device pair, per TSCI, and / or per CIC / TSCIC / component. The granularity level of the MS may be selected to achieve different trade-offs between the performance of the sensing task and its sensitivity to changes in (a) the spatial arrangement of device pairs, (b) the number of spatial streams (or radio links, or antenna pairs), or (c) the number of CICs (or channel bandwidth). The MS may be calculated as, or based on, a feature of the one-dimensional transformation of N11 points.

[0308] In some embodiments, component-specific / TSCIC-specific MS may be calculated for each TSCIC of each TSCI in each device pair (or for each N4 TSCIC of each N1 TSCI in each of the N1 device pairs) based on the CICs of multiple CIs of the TSCI in a sliding time window. There may be N4 component-specific MS for each TSCI in a sliding time window. There may be N1*N4 component-specific MS for each device pair in a sliding time window. In some embodiments, TSCI-specific MS may be calculated for each TSCI in each device pair. TSCI-specific MS for a particular TSCI of a particular device pair may be calculated as an aggregate of all N4 component-specific MS associated with the N4 CICs of the CI of that particular TSCI of that device pair. There may be N1 TSCI-specific MS for a device pair, with each TSCI-specific MS associated with one of each of the N1 TSCIs. In some embodiments, the first MS per device pair for a device pair may be calculated as an aggregate of the MS for each N1 TSCI associated with each N1 TSCI of that device pair. The second MS per device pair for that device pair may be calculated as another aggregate of the MS for each N1*N4 component associated with each N1 TSCI of that device pair, each N4 CIC of the CI. There may be an MS for N10 device pairs associated with N10 device pairs. In some embodiments, the first MS for all device pairs may be calculated as an aggregate of the MS for all N10 device pairs associated with N10 device pairs. The second MS for all device pairs may be calculated as another aggregate of the MS for each N10*N1 TSCI associated with each N1 TSCI of all N10 device pairs (assuming N1 is the same for all device pairs).The third MS for all device pairs may be yet another aggregation of the MS for all N10*N1*N4 components associated with all N1 TSCIs for all N4 TSCICs in all N10 device pairs (assuming N1 is the same for all device pairs and N4 is the same for all TSCIs).

[0309] In some embodiments, any MS at a certain granularity level may be calculated based on at least one one-dimensional / k-dimensional transformation matrix at that granularity level. Any MS may include, or may not include, a time-reverse resonance intensity (TRRS), the dot product of two adjacent CI vectors, a similarity score between two adjacent CI vectors, or a feature of a one-dimensional transformation. The feature may be a one-dimensional transformation evaluated at a specific transformation index / coefficient (e.g., a specific time lag such as one sampling period, or an ACF evaluated at tau = 1 / Fs). The feature may be close to zero in a stationary environment and positive in a dynamic environment where the object is in motion. Thus, motion may be detected when the MS is greater than a certain threshold.

[0310] In some embodiments, the system may utilize statistics of all device pairs and sensitivity to the spatial arrangement of N10 device pairs. When a sensing task is performed using statistics of all device pairs of the sensing system (e.g., using statistics of all device pairs as input to a neural network) (e.g., using a 1D / kD transformation / MS / TSCI of all device pairs), the sensing system may tend to be sensitive to the spatial arrangement of N10 device pairs. The advantage of using statistics of all device pairs is that, assuming the spatial arrangement of N10 device pairs does not change in the future, the statistics of all device pairs allow the system to have very good / excellent sensing task performance. The disadvantage of using statistics of all device pairs is that if the spatial arrangement of N10 device pairs changes / moves / modifies (e.g., moving one or more device pairs within the same venue, or simply being in a different / new venue), the sensing task performance / characteristics of the sensing system may be significantly affected / degraded / worsened.

[0311] In some embodiments, the system may utilize per-device-pair MS and sensitivity to the spatial arrangement of N10 device pairs. If the sensing task is performed / implemented using per-device-pair statistics of the sensing system (e.g., using per-device-pair statistics as input to a neural network) (e.g., using per-device-pair 1D / kD transformation / MS / TSCI), the sensing system may not be sensitive to the spatial arrangement of N10 device pairs. This is because many / varied / excessive spatial arrangements of N10 device pairs may be used to train / tune / configure the sensing system. The advantage of using per-device-pair statistics is that the sensing system should perform well when there is no change in the number of antenna pairs or spatial streams. The disadvantage is that it may still be sensitive to any changes / modifications / differences in the antenna pairs / spatial streams in each device pair.

[0312] In some embodiments, the system may perform sensing using MS over a certain time period. Sometimes, sensing tasks (monitoring / detecting / estimating / recognizing / counting of an object's presence / motion / respiration / vital signs / falls / abnormalities / movement / position / activities of daily living) may be performed incorrectly / sporadically due to noise, potentially leading to false positives / false negatives. To avoid false positives / false negatives, monitoring / detecting / estimating / recognizing / counting of an object may be determined to be positive / negative in that time period only if it is valid / repeated over a sufficient number of sliding time windows in that time period (i.e., when the percentage of sliding time windows in which object monitoring / detection / estimation / recognition / counting is positive / negative is greater than a threshold (e.g., 30%, 50%, 70%, 90%)). Additional (temporarily adjacent / neighboring) time periods may exist in which the motion of an object is monitored / detected / estimated / recognized / counted. Alternative methods for detecting the motion of an object over a time period. Alternatively, an extended time period may be found during which the motion of an object can be detected. This extended time period may be divided into that time period and an additional time period.

[0313] In one embodiment / example, multiple classification / detection results may include the presence or absence of a user / intruder. The classification / detection may be performed when a moving object is detected during that time period.

[0314] In some embodiments, the system may perform the task of assembling / combining / concatenating one-dimensional transformations to form a two-dimensional transformation. Multiple one-dimensional transformations associated with the TSCI may be concatenated / grouped / combined with one another to form a k-dimensional ("k-dimensional", e.g., 2, 3, 4, or more dimensions) transformation matrix for computation / generation / preparation for input to a DNN / model / FM / LLM. There may be N9 sliding time windows in that time period. For each of the N9 sliding time windows in the time period, a one-dimensional N11-point transformation for each TSCI may be computed based on the TSCI, and it should be noted that each N11-point one-dimensional transformation is represented as an N11-tuple vector. For a TSCI, N9 vectors associated with N9 one-dimensional N11-point transformations (at the granularity of each TSCI) during that time period may be assembled / combined / concatenated to form a two-dimensional (2D) matrix of size N11xN9 (at the granularity of each TSCI) (called the "transformation matrix" or "spectral matrix"), where each one-dimensional transformation (vector) is a column of the two-dimensional matrix. The horizontal axis of the two-dimensional transformation matrix may be the time axis, and the vertical axis may be the transformation axis (for example, the frequency axis if the one-dimensional transformation is STFT, or the time lag axis if the one-dimensional transformation is ACF). For each TSCI of each device pair, an N11xN9 two-dimensional transformation matrix may be calculated. Alternatively, for each of the N9 sliding time windows during that time period, a one-dimensional N11-point transformation for each TSCIC may be calculated based on the TSCIC of the TSCI, where each N11-point one-dimensional transformation is represented as an N11 tuple vector. For a TSCIC of a TSCI, the N9 vectors associated with N9 one-dimensional N11-point transformations (at the granularity level of each TSCIC) over that time period may be assembled / combined / concatenated to form a two-dimensional (2D) matrix (transformation / spectral matrix) of size N11xN9 (at the granularity of each TSCIC), where each one-dimensional transformation (vector) is a column of the two-dimensional matrix. The horizontal axis of the two-dimensional transformation matrix may be the time axis, and the vertical axis may be the transformation axis.For each TSCIC in a TSCI, an N11xN9 two-dimensional transformation matrix may be calculated. Since each CI in a TSCI may have N4 CICs, there may be N4 such N11xN9 two-dimensional transformation matrices for that TSCI. Note that N4 may be different for each device pair.

[0315] In some embodiments, the system may perform the construction of a k-dimensional transformation matrix. For each device pair, there may be N1 TSCIs, and therefore N1 associated 2D transformation matrices of size N11xN9. The N1 2D transformation matrices may be combined / assembled / concatenated to form a 3D transformation matrix of size N1xN11xN9. For a venue, there may be N10 device pairs, and therefore N10 associated 3D transformation matrices of size N1xN11xN9. The N10 3D transformation matrices may be combined / assembled / concatenated to form a 4D transformation matrix of size N10xN1xN11xN9. In some situations, multiple venues (e.g., living room, dining room, kitchen, second floor, etc.) may be considered together for a given task. Multiple 4D transformation matrices may be further combined / assembled / concatenated (e.g., recursively) to form a transformation matrix of 5 dimensions or more. Any k-dimensional transformation matrix may be computed at one of the following granularity levels: per TSCIC / per TSCI / per device pair / for all device pairs.

[0316] For each k-dimensional transformation matrix for each TSCIC, when a one-dimensional transformation (per component) for each TSCIC is used, the resulting k-dimensional transformation matrix (e.g., a two-dimensional transformation matrix of size N11xN9) is a k-dimensional transformation matrix for each component. For each of the N9 sliding time windows in that time period, there may be N4 one-dimensional N11-point transformations for each component calculated based on each CIC of the CI of each TSCI. For each TSCI, there may be N4 two-dimensional transformation matrices for each component of size N11xN9. For each device pair, there may be N4 three-dimensional transformation matrices for each component of size N1xN11xN9. For each venue, there may be N4 four-dimensional transformation matrices for each component of size N10xN1xN11xN9.

[0317] For each k-dimensional transformation matrix for each TSCI, when a one-dimensional transformation is used for each TSCI, the resulting k-dimensional transformation matrix (e.g., a two-dimensional transformation matrix of size N11xN9) is the k-dimensional transformation matrix for each TSCI. For each of the N9 sliding time windows in that time period, there may be one one-dimensional N11-point transformation for each TSCI, calculated based on the CI of each TSCI. For each TSCI, there may be one two-dimensional transformation matrix for each TSCI, each of size N11xN9. For each device pair, there may be one three-dimensional transformation matrix for each TSCI, each of size N1xN11xN9. For each venue, there may be one four-dimensional transformation matrix for each TSCI, each of size N10xN1xN11xN9.

[0318] In some embodiments, regardless of any / all differences between different device pairs, including differences in bandwidth, number of TX antennas, number of RX antennas, etc., all 2D, 3D, and 4D transformation matrices may be of size N11xN9, N1xN11xN9, and N10xN1xN11xN9, respectively, for all granularities. Different device pairs may have different N1 and N4, but the size of the 2D transformation matrix may be independent of N1 and N4. The dimensional invariance of such 2D transformation matrices may make them suitable as inputs to a CNN. This also makes it possible to use 2D transformation matrices at a first granularity (e.g., a lower granularity, such as per TSCI) as additional / supplementary / appropriate 2D transformation matrices to train a deep learning network that expects 2D transformation matrices at a second granularity (e.g., a higher granularity, such as per device pair). In particular, when the training data (e.g., 1D transformations, 2D transformation matrices) at the second granularity is insufficient, some of the data at the first granularity may be used as supplemental / additional / reasonable training data for training the DNN.

[0319] In some embodiments, the system may perform the construction of a “d-transform” matrix as input to a DNN / model / FM / LLM (instead of a transformation matrix). Alternatively, the derivatives (e.g., first-order / second-order / cubic) or differentials (called “1D d-transform” or “1D Δ-transform”) of a one-dimensional transformation may be computed at any granularity (e.g., per TSCIC / per TSCI / per device pair / overall device pair). A one-dimensional d-transform may be used instead of a one-dimensional transformation to generate / construct a k-dimensional transformation, and the resulting k-dimensional transformation is called a k-dimensional d-transform. In particular, all one-dimensional d-transforms over a time period may be assembled / concatenated to form a two-dimensional d-transform matrix of size N11xN9, where each one-dimensional d-transform is a column of the two-dimensional d-transform matrix. The horizontal axis of the two-dimensional d-transform matrix may be the time axis, and the vertical axis may be the transformation axis. Then, recursively, multiple k-dimensional d-transform matrices may be assembled / combined / concatenated to form a (k+1)-dimensional d-transform matrix. A k-dimensional matrix may be supplied as input, or it may be used to construct inputs to a DNN / model / FM / LLM.

[0320] In some embodiments, the system may use a CNN as the feature extraction module of a DNN / Model / FM / LLM. A k-dimensional transformation matrix and / or an additional k-dimensional transformation matrix (or, a k-dimensional d-transformation matrix and an additional k-dimensional d-transformation matrix) may be supplied as input to the DNN feature extraction module to compute / derive / extract features from the k-dimensional transformation matrix (or k-dimensional d-transformation matrix). The feature extraction module may be a convolutional neural network (CNN) or an encoder of a transformer architecture. In each layer of the CNN, multiple convolutional filters may be applied to each of the multiple transformation / d-transformation matrices in parallel / simultaneously / sequentially / independently (e.g., in any order). The output of the CNN may be rearranged / flattened / concatenated / reorganized to form a data structure for input to a Stage 2 network. The Stage 2 network may compute an output analysis for each outcome class. Among multiple output analyses, the largest one may be identified and the associated outcome class may be selected as the classifier output.

[0321] In some embodiments, during the training phase, both the Stage 1 network and the Stage 2 network may be trained using training data (e.g., labeled training data). For each outcome / event / action / situation class, each collection of training TSCIs may be obtained based on training radio signals transmitted from a Type 1 training device to a Type 2 training device when the situation / event / action associated with the outcome class occurs. However, for some reason, there may be insufficient / not enough training data available for a particular outcome / event / action / situation class. During the action / inference phase, a one-dimensional transformation and a subsequent two-dimensional transformation matrix may be computed at a target granularity level (e.g., per-device-pair level) based on the TSCIs obtained based on the radio signals transmitted from the Type 1 device to the Type 2 device.

[0322] During the training phase, a one-dimensional transformation and its subsequent two-dimensional transformation matrix may be computed at two or more granularity levels, including the target granularity level and at least one other granularity level. For example, at least one other granularity level (e.g., per TSCI level) may include levels lower than the target granularity level (e.g., per device pair level). For outcome classes with sufficient training data, the one-dimensional transformation and its subsequent two-dimensional transformation matrix may be computed only at the target granularity level. However, for at least one outcome class with insufficient training data, the one-dimensional transformation and its subsequent two-dimensional transformation matrix may be computed at both the target granularity level (e.g., per device pair) and at least one other level (e.g., per TSCI level). In addition to the two-dimensional transformation matrix at the target level, at least one other level's two-dimensional transformation matrix may be used as training data to train the deep learning network.

[0323] In some embodiments, the base model may be used for wireless sensing. The base model may be a type of large-scale machine learning model (e.g., an AI model) trained on a massive amount of wireless sensing data (e.g., N1 TSCIs for each device pair) to perform a wide range of wireless sensing tasks. These models may be designed to serve as a general starting point for various downstream applications such as wireless sensing, motion detection, presence detection, intrusion detection, respiration / vital sign detection, sleep monitoring, occupancy detection, daily living activity monitoring, fall detection, motion recognition, gesture recognition, gait recognition, localization / positioning, navigation, multimodal tasks, speech enhancement, speech activity detection, natural language processing, images, video, speech, audio, etc. The base model is typically pre-trained on a large dataset and then fine-tuned for a specific task.

[0324] The foundational models may have the following key characteristics: (1) Scale. They may be trained on large datasets and may have billions of parameters. (2) Generalization. They may be adapted to multiple tasks across different domains. (3) Transfer learning. They may leverage knowledge from pre-training to improve performance on specific tasks with minimal additional training. (4) Generalizability. They may be used for tasks such as text generation, translation, summarization, image recognition, and wireless sensing tasks such as motion / presence / location estimation, breathing detection, fall detection, and more. Some examples of foundational models for natural language processing (NLP) may include (1) the GPT® (Generative Pre-trained Transformer) series such as GPT-3® and GPT-4®, (2) BERT (Bidirectional Encoder Representations from Transformers), and (3) T5 (Text-to-Text Transfer Transformer). Some examples of multimodal foundational models may include (1) CLIP (Contrastive Language Image Pretraining) and (2) DALL-E (Generating Images from Text Prompts). Some examples of foundational models for computer vision may include (1) Vision Transformers (ViT). Some challenges for foundational models may include (1) high computational costs for training, (2) potential bias in the data, and (3) environmental impact due to energy consumption.

[0325] In some embodiments, the underlying models may be constructed using various architectures depending on the type of data (e.g., text, images, audio, TSCI, etc.) and the tasks they may be designed to perform. Some exemplary architectures used for the underlying models include transformer architectures, pre-training and fine-tuning, scalability, and transfer learning.

[0326] Transformer Architecture: The underlying model may be based on a transformer architecture. Key features may include: (a) a self-aware mechanism, which may allow the model to handle long-range dependencies and dynamically weight the importance of different parts of the input data; (b) scalability, which may allow the transformer to handle large amounts of data and efficiently parallelize computations; and (c) a layered structure, which may include multiple encoder and decoder layers (although some may use only encoders or decoders). Encoders may have a stack of layers, and decoders may have a stack of identical layers. Each layer may have two sub-layers, namely a multi-head self-aware mechanism and a position-based fully coupled feedforward network.

[0327] Pre-training and fine-tuning: The foundational model may be pre-trained on a large dataset using unsupervised or self-supervised learning objectives (e.g., masked language modeling, next token prediction). After pre-training, the model may be fine-tuned on a specific downstream task using smaller, task-specific datasets. This two-phase approach may allow the model to learn general features from a large dataset and then specialize for a specific task.

[0328] Scalability: The underlying model may be designed to scale with increasing computing resources and data. This may include scaling up: (a) the model size, which is the number of parameters (e.g., billions or trillions); (b) the data size, which is the amount of training data (e.g., terabytes of text or images); and (c) the computing resources used for training on large GPUs / TPUs. The underlying model may also be designed to scale with more data and more parameters.

[0329] Transfer Learning: A key function of a foundational model may be its ability to transfer knowledge from one task to another. This may be facilitated by a pre-training phase, during which the model can learn a wide range of features and patterns that may be generally useful across many tasks.

[0330] In some examples, for each training data of the base model (e.g., each training data includes at least one of the following: a 1D / kD transformation matrix, a 1D / kD matrix of CI measurements, a 1D / kD matrix of other measurements, or at least one sequence of 1D / kD matrices; the other measurements may include any of the following: RSSI, interference information, other radio measurements such as system state / settings / parameters, audio, imaging, video, pressure, etc.), multiple related derived / augmented training data may be computed by performing data augmentation on the training data and / or additional training data. Multiple methods for performing data augmentation may exist. The system may perform swapping / rearrangement / interchange / reorganization / modification of multiple submatrices of a 1D / 2D / kD matrix. For a 1D matrix / vector, a submatrix may simply be a 1-D submatrix containing some of the elements / components of the 1D matrix / vector. The elements / components (i.e., corresponding component indices) may be continuous or discontinuous. For a 2-D matrix, a submatrix may be a 1-D submatrix such as a row vector containing a row or part of a row, a column vector containing a column or part of a column, a diagonal / anti-diagonal vector containing a diagonal (or off-diagonal) or anti-diagonal (or off-anti-diagonal) element / component, or a direction vector containing a matrix element / component in a certain direction. A submatrix may also be a 2-D submatrix containing multiple row and column components / elements (i.e., components in two directions, each direction being either row, column, diagonal, anti-diagonal, or any direction). The 2-D submatrix may contain multiple continuous / discontinuous rows / columns. The 2-D submatrix may include sampling of matrix elements / components (e.g., subsampling, horizontal / vertical / diagonal / anti-diagonal / directional subsampling, periodic / aperiodic / random sampling). The 2-D submatrix may also be a rectangular submatrix (a "full" rectangular submatrix containing all the submatrix elements obtained from the 2-D matrix). The 2-D submatrix may be a set of submatrix rows / columns of different lengths (some submatrix elements may be empty, i.e., they may not be obtainable from the 2-D matrix, forming a "non-full" rectangular submatrix).Similarly, for a kD matrix, submatrices may be 1-D submatrices, 2-D submatrices, 3-D submatrices, ..., (k-1)-D submatrices, or kD submatrices. A 1-D submatrice may be a vector / 1D submatrice containing elements / components in rows / columns / directions. A 2-D submatrice may contain elements / components in two directions (e.g., any two of the k dimensions, or any two "diagonal" or "anti-diagonal" directions, or any two arbitrary directions). A (k2)-D submatrice may contain elements / components in k2 directions for any k2 <= k (e.g., any k2 of the k dimensions, any k2 "diagonal" or "anti-diagonal" directions, or any k2 arbitrary directions). A k2-D submatrice may contain sampling of matrix elements / components in k2 directions (e.g., subsampling, directional subsampling, random sampling). The (k2)-D submatrix may be a "rectangular" submatrix (a "full" submatrix where all submatrix elements are derived from the kD matrix). The (k2)-D submatrix may also be a set of smaller-dimensional submatrixes of different submatrix sizes (forming a "non-full" submatrix where some elements are empty, i.e., not derived from the kD matrix).

[0331] In some cases, a single submatrix may be reorganized / modified. A submatrix may be divided into multiple partitions, each partition containing some of the elements / components of the submatrix. Partitions may be disjoint or disjoint. A submatrix (or a partition of a submatrix) may be rotated, with its components / elements rotated in a manner similar to how bits are rotated in bitwise rotation of a byte / word in a computer. A submatrix (or a partition of a submatrix) may be shifted, with its components / elements shifted in a manner similar to bitwise shifts. A submatrix (or a partition of a submatrix) may be permuted / shuffled, with its components / elements being permuted / shuffled / rearranged in many possible orders.

[0332] Scaling. Submatrices (or partitions of submatrices) may be scaled. Each submatrix component / element may be scaled by its respective scaling factor. The scaling factors for different submatrix / element components may be the same or different. Submatrices (or partitions of submatrices) may be replaced by similar patches / submatrices / partitions. Submatrices (or partitions of submatrices) may be resized (made larger or smaller). The resizing factor may be different for different dimensions. Submatrices (or partitions of submatrices) may be moved to another location. Submatrices (or partitions of submatrices) may be distorted along the axes. Noise (e.g., Gaussian, impulse, salt and pepper) may be added to submatrices (or partitions of submatrices). Submatrices (or partitions of submatrices) may be filtered (e.g., low-pass filtering, high-pass filtering, contrast enhancement, edge detection, smoothing). Submatrices (or partitions of submatrices) may be removed. It may be replaced / masked by a predefined value (e.g., zero) or pattern (e.g., noise patch / submatrix / partition). A submatrix (or partition of a submatrix) may be replaced by blending (e.g., weighted average) of two or more patches / submatrixes / partitions. A submatrix (or partition of a submatrix) may be replaced by a patch / submatrix / partition from another matrix. Patches / submatrixes / partitions may be synthesized based on that other matrix.

[0333] In some embodiments, two (or more) submatrices may be swapped / shuffled. For a 1D vector / matrix, two vector components may be swapped. The two components may be adjacent to each other with a distance of 1 (of their respective component indices). The two components may not be adjacent to each other with a distance greater than 1. The component indices of the two components may be predefined constants or dynamically generated (adaptively modified) quantities. For a 2-D matrix, two 1-D submatrices may be adjacent to each other with a distance of 1 (e.g., one submatrix is ​​row 2 and the other is row 3, with a row distance or offset distance of 1), or they may not be adjacent to each other with a distance greater than 1 (e.g., one is row 2 and the other is row 5, with a distance of 3). For that 2-D matrix, the two submatrices may be 2-D submatrices. The two 2-D submatrices may have an "offset distance" of 1 or greater than 1. Similarly, for a kD matrix, the (k2)-D submatrix may have an offset distance of 1 or greater than 1.

[0334] Figure 1 shows an exemplary framework of a system 100 for wireless sensing using a base model, according to several embodiments of the present disclosure. As shown in Figure 1, the system 100 may collect CSI 111 based on a radio signal transmitted from one or more IoT devices 102 to a router 104. Generally, CSI 111 may be any channel information (e.g., CSI, CFR, CIR, etc.) collected based on a radio signal transmitted from a transmitter to a receiver. In some embodiments, the transmitter may function as a bot (e.g., a type 1 device), while the receiver may function as an origin (e.g., a type 2 device). The bot may transmit a radio signal to an origin in a venue (e.g., a house) to obtain channel information of a wireless multipath channel based on the radio signal, where the channel information of the wireless multipath channel may be affected by the movement / presence of any object / user in the venue.

[0335] In some embodiments, the edge device 110 (e.g., a local device or local server) may process the CSI 111 to generate a processed CSI 118. In the example shown in Figure 1, the edge device 110 includes a base engine 112 configured to determine whether a trigger event (e.g., motion detection) has occurred based on the CSI 111. If motion is detected by the base engine 112, the edge device 110 may, in operation 114, extract the most recent past portion of the CSI 111 within the most recent past time period (e.g., the last 5 seconds, the last 10 seconds). In operation 116, the edge device 110 may process the extracted CSI data based on data augmentation. For example, the edge device 110 may select subcarriers for at least one of the extracted CSI data to ensure that all extracted CSI data have the same number of subcarriers according to a standardized format suitable for the underlying model 122, and each of the extracted CSI data may be resampled to a predetermined time rate according to the standardized format. In some embodiments, the data augmentation performed may further include at least one of the following: adding random noise to the extracted CSI data; randomizing selected subcarriers within a block; performing time scaling or time warping on the extracted CSI data; simulating at least one environmental parameter related to multipath changes or shielding; and normalizing the amplitude of the extracted CSI data to mitigate power fluctuations. In this way, the edge device 110 may generate a processed CSI 118 having a standardized format readable by the base model 122.

[0336] The edge device 110 may send the processed CSI 118 to the cloud server 120. As shown in Figure 1, the cloud server 120 includes a base model 122, a plurality of task-specific models 124, and a user interface 126. For example, the processed CSI 118 may be sent to the base model 122 for training and / or execution. In some examples, with the processed CSI 118 as input, the base model 122 may output a feature map representing channel-related functions and / or sensing-related functions. The feature map may be used by each of the plurality of task-specific models 124 to perform a corresponding wireless sensing task, such as motion detection, user presence detection, respiration / heart rate detection, fall detection, intruder detection, etc. In some embodiments, one or more of the task-specific models 124 are selected to perform the corresponding wireless sensing task using the feature map generated by the base model 122, for example, based on a user instruction during the inference phase. Thus, the base model 122 is always used to perform one of the wireless sensing tasks, while each of the task-specific models 124 is used to perform only one of the corresponding wireless sensing tasks.

[0337] In some embodiments, multiple task-specific models 124, which are downstream task models, may be trained based on the base model 122. For each downstream task, the system may fine-tune the downstream task models and the base model together using supervised data, or it may freeze the base model for all tasks and tune only the downstream task models. In some examples, the multiple task-specific models 124 may be trained by freezing all model parameters of the base model 122 while training the multiple task-specific models 124. In some examples, the multiple task-specific models 124 may be trained by fine-tuning all model parameters of the base model 122 based on at least one task-specific prediction loss while training the multiple task-specific models 124. In some examples, the multiple task-specific models 124 may be trained by freezing the model parameters of the upstream layer of the base model 122 and fine-tuning the model parameters of the downstream layer of the base model 122. Each of the multiple task-specific models 124 may be a downstream model compared to the base model 122.

[0338] In some embodiments, the base model 122 may be trained when the model design is finalized and a certain amount of data has been collected. In some embodiments, the downstream task models may be trained after the base model 122, which can capture several high-level representations, has been trained and the designed models for each downstream task are ready. Labeled data may be used to train the downstream models, where the output of the encoder of the base model 122 may be the input to the downstream models. The base model 122 may be updated on the cloud server 120. For example, when more data is collected, the base model 122 may be retrained, and its weights may be updated and stored on the cloud server 120.

[0339] In some embodiments, the base model 122 may be trained on self-supervised machine learning without labeled data. Each of the multiple task-specific models 124 may be trained on labeled data.

[0340] In some embodiments, each of the multiple task-specific models 124 may be used together with the base model 122 to perform a corresponding one of a plurality of wireless sensing tasks. The sensing results of the wireless sensing task may be presented to one or more users via a user interface 126, which may be a website-based user interface or an app-based user interface. In some embodiments, different sensing results of the wireless sensing task may be presented via different user interfaces.

[0341] In some embodiments, during the training phase, the training dataset may be generated by the edge device 110 and sent from the edge device 110 to the cloud server 120. The foundation model 122 and multiple task-specific models 124 may be trained by the cloud server 120.

[0342] During the inference phase, multiple wireless sensing tasks may be performed by: collecting and processing real-time CI data by at least one local or edge device to generate processed real-time CI data; determining, by at least one local device, whether a trigger event has occurred based on the processed real-time CI data; transmitting the most recent past portion of the processed real-time CI data within the most recent past time period from at least one local device to a cloud server in accordance with the determination that a trigger event has occurred; and by the cloud server executing a wireless sensing task corresponding to the trigger event based on the most recent past portion of the processed real-time CI data, using a trained foundational model and trained task-specific models corresponding to the wireless sensing tasks.

[0343] In some embodiments, the CSI 111 may be a raw CSI generated by the router 104 based on data sent to the router 104 from one or more IoT devices 102. In some examples, the edge device 110 may be a device coupled to or integrated with the router 104. If there is a trigger (e.g., motion detection) on the edge device 110, the edge device 110 can send processed CSI 118 (following a unified format) from the past few seconds to the cloud server 120 as input to the base model 122.

[0344] Figure 2 shows exemplary processes 210, 220 for learning and running a base model for wireless sensing according to some embodiments of the present disclosure. In some embodiments, the base model may be implemented as base model 122 in Figure 1. In the example in Figure 2, the base model may be implemented as an encoder. In some examples, process 210 represents a process for learning the base model, while process 220 represents a process for running the base model.

[0345] During process 210, the training dataset (e.g., CSI data used for training) may be processed by CNN211 based on some data augmentation techniques, such as those described above, before being used to train the base model. In this example shown in Figure 2, two copies of the base model are represented by two encoders 212 and 213, respectively, and are trained to minimize the contrast loss function 216 and the reconstruction loss function 217, respectively, and simultaneously. In some embodiments, the entire base model, including CNN211 and encoders 212 and 213, may be located on a cloud server. In some embodiments, to reduce data transmission costs, CNN211 may be located on an edge device, while encoders 212 and 213 may be located on a cloud server.

[0346] In some embodiments, the training dataset may include multiple CI pairs, original CI data, and masks. In some embodiments, the multiple CI pairs include positive CI pairs formed by a preprocessed CI and its associated extended CI, positive CI pairs formed by two preprocessed CIs, positive CI pairs formed by two extended CIs, negative CI pairs formed by two CIs obtained from two different radio channels, negative CI pairs formed by two CIs obtained from two different venues, and negative CI pairs formed by two CIs associated with two different sensing events.

[0347] The contrast loss function 216 may be determined by the contrast head 214 of the underlying model based on a first similarity metric (e.g., embedding distance) between the CI data of each CI pair in the training dataset. In some embodiments, determining the contrast loss function involves mapping each CI in the training dataset to a corresponding embedding point in the embedding space using the underlying model; for each CI pair containing two CIs in the training dataset, generating a distance score between the two embedding points corresponding to the two CIs of the CI pair based on the first similarity metric; and determining the contrast loss function based on the distance score. For example, the distance score will be smaller when the CI pair is a positive CI pair; and larger when the CI pair is a negative CI pair.

[0348] The reconstruction loss function 217 may be determined by the reconstruction head 215 of the underlying model based on a second similarity metric between the original CI data and the predicted CI data generated based on the mask. In some embodiments, determining the reconstruction loss function includes: generating at least partially masked CI data by applying a mask to the original CI data to remove at least a portion of the original CI data along the time dimension or subcarrier dimension; generating predicted CI data based on the masked CI data using the underlying model; generating an error function between the original CI data and the predicted CI data based on the second similarity metric; and determining the reconstruction loss function based on the error function.

[0349] In some embodiments, the total loss function may be determined based on an aggregation of the contrast loss function and the reconstruction loss function. The model parameters of the base model may be determined, learned, or acquired to minimize the total loss function. For example, the aggregation of the contrast loss function and the reconstruction loss function may include a weighted combination of the contrast loss function and the reconstruction loss function. The weights used in the weighted combination may also be included in the model parameters of the base model and may be adjusted during learning to minimize the total loss function through an iterative backpropagation process.

[0350] During process 220, the trained foundational model (implemented as encoder 222) may be executed to infer the decision results of the wireless sensing task during the inference phase. In this example shown in Figure 2, real-time CI data (e.g., real-time CSI) is collected and then processed by CNN 221 based on some data augmentation techniques, such as those described above, before being used as input for running the foundational model. Based on the real-time collected CI data, the trained foundational model (implemented as encoder 222) may generate a feature map. The feature map may be input to a downstream task classifier 223, which may be one of several task-specific models, to execute one of several wireless sensing tasks. The decision results generated by the downstream task classifier 223 may be based on the feature map and may indicate the classification of the corresponding wireless sensing task (e.g., whether motion was detected, the type of moving subject, whether occupied or empty, whether it fell over, whether a user is present, etc.).

[0351] Figure 3 shows an exemplary method 300 for combining determination of a multilink-based wireless sensing task according to several embodiments of the present disclosure. In some embodiments, there are multiple device pairs in at least one venue. Each device pair is formed by a transmitter and a receiver. For each of the multiple device pairs, a wireless signal is transmitted by the transmitter to the receiver via a wireless channel. The received wireless signal is different from the transmitted wireless signal due to the wireless channel and any sensing event in at least one venue. Based on the received wireless signal, a time-series of channel information (TSCI) of the wireless channel may be obtained. In some embodiments, real-time CI data may be collected based on all the TSCIs obtained for the device pairs. That is, real-time CI data may be collected from multiple wireless links during the inference phase. Each wireless link may correspond to a wireless channel between a transmitter and a receiver, or a wireless channel between a transmitting antenna and a receiving antenna. The real-time CI data may be used to perform at least one of the multiple wireless sensing tasks.

[0352] For each task of at least one task, the system may process each real-time CI data collected from one of multiple wireless links using a corresponding CNN301 to generate a processed CI, and then use a corresponding copy of the underlying model 302 to generate a corresponding feature map based on the processed CI. That is, real-time CI data collected from multiple links may be processed in parallel and used to run the underlying model 302 in parallel. In some embodiments, the same CNN 301 may process all real-time CI data in series, and the same copy of the underlying model 302 may be run in series based on the processed CI.

[0353] In some embodiments, the system may merge multiple feature maps generated by the base model 302 for multiple links, for example, along the subcarrier dimension, or according to the index of each of the multiple radio links, to generate a merged feature map. The system may input the merged feature map to a downstream task classifier 303, which may be a task-specific model corresponding to a radio sensing task, for generating a decision result for the radio sensing task. In some embodiments, the downstream task classifier 303 may be a small transformer model or RNN model that can take a variable length input.

[0354] Figure 4 shows another exemplary method 400 for combining determination of a multilink-based wireless sensing task according to some embodiments of the present disclosure. Similar to Figure 3, real-time CI data may be collected from multiple wireless links, where each wireless link may correspond to a wireless channel between a transmitter and a receiver, or a wireless channel between a transmitting antenna and a receiving antenna. The real-time CI data may be used to perform at least one of the multiple wireless sensing tasks.

[0355] For each task of at least one task, the system may process each real-time CI data collected from one of multiple wireless links using a corresponding CNN401 to generate a processed CI, and then use a corresponding copy of the underlying model 402 to generate a corresponding feature map based on the processed CI. That is, real-time CI data collected from multiple links may be processed in parallel and used to run the underlying model 402 in parallel. In some embodiments, the same CNN 401 may process all real-time CI data in series, and the same copy of the underlying model 402 may be run in series based on the processed CI.

[0356] In some embodiments, the system may input each of the feature maps generated by the base model 402 into a copy of a downstream task classifier 403, which may be a task-specific model corresponding to a wireless sensing task, to generate a candidate decision result for a task. The system may utilize a fusion model or algorithm 404 to fuse all the candidate decision results generated for that task in order to generate a final decision result for the task.

[0357] In some examples, for a motion detection task, if X candidate judgment results indicate motion detectio...

Claims

1. A method for wireless sensing, To obtain channel information (CI) data generated based on at least one radio channel, The process involves generating a training dataset based on the aforementioned CI data, wherein the training dataset comprises a plurality of CI pairs, original CI data, and a mask. Training a base model using the aforementioned training dataset, at least partially, The contrast loss function is determined based on a first similarity metric between the CI data of each CI pair in the training dataset. Determining a reconstruction loss function based on a second similarity metric between the original CI data and the predicted CI data generated based on the mask, The total loss function is determined based on the sum of the comparison loss function and the reconstruction loss function, and Determining the model parameters of the base model so as to minimize the total loss function, Training by, Training multiple task-specific models, The process involves executing a plurality of wireless sensing tasks based on the aforementioned base model and the plurality of task-specific models, wherein each of the plurality of task-specific models is used together with the base model to execute a corresponding one of the plurality of wireless sensing tasks. A method for providing this.

2. The method according to claim 1, wherein the acquisition of the CI data is To determine multiple device pairs in at least one venue, where each of the multiple device pairs is formed by a transmitter and a receiver, For each of the aforementioned plurality of device pairs, The transmitter transmits a wireless signal through a wireless channel. The receiver receives the radio signal, wherein the received radio signal is different from the transmitted radio signal due to the radio channel and any sensing event in the at least one venue. Based on the received radio signal, the time series (TSCI) of the channel information of the radio channel is obtained, The CI data is acquired based on all the TSCIs obtained for the aforementioned plurality of device pairs, A method that includes [a certain feature].

3. A method according to claim 2, wherein generating the training dataset is: The process involves processing the CI data in order to generate pre-processed CI data according to a standardized format readable by the aforementioned base model, To generate an extended CI, data augmentation is performed on the preprocessed CI in the preprocessed CI data, wherein the plurality of CI pairs are A positive CI pair formed by a pre-treated CI and the associated extended CI, A positive CI pair formed by two pre-treated CIs, A positive CI pair formed by two extended CIs, A negative CI pair formed by two CIs obtained from two different radio channels, A negative CI pair formed by two CIs obtained from two different venues, A negative CI pair formed by two CIs associated with two different sensing events, A method that includes the act of performing and the means of performing.

4. The method according to claim 3, wherein processing the CI data is: In order to generate the same number of subcarriers for all CIs in the CI data according to the standardized format, a subcarrier is selected for at least one CI in the CI data, The process involves resampling each CI in the CI data to a predetermined time rate according to the standardized format, A method that includes [a certain feature].

5. The method according to claim 4, wherein performing the data augmentation is Adding random noise to the aforementioned pre-processed CI, Randomizing the subcarriers selected within the block, Performing time scaling or time warping on the aforementioned pre-processed CI, To simulate at least one environmental parameter related to multipath changes or occlusion, To mitigate power fluctuations, the amplitude of the pre-processed CI is normalized, A method that includes [a certain feature].

6. The method according to claim 3, wherein determining the comparative loss function is Using the aforementioned base model, each CI in the training dataset is mapped to the corresponding embedding point in the embedding space, For each CI pair having two CIs in the training dataset, a distance score is generated between the two embedding points corresponding to the two CIs of the CI pair, based on the first similarity metric. The distance score is smaller when the CI pair is a positive CI pair. The aforementioned distance score is greater when the CI pair is a negative CI pair. To generate, The comparison loss function is determined based on the distance score, A method that includes [a certain feature].

7. The method according to claim 1, wherein determining the reconstruction loss function is The process involves generating masked CI data by applying the mask to the original CI data to remove at least a portion of the original CI data along the time dimension or subcarrier dimension, Using the aforementioned base model, predictive CI data is generated based on the masked CI data, Based on the second similarity metric, an error function is generated between the original CI data and the predicted CI data. Determining the reconstruction loss function based on the error function, A method that includes [a certain feature].

8. The method according to claim 1, The aggregation of the contrast loss function and the reconstruction loss function comprises a weighted combination of the contrast loss function and the reconstruction loss function, The weights used in the aforementioned weight combination are included in the model parameters of the base model and are adjusted during training to minimize the total loss function through an iterative backpropagation process. method.

9. The method according to claim 1, wherein training the plurality of task-specific models is Freezing all model parameters of the base model during training of the multiple task-specific models, During the training of the multiple task-specific models, all model parameters of the base model are fine-tuned based on at least one task-specific predictive loss, or During the training of the aforementioned multiple task-specific models, Freezing the model parameters of the upstream layer of the aforementioned base model, and The process involves fine-tuning the model parameters of the downstream layer of the base model, wherein each of the multiple task-specific models is a downstream model compared to the base model. A method comprising at least one of the following.

10. The method according to claim 1, wherein the plurality of wireless sensing tasks are performed Based on CI data collected in real time, a feature map is generated using the aforementioned base model, In order to perform each of the aforementioned multiple wireless sensing tasks, the feature maps are input into the respective task-specific models. A method that includes [a certain feature].

11. The method according to claim 1, wherein the plurality of wireless sensing tasks are performed To collect real-time CI data from multiple wireless links for at least one of the aforementioned multiple wireless sensing tasks, For each of the tasks of the at least one task mentioned above, Using the aforementioned base model, multiple feature maps are generated based on real-time CI data collected from one of the multiple wireless links. To generate a merged feature map by merging the feature maps at least partially along the subcarrier dimension or according to the index of each of the multiple radio links, and To generate a decision result for the task, input the fused feature map into a task-specific model corresponding to the task. A method that includes [a certain feature].

12. The method according to claim 1, wherein the plurality of wireless sensing tasks are performed To collect real-time CI data from multiple wireless links for at least one of the aforementioned multiple wireless sensing tasks, For each of the tasks of the at least one task mentioned above, Using the aforementioned base model, multiple feature maps are generated based on real-time CI data collected from one of the multiple wireless links. To generate candidate decision results for the task, each of the multiple feature maps is input into a task-specific model corresponding to the task, and To generate a final decision result for the task, all candidate decision results generated for the task based on the fusion model are fused. A method that includes [a certain feature].

13. The method according to claim 1, The aforementioned base model is trained based on self-supervised machine learning without labeled data. Each of the aforementioned task-specific models is trained on labeled data.

14. The method according to claim 1, The aforementioned training dataset is generated by a local device and sent from the local device to a cloud server. The aforementioned base model and the multiple task-specific models are trained by the cloud server. Performing the aforementioned multiple wireless sensing tasks is, To generate processed real-time CI data, real-time CI data is collected and processed by at least one local device. Based on the processed real-time CI data, at least one local device determines whether or not a trigger event occurs. In response to the determination that the trigger event has occurred, the most recent past portion of the processed real-time CI data within the most recent past period is sent from the at least one local device to the cloud server, and The cloud server executes a wireless sensing task corresponding to the trigger event, using the base model and a task-specific model corresponding to the wireless sensing task, based on the most recent historical portion of the processed real-time CI data. A method that includes [a certain feature].

15. A system for wireless sensing, At least one local device, To obtain channel information (CI) data generated based on at least one radio channel, The process involves generating a training dataset based on the aforementioned CI data, wherein the training dataset comprises a plurality of CI pairs, original CI data, and a mask. A local device configured to run, It is a cloud server, Training a base model using the aforementioned training dataset, The contrast loss function is determined based on a first similarity metric between the CI data of each CI pair in the training dataset. Determining a reconstruction loss function based on a second similarity metric between the original CI data and the predicted CI data generated based on the mask, The total loss function is determined based on the sum of the comparison loss function and the reconstruction loss function, and Determining the model parameters of the base model so as to minimize the total loss function, To train, at least partially. Training multiple task-specific models, A cloud server configured to perform the following actions: Equipped with, The at least one local device and the cloud server are further configured to perform a plurality of wireless sensing tasks based on the base model and the plurality of task-specific models. Each of the aforementioned task-specific models is used together with the base model to perform a corresponding one of the aforementioned wireless sensing tasks. system.

16. The system according to claim 15, wherein the at least one local device is at least partially The process involves processing the CI data in order to generate pre-processed CI data according to a standardized format readable by the aforementioned base model, To generate an extended CI, data extension is performed on the pre-processed CI in the pre-processed CI data, The system is configured to generate the aforementioned training dataset, The aforementioned plurality of CI pairs are A positive CI pair formed by the pre-treated CI and the associated extended CI, A positive CI pair formed by two pre-treated CIs, A positive CI pair formed by two extended CIs, A negative CI pair formed by two CIs obtained from two different radio channels, A negative CI pair formed by two CIs obtained from two different venues, and A negative CI pair formed by two CIs associated with two different sensing events, A system that includes this.

17. The system according to claim 16, Processing the aforementioned CI data is Selecting a subcarrier for at least one CI in the CI data in order to generate the same number of subcarriers for all CIs in the CI data according to the standardized format, and Resampling each CI in the CI data to a predetermined time rate according to the standardized format, Equipped with, Performing the aforementioned data expansion means Adding random noise to the aforementioned pre-processed CI, Randomizing the subcarriers selected within the block, Perform time scaling or time warping on the aforementioned pre-processed CI. To simulate at least one environmental parameter related to multipath changes or occlusion, and To mitigate power fluctuations, the amplitude of the pre-processed CI is normalized. A system comprising at least one of the following.

18. The system according to claim 16, wherein determining the comparative loss function is Using the aforementioned base model, each CI in the training dataset is mapped to the corresponding embedding point in the embedding space, For each CI pair having two CIs in the training dataset, a distance score is generated between the two embedding points corresponding to the two CIs of the CI pair, based on the first similarity metric. The distance score is smaller when the CI pair is a positive CI pair. The aforementioned distance score is greater when the CI pair is a negative CI pair. To generate, A system comprising determining the comparison loss function based on the distance score.

19. The system according to claim 15, wherein determining the reconstruction loss function is The process involves generating masked CI data by applying the mask to the original CI data to remove at least a portion of the original CI data along the time dimension or subcarrier dimension, Using the aforementioned base model, predictive CI data is generated based on the masked CI data, Based on the second similarity metric, an error function is generated between the original CI data and the predicted CI data. Determining the reconstruction loss function based on the error function, A system equipped with these features.

20. A device for wireless sensing, At least one processor, When executed, at least one of the processors will To obtain channel information (CI) data generated based on at least one radio channel, The process involves generating a training dataset based on the aforementioned CI data, wherein the training dataset comprises a plurality of CI pairs, original CI data, and a mask. At least partially, The contrast loss function is determined based on a first similarity metric between the CI data of each CI pair in the training dataset. Determining a reconstruction loss function based on a second similarity metric between the original CI data and the predicted CI data generated based on the mask, The total loss function is determined based on the sum of the comparison loss function and the reconstruction loss function, and Determine the model parameters of the base model so as to minimize the aforementioned total loss function. This involves training the base model using the training dataset, Training multiple task-specific models, The process involves executing a plurality of wireless sensing tasks based on the aforementioned base model and the plurality of task-specific models, wherein each of the plurality of task-specific models is used together with the base model to execute a corresponding one of the plurality of wireless sensing tasks. A device comprising at least one memory that stores instructions for performing an operation comprising the following:

Images