Target detection method and device based on channel state information, equipment and medium

Abstract

The application discloses a target detection method and device based on channel state information, equipment and medium. The original channel state information sequence is preprocessed to obtain a target channel state information sequence, so as to remove invalid interference noise of the channel state information, which is conducive to improving the accuracy of subsequent target detection. Then, the target model after training is used for mapping and representing the target channel state information sequence in a unified dimension, and the target detection result in a target scene is predicted based on the feature length dependency relationship of the unified dimension feature sequence. In this way, the target model has high adaptation ability and strong cross-environment generalization ability, is suitable for target detection of channel state information in any input form, has high applicability, reduces deployment cost, enhances the universality of the target detection scene based on the channel state information, meets the actual application requirements of large scale and generalization, and improves the target detection efficiency based on the channel state information.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a target detection method, apparatus, device and medium based on channel state information. Background Technology

[0002] With the development of wireless communication and IoT technologies, non-contact sensing based on channel state information (CSI) sequences from wireless networks has been widely applied in smart homes, smart healthcare, and indoor behavior analysis due to its advantages such as being wearable-free, unaffected by lighting conditions, and able to penetrate obstructions. Existing technologies mostly employ deep learning models to extract features and identify behaviors from the time-series data of CSI sequences, achieving high accuracy in fixed devices and single environments.

[0003] However, in related technologies, channel state information sequences suffer from significant device heterogeneity, environmental heterogeneity, and format heterogeneity. Signals generated by different acquisition devices, deployment environments, and datasets vary greatly in dimensionality, distribution, and structure. Traditional model structures are highly coupled with data input formats, resulting in poor model device adaptability and weak cross-environment generalization ability. The trained models are only suitable for target detection of channel state information with fixed input formats, exhibiting low applicability and consequently high deployment costs, making it difficult to meet the needs of large-scale and general-purpose practical applications. Summary of the Invention

[0004] This application provides a target detection method, apparatus, device, and medium based on channel state information, which is applicable to target detection with channel state information in any input form, reduces deployment costs, enhances the universality of target detection scenarios based on channel state information, and improves the efficiency of target detection based on channel state information.

[0005] In a first aspect, this application provides a target detection method based on channel state information, including: Obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scene; The original channel state information sequence is preprocessed to obtain the processed target channel state information sequence; The target channel state information sequence is input into the trained target model. The target model maps and represents the target channel state information sequence to obtain a unified dimension feature sequence. The target detection result is output based on the time-series features corresponding to the unified dimension feature sequence. The target model is obtained by training a preset model to minimize the loss of the joint reconstruction loss and the prediction loss. The preset model is based on the predicted detection result and the reconstructed channel state information sequence output by the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

[0006] Secondly, this application provides a target detection device based on channel state information, comprising: The acquisition unit is used to acquire the original channel state information sequence corresponding to the wireless transmitting device in the target scene; The preprocessing unit is used to preprocess the original channel state information sequence to obtain the processed target channel state information sequence. The input unit is used to input the target channel state information sequence into the trained target model, map the target channel state information sequence through the target model to obtain a unified dimension feature sequence, and output the target detection result according to the time-series features corresponding to the unified dimension feature sequence. The target model is obtained by training a preset model to minimize the loss of the joint reconstruction loss and the prediction loss. The preset model is based on the predicted detection result and the reconstructed channel state information sequence output by the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

[0007] In some implementations, the target model includes an embedded unified representation network, the embedded unified representation network comprising a first convolutional layer and a second convolutional layer, and the input unit is further configured to: Through the first convolutional layer in the target model, local time variation features are extracted for the target channel state information sequence in the time dimension, and frequency domain variation features are extracted for the channel state information corresponding to adjacent subcarriers in the target channel state information sequence in the frequency domain dimension. The local temporal variation features and the frequency domain variation features are fused and mapped through the second convolutional layer in the target model to generate a unified dimension feature sequence.

[0008] In some embodiments, the input unit is further configured to: By using the first convolutional layer in the target model, a one-dimensional convolution operation is performed on the target channel state information sequence along the time dimension to obtain the local time variation features of the target channel state information sequence in continuous time segments. The first convolutional layer in the target model extracts features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence along the frequency domain dimension to obtain the corresponding channel state features. The similarity change features among the channel state features are then aggregated to obtain the frequency domain change features of the target channel state information sequence.

[0009] In some implementations, the target model includes a shared backbone network, and the input unit is further configured to: By using the shared backbone network, the temporal relationship features corresponding to the unified dimension feature sequence are obtained according to the attention mechanism to obtain general temporal features; The target detection result is output based on the general temporal features through the prediction output head in the target model.

[0010] In some embodiments, the input unit is further configured to: Through the shared backbone network, the temporal dependency modeling of the unified dimension feature sequence is performed according to the multi-head self-attention mechanism to obtain the correlation strength and dependency relationship features between features at different time positions in the unified dimension feature sequence. Based on the correlation strength and dependency features, the corresponding attention weights are calculated, and the features at different time positions are weighted and fused according to the attention weights to obtain fused features that contain global temporal change patterns. Based on the fusion features, long-distance temporal relationship features are extracted to obtain general temporal features.

[0011] In some embodiments, the preprocessing unit is further configured to: The original channel state information sequence is subjected to abnormal data removal to obtain the removed initial channel state information sequence; The amplitude information in the initial channel state information sequence is normalized, and the phase information in the initial channel state information sequence is corrected to obtain the candidate channel state information sequence. The signal noise in the candidate channel state information sequence is suppressed to obtain the processed target channel state information sequence.

[0012] In some embodiments, the target detection device based on channel state information further includes a model training unit, used for: Obtain a local preset channel state information sequence and the corresponding sample detection results, and preprocess the preset channel state information sequence to obtain a sample channel state information sequence; The sample channel state information sequence is input into a preset model to obtain the reconstructed channel state information sequence output by the reconstruction output header of the preset model; The reconstruction loss is constructed based on the difference between the reconstructed channel state information sequence and the sample channel state information sequence; Based on the reconstruction loss, the model parameters of the preset model are adjusted starting from the reconstruction output head using backpropagation, and iterative training is performed to obtain the trained candidate model. The sample channel state information sequence is input into the candidate model to obtain the prediction detection result output by the prediction output head; A prediction loss is constructed based on the difference between the predicted detection result and the sample detection result, and the candidate model is fine-tuned based on the prediction loss to obtain the trained target model.

[0013] In some implementations, the preset model includes at least an embedded unified representation network, a shared backbone network, and a reconstructed output head; the model training unit is further configured to: Based on the reconstruction loss, the gradients of the reconstruction output head, the shared backbone network, and the embedded unified representation network are calculated sequentially according to the backpropagation method; The network parameters of the reconstructed output head, the shared backbone network, and the embedded unified representation network are adjusted according to the corresponding gradients, and iterative training is performed until convergence to obtain the trained candidate model.

[0014] Furthermore, this application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described target detection method based on channel state information.

[0015] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the aforementioned target detection method based on channel state information.

[0016] This application embodiment can obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scene; preprocess the original channel state information sequence to obtain the processed target channel state information sequence; input the target channel state information sequence into the trained target model, and map the target channel state information sequence to obtain a unified dimension feature sequence through the target model, and output the target detection result according to the time series features corresponding to the unified dimension feature sequence; wherein, the target model is obtained by minimizing the loss of a preset model through joint reconstruction loss and prediction loss, the preset model outputs the predicted detection result and the reconstructed channel state information sequence based on the sample channel state information sequence, the reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result. In this way, the channel state information sequence of any input form can be mapped to a unified dimension feature sequence through the trained target model, and target detection can be performed, which enhances the universality of target detection scenarios based on channel state information.

[0017] As can be seen from the above, compared with the traditional model structure and data input form in related technologies, which are highly coupled, resulting in poor model device adaptability and weak cross-environment generalization ability, and the trained model is only applicable to target detection of channel state information with fixed input form, the applicability is low, and thus the deployment cost is high, the stability is insufficient, and it is difficult to meet the needs of large-scale and generalized practical applications, this application obtains the target channel state information sequence by preprocessing the original channel state information sequence to remove invalid interference noise of the channel state information, which is conducive to improving the accuracy of subsequent target detection. Then, the target channel state information sequence is mapped to a unified dimension through the trained target model, and the target detection result in the target scene is predicted based on the feature length dependency relationship of the unified dimension feature sequence. In this way, the target model has high adaptability and strong cross-environment generalization ability, and is applicable to target detection of channel state information with arbitrary input form. It has high applicability, reduces deployment cost, enhances the universality of target detection scenarios based on channel state information, can meet the needs of large-scale and generalized practical applications, and improves the efficiency of target detection based on channel state information. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a schematic diagram of a target detection system based on channel state information provided in an embodiment of this application. Figure 2 This is a flowchart illustrating the steps of the target detection method based on channel state information provided in the embodiments of this application. Figure 3 An example diagram of the architecture of the target model provided in the embodiments of this application; Figure 4 Example diagrams comparing the performance of the target model with other models provided in this application embodiment; Figure 5 This is a schematic diagram of the target detection device based on channel state information provided in an embodiment of this application; Figure 6 This is a schematic diagram of the terminal structure provided in the embodiments of this application; Figure 7 This is a partial structural block diagram of the server provided in an embodiment of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of this application.

[0021] It is understood that in the specific implementation of this application, data such as the original channel state information sequence, the target channel state information sequence, the unified dimension feature sequence, and the target detection result are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the target is required, and the collection, use and processing of related data must comply with relevant laws, regulations and standards.

[0022] Furthermore, when this application embodiment needs to obtain relevant data, it will obtain separate permission or separate consent for the original channel state information sequence, target channel state information sequence, unified dimension feature sequence, target detection results and other related data through pop-up windows or jumps to a confirmation page. After clearly obtaining separate permission or separate consent for the original channel state information sequence, target channel state information sequence, unified dimension feature sequence, target detection results and other related data, it will then obtain the necessary data for this application embodiment to operate normally.

[0023] It should be noted that while some processes described in the specification, claims, and accompanying drawings include multiple steps appearing in a specific order, it should be clearly understood that these steps may not be performed in the order they appear herein, or may be performed in parallel. The step numbers are merely used to distinguish different steps and do not represent any execution order. Furthermore, descriptions such as "first," "second," or "objective" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of this application.

[0025] This application provides a target detection method, apparatus, device, and medium based on channel state information. Specifically, the target detection method based on channel state information in this application can be implemented in a computer device, which can be a server or a terminal device. The server can be a standalone physical server, a server cluster or service cluster composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The user terminal device can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart home appliance, vehicle terminal, smart voice interaction device, aircraft, drone, etc., but is not limited to these.

[0026] For ease of understanding, this application will describe the implementation process of the target detection method based on channel state information through several embodiments, as follows: This application provides a target detection method based on channel state information (CSO). The method acquires the original CSO sequence corresponding to a wireless transmitting device within a target scenario. It preprocesses the original CSO sequence to obtain a processed target CSO sequence. This target CSO sequence is then input into a trained target model. The target model maps the target CSO sequence to a unified-dimensional feature sequence, and outputs the target detection result based on the temporal features corresponding to the unified-dimensional feature sequence. The target model is trained by minimizing the loss of a preset model using a joint reconstruction loss and a prediction loss. The preset model outputs the predicted detection result and the reconstructed CSO sequence based on the sample CSO sequence. The reconstruction loss is determined by the difference between the reconstructed CSO sequence and the sample CSO sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result. Thus, the trained target model can map any input form of CSO sequence to a unified-dimensional feature sequence for target detection, enhancing the universality of target detection scenarios based on CSO. Please refer to the following specific embodiments for details.

[0027] For example, see Figure 1 This is a schematic diagram of an information push system provided in an embodiment of this application. The system includes a terminal 110 and a server 120.

[0028] Taking the example of execution by a single terminal, each terminal 110 can have a target application installed, and the corresponding application service can be run through the target application. This target application can be called a client. Taking the terminal 110 as a signal acquisition device (Wi-Fi collector) of a wireless local area network, the signal acquisition device can collect the original channel state information sequence emitted by the wireless transmitter (such as a router) in the target scene (such as a space or room) and send the original channel state information sequence to the server 120.

[0029] The server 120 can be a single service node, a distributed system composed of multiple service nodes, or a service node within a distributed system. The server 120 executes the steps of distributed training of a target detection method based on channel state information. Specifically, it acquires the original channel state information sequence corresponding to the wireless transmitting device in the target scene; preprocesses the original channel state information sequence to obtain a processed target channel state information sequence; inputs the target channel state information sequence into the trained target model, maps the target channel state information sequence to obtain a unified dimension feature sequence through the target model, and outputs the target detection result based on the temporal features corresponding to the unified dimension feature sequence. The target model is obtained by minimizing the loss of a preset model using joint reconstruction loss and prediction loss. The preset model outputs the predicted detection result and the reconstructed channel state information sequence based on the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

[0030] Therefore, this application obtains the target channel state information sequence by preprocessing the original channel state information sequence to remove invalid interference noise from the channel state information, which is beneficial to improving the accuracy of subsequent target detection. Furthermore, the target channel state information sequence is mapped to a unified dimension through the trained target model, and the target detection results in the target scene are predicted based on the feature length dependency relationship of the unified dimension feature sequence. Thus, the target model has high adaptability and strong cross-environment generalization ability, and is suitable for target detection of channel state information in any input form. It has high applicability, reduces deployment costs, enhances the universality of target detection scenarios based on channel state information, meets the needs of large-scale and generalized practical applications, and improves the efficiency of target detection based on channel state information.

[0031] For ease of understanding, the steps of the target detection method based on channel state information will be described in detail below. It should be noted that the order of the following embodiments is not intended to limit the preferred order of the embodiments.

[0032] See Figure 2 , Figure 2 This is a flowchart illustrating the steps of a target detection method based on channel state information provided in an embodiment of this application. In this embodiment, the target detection method based on channel state information can be executed by a computer device. Taking a computer device as a server as an example, the specific process of executing the target detection method based on channel state information is as follows: 101. Obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scene.

[0033] With the development of wireless communication and IoT technologies, non-contact sensing based on channel state information (CSI) sequences from wireless networks has been widely applied in smart homes, smart healthcare, and indoor behavior analysis due to its advantages such as being wearable-free, unaffected by lighting conditions, and able to penetrate obstructions. Existing technologies mostly employ deep learning models to extract features and identify behaviors from the time-series data of CSI sequences, achieving high accuracy in fixed devices and single environments.

[0034] However, in related technologies, channel state information sequences suffer from significant device heterogeneity, environmental heterogeneity, and format heterogeneity. Signals generated by different acquisition devices, deployment environments, and datasets vary greatly in dimensionality, distribution, and structure. Traditional model structures are highly coupled with data input formats, resulting in poor model device adaptability and weak cross-environment generalization ability. The trained models are only suitable for target detection of channel state information with fixed input formats, exhibiting low applicability and consequently high deployment costs, making it difficult to meet the needs of large-scale and general-purpose practical applications.

[0035] To address the above issues, this application embodiment can obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scenario; preprocess the original channel state information sequence to obtain the processed target channel state information sequence; input the target channel state information sequence into the trained target model, and use the target model to map the target channel state information sequence to obtain a unified dimension feature sequence; and output the target detection result based on the temporal features corresponding to the unified dimension feature sequence. The target model is obtained by minimizing the loss of a preset model using joint reconstruction loss and prediction loss. The preset model outputs the predicted detection result and the reconstructed channel state information sequence based on the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result. Thus, the trained target model can map channel state information sequences of any input form to a unified dimension feature sequence, and capture features representing global temporal relationships in the target channel state information sequence for target detection, thereby enhancing the universality of target detection scenarios based on channel state information.

[0036] In this embodiment, to perform target detection in a target scene, it is necessary to obtain the original channel state information sequence transmitted by the wireless transmitting device within the target scene. This sequence is then used for subsequent target detection. Specifically, after a router (the transmitter of a Wi-Fi / wireless network) emits a Wi-Fi signal, a signal acquisition device (Wi-Fi collector) can collect data from the Wi-Fi signal within the target scene to obtain the original channel state information sequence. The signal acquisition device (Wi-Fi collector) then sends the collected original channel state information sequence to the server 120. In this way, the server 120 obtains the original channel state information sequence corresponding to the wireless transmitting device within the target scene. This sequence can then be used as the basis for target detection, thus achieving reliable target detection within the target scene.

[0037] The target scene can be an indoor or outdoor scene. For example, an indoor scene can be any indoor place such as a living room, bedroom, office, basement, parking lot, shopping mall, gym, etc. An outdoor scene can be any outdoor place such as a square, stadium, park, street, etc.

[0038] The wireless transmitting device can be a router, such as a Wi-Fi / wireless network transmitter.

[0039] The original channel state information sequence can be a channel response data sequence containing multiple antennas and subcarrier dimensions, obtained by the wireless signal collector in the target scene after performing air interface detection and real-time acquisition of the wireless radio frequency signal broadcast by the wireless transmitting device. The CSI data includes complex channel response information corresponding to different subcarriers, different antennas, and different timestamps. That is, it includes the amplitude response and phase response information of all subcarriers corresponding to each antenna. It is the original perception data formed after the signal propagates in the target scene and is affected by environmental structure, obstructions, human activities, and multipath reflection. It has not undergone any preprocessing operations such as anomaly removal, normalization, and noise suppression. It can truly and completely reflect the dynamic change characteristics of the wireless channel and provide basic original data support for subsequent channel state sequence preprocessing, feature mapping, and target detection.

[0040] Specifically, when acquiring the original channel state information sequence corresponding to the wireless transmitting device in the target scene, a WiFi signal collector deployed in the target scene is used to perform air interface detection on the wireless radio frequency signal broadcast by the wireless transmitting device and propagated through multipath in space, and to collect the original channel state information sequence corresponding to multiple antennas and multiple subcarriers. This original channel state information sequence contains the amplitude response and phase response information of different antennas on each subcarrier, which can fully reflect the attenuation, delay and phase shift characteristics of the signal during propagation in the target scene caused by environmental structure, obstructions and human activities. This provides real and comprehensive original perception data for subsequent preprocessing, feature mapping and target detection, and ensures that the subsequent model can achieve stable and reliable detection output based on complete channel change information.

[0041] Through the above methods, the server can obtain a sequence of original channel state information that truly and completely reflects the dynamic changes of the wireless channel. This provides real and comprehensive raw perception data for subsequent preprocessing, feature mapping, and target detection, ensuring that the subsequent model can achieve stable and reliable detection output based on complete channel change information. This enables reliable target detection within the target scene.

[0042] 102. Preprocess the original channel state information sequence to obtain the processed target channel state information sequence.

[0043] In this embodiment, after obtaining the original channel state information sequence in the target scene, the original channel state information sequence includes complex channel response information corresponding to different subcarriers, different antennas, and different timestamps. Since different devices have differences in center frequency, bandwidth, number of antennas, and sampling rate, the original channel state information sequence can be preprocessed. Specifically, it can be a preprocessing process such as removing abnormal data, amplitude normalization, phase correction, and noise suppression, thereby reducing the impact of device differences and acquisition errors on subsequent modeling. The preprocessed target channel state information sequence (CS sequence) maintains the time structure and frequency domain structure of the original signal, providing standardized input for subsequent unified modeling.

[0044] The target channel state information sequence can be a standardized, highly reliable channel state data sequence obtained by sequentially performing preprocessing operations such as abnormal data removal, amplitude normalization, phase correction, and noise suppression on the original channel state information sequence. It consists of valid and abnormal channel response data from the original channel state information sequence, retaining the core time-frequency domain features (including the frequency domain correlation of adjacent subcarriers and the time-domain variation features of continuous time segments) that reflect environmental changes and human activities within the target scene in the original sequence. It eliminates abnormal data, noise interference, and amplitude and phase distortion in the original sequence, unifies the distribution and format of channel state data under different acquisition devices and environments, and can adapt to the input requirements of the target model. It provides a high-quality data foundation for subsequent feature mapping embedded in a unified representation network, time-dependent modeling of shared backbone networks, and accurate output of target detection results.

[0045] In some implementations, the original channel state information sequence is subjected to outlier removal, amplitude normalization, phase correction, and noise suppression to obtain a processed target channel state information sequence. For example, step 102 may include: (102.1) Remove abnormal data from the original channel state information sequence to obtain the removed initial channel state information sequence; (102.2) Normalize the amplitude information in the initial channel state information sequence and correct the phase information in the initial channel state information sequence to obtain the candidate channel state information sequence. (102.3) The signal noise in the candidate channel state information sequence is suppressed to obtain the processed target channel state information sequence.

[0046] Specifically, firstly, the acquired raw channel state information sequence undergoes anomaly removal processing. Using a preset anomaly detection threshold, abnormal jump data, invalid or missing data, and data exceeding reasonable value ranges caused by equipment malfunctions, sudden signal interference, or acquisition errors are filtered and removed from the sequence, resulting in a clean initial channel state information sequence. This effectively removes invalid interference data from the original sequence, preventing anomalies from interfering with subsequent feature extraction and model training, ensuring the validity and authenticity of the data for subsequent processing, and laying a reliable data foundation for the entire preprocessing workflow.

[0047] Then, after obtaining the initial channel state information sequence, the amplitude information in the sequence is normalized. A preset normalization algorithm is used to map the amplitude response values, which vary greatly from different acquisition devices and environments, to a unified value range, eliminating the amplitude value dispersion caused by differences in device hardware and signal propagation distance. Simultaneously, the phase information in the initial channel state information sequence is corrected to compensate for phase distortion caused by multipath reflection and device phase offset during signal propagation, and to correct phase entanglement, thus obtaining a candidate channel state information sequence. In this way, amplitude normalization unifies the numerical distribution of channel state data from different sources, and phase correction restores the true phase characteristics of the channel response, effectively mitigating the data format differences caused by heterogeneous devices and environments. This provides the processed candidate sequence with a unified numerical benchmark, facilitating subsequent embedding into a unified representation network for feature mapping and improving the model's adaptability to heterogeneous data.

[0048] Finally, the obtained candidate channel state information sequence is subjected to signal and noise suppression processing. A lightweight filtering algorithm is used to suppress high-frequency noise and random interference signals in the sequence, while retaining the core time-frequency domain features (including the frequency domain correlation of adjacent subcarriers and the time-domain variation features of continuous time segments) of effective signals reflecting environmental changes and human activities within the target scene. This yields the processed target channel state information sequence. This effectively filters out useless noise interference in the candidate sequence, further improving the purity and reliability of the data, highlighting the characteristic information of effective signals, and ensuring that the subsequent embedding of the unified representation network can accurately extract local time-varying features and frequency-varying features. The shared backbone network can accurately capture temporal dependencies, ultimately improving the accuracy and stability of target detection results. Simultaneously, it provides high-quality standardized input data for the model to generalize across signal acquisition devices and environments.

[0049] The above methods can be used to preprocess the original channel state information sequence, specifically by removing outlier data, normalizing amplitude, correcting phase, and suppressing noise. This reduces the impact of equipment differences and acquisition errors on subsequent modeling. The preprocessed target channel state information sequence maintains the time and frequency domain structure of the original signal, providing standardized input for subsequent unified modeling.

[0050] 103. Input the target channel state information sequence into the trained target model, map the target channel state information sequence through the target model to obtain a unified dimension feature sequence, and output the target detection result according to the time series features corresponding to the unified dimension feature sequence.

[0051] In this embodiment, after obtaining the preprocessed target channel state information sequence, the server can input the target channel state information sequence into the trained target model. The target model then maps the target channel state information sequence to obtain a unified-dimensional feature sequence, achieving a unified-dimensional mapping of the target channel state information sequence. Specifically, a convolutional embedding network is used to extract features from the target channel state information sequence, capturing local signal change features in the time dimension and aggregating the frequency domain correlation between adjacent subcarriers to generate a unified-dimensional feature sequence representation. Thus, this mapping process maintains the original physical structure of the CSI signal while achieving feature alignment between different data sources, mapping CSI signals with different structures to a unified feature representation space. Furthermore, the target detection result is output based on the temporal features corresponding to the unified-dimensional feature sequence. Specifically, a self-attention mechanism is used to model the temporal dependencies in the target channel state information sequence, enabling the system to capture the dynamic changes in the channel caused by human movement or environmental changes, thereby performing target detection. In this way, channel state information sequences of any input form can be mapped to a unified dimension feature sequence, and target detection can be performed by capturing the global temporal relationship features of the channel dynamic change patterns caused by human movement or environmental changes. This enhances the universality of target detection scenarios based on channel state information.

[0052] The target detection results can be the results of human action recognition, gesture recognition, or identity recognition within the target scene. For example, human action recognition results can be "falling down", "standing", "sitting", "lying down", etc.; gesture recognition results can be "opening palm", "scissor hand", "making a fist", "extending four fingers", "extending three fingers", "extending the middle finger", etc.; and identity recognition results can be "object 1", "object 2", "object 3", "object 4", etc. from a preset set of identity labels. The specific results depend on the actual scene and are not limited here.

[0053] In some implementations, the target channel state information sequence can be mapped to a unified dimension to produce a unified dimension feature sequence representation. For example, the target model includes an embedded unified representation network, which contains a first convolutional layer and a second convolutional layer. Step 103, "mapping the target channel state information sequence through the target model to obtain a unified dimension feature sequence," may include: (103.a.1) Through the first convolutional layer in the target model, local time variation features are extracted for the target channel state information sequence in the time dimension, and frequency domain variation features are extracted for the channel state information corresponding to adjacent subcarriers in the target channel state information sequence in the frequency domain dimension. (103.a.2) The local time-varying features and frequency domain-varying features are fused and mapped through the second convolutional layer in the target model to generate a unified dimension feature sequence.

[0054] Among them, the local temporal variation features can be extracted by performing a one-dimensional convolution operation on the preprocessed target channel state information sequence along the time dimension using the first convolutional layer of the unified representation network embedded in the target model. Its core reflects the dynamic fluctuation pattern of the target channel state information sequence in continuous time segments. It can accurately capture the details of channel state temporal changes introduced during target activities, environmental changes, and signal propagation. It is an important temporal basis for characterizing target behavior and state changes, providing effective temporal feature support for subsequent feature fusion and temporal dependency modeling. At the same time, it can retain the general rules of channel temporal changes in different scenarios, helping to improve the model's cross-environment generalization ability.

[0055] Among them, the frequency domain variation features can be obtained by the first convolutional layer of the unified representation network embedded in the target model, extracting features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence along the frequency domain dimension, and aggregating the similar variation features between adjacent subcarriers. Its core reflects the frequency domain correlation between adjacent subcarriers and the inherent physical characteristics of the channel state in the frequency domain. It can effectively avoid the feature heterogeneity problem caused by the difference in the number of subcarriers of different acquisition devices, retain the frequency domain feature rules that are universal across signal acquisition devices and environments, provide stable frequency domain feature support for subsequent feature fusion and unified dimension mapping, and further enhance the model's adaptability to heterogeneous channel data.

[0056] Among them, the unified dimension feature sequence can be a fixed dimension feature sequence generated by the second convolutional layer of the unified representation network embedded in the target model, which deeply fuses local time-varying features and frequency-domain variation features, and then performs dimension compression and mapping. It integrates the complete time-frequency domain features of the target channel state information sequence, eliminates the differences in dimension, distribution and format of channel state data in different acquisition devices and deployment environments, realizes the standardized transformation of heterogeneous channel state data, and can accurately adapt to the input requirements of the shared backbone network in the target model. It provides a standardized and high-quality feature foundation for the subsequent time-dependent modeling and prediction output head output of target detection results of the shared backbone network. It is the core feature carrier for realizing the model's generalized detection across signal acquisition devices and environments.

[0057] It should be noted that the process of mapping the target channel state information sequence to obtain a unified-dimensional feature sequence through the target model is specifically executed by the unified representation network embedded in the target model. This unified representation network contains a first convolutional layer and a second convolutional layer that work together. Both the first and second convolutional layers are one-dimensional convolutional layers. The convolutional kernel of the first convolutional layer performs convolution operations on the target channel state sequence along the time dimension, while the convolutional kernel of the second convolutional layer is used to compress and fuse the features to output a fixed-dimensional feature vector. In this way, the two convolutional layers work together to complete feature extraction, feature fusion, and dimension mapping in sequence, ultimately realizing the transformation of target channel state information sequences from different sources and in different formats into a unified-dimensional feature sequence, providing standardized feature input for subsequent temporal dependency modeling and target detection.

[0058] Specifically, firstly, the preprocessed target channel state information sequence is input into the first convolutional layer of the embedded unified representation network. This first convolutional layer performs feature extraction operations simultaneously along the time and frequency domain dimensions: In the time dimension, the first convolutional layer captures local features of the target channel state information sequence through one-dimensional convolution operations, extracting local time change features of the sequence within continuous time segments. This operation can accurately capture the dynamic fluctuation pattern of the channel state over time, restore the temporal feature differences brought about by target activities and environmental changes, and provide effective temporal feature support for subsequent temporal relationship modeling; In the frequency domain dimension, the first convolutional layer extracts features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence, and simultaneously aggregates similar change features between adjacent subcarriers to obtain frequency domain change features. This operation can fully explore the inherent physical characteristics of the channel state in the frequency domain, retain the core feature of frequency domain correlation between adjacent subcarriers that is universal across signal acquisition devices and environments, effectively weaken the feature heterogeneity problem caused by the difference in the number of subcarriers in different devices, and lay the foundation for unified dimensional mapping of features.

[0059] Furthermore, after the first convolutional layer completes the dual-dimensional feature extraction, the extracted local temporal variation features and frequency domain variation features are simultaneously input into the second convolutional layer embedded in the unified representation network. The second convolutional layer performs deep fusion and dimensional mapping processing on the two types of features: on the one hand, the convolution operation realizes the organic fusion of local temporal variation features and frequency domain variation features, breaking the separation between time domain and frequency domain features, integrating the complete time-frequency domain information of the channel state, so that the fused features can comprehensively reflect the comprehensive impact of target activities and environmental changes on the channel; on the other hand, through the dimensional compression and mapping effect of the second convolutional layer, the fused features of different dimensions and different distributions are uniformly mapped to a preset fixed dimension, generating a unified dimensional feature sequence. The technical effect of this step is that it effectively solves the technical problem of inconsistent dimensions and formats of target channel state information sequences under different acquisition devices and deployment environments, realizes the standardized processing of heterogeneous channel state data, enables the subsequent shared backbone network to carry out time-dependent modeling based on feature sequences of unified specifications, significantly improves the cross-signal acquisition device and cross-environment adaptability of the target model, and simplifies the subsequent processing flow of the model, providing standardized and high-quality feature support for the accurate output of target detection results.

[0060] In some implementations, the first convolutional layer can capture the local temporal variation features of the target channel state information sequence within a continuous time segment, as well as the frequency domain variation features obtained by aggregating similar variation features between adjacent subcarriers. For example, step (103.a.1) may include: performing a one-dimensional convolution operation on the target channel state information sequence along the time dimension through the first convolutional layer in the target model to obtain the local temporal variation features of the target channel state information sequence within a continuous time segment; extracting features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence along the frequency domain dimension through the first convolutional layer in the target model to obtain the corresponding channel state features; and aggregating the similar variation features between the channel state features to obtain the frequency domain variation features of the target channel state information sequence.

[0061] Specifically, after inputting the target channel state information sequence into the trained target model, the first convolutional layer of the target model performs a one-dimensional convolution operation along the time dimension on the preprocessed target channel state information sequence. This captures the signal changes of the target channel state information sequence within continuous time segments, extracting local time variation features. The technical advantage of this step is that it accurately captures the dynamic fluctuations of the target channel state sequence over time, clearly presenting the changing patterns of the channel state at different time points, providing reliable temporal feature support for subsequent feature fusion. Simultaneously, it effectively filters out invalid fluctuations in the time dimension, ensuring the targeted nature of the extracted local time variation features, avoiding interference from irrelevant time signals in subsequent feature mapping, and improving the overall accuracy of feature extraction.

[0062] Subsequently, through this first convolutional layer, features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence are extracted along the frequency domain dimension. Simultaneously, similarity analysis and aggregation processing are performed on the extracted channel state features corresponding to each adjacent subcarrier, ultimately obtaining the frequency domain variation features of the target channel state information sequence. The technical effect of this operation is that it can fully explore the frequency domain correlation between adjacent subcarriers, capture the synchronous change patterns of the channel states of adjacent subcarriers, strengthen the integrity and correlation of frequency domain features, and effectively weaken the differences in frequency domain features caused by different devices and environments. This lays the foundation for subsequent unified dimension mapping, ensures that frequency domain features and time domain features can be effectively integrated, further improves the adaptability of the target model to heterogeneous channel data, and guarantees the accuracy and stability of the subsequent unified dimension feature sequence generation. Thus, the first convolutional layer achieves comprehensive capture of the core features of the target channel state information sequence through dual feature extraction in both the time and frequency domains. It not only preserves the dynamic characteristics of the channel state changing over time, but also mines the frequency domain correlation of adjacent subcarriers, providing a high-quality feature foundation for subsequent feature fusion and unified dimension mapping. This effectively solves the problems of separation between the time and frequency domains and incomplete features in traditional feature extraction, and improves the accuracy of subsequent target model mapping and detection results.

[0063] In some implementations, the target model outputs target detection results based on temporal relationship features corresponding to a unified-dimensional feature sequence. For example, if the target model includes a shared backbone network, step 103, "outputting target detection results based on temporal features corresponding to a unified-dimensional feature sequence," may include: (103.b.1) By sharing the backbone network, the temporal relationship features corresponding to the unified dimension feature sequence are obtained according to the attention mechanism to obtain the general temporal features; (103.b.2) Output target detection results based on general temporal features through the prediction output head in the target model.

[0064] The shared backbone network is the core functional module in the target model, connecting and embedding the unified representation network and the prediction output head. It is primarily used for temporal feature modeling of the preprocessed target channel state information, serving as the core support module for target detection. Its core structure includes an attention mechanism module, a feature fusion module, and a parameter adjustment unit. The attention mechanism module accurately captures the dynamic changes in the target channel state, enabling real-time monitoring of channel state changes within continuous time segments. The feature fusion module integrates feature information from different dimensions, eliminating feature biases caused by heterogeneous devices and environmental differences. The parameter adjustment unit flexibly adjusts the feature extraction strategy based on the input channel state data, ensuring the accuracy and universality of feature extraction.

[0065] The prediction output head is the functional module in the target model responsible for outputting detection results. It connects the shared backbone network with the target detection result output and is mainly used to decode and classify the general temporal features output by the shared backbone network. Its core function is to output clear target detection results based on general temporal features, through feature mapping and category judgment, including key information such as the target's state and behavior. At the same time, this module can adapt to the detection needs of different scenarios. By combining the core information of general temporal features, it avoids detection bias caused by heterogeneous devices and environments, ensuring the accuracy and stability of the detection results. It is a key link connecting feature extraction and target detection output, directly determining the accuracy and applicability of the model's detection results.

[0066] Among them, the general temporal feature is a universal feature information extracted and generated by the shared backbone network after performing temporal modeling on a unified dimension feature sequence. It is the core carrier connecting channel state and target detection results. It integrates the time-domain dynamic change features and frequency-domain inherent characteristics in the target channel state information, eliminates feature interference caused by device and environmental differences, retains key features related to target activity, and has adaptability across signal acquisition devices and scenarios. It does not depend on specific devices or environments and can accurately reflect the correlation between target activity and channel state changes. It provides reliable feature input for the prediction output head and is the core feature foundation for realizing cross-scenario generalization of the model and improving detection accuracy.

[0067] It should be noted that the process of outputting target detection results based on the temporal features corresponding to the unified dimension feature sequence is executed collaboratively by the shared backbone network and the prediction output head in the target model. The two work together in a step-by-step manner to complete the temporal feature modeling and detection result output in sequence, making full use of the core time-frequency domain information of the unified dimension feature sequence to ensure the accuracy and generalization of the target detection results.

[0068] Specifically, firstly, the unified-dimensional feature sequence generated by the embedded unified representation network is input into the shared backbone network of the target model. This shared backbone network uses an attention mechanism to model the temporal dependencies of the unified-dimensional feature sequence. By calculating the correlation strength and dependency between features at different time positions in the unified-dimensional feature sequence, the attention weight of each feature at each time position is determined. Then, based on this attention weight, the features at different time positions are weighted and fused, focusing on strengthening key temporal features that can reflect target activities and environmental changes, while weakening the interference of invalid and redundant features. Finally, a general temporal feature containing global temporal change patterns and adaptable to cross-signal acquisition devices and cross-environments is obtained. The technical effect of this step is that the attention mechanism can accurately capture long-distance temporal dependencies in the unified-dimensional feature sequence, breaking the limitation of traditional temporal feature extraction that is difficult to take into account both local and global features. It fully preserves the core patterns of channel state temporal changes. At the same time, relying on the standardized characteristics of the unified-dimensional feature sequence, the shared backbone network can stably process heterogeneous channel data from different sources, significantly improving the universality and reliability of the general temporal feature, and providing high-quality temporal feature support for the accurate output of subsequent target detection results.

[0069] Furthermore, after obtaining the general temporal features, they are input into the prediction output head of the target model. The prediction output head further decodes and classifies the general temporal features, and combines the correspondence between the general temporal features learned during training and the target detection results to perform category judgment and probability calculation on the general temporal features. Finally, it outputs a target detection result adapted to the target scene, which can accurately represent key information such as the target state and behavior within the target scene. The technical effect of this step is that the prediction output head can accurately receive the general temporal features output by the shared backbone network. Through targeted decoding and classification mapping, it transforms the abstract temporal features into intuitive and applicable target detection results, effectively avoiding the problem of the temporal features being out of sync with the detection results. At the same time, relying on the cross-signal acquisition device and cross-environment adaptability of the general temporal features, the prediction output head can stably output accurate detection results, weakening the detection bias caused by heterogeneous signal acquisition devices and environments, improving the adaptability and detection stability of the target model in real-world scenarios, and ultimately achieving accurate perception and detection of targets within the target scene, meeting the needs of large-scale and generalized practical applications.

[0070] In some implementations, step (103.b.1) is preceded by: performing a padding operation on the unified dimension feature sequence to align unified dimension feature sequences of different time lengths to a preset time length, thereby obtaining an aligned unified dimension feature sequence. Then step (103.b.1) includes: obtaining the temporal relationship features corresponding to the aligned unified dimension feature sequence through a shared backbone network using an attention mechanism, thereby obtaining general temporal features.

[0071] Specifically, before obtaining the temporal relationship features corresponding to the unified-dimensional feature sequence through the shared backbone network using the attention mechanism to obtain general temporal features, a padding operation is first performed on the unified-dimensional feature sequence. Specifically, zero-padding or other feature padding is applied to the time dimension of the unified-dimensional feature sequence according to a preset time length. This aligns the unified-dimensional feature sequences of different time lengths obtained from different acquisition devices and scenarios to the same fixed preset time length, thus forming an aligned unified-dimensional feature sequence. The technical effect of this step is to effectively solve the problem of inconsistent channel state sequence lengths under different devices and environments, eliminate temporal modeling gaps caused by time length differences, and ensure that the shared backbone network can perform stable and continuous temporal relationship learning based on feature sequences of uniform length. This provides an input foundation with consistent format and length for the subsequent accurate extraction of general temporal features. Furthermore, after alignment, the aligned unified-dimensional feature sequence is input into the shared backbone network. The network's internal attention mechanism module calculates the correlation strength and dependency between features at different time points in the sequence. By dynamically allocating attention weights, key temporal features reflecting target behavior and channel dynamic changes are strengthened, while the impact of invalid noise and redundant features is weakened. Weighted fusion and global temporal information integration are performed on features at each time point, ultimately extracting universal temporal features that characterize target activity and environmental changes and possess cross-signal acquisition device and cross-environment universality. The technical effect of this step is that the attention mechanism can accurately capture long-distance temporal correlations, fully utilize the complete time-frequency domain information contained in the unified-dimensional feature sequence, and ensure that the generated universal temporal features not only contain detailed patterns of channel changes but also possess good adaptability and robustness. This provides high-quality, stable, and reliable temporal feature support for the subsequent prediction output head to accurately output target detection results.

[0072] In some implementations, the attention mechanism is a multi-head self-attention mechanism. Based on this mechanism, long-distance temporal relationship features are extracted from a unified-dimensional feature sequence to obtain general temporal features. For example, step (103.b.1) may include: using a shared backbone network, performing temporal dependency modeling on the unified-dimensional feature sequence according to the multi-head self-attention mechanism to obtain the association strength and dependency features between features at different time positions in the unified-dimensional feature sequence; calculating the corresponding attention weights based on the association strength and dependency features, and weighting and fusing the features at different time positions according to the attention weights to obtain fused features containing global temporal change patterns; and extracting long-distance temporal relationship features based on the fused features to obtain general temporal features.

[0073] Among them, the general time-series feature is the hidden layer feature sequence output by the shared backbone network after performing feature encoding on the feature sequence of the same dimension. The hidden layer feature sequence simultaneously encodes the time-domain dynamic information and frequency-domain structural information of the target channel state sequence.

[0074] The shared backbone network comprises multiple adaptive coding modules, each consisting of a multi-head self-attention layer, a feed-forward network layer, residual connections, and layer normalization. The attention mechanism is a multi-head self-attention mechanism, used to capture long-distance temporal dependencies at different time points in a feature sequence of the same dimension. The adaptive coding modules work collaboratively, with the multi-head self-attention mechanism at its core, progressively completing temporal dependency modeling, feature weighted fusion, and long-distance temporal feature extraction. Each step is progressive, ensuring both the completeness and accuracy of temporal feature extraction while enhancing the cross-device and cross-environment adaptability of general temporal features.

[0075] Specifically, firstly, the unified-dimensional feature sequence is input into the first adaptive encoding module of the shared backbone network. The multi-head self-attention layer of this module serves as the core computational unit, employing a multi-head self-attention mechanism to model the temporal dependency of the unified-dimensional feature sequence. By using multiple attention heads set in parallel, the correlation between features at different time positions in the unified-dimensional feature sequence is captured simultaneously from different dimensions and scales. The correlation strength of the interaction between features at different time positions and the temporal dependency between features at different time positions are accurately calculated, thus obtaining correlation strength and dependency features that can clearly characterize the correlation of features at each time position. After the multi-head self-attention layer completes the modeling, the original unified-dimensional features input to this module are residually fused with the output features of the multi-head self-attention layer through residual connections. Subsequently, layer normalization is performed to eliminate feature distribution offset, ensuring the stability of the feature data and providing reliable support for subsequent processing. The technical benefits of this step are that the cascaded design of multiple adaptive encoding modules enables hierarchical and progressive extraction of temporal features. Compared to a single module, it can gradually uncover deeper and more comprehensive temporal correlations. The multi-head self-attention layer can mine temporal correlations from multiple dimensions, avoiding the limitations of single-dimensional feature extraction, accurately capturing subtle temporal changes and correlation patterns in a unified-dimensional feature sequence, and adapting to standardized unified-dimensional feature sequences under different devices and environments. Residual connections can effectively alleviate the gradient vanishing problem during deep network training, and layer normalization can stabilize feature distribution. The synergistic effect of these two methods ensures the stability and accuracy of temporal dependency modeling, providing a comprehensive and accurate foundation for subsequent attention weight calculation and feature fusion.

[0076] After the first adaptive coding module completes the initial temporal modeling, the features processed by layer normalization are input into the feedforward neural network layer of this module. The features undergo nonlinear transformation and dimensionality optimization to enhance the representational ability of effective features. Subsequently, the input and output features of the feedforward neural network layer are fused again through residual connections, and a second layer normalization process is performed to obtain a feature sequence that has been initially optimized. This feature sequence is then passed into subsequent adaptive coding modules that share a backbone network. Each subsequent module repeats the above process of "multi-head self-attention modeling, residual connection, layer normalization, feedforward neural network optimization, residual connection, and layer normalization" to continuously deepen the temporal dependency modeling and further refine the correlation strength and dependency features between features at different time positions. Based on this, according to the correlation strength and dependency features output by each adaptive coding module, attention weights are calculated for features at different time positions in the unified dimension feature sequence through preset weight calculation rules. Among them, key features representing target activities and channel dynamic changes are assigned higher attention weights, while invalid and redundant features are assigned lower attention weights. Then, according to the calculated attention weights, the features at different time positions are weighted and fused to integrate the effective feature information at each time position, and finally obtain fused features containing global temporal change patterns. The technical benefits of this step are that the feedforward neural network layer can perform nonlinear optimization on temporal features, enhance the representational ability of features, and further filter out core effective features; the repeated application of residual connections and layer normalization can continuously stabilize the feature distribution, avoid the degradation of deep features, and ensure the feature output quality of each module; the hierarchical processing of multiple adaptive coding modules can gradually deepen the temporal correlation mining, make the calculation of attention weights more targeted, and the fused features after weighted fusion can centrally reflect the core temporal information of the feature sequence of the same dimension, realize the effective integration of global temporal features, break the limitations of local temporal features, lay a solid foundation for the extraction of subsequent long-distance temporal relationship features, and further improve the universality and effectiveness of features.

[0077] Finally, after layered processing and feature optimization by all adaptive coding modules, based on the fused features that contain global temporal variation patterns, the output layer of the shared backbone network is used to further filter and refine long-distance temporal relationship features that reflect general patterns across devices and environments. Local redundant features related to device and environmental characteristics are eliminated, ultimately yielding general temporal features adapted to cross-device and cross-environment scenarios. The technical effect of this step is that the synergistic effect of multiple adaptive coding modules can fully explore long-distance temporal dependencies, completely preserve the general patterns of channel state temporal changes, and effectively weaken the feature differences caused by device and environmental heterogeneity. The cooperation of residual connections, layer normalization, and feedforward neural network layers ensures the stability and accuracy of the feature extraction process. This results in general temporal features that possess both good universality, adapting to unified-dimensional feature sequences from different sources, and strong specificity, accurately representing the correlation between target activity and channel state changes. This provides high-quality, stable, and reliable temporal feature support for subsequent prediction output head decoding to generate accurate target detection results, further improving the generalization ability and detection accuracy of the target model.

[0078] In this embodiment, the target model is obtained by training a preset model to minimize the loss of the joint reconstruction loss and the prediction loss. The preset model outputs the predicted detection result and the reconstructed channel state information sequence based on the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

[0079] Specifically, the training of the target model adopts a two-stage joint loss training strategy. It uses a pre-set model as the training basis, which includes an embedded unified representation network, a shared backbone network, a reconstruction output head, and a prediction output head. Based on the input sample channel state information sequence, it can simultaneously output a reconstructed channel state information sequence for self-supervised training and a prediction detection result for supervised training. Through the joint constraint of the joint reconstruction loss and the prediction loss, the pre-set model is subjected to loss minimization iterative training, and finally a target model that meets the requirements of target detection accuracy and generalization is obtained.

[0080] During training, firstly, a reconstruction loss is constructed based on the reconstructed channel state information sequence output by the preset model and the input sample channel state information sequence. This reconstruction loss is determined by calculating the numerical difference between the two and is used to quantify the preset model's ability to reconstruct the original sample channel state information sequence. Its core function is to force the embedded unified representation network and the shared backbone network in the preset model to learn the inherent physical characteristics of the sample channel state information sequence through self-supervised constraints. These characteristics include the frequency domain correlation of adjacent subcarriers, the time domain variation characteristics of continuous time segments, and the channel dynamic laws that are universal across signal acquisition devices and environments. This effectively avoids the model only fitting the local features of a single device or environment, thereby improving the universality of the features learned by the model and laying a solid foundation of universal features for subsequent target detection tasks. Meanwhile, a prediction loss is constructed based on the predicted detection results output by the preset model and the sample detection results corresponding to the sample channel state information sequence. This prediction loss is determined by calculating the class difference between the two and is used to quantify the target detection accuracy of the preset model. Its core function is to guide the prediction output head in the preset model to learn the mapping relationship between general temporal features and target detection results through supervision constraints, so that the model can accurately adapt to specific target detection tasks based on mastering general channel features, thereby improving the model's recognition accuracy of target behavior and state. Furthermore, during iterative training, the reconstruction loss and prediction loss are fused to form a joint loss. With the goal of minimizing the joint loss, the network parameters of the embedded unified representation network, shared backbone network, reconstruction output head, and prediction output head in the preset model are updated sequentially through the backpropagation algorithm. The pre-training stage focuses on updating parameters with the reconstruction loss constraint to ensure the effective learning of general features. The fine-tuning stage focuses on updating parameters with the prediction loss constraint to improve detection accuracy without destroying the learned general features. Through two-stage collaborative training and joint loss constraint, a target model that balances cross-signal acquisition device and cross-environment generalization ability with target detection accuracy is finally obtained. This effectively solves the technical problems of poor adaptability and weak generalization ability of traditional models, reduces model deployment costs, and improves the stability and applicability of the model in real-world scenarios.

[0081] In some implementations, a preset model is first pre-trained based on local sample channel state information sequences to obtain candidate models. Then, the pre-trained candidate models are fine-tuned by combining the sample channel state information sequences and sample detection results to obtain the trained target model. For example, the training process of the target model is as follows: (A.1) Obtain the local preset channel state information sequence and the corresponding sample detection results, and preprocess the preset channel state information sequence to obtain the sample channel state information sequence; (A.2) Input the sample channel state information sequence into the preset model to obtain the reconstructed channel state information sequence output by the reconstruction output head of the preset model; (A.3) Construct the reconstruction loss based on the difference between the reconstructed channel state information sequence and the sample channel state information sequence; (A.4) Based on the reconstruction loss, adjust the model parameters of the preset model starting from the reconstruction output head according to the backpropagation method, and perform iterative training to obtain the trained candidate model; (A.5) Input the sample channel state information sequence into the candidate model to obtain the prediction detection result output by the prediction output head; (A.6) Construct a prediction loss based on the difference between the predicted detection results and the sample detection results, and fine-tune the candidate model based on the prediction loss to obtain the trained target model.

[0082] The preset channel state sequence comprises channel state sequences collected from multiple heterogeneous acquisition devices and in multiple heterogeneous deployment environments. These heterogeneous acquisition devices differ in the number of antennas, subcarriers, and sampling rates. Specifically, the preset channel state information sequence refers to the raw sample data used for model pre-training and fine-tuning. It is a locally stored channel state information sequence originating from different acquisition devices and deployment environments, without any preprocessing operations such as anomaly removal or normalization. It contains the raw amplitude and phase information generated during channel propagation, as well as various interference data. It forms the basis for preparing the sample channel state information sequence, providing raw data support covering different scenarios and devices for model training, ensuring that the trained model has cross-device and cross-environment generalization capabilities.

[0083] Among them, the sample channel state information sequence refers to the standardized channel state information sequence obtained by sequentially performing abnormal data removal, amplitude normalization, phase correction and noise suppression on the preset channel state information sequence according to the preset preprocessing process. It eliminates anomalies, noise and distortion in the original data, unifies the distribution and format of sample data, and retains the core time-frequency domain characteristics of the channel state. It can accurately adapt to the input requirements of the preset model and candidate model. It is the core input data for reconstruction loss calculation in the model pre-training stage and prediction loss calculation in the fine-tuning stage, providing high-quality and standardized sample support for model parameter optimization.

[0084] Among them, the reconstructed channel state information sequence refers to the channel state information sequence output by the reconstruction output head after the sample channel state information sequence is input into the preset model during the model pre-training stage, processed by the embedded unified representation network feature mapping and shared backbone network, and the features are decoded and restored by the reconstruction output head. Its core is used to simulate the original feature distribution of the sample channel state information sequence. The difference between it and the sample channel state information sequence is used to construct the reconstruction loss, providing a self-supervised constraint basis for the adjustment of model parameters during the pre-training stage, and helping the model learn the inherent physical characteristics and general laws of the channel state.

[0085] It should be noted that the process of pre-training the preset model to obtain the candidate model and then fine-tuning it to obtain the target model adopts a two-stage training strategy of "pre-training-fine-tuning". The reconstruction loss and prediction loss are used as the core constraints of the two stages, respectively. Based on the sample channel state information sequence and the corresponding sample detection results, the model parameters are gradually optimized to ensure that the target model has both cross-device and cross-environment generalization ability and can achieve accurate target detection.

[0086] Specifically, firstly, a pre-defined channel state information sequence stored locally is acquired, along with corresponding sample detection results. The sample detection results characterize the actual target situation corresponding to the pre-defined channel state information sequence, providing a standard basis for supervised training in the subsequent fine-tuning stage. Then, following a pre-defined preprocessing procedure, the acquired pre-defined channel state information sequence undergoes abnormal data removal, amplitude normalization, phase correction, and noise suppression to obtain a standardized and highly reliable sample channel state information sequence. The technical effect of this step is that preprocessing eliminates abnormal data, noise interference, and amplitude and phase distortion in the pre-defined channel state information sequence, unifies the distribution and format of sample data, avoids interference from invalid data on model training, and ensures that the sample channel state information sequence accurately adapts to the input requirements of the pre-defined model. This provides a high-quality, standardized sample foundation for model parameter optimization in the subsequent pre-training and fine-tuning stages, guaranteeing the stability and effectiveness of the training process.

[0087] Then, in the model pre-training stage, the preprocessed sample channel state information sequence is input into the preset model. The embedded unified representation network in the preset model performs feature mapping on the sample channel state information sequence to obtain a unified dimension feature sequence. This sequence is then processed by the shared backbone network and fed into the reconstruction output head. The reconstruction output head decodes and restores the features, outputting a reconstructed channel state information sequence. This reconstructed channel state information sequence is used to simulate the original feature distribution of the sample channel state information sequence. Next, based on the numerical difference between the reconstructed channel state information sequence and the input sample channel state information sequence, a preset loss calculation method is used to construct the reconstruction loss. This reconstruction loss is used to quantify the accuracy of the preset model in restoring the sample channel state information sequence. Subsequently, with the goal of minimizing the reconstruction loss, the gradient is backpropagated sequentially from the reconstruction output head to the shared backbone network and the embedded unified representation network, gradually adjusting the network parameters of each module of the preset model and performing multiple rounds of iterative training until the reconstruction loss converges to a preset threshold, resulting in the trained candidate model. The technical effect of this pre-training stage is that, through the self-supervised constraint of the reconstruction loss, the pre-set model is forced to learn the inherent physical characteristics of the sample channel state information sequence, including the frequency domain correlation of adjacent subcarriers, the time domain variation characteristics of continuous time segments, and the channel dynamic laws that are common across devices and environments. This effectively avoids the model only fitting the local features of a single device or environment, and improves the generality of the features learned by the model. At the same time, the method of adjusting the parameters in reverse from the reconstruction output head can prioritize the reconstruction accuracy and ensure the model's ability to restore the channel state sequence. This lays a solid foundation for the detection accuracy optimization in the subsequent fine-tuning stage and effectively improves the generalization ability of the candidate model.

[0088] Furthermore, after pre-training, the model fine-tuning stage begins. The preprocessed sample channel state information sequence is input again into the trained candidate model. The candidate model's embedding unified representation network and shared backbone network sequentially perform feature mapping and temporal modeling on the sample channel state information sequence to obtain general temporal features, which are then fed into the prediction output head. The prediction output head decodes and classifies the general temporal features, outputting the predicted detection result. Subsequently, based on the category difference between the predicted detection result and the preset sample detection result, a prediction loss is constructed. This prediction loss is used to quantify the target detection accuracy of the candidate model and provide supervision for model parameter fine-tuning. Finally, with the goal of minimizing the prediction loss, the network parameters of the candidate model are fine-tuned and optimized. The parameters of the embedding unified representation network and the prediction output head are adjusted, while the core parameters of the shared backbone network are fixed to avoid destroying the general features learned in the pre-training stage during the fine-tuning process. After multiple rounds of iterative fine-tuning until the prediction loss converges, the trained target model is finally obtained. The technical effect of this fine-tuning stage is that, through the supervised constraints of the prediction loss, it guides the candidate model to accurately learn the mapping relationship between general temporal features and target detection results. While retaining the generalization ability of the pre-training stage, it significantly improves the target detection accuracy of the model. The key adjustment is to the way the unified representation network and the prediction output head parameters are embedded, which not only ensures the model's adaptability to heterogeneous samples, but also optimizes the detection performance in a targeted manner. Ultimately, the target model can take into account both cross-device and cross-environment generalization ability and accurate target detection ability, solve the problems of poor adaptability, weak generalization and insufficient detection accuracy of traditional models, and meet the needs of large-scale and generalized practical applications.

[0089] The reconstruction loss, specifically the mean squared error loss, characterizes the numerical difference between the reconstructed channel state sequence and the sample channel state sequence. The calculation process for the reconstruction loss is as follows:

[0090] in, Indicates the reconstruction loss. This represents the sequence of reconstructed channel state information. This represents a sequence of sample channel state information.

[0091] The prediction loss, specifically the cross-entropy loss, characterizes the difference in class probabilities between the predicted detection result and the sample detection result. The calculation process for the prediction loss is as follows:

[0092] in, Indicates the predicted loss. Indicates the predicted detection result. This indicates the sample test results.

[0093] In some implementations, based on the reconstruction loss, the gradients of the reconstructed output head, the shared backbone network, and the embedded unified representation network are calculated using backpropagation. The model parameters of these three networks are then adjusted to obtain a candidate model. For example, if the preset model includes at least an embedded unified representation network, a shared backbone network, and a reconstructed output head, step (A.4) may include: calculating the gradients of the reconstructed output head, the shared backbone network, and the embedded unified representation network sequentially using backpropagation based on the reconstruction loss; adjusting the network parameters of the reconstructed output head, the shared backbone network, and the embedded unified representation network according to the corresponding gradients; and performing iterative training until convergence to obtain the trained candidate model.

[0094] Specifically, the pre-defined model includes at least an embedded unified representation network, a shared backbone network (containing a multi-head self-attention mechanism), and a reconstruction output head. The embedded unified representation network is used for feature mapping, the shared backbone network for temporal relationship modeling, and the reconstruction output head for generating reconstructed channel state information. These three components work together to complete model training and parameter tuning. Based on this, according to the reconstruction loss, and following backpropagation, the gradient information of the reconstruction output head, the shared backbone network, and the embedded unified representation network is calculated sequentially, starting from the reconstruction output head. Then, based on the gradient information, the model parameters of the reconstruction output head, the shared backbone network, and the embedded unified representation network are adjusted respectively. This is followed by multiple rounds of iterative training until the reconstruction loss converges to a preset threshold, ultimately yielding the trained candidate model. In this way, by using backpropagation, the parameter deviations of each network module can be accurately located, and the feature mapping accuracy of the embedded unified representation network, the temporal modeling capability of the shared backbone network, and the reconstruction accuracy of the output head can be optimized in a targeted manner. This avoids model training failure caused by parameter deviations and improves the stability and generalization ability of the model. In addition, through multiple rounds of iterative training, the model parameters can be gradually optimized, enabling the model to continuously learn the temporal patterns and feature correlations of the channel state, correct parameter deviations, and ensure that the candidate model can accurately capture the dynamic changes of the channel and the characteristics of target activity. This provides reliable model support for subsequent target detection, while strengthening the model's adaptability to different devices and environments, and solving the problem of insufficient generalization of traditional models.

[0095] By using the above methods, channel state information sequences of any input form can be mapped to a unified dimension feature sequence. By capturing the global temporal relationship features of the channel dynamic change patterns caused by human movement or environmental changes, target detection can be performed. This enhances the universality of target detection scenarios based on channel state information.

[0096] To better understand the embodiments of this application, Figure 3 This is an example diagram of the architecture of the target model provided in the embodiments of this application. Figure 4This application provides an example diagram comparing the performance of the target model with other models in its embodiments. Figure 3 and Figure 4 As shown, an example is given to illustrate the target detection process based on channel state information, as follows: like Figure 3 As shown, this embodiment constructs a unified CSI perception modeling system for heterogeneous WiFi devices and environments. The system mainly includes a CSI data acquisition module, a data preprocessing and unified interface module, a CSI signal embedding module, a unified temporal representation learning model, and a task adaptation and output module. The model operation process is described below, based on these modules: First, CSI signal data from the environment is collected using different WiFi devices. Due to differences in center frequency, bandwidth, number of subcarriers, and antenna configuration among different devices, the collected CSI data exhibits significant differences in dimensional structure and sampling method. Then, the data preprocessing module performs outlier removal, amplitude normalization, phase correction, and noise suppression on the collected CSI signals to reduce the impact of device differences and organizes data from different sources into a unified format through a unified data interface. Subsequently, the preprocessed CSI sequence is input into the CSI signal embedding module, mapping the CSI data with different structures into a unified dimensional feature representation. This feature sequence is further input into the unified temporal representation learning model for modeling to extract the dynamic change features of the wireless channel. Finally, the task adaptation module outputs specific perception results, such as results for tasks like human motion recognition, gesture recognition, or identity recognition.

[0097] To make it easier to understand, the following steps are explained using a separate example: 1) Heterogeneous CSI data acquisition and preprocessing steps First, channel state information (CSI) data is collected from the environment using a WiFi device. The CSI data includes complex channel response information corresponding to different subcarriers, antennas, and timestamps. Since different devices vary in center frequency, bandwidth, number of antennas, and sampling rate, this step preprocesses the raw CSI data, including outlier removal, amplitude normalization, phase correction, and noise suppression, thereby reducing the impact of device differences and acquisition errors on subsequent modeling. The preprocessed CSI sequence preserves the time and frequency domain structure of the original signal, providing standardized input for subsequent unified modeling. The input representation is as follows:

[0098] in, Indicates the batch size. Indicates the length of time. This represents the number of channels (number of antennas x subcarriers).

[0099] 2) CSI signal unified representation mapping steps To address the issue of variations in subcarrier numbers, antenna configurations, and sampling lengths across different datasets, this example constructs a dataset adaptation embedding module to map CSI signals with different structures to a unified feature representation space. Specifically, a convolutional embedding network extracts features from the CSI sequences, capturing local signal variation features in the time dimension and aggregating the frequency domain correlations between adjacent subcarriers to generate a unified feature sequence representation. This mapping process achieves feature alignment between different data sources while preserving the original physical structure of the CSI signals.

[0100] Because CSI data collected by different WiFi devices varies in the number of antennas, subcarriers, and sampling length, the original CSI data structure is usually different. Therefore, this embodiment first uses a CSI signal embedding module to uniformly represent and map the original CSI sequence as [X\in\mathbb{C}^{N_a\times N_s\times T}], where N_a represents the number of antennas; N_s represents the number of subcarriers; and T represents the time sampling length. The specific process is as follows: Specifically, firstly, the antenna dimension and subcarrier dimension are flattened to obtain data in time series form, represented as [X'\in\mathbb{R}^{T\times F}], where F=N_a\times N_s.

[0101] Subsequently, the time series is input into a one-dimensional convolutional embedding network for feature extraction. The convolutional embedding network mainly performs convolution operations along the time dimension to capture local temporal variation features in the CSI signal, while aggregating the frequency domain correlation between adjacent subcarriers. It should be noted that the embedding module includes two one-dimensional convolutional layers: the first convolutional layer is used to extract local temporal features, and the second convolutional layer is used to generate a uniform-dimensional embedding representation.

[0102] First, perform a dimensional transformation, represented as

[0103] After processing by the convolutional embedding module, the original CSI sequence is mapped to a unified-dimensional feature sequence as [Z\in \mathbb{R}^{T'\times d}\], where T' is the new time length and d is the unified feature dimension. For example, the embedding representation process is as follows:

[0104]

[0105] Then convert it to sequence form, which is represented as: ; The essence is ; Therefore, a uniform length is obtained, represented as: ; Through this mapping, heterogeneous CSI data generated by different devices can be converted into a feature representation with a unified dimension, thereby achieving unified modeling across devices and datasets.

[0106] 3) Unified temporal representation learning steps The mapped CSI feature sequences are input into a unified temporal representation learning model for modeling. This model uses a self-attention mechanism to model the temporal dependencies in the CSI sequences, enabling the system to capture the dynamic channel changes caused by human movement or environmental variations. Through joint training under multiple datasets, the model learns a universally applicable representation of wireless channel dynamic features, thereby improving its generalization ability under different devices and environmental conditions.

[0107] Specifically, after obtaining the unified-dimensional feature sequence, the unified-dimensional feature sequence is input into the unified temporal representation learning module for modeling. The unified temporal representation learning module adopts a Transformer model based on the GPT-2 architecture. This model consists of a multi-layer self-attention network, used to capture long-range temporal dependencies in the CSI sequence.

[0108] Let the unified dimension feature sequence after embedding be: Z={z_1,z_2,...,z_{T'}\}, where (z_i\in\mathbb{R}^{d}).

[0109] In each Transformer layer, the correlation between different time positions in the sequence is first calculated using a multi-head self-attention mechanism. The calculation formula is as follows: [Attention(Q,K,V)=softmax (frac{QK^T}{sqrt{d_k}} )V], where Q is the query matrix; K is the key matrix; V is the value matrix; and d_k is the dimension of the key vector.

[0110] Through this self-attention mechanism, the model can automatically learn the dependencies between different time slices in the CSI sequence, thereby capturing the dynamic characteristics of the wireless channel caused by human actions or environmental changes.

[0111] It should be noted that the GPT-2 structure is composed of multiple stacked Transformer Blocks. Each layer contains a multi-head self-attention layer, a feedforward neural network layer, a residual connection, and layer normalization. After multi-layer Transformer computation, a unified CSI representation with rich temporal semantic information can be obtained as [H\in\mathbb{R}^{T'times d}]. This representation can simultaneously encode temporal dynamic information and frequency domain structure information, thereby achieving effective modeling of the changing patterns of wireless channels.

[0112] The processing procedure for each Transformer Block is as follows: ( ), .

[0113] 4) Cross-dataset joint training and task adaptation steps A cross-dataset joint training mechanism is built upon the unified representation learning model, enabling collaborative learning of data from different task scenarios within the same training framework. The model learns a general signal representation by sharing a feature extraction layer and incorporates a lightweight task module at the output to adapt to various perceptual tasks, such as human motion recognition, gesture recognition, or identity recognition. Joint training with multi-source data enhances the model's adaptability to different environmental conditions and reduces reliance on training on a single dataset.

[0114] Task adaptation module: After obtaining a unified CSI representation, the task adaptation module completes the prediction of specific perception tasks.

[0115] The task adaptation module adopts a multilayer perceptron (MLP) structure, with a unified representation H as input and the prediction result of the corresponding task as output.

[0116] For example, in human behavior recognition tasks, the output is the probability of behavior category: hat{y}=softmax(W*h+b), where h is the Transformer output feature, W is the weight matrix, and b is the bias term.

[0117] It should be noted that the model training process employs a cross-dataset joint training approach. The training data originates from multiple CSI datasets collected from different devices and environments. Data from different datasets is loaded through a unified data interface, and training batches are formed through random sampling during the training process.

[0118] Model training consists of two phases: The first stage is the representation learning pre-training stage: During the pre-training phase, a unified temporal representation learning model is trained through the CSI signal reconstruction task, enabling the model to learn a general representation of the dynamic features of wireless channels.

[0119] Let the original CSI signal be (X\) and the model reconstruction output be hat{X}, then the reconstruction loss function is the mean squared error loss: L_{rec}=frac{1}{N}sum_{i=1}^{N}(X_i - hat{X}_i)^2, where N is the number of samples.

[0120] Then reconstruct the header: , ; Right now,

[0121] in, Indicates the reconstruction loss. This represents the sequence of reconstructed channel state information. This represents a sequence of sample channel state information.

[0122] The second phase is the downstream task training phase: After pre-training is completed, a task adaptation module is added to the model output, and fine-tuning training is performed for specific perception tasks.

[0123] For classification tasks, the cross-entropy loss function is used: L_{cls}=sum_{i=1}^{C}y_i log(hat{y}_i)], where y_i is the true label, hat{y}_i is the model prediction probability, and C is the number of categories.

[0124] Task Category: ; Right now ; in, Indicates the predicted loss. Indicates the predicted detection result. This indicates the sample test results.

[0125] The final model optimization objective is: L = L_{rec} + ambda * L_{cls}, where lambda is the loss weight coefficient.

[0126] During training, the Adam optimization algorithm is used to update the model parameters, thereby obtaining a unified wireless sensing model that can adapt to multiple devices and environmental conditions.

[0127] Example illustration, using CSI data collected from two different WiFi devices: The data acquired by signal acquisition device A has the following dimensions: [3 times 30 times 500]. The data collected by device B has the following dimensions: [2 times 56 times 800]. The CSI signal embedding module proposed in this example can uniformly map CSI data with different structures into a feature sequence representation of [T' \times d]. This feature sequence is then input into a unified temporal representation learning module for modeling, thereby achieving unified CSI perception modeling across devices and environments.

[0128] To understand the above examples, the process is summarized in detail below: First, a lightweight CNN embedding module specific to the dataset is introduced at the input end to uniformly complete local feature extraction (including subcarrier correlation and short-term dynamics), mapping CSI data of different formats into feature sequences of consistent dimensions. Then, this sequence is input into a shared Transformer backbone specifically for modeling long-term temporal dependencies and dynamic patterns across time. In this way, the task of "adapting to different data formats," originally undertaken by the model structure, is partially absorbed by the front-end embedding module, while the backbone network focuses only on the spatiotemporal structural patterns shared across datasets.

[0129] Furthermore, through joint training across datasets, the model is exposed to multiple signal distributions simultaneously during the training phase (such as different subcarrier configurations, sampling rates, and environmental conditions), thereby learning a more universal representation. This design achieves a division of labor mechanism of "local adaptation + global sharing," enabling the model to maintain high recognition accuracy on different datasets without relying on a specific data format or dedicated structure, fundamentally alleviating the problem of strong coupling between data and model structure.

[0130] For example, dataset a consists of 3 antenna pairs (generally, m transmit and n receive antennas represent m*n pairs), each pair has 10 subcarriers, and each data point has 70 timestamps. Therefore, it will be processed into a data format of [3*10, 70]. Dataset b is processed into a format of [50, 100].

[0131] Then, the data will have features extracted by their respective feature extraction layers, i.e., CNN networks. For example, if the CNN extracts 25 features, they will become [25, 70] and [25, 100] respectively. Here, the total number of features is kept constant, which is 256 in this case. Each CNN is a 2-layer CNN, designed according to the size of its respective dataset, which is [antenna pair * subcarrier, timestamp].

[0132] Next, the data [25, 70] and [25, 100] are processed through a transformer-structured network. All datasets share the same network, which pads all sequences to a uniform length using padding (filling in empty tokens). The processed high-dimensional vectors are then set as feature_a and feature_b. Each dataset is connected to a lightweight neural network called the reconstruction head, which reconstructs the original data using the features and calculates the loss function based on the reconstruction results. There is also a task head, which outputs classification task results based on the features.

[0133] When updating the network, for example, if the current batch is of dataset a, then the network CNNa, transformer, reconstruct_a, and task_a will be updated with gradients, while the others will be frozen.

[0134] It's important to note that during the pre-training phase, all datasets are updated together. During the fine-tuning phase, to obtain a model for a specific dataset, the model is updated only with that dataset; the heads of other datasets are not updated.

[0135] In a CNN, the layer is called the embed layer, and the data that is embedded and mapped is called the feature.

[0136] Furthermore, when transferring to other datasets not seen during the pre-training phase, the transformer remains, while the other CNN_x, reconstruct_x, and tasl_x are retrained.

[0137] The convolution is a one-dimensional convolution. Both the reconstruction head and the task head use a single MLP layer, directly outputting the reconstruction or classification results.

[0138] Therefore, by implementing the above target detection scenario examples based on channel state information, combined with Figure 4 As shown, under full training conditions across multiple datasets, the performance of traditional deep learning models and the method of this invention is compared on multiple WiFi sensing datasets. Several common models were selected for comparison, including multilayer perceptron models, convolutional neural network models, recurrent neural network models, and the standard Transformer model. Experimental results show that traditional methods exhibit significant performance differences across different datasets. Some models achieve high recognition accuracy on certain datasets but show a significant performance drop on others, indicating that their model structure is strongly dependent on specific signal distributions. In contrast, Figure 4The “Ours” in the text demonstrates that the unified representation learning model proposed in this application can maintain relatively stable recognition performance on multiple datasets, and its average recognition accuracy is significantly higher than that of traditional models, indicating that the method can effectively learn universal CSI feature representations under cross-device and cross-environment conditions.

[0139] As can be seen from the above embodiments, this application embodiment can obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scene; preprocess the original channel state information sequence to obtain the processed target channel state information sequence; input the target channel state information sequence into the trained target model, and map the target channel state information sequence to obtain a unified dimension feature sequence through the target model, and output the target detection result according to the temporal features corresponding to the unified dimension feature sequence; wherein, the target model is obtained by minimizing the loss of a preset model through joint reconstruction loss and prediction loss, the preset model outputs the predicted detection result and the reconstructed channel state information sequence based on the sample channel state information sequence, the reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result. In this way, through the trained target model, channel state information sequences of any input form can be mapped to a unified dimension feature sequence and target detection can be performed, enhancing the universality of target detection scenarios based on channel state information.

[0140] Therefore, compared to traditional models in related technologies that are highly coupled with data input formats, resulting in poor model device adaptability, weak cross-environment generalization ability, and trained models that are only applicable to target detection of channel state information with fixed input formats, this application preprocesses the original channel state information sequence to obtain the target channel state information sequence, thereby removing invalid interference noise and improving the accuracy of subsequent target detection. Furthermore, the trained target model maps the target channel state information sequence to a unified dimension and predicts the target detection results in the target scene based on the feature length dependency relationship of the unified dimension feature sequence. Thus, the target model has high adaptability and strong cross-environment generalization ability, applicable to target detection of channel state information with any input format, has high applicability, reduces deployment costs, enhances the universality of target detection scenarios based on channel state information, meets the needs of large-scale and generalized practical applications, and improves the efficiency of target detection based on channel state information.

[0141] To facilitate better implementation of the target detection method based on channel state information provided in this application, this application also provides a target detection device based on the aforementioned channel state information. The meanings of the terms used are the same as in the target detection method based on channel state information described above, and specific implementation details can be found in the descriptions within the method embodiments.

[0142] Please see Figure 5 , Figure 5 This is a schematic diagram of the target detection device based on channel state information provided in an embodiment of this application. The target detection device based on channel state information is integrated into the computer equipment of this application, such as a server. The target detection device based on channel state information may include an acquisition unit 401, a preprocessing unit 402, and an input unit 403.

[0143] The acquisition unit 401 is used to acquire the original channel state information sequence corresponding to the wireless transmitting device in the target scene; Preprocessing unit 402 is used to preprocess the original channel state information sequence to obtain the processed target channel state information sequence; The input unit 403 is used to input the target channel state information sequence into the trained target model, map the target channel state information sequence through the target model to obtain a unified dimension feature sequence, and output the target detection result according to the time series features corresponding to the unified dimension feature sequence. The target model is obtained by training a preset model by minimizing the joint reconstruction loss and prediction loss. The preset model is based on the predicted detection result and the reconstructed channel state information sequence output by the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

[0144] In some implementations, the target model includes an embedded unified representation network, which comprises a first convolutional layer and a second convolutional layer. The input unit 403 is further configured to: Through the first convolutional layer in the target model, local time variation features are extracted for the target channel state information sequence in the time dimension, and frequency domain variation features are extracted for the channel state information corresponding to adjacent subcarriers in the target channel state information sequence in the frequency domain dimension. The second convolutional layer in the target model is used to fuse and map local time-varying features and frequency domain-varying features to generate a unified-dimensional feature sequence.

[0145] In some embodiments, the input unit 403 is further configured to: By using the first convolutional layer in the target model, a one-dimensional convolution operation is performed on the target channel state information sequence along the time dimension to obtain the local time variation characteristics of the target channel state information sequence in continuous time segments. The first convolutional layer in the target model extracts features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence along the frequency domain dimension, obtaining the corresponding channel state features. Then, the similar change features between the channel state features are aggregated to obtain the frequency domain change features of the target channel state information sequence.

[0146] In some implementations, the target model includes a shared backbone network, and the input unit 403 is further configured to: By sharing a backbone network and using an attention mechanism, we obtain temporal relationship features corresponding to a unified dimension feature sequence, thus obtaining general temporal features. The target detection results are output based on general temporal features through the prediction output head in the target model.

[0147] In some embodiments, the input unit 403 is further configured to: By sharing a backbone network and using a multi-head self-attention mechanism, temporal dependency modeling is performed on a unified dimension feature sequence to obtain the correlation strength and dependency features between features at different time positions in the unified dimension feature sequence. The attention weights are calculated based on the correlation strength and dependency features, and the features at different time positions are weighted and fused according to the attention weights to obtain fused features that contain global temporal change patterns. Based on the fusion feature extraction, long-distance temporal relationship features are obtained to obtain general temporal features.

[0148] In some embodiments, the preprocessing unit 402 is further configured to: The original channel state information sequence is subjected to abnormal data removal to obtain the initial channel state information sequence after removal. The amplitude information in the initial channel state information sequence is normalized, and the phase information in the initial channel state information sequence is corrected to obtain the candidate channel state information sequence. The signal noise in the candidate channel state information sequence is suppressed to obtain the processed target channel state information sequence.

[0149] In some implementations, the target detection device based on channel state information further includes a model training unit, used for: Acquire the local preset channel state information sequence and the corresponding sample detection results, and preprocess the preset channel state information sequence to obtain the sample channel state information sequence. The sample channel state information sequence is input into the preset model to obtain the reconstructed channel state information sequence output by the reconstruction output header of the preset model; The reconstruction loss is constructed based on the difference between the reconstructed channel state information sequence and the sample channel state information sequence; Based on the reconstruction loss, the model parameters of the preset model are adjusted starting from the reconstruction output head using backpropagation, and iterative training is performed to obtain the trained candidate model. The sample channel state information sequence is input into the candidate model to obtain the prediction detection result output by the prediction output head; The prediction loss is constructed based on the difference between the predicted detection results and the sample detection results. The candidate model is then fine-tuned based on the prediction loss to obtain the trained target model.

[0150] In some implementations, the pre-defined model includes at least an embedded unified representation network, a shared backbone network, and a reconstructed output head, and a model training unit, and is further used for: Based on the reconstruction loss, the gradients of the reconstructed output head, the shared backbone network, and the embedded unified representation network are calculated sequentially according to the backpropagation method; The network parameters of the reconstructed output head, the shared backbone network, and the embedded unified representation network are adjusted according to the corresponding gradients, and iterative training is performed until convergence to obtain the trained candidate model.

[0151] As can be seen from the above, this application obtains the target channel state information sequence by preprocessing the original channel state information sequence to remove invalid interference noise from the channel state information, which is beneficial to improving the accuracy of subsequent target detection. Furthermore, the target channel state information sequence is mapped to a unified dimension through the trained target model, and the target detection results in the target scene are predicted based on the feature length dependency relationship of the unified dimension feature sequence. Thus, the target model has high adaptability and strong cross-environment generalization ability, and is suitable for target detection of channel state information in any input form. It has high applicability, reduces deployment costs, enhances the universality of target detection scenarios based on channel state information, meets the needs of large-scale and generalized practical applications, and improves the efficiency of target detection based on channel state information.

[0152] The specific implementation of each of the above units can be found in the previous embodiments, and will not be repeated here.

[0153] Figure 6To implement the structural block diagram of a terminal in this embodiment of the application, the terminal 110 includes: a radio frequency (RF) circuit 510, a memory 515, an input unit 530, a display unit 540, a sensor 550, an audio circuit 560, a wireless fidelity (WiFi) module 570, a processor 580, and a power supply 590, among other components. Those skilled in the art will understand that the terminal 110 structure shown in the figures does not constitute a limitation on a mobile phone or computer, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0154] The RF circuit 510 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 580; in addition, it transmits uplink data to the base station.

[0155] The memory 515 can be used to store software programs and modules. The processor 580 executes various functional applications and data processing of the terminal by running the software programs and modules stored in the memory 515.

[0156] The input unit 530 can be used to receive input numeric or character information, and to generate key signal inputs related to the terminal's settings and function control. Specifically, the input unit 530 may include a touch panel 531 and other input devices 532.

[0157] The display unit 540 can be used to display input or provided information, as well as various menus of the terminal. The display unit 540 may include a display panel 541.

[0158] Audio circuit 560, speaker 561, and microphone 562 provide an audio interface.

[0159] In this embodiment, the processor 580 included in the terminal 110 can execute the target detection method based on channel state information in the previous embodiment.

[0160] The terminal 110 in this application embodiment includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. This application embodiment can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0161] Figure 7This is a partial structural block diagram of a server embodying this embodiment of the application. The server 120 can vary significantly due to different configurations or performance characteristics, and may include one or more Central Processing Units (CPUs) 622 (e.g., one or more processors) and a memory 632, and one or more storage media 620 (e.g., one or more mass storage devices) for storing application programs 642 or data 644. The memory 632 and storage media 620 may be temporary or persistent storage. The program stored in the storage media 620 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server 120. Furthermore, the CPU 622 may be configured to communicate with the storage media 620 and execute the series of instruction operations in the storage media 620 on the server 120.

[0162] Server 120 may also include one or more power supplies 626, one or more wired or wireless network interfaces 650, one or more input / output interfaces 658, and / or one or more operating systems 641, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0163] The central processing unit 622 in server 120 can be used to execute the target detection method based on channel state information according to the embodiments of this application.

[0164] This application also provides a computer-readable storage medium for storing program code for executing the target detection method based on channel state information in the foregoing embodiments.

[0165] This application also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the target detection method based on channel state information described above.

[0166] Furthermore, the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, apparatus, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are expressly listed, but may include other steps or units that are not expressly listed or that are inherent to such process, method, product or device.

[0167] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0168] It should be understood that in the description of the embodiments of this application, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed between them may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] It should also be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0174] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0175] The above is a detailed description of the embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A target detection method based on channel state information, characterized in that, include: Obtain the original channel state information sequence corresponding to the wireless transmitting device in the target scene; The original channel state information sequence is preprocessed to obtain the processed target channel state information sequence; The target channel state information sequence is input into the trained target model. The target model maps and represents the target channel state information sequence to obtain a unified dimension feature sequence. The target detection result is output based on the time-series features corresponding to the unified dimension feature sequence. The target model is obtained by training a preset model to minimize the loss of the joint reconstruction loss and the prediction loss. The preset model is based on the predicted detection result and the reconstructed channel state information sequence output by the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

2. The method according to claim 1, characterized in that, The target model includes an embedded unified representation network, which contains a first convolutional layer and a second convolutional layer. The process of mapping the target channel state information sequence using the target model to obtain a unified dimension feature sequence includes: Through the first convolutional layer in the target model, local time variation features are extracted for the target channel state information sequence in the time dimension, and frequency domain variation features are extracted for the channel state information corresponding to adjacent subcarriers in the target channel state information sequence in the frequency domain dimension. The local temporal variation features and the frequency domain variation features are fused and mapped through the second convolutional layer in the target model to generate a unified dimension feature sequence.

3. The method according to claim 2, characterized in that, The step of extracting local temporal variation features of the target channel state information sequence in the time dimension through the first convolutional layer in the target model, and extracting frequency domain variation features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence in the frequency domain dimension, includes: By using the first convolutional layer in the target model, a one-dimensional convolution operation is performed on the target channel state information sequence along the time dimension to obtain the local time variation features of the target channel state information sequence in continuous time segments. The first convolutional layer in the target model extracts features of the channel state information corresponding to adjacent subcarriers in the target channel state information sequence along the frequency domain dimension to obtain the corresponding channel state features. The similarity change features among the channel state features are then aggregated to obtain the frequency domain change features of the target channel state information sequence.

4. The method according to claim 1, characterized in that, The target model includes a shared backbone network, and the step of outputting the target detection result based on the temporal features corresponding to the unified dimension feature sequence includes: By using the shared backbone network, the temporal relationship features corresponding to the unified dimension feature sequence are obtained according to the attention mechanism to obtain general temporal features; The target detection result is output based on the general temporal features through the prediction output head in the target model.

5. The method according to claim 4, characterized in that, The step of obtaining the temporal relationship features corresponding to the unified dimension feature sequence through the shared backbone network according to the attention mechanism to obtain general temporal features includes: Through the shared backbone network, the temporal dependency modeling of the unified dimension feature sequence is performed according to the multi-head self-attention mechanism to obtain the correlation strength and dependency relationship features between features at different time positions in the unified dimension feature sequence. Based on the correlation strength and dependency features, the corresponding attention weights are calculated, and the features at different time positions are weighted and fused according to the attention weights to obtain fused features that contain global temporal change patterns. Based on the fusion features, long-distance temporal relationship features are extracted to obtain general temporal features.

6. The method according to any one of claims 1 to 5, characterized in that, The training process of the target model is as follows: Obtain a local preset channel state information sequence and the corresponding sample detection results, and preprocess the preset channel state information sequence to obtain a sample channel state information sequence; The sample channel state information sequence is input into a preset model to obtain the reconstructed channel state information sequence output by the reconstruction output header of the preset model; The reconstruction loss is constructed based on the difference between the reconstructed channel state information sequence and the sample channel state information sequence; Based on the reconstruction loss, the model parameters of the preset model are adjusted starting from the reconstruction output head using backpropagation, and iterative training is performed to obtain the trained candidate model. The sample channel state information sequence is input into the candidate model to obtain the prediction detection result output by the prediction output head; A prediction loss is constructed based on the difference between the predicted detection result and the sample detection result, and the candidate model is fine-tuned based on the prediction loss to obtain the trained target model.

7. The method according to claim 6, characterized in that, The preset model includes at least an embedded unified representation network, a shared backbone network, and a reconstruction output head. Based on the reconstruction loss, the model parameters of the preset model are adjusted starting from the reconstruction output head using backpropagation, and iterative training is performed to obtain a trained candidate model, including: Based on the reconstruction loss, the gradients of the reconstruction output head, the shared backbone network, and the embedded unified representation network are calculated sequentially according to the backpropagation method; The network parameters of the reconstructed output head, the shared backbone network, and the embedded unified representation network are adjusted according to the corresponding gradients, and iterative training is performed until convergence to obtain the trained candidate model.

8. A target detection device based on channel state information, characterized in that, include: The acquisition unit is used to acquire the original channel state information sequence corresponding to the wireless transmitting device in the target scene; The preprocessing unit is used to preprocess the original channel state information sequence to obtain the processed target channel state information sequence. The input unit is used to input the target channel state information sequence into the trained target model, map the target channel state information sequence through the target model to obtain a unified dimension feature sequence, and output the target detection result according to the time-series features corresponding to the unified dimension feature sequence. The target model is obtained by training a preset model to minimize the loss of the joint reconstruction loss and the prediction loss. The preset model is based on the predicted detection result and the reconstructed channel state information sequence output by the sample channel state information sequence. The reconstruction loss is determined by the difference between the reconstructed channel state information sequence and the sample channel state information sequence, and the prediction loss is determined by the difference between the predicted detection result and the sample detection result.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the target detection method based on channel state information as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the target detection method based on channel state information as described in any one of claims 1 to 7.

Images