Integrated sensing and communication enabled by networked hybrid quantum-classical machine learning

Abstract

Devices with communication capabilities, such as commercial Wi-Fi devices, can be used in an Integrated Sensing and Communication (ISAC) system to collaboratively exchange data and monitor the environment. Such devices typically require diverse signal processing, such as machine learning inference, that requires high-power operation for real-time sensing and computing. The present invention provides a method for achieving energy-efficient computing by utilizing the capabilities of data communication to access distributed computing resources, including classical and quantum computers, over a network. The system and method are based on the recognition that computationally intensive processing is offloaded to a networked classical-quantum hybrid computing to build a dynamic computing graph. Some embodiments use automated classical-quantum machine learning, whose circuits and hyperparameters are automatically tuned through gradient or heuristic optimization for Wi-Fi indoor monitoring and human tracking. In some embodiments, the system and method can reduce power consumption and the number of trainable parameters by integrating classical and quantum neural networks. In some embodiments, signal processing, such as noise removal, filtering, and detection, is achieved using a classical-quantum hybrid processor on the network.

Description

[Technical field]

[0001] The present invention relates to machine learning and inference systems for sensing via signals for data communication, and more specifically to hybrid quantum-classical machine learning systems for wireless sensor networks, Wi-Fi® indoor monitoring, body area sensor networks, terahertz spectroscopy, coherent optical sensing, mmWave localization, driving navigation, and object tracking. [Background technology]

[0002] Communication devices such as mobile phones, Wi-Fi access points, and Internet of Things (IoT) appliances have become widespread almost everywhere in a normal living environment. Communication devices are primarily used to exchange data from one point to another by modulating signals such as electromagnetic waves through wired or wireless medium channels. To exchange data securely and reliably, various signal processing methods such as error control coding, data compression, pulse shaping, channel estimation, etc. are implemented.

[0003] In the past decade, communication devices with sensing capabilities have attracted tremendous attention as a new technology framework called integrated sensing and communications (ISAC). Since data communication relies on a channel link that changes with environmental changes such as the movement of surrounding objects, communication devices can essentially sense the environment state by estimating the channel during data communication without relying on other external sensors. For example, Wi-Fi sensing or wireless local area network (WLAN) sensing uses Wi-Fi access points or Wi-Fi mobile devices to realize indoor monitoring in addition to data communication. Several research groups have focused on WLAN sensing by taking better advantage of 802.11 standard technology for novel industrial and commercial applications. Some WLAN sensing frameworks use either channel state information (CSI) from the physical (PHY) layer or received signal strength indicator (RSSI) measurements from the medium access control (MAC) layer. RSSI measurements suffer from measurement instability and coarse granularity of channel information, which leads to limited sensing accuracy. CSI measurements have high sensing granularity but require access to a PHY layer interface and high computing power to process large amounts of subcarrier data.

[0004] The technical field of indoor localization, as one of the WLAN sensing applications, provides a way to locate an object in an enclosed area. The object may be a device that transmits and / or receives signals to and from some other devices, or it may be an entity that does not have such capabilities. Some indoor localization industrial applications, such as locating objects in hospitals, warehouses, shopping malls, factories, to name a few, require accurate indoor localization. Some traditional indoor localization approaches require dedicated hardware to be installed in the indoor area. However, these types of traditional approaches are undesirable because expensive dedicated hardware is required for indoor localization systems. One example of this approach is the Ultra Wideband (UWB) wireless localization system, which is expensive and used as a last choice for communities. Other examples include systems based on Light Detection and Ranging (LIDAR), radar, or ultrasound, which require high installation and maintenance costs similar to UWB localization systems. Therefore, ISAC, which allows accurate sensing without the need for external sensors, is a viable solution for various applications.

[0005] However, the dependency and correlation between the channel state and the target sensing factor is often weak, making it difficult to accurately estimate the sensing value from the communication signal. To solve this problem, machine learning and deep learning methods based on deep neural networks (DNNs) are used in ISAC. For example, WLAN sensing first collects a large amount of data during communication, and this data is used as a fingerprinting database to train a DNN model to predict the sensing factor in an online manner. DNNs can be used for image classification, speech recognition, computational sensing, compressed sensing, data analysis, feature extraction, signal processing, and artificial intelligence. There are many DNN architectures, such as multi-layer perceptrons, recurrent networks, convolutional networks, transformer networks, attention networks, long short-term memory, generative adversarial networks, autoencoders, U-shaped networks, residual networks, reversible networks, loop networks, clique networks, implicit layers, and their variants. Although the high performance of DNNs has been demonstrated in many practical systems, training and deploying DNNs on resource-limited hardware with small memory and processing power for real-time applications has remained a challenge. Part of the reason is that typical DNNs have a large number of trainable parameters, such as weights and biases, and many artificial neurons throughout deep hidden layers, which also require a large number of arithmetic multiplications and floating-point operations. In addition, they require large amounts of training data and training iterations, which generally leads to high power requirements.

[0006] In addition to standard DNNs, the emerging framework of quantum neural networks (QNNs) leveraging quantum processing units (QPUs) has been introduced as an alternative computational paradigm for the future era of quantum supremacy. QNNs can solve some problems of classical DNN approaches through exponentially large parallel computations with fewer parameters. Quantum computers have the potential to achieve computationally efficient signal processing compared to classical digital computers, not only in terms of execution time but also in terms of energy consumption, by exploiting quantum mechanisms, e.g., superposition and entanglement. Nevertheless, quantum processors are difficult to embed in mobile devices, as they require careful maintenance.

[0007] Quantum machine learning (QML) is considered a potential driver in sixth generation (6G) applications. In the past few years, several companies, including IBM®, Google®, and Honeywell®, have produced commercial quantum computers. As an example, IBM announced a 433-qubit QPU that will be publicly available via its cloud service in 2022, and plans to produce a 1121-qubit QPU by 2023. Although several groups have reported achieving quantum supremacy for specific problems, quantum advantage has not been fully verified for general problems. It is not in the distant future that noisy intermediate-scale quantum (NISQ) computers will be widely used for a variety of practical applications. Quantum-enabled algorithms for wireless communication systems have been investigated, but most existing work assumes fault-tolerant QPUs, which are beyond the capabilities of near-term NISQ devices.

[0008] In addition to machine learning inference, communication devices require a variety of signal processing methods that require high power operation for real-time sensing and computing. Therefore, there is a need to develop ISAC systems and methods that enable low-cost, low-power, and reduced hardware deployment. Summary of the Invention

[0009] The present invention is based on the recognition that communication devices can access external computing resources without implementing them on-premise, and the hybrid use of quantum computing processors and classical computing processors provides high performance, low power and flexible signal processing. The present invention is based on the networked use of both classical deep neural networks (DNNs) and quantum neural networks (QNNs) with proper load balancing by distributing operational resources on the network, and the sensing equipment has data access points that are inherently easy to communicate with remote servers to execute any heterogeneous computing graphs through DNNs and QNNs. QNNs include quantum convolution, quantum graph neural networks (QGNNs), tensor networks, quantum autoencoders, quantum generative adversarial networks, quantum reservoir networks, quantum implicit layers, etc.

[0010] Thus, the present invention is based on the recognition that an Integrated Sensing and Communications (ISAC) has the capability to access remote quantum computing servers such as Amazon braket and IBM quantum. The present invention uses hybrid classical-quantum machine learning to exploit this capability for sensing and data analysis so that the ISAC system can be deployed with low cost and power overhead. In some embodiments, several Wi-Fi access points monitor the indoor / outdoor environment and surrounding users during communication, channel acquisition, and beam tracking. The Wi-Fi access points employ hybrid use of DNNs and QNNs on cloud or on-premise when available.

[0011] In some embodiments, the terahertz time-domain spectroscopy system, biosignal sensing system, and LIDAR / radar / sonar imaging system can access the server using both classical and quantum computers, and the measurement data is analyzed using networked classical-quantum hybrid machine learning for feature extraction. In some embodiments, the hybrid architecture on the computing graph in the networked DNN and QNN is adaptively tuned through an automated machine learning (AutoML) framework with either deep reinforcement learning, Bayesian optimization (BO), or metaheuristic optimization to design quantum ansatz, embedding, measurement circuit, specification, quantization, and hyper-network architecture to configure the variational quantum circuit (VQC). In some embodiments, signal processing for data communication and analysis, such as encoding, decoding, modulation, demodulation, equalization, detection, pulse shaping, compression, decompression, noise removal, compressed sensing, etc., is further offloaded towards the networked quantum-classical hybrid computer. Networked quantum-classical hybrid machine learning has the non-trivial advantage that communication capabilities are utilized to simultaneously realize communication and sensing in conjunction with distributed computing load balancing for classical and quantum computers, which can significantly improve the performance of signal processing and inference without additional hardware embedding. In particular, managing multiple CPUs and QPUs simultaneously is difficult in conventional techniques, because the potential number of distributed patterns of operations going to multiple CPUs and QPUs is huge, and the accuracy, latency, power requirements, and fidelity of different CPUs and QPUs vary widely. The present invention provides a method for controlling a computing graph over a network to manage the appropriate load balance of individual CPUs and QPUs such that the total cost and efficiency are well maintained. In addition, the present invention provides a method for integrating QNNs into DNNs in a flexible manner by utilizing the capabilities of distributed computing that can be seamlessly reconfigured over a network.For example, some parts of signal processing, such as convolutional layers and activations in DNNs, are adaptively distributed across different central computing units (CPUs), graphics computing units (GPUs), field programmable gate arrays (FPGAs), and quantum processing units (QPUs).

[0012] Some embodiments of the present invention provide a system deployed for integrated computing, sensing, and communication, the system having at least one communication link, at least one classical computing processor configured with a set of trainable parameters, at least one quantum computing processor configured with a variation circuit according to a set of variational parameters, and at least one memory bank coupled to the at least one classical computing processor and the at least one quantum computing processor via the at least one communication link, the at least one memory bank storing instructions implementing a set of classical-quantum hybrid computing methods for data communication, machine learning, and environmental sensing. As an example, the trainable parameters are weights and biases, and a large number of artificial neurons are throughout a deep hidden layer requiring a large number of arithmetic multiplications and floating point operations. The instructions, when executed by at least one classical computing processor and at least one quantum computing processor, perform steps including: causing the at least one classical processor to exchange data through at least one communication link to adjust the set of trainable parameters and the set of variational parameters and update the data in the at least one memory bank according to the set of computing methods, and causing the at least one quantum processor to perform a measurement of the quantum state taking into account the set of variational parameters and the data in the at least one memory bank according to the set of computing methods. By way of example, the set of computing methods include an adaptive filtering algorithm, a gradient descent algorithm, a feature extraction algorithm, a classification algorithm, a regression algorithm, a prediction algorithm, a compressive sensing algorithm, an estimation algorithm, an inference algorithm, a deep neural network, machine learning, a denoising algorithm, encoding, decoding, modulation, demodulation, or equalization.

[0013] According to some embodiments of the present invention, there is provided a computer-implemented method for signal processing, communication, and sensing performed by at least one classical computing processor and at least one quantum computing processor. The at least one classical computing processor and the at least one quantum computing processor are coupled through at least one communication link to a memory bank storing instructions implementing a set of classical-quantum hybrid computing methods. The instructions perform the following steps, including exchanging a set of data between the at least one memory bank and the at least one classical computing processor and the at least one quantum computing processor through at least one communication link, distributing a set of sub-instructions to the at least one classical computing processor and the at least one quantum computing processor, adjusting a set of variational parameters for the at least one quantum computing processor, adjusting a set of trainable parameters for the at least one classical computing processor, and modifying the set of data according to the set of sub-instructions, the set of variational parameters, and the set of trainable parameters.

[0014] Thus, embodiments can achieve low-power, reduced-hardware signal processing for data communication and sensing with accurate DNN and QNN inference by using networked hybrid quantum-classical computing over communications, which can be used for real-time applications requiring low-power and resource-constrained deployment.

[0015] The accompanying drawings, which are included to provide a further understanding of the invention, illustrate embodiments of the invention and, together with the description, explain the principles of the invention. [Brief description of the drawings]

[0016] [Figure 1A] FIG. 1 illustrates an exemplary system for networked hybrid quantum-classical machine learning enabled Wi-Fi human surveillance according to some embodiments. [Figure 1B] FIG. 1 illustrates an exemplary deployment of a networked hybrid quantum-classical machine learning enabled Wi-Fi indoor surveillance system according to some embodiments. [Figure 2A] FIG. 1 illustrates an exemplary scenario for a Wi-Fi indoor surveillance system according to some embodiments. [Figure 2B] FIG. 1 illustrates an exemplary scenario for a Wi-Fi indoor surveillance system according to some embodiments. [Figure 3A] FIG. 1 illustrates an exemplary transmitter-receiver system for integrated communication and sensing according to some embodiments. [Figure 3B] FIG. 1 illustrates an exemplary transmitter-receiver system for integrated communication and sensing according to some embodiments. [Figure 4A] 1 illustrates an exemplary diagram of communication channels for collaborative data exchange and sensing according to some embodiments. [Figure 4B] 1 illustrates an exemplary diagram of communication channels for collaborative data exchange and sensing according to some embodiments. [Figure 4C] 1 illustrates an exemplary diagram of communication channels for collaborative data exchange and sensing according to some embodiments. [Diagram 5] FIG. 1 shows an exemplary schematic diagram of a quantum neural network for use based on variational quantum circuits for Wi-Fi sensing, according to some embodiments. [Figure 6A] FIG. 2 illustrates exemplary types (a)-(i) of exemplary Ansatz-based variational quantum circuits according to some embodiments. [Figure 6B]FIG. 1 illustrates an exemplary method for designing QNN Ansatz and hyper-parameters through an AutoML framework, according to some embodiments. [Figure 6C] FIG. 13 illustrates exemplary performance plots of different inference methods compared to a hybrid DNN·QNN designed by Auto-Ansatz for a Wi-Fi sensing system, according to some embodiments. [Figure 7] FIG. 1 illustrates an exemplary block diagram of a computing graph representing signal processing distributed across multiple classical computing processors and quantum computing processors in networked quantum-classical hybrid computing, according to some embodiments. [Figure 8A] FIG. 1 illustrates an exemplary system for a networked hybrid quantum-classical machine learning enabled IoT wireless monitoring system according to some embodiments. [Figure 8B] FIG. 1 illustrates an exemplary block diagram of networked hybrid quantum-classical computing according to some embodiments. [Figure 8C] FIG. 1 illustrates an exemplary block diagram of networked hybrid quantum-classical computing according to some embodiments. [Figure 8D] FIG. 1 illustrates an exemplary block diagram of networked hybrid quantum-classical computing according to some embodiments. [Figure 9A] FIG. 1 illustrates an exemplary system for integrated communications and sensing enabled by networked hybrid quantum-classical computing, according to some embodiments. [Figure 9B] FIG. 1 illustrates an exemplary system for integrated communications and sensing enabled by networked hybrid quantum-classical computing, according to some embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Various embodiments of the present invention will now be described with reference to the drawings. Please note that the drawings are not drawn to scale, and elements of similar structure or function are indicated by similar reference numerals throughout the drawings. Please also note that the drawings are intended only to facilitate the description of certain embodiments of the present invention. They are not intended to exhaustively describe the present invention, nor are they intended to limit the scope of the present invention. In addition, aspects described in connection with a specific embodiment of the present invention are not necessarily limited to that embodiment, and may be implemented in any other embodiment of the present invention.

[0018] Some embodiments of the present disclosure provide systems and methods for integrated sensing and communication (ISAC) enabled by networked quantum-classical hybrid computing to improve signal processing and machine learning performance while reducing power consumption and hardware resources in communication devices including mobile phones, Wi-Fi access points, and Internet of Things (IoT) appliances. Networked quantum-classical hybrid computing is implemented with a computing graph over a data communication network using at least one classical computing processor and at least one quantum computing processor, and load balances across different processors to meet certain requirements including power efficiency, computation accuracy, and latency to completion. Some other embodiments include, but are not limited to, inference and prediction for digital pre-distortion, channel equalization, channel estimation, nonlinear turbo equalization, speech recognition, image processing, biosignal sensing, LIDAR / radar / sonar / terahertz imaging, compressive sensing, deep learning, deep reinforcement learning, and the like. Wi-Fi Sensing for Indoor Surveillance

[0019] In the case of Wi-Fi sensing, an access point in the 802.11 standard can utilize the capabilities of Wi-Fi devices to easily access distributed computers to collect beam scanning measurements that can be associated with classes of human gestures, object behaviors, and environmental changes as fingerprinting data for learning deep neural network (DNN) and quantum neural network (QNN) models. Specifically, the present invention uses at least one classical computing processor and at least one quantum computing processor on the network. Thus, some computationally heavy signal processing can be offloaded towards an external central processing unit (CPU) and quantum processing unit (QPU). In addition, the use of quantum computing devices that utilize entanglement and superposition can improve computational efficiency to exceed CPUs. Depending on the type of operation, the present disclosure assigns parts of the signal processing diagram to distributed CPUs and QPUs with appropriate load balancing for real-time applications. Figure 1 shows a Wi-Fi human monitoring system enabled by networked quantum-classical hybrid machine learning. This is an extension of the Wi-Fi sensing system that leverages data communication signals to sense the surrounding environment by further utilizing communication capabilities to access distributed CPUs and QPUs.

[0020] FIG. 1A illustrates an exemplary Wi-Fi sensing system according to the present invention. The exemplary configuration uses several Wi-Fi stations 13 and 17 in an indoor environment or in a certain outdoor area. This embodiment of the Wi-Fi sensing system enables data communication via the Wi-Fi stations as well as environmental sensing, including but not limited to human identity recognition, object recognition, human / object counting, human activity / motion detection, human posture / location recognition, patient vital signs detection, and the like. For example, as shown in FIG. 1A, a subject performs several different postures including different gestures 15 such as "sit", "stand with left arm up", and the like. For each posture, the communication device records training and test data for different periods of time with sufficient time intervals between successive sessions. The data can be expanded to improve Wi-Fi sensing performance for some embodiments.

[0021] By employing networked classical and quantum machine learning, the Wi-Fi sensing system can efficiently predict human posture using Wi-Fi signals without relying on other sensing modalities such as cameras. While the Wi-Fi stations 13 and 17 establish data access with other wireless devices, the Wi-Fi stations acquire measured wireless signals such as beam signal-to-noise ratio (SNR) 14 and channel state information (CSI). The measured wireless signals 14 are further used to predict environmental changes such as human posture 15. The prediction is based on networked machine learning methods executed on distributed classical computers 12 and quantum computers 11 through data links 16 and 18 for accessing the cloud network 10. The estimated environmental changes are sent back to the Wi-Fi stations.

[0022] Since ultra-granular mmWave channel state information (CSI) is typically not accessible from commercial off-the-shelf (COTS) devices without additional overhead, some embodiments use medium-grained Wi-Fi measurements in the beam angle domain-beam signal-to-noise ratio (SNR) generated from a beam training (also known as beam alignment) phase. For each probing beam pattern (also known as beam sector), the beam SNR is collected by the 802.11ad device as a measure of beam quality. Such beam training is performed periodically to adapt the beam sector to environmental changes.

[0023] For example, the system can collect beam SNR at 60 GHz using an 802.11ad compliant router. The router supports single-stream communication using analog beamforming on a multi-element planar array. From one beam training, one Wi-Fi station can collect several beam SNRs across discrete transmit beam patterns. The measured beam SNR is shared with a networked CPU and QPU via an Ethernet cable or different communication link to train DNNs and QNNs (e.g., through IBM Quantum Cloud Services). The exemplary system can be deployed in a standard indoor environment setting.

[0024] The hybrid use of distributed CPUs and QPUs allows flexible control of computation load distribution in the present invention. Since there are various specifications of CPUs and QPUs in the cloud and the size of possible methods is huge, it is difficult to assign specific operations to different processors. For example, IBM QPUs have chips of different sizes corresponding to the number of quantum bits (qubits) ranging from 5 to 433. The fidelity, coherent time, quantum volume, and required power consumption also vary widely. In addition, the availability of cloud CPUs and QPUs changes over time depending on the demand. The present invention uses a controller to assign specific operations to distributed CPUs and QPUs depending on these factors, and the total resource budget is kept at a target value depending on the acceptable accuracy error. In addition, the QNN architecture is automatically designed by using Bayesian optimization, reinforcement learning, or metaheuristic optimization methods to explore different quantum ansatz to maximize inference accuracy while minimizing circuit size. The compact QNN architecture achieves power-efficient inference. Quantum Machine Learning (QML) and Transfer Learning

[0025] A key challenge in Wi-Fi sensing is that Wi-Fi measurements or the surrounding environment may change over a measurement session, and these changes reduce the efficiency of machine learning models due to domain shifts. Some embodiments use a networked hybrid quantum-classical transfer learning framework to address the domain shift problem.

[0026] The present invention is based on the recognition that the Quantum Machine Learning (QML) framework is suitable for Wi-Fi sensing systems and the development of QPU is growing rapidly. Some modern DNN methods have already migrated to the quantum domain, such as quantum convolutional layers, quantum autoencoders, quantum graph neural networks (GNNs), and quantum generative adversarial networks (GANs). More importantly, QML is more suitable for ISAC and Wi-Fi sensing, since cloud quantum servers such as IBM Quantum and Amazon braket are easily accessible without additional hardware.

[0027] Similar to DNNs, it has been proven that QNNs hold the universal approximation property (UAP). Thus, by increasing the number of qubits and quantum entanglement layers, they can enjoy the performance of prior art DNNs. In addition, quantum circuits can be analytically distinguished, which allows for stochastic gradient descent (GDS) optimization of QNNs. Nevertheless, QNNs suffer from a gradient vanishing problem called the barren plateau. To mitigate this problem, the simplified two-design (S2D) ansatz was proposed to realize a shallow entangler for nearly any unit approximation. As classical deep learning has become extremely energy intensive, it is important to offload heavy signal processing from the CPU to the QPU for future sustainable computing.

[0028] 1B illustrates various other scenarios for the ISAC system in some embodiments, such as intruder detection for security surveillance, indoor localization for moving people, posture recognition, fall detection for elderly surveillance, motion detection, occupancy monitoring, etc.

[0029] 2A illustrates exemplary scenarios of posture recognition, occupancy detection, and localization in Wi-Fi sensing, and FIG. 2B illustrates exemplary scenarios of security monitoring, energy management, elderly care, remote operation, gesture recognition, and bio-signal processing, according to some embodiments.

[0030] 3A and 3B show an exemplary communication channel that depends on environmental conditions, which enables environmental sensing based on communication channel measurements without relying on external sensors. A communication device 101 uses multiple antenna elements 203 to transmit radio waves through a transmitter 201 to form a beam 205 toward the channel. Another communication device 102 uses multiple antenna elements 204 to receive radio waves through a receiver 202 to form a beam 206. This exemplary system uses the beam information and channel measurements to enable environmental sensing as well as the primary function of data communication. The transmitter 101 may be, for example, a handheld phone and the receiver 102 may be a Wi-Fi access point. The Wi-Fi access point receives a signal waveform 1701 from the channel, which is modified by physical layer operations 1703 such as equalization and demodulation. The waveform is further modified by medium access control (MAC) operations 1705 such as packet synchronization and scheduling. At the physical layer, the full CSI information can be used for environmental sensing. At the MAC layer, partial CSI 1710 or beam SNR can be used instead. Environmental sensing includes indoor localization 1707 of the handheld phone 101, for example. In particular, radio waves from multiple antennas in the millimeter wave (mmWave) band have rich spatial characteristics, as shown in Figures 4A, 4B, and 4C. Through Wi-Fi stations 41, fine-grained channel state information (CSI) 42 and medium-grained beam signal-to-noise ratio (SNR) 43 across the subcarrier spectrum can be jointly used in some embodiments to complement the physical characteristics of different radio bands.

[0031] FIG. 5 shows an exemplary QNN model used for Wi-Fi sensing with S2D Ansatz consisting of Pauli-Y rotation and staggered controlled Z entanglers. This Ansatz is a variational quantum circuit (VQC) 50 based on a simplified variant of a 2-design circuit, whose statistical properties are identical to the ensemble random unitary in terms of the Haar measure up to the first two moments. In the n-qubit variational quantum circuit 50, there are 2(n-1)L variational parameters {θ} 58 on the L-layer S2D Ansatz. The VQC has a quantum state preparation module, a state evolution module, and a state measurement module. The beam SNR measurement data 14 is first given to the VQC through a preprocessing layer 51, and the preprocessed data is embedded in the VQC through an angle embedding layer 52. The embedding layer prepares the quantum state. This includes angle embedding, amplitude embedding, quantum approximation optimization algorithm embedding, IQP embedding, Mottonen state preparation, etc. After the state preparation module, a state evolution module is executed. As the state evolution module, the entanglement layers 53, 54 are repeated L times. After the entanglement layers, a state measurement module 55 is used to convert the quantum state to classical data. The quantum measurement module includes expectation, probability, sample, variance, etc. For quantum state preparation, angle embedding or amplitude embedding is used in some embodiments. The quantum state evolution module can use different Ansatz as shown in FIG. 6, for example, Quantum Approximation Optimization Algorithm (QAOA) Ansatz, Tree Tensor Network (TTN) Ansatz, Matrix Product State (MPS) Ansatz, Multi-scale Entanglement Renormalization Ansatz (MERA), Strong Entanglement Layer Ansatz, Basic Entangler Layer Ansatz, Random Layer Ansatz, Continuous Variable (CV) Neural Network Layer, etc. After the quantum measurement 55, the measured data is post-processed by an output layer 56, which provides a score for prediction such as human pose 59. The predicted classes are then quantified by a loss function, such as cross-entropy loss 57, to update the quantum variational parameters 58 via gradient methods, metaheuristic optimization methods, or variants thereof.

[0032] In some embodiments, to provide multi-dimensional beam SNR, an input linear layer is used in the CPU to initialize quantum states in the QPU for the rotation angles of the Pauli-Y gates. Multi-class pose estimation is provided by quantum measurements on the Hamiltonian observables of the Pauli-Z operations, followed by an output layer that aligns the dimensions in some embodiments. The variational and trainable parameters for the input / output layers are optimized by adaptive momentum gradient methods to minimize the softmax cross entropy loss. QNNs are not necessarily better than DNNs in prediction accuracy, but they can be used in parallel with a small number of quantum gates to achieve high accuracy. n Computational efficiency can be improved by manipulating individual quantum states.

[0033] The best QNN hyperparameters (e.g., qubit size n and layer size L) are highly dependent on the dataset, so a significant amount of manual effort is typically required to tune them. In addition, there are many different types of quantum ansatz in the literature to explore. The present invention provides a method to automatically tune the hyperparameters of a networked DNN-QNN hybrid architecture by exploring various ansatz options suitable for distributed CPUs and QPUs. Figure 6A shows some possible ansatz: angle embedding, IQP embedding, tensor networks such as QAOA, TTN, MPS and MERA, basic entanglers, strong entanglement layers, and random layers. Some embodiments use an automated machine learning framework (AutoML) to optimize the QNN ansatz and hyperparameters.

[0034] FIG. 6B illustrates an exemplary method for designing QNN Ansatz and hyperparameters through an AutoML framework, according to some embodiments. The automated design method of QNN Ansatz and hyperparameters searches for an optimal QNN circuit by using training data 60, which is further used to train candidate QNN circuits 61. Using validation data 62, the candidate QNN circuits are profiled, and a pruning mechanism 63 can stop the training operation if the candidate profile is unlikely. If pruning is not required, the profile information is used to determine how to sample the next candidate 64. According to the sampled candidate values, a new QNN circuit is constructed (65) and retrained. For example, some embodiments use Bayesian optimization (BO), based on a tree-Parzen estimator (TPE) and hyperband pruning, for efficient exploration of different hyperparameters. This automatically explores a diverse set of quantum Ansatz, qubit sizes, and layer sizes as well as learning rates, in some embodiments, without requiring human effort in Ansatz / hyperparameter tuning.

[0035] The present invention is based on the recognition that designing QNN Ansatz is not straightforward due to the Baren Plateau problem, in which the gradient variance of a randomly selected QNN circuit decays exponentially. The present invention is also based on the recognition that QNNs can be more compact than DNNs because an exponentially large state space can be expressed in terms of the number of qubits. The method and system of the present invention provide a method for achieving compact and power-efficient QNN design by simultaneously considering the trainability and size of the QNN Ansatz. Specifically, to address the Baren Plateau problem, shallower and smaller QNN Ansatz are explored by BO and RL optimization methods. In addition, the networked hybrid quantum computing process of the present invention controls the allocation of QNN and DNN operations to multiple QPUs and CPUs such that the total resource budget is maintained within a target range. The use of a QPU-CPU hybrid on a network can provide high flexibility to adjust the workload according to demand and acceptable accuracy. In addition, the QNN layer can be seamlessly integrated with the DNN layer by using the power of the networked CPUs and QPUs. In some embodiments, the AutoML framework of FIG. 6B is further extended to automatically design the allocation of operations to multiple CPUs and QPUs in a network, in addition to quantum ansatz.

[0036] FIG. 6C shows an exemplary performance plot of different inference methods compared to a hybrid DNN·QNN designed by automated Ansatz for Wi-Fi sensing systems according to some embodiments. This illustrates the advantages of the present invention in an experiment. The experimental setup uses two Wi-Fi stations, one in front of the subject and the other behind the subject. Both stations are placed on a stand with a height of 1.2 meters, with a distance of about 2 meters. The subject is asked to perform a total of 8 postures. For each posture, 7 independent measurement sessions are recorded with different durations. The first 4 sessions are used as training data, and the last 3 sessions are used as test data. The total number of measurement samples is 42915 and 1040 in training and testing, respectively. We collect the 60 GHz beam SNR using an 802.11ad compliant TP-Link Talon AD7200 router. This router supports single-stream communication using analog beamforming on a 32-element planar array. From one beam training, one Wi-Fi station can collect 36 beam SNRs across discrete transmit beam patterns. AutoQML explores 2000 trials of hyperparameter tuning, and each model is trained for up to 100 epochs using adaptive momentum gradient methods. Bayesian optimization uses categorical sampling of different Ansatz (Figure 6A (c)-(i)) and different embedding methods (Figure 6A (a) and (b)). The number of qubits and the number of entangler layers are also sampled from the ranges of 5-15 and 1-5, respectively. In addition, an initial learning rate of 10 -3 ~10 -1The learning rate is adaptively decreased with a plateau of 0.5-fold training loss over 10 epochs of patients. Figure 6C shows the test accuracy as a function of the number of labeled training samples. Performance can exceed 90% accuracy for DNN, QNN, and SVM when a sufficient amount of labeled data is available. It is confirmed that the small-scale QNN designed by AutoQML can improve the baseline QNN model and achieve slightly better state-of-the-art performance than the large-scale DNN. More specifically, the QNN model consists of 12 variational parameters, while the DNN model requires 3500 trainable parameters. Using QNNs can significantly reduce the computational complexity for Wi-Fi sensing and human monitoring systems. According to functional analysis of variance (fANOVA) scores, it is found that the choice of quantum Ansatz is the most important hyperparameter and has a larger impact than the learning rate parameter. Compressive Sensing and Communications

[0037] The present invention provides a method for distributing operations in a signal processing method for data communication and sensing across multiple CPUs and QPUs in a network. FIG. 7 shows an exemplary diagram of signal processing, in which some of the computation modules, such as quantization 71, normalization 72, fast Fourier transform (FFT) 73, QNN 74, cosine 76, convolutional neural network (CNN) 75, QNN 74, average 77, and comparison 78, are assigned to different CPUs and QPUs. The signal processing diagram determines a computing graph across the networked quantum-classical hybrid computing. In some embodiments, the trainable parameters of the CPUs 71, 73, 75, 77 and the variational parameters of the QPUs 72, 74, 76, 78 can be adjusted by backpropagating loss gradients along the computing graph. The present invention provides a method for controlling the allocation of operations across distributed CPUs and QPUs according to processor specifications, such as required power consumption, latency to complete operations, and accuracy of operations. The controller explores the optimal combination that enables power-efficient quantum and classical computing through the use of Bayesian optimization, reinforcement learning, or metaheuristic optimization methods.

[0038] FIG. 8A also shows an exemplary system used for multi-user multi-device communication where grant-free Internet of Things (IoT) devices are connected to a communication device. In this embodiment, multiple IoT devices 27 access a wireless base station 25 simultaneously. The base station 25 offloads time-consuming signal processing to a distributed CPU 22 and QPU 21 via communication link 26. The CPU 22 and QPU 21 perform cooperative operations by exchanging messages through links 23 and 24, enabling channel estimation and user identification for grant-free IoT communication with reduced computational power. Using networked quantum-classical hybrid computing 20, many devices can be jointly accessed with quantum multi-user detection, turbo equalization, and compressed sensing.

[0039] In some embodiments, such hybrid quantum-classical computing is used in grant-free wireless access systems with an indefinite number of devices communicating with a quantum processor-enabled base station. Such wireless networks are subject to the Compressive Sensing (CS) problem to estimate the presence of devices, e.g., Narrowband Internet of Things (NB-IoT) systems.

[0040] Although most CS methods assume i.d. device activity, a typical grant-free network may face correlated user activity due to a shared medium environment. One of the prior art CS techniques is the well-known Orthogonal Approximation Message Passing (OAMP) algorithm. The OAMP algorithm can usually achieve better convergence performance than traditional CS techniques such as Fast Iterative Soft Thresholding Algorithm (FISTA) and standard AMP algorithm. The OAMP algorithm inherits a two-step iterative process: a decorrelated linear estimation (LE) step and a divergence-free nonlinear estimation (NLE) step. The OAMP alternating process maintains the orthogonality of the LE and NLE estimation errors and generally achieves better convergence performance for solving linear recovery problems.

[0041] The present invention provides a way to improve CS methods by using networked hybrid quantum-classical processing. Specifically, some embodiments integrate quantum circuits into the compressed sensing algorithm so that correlations between device activities can be exploited by VQC-based denoising in the NLE step.

[0042] 8B and 8C show an exemplary schematic diagram of such a VQC-based compressed sensing, in which the NLE step 820 uses four sub-processes: embedding 822, denoising 824, state preparation based on estimated error 825, and VQC 823. The LE estimation 821 is refined by a denoising function 824 in the NLE step using a VQC-based scaling operation. In this embodiment, the use of the QPU for VQC enables an efficient NLE step for compressed sensing, especially in the presence of unknown correlation in grant-free device communication. In another embodiment, turbo decoding is realized by a hybrid use of CPU and QPU in addition to compressed sensing. In this case, the VQC-based decoding is realized by the QPU in the loop of the LE step 811 and the NLE step 812. The turbo decoding 813 includes a QAOA-based decoding 814 and a log-likelihood ratio (LLR) calculation 815.

[0043] Instead of using traditional stochastic gradient descent methods to gradually tune trainable parameters in VQC, which requires a large number of information exchanges between classical computers and quantum processors, some embodiments use a learning-to-learn (L2L) method based on a reinforcement learning framework to find variational parameters in a small number of information exchanges. The L2L method can accelerate the optimization step while finding trainable / variational parameters in VQC. For example, in a scenario of learning parameters used in QAOA, the input of the classical neural network is the initialization and expectation over the cost Hamiltonian of QAOA. FIG. 8D shows an exemplary signal processing that utilizes a CPU 802 for a long short-term memory (LSTM) module and a QPU 801 for QAOA decoding. QAOA decoding is used to identify communication devices in an iterative approach, and multiple QAOA operations 804, 806, 808 are performed to refine the device identification. QAOA decoding obtains multiple variational parameters beta and gamma, which are optimized by multiple LSTM cells 805, 807, 809 for each QAOA decoding iteration. After hybrid use of the CPU and QPU for iterations 803, a final estimate of the device identity is provided. This L2L-based method allows for alternative optimization of the variational parameters without the need for gradient backpropagation. System Implementation

[0044] 9A is a diagram of some components of a processing system that can be used for some operations for one or all embodiments of the present disclosure, according to some embodiments of the present disclosure. The processing system 1000A is interconnected via a bus system 1066 that interconnects the components of the system. The system 1000A may include one or more central processing units ("processors") 1061. The processor 1061 may be a server computer, a third party computer, a personal computer (PC), a client computer, a user device, a tablet PC, a laptop computer, a mobile phone, a smartphone, a web device, a network router, a switch or bridge, a console, or any machine configured to execute instructions according to operations performed by a computing system. Additionally, the processing system 1000A may include a data storage 1062, a stored module 1063, and a set of instructions 1064.

[0045] The processing system 1000A may include a primary memory 1071, which may include other types of memory (not shown), such as databases, non-volatile memory machine-readable media, servers, etc., depending on the requirements of the user and the operating components. The primary memory may store the above components as well as instructions, applications, programs, modules, computer programs, all of which may be executed by one or more processors of the system.

[0046] With continued reference to FIG. 9A, the system 1000A may include a human machine interface (HMI) 1081, which is a user interface or dashboard that connects a person to a machine, system, or device. Other terms for HMI include man-machine interface (MMI), operator interface terminal (OIT), local operator interface (LOI), or operator terminal (OT). HMI and graphical user interface (GUI) are similar, and thus GUI is often utilized within HMI for visualization functions. HMI can be used to visually display data, track production time, trends, and tags, oversee KPIs, and monitor machine inputs and outputs, depending on the user's system and operational requirements. Some components that can be connected leading to the HMI or to the bus system 1066 may be a keyboard 1082, a display device 1083, a controller 1084, and an input / output device 1085, as well as other similar components related to the above terms known to function in a particular industry.

[0047] Other components of the system 1000A may include a controller interface 1086, a controller 1087, an external computer and computer system 1072, a network interface 1088, and at least one network 1089. The network interface 1088 may include a network adapter, which assists the processing system 1000A to manage data in the network 1089 with entities that may be external to the processing system 1000A (i.e., network equipment, etc.), typically, but not necessarily, configured through any known and / or convenient communication protocol supported by the processing system 1000A and the external entities. The network adapter 1089 may include a network adapter card, a wireless network interface card, a router, an access point, a wireless router, a switch, a multi-layer switch, a protocol converter, a gateway, a bridge, a bridge router, a hub, a digital media receiver, and / or a repeater. The network adapter 1089 may include a firewall that can govern and / or manage permissions to access data in a computer network and track levels of trust between different machines and / or applications. A firewall may be any number of modules having any combination of hardware and / or software components capable of enforcing a predefined set of access rights between a particular set of machines and applications, machines and / or applications, for example regulating traffic flow and resource sharing between these varying entities. A firewall may further manage and / or have access to access control lists that detail permissions, including, for example, access and operation rights of objects by individuals, machines and / or applications, and the circumstances under which permission rights are established.

[0048] 9A, the network 1089 may include a communication system that may wirelessly communicate or be wired to other components, such as equipment, machines, etc., that may be used in embodiments of the present disclosure. For example, the network 1089 may communicate with a cloud 1090, which may include data storage 1091, modules 1092, a set of instructions 1093, and a quantum processing unit (QPU) 1094, as well as other components and aspects known in the understanding of clouds within the cloud industry. It should be noted that the bus system 1066 may be a separate bus, a specific connection from one component or subsystem to another, or both connected by a bridge, adapter, or controller. The bus 1066 may be configured to relay data packets between components of the network equipment (i.e., network ports, other ports, etc.).

[0049] 9B is a block diagram illustrating some components of a computer that may be used with or in combination with some or all of the components of FIG. 9A, according to some embodiments of the present disclosure. Component 1000B may include a computer 1011 having a processor 1040, a computer readable memory 1012, storage 1058, and a user interface 1049 to a display device 1052 and a keyboard 1051, connected via a bus 1056. For example, a user interface 1064 that interacts with the processor 1040 and computer readable memory 1012 receives input from a user on a surface of the user interface 1057, a keyboard 1053, and retrieves and stores data in the computer readable memory 1012.

[0050] The computer 1011 may include a power supply 1054, which may optionally be located external to the computer 1011 depending on the application. A user input interface 1057 may be connected through the bus 1056, adapted to connect to a display device 1048, which may include, among other things, a computer monitor, a camera, a television, a projector, or a mobile device. A printer interface 1059 may also be connected through the bus 1056, adapted to connect to a printing device 1032, which may include, among other things, a liquid inkjet printer, a solid ink printer, a large scale commercial printer, a thermal printer, a UV printer, or a dye sublimation printer. A network interface controller (NIC) 1034 is adapted to connect to a network 1036 through the bus 1056, and may, among other things, render data or other data on a third party display device, a third party imaging device, and / or a third party printing device external to the computer 1011. The computer / processor 1011 may include a GPS 1001 connected to the bus 1056. Through the network 1036, at least one classical computing processor, such as a central processing unit (CPU) 1091 and a field programmable gate array (FPGA) 1092, and at least one quantum computing processor (QPU) 1093 are used to execute signal processing methods for data communication and sensing.

[0051] Continuing to refer to FIG. 9B, among other things, the data or other data can be transmitted through a communication channel of the network 1036 and / or stored in a storage system 1058 for storage and / or further processing. Furthermore, the time series data or other data can be received wirelessly or wired from a receiver 1046 (or an external receiver 1038) or transmitted wirelessly or wired via a transmitter 1047 (or an external transmitter 1039), both of which are connected through a bus 1056. The computer 1011 can be connected to an external sensing device 1044 and an external input / output device 1041 via an input interface 1008. The input interface 1008 can be connected to one or more input / output devices 1041, an external memory 1006, an external sensor 1004, which can be connected to a device 1002, such as a machine. The controller 1042 can be connected to the device 1043. Furthermore, other computers 1045 can be connected to the bus 1056. For example, the external sensing device 1044 may include sensors that collect data before, during, and after the collected time series data of the machine. The computer 1011 may be connected to other external computers 1042. The output interface 1009 may be used to output the processed data from the processor 1040. Note that the user interface 1049 in communication with the processor 1040 and the non-transitory computer readable storage medium 1012 obtains and stores area data in the non-transitory computer readable storage medium 1012 upon receiving input from a user through a surface 1052 of the user interface 1049. Further, a controller 1061 may be connected to the bus 1056 to control devices related to the embodiments of the system and method of the present disclosure. Additional Features

[0052] In an aspect, the at least one communication link comprises an interface, gateway, switching hub, router, access point, or variations thereof, and the at least one communication link is based on electrical wired communication, optical fiber communication, wireless wireless communication, free space optical communication, visible light communication, power line communication, magnetic coupling communication, acoustic communication, millimeter wave / terahertz communication, etc.

[0053] In an aspect, the at least one classical computing processor is based on a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a field programmable gate array (FPGA), a microprocessor (uP), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a complex programmable logic device (CPLD), or the like.

[0054] In an embodiment, the at least one quantum computing processor is based on superconducting quantum physics, trapped ion quantum physics, photonic quantum physics, neutral atoms, nuclear magnetic resonance, quantum annealing, boson sampling, quantum physics emulators, or variations thereof.

[0055] In an embodiment, the variational quantum circuit further includes a quantum state preparation module, a quantum state evolution module, and a quantum state measurement module. The quantum state preparation module is based on an angle embedding, an amplitude embedding, a basis embedding, a displacement embedding, a quantum approximate optimization algorithm (QAOA) embedding, a squeezing embedding, an instantaneous quantum polynomial (IQP) time embedding, a Mottonen state preparation Ansatz, or a variant thereof. The quantum state evolution module is based on a simplified two-design Ansatz, a QAOA Ansatz, a tree-tensor network (TTN) Ansatz, a matrix product state (MPS) Ansatz, a multi-scale entanglement renormalization Ansatz (MERA), a strongly entangled layer Ansatz, a basic entangler Ansatz, a random layer Ansatz, a continuous variable (CV) neural network layer, or a variant thereof. The quantum state measurement module is based on an expectation value, a sampling value, a variance value, a probability, a quantum state, a computation-based quantum density matrix, or a variant thereof.

[0056] In another aspect, the classical-quantum hybrid computing method is based on adaptive filtering algorithms, gradient descent algorithms, feature extraction algorithms, classification algorithms, regression algorithms, prediction algorithms, compressive sensing algorithms, estimation algorithms, inference algorithms, deep neural networks, machine learning, noise reduction algorithms, encoding, decoding, modulation, demodulation, equalization, or variations thereof for signal processing for data communication and sensing.

[0057] In an embodiment, trainable parameters for classical computing processors are adjusted based on stochastic gradient descent (SGD), elastic backpropagation, root mean square (RMS) propagation, Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm, adaptive momentum (Adam) optimization, metaheuristic optimization, simulated annealing (SA), genetic algorithms (GA), evolutionary algorithms such as simplex methods like Nelder-Mead, adaptive subgradient, adaptive delta, covariance matrix adaptation (CMA), evolutionary strategies (ES), swarm intelligence like ant colony optimization (ACO), particle swarm optimization (PSO), etc., deep learning, reinforcement learning (RL) like policy gradient methods, Bayesian optimization (BO) like Gaussian processes (GP), and variations of such.

[0058] In an embodiment, the variational parameters for the quantum processing device are adjusted based on SGD, parameter shift rule, adjoint method, finite difference, policy gradient, elastic backpropagation, RMS propagation, BFGS algorithm, Adam, simplex method, adaptive subgradient, adaptive delta, metaheuristic optimization, SA, GA, evolutionary algorithm, CMA-ES, differential evolution (DE), swarm intelligence, deep learning, RL, BO, or variations thereof.

[0059] In an aspect, the exchange of data is based on encoding, decoding, modulation, demodulation, compression, decompression, equalization, transmitting, receiving, relaying, filtering, measuring, synchronization, queuing, scheduling, delay, multiplexing, demultiplexing, sampling, resampling, authentication, scrambling, interleaving, beamforming, shaping, and the like.

[0060] In certain embodiments, the classical-quantum hybrid computing method is performed based on an adaptive filtering algorithm, a gradient descent algorithm, a feature extraction algorithm, a classification algorithm, a regression algorithm, a prediction algorithm, a compressive sensing algorithm, an estimation algorithm, an inference algorithm, a deep neural network, machine learning, a denoising algorithm, encoding, decoding, modulation, demodulation, equalization, or variations thereof.

[0061] In another embodiment, the sub-instructions are distributed to the networked CPUs and QPUs to determine resource allocation and operation types. At least one controller including a memory and a processor that stores the distributed sub-instructions can construct a computing graph and back-propagate the computing graph to calculate loss gradients across the multiple CPUs and QPUs so that the trainable parameters and the variational parameters are adjusted. In a classical computing processor, a deep neural network is constructed by a set of trainable parameters. In a quantum computing processor, a variational quantum circuit is specified by a set of variational parameters.

[0062] In an aspect, the training data and test data are based on time stamps, temperature measurements, sound volume, light measurements, image data, video data, magnetic flux measurements, accelerometer output measurements, pressure measurements, vibration measurements, infrared red light measurements, humidity measurements, power measurements, particle density measurements, odor measurements, radiation measurements, other digital data, and other analog data. The data is related to the environmental conditions to be estimated, such as the behavior type, state, attitude, and position of an object, such as stationary, moving, animate, or non-animate objects, such as surrounding furniture, surrounding materials, surrounding buildings, surrounding plants, surrounding pets, surrounding computers, surrounding robots, surrounding users, or communication devices.

[0063] In an aspect, the wireless channel attribute data is based on one or a combination of a signal strength amount, a noise strength amount, an interference strength amount, a signal-to-noise ratio (SNR) amount, a signal-to-interference-with-noise ratio (SINR) amount, a set of channel state information, a time-of-arrival amount, an angle of arrival (AoA), an angle of departure (AoD), a power delay profile (PDP), a power spectral density (PSD), a delay Doppler spectrum, an angular power spectrum, a beam power profile, or other data.

[0064] In an embodiment, the disparate data is acquired by one or a combination of visual / imaging sensors, temperature sensors, radiation sensors, proximity sensors, pressure sensors, position sensors, photoelectric sensors, particle sensors, motion sensors, metal sensors, level sensors, leak sensors, humidity sensors, gas / chemical sensors, force sensors, flow sensors, flaw sensors, flame sensors, electrical sensors, contact sensors, non-contact sensors, or other sensor devices.

[0065] In another aspect, the estimated environmental state is determined over a period of time and includes one or a combination of at least one object behavior type, at least one object state, at least one object pose, or at least one object position, where the at least one object is one of stationary objects, moving objects, living things, non-living things such as one or a combination of surrounding furniture, surrounding materials, surrounding buildings, surrounding plants, surrounding pets, surrounding computers, surrounding robots, surrounding users or wireless devices.

[0066] In another aspect, the environmental state includes one or a combination of at least one activity of a living being, at least one posture of an object or living being, occupancy of an object or living being, amount of objects or living beings in an area of ​​the environment, indoor traffic volume, location of an object or living being, location of an outdoor object or outdoor living being within range of the wireless device, a stationary or moving robot or non-living being, a set of instructions related to indoor navigation, or a set of instructions related to indoor parking assistance.

[0067] In yet another aspect, the parameterized model is based on a trained deep neural network, the deep neural network being trained online by accessing a communication network or offline by accessing stored measurement data to obtain at least one type of measurement from a first radio band and at least one type of measurement from a second radio band. The set of fused measurements is input to the deep neural network to generate an estimated environmental state in the environment. A loss value is calculated based on a difference between the generated estimated environmental state and the stored estimated environmental state accessed from the stored measurement data. The set of trainable parameters of the deep neural network is updated by a set of rules for training, the set of rules for training being based on a gradient method. The deep neural network includes one or a combination of first network blocks that encode the set of fused measurements into an encoded vector depending on whether the two types of measurements to be fused are different types of measurement formats, such as uniformly aligned asynchronous and non-cooperative measurement formats. The second network block converts the encoded vector into a feature vector such that the feature vector is insensitive to the different types of measurement formats. The third network block generates an estimated environment state based on the feature vector and a set of fusion measurements. The fourth network block reconstructs the set of fusion measurements from the feature vector. The fifth network block regularizes the feature vector for disentanglement in an adversarial fashion.

[0068] In an aspect, the at least one type of measurement from the first radio band is a medium-grain beam signal-to-noise ratio (SNR) measurement in the 60 GHz millimeter wave band, and the at least one type of measurement from the second radio band is a fine-grain channel state information (CSI) measurement in sub-6 GHz from multiple spatial streams, and thus the fusion includes fusing the fine-grain CSI measurement in sub-6 GHz from multiple spatial streams with the fine-grain CSI measurement in sub-6 GHz from multiple spatial streams. The fine-grain CSI measurement includes complex-valued amplitudes at orthogonal frequency division multiplexing (OFDM) subcarrier tones, such that the fine-grain CSI measurement is equivalent to a power delay profile (PDP) in the time domain, reflecting the power distribution along the propagation path, and the medium-grain beam SNR provides a spatial domain channel measurement across multiple beamforming directions or beam spaces.

[0069] In another aspect, the coverage of the environmental conditions is automatic based on the generation of the estimated environmental conditions. The computing processor is configured to access each of the modules through the data storage through an executable program, each module including one or more predefined rules. Iteratively input the estimated environmental conditions into each model to generate an output of either an alert action or no action, and if an alert action is generated by the module, it indicates a violation of at least one predefined rule, and the computing processor converts the alert action into an alert signal and transmits the alert signal via a transceiver to a communication network. The alert signal is received and input into an alert processing module associated with the system to generate an output including one or more corrective actions along with a set of instructions based on the estimated environmental conditions that are transmitted to a work team for completion. At least the module is a proximity module including inputting the estimated environmental conditions into the proximity module to generate a proximity alert action or not generate a proximity alert action, where if a proximity alert (PA) action is generated, a violation of at least one predefined proximity rule is indicated, for example, an object moving within the environment has entered a predefined object-prohibited designated area within the environment, and a computing processor converts the PA action into a PA signal and transmits the PA signal via a transceiver to a communication network. The PA signal is received and input into an alert processing module associated with the system to generate an output including one or more corrective proximity actions having a set of instructions based on the estimated environmental conditions that are transmitted to a work team for completion, where the one or more corrective proximity actions include directing the moving object to move out of the predefined object-prohibited designated area.

[0070] In another aspect, each estimated environmental state for each time period of the plurality of time periods determines a position of at least one object or organism within the environment, and each estimated environmental state is displayed on a display device to provide a visual tracking indication of the determined position of the at least one object or organism for that time period.

[0071] In an aspect, an executable program includes instructions for performing coordination between at least one wireless device using a plurality of antenna elements across a first communication channel of a first wireless band and other wireless devices including instructions using respective plurality of antenna elements across a first communication channel of the first wireless band, the instructions, when executed by a computing processor, performing coordination to time synchronize the at least one wireless device with the other wireless devices.

[0072] In another aspect, the stored data includes values ​​indicative of signal-to-noise ratio (SNR) measurements of a set of beams radiated at different beam angles by a phased antenna array and measured at a set of locations, which provides a mapping between different combinations of SNR values ​​of the set of beams and the set of locations, such that a location in the set of locations is a location for a period of time that maps to a unique combination of SNR values ​​of the set of beams, and further provides information associated with a type of behavior of the device at that location for that period of time, an attitude of the device at that location for that period of time, locations of physical objects in the environment for that period of time, and a type of behavior of surrounding users in the environment for that period of time.

[0073] Further, in an aspect, the stored data includes values ​​indicative of link attributes including one of a beam received signal strength indicator (RSSI) measurement, a beam channel state information (CSI) measurement, a beam pattern or beam sequence, each link attribute including measurements of a set of beams radiated at different beam angles by a phased array of antennas and measured at a set of locations in the environment, and the stored values ​​provide a mapping between different combinations of at least one link attribute value of the set of beams and the set of locations, a location in the set of locations being mapped to a unique combination of at least one link attribute value of the set of beams. The link attributes further include one or a combination of RSSI measurements or CSI measurements.

[0074] An aspect includes using a communication system having beamforming transmission of a millimeter wave spectrum in an environment, including a phased antenna array configured to perform beamforming to establish millimeter wave channel links with devices at different locations within the environment. The method includes performing beam training with a target device located within the environment to estimate SNR measurements of different beams transmitted over different beam angles using a control circuit connected to the antenna. The control circuit is configured to select at least one primary angle for beamforming communication with the target device in response to the beam training. A memory connected to the phased antenna array is accessed, the memory having data stored therein. The stored data includes values ​​indicative of SNR measurements of a set of beams radiated by the phased antenna array at different beam angles and measured at a set of locations within the environment. The stored values ​​provide a mapping between different combinations of SNR values ​​of the set of beams and a set of locations. Thus, a location within the set of locations is mapped to a unique combination of SNR values ​​of the set of beams. From the mapping stored in the memory, a location of the target device corresponding to the SNR values ​​of the different beams estimated during beam training is estimated. A phased antenna array is used to transmit an estimated location of a target device via beamforming transmission over at least one principal angle.

[0075] In another aspect, the stored data from the data storage includes fingerprinting data, where the fingerprinting data includes each location from a set of locations mapped to a unique combination of SNR values ​​of a set of beams for a period of time, and thus the unique combination of SNR values ​​of the set of beams for that period of time further provides information associated with a type of device behavior at that location for that period of time, a device attitude at that location for that period of time, locations of physical objects in the environment for that period of time, and a type of surrounding user behavior in the environment for that period of time.

[0076] Some aspects include a communication system using beamforming transmission of a millimeter wave spectrum in an environment, the communication system including a phased antenna array configured to perform beamforming to establish millimeter wave channel links with devices at different locations in the environment. The communication system includes a memory coupled to the phased antenna array and having stored data. The stored data includes values ​​indicative of signal-to-noise ratio (SNR) measurements of a set of beams radiated at different beam angles by the phased antenna array and measured at the set of locations, providing a mapping between different combinations of SNR values ​​of the set of beams and the set of locations. Thus, a location in the set of locations is mapped to a unique combination of SNR values ​​of the set of beams. A control circuit is communicatively coupled to the phased antenna array and the memory and configured to estimate SNR values ​​of different beams transmitted over different beam angles by performing beam training at a target device located in the environment. In response to the beam training, the control circuit selects at least one primary angle for beamforming communication with the target device. From the mapping stored in the memory, estimate a location of the target device corresponding to the SNR values ​​of the different beams estimated during beam training. A phased antenna array is used to transmit an estimated location of a target device via beamforming transmission over at least one principal angle.

[0077] In other aspects, the stored data includes each location of the set of locations mapped to a unique combination of SNR values ​​of the set of beams for a period of time, which unique combination of SNR values ​​of the set of beams for that period of time thus further provides information associated with a type of device behavior at the location for that period of time, a device attitude at the location for that period of time, locations of physical objects in the environment for that period of time, and a type of surrounding user behavior in the environment for that period of time.

[0078] The above-described embodiments of the present invention may be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software, or a combination thereof. If implemented in software, the software code may be executed on any suitable processor or collection of processors, which may be located in one computer or distributed across multiple computers. Such a processor may be implemented as an integrated circuit, with one or more processors in an integrated circuit component. However, the processor may be implemented using circuitry in any suitable format.

[0079] Also, embodiments of the invention may be implemented as a method, an example of which is provided. The order of operations performed as part of the method may be determined in any suitable manner. Thus, embodiments may be configured to perform operations in an order different from that illustrated, which may include performing some operations simultaneously, although in the illustrated embodiment they are shown as a sequence of operations.

[0080] In the claims, ordinal terms such as "first" or "second" modifying a claim element do not, by themselves, imply a priority, precedence, or order of the claim element over other elements, or a chronological order of performing the operations of a method, but are merely used as a label to distinguish the claim elements from other elements having the same name (when no ordinal terms are used).

[0081] Although the invention has been described in terms of examples of preferred embodiments, it is to be understood that various other adaptations and modifications are possible within the spirit and scope of the invention.

[0082] Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.

Claims

1. 1. A system for integrated computing, sensing, and communication, the system comprising: At least one communication link; at least one classical computing processor configured with a set of trainable parameters; At least one quantum computing processor configured with a set of variational parameters; at least one controller that generates a computing graph on a network connecting the at least classical computing processor and the at least quantum computing processor through the at least one communication link; and at least one memory bank coupled to the at least one classical computing processor and the at least one quantum computing processor through the at least one communication link, the at least one memory bank storing instructions for implementing a set of computing methods based on networked classical-quantum hybrid processing for data communication, machine learning, and environmental sensing according to the computing graph, the instructions, when executed by the at least one classical computing processor and the at least one quantum computing processor, perform the following steps, the steps comprising: causing the at least one classical processor to exchange data over the at least one communications link to adjust the set of trainable parameters and the set of variational parameters and to update a set of data in the at least one memory bank according to the set of computing methods; and causing the at least one quantum processor to perform a measurement of a quantum state according to the set of computing methods and taking into account the set of variational parameters and the set of data in the at least one memory bank.

2. 10. The system of claim 1, wherein the at least one communication link is configured using an interface, a gateway, a switching hub, a router, an access point, or a variation thereof, and the at least one communication link is based on electrical wired communication, optical fiber communication, wireless communication, free space optical communication, magnetic coupling communication, acoustic communication, mmWave / Terahertz communication, or a variation thereof.

3. 2. The system of claim 1, wherein the at least one classical computing processor is based on a central processing unit, a graphics processing unit, a tensor processing unit, a field programmable gate array, a microprocessor, a digital signal processor, an application specific integrated circuit, a complex programmable logic device, or a variation thereof.

4. 10. The system of claim 1, wherein the at least one quantum computing processor is based on a superconducting quantum device, a trapped ion quantum device, a photonic quantum device, a neutral atom device, a nuclear magnetic resonance device, a quantum annealing device, a boson sampling device, a quantum emulator device, or a variation thereof.

5. The system of claim 1 , wherein the quantum computing processor further comprises a quantum state preparation module, a quantum state evolution module, and a quantum state measurement module.

6. 6. The system of claim 5, wherein the quantum state preparation module is based on angle embedding, amplitude embedding, basis embedding, displacement embedding, quantum approximate optimization algorithm embedding, squeezing embedding, instantaneous quantum polynomial time embedding, Mottonen state preparation Ansatz, or a variation thereof.

7. 6. The system of claim 5, wherein the quantum state evolution module is based on a simplified two-design Ansatz, a quantum approximate optimization algorithm Ansatz, a tree tensor network Ansatz, a matrix product state Ansatz, a multi-scale entanglement renormalization Ansatz, a strong entanglement layer Ansatz, a basic entangler Ansatz, a random layer Ansatz, a continuously variable neural network layer, or a variation thereof.

8. 6. The system of claim 5, wherein the quantum state measurement module is based on expectation values, sampling values, variance values, probabilities, quantum states, von Neumann entropy, mutual information, classical shadow protocols, computation-based quantum density matrices, or variations thereof.

9. 10. The system of claim 1, wherein the set of computing methods is based on an adaptive filtering algorithm, a gradient descent algorithm, a feature extraction algorithm, a classification algorithm, a regression algorithm, a prediction algorithm, a compressive sensing algorithm, an estimation algorithm, an inference algorithm, a deep neural network, machine learning, a denoising algorithm, encoding, decoding, modulation, demodulation, equalization, or variations thereof.

10. 2. The system of claim 1, wherein the set of data is associated with an environmental state to be estimated, the environmental state including one or a combination of behavior type, state, posture, movement, and position of at least one object, the at least one object being stationary, moving, living, or non-living such as one or a combination of surrounding furniture, surrounding materials, surrounding buildings, surrounding plants, surrounding pets, surrounding computers, surrounding robots, surrounding users, or communication devices.

11. A computer-implemented method for signal processing, communication and sensing performed by at least one classical computing processor and at least one quantum computing processor, the at least one classical computing processor and the at least one quantum computing processor being coupled to a memory bank through at least one communication link, the memory bank storing instructions implementing a set of computing methods based on networked hybrid classical-quantum processing, the instructions performing the steps of: exchanging a set of data between the at least one memory bank, the at least one classical computing processor, and the at least one quantum computing processor over the at least one communication link; Distributing a set of sub-instructions across the at least one classical computing processor and the at least one quantum computing processor; constructing a set of computing graphs on the at least one classical computing processor and the at least one quantum computing processor, such that gradients are backpropagated through the set of computing graphs; adjusting a set of variational parameters for the at least one quantum computing processor according to the set of computing graphs; adjusting a set of trainable parameters for the at least one classical computing processor according to the set of computing graphs; modifying the set of data according to the set of subinstructions, the set of variational parameters, the set of trainable parameters, and the set of computing graphs.

12. 12. The method of claim 11, wherein tuning the set of trainable parameters is based on stochastic gradient descent, elastic backpropagation, root mean square propagation, Broyden-Fletcher-Goldfarb-Shanno algorithm, adaptive momentum optimization, metaheuristic optimization, simulated annealing, genetic algorithms, simplex methods, adaptive subgradients, adaptive delta, evolutionary algorithms, evolutionary strategies, swarm intelligence, deep learning, reinforcement learning, Bayesian optimization, or variations thereof.

13. 12. The method of claim 11, wherein adjusting the set of variational parameters is based on stochastic gradient descent, parameter shifting rules, adjoint methods, finite differences, policy gradients, elastic backpropagation, root mean square propagation, Broyden-Fletcher-Goldfarb-Shanno algorithm, adaptive momentum optimization, simplex methods, adaptive subgradients, adaptive delta, metaheuristic optimization, simulated annealing, genetic algorithms, evolutionary algorithms, evolutionary strategies, swarm intelligence, deep learning, reinforcement learning, Bayesian optimization, or variations thereof.

14. 12. The method of claim 11, wherein the exchanging step further comprises the steps of encoding, decoding, modulating, demodulating, compressing, decompressing, equalizing, transmitting, receiving, relaying, filtering, measuring, synchronizing, queuing, scheduling, delaying, multiplexing, demultiplexing, sampling, resampling, authenticating, scrambling, interleaving, beamforming, shaping, or combinations thereof.

15. 12. The method of claim 11, wherein the set of computing methods is based on an adaptive filtering algorithm, a gradient descent algorithm, a feature extraction algorithm, a classification algorithm, a regression algorithm, a prediction algorithm, a compressive sensing algorithm, an estimation algorithm, an inference algorithm, a deep neural network, machine learning, a denoising algorithm, encoding, decoding, modulation, demodulation, equalization, or variations thereof.

16. 12. The method of claim 11, wherein distributing the set of subinstructions determines resource allocations and types of operations for the at least one classical computing processor and the at least one quantum computing processor.

17. The method of claim 11 , wherein modifying the set of data is based on a deep neural network specified by the set of trainable parameters for the at least one classical computing processor.

18. The method of claim 11 , wherein modifying the set of data is based on a variational quantum circuit specified by the set of variational parameters for the at least one quantum processor.

19. 12. The method of claim 11, wherein the set of data is based on one or a combination of a timestamp, a temperature amount, a sound volume, a light amount, image data, video data, a magnetic flux amount, an accelerometer output amount, a pressure amount, a vibration amount, an infrared red light amount, a humidity amount, a power amount, a particle density amount, an odor amount, a radiation amount, other digital data, and other analog data.

20. 12. The method of claim 11, wherein the set of data is associated with an environmental state to be estimated, the environmental state including one or a combination of behavior type, state, pose, movement, and position of at least one object, the at least one object being stationary, moving, living, or non-living, such as one or a combination of surrounding furniture, surrounding materials, surrounding buildings, surrounding plants, surrounding pets, surrounding computers, surrounding robots, surrounding users, or communication devices.

Images