Low sampling rate wi-fi sensing method based on compressed sensing depth unfolding network

Abstract

The application relates to a low-sampling-rate WiFi sensing method based on a compressed sensing depth unfolding network, which comprises the following steps: collecting observed non-uniform signal data; up-sampling the observed non-uniform signal data to an initial state of a preset resolution; inputting the initial state of the preset resolution into a Doppler-DCNet network to obtain a high-precision channel state information signal; wherein the Doppler-DCNet network unfolds an iterative optimization algorithm into a plurality of cascaded stages, each stage parallelly processes time domain features and frequency domain features, the processed time domain features and frequency domain features are fused, the fused time domain features and frequency domain features are corrected through physical constraints, and an intermediate signal is output, and after a plurality of stages, a high-precision channel state information signal is obtained; and the high-precision channel state information signal is input into a behavior recognition model to obtain a human activity category. The application guarantees the robustness and accuracy of a sensing system.

Description

Technical Field

[0001] This invention relates to the field of wireless sensing technology, and in particular to a low-sampling-rate WiFi sensing method based on compressed sensing depth unfolding networks. Background Technology

[0002] With the evolution of wireless communication technology, WiFi technology has gradually expanded from simple data communication to integrated sensing and communication (ISAC). WiFi sensing technology utilizes Channel State Information (CSI) for human activity identification and has wide applications in health monitoring and smart homes. In WiFi sensing, both high-precision sensing tasks and efficient data communication processes are constrained by channel time and spectrum resources. Therefore, a large amount of research focuses on how to reduce communication overhead while ensuring sensing accuracy. In addition, due to multipath effects, co-channel interference, and hardware processing delays, low sampling rates and non-uniform arrival can easily occur during data acquisition at the receiving end, especially under high communication loads or limited device resources.

[0003] Currently, typical signal distortion modes are categorized into low sampling rate and non-uniform sampling, with random packet loss being one of the most common time-domain signal distortion methods. In random packet loss, the channel response acquired by the receiver is lost due to discontinuous time intervals, resulting in the loss of sensing features. To address random packet loss, many researchers have proposed relevant solutions and applied them to WiFi sensing systems. Signal recovery based on linear or nonlinear fitting refers to estimating the signal amplitude and phase at missing moments using known sampling points. Therefore, many studies utilize cubic splines or Archima interpolation to achieve data recovery, as these methods can complete the continuity of the waveform in the time dimension. In addition, other researchers have proposed methods based on compressed sensing sparse reconstruction or generative adversarial networks (GANs) and Transformers to achieve high-precision signal recovery and improve the robustness of sensing systems.

[0004] Compressed sensing is a commonly used sparse signal reconstruction algorithm. However, due to the extremely high real-time requirements of CSI sensing tasks, and the fact that traditional compressed sensing algorithms (such as ISTA or ADMM) involve a large number of iterative operations and matrix multiplications, resulting in high computational complexity, they cannot be directly used on edge WiFi devices with limited computing power. Summary of the Invention

[0005] This invention provides a low-sampling-rate WiFi sensing method based on compressed sensing depth unfolding network to solve the problem of high-precision signal recovery under limited communication resources.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a low sampling rate WiFi sensing method based on compressed sensing depth unfolding network, including the following steps:

[0007] Collect and observe non-uniform signal data;

[0008] The observed non-uniform signal data is upsampled to an initial state with a preset resolution;

[0009] The initial state at the preset resolution is input into the Doppler-DCNet network to obtain a high-precision channel state information signal; wherein, the Doppler-DCNet network expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained;

[0010] The high-precision channel state information signal is input into the behavior recognition model to obtain the category of human activity.

[0011] When upsampling the observed non-uniform signal data to an initial state with a preset resolution, the observed non-uniform signal data is upsampled to an initial state with a preset resolution through bilinear interpolation.

[0012] Each of the stages includes:

[0013] The time-domain reconstruction module is used to capture the temporal envelope features of the channel state information signal from the input signal to obtain the time-domain features;

[0014] The frequency domain reconstruction module is used to denoise the input signal by utilizing the sparsity of the Doppler domain to obtain frequency domain features;

[0015] The feature fusion module is used to concatenate the time-domain features and frequency-domain features along the channel dimension, and perform feature dimensionality reduction and fusion through a multi-layer convolutional network to obtain the intermediate reconstructed signal.

[0016] The data consistency layer is used to correct the intermediate reconstructed signal to obtain an intermediate signal.

[0017] The time-domain reconstruction module includes:

[0018] The one-dimensional convolutional feature extraction part is used to extract features from the input signal;

[0019] The local modulation module is used to enhance the extracted features;

[0020] The residual connection part is used to output temporal features based on the enhanced features.

[0021] The frequency domain reconstruction module includes:

[0022] The time-frequency transformation section is used to perform a fast Fourier transform on the input signal along the time dimension, transforming the input signal from the time domain to the Doppler frequency domain, and splicing the real and imaginary parts of the complex spectrum in the channel dimension to construct a real tensor;

[0023] The local modulation module is used to extract cross-subcarrier correlation features from the real tensor in the frequency domain, remove noise interference, and obtain the enhanced real and imaginary parts.

[0024] The inverse time-frequency transform section is used to recombine the enhanced real and imaginary parts into a complex number, perform an inverse fast Fourier transform, and output the frequency domain features.

[0025] The local modulation module generates query components, key components, and value components from the input features through convolutional layers, and performs parallel spatial and channel operations on the query and key components. The features obtained after the spatial and channel operations are fused with a projection layer through depthwise convolution to obtain enhanced features. Specifically, the spatial operation involves extracting local features through grouped convolution, generating a spatial attention mask through a sigmoid activation function, and weighting the spatial dimensions of the features. Specifically, the channel operation involves generating a channel attention mask through global average pooling and convolution to weight the subcarrier channels.

[0026] The data consistency layer downsamples the intermediate reconstructed signal and calculates its residual with the non-uniform signal data; then it upsamples the residual and subtracts it from the intermediate reconstructed signal.

[0027] The technical solution adopted by this invention to solve its technical problem is: to provide a low sampling rate WiFi sensing device based on compressed sensing depth unfolding network, comprising:

[0028] The acquisition module is used to acquire observed non-uniform signal data;

[0029] An upsampling module is used to upsample the observed non-uniform signal data to an initial state with a preset resolution;

[0030] A conversion module is used to input the initial state at the preset resolution into the Doppler-DCNet network to obtain a high-precision channel state information signal; wherein, the Doppler-DCNet network expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained;

[0031] The recognition module is used to input the high-precision channel state information signal into the behavior recognition model to obtain the category of human activity.

[0032] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the above-mentioned low sampling rate WiFi sensing method based on compressed sensing depth unfolding network.

[0033] The technical solution adopted by the present invention to solve its technical problem is: to provide a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it implements the steps of the above-mentioned low sampling rate WiFi sensing method based on compressed sensing depth unfolding network.

[0034] Beneficial effects

[0035] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art: The present invention constructs a Doppler-DCNet network that includes a physical constraint data consistency module and a dual-domain denoising module, which realizes high-precision recovery of the original signal from non-uniform CSI data under the condition of low sampling rate and random packet loss, thus ensuring the robustness and accuracy of the sensing system. Attached Figure Description

[0036] Figure 1 This is a flowchart of the low sampling rate WiFi sensing method based on compressed sensing depth unfolding network according to the first embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the stage portion of the Doppler-DCNet network in the first embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0039] The first embodiment of the present invention relates to a low sampling rate WiFi sensing method based on a compressed sensing deep unfolded network. The method constructs a Doppler-DCNet network that includes time-frequency dual-domain collaborative processing and local modulation mechanism. The Doppler-DCNet network, combined with a data consistency layer with physical constraints, realizes the reconstruction of high-resolution CSI signals from low sampling rate non-uniform signals, and then completes the identification of human activity categories through a pre-trained recognition model.

[0040] like Figure 1 As shown, the low sampling rate WiFi sensing method based on compressed sensing depth unfolding network in this embodiment includes the following steps:

[0041] Step 1: Collect the observed non-uniform signal data.

[0042] The method described in this embodiment is applicable to WiFi sensing systems based on Orthogonal Frequency Division Multiplexing (OFDM), including a transmitter and a receiver. This WiFi sensing system acquires Channel State Information (CSI) during continuous observation time, and the original complete CSI matrix can be represented as: ,in, and These represent the number of subcarriers and the number of sampling packets within the continuous observation time, respectively. Considering the dynamic multipath mechanism in indoor environments, the first... The subcarrier at the _ ... Channel frequency response (CFR) at each sampling time It is composed of the superposition of static background component and dynamic multipath component, and is represented as:

[0043] (1)

[0044] in, Represents static components that do not change over time; It is a set of dynamic paths with non-zero Doppler frequency shift; and The first The amplitude and initial delay of the dynamic path; The Doppler frequency shift is caused by human motion. Equation (1) shows that the CSI signal has a linear phase structure in the frequency domain and significant sparsity characteristics in the Doppler frequency domain.

[0045] Due to random packet loss or low sampling rate strategies, the data acquired by the receiver exhibits non-uniform characteristics. Define a binary mask matrix. ,in, This indicates that the data is valid. This indicates data loss. The non-uniform signal actually observed at the receiver. It can be represented as:

[0046] (2)

[0047] in, , It represents the Hadamard Product. This is additive noise that follows a complex Gaussian distribution. The goal of this implementation is to reduce noise from sparse observations. Recover the original CSI matrix Based on the sparsity of human activity in the Doppler domain, the signal recovery problem is modeled as a compressed sensing optimization problem. Let... Represents the original CSI matrix Sparse representation in the Doppler domain, i.e. ,in, This is the normalized discrete Fourier transform matrix. It restores the complete original CSI matrix. The objective function is defined as follows:

[0048] (3)

[0049] The first term is the Data Consistency (DC) term, which constrains the reconstructed signal to remain consistent with physical measurements at valid observation locations; the second term is the Sparsity Regularization term, which utilizes... Norm-constrained sparsity of signals in the Doppler domain; This is the regularization parameter.

[0050] Step 2: Upsample the observed non-uniform signal data to an initial state with a preset resolution. In this step, bilinear interpolation is used to upsample the observed non-uniform signal data. Upsampled to high resolution initial state .

[0051] Step 3: Input the initial state of the preset resolution into the Doppler-DCNet network to obtain a high-precision channel state information signal.

[0052] The Doppler-DCNet network in this step expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained.

[0053] like Figure 2 As shown, each stage includes:

[0054] The temporal reconstruction module is used to capture the temporal envelope features of the channel state information signal from the input signal, thereby obtaining the temporal features. This module comprises a one-dimensional convolutional feature extraction section, a local modulation module (LMB), and a residual connection section, connected sequentially. After the input signal undergoes feature mapping through the convolutional layers of the one-dimensional convolutional feature extraction section, it is input to the local modulation module for feature enhancement, and finally, the temporal features are output through the residual connection section. .

[0055] The frequency domain reconstruction module utilizes the sparsity of the Doppler domain to denoise the input signal and obtain frequency domain features. The frequency domain reconstruction module sequentially includes a time-frequency transformation section, a local modulation module section, and a time-frequency inverse transformation section. The time-frequency transformation section performs a Fast Fourier Transform (FFT) on the input signal along the time dimension, transforming the input signal from the time domain to the Doppler frequency domain, and concatenates the real and imaginary parts of the complex spectrum along the channel dimension to construct a real tensor. The local modulation module extracts cross-subcarrier correlation features from the real tensor in the frequency domain, removes noise interference, and obtains enhanced real and imaginary parts. The time-frequency inverse transformation section reassembles the enhanced real and imaginary parts into a complex number, performs an inverse Fast Fourier Transform, and outputs the frequency domain features. .

[0056] The local modulation module, as the core of the time-domain reconstruction module and the frequency-domain reconstruction module, generates query (Q) components, key (K) components and value (V) components from the input features when processing the input, and performs parallel spatial and channel operations on the query (Q) components and key (K) components.

[0057] Spatial operations refer to extracting local features through grouped convolutions, generating spatial attention masks using the sigmoid activation function, and then weighting the spatial dimensions of the features. This can be represented as:

[0058] (4)

[0059] Channel operations refer to the weighting of subcarrier channels by generating channel attention masks through global average pooling and convolution, which can be expressed as:

[0060] (5)

[0061] in, Features after spatial operations Features after channel operation For input features, It is the Sigmoid activation function. For convolution processing, It is a linear rectified function. For normalization operations, Features after channel operation This is a global average pooling operation.

[0062] The feature fusion module is used to fuse the time-domain features. and frequency domain features The signals are concatenated along the channel dimension and then subjected to feature dimensionality reduction and fusion through a multi-layer convolutional network to obtain the intermediate reconstructed signal. .

[0063] A data consistency layer is used for the intermediate reconstructed signal. The intermediate signal is obtained by making corrections. To ensure the reconstructed signal conforms to physical measurement constraints, this implementation designes a data consistency layer based on interpolation error correction. Specifically, it involves: processing the intermediate reconstructed signal... Downsampling is performed to a low-resolution scale, and the result is compared with the true observation. The residual is then upsampled back to a high-resolution scale, and the signal is reconstructed from the intermediate value. Subtracting from the middle forces the network output to maintain physical consistency. This can be expressed by the following formula:

[0064]

[0065] in, This is an intermediate signal. For upsampling operation, This is a downsampling operation.

[0066] initial state go through After this processing, the final high-precision CSI signal can be output. .

[0067] Step 4, transfer the high-precision channel state information signal The data is input into the behavior recognition model to obtain the categories of human activities. The behavior recognition model in this step is trained based on training data consisting of various channel state information signals and corresponding human activity categories.

[0068] It is easy to see that the Doppler-DCNet network constructed in this embodiment, which includes a physical constraint data consistency module and a dual-domain denoising module, can recover the original signal with high precision from non-uniform CSI data under the condition of low sampling rate and random packet loss, thus ensuring the robustness and accuracy of the sensing system.

[0069] A second embodiment of the present invention relates to a low-sampling-rate WiFi sensing device based on a compressed sensing depth unfolding network, comprising:

[0070] The acquisition module is used to acquire observed non-uniform signal data;

[0071] An upsampling module is used to upsample the observed non-uniform signal data to an initial state with a preset resolution;

[0072] A conversion module is used to input the initial state at the preset resolution into the Doppler-DCNet network to obtain a high-precision channel state information signal; wherein, the Doppler-DCNet network expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained;

[0073] The recognition module is used to input the high-precision channel state information signal into the behavior recognition model to obtain the category of human activity.

[0074] The upsampling module upsamples the observed non-uniform signal data to an initial state with a preset resolution using bilinear interpolation.

[0075] Each of the stages includes:

[0076] The time-domain reconstruction module is used to capture the temporal envelope features of the channel state information signal from the input signal to obtain the time-domain features;

[0077] The frequency domain reconstruction module is used to denoise the input signal by utilizing the sparsity of the Doppler domain to obtain frequency domain features;

[0078] The feature fusion module is used to concatenate the time-domain features and frequency-domain features along the channel dimension, and perform feature dimensionality reduction and fusion through a multi-layer convolutional network to obtain the intermediate reconstructed signal.

[0079] The data consistency layer is used to correct the intermediate reconstructed signal to obtain an intermediate signal.

[0080] The time-domain reconstruction module includes:

[0081] The one-dimensional convolutional feature extraction part is used to extract features from the input signal;

[0082] The local modulation module is used to enhance the extracted features;

[0083] The residual connection part is used to output temporal features based on the enhanced features.

[0084] The frequency domain reconstruction module includes:

[0085] The time-frequency transformation section is used to perform a fast Fourier transform on the input signal along the time dimension, transforming the input signal from the time domain to the Doppler frequency domain, and splicing the real and imaginary parts of the complex spectrum in the channel dimension to construct a real tensor;

[0086] The local modulation module is used to extract cross-subcarrier correlation features from the real tensor in the frequency domain, remove noise interference, and obtain the enhanced real and imaginary parts.

[0087] The inverse time-frequency transform section is used to recombine the enhanced real and imaginary parts into a complex number, perform an inverse fast Fourier transform, and output the frequency domain features.

[0088] The local modulation module generates query components, key components, and value components from the input features through convolutional layers, and performs parallel spatial and channel operations on the query and key components. The features obtained after the spatial and channel operations are fused with a projection layer through depthwise convolution to obtain enhanced features. Specifically, the spatial operation involves extracting local features through grouped convolution, generating a spatial attention mask through a sigmoid activation function, and weighting the spatial dimensions of the features. Specifically, the channel operation involves generating a channel attention mask through global average pooling and convolution to weight the subcarrier channels.

[0089] The data consistency layer downsamples the intermediate reconstructed signal and calculates its residual with the non-uniform signal data; then it upsamples the residual and subtracts it from the intermediate reconstructed signal.

[0090] A third embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the low sampling rate WiFi sensing method based on compressed sensing depth unfolding network described above.

[0091] The fourth embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the above-described low-sampling-rate WiFi sensing method based on compressed sensing depth unfolding network.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A low-sampling-rate WiFi sensing method based on compressed sensing depth unfolded network, characterized in that, Includes the following steps: Collect and observe non-uniform signal data; The observed non-uniform signal data is upsampled to an initial state with a preset resolution; The initial state at the preset resolution is input into the Doppler-DCNet network to obtain a high-precision channel state information signal; wherein, the Doppler-DCNet network expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained; The high-precision channel state information signal is input into the behavior recognition model to obtain the category of human activity.

2. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 1, characterized in that, When upsampling the observed non-uniform signal data to an initial state with a preset resolution, the observed non-uniform signal data is upsampled to an initial state with a preset resolution through bilinear interpolation.

3. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 1, characterized in that, Each of the stages includes: The time-domain reconstruction module is used to capture the temporal envelope features of the channel state information signal from the input signal to obtain the time-domain features; The frequency domain reconstruction module is used to denoise the input signal by utilizing the sparsity of the Doppler domain to obtain frequency domain features; The feature fusion module is used to concatenate the time-domain features and frequency-domain features along the channel dimension, and perform feature dimensionality reduction and fusion through a multi-layer convolutional network to obtain the intermediate reconstructed signal. The data consistency layer is used to correct the intermediate reconstructed signal to obtain an intermediate signal.

4. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 3, characterized in that, The time-domain reconstruction module includes: The one-dimensional convolutional feature extraction part is used to extract features from the input signal; The local modulation module is used to enhance the extracted features; The residual connection part is used to output temporal features based on the enhanced features.

5. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 3, characterized in that, The frequency domain reconstruction module includes: The time-frequency transformation section is used to perform a fast Fourier transform on the input signal along the time dimension, transforming the input signal from the time domain to the Doppler frequency domain, and splicing the real and imaginary parts of the complex spectrum in the channel dimension to construct a real tensor; The local modulation module is used to extract cross-subcarrier correlation features from the real tensor in the frequency domain, remove noise interference, and obtain the enhanced real and imaginary parts. The inverse time-frequency transform section is used to recombine the enhanced real and imaginary parts into a complex number, perform an inverse fast Fourier transform, and output the frequency domain features.

6. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 3 or 4, characterized in that, The local modulation module generates query components, key components, and value components from the input features through convolutional layers, and performs parallel spatial and channel operations on the query components and key components. The features obtained after the spatial and channel operations are fused with the projection layer through deep convolution to obtain enhanced features. The spatial operation specifically involves: extracting local features through grouped convolution, and then generating a spatial attention mask using a Sigmoid activation function to weight the spatial dimensions of the features; the channel operation specifically involves: generating a channel attention mask using global average pooling and convolution to weight the subcarrier channels.

7. The low sampling rate WiFi sensing method based on compressed sensing depth unfolded network according to claim 3, characterized in that, The data consistency layer downsamples the intermediate reconstructed signal and calculates its residual with the non-uniform signal data; then it upsamples the residual and subtracts it from the intermediate reconstructed signal.

8. A low-sampling-rate WiFi sensing device based on compressed sensing depth unfolding network, characterized in that, include: The acquisition module is used to acquire observed non-uniform signal data; An upsampling module is used to upsample the observed non-uniform signal data to an initial state with a preset resolution; A conversion module is used to input the initial state at the preset resolution into the Doppler-DCNet network to obtain a high-precision channel state information signal; wherein, the Doppler-DCNet network expands the iterative optimization algorithm as follows: Each stage is a cascaded process, handling both time-domain and frequency-domain features in parallel. The processed time-domain and frequency-domain features are then fused and corrected using physical constraints to output an intermediate signal. After several stages, a high-precision channel state information signal is obtained; The recognition module is used to input the high-precision channel state information signal into the behavior recognition model to obtain the category of human activity.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the low sampling rate WiFi sensing method based on compressed sensing depth unrolling network as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the low sampling rate WiFi sensing method based on compressed sensing depth unfolding network as described in any one of claims 1-7.

Images