Implementation method for restoring channel state information sample sparsity

Abstract

The invention discloses an implementation method for recovering channel state information sample sparsity, which comprises the following steps of: collecting CSI (Channel State Information) data, and carrying out denoising processing on the collected CSI data by adopting a subcarrier dimension PCA (Principal Component Analysis) denoising algorithm; dividing the CSI data after denoising processing into a plurality of CSI fragment sequences according to an original amplitude sequence, and resampling each CSI fragment sequence to a predetermined frequency; and carrying out recovery processing on the resampled CSI data by using a sparse sample recovery network model generated based on training of a generative adversarial network GANs. According to the method, low-sampling-rate and missing signals can be recovered by utilizing the characteristics of the original signals to the maximum extent, so that the stability and the precision of a downstream sensing task are improved; experimental results show that the technical scheme provided by the embodiment of the invention can improve the gait recognition accuracy and the gesture recognition accuracy by 15% and 9% respectively at 10Hz.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent Internet of Things, and in particular to a method for realizing recovery of sparse channel state information samples. Background Art

[0002] Currently, infrastructure based on Wi-Fi technology has been widely deployed in various indoor scenarios to achieve Internet connection and has become one of the key wireless technologies. In Wi-Fi scenarios, by analyzing the wireless signal pattern of Wi-Fi signals, it can show considerable potential in device-free human perception tasks, including activity recognition, gesture recognition, gait recognition, etc. However, in the process of analyzing the wireless signal pattern of Wi-Fi signals, it is found that the reliability and availability of the corresponding channel state information (CSI) are problematic, resulting in the inability to perform the corresponding device-free human perception tasks well.

[0003] Specifically, the corresponding CSI mainly has the following two problems:

[0004] (1) Problem of low sampling

[0005] CSI can only be estimated during data communication, so when data communication is inactive, the sampling rate of CSI is significantly reduced;

[0006] At the same time, since CSI data is high-dimensional data, maintaining a high frame rate for real-time applications may put a lot of pressure on the Wi-Fi communication network, which may interrupt the normal function of Wi-Fi.

[0007] (2) Irregularity issues

[0008] Due to the contention-based multiple access characteristics of Wi-Fi, the time intervals between adjacent frames often vary significantly, making the frame arrival rate of each link irregular. Moreover, the data caching and rate control processing of the upper-layer protocol further aggravate its irregularity. The corresponding inconsistent frame rate is converted into a fluctuating sampling rate, which will greatly affect Wi-Fi-based perception applications.

[0009] Based on the above two problems, interpolation technology can be considered for processing; however, interpolation technology only uses local information to fill in data gaps to ensure uniform distribution in the time domain; when part of the signal is missing, its processing may be ineffective.

[0010] In view of this, the present invention is proposed. Summary of the invention

[0011] The purpose of the present invention is to provide a method for realizing recovery of sparse channel state information samples, so as to utilize the characteristics of the original signal to recover the low sampling rate and missing signals, and solve the problems existing in the prior art.

[0012] The objective of the present invention is achieved through the following technical solutions:

[0013] A method for recovering sparse channel state information samples, comprising:

[0014] Collect channel state information (CSI) data and use the principal component analysis (PCA) denoising algorithm of the subcarrier dimension to denoise the collected CSI data;

[0015] Dividing the denoised CSI data into a plurality of CSI segment sequences according to the original amplitude sequence, and resampling each of the CSI segment sequences to a predetermined frequency;

[0016] The resampled CSI data is restored using a sparse recovery model generated by generative adversarial network (GANs) training to obtain missing CSI data, thereby achieving CSI data recovery.

[0017] The method further includes: performing automatic gain control (AGC) removal processing on the CSI data after the denoising processing.

[0018] The denoising process comprises:

[0019] The CSI matrix of the transposed CSI data is used to represent the reflection paths of all subcarriers in the environment over a period of time;

[0020] In the transposed CSI matrix, the first principal component is kept as being related to motion, and the other principal components are regarded as irrelevant components representing environmental noise, and the CSI data is reconstructed by selecting the eigenvector corresponding to the maximum eigenvalue therein to remove the environmental noise.

[0021] The resampling to a predetermined frequency comprises:

[0022] The CSI segment sequence is resampled to a frequency corresponding to a target perception task, where the target perception task is a perception task that needs to be performed based on the CSI data.

[0023] The sparse recovery model includes a generator and a discriminator, wherein:

[0024] The generator is used to restore missing CSI data on the input resampled CSI data, and includes a plurality of residual blocks with the same structure. The input CSI data is processed through a normalization layer and an activation function layer as the input of the residual block; the output of the generator is transmitted to the discriminator;

[0025] The discriminator is used to identify the output of the generator. If it is judged to be true, it represents real data, and if it is judged to be false, it represents generated data; the discriminator includes multiple downsampling blocks and corresponding multiple upsampling blocks, as well as an average pooling layer and an activation function layer connected to the sampled output.

[0026] The inputs in the sparse recovery model training process include:

[0027] Extract the amplitude sequence X = [x1, x2, ..., x n ], and randomly generate a mask sequence M=[m1,m2,…,m n ],m i ∈{0,1}; then multiply the amplitude sequence and the mask sequence as the input Y of the sparse recovery model, and:

[0028] Y=X⊙M=[y1,y2,…,y n ],y i =x i ·m i for i=1,2,…,n.

[0029] The method further includes:

[0030] According to the determination result of the discriminator, the loss function of the sparse recovery model is calculated based on the output of the generator, and the loss function The calculation formula is:

[0031]

[0032] in, is the time domain loss function, is the frequency domain loss function, α is a hyperparameter; and the frequency domain loss function is obtained by fast Fourier transform STFT calculation.

[0033] The time domain loss function is the mean square error of extracting low-level features, and its calculation formula is:

[0034]

[0035] Among them, E is the mean value, Y is the original CSI true value data, CSI data generated for the GAN network.

[0036] The calculation formula of the frequency domain loss function is:

[0037]

[0038] in, and Z represent the spectrum amplitudes of the signal generated by the generator and the original standard signal, respectively, and and Z are obtained by converting the corresponding CSI data into the frequency domain based on STFT, and U and V represent the number of time and frequency components, respectively.

[0039] The discriminator uses a leaky rectified linear activation function LeakyReLU activation, and after sampling the feature map, it is processed by an average pooling layer and an activation function layer to obtain the probability of each signal point being classified as true or false, and the average value of the probability of all signal points being classified as true or false is used to determine whether the input is original standard data or generated data.

[0040] Compared with the prior art, the technical solution provided by the present invention introduces a PCA denoising method in the subcarrier dimension to reduce the impact of noise on perception, thereby improving the quality of signal recovery under low sampling; and provides a processing method for training the model by combining the loss function of time domain and frequency domain information, which further effectively improves the signal recovery effect in the low frequency range (0-40Hz); therefore, the implementation of the embodiment of the present invention can maximize the use of the characteristics of the original signal to restore low sampling rate and missing signals, thereby improving the stability and accuracy of downstream perception tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0042] Figure 1 A flowchart of a method according to an embodiment of the present invention;

[0043] Figure 2 A comparison diagram of denoising algorithms provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the structure of a generator provided by an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of the structure of a discriminator provided in an embodiment of the present invention;

[0046] Figure 5A schematic diagram of comparing time-frequency domain results provided by an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of gait recognition results provided by an embodiment of the present invention;

[0048] Figure 7 A schematic diagram of a gesture recognition result provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in combination with the specific content of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, which does not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0050] First, the terms that may be used in this article are explained as follows:

[0051] The term “and / or” means that either or both of them can be realized at the same time. For example, X and / or Y means both “X” or “Y” and “X and Y”.

[0052] The terms "include", "comprises", "contains", "has" or other descriptions with similar semantics should be interpreted as non-exclusive inclusion. For example, including certain technical feature elements (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products or products, etc.) should be interpreted as including not only certain technical feature elements explicitly listed, but also other technical feature elements known in the art that are not explicitly listed.

[0053] The term "consisting of..." means excluding any technical feature elements not explicitly listed. If this term is used in a claim, it will make the claim closed, so that it does not contain technical feature elements other than the technical feature elements explicitly listed, except for the conventional impurities related to them. If this term only appears in a clause of a claim, it only limits the elements explicitly listed in the clause, and the elements recorded in other clauses are not excluded from the overall claim.

[0054] Unless otherwise specified or limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this article can be understood according to specific circumstances.

[0055] When concentration, temperature, pressure, size or other parameters are expressed in the form of a numerical range, the numerical range should be understood to specifically disclose all ranges formed by the pairing of any upper limit, lower limit, and preferred value in the numerical range, regardless of whether the range is explicitly stated; for example, if a numerical range of "2 to 8" is stated, the numerical range should be interpreted as including ranges such as "2 to 7", "2 to 6", "5 to 7", "3 to 4 and 6 to 7", "3 to 5 and 7", "2 and 5 to 7", etc. Unless otherwise specified, the numerical ranges stated herein include both their end values ​​and all integers and fractions within the numerical range.

[0056] The orientation or position relationship indicated by terms such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. are based on the orientation or position relationship shown in the drawings and are only for the convenience and simplification of description, and do not explicitly or implicitly indicate that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation of this document.

[0057] In the implementation process of the present invention, the image restoration and super-resolution image restoration technologies based on generative adversative networks (GANs) technology are mainly used to implement the present invention. Furthermore, in an embodiment of the present invention, a corresponding method for recovering CSI data based on a GAN network is provided to fill in missing pixels based on the learned data distribution based on the GANs technology, thereby realizing the recovery of sparse CSI samples. Specifically, the present invention provides an implementation method for recovering sparse channel state information samples, which can reduce the influence of noise through PCA denoising technology and generate missing signal values ​​using a generative adversarial network; therefore, the embodiment of the present invention can maximize the use of the characteristics of the original signal to recover low sampling rate and missing signals, thereby improving the stability and accuracy of downstream perception tasks.

[0058] The specific implementation process of a method for recovering sparse channel state information samples provided by an embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] like Figure 1 As shown, in the embodiment of the present invention, a new generation model for directly processing the original CSI waveform is introduced to recover the intermittent sparse samples caused by the Wi-Fi characteristics; specifically, the corresponding processing process may include the following processing steps:

[0060] (1) De-noising and gain processing of CSI data;

[0061] That is, collecting CSI (channel state information) data (i.e., sparse CSI data to be recovered), and using a PCA (principal component analysis) denoising algorithm of subcarrier dimension to denoise the collected CSI data;

[0062] Specifically, the PCA denoising algorithm of the subcarrier dimension and the AGC (automatic gain control) removal algorithm can be applied to the collected CSI data for denoising and gain processing, so as to reduce the influence of noise on perception through the PCA denoising method, that is, minimize the influence of noise in the CSI data on the perception processing process, thereby improving the quality and reliability of the recovered signal (that is, CSI data);

[0063] (2) Segmenting and resampling the CSI data after denoising and gain processing;

[0064] That is, the denoised CSI data is divided into a plurality of CSI segment sequences according to the original amplitude sequence, and each of the CSI segment sequences is resampled to a predetermined frequency;

[0065] Furthermore, the CSI data of the original amplitude sequence after denoising and gain processing can be divided into segments of a predetermined time period (such as 2 seconds, etc.), and each signal sequence (i.e., each CSI segment) can be resampled to a suitable frequency for the target perception task, thereby achieving a balance between computing resources and contextual information.

[0066] (3) CSI data recovery

[0067] Specifically, the resampled CSI data is restored using a sparse recovery model generated by generative adversarial network GANs training;

[0068] Furthermore, after the above denoising and resampling processing, the corresponding CSI data can be input into the sparse recovery model to achieve the recovery processing for the CSI data, thereby generating complete perception data. Specifically, the corresponding restored CSI data can be further processed to perform corresponding target perception tasks, such as time-frequency analysis, to identify the subject's gait or posture and other target perception tasks.

[0069] It can be seen from the description of the above processing that during the implementation of the present invention, the GAN network can be used to recover the collected CSI data packets in view of the current situation of low sampling rate and uneven data packet interval in actual application scenarios, so as to improve the accuracy of downstream perception tasks under poor data conditions. Moreover, during the implementation of the embodiment of the present invention, a loss function combining time domain and frequency domain information is also used to train the model, that is, the real waveform and the restored waveform are converted to the frequency domain using the short-time Fourier transform (STFT), and the frequency domain loss function used in the frequency domain directly emphasizes the lack of high-frequency information components, thereby providing global guidance for missing information during training, so that CSI data with sparse samples can be effectively recovered.

[0070] In the processing process (1) of the above-mentioned embodiment of the present invention, the CSI data is subjected to corresponding denoising processing, and specifically adopts the implementation method of PCA denoising in the subcarrier dimension. The implementation process of the denoising processing will be analyzed and explained in detail below.

[0071] Typically, the raw CSI data collected by Wi-Fi network cards contains various types of noise. When the target is far away from the transceiver or the walking direction is close to parallel to the propagation path, the signal pattern may be overwhelmed by the noise. Moreover, the presence of noise can cause signal distortion, reduce the signal-to-noise ratio, and damage the reliability of the signal, resulting in a limited perception range. In addition, during the low sampling rate signal recovery process, the noise is regarded as part of the signal itself, resulting in a large error between the recovered signal and the original signal.

[0072] In order to remove the corresponding noise, if the traditional denoising method (such as median filtering, wavelet denoising, etc.) is used, it may cause the loss of high-frequency information, subtle signal changes and shape changes, and possible time domain offset and distortion. The emergence of these effects may reduce the details and frequency domain characteristics of the signal, thereby affecting the recovery of low sampling rate signals. To this end, the embodiment of the present invention avoids the traditional noise removal method and instead uses principal component analysis (PCA) on the subcarrier dimension and extracts the first principal component to remove the noise, so that the environmental noise of the CSI data can be better removed, while retaining the dynamic characteristics of the signal as much as possible; that is, the embodiment of the present invention uses PCA denoising along the subcarrier dimension to minimize the impact of noise and improve the quality of the recovered signal, thereby ensuring the stability and reliability of subsequent personnel detection tasks.

[0073] The PCA denoising method depends on the selection of the number of principal components in the time dimension. To this end, the embodiment of the present invention transposes the CSI matrix to represent the reflection paths of all subcarriers in the environment over a period of time. Specifically, the first principal component can be kept as relevant to the perception task (such as motion-related), and the other principal components can be regarded as irrelevant components representing environmental noise; and then the signal can be reconstructed by selecting the eigenvector corresponding to the maximum eigenvalue to remove the environmental noise.

[0074] In order to further prove the effectiveness of the PCA denoising algorithm in the subcarrier dimension, the present invention also compares the effects of different denoising methods on the signal, including the time domain and the frequency domain. Figure 2 As shown. In the time-frequency spectrum of PCA denoising and wavelet threshold denoising, the noise area is significantly attenuated and the details of the signal are well preserved. In contrast, median filtering reduces the noise to a certain extent, but the retention of signal details is not as effective as principal component analysis and wavelet denoising, especially for high frequencies. Compared with wavelet threshold denoising, the PCA denoising used in the embodiment of the present invention can clearly separate the signal and noise areas in the spectrum, making the subsequent process of restoring the frequency domain characteristics of the signal more effective.

[0075] In the processing process (3) of the above-mentioned embodiment of the present invention, a process of recovering CSI data is executed. This process mainly performs recovery processing on the sparse recovery model generated by GANs network training. The implementation of this recovery process will be analyzed and explained in detail below.

[0076] That is, in the embodiment of the present invention, a generative adversarial network (GANs) is used to perform a corresponding data recovery algorithm, and GANs can use the learned data distribution to fill in the missing pixels. That is, the embodiment of the present invention uses the GANs network to learn the local information and global distribution of the signal to perform high-quality recovery of the missing CSI data.

[0077] Specifically, a new generative model (i.e., generator) is introduced in an embodiment of the present invention, which directly operates on the original CSI waveform (i.e., original CSI data) to recover continuous channel changes from CSI data that are intermittent sparse samples. Furthermore, the implementation process of the embodiment of the present invention not only considers the time domain of the CSI data, but also its frequency domain. The key to performing sparse recovery processing is to explore the generation of missing frequencies in the sparse part. To this end, an embodiment of the present invention proposes a new loss function in the frequency domain by using a short fast Fourier transform (STFT) to calculate the frequency components of the CSI data. That is, the corresponding generative model directly refers to the time domain and frequency domain of the original CSI data for recovery, thereby improving the accuracy of the amplitude sequence sampling and paying balanced attention to the spectral details of the data.

[0078] Furthermore, in the embodiment of the present invention, for the recovery of CSI data, the amplitude sequence X=[x1, x2, ..., x n ], and randomly generate a mask sequence M=[m1, m2, ..., m n ],m i ∈{0, 1} to simulate the situation of missing data in actual scenes; then the original amplitude sequence is multiplied by the mask sequence as the input Y of the sparse recovery model, and:

[0079] Y=X⊙M=[y1,y2,...,y n ],y i =x i ·m i for i=1,2,...,n;

[0080] The above formula can be used to obtain the input data for training the generated model.

[0081] The ultimate goal of the embodiment of the present invention is to train a generating function G, that is, to establish a corresponding generating model, so as to estimate the missing part of the original CSI signal through the generating function G, generate a standard CSI signal, and then realize the recovery of the CSI signal. To this end, the embodiment of the present invention trains the generator network as a feedforward convolutional neural network (CNN) and defines a corresponding discriminator network D to form a corresponding sparse recovery model. That is, the sparse recovery model may include a generator and a discriminator, wherein:

[0082] (I) Generator

[0083] The generator is used to perform missing recovery processing on the input CSI data after the resampling, and includes a plurality of residual blocks with the same structure. The input CSI data is processed through a normalization layer and an activation function layer as the input of the residual block; the output of the generator is transmitted to the discriminator for subsequent processing;

[0084] Specifically, considering the advantages of CNN in extracting high-dimensional feature information, the embodiment of the present invention designs a CNN model to achieve model generation, that is, to build a corresponding generator; the architecture of the generator can be as follows Figure 3 As shown, the core of the generator is the above-mentioned generating function G (also called the generator network G), which is a plurality of (assuming B) residual blocks with the same structure; and in the generator, two convolutional layers with sizes of 9×9 and 64 feature maps can also be used, and then the batch normalization layer and ParametricReLU (parametric rectified linear activation function) are used as activation functions to perform corresponding data generation processing;

[0085] (II) Discriminator

[0086] In order to distinguish the real signal (i.e., the real CSI data) from the generated samples, the embodiment of the present invention further trains a discriminator network, referred to as the discriminator;

[0087] The discriminator is used to identify the output of the generator, and if it is judged to be true, it represents real data, and if it is judged to be false, it represents generated data; the discriminator includes multiple downsampling blocks and corresponding multiple upsampling blocks, as well as an average pooling layer and an activation function layer connected to the sampled output;

[0088] Its specific implementation architecture remains the same Figure 3 As shown, LeakyReLU (leaky rectified linear activation function) is used in the discriminator to activate and avoid the maximum pooling of the entire network; it includes four down-sampling blocks Down Sample Block and four up-sampling blocks Up Sample Block. After the feature map is sampled, it is processed by the average pooling AvgPool1d layer and the activation function layer to obtain the probability of each signal point being classified as true or false. The average value of the probability of all signal points being classified as true or false is used to determine whether it is the original standard data or the generated data.

[0089] (III) Loss Function

[0090] According to the determination result of the differentiator, the loss function of the sparse recovery model is calculated based on the output of the generator, and the loss function The definition of is crucial to the performance of the sparse recovery network (i.e., the sparse recovery model), which performs recovery processing of missing CSI data based on the loss function, and in the sparse recovery model, the loss function can be formulated as the time domain loss and frequency domain loss The weighted sum of is as follows:

[0091]

[0092] Here, α is a hyperparameter.

[0093] Among them, the loss function in the time domain can be obtained by extracting the mean square error (MSE) of low-level features. Specifically, the corresponding loss function in the time domain is for:

[0094]

[0095] Among them, E is the mean value, and Y represent the amplitudes of the signal generated by the generator and the original standard signal respectively.

[0096] In addition to the loss in the time domain, the present invention also proposes a method for determining the loss function in the frequency domain so as to recover the high-frequency information of the sparse part; the key to sparse recovery is to explore the generation of missing frequencies from low frequency to high frequency, and compared with the spatial domain, specific frequencies can be clearly separated in the frequency domain; in addition, frequency components can provide global information about the signal; to this end, the embodiment of the present invention uses a short-time Fourier transform (STFT) to convert the real waveform and the restored waveform into the frequency domain, and calculates the L1 loss of the spectrum amplitude difference between the signal generated by the generator and the original standard signal

[0097]

[0098] in, and Z represent the spectrum amplitudes of the signal generated by the generator and the original standard signal, respectively, and and Z are obtained by converting the corresponding CSI data into the frequency domain based on STFT, and U and V represent the number of time and frequency components, respectively;

[0099] The frequency domain loss function used in the frequency domain directly emphasizes the lack of high-frequency information components, thereby providing global guidance on the missing information during training.

[0100] Based on the above generator, discriminator and corresponding loss function, the sparse recovery model trained by the embodiment of the present invention can reliably recover the low-sampled CSI signal, thereby providing a reliable data basis for the realization of subsequent perception tasks.

[0101] In summary, the technical solution provided by the embodiment of the present invention has at least the following technical advantages and effects:

[0102] (1) In the embodiment of the present invention, the proportion of the first principal component at different distances in the subcarrier dimension is analyzed according to the characteristics of the Wi-Fi signal, and a PCA denoising method in the subcarrier dimension is introduced to reduce the impact of noise on perception, thereby improving the quality of signal recovery under low sampling;

[0103] (2) The embodiment of the present invention also provides a processing method for training the model by combining the loss function of time domain and frequency domain information, which further effectively improves the signal recovery effect in the low frequency range (0-40Hz);

[0104] (3) In the embodiment of the present invention, the sparse recovery system can be applied to gait recognition and gesture recognition data sets, and experimental results show that compared with other existing technologies, the technical solution provided by the embodiment of the present invention can effectively improve the accuracy of perception tasks. Specifically, the gait recognition and gesture recognition accuracies are increased by 15% and 9% respectively at 10Hz.

[0105] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A method for recovering sparse channel state information samples, characterized in that: include: Collect channel state information (CSI) data and use the principal component analysis (PCA) denoising algorithm of the subcarrier dimension to denoise the collected CSI data; Dividing the denoised CSI data into a plurality of CSI segment sequences according to the original amplitude sequence, and resampling each of the CSI segment sequences to a predetermined frequency; The resampled CSI data is restored using a sparse recovery model generated by generative adversarial network (GANs) training to obtain missing CSI data, thereby achieving CSI data recovery.

2. The method according to claim 1, characterized in that The method further includes: performing automatic gain control (AGC) removal processing on the CSI data after the denoising processing.

3. The method according to claim 1, characterized in that The denoising process comprises: The CSI matrix of the transposed CSI data is used to represent the reflection paths of all subcarriers in the environment over a period of time; In the transposed CSI matrix, the first principal component is kept as being related to motion, and the other principal components are regarded as irrelevant components representing environmental noise, and the CSI data is reconstructed by selecting the eigenvector corresponding to the maximum eigenvalue therein to remove the environmental noise.

4. The method according to claim 1, characterized in that: The resampling to a predetermined frequency comprises: The CSI segment sequence is resampled to a frequency corresponding to a target perception task, where the target perception task is a perception task that needs to be performed based on the CSI data.

5. The method according to any one of claims 1 to 4, characterized in that: The sparse recovery model includes a generator and a discriminator, wherein: The generator is used to restore missing CSI data on the input resampled CSI data, and includes a plurality of residual blocks with the same structure. The input CSI data is processed through a normalization layer and an activation function layer as the input of the residual block; the output of the generator is transmitted to the discriminator; The discriminator is used to identify the output of the generator. If it is judged to be true, it represents real data, and if it is judged to be false, it represents generated data; the discriminator includes multiple downsampling blocks and corresponding multiple upsampling blocks, as well as an average pooling layer and an activation function layer connected to the sampled output.

6. The method according to claim 5, characterized in that The inputs in the sparse recovery model training process include: Extract the amplitude sequence X=[x1,x2,…,x n ], and randomly generate a mask sequence M=[m1,m2,…,m n ],m i ∈{0,1}; then multiply the amplitude sequence and the mask sequence as the input Y of the sparse recovery model, and: Y=X⊙M=[y1,y2,…,y n ],y i =x i ·m i for i=1,2,…,n。 7. The method according to claim 5, characterized in that The method further includes: According to the determination result of the discriminator, the loss function of the sparse recovery model is calculated based on the output of the generator, and the loss function The calculation formula is: in, is the time domain loss function, is the frequency domain loss function, α is a hyperparameter; and the frequency domain loss function is obtained by fast Fourier transform STFT calculation.

8. The method according to claim 7, characterized in that The time domain loss function is the mean square error of extracting low-level features, and its calculation formula is: Among them, E is the mean value, Y is the original CSI true value data, CSI data generated for the GAN network.

9. The method according to claim 7, characterized in that: The calculation formula of the frequency domain loss function is: in, and Z represent the spectrum amplitudes of the signal generated by the generator and the original standard signal, respectively, and and Z are obtained by converting the corresponding CSI data into the frequency domain based on STFT, and U and V represent the number of time and frequency components, respectively.

10. The method according to claim 5, characterized in that The discriminator uses a leaky rectified linear activation function LeakyReLU activation, and after sampling the feature map, it is processed by an average pooling layer and an activation function layer to obtain the probability of each signal point being classified as true or false, and the average value of the probability of all signal points being classified as true or false is used to determine whether the input is original standard data or generated data.

Images