A channel state information reconstruction method and system for channel state interference detection

Abstract

This invention proposes a channel state information reconstruction method and system for channel state interference detection. The invention generates effective feature embedding based on collected channel state information; uses a diffusion model to generate channel state information data for missing locations; calculates a spectral correlation function based on the high-sampling rate channel state information data, and performs interference detection. This invention addresses the problem in existing low-sampling Wi-Fi channel state information reconstruction techniques, which suffer from low accuracy and fail to fully consider the characteristics of the channel state information interference detection task, resulting in the inability to directly use Wi-Fi channel state information for radio frequency interference detection. This invention improves the channel state information reconstruction effect and interference detection accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of WiFi interference detection and signal processing, and in particular relates to a channel state information reconstruction method and system for channel state interference detection. Background Art

[0002] With the rapid development of wireless communication technology, wireless networks have been widely deployed in cities. Among them, WiFi networks, as the most widely used wireless communication network in daily life, have brought tremendous convenience to human life. However, the frequency band in which WiFi operates is very crowded and susceptible to radio frequency interference from devices such as Bluetooth and ZigBee, which greatly reduces the stability of communication. On the other hand, with the development of integrated telepathy technology, WiFi is widely used for non-invasive perception. However, the channel state information (CSI) used for WiFi perception carries information about radio frequency interference, which seriously affects the sensitivity and accuracy of perception applications. Therefore, detecting the presence of radio frequency interference in CSI data is of great significance for ensuring WiFi communication and perception performance, as well as interference mitigation.

[0003] Existing research has proposed methods, such as a spectral correlation function-based channel state information interference detection algorithm, to identify radio frequency interference in channel state information data. Although existing methods can effectively detect interference, real-world applications are prone to problems such as WiFi packet loss and hardware defects, resulting in discontinuous channel state information data and low sampling rates, which negatively impact interference detection.

[0004] A common approach to solving this problem is to reconstruct complete channel state information based on low-sampled channel state information data and use the complete channel state information for interference detection. Existing methods for reconstructing channel state information data are traditional interpolation methods, which use adjacent known data points and mathematical principles such as linear interpolation to infer unknown data points within a range. This method is simple, but has low accuracy when faced with high data loss rates. With the development of generative artificial intelligence technology, some studies have used generative adversarial networks (GANs) to reconstruct channel state information data in an image restoration manner, achieving high accuracy and fast speed. However, the above methods all focus on channel state information data recovery, ignoring the purpose of channel state information interference detection, and cannot be directly applied to the data reconstruction problem under the channel state information interference detection task.

[0005] Therefore, it is necessary to design a high-accuracy low-sampling WiFi channel state information reconstruction method for channel state information interference detection to solve the above problems. Summary of the Invention

[0006] To address the low accuracy of existing low-sampling WiFi channel state information reconstruction technologies and the inability to fully consider the characteristics of channel state information interference detection tasks, resulting in the inability to directly use WiFi channel state information for radio frequency interference detection, and to achieve the goal of improving the channel state information reconstruction effect and interference detection accuracy, the present invention proposes a channel state information reconstruction method and system for channel state interference detection.

[0007] The technical solution of the method of the present invention is a channel state information reconstruction method for channel state interference detection, comprising the following steps:

[0008] Step 1: Generate effective feature embedding based on the collected channel state information, where the channel state information data is low-sampling rate channel state information data collected by the WiFi signal receiving end;

[0009] Step 2: Generate channel state information data at the missing position using a diffusion model, wherein the channel state information data at the missing position and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data;

[0010] Step 3: Calculating a spectral correlation function based on the high sampling rate channel state information data and performing interference detection;

[0011] Preferably, the channel state information collected in step 1 is specifically as follows:

[0012] For a WiFi channel with M transmit antennas, N receive antennas, and F subcarriers, the channel state information data collected at T time moments is composed of low-sampling-rate channel state information. The low-sampling-rate channel state information of each transmit-receive antenna pair can be regarded as a three-dimensional matrix of shape F×T×2. This matrix is ​​composed of two two-dimensional matrices of shape F×T, which respectively include the real part data and imaginary part data of the low-sampling-rate channel state information of the transmit-receive antenna pair.

[0013] Initial channel state information

[0014] Where, C = 2 × M × N;

[0015] Among them, the channel state information value at the moment when the sampled data is missing is 0, m∈{0,1} F×T×C is the channel state information data mask, the value of m is 0 when data is missing, otherwise it is 1;

[0016] Preferably, the effective feature embedding is generated in step 1 as follows:

[0017] The temporal feature extraction unit uses the self-attention block to extract the temporal feature s from the i-th subcarrier data of the channel state information data i , then the time series characteristics

[0018]

[0019]

[0020] Among them, LP is the linear projection operation, Norm is the layer normalization operation, X i,:,: represents the data of all transmit and receive antenna pairs at all times of the i-th subcarrier, and softmax is the softmax function:

[0021]

[0022] x n is the nth element of vector x, X i,:,: The corresponding Query, Key, Value matrix in the attention mechanism, the hyperparameter d k is the scaling factor, is the set of real numbers, represents a real matrix of size F×T×C;

[0023] The frequency domain feature extraction unit uses the self-attention block to extract the frequency domain feature s of the j-th moment data of the channel state information data j ,

[0024]

[0025] Among them, the hyperparameter d k is the scaling factor, X :,j,: Represents the data of all subcarriers and all transmit and receive antenna pairs at the jth moment;

[0026] X :,j,: The corresponding Query, Key, and Value matrices in the attention mechanism, represents a real matrix of size (F×1×C)×T;

[0027] Cross-fusion unit, used to extract features between time domain and frequency domain, and use cross-attention block to fuse the time series features and the frequency domain features to obtain time-frequency features

[0028] K3,V3=LP(Norm(S freq ))

[0029] Q3=LP(Norm(S time ))

[0030]

[0031] Among them, the hyperparameter d k is the scaling factor,

[0032] K3 and V3 are the Key and Value matrices in the attention mechanism corresponding to the frequency domain features, respectively, and Q3 is the Query matrix in the attention mechanism corresponding to the time series features;

[0033] The mask fusion unit is used to fuse the time-frequency features and the channel state information data mask to obtain effective feature embedding

[0034] S condition =SiLU(Concat(S cross ,m))

[0035] Among them, Concat represents data concatenation, the activation function SiLu(·) is a sigmoid weighted linear unit, m∈{0,1} F×T×C is the channel state information data mask. The value of m is 0 when data is missing and 1 otherwise.

[0036] Preferably, the channel state information data of the missing position is generated using a diffusion model in step 2, specifically as follows:

[0037] Generate random Gaussian noise And add it to the missing position of the channel state information data, that is:

[0038]

[0039] in, is the bitwise inversion of the mask m, ⊙ is the bitwise multiplication of the elements, is the channel state information with Gaussian noise added at the data missing position, and X0 is the initial channel state information with data missing;

[0040] Select a time step t and calculate the embedding of the current time step

[0041]

[0042] S condition As a condition, and S t As input, use the neural network to predict the added noise Then the data after t steps of denoising is The distribution is:

[0043]

[0044] Wherein, θ is the parameter of the neural network, Noise data The distribution of is the denoising result of the tkth time step.

[0045] in, Indicates and S condition The result after the condition is denoised by the neural network Satisfies the conditional Gaussian distribution, that is Conditional Expectation

[0046]

[0047] Among them, := means that β k is the noise level at step k, For neural networks, S condition As a condition, and S k is the noise predicted by input step t-k+1, is the covariance matrix and

[0048] Reconstructed channel state information

[0049] Preferably, the neural network comprises:

[0050] The network adopts a residual structure, which consists of N residual blocks. The input of each residual block includes the time step embedding S k , the feature embedding S condition and n=1,2,…,N three parts, among which, is the output of the n-1th residual block in the t-k+1th step denoising process

[0051] In the t-k+1th step denoising process, for the nth residual block, its output for:

[0052]

[0053] in, It is the intermediate variable calculated by the nth residual block according to the input in the denoising process of the t-k+1th step, The nth residual block in the t-k+1th step denoising process is based on the intermediate variable The Key, Query, and Value matrices in the calculated attention mechanism, Add(·) is the summation function, and Conv(·) is a 1×1 convolution operation. is the output of the nth residual block in the denoising process of the t-k+1th step; then the noise predicted in the t-k+1th step for:

[0054]

[0055] Predict the added noise, which is the network output X in the last step of denoising out for:

[0056]

[0057] Therefore, the noise at the location where the channel state information data is missing can be calculated based on the mask:

[0058]

[0059] The time series feature extraction unit, frequency domain feature extraction unit, cross fusion unit, mask fusion unit, and neural network are trained end-to-end in a self-supervised manner. The specific steps include:

[0060] Step 2.1: Collect channel state information at a high sampling rate As training data, randomly select the loss rate γ in the range of [10%, 90%] and generate the mask m∈{0,1} F×T×C , where the value of γ×T moments is 0 and the rest are 1, then the low sampling channel state information data X0=X hiigh ⊙m;

[0061] Step 2.2: Calculate the feature embedding S according to the feature embedding module condition ;

[0062] Step 2.3: Randomly sample diffusion time step t and generate random Gaussian noise Add the Gaussian noise to X hiigh Get

[0063]

[0064] in, Each component of ∈ is independent and identically distributed and follows a Gaussian distribution with a mean of 0 and a variance of 1, and I is the identity matrix;

[0065] in, β i is the noise level at step i, The mask m is bitwise inverted, and ⊙ is the bitwise multiplication of the elements;

[0066] Step 2.4: Based on the current time step embedding, predict the added noise through the neural network and calculate the predicted noise

[0067] Step 2.5: Calculate the loss And using the gradient descent method to update the model parameters of the feature embedding module and the data generation module;

[0068] described is the noise prediction loss, and its calculation expression is

[0069]

[0070] ∈ is the Gaussian noise added in step 2.3, ‖*‖2 is the L2 norm

[0071] described is the cyclic spectrum loss, and its calculation expression is

[0072]

[0073] in, and They are The spectral correlation function and The spectral correlation function, X′ high and X′ out are the high sampling rate channel state information X hiigh The data generation module denoises and reconstructs the channel state information X out According to the complex matrix composed of real and imaginary data, α is the cyclic frequency;

[0074] The spectral correlation function

[0075] Among them, (X′ high ) * Represents X′ high Perform conjugate operation on each element of (X o ' ut ) * Indicates X o ' ut Perform conjugate operation on each element of ;

[0076] Among them, X′ high [k,1:T,1] means taking X′ high The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ high ) * [k-α,1:T,1] means taking (X′ high ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier, X o 'ut [k,1:T,1] means taking X o ' ut The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X o ' ut ) * [k-α,1:T,1] means taking (X o ' ut ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier;

[0077] Step 2.6: Repeat steps 2.2 to 2.5 until the training is complete.

[0078] Preferably, the step 3 calculates the spectral correlation function according to the high sampling rate channel state information data and performs interference detection, specifically as follows:

[0079] The reconstructed channel state information X is calculated according to the calculation formula of the spectral correlation function o ' ut Spectral correlation function And calculate the magnitude of each element;

[0080] The difference between each element and the adjacent element in the spectrum correlation function amplitude is calculated to obtain a gradient matrix of the spectrum correlation function amplitude, and an appropriate threshold is selected to detect whether the channel state information is interfered.

[0081] Perform singular value decomposition on the gradient matrix. If the largest singular value is greater than the threshold, there is interference; otherwise, there is no interference.

[0082] The technical solution of the system of the present invention is a channel state information reconstruction system for channel state interference detection, comprising:

[0083] A feature embedding module is used to generate effective feature embedding based on the collected channel state information, wherein the channel state information data is low-sampling rate channel state information data collected by the WiFi signal receiving end;

[0084] a data generation module, configured to generate channel state information data at missing locations using a diffusion model based on the feature embedding, wherein the channel state information data at missing locations and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data;

[0085] An interference detection module, configured to calculate a spectral correlation function based on the high sampling rate channel state information data and perform interference detection;

[0086] The low-sampling WiFi channel state information reconstruction method for channel state information interference detection provided by the present invention uses an end-to-end self-supervised approach to train a neural network model, which can substantially reduce the workload of data collection and labeling. The present invention uses the characteristics of sampled data as a condition to guide the diffusion model for conditional data generation, which can substantially improve the performance of channel state information data reconstruction. The present invention introduces cyclic spectrum loss, which can fully meet the requirements of interference detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 : A schematic diagram of a method flow diagram of an embodiment of the present invention;

[0088] Figure 2 : Schematic diagram of a time series feature extraction unit and a frequency domain feature extraction unit according to an embodiment of the present invention;

[0089] Figure 3 : A schematic diagram of a cross-fusion unit according to an embodiment of the present invention;

[0090] Figure 4 : Schematic diagram of a neural network structure for predicting noise according to an embodiment of the present invention;

[0091] Figure 5 : System schematic diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0092] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0093] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0094] In this embodiment of the present invention, the transmitter is a WiFi router equipped with three antennas, and the receiver is a desktop computer equipped with three antennas and an Intel 5300 network card. The transmitter and receiver are 2 meters apart, and the interference signals include Bluetooth and ZigBee interference. The receiver is equipped with an independent physical server for running the method proposed in the technical solution of the present invention. In some embodiments, the initial channel state information data collected at the receiver can be other WiFi communication devices with channel state information acquisition capabilities, and are not limited to the above configuration.

[0095] The following is combined with Figure 1-5 The specific embodiment of the present invention is a channel state information reconstruction method and system for channel state interference detection, which is as follows:

[0096] Please see Figure 1 , Figure 1 : is a flow chart of a method according to an embodiment of the present invention. The embodiment of the method according to the present invention is a channel state information reconstruction method for channel state interference detection, which is specifically as follows:

[0097] Step 1: Generate effective feature embedding based on the collected channel state information, where the channel state information data is low-sampling rate channel state information data collected by the WiFi signal receiving end;

[0098] The channel state information collected in step 1 is as follows:

[0099] For a WiFi channel with M = 3 transmitting antennas, N = 3 receiving antennas, and F = 30 subcarriers, the channel state information data at T = 1000 times is collected to form the low-sampling rate channel state information. The low-sampling rate channel state information of each transmitting and receiving antenna pair can be regarded as a three-dimensional matrix of shape F × T × 2. This matrix is ​​composed of two two-dimensional matrices of shape F × T, which respectively include the real part data and imaginary part data of the low-sampling rate channel state information of the transmitting and receiving antenna pair.

[0100] Initial channel state information

[0101] Where, C = 2 × M × N;

[0102] Among them, the channel state information value at the moment when the sampled data is missing is 0, m∈{0,1} F×T×C is the channel state information data mask, the value of m is 0 when data is missing, otherwise it is 1;

[0103] Generate valid feature embeddings as described in step 1, as follows:

[0104] Temporal feature extraction unit, such as Figure 2 As shown, the self-attention block is used to extract the temporal features s of the i-th subcarrier data of the channel state information data. i , then the time series characteristics

[0105]

[0106]

[0107] Among them, LP is the linear projection operation, Norm is the layer normalization operation, X i,:,:represents the data of all transmit and receive antenna pairs at all times of the i-th subcarrier, and softmax is the softmax function:

[0108]

[0109] x n is the nth element of vector x, X i,:,: The corresponding Query, Key, Value matrix in the attention mechanism, the hyperparameter d k is the scaling factor, is the set of real numbers, represents a real matrix of size F×T×C;

[0110] Frequency domain feature extraction unit, such as Figure 2 As shown, the self-attention block is used to extract the frequency domain feature s of the j-th moment data of the channel state information data. j ,

[0111]

[0112] Among them, the hyperparameter d k is the scaling factor, X :,j,: Represents the data of all subcarriers and all transmit and receive antenna pairs at the jth moment;

[0113] X :,j,: The corresponding Query, Key, and Value matrices in the attention mechanism, represents a real matrix of size (F×1×C)×T;

[0114] Cross-fusion units, such as Figure 2 As shown, it is used to extract features between the time domain and the frequency domain, and the cross attention block is used to fuse the time series features and the frequency domain features to obtain the time-frequency features.

[0115] K3,V3=LP(Norm(S freq ))

[0116] Q3=LP(Norm(S time ))

[0117]

[0118] Among them, the hyperparameter d k is the scaling factor,

[0119] K3 and V3 are the Key and Value matrices in the attention mechanism corresponding to the frequency domain features, respectively, and Q3 is the Query matrix in the attention mechanism corresponding to the time series features;

[0120] The mask fusion unit is used to fuse the time-frequency features and the channel state information data mask to obtain effective feature embedding

[0121] S condition =SiLU(Concat(S cross ,m))

[0122] Among them, Concat represents data concatenation, the activation function SiLu(·) is a sigmoid weighted linear unit, m∈{0,1} F×T×C is the channel state information data mask. The value of m is 0 when data is missing and 1 otherwise.

[0123] Step 2: Generate channel state information data at the missing position using a diffusion model, wherein the channel state information data at the missing position and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data;

[0124] The channel state information data of the missing position is generated using the diffusion model described in step 2 as follows:

[0125] Generate random Gaussian noise And add it to the missing position of the channel state information data, that is:

[0126]

[0127] in, is the bitwise inversion of the mask m, ⊙ is the bitwise multiplication of the elements, is the channel state information with Gaussian noise added at the data missing position, and X0 is the initial channel state information with data missing;

[0128] Select a time step t=50 and calculate the embedding of the current time step

[0129]

[0130] S condition As a condition, and S t As input, use the neural network to predict the added noise Then the data after t steps of denoising is The distribution is:

[0131]

[0132] Wherein, θ is the parameter of the neural network, Noise data The distribution of is the denoising result of the tkth time step.

[0133] in, Indicates and S condition The result after the condition is denoised by the neural network Satisfies the conditional Gaussian distribution, that is Conditional Expectation

[0134]

[0135] Among them, := means that β k is the noise level at step k, For neural networks, S condition As a condition, and S k is the noise predicted by input step t-k+1, is the covariance matrix and

[0136] Reconstructed channel state information

[0137] The neural network comprises:

[0138] The network adopts a residual structure, such as Figure 4 As shown, it consists of N = 4 residual blocks, and the input of each residual block includes the time step embedding S k , the feature embedding S condition and n=1,2,...,N three parts, among which, is the output of the n-1th residual block in the t-k+1th step denoising process

[0139] In the t-k+1th step denoising process, for the nth residual block, its output for:

[0140]

[0141]

[0142]

[0143] in, It is the intermediate variable calculated by the nth residual block according to the input in the denoising process of the t-k+1th step, The nth residual block in the t-k+1th step denoising process is based on the intermediate variable The Key, Query, and Value matrices in the calculated attention mechanism, Add(·) is the summation function, and Conv(·) is a 1×1 convolution operation. is the output of the nth residual block in the denoising process of the t-k+1th step; then the noise predicted in the t-k+1th step for:

[0144]

[0145] Predict the added noise, which is the network output X in the last step of denoising out for:

[0146]

[0147] Therefore, the noise at the location where the channel state information data is missing can be calculated based on the mask:

[0148]

[0149] The time series feature extraction unit, frequency domain feature extraction unit, cross fusion unit, mask fusion unit, and neural network are trained end-to-end in a self-supervised manner. The specific steps include:

[0150] Step 2.1: Collect channel state information at a high sampling rate As training data, randomly select the loss rate γ in the range of [10%, 90%] and generate the mask m∈{0,1} F×T×C , where the value of γ×T moments is 0 and the rest are 1, then the low sampling channel state information data X0=X high ⊙m;

[0151] Step 2.2: Calculate the feature embedding S according to the feature embedding module condition ;

[0152] Step 2.3: Randomly sample diffusion time step t and generate random Gaussian noise Add the Gaussian noise to X high Get

[0153]

[0154] in, Each component of ∈ is independent and identically distributed and follows a Gaussian distribution with a mean of 0 and a variance of 1, and I is the identity matrix;

[0155] in, β i is the noise level at step i, The mask m is bitwise inverted, and ⊙ is the bitwise multiplication of the elements;

[0156] Step 2.4: Based on the current time step embedding, calculate the predicted noise through the neural network and calculate the predicted noise

[0157] Step 2.5: Calculate the loss And using the gradient descent method to update the model parameters of the feature embedding module and the data generation module;

[0158] described is the noise prediction loss, and its calculation expression is

[0159]

[0160] ∈ is the Gaussian noise added in step 2.3, ‖*‖2 is the L2 norm

[0161] described is the cyclic spectrum loss, and its calculation expression is

[0162]

[0163] in, and They are The spectral correlation function and The spectral correlation function, X′ high and X′ out are the high sampling rate channel state information X high The data generation module denoises and reconstructs the channel state information X out According to the complex matrix composed of real and imaginary data, α is the cyclic frequency;

[0164] The spectral correlation function

[0165] Among them, (X′ high ) * Represents X′ high Perform conjugate operation on each element of (X o ' ut ) * Indicates X o ' ut Perform conjugate operation on each element of ;

[0166] Among them, X′ high [k,1:T,1] means taking X′ highThe data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ high ) * [k-α,1:T,1] means taking (X′ high ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier, X o ' ut [k,1:T,1] means taking X o ' ut The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X o ' ut ) * [k-α,1:T,1] means taking (X o ' ut ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier;

[0167] Step 2.6: Repeat steps 2.2 to 2.5 until the training is complete.

[0168] Step 3: Calculating a spectral correlation function based on the high sampling rate channel state information data and performing interference detection;

[0169] Step 3 calculates the spectral correlation function based on the high sampling rate channel state information data and performs interference detection, as follows:

[0170] The reconstructed channel state information X is calculated according to the calculation formula of the spectral correlation function o ' ut Spectral correlation function And calculate the magnitude of each element;

[0171] The difference between each element and the adjacent element in the spectrum correlation function amplitude is calculated to obtain a gradient matrix of the spectrum correlation function amplitude, and an appropriate threshold is selected to detect whether the channel state information is interfered.

[0172] Perform singular value decomposition on the gradient matrix. If the largest singular value is greater than the threshold, there is interference; otherwise, there is no interference.

[0173] The system embodiment of the present invention is a channel state information reconstruction system for channel state interference detection, comprising:

[0174] like Figure 5 As shown, the feature embedding module is used to generate effective feature embedding according to the collected channel state information, where the channel state information data is low sampling rate channel state information data collected by the WiFi signal receiving end;

[0175] a data generation module, configured to generate channel state information data at missing locations using a diffusion model based on the feature embedding, wherein the channel state information data at missing locations and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data;

[0176] An interference detection module, configured to calculate a spectral correlation function based on the high sampling rate channel state information data and perform interference detection;

[0177] Table 1: Comparison of experimental results

[0178]

[0179] Table 1 compares the experimental results of the method proposed in the present invention with those of the existing technology. It respectively calculates the mean square error (MSE) of the channel state information reconstructed using linear interpolation, generative adversarial network (GAN), and the diffusion model method adopted by the present invention. It also calculates the accuracy of interference detection using the original low-sampling rate data and the channel state information reconstructed by the three technologies.

[0180] Experimental results show that the proposed channel state information reconstruction method and system have lower reconstruction errors than existing technologies and can accurately reconstruct low-sampling-rate channel state information. Furthermore, the reconstructed channel state information can better preserve spectral information, improving interference detection accuracy.

[0181] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0182] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A channel state information reconstruction method for channel state interference detection, characterized in that: The following steps are involved: Step 1: Generate effective feature embedding based on the collected channel state information, where the channel state information data is low-sampling rate channel state information data collected by the WiFi signal receiving end; Step 2: Generate channel state information data at the missing position using a diffusion model, wherein the channel state information data at the missing position and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data; Step 3: Calculating a spectral correlation function based on the high sampling rate channel state information data and performing interference detection; The time series feature extraction unit, frequency domain feature extraction unit, cross fusion unit, mask fusion unit, and neural network are trained end-to-end using a self-supervised approach. The specific steps include: Step 2.1: Collect channel state information at a high sampling rate As training data, randomly select the loss rate γ in the range of [10%, 90%] and generate the mask m∈{0,1} F×T×C , where the value of γ×T moments is 0 and the rest are 1, then the low sampling channel state information data X0=X high ⊙m; Step 2.2: Calculate feature embedding S according to the feature embedding module condition ; Step 2.3: Randomly sample diffusion time step t and generate random Gaussian noise Add the Gaussian noise to X high Get in, Each component of ∈ is independent and identically distributed and follows a Gaussian distribution with a mean of 0 and a variance of 1, and I is the identity matrix; in, β i is the noise level at step i, The mask m is bitwise inverted, and ⊙ is the bitwise multiplication of the elements; Step 2.4: Based on the current time step embedding, calculate the predicted noise through the neural network Step 2.5: Calculate the loss And using the gradient descent method to update the model parameters of the feature embedding module and the data generation module; described is the noise prediction loss, and its calculation expression is ∈ is the Gaussian noise added in step 2.3, ‖*‖2 is the L2 norm described is the cyclic spectrum loss, and its calculation expression is in, and They are The spectral correlation function and The spectral correlation function, X′ high and X′ out are the high sampling rate channel state information X high The data generation module denoises and reconstructs the channel state information X out According to the complex matrix composed of real and imaginary data, α is the cyclic frequency; The spectral correlation function Among them, (X′ high ) * Represents X′ high Each element of out ) * Represents X′ out Perform conjugate operation on each element of ; Among them, X′ high [k,1:T,1] means taking X′ high The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ high ) * [k-α,1:T,1] means taking (X′ high ) * The data of the first transmitting and receiving antenna pair at time 1 to T of the k-αth subcarrier, X′ out [k,1:T,1] means taking X′ out The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ out ) * [k-α,1:T,1] means taking (X′ out ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier; Step 2.6: Repeat steps 2.2 to 2.5 until the training is complete.

2. The channel state information reconstruction method for channel state interference detection according to claim 1, characterized in that: The channel state information collected in step 1 is as follows: Initial channel state information Where, C = 2 × M × N; Among them, the channel state information value at the moment when the sampled data is missing is 0, m∈{0,1} F×T×C is the channel state information data mask. The value of m is 0 when data is missing and 1 otherwise.

3. The channel state information reconstruction method for channel state interference detection according to claim 2, characterized in that: Generate valid feature embeddings as described in step 1, as follows: The generation of effective feature embedding is achieved through a temporal feature extraction unit, a frequency domain feature extraction unit, a cross fusion unit, and a mask fusion unit.

4. The channel state information reconstruction method for channel state interference detection according to claim 3, characterized in that: The temporal feature extraction unit uses the self-attention block to extract the temporal feature s from the i-th subcarrier data of the channel state information data i , then the time series characteristics Among them, LP is the linear projection operation, Norm is the layer normalization operation, X i,:,: represents the data of all transmit and receive antenna pairs at all times of the i-th subcarrier, and softmax is the softmax function: x n is the nth element of vector x, X i,:,: The corresponding Query, Key, Value matrix in the attention mechanism, the hyperparameter d k is the scaling factor, is the set of real numbers, represents a real matrix of size F×T×C; The frequency domain feature extraction unit uses the self-attention block to extract the frequency domain feature s of the j-th moment data of the channel state information data j , Among them, the hyperparameter d k is the scaling factor, X :,j,: Represents the data of all subcarriers and all transmit and receive antenna pairs at the jth moment; X :,j,: The corresponding Query, Key, and Value matrices in the attention mechanism, represents a real matrix of size (F×1×C)×T; Cross-fusion unit, used to extract features between time domain and frequency domain, and use cross-attention block to fuse the time series features and the frequency domain features to obtain time-frequency features K3,V3=LP(Norm(S freq )) Q3=LP(Norm(S time )) Among them, the hyperparameter d k is the scaling factor, K3 and V3 are the Key and Value matrices in the attention mechanism corresponding to the frequency domain features, respectively, and Q3 is the Query matrix in the attention mechanism corresponding to the time series features; The mask fusion unit is used to fuse the time-frequency features and the channel state information data mask to obtain effective feature embedding With condition =SiLU(Concat(S cross ,m)) Among them, Concat represents data concatenation, the activation function SiLu(·) is a sigmoid weighted linear unit, m∈{0,1} F×T×C is the channel state information data mask. The value of m is 0 when data is missing and 1 otherwise.

5. The channel state information reconstruction method for channel state interference detection according to claim 4, characterized in that: The channel state information data of the missing position is generated using the diffusion model described in step 2 as follows: Generate random Gaussian noise And add it to the missing position of the channel state information data, that is: in, is the bitwise inversion of the mask m, ⊙ is the bitwise multiplication of the elements, is the channel state information with Gaussian noise added at the data missing position, and X0 is the initial channel state information with data missing; Select a time step t and calculate the embedding of the current time step S condition As a condition, and S t As input, use the neural network to predict the added noise Then the data after t steps of denoising is The distribution is: Wherein, θ is the parameter of the neural network, Noise data The distribution of is the denoising result of the tkth time step; in, Indicates and S condition The result after the condition is denoised by the neural network Satisfies the conditional Gaussian distribution, that is Conditional Expectation Among them, := means that β k is the noise level at step k, For neural networks, S condition As a condition, and S k is the noise predicted by input step t-k+1, is the covariance matrix and Reconstructed channel state information 6. The channel state information reconstruction method for channel state interference detection according to claim 5, characterized in that: The neural network comprises: The network adopts a residual structure, which consists of N residual blocks. The input of each residual block includes the time step embedding S k , the feature embedding S condition and n=1,2,…,N three parts, among which, is the output of the n-1th residual block in the t-k+1th step denoising process In the t-k+1th step denoising process, for the nth residual block, its output for: in, It is the intermediate variable calculated by the nth residual block according to the input in the denoising process of the t-k+1th step, The nth residual block in the t-k+1th step denoising process is based on the intermediate variable The Key, Query, and Value matrices in the calculated attention mechanism, Add(·) is the summation function, and Conv(·) is a 1×1 convolution operation. is the output of the nth residual block in the denoising process of the t-k+1th step; then the noise predicted in the t-k+1th step for: Predict the added noise, which is the network output X in the last step of denoising out for: Therefore, the noise at the location where the channel state information data is missing can be calculated based on the mask:

7. The channel state information reconstruction method for channel state interference detection according to claim 6, characterized in that: Step 3 calculates the spectral correlation function based on the high sampling rate channel state information data and performs interference detection, as follows: The reconstructed channel state information X′ is calculated according to the calculation formula of the spectral correlation function out Spectral correlation function And calculate the magnitude of each element; Calculating the difference between each element and adjacent elements in the spectral correlation function amplitude to obtain a gradient matrix of the spectral correlation function amplitude, and selecting an appropriate threshold to detect whether the channel state information is interfered with; Perform singular value decomposition on the gradient matrix. If the largest singular value is greater than the threshold, there is interference; otherwise, there is no interference.

8. A channel state information reconstruction system for channel state interference detection using the channel state information reconstruction method for channel state interference detection according to claim 1, comprising: A feature embedding module is used to generate effective feature embedding based on the collected channel state information, wherein the channel state information data is low-sampling rate channel state information data collected by the WiFi signal receiving end; a data generation module, configured to generate channel state information data at missing locations using a diffusion model based on the feature embedding, wherein the channel state information data at missing locations and the low sampling rate channel state information data together constitute reconstructed high sampling rate channel state information data; An interference detection module, configured to calculate a spectral correlation function based on the high sampling rate channel state information data and perform interference detection; The time series feature extraction unit, frequency domain feature extraction unit, cross fusion unit, mask fusion unit, and neural network are trained end-to-end using a self-supervised approach. The specific steps include: Step 2.1: Collect channel state information at a high sampling rate As training data, randomly select the loss rate γ in the range of [10%, 90%] and generate the mask m∈{0,1} F×T×C , where the value of γ×T moments is 0 and the rest are 1, then the low sampling channel state information data X0=X high ⊙m; Step 2.2: Calculate feature embedding S according to the feature embedding module condition ; Step 2.3: Randomly sample diffusion time step t and generate random Gaussian noise Add the Gaussian noise to X high Get in, Each component of ∈ is independent and identically distributed and follows a Gaussian distribution with a mean of 0 and a variance of 1, and I is the identity matrix; in, β i is the noise level at step i, The mask m is bitwise inverted, and ⊙ is the bitwise multiplication of the elements; Step 2.4: Based on the current time step embedding, calculate the predicted noise through the neural network Step 2.5: Calculate the loss And using the gradient descent method to update the model parameters of the feature embedding module and the data generation module; described is the noise prediction loss, and its calculation expression is: ∈ is the Gaussian noise added in step 2.3, ‖*‖2 is the L2 norm described is the cyclic spectrum loss, and its calculation expression is in, and They are The spectral correlation function and The spectral correlation function, X′ high and X′ out are the high sampling rate channel state information X high The data generation module denoises and reconstructs the channel state information X out According to the complex matrix composed of real and imaginary data, α is the cyclic frequency; The spectral correlation function Among them, (X′ high ) * Represents X′ high Each element of out ) * Represents X′ out Perform conjugate operation on each element of ; Among them, X′ high [k,1:T,1] means taking X′ high The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ high ) * [k-α,1:T,1] means taking (X′ high ) * The data of the first transmitting and receiving antenna pair at time 1 to T of the k-αth subcarrier, X′ out [k,1:T,1] means taking X′ out The data of the first transmitting and receiving antenna pair at the 1st to Tth time of the kth subcarrier, (X′ out ) * [k-α,1:T,1] means taking (X′ out ) * The data of the first transmitting and receiving antenna pair at time 1 to time T of the k-αth subcarrier; Step 2.6: Repeat steps 2.2 to 2.5 until the training is complete.

Images