Human body activity CSI identification method based on BLSTM-CNN network

Abstract

The invention relates to a human body activity CSI recognition method based on a BLSTM-CNN network, belongs to the field of intelligent sensing, and solves the problems that an existing machine learning model is large in calculation amount and a deep learning model is low in accuracy. Comprising the following steps: S1, collecting CSI original data H'when a target executes an action, and representing amplitude and phase information of different subcarriers at different moments; s2, obtaining amplitude matrixes F of different subcarriers at different moments according to a signal amplitude calculation formula on the basis of the CSI original data H '; s3, preprocessing the amplitude matrix F, and screening out an amplitude sub-matrix FT representing that the target is in an active state from the matrix amplitude F according to a preset condition; s4, inputting the amplitude sub-matrix FT into a trained classification model for feature extraction, and obtaining an activity category represented by the CSI original data based on the extracted features; and the classification model adopts a BLSTM-CNN network. The human body activity CSI identification method with simplified steps and high accuracy is realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent perception technology, and in particular to a human activity CSI recognition method based on a BLSTM-CNN network. Background Art

[0002] Compared with vision-based human activity recognition methods that are only applicable to line-of-sight environments and infringe on personal privacy, as well as wearable device-based human activity recognition methods with limited application scenarios, the recognition method based on Wi-Fi channel state information (CSI) has shown great application potential in the field of intelligent perception due to its advantages such as convenient data collection, no need for additional equipment deployment, and insensitivity to ambient light.

[0003] Channel state information (CSI) is a kind of wireless communication status data, which contains the amplitude and phase of the subcarrier in the orthogonal frequency division multiplexing system. When the target between the transmitter and the receiver is active, the channel condition of the wireless signal transmission changes, that is, the diffraction, reflection, scattering and other processes of the signal during transmission change with the change of the target activity, and these changes are reflected in the subcarrier received by the receiving end. The process of the CSI-based identification method is as follows: Figure 1 As shown, it generally includes: 1) collecting CSI data reflecting the target activity; 2) performing preprocessing such as denoising on the CSI data; 3) extracting features from the preprocessed data; 4) inputting the extracted features into a trained machine learning or deep learning model to classify and identify the target activity category.

[0004] In order to improve the accuracy of the CSI-based recognition method, the existing improvement ideas focus on considering the interference of experimental environment and conditions on wireless signals, and then try to filter the interference factors to avoid the influence of this factor, such as the interference of co-channel or adjacent channels between Wi-Fi signals, the interference of position changes on the recognition results when the target performs the same activity, and the interference of different antenna sensitivities on the recognition results; on the other hand, they focus on improving machine learning or deep learning training models. For example, the literature "A Fall Detection Method Based on CSI" and the literature "Daily Behavior Recognition Method Based on CSI" respectively use the AFSA modified SVM algorithm and the MIFS combined with PSO optimized SVM algorithm to recognize human actions, with an accuracy of 94.64% and 96.1%, respectively. However, the above machine learning algorithms are difficult to extract features, the training process is complicated and the cost is high; deep learning is gradually applied to human activity recognition methods because it can automatically extract features, including: 1D-CNN and BLSTM networks with a single structure. Due to the single structure, the recognition accuracy of such models is not high. In addition, patent number CN111079599B discloses a method for identifying human behavior based on a CNN-LSTM fusion network, which targets human activity signals collected by sensors or smart wearable devices, such as the UniMib dataset, which is different from the CSI data collected by the present invention. Therefore, the application scenarios of this model are limited.

[0005] In summary, the human activity recognition method based on Wi-Fi CSI data has broad application prospects, so it is very necessary to explore a deep learning algorithm with a simple training process and higher accuracy. Summary of the invention

[0006] In view of the above analysis, an embodiment of the present invention aims to provide a human activity CSI recognition method based on a BLSTM-CNN network, so as to solve the problems that the existing human activity CSI recognition method based on machine learning has a complex processing process and requires manual feature extraction, and the accuracy of the human activity CSI recognition method based on deep learning still needs to be improved.

[0007] On the one hand, an embodiment of the present invention provides a human activity CSI recognition method based on a BLSTM-CNN network, comprising:

[0008] Step S1, collecting CSI raw data H′ when the target performs an action, representing the amplitude and phase information of different subcarriers at different times;

[0009] Step S2, based on the CSI original data H′, according to the signal amplitude calculation formula, obtain the amplitude matrix F of different subcarriers at different times;

[0010] Step S3: pre-process the amplitude matrix F, and select the amplitude sub-matrix F representing the target being in an active state from the matrix amplitude F according to the preset conditions. T ;

[0011] Step S4: Substituting the amplitude submatrix F T The trained classification model is input to perform feature extraction, and based on the extracted features, the activity category represented by the CSI raw data is obtained; the classification model adopts a BLSTM-CNN network.

[0012] Specifically, the amplitude matrix F of different subcarriers at different times in step S2 is M×N in size, and the element f mn (m=1,2,…,M;n=1,2,…,N) represents the amplitude of subcarrier n at time m; f mn Obtained by the amplitude calculation formula:

[0013]

[0014] Among them, a mn ,b mn Obtained from the CSI original data matrix H′; the size of the matrix H′ is M×N, and the element h′ mn (m=1,2,…,M; n=1,2,…,N)=a mn +ib mn , i is an imaginary number.

[0015] Specifically, the step S3 selects the amplitude submatrix F representing the target being in an active state from the matrix amplitude F according to the preset conditions. T ,include:

[0016] Step S301: Use a sliding window with a window size of C×N to slide along the time dimension of the matrix F, starting from the time point m that meets the preset conditions, and select a C×N dimensional sub-matrix F that meets the preset conditions. C , where C<M; the subcarrier amplitude at the time point m that meets the preset condition satisfies:

[0017]

[0018] Among them, f mn ,f (m-1)n ,f (m-2)n Respectively represent the subcarrier amplitude at time m, time (m-1), and time (m-2);

[0019] Step S302: select the C×N dimensional sub-matrix F C Downsample and compress it into a T×N dimensional matrix F T ;in d (d = 1, 2, ..., S; S < < C) is the sampling frequency, that is, a value is extracted every d time points.

[0020] Specifically, the classification model described in step S4 includes a BLSTM network, a 1D-CNN network, a global average pooling layer, a fully connected layer and a Softmax layer; wherein the BLSTM network, the 1D-CNN network and the global average pooling layer are used to extract feature vectors, and the fully connected layer and the Softmax layer are used to output corresponding activity categories and probabilities according to the feature vectors.

[0021] Specifically, the structure of the 1D-CNN network is a first 1D convolutional layer, a first maximum pooling layer, a second 1D convolutional layer, a second maximum pooling layer, a third 1D convolutional layer and a third maximum pooling layer.

[0022] Specifically, the feature extraction using a BLSTM network, a 1D-CNN network and a global average pooling layer includes:

[0023] Step S401: transform the T×N dimensional matrix F T Decompose into T column matrices, any column matrix in, is the matrix F T The elements of the t-th row represent the amplitudes of the N subcarriers at time t;

[0024] Step S402: The BLSTM network includes T forward LSTM modules and T backward LSTM modules. t (t=1,2,…,T) are input into T forward LSTM modules in chronological order, and T forward hidden layer state matrices are output. At the same time, the T column matrices x t (t=1,2,…,T) are input into T backward LSTM modules in reverse time order, and T backward hidden layer state matrices are output The forward hidden layer state matrix and the backward hidden layer state matrix Merge to get T hidden layer state matrices h t (t=1,2,…,T), where,

[0025] Step S403: transform T column matrices h t After transposing each of them, they are merged into a T×2N-dimensional matrix H, which is used as the input matrix of the 1DCNN network; the T×2N-dimensional matrix H is processed by the first 1D convolution layer, the first maximum pooling layer, the second 1D convolution layer, the second maximum pooling layer, the third 1D convolution layer and the third maximum pooling layer in sequence, and then input into the global average pooling layer, and the output result is used as the feature vector.

[0026] Specifically, the training process of the classification model includes:

[0027] Step SS1: obtain CSI raw data corresponding to known active tags, and extract amplitude matrices of different subcarriers at different times;

[0028] Step SS2, selecting an amplitude submatrix within the range of the start time and the end time of the target execution activity from the amplitude matrix, and downsampling it to obtain a compressed amplitude submatrix; wherein the start time and the end time of the target execution activity are known;

[0029] Step SS3, decompose the compressed amplitude submatrix into a column matrix describing the amplitudes of all subcarriers at different times, input the matrix into the BLSTM network, obtain the hidden state matrix, input the hidden state matrix into the 1D-CNN network to obtain the matrix after three convolutional layers and three maximum pooling layers, input the global average pooling layer, and obtain the extracted feature vector; input the feature vector into the fully connected layer, and output the activity label represented by the input data;

[0030] Step SS4, calculate the category cross entropy error between the output label and the actual label in the previous step, and based on the category cross entropy error, use the gradient-based optimization method to perform back propagation to update the BLSTM-CNN network update parameters; until the output result is consistent with its actual label, fix the parameters of the network as the trained classification model.

[0031] Furthermore, in step SS4, based on the category cross entropy error, a gradient-based optimization method is used for back propagation, and in the process of updating the BLSTM-CNN network parameters, an ADAM optimizer and a dropout method are used to prevent overfitting.

[0032] Specifically, the CSI raw data corresponding to the known activity labels adopts Yousefi's 7-label dataset; the dataset includes CSI data corresponding to 7 types of activity labels performed by the target, namely, going to bed, falling, picking up things, running, sitting down, standing up and walking.

[0033] Specifically, the CSI raw data H′ when the target performs an action is collected by using a computer with an Intel 5000 network card and a Linux 802.11n CSI Tool installed.

[0034] Compared with the human activity CSI recognition method based on machine learning, the method of the present invention does not need to denoise the original data or manually extract features, which simplifies the calculation steps and reduces the calculation cost; compared with other human activity CSI recognition methods based on deep learning, the recognition accuracy of the method of the present invention is improved overall, that is, the performance is better.

[0035] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0037] Figure 1 A schematic diagram of the CSI human activity recognition process based on machine learning or deep learning;

[0038] Figure 2 This is a flow chart of the CSI human activity recognition method based on the BLSTM-CNN network;

[0039] Figure 3 Schematic diagram of the training process of the BLSTM-CNN network;

[0040] Figure 4 This is the algorithm principle diagram of the BLSTM network;

[0041] Figure 5 Schematic diagram of the process of extracting feature vectors for 1D-CNN network;

[0042] Figure 6 Schematic diagram of the collection scenario of CSI raw data;

[0043] Figure 7 The recognition accuracy of the BLSTM-CNN network for 7 types of human activities;

[0044] Figure 8 The recognition accuracy of 1D-CNN network for 7 types of human activities;

[0045] Fig. 9 The recognition accuracy of the BLSTM-Attention network for seven types of human activities;

[0046] Fig.10 It is the recognition accuracy of BLSTM network for 7 types of human activities. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0048] A specific embodiment of the present invention discloses a human activity CSI recognition method based on a BLSTM-CNN network, such as Figure 2 As shown, including:

[0049] Step S1, collecting CSI raw data H′ when the target performs an action, representing the amplitude and phase information of different subcarriers at different times; the target refers to a human body performing an action.

[0050] Step S2, based on the CSI original data H′, according to the signal amplitude calculation formula, obtain the amplitude matrix F of different subcarriers at different times;

[0051] Step S3: pre-process the amplitude matrix F, and select the amplitude sub-matrix F representing the target being in an active state from the matrix amplitude F according to the preset conditions. T ;

[0052] Step S4: Substituting the amplitude submatrix F T The trained classification model is input to perform feature extraction, and based on the extracted features, the activity category represented by the CSI raw data is obtained; the classification model adopts a BLSTM-CNN network.

[0053] Specifically, the CSI raw data collection device includes a wireless signal transceiver, wherein the transceiver includes a transceiver and a transceiver antenna, and the transceiver is a computer with an Intel 5000 network card, which is 3m to 5m apart, and the receiver is equipped with Linux 802.11n CSI Tool software.

[0054] Specifically, in any scenario, when the target performs a certain action, the CSI raw data H′ is of size M×N, and the element h′ mn (m=1,2,…,M; n=1,2,…,N)=a mn +ib mn , i is an imaginary number; it represents the amplitude of N subcarriers at M moments. Exemplarily, the number of subcarriers in the present invention is: N=90.

[0055] Specifically, in step S2, based on the CSI raw data H′, the signal amplitude is calculated according to the following formula:

[0056]

[0057] Get the amplitude matrix F of different subcarriers at different times, the size is M×N, the element f mn (m=1,2,…,M; n=1,2,…,N) represents the amplitude of subcarrier n at time m.

[0058] Specifically, in step S3, according to the preset conditions, the amplitude sub-matrix F representing that the target is in an active state is screened out from the matrix amplitude F T , including:

[0059] Step S301: Use a sliding window with a window size of C×N to slide along the time dimension of the matrix F, and start from the time point m that meets the preset conditions to frame out a C×N-dimensional sub-matrix F that meets the preset conditions C , where C<M; the sub-carrier amplitude at the time point m that meets the preset conditions satisfies:

[0060]

[0061] where f mn , f (m-1)n , f (m-2)n respectively represent the sub-carrier amplitudes at time m, time (m - 1), and time (m - 2);

[0062] Step S302: Downsample the framed C×N-dimensional sub-matrix F C and compress it into a T×N-dimensional matrix F T ; where is the sampling frequency, that is, one value is extracted every d time points.

[0063] Exemplarily, the sliding window size adopted in the present invention is 1000×90, where M>>1000, and the size of the matrix representing that the target is in an active state framed out is 1000×90; the sampling frequency d = 2, and the size of the matrix F T obtained after downsampling is 500×90, representing the amplitudes of 90 sub-carriers at 500 time points.

[0064] Specifically, as Figure 3 shown, the classification model in step S4 is composed of a BLSTM network, a 1D-CNN network, a global average pooling layer, a fully connected layer, and a Softmax layer; where the BLSTM network, the 1D-CNN network, and the global average pooling layer are used to extract feature vectors, and the fully connected layer and the Softmax layer are used to output the corresponding activity categories and probabilities according to the feature vectors.

[0065] Specifically, the BLSTM network includes T forward LSTM modules and T backward LSTM modules; the 1D-CNN network includes three 1D convolutional layers and three max pooling layers.

[0066] Using the BLSTM network, the 1D-CNN network, and the global average pooling layer to extract feature vectors, as Figure 4 and Figure 5 shown, including:

[0067] Step S401: transform the T×N dimensional matrix F T Decompose into T column matrices, any column matrix in, is the matrix F T The elements of the tth row represent the amplitudes of the N subcarriers at time t;

[0068] Step S402: Figure 4 As shown, the T column matrices x t (t=1,2,…,T) are input into T forward LSTM modules in chronological order, and T forward hidden layer state matrices are output. At the same time, the T column matrices x t (t=1,2,…,T) are input into T backward LSTM modules in reverse time order, and T backward hidden layer state matrices are output The forward hidden layer state matrix and the backward hidden layer state matrix Merge to get T hidden layer state matrices h t (t=1,2,…,T), where,

[0069] Step S403: transform T column matrices h t After transposing each, they are merged into a T×2N dimensional matrix H, which serves as the input matrix of the 1DCNN network; Figure 5 As shown in FIG. 1 , the T×2N dimensional matrix H is processed by the first 1D convolution layer, the first maximum pooling layer, the second 1D convolution layer, the second maximum pooling layer, the third 1D convolution layer and the third maximum pooling layer in sequence, and then input into the global average pooling layer, and the output result is used as the feature vector.

[0070] Exemplarily, the amplitude submatrix F input to the BLSTM network in the present invention is T The size is 500×90, and the size of the matrix H output by the BLSTM network is 500×180; Figure 5 As shown in the figure, the matrix output by the 1D-CNN network is 15×128, which is input into the global average pooling layer to obtain a 1×128 feature vector, which is the feature vector automatically extracted by the BLSTM and 1D-CNN networks.

[0071] Specifically, the training process of the BLSTM-CNN classification model is as follows Figure 3 As shown, including:

[0072] Step SS1: Obtain the corresponding CSI raw data H of the known active tag 0 , get the amplitude matrix F of different subcarriers at different times 0 ;

[0073] Step SS2: Select the amplitude submatrix representing the start time and end time of the target execution activity from the above amplitude matrix, and downsample it to obtain a compressed amplitude submatrix The starting time and ending time of the target execution activity are known;

[0074] Step SS3, decompose the compressed amplitude submatrix into a column matrix describing the amplitudes of all subcarriers at different times, input the matrix into the BLSTM network, obtain the hidden state matrix, input the hidden state matrix into the 1D-CNN network to obtain the matrix after three convolutional layers and three maximum pooling layers, input the global average pooling layer, and obtain the extracted feature vector; input the feature vector into the fully connected layer, and output the activity label represented by the input data;

[0075] Step SS4, calculate the category cross entropy error between the output label and the actual label in the previous step, and based on the category cross entropy error, use the gradient-based optimization method to perform back propagation to update the parameters of the BLSTM-CNN network; until the output result is consistent with its actual label, fix the parameters of the network as the trained classification model.

[0076] For example, the present invention selects the 7-label dataset of the public dataset Yousefi for training, and its collection environment is as follows: Figure 6 As shown: the data is collected indoors without obstacles, that is, in a LOS channel environment, the transceiver is on the same horizontal plane, 3 meters apart, there is one transmitting antenna, three receiving antennas, and the transceivers are all computers with Intel 5000 network cards. The receiver is installed with Linux 802.11n CSI Tool to complete the collection of CSI data. The target is between the transceiver and the receiver to perform activities; the data set includes: the target performs CSI data corresponding to 7 types of activity labels: going to bed, falling, picking up things, running, sitting down, standing up and walking; specifically, the present invention uses 90% of the data set for training.

[0077] Preferably, when training the classification model, the gradient-based optimization method is used for back propagation based on the category cross entropy error in step SS4, and in the process of updating the parameters of the BLSTM-CNN network, an ADAM optimizer and a dropout method are used to prevent overfitting; wherein, the ADAM optimizer can update the variables according to the oscillation of the historical gradient and the true historical gradient after filtering the oscillation, and can effectively calculate the adaptive learning rate of each parameter in the model during the optimization process, which can adapt to sparse gradients and alleviate the problem of gradient oscillation.

[0078] Exemplarily, in order to verify the performance of the method of the present invention, the remaining 10% of the data in the above-mentioned Yousefi 7-label data set is used as test data, and the trained BLSTM-CNN network, 1D-CNN network, BLSTM-Attention network and BLSTM network of the present invention are respectively input to perform human activity recognition, and the performance of different models is compared through the recognition results.

[0079] The recognition results of BLSTM-CNN network, 1D-CNN network, BLSTM-Attention network and BLSTM network for the above 7-label datasets are as follows: Figure 7 As shown in (a), (b), (c) and (d), it can be seen from the figure that the recognition accuracy of the data of 7 labels by the method of the present invention exceeds 96%, and the accuracy is improved compared with the other three types of models; the overall average accuracy of the above four networks for 7-label data is shown in Table 1. It can be seen that compared with the other three networks, the accuracy of the method of the present invention can be improved by 2% to 5%.

[0080] Network Model Accuracy BLSTM 92％ 1D-CNN 96.14％ BLSTM-Attention 97.28％ BLSTM-CNN 98.28％

[0081] Table 1

[0082] In general, the BLSTM and 1D-CNN networks are applied to CSI human activity recognition. Compared with machine learning methods, the method of the present invention does not need to perform denoising and other processing on CSI data, and does not need to perform manual feature extraction, which greatly reduces the workload; in addition, it can be seen from the test results that compared with other deep learning methods, the average recognition accuracy of the method of the present invention is higher. Therefore, the method of the present invention has broad application prospects in the field of human activity recognition.

[0083] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0084] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A human activity CSI recognition method based on BLSTM-CNN network, characterized in that: include: Step S1, collecting CSI raw data H′ when the target performs an action, representing the amplitude and phase information of different subcarriers at different times; Step S2, based on the CSI original data H′, according to the signal amplitude calculation formula, obtain the amplitude matrix F of different subcarriers at different times; Step S3: preprocess the amplitude matrix F, and select the amplitude submatrix F representing the target being in an active state from the matrix amplitude F according to the preset conditions. T ; Step S4: Substituting the amplitude submatrix F T The trained classification model is input to perform feature extraction, and based on the extracted features, the activity category represented by the CSI raw data is obtained; the classification model adopts a BLSTM-CNN network.

2. According to the BLSTM-CNN network-based human activity CSI recognition method of claim 1, it is characterized in that: The amplitude matrix F of different subcarriers at different times in step S2 is M×N in size, and the element f mn represents the amplitude of subcarrier n at time m, where m = 1, 2, ..., M, n = 1, 2, ..., N; f mn Obtained by the amplitude calculation formula: Among them, a mn , b mn Obtained from the CSI original data matrix H′; the size of the matrix H′ is M×N, and the element h′ mn =a mn +ib mn , i is an imaginary number.

3. According to the BLSTM-CNN network-based human activity CSI recognition method of claim 2, it is characterized in that: The step S3 selects the amplitude submatrix F representing the target being in an active state from the matrix amplitude F according to the preset conditions. T ,include: Step S301: Use a sliding window with a window size of C×N to slide along the time dimension of the matrix F, starting from the time point m that meets the preset conditions, and select a C×N dimensional sub-matrix F that meets the preset conditions. C , where C<M; the subcarrier amplitude at the time point m that meets the preset condition satisfies: Among them, f mn , f (m-1)n , f (m-2)n Respectively represent the subcarrier amplitudes at time m, time m-1, and time m-2; Step S302: select the C×N dimensional sub-matrix F C Downsample and compress it into a T×N dimensional matrix F T ;in The sampling frequency d=1, 2, ..., S, S<<C, that is, a value is extracted every d time points.

4. According to the BLSTM-CNN network-based human activity CSI recognition method of claim 3, it is characterized in that: The classification model described in step S4 includes a BLSTM network, a 1D-CNN network, a global average pooling layer, a fully connected layer and a Softmax layer; wherein the BLSTM network, the 1D-CNN network and the global average pooling layer are used to extract feature vectors, and the fully connected layer and the Softmax layer are used to output corresponding activity categories and probabilities according to the feature vectors.

5. According to the BLSTM-CNN network-based human activity CSI recognition method of claim 3, it is characterized in that: The structure of the 1D-CNN network is sequentially a first 1D convolutional layer, a first maximum pooling layer, a second 1D convolutional layer, a second maximum pooling layer, a third 1D convolutional layer and a third maximum pooling layer.

6. The human activity CSI recognition method based on BLSTM-CNN network according to claim 5 is characterized in that: The feature extraction using the BLSTM network, 1D-CNN network and global average pooling layer includes: Step S401: transform the T×N dimensional matrix F T Decompose into T column matrices, any column matrix Wherein, t=1, 2, ..., T; is the matrix F T The elements of the tth row represent the amplitudes of the N subcarriers at time t; Step S402: The BLSTM network includes T forward LSTM modules and T backward LSTM modules. t Input into T forward LSTM modules in chronological order, and output T forward hidden layer state matrices At the same time, the T column matrices x t Input them into T backward LSTM modules in reverse time order, and output T backward hidden layer state matrices The forward hidden layer state matrix and the backward hidden layer state matrix Merge to get T hidden layer state matrices h t ,in, Step S403: transform T column matrices h t After transposing each of them, they are merged into a T×2N-dimensional matrix H, which is used as the input matrix of the 1DCNN network; the T×2N-dimensional matrix H is processed by the first 1D convolution layer, the first maximum pooling layer, the second 1D convolution layer, the second maximum pooling layer, the third 1D convolution layer and the third maximum pooling layer in sequence, and then input into the global average pooling layer, and the output result is used as the feature vector.

7. The human activity CSI recognition method based on BLSTM-CNN network according to claim 6 is characterized in that: The training process of the classification model includes: Step SS1: obtain CSI raw data corresponding to known active tags, and extract amplitude matrices of different subcarriers at different times; Step SS2, selecting an amplitude submatrix within the range of the start time and the end time of the target execution activity from the amplitude matrix, and downsampling it to obtain a compressed amplitude submatrix; wherein the start time and the end time of the target execution activity are known; Step SS3, decompose the compressed amplitude submatrix into a column matrix describing the amplitudes of all subcarriers at different times, input the matrix into the BLSTM network, obtain the hidden state matrix, input the hidden state matrix into the 1D-CNN network, input the matrix output by the 1D-CNN network into the global average pooling layer, and obtain the extracted feature vector; input the feature vector into the fully connected layer, and output the activity label represented by the input data; Step SS4, calculate the category cross entropy error between the output label of step SS3 and the actual label, based on the category cross entropy error, use the gradient-based optimization method to perform back propagation, and update the parameters of the BLSTM-CNN network; until the output result is consistent with its actual label, fix the parameters of the network, and obtain a trained classification model.

8. The human activity CSI recognition method based on BLSTM-CNN network according to claim 7 is characterized in that: In step SS4, based on the category cross entropy error, a gradient-based optimization method is used for back propagation. In the process of updating the BLSTM-CNN network parameters, an ADAM optimizer and a dropout method are used to prevent overfitting.

9. The human activity CSI recognition method based on BLSTM-CNN network according to claim 7 is characterized in that: The CSI raw data corresponding to the known activity labels adopts Yousefi's 7-label dataset; the dataset includes CSI data corresponding to 7 types of activity labels including the target performing going to bed, falling, picking up things, running, sitting down, standing up and walking.

10. The human activity CSI recognition method based on BLSTM-CNN network according to claim 1 is characterized in that: The acquisition of the CSI raw data H′ when the target performs an action is completed using a computer with an Intel 5000 network card and a Linux 802.11n CSI Tool installed.

Images