Csi action recognition method and system based on deep non-negative matrix factorization
By combining deep nonnegative matrix factorization and self-attention mechanism, the problem of insufficient global feature capture in CSI action recognition is solved, achieving high-precision action recognition and improving the system's performance and interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2023-08-14
- Publication Date
- 2026-06-02
AI Technical Summary
In existing CSI action recognition systems, traditional machine learning methods rely on human experience in feature extraction, while deep learning methods struggle to effectively capture the global features of actions, resulting in insufficient recognition accuracy, especially when the action lasts for a long time and information is severely lost.
A method combining deep nonnegative matrix factorization and self-attention mechanism is adopted. By calculating subcarrier amplitude and extracting features from CSI signals, the deep nonnegative matrix factorization method is used to mine subcarrier relationships, and a self-attention mechanism classifier is used to capture global features to achieve action recognition.
It improves the accuracy of action recognition, achieving a classification effect of 95%, avoids the shortcomings of manually designed features, reduces the number of parameters and memory usage, and enhances the interpretability and global feature capture capability of the method.
Smart Images
Figure CN117113140B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of human motion recognition technology, and in particular to a CSI motion recognition method and system based on deep nonnegative matrix factorization. Background Technology
[0002] The purpose of human action recognition is to automatically identify the type of human action by analyzing the obtained human action feature parameters, based on successful action tracking and feature extraction. Existing systems that use Channel State Information (CSI) to recognize actions generally fall into two categories. One type uses traditional machine learning methods, manually extracting features from the data, processing the data in various ways, and then using methods such as Support Vector Machine (SVM), k-Nearest Neighbor (KNN), and Random Forest to process the extracted features before finally classifying the action. The other type uses existing deep learning techniques, performing less processing on the raw data, and mostly using modern neural networks such as Multiple Layers Machine (MLP), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) to allow the network itself to extract features from the data and classify the actions.
[0003] However, the classification performance of traditional machine learning methods largely depends on the experience of the researchers. Designing feature extraction algorithms requires considerable experience, posing significant challenges for many researchers in this field. Furthermore, with the advancements in deep learning, traditional machine learning methods have become less effective than deep learning methods in various tasks. However, current deep learning methods for CSI action recognition systems also have many problems. For example, using CNN algorithms, the receptive field depends on the size of the convolutional kernels designed by the designer to find local features, but it's difficult to focus on global features. Systems using RNN networks, such as the typical Long Short-Term Memory (LSTM) network, also struggle to focus on global features. If an action lasts for a long time, the correlation between each time point is high, and LSTM is prone to losing information from earlier time points, leading to information loss and incorrect action recognition. Summary of the Invention
[0004] To address the aforementioned technical problems, the present invention aims to provide a CSI action recognition method and system based on deep nonnegative matrix factorization. By combining deep nonnegative matrix factorization with a self-attention mechanism classifier, the method effectively captures global features of CSI signal data, thereby improving classification accuracy.
[0005] The first technical solution adopted in this invention is: a CSI action recognition method based on deep nonnegative matrix factorization, comprising the following steps:
[0006] Collect CSI signal data and perform subcarrier amplitude calculation on the CSI signal data to obtain the CSI subcarrier amplitude;
[0007] The subcarrier amplitude of the CSI is processed by deep nonnegative matrix factorization to obtain a low-dimensional feature coefficient matrix.
[0008] Position encoding is introduced, and a self-attention mechanism classifier is used to classify the low-dimensional feature coefficient matrix to obtain the CSI signal data classification result.
[0009] Furthermore, the step of acquiring CSI signal data and performing subcarrier amplitude calculation processing on the CSI signal data to obtain the CSI subcarrier amplitude specifically includes:
[0010] CSI signal data is acquired via Bluetooth module;
[0011] Based on the CSI signal data, the time value of each subcarrier in the CSI signal data is extracted to obtain the complex value of the CSI subcarrier.
[0012] The real and imaginary parts of the complex value of the CSI subcarrier are successively squared and squared to obtain the preliminary CSI subcarrier amplitude.
[0013] The subcarrier amplitude of the preliminary CSI is normalized to obtain the subcarrier amplitude of the CSI.
[0014] Furthermore, the step of extracting features from the subcarrier amplitude of the CSI using a deep nonnegative matrix factorization method to obtain a low-dimensional feature coefficient matrix specifically includes:
[0015] Set the number of layers in the depth nonnegative matrix factorization method;
[0016] The subcarrier amplitude of the CSI is subjected to NMF decomposition based on the aforementioned deep nonnegative matrix decomposition method to obtain the feature matrix and the first coefficient matrix of the first decomposition.
[0017] Based on the deep nonnegative matrix decomposition method, the first coefficient matrix is subjected to a second NMF decomposition to obtain the feature matrix and the second coefficient matrix of the second decomposition.
[0018] The NMF decomposition process described above is performed on the second coefficient matrix until the number of NMF decompositions reaches the number of NMF layers, at which point the NMF decomposition stops.
[0019] The low-dimensional feature coefficient matrix is constructed based on the feature matrix corresponding to each NMF decomposition and the coefficient matrix corresponding to the last NMF decomposition.
[0020] Furthermore, the low-dimensional feature coefficient matrix comprises several feature matrices and a coefficient matrix, and its expression is as follows:
[0021] X≈W1W2W3...W n H n
[0022] In the above formula, W1W2W3...W n H represents the characteristic matrix of the nth level decomposition. n Let X represent the coefficient matrix of the nth level decomposition, and let X represent the low-dimensional characteristic coefficient matrix.
[0023] Furthermore, the depth nonnegative matrix factorization (NMF) method has three layers, wherein:
[0024] Based on the first-layer NMF, it is used to obtain the relationship between subcarriers of multiple CSIs at each time point;
[0025] Based on the second-layer NMF, abstract feature extraction is performed according to the relationship between the subcarriers of multiple CSIs at each time point;
[0026] The third-layer NMF is used to perform data dimensionality reduction processing on the abstract features.
[0027] Furthermore, the step of introducing position encoding and classifying the low-dimensional feature coefficient matrix using a self-attention mechanism classifier to obtain the CSI signal data classification result specifically includes:
[0028] A positional encoding is introduced, and the positional encoding is added to the low-dimensional feature coefficient matrix to obtain the added low-dimensional feature coefficient matrix.
[0029] The summed low-dimensional feature coefficient matrix is input into the self-attention mechanism classifier for classification processing, and the CSI signal data classification result is output, which represents the different actions in the collected CSI signal data.
[0030] Furthermore, the loss function of the self-attention mechanism classifier is the cross-entropy loss function, and the update function of the self-attention mechanism classifier is the Adam function.
[0031] Furthermore, the step of inputting the summed low-dimensional feature coefficient matrix into the self-attention mechanism classifier for classification processing and outputting the CSI signal data classification result specifically includes:
[0032] The summed low-dimensional feature coefficient matrix is input into the self-attention mechanism classifier, which includes a Self-Attention layer, a first normalization layer, an MLP network, and a Dropout layer.
[0033] Based on the Self-Attention layer of the self-attention mechanism classifier, feature extraction processing is performed on the summed low-dimensional feature coefficient matrix to obtain the temporal relationship of the CSI signal.
[0034] Based on the first normalization layer of the self-attention mechanism classifier, the temporal relationship of the CSI signal is normalized to obtain the normalized CSI signal temporal relationship.
[0035] Based on the MLP network of the self-attention mechanism classifier, the time relationship of the normalized CSI signal is identified and processed, and a preliminary CSI signal data classification result is output.
[0036] Based on the Dropout layer of the self-attention mechanism classifier, the preliminary CSI signal data classification result is regularized to obtain the CSI signal data classification result.
[0037] Furthermore, the step of using the MLP network based on the self-attention mechanism classifier to identify the temporal relationship of the normalized CSI signal and output preliminary CSI signal data classification results specifically includes:
[0038] The MLP network of the self-attention mechanism classifier includes a residual layer, a second normalization layer, and a classification layer;
[0039] Based on the residual layer of the MLP network, the normalized CSI signal time relationship is mapped to obtain the mapped CSI signal time relationship.
[0040] Based on the second normalization layer of the MLP network, the mapped CSI signal time relationship is normalized to obtain the CSI signal time relationship to be classified.
[0041] Based on the classification layer of the MLP network, the temporal relationship of the CSI signal to be classified is classified to obtain preliminary CSI signal data classification results.
[0042] The second technical solution adopted in this invention is: a CSI action recognition system based on deep nonnegative matrix factorization, comprising:
[0043] An amplitude calculation module is used to acquire CSI signal data and perform subcarrier amplitude calculation processing on the CSI signal data to obtain the subcarrier amplitude of CSI.
[0044] The feature extraction module is used to extract features from the subcarrier amplitude of the CSI using a deep nonnegative matrix factorization method to obtain a low-dimensional feature coefficient matrix.
[0045] The classification module is used to introduce position encoding and classify the low-dimensional feature coefficient matrix through a self-attention mechanism classifier to obtain the CSI signal data classification result.
[0046] The beneficial effects of the method and system of this invention are as follows: This invention collects CSI signal data and performs subcarrier amplitude calculation on the CSI signal data. Based on the subcarrier amplitude of CSI, feature extraction is performed using the deep non-negative matrix factorization method to mine the relationship between each subcarrier and the relationship between CSI data at each time point and other time points. This feature is then used to classify the data using a self-attention mechanism classifier. This avoids many manual feature design processes and prevents insufficient classification performance due to inadequate manual feature design. This invention combines the deep non-negative matrix factorization method with the self-attention mechanism classifier. The deep non-negative matrix factorization method can find the variation characteristics of a data over its time length, while the self-attention mechanism classifier can find the correlation between each time point and all time points in the entire time series. This effectively captures the global features of CSI signal data and improves the accuracy of classification. Attached Figure Description
[0047] Figure 1 This is a flowchart of the steps of the CSI action recognition method based on deep nonnegative matrix factorization in an embodiment of the present invention;
[0048] Figure 2 This is a structural block diagram of the CSI action recognition system based on deep nonnegative matrix factorization according to an embodiment of the present invention;
[0049] Figure 3 This is a schematic diagram illustrating a scenario of receiving and transmitting CSI data according to a specific embodiment of the present invention;
[0050] Figure 4 This is a schematic diagram of the network structure for recognizing CSI actions according to a specific embodiment of the present invention;
[0051] Figure 5 This is a schematic diagram of the process of CSI action recognition based on a model in a specific embodiment of the present invention. Detailed Implementation
[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
[0053] Reference Figure 1 and Figure 4 This invention provides a CSI action recognition method based on deep nonnegative matrix factorization, which includes the following steps:
[0054] S1. Acquire CSI signal data and perform subcarrier amplitude calculation on the CSI signal data to obtain the CSI subcarrier amplitude;
[0055] Specifically, CSI signal data is acquired via a Bluetooth module; based on the CSI signal data, the time value of each subcarrier in the CSI signal data is extracted to obtain the complex value of the CSI subcarrier; the real and imaginary parts of the CSI subcarrier complex value are successively squared and squared to obtain the preliminary CSI subcarrier amplitude; the preliminary CSI subcarrier amplitude is normalized to obtain the final CSI subcarrier amplitude.
[0056] Furthermore, refer to Figure 3This invention uses the Espressif ESP32 chip to transmit and receive CSI signals. CSI signals are a type of information used in wireless communication systems to describe the state and characteristics of the channel. Two chips were deployed in the experimental setting: one for transmitting CSI data and the other for receiving it. In this experiment, five actions were collected using this device: falling, jumping, running, walking, and remaining still. Six volunteers participated in each action, and each action was recorded 40 times. Six people collected a total of 1440 CSI data points. Based on the collected data, the following CSI action recognition system based on deep nonnegative matrix factorization and self-attention mechanism was developed. The system collects data and calculates the amplitude of each subcarrier at each time. In an Orthogonal Frequency Division Multiplexing (OFDM) system, a subcarrier refers to dividing the spectrum resources into multiple non-overlapping sub-channels. Each sub-channel is called a subcarrier. The subcarrier of a CSI signal refers to a specific subcarrier used to transmit CSI information. The value of each subcarrier in time is a complex value. The square root of the sum of the square of its real part and the square of its imaginary part is the amplitude. Then, the amplitude of each subcarrier is normalized to avoid problems such as data overflow and overfitting in subsequent processing.
[0057] S2. The subcarrier amplitude of CSI is processed by deep nonnegative matrix factorization to obtain a low-dimensional feature coefficient matrix.
[0058] Specifically, refer to Figure 5 The method sets the number of NMF layers in the deep nonnegative matrix factorization (DNF) method. It then performs a first NMF decomposition on the subcarrier amplitudes of the CSI based on the DNF method, obtaining a first feature matrix and a first coefficient matrix. Next, it performs a second NMF decomposition on the first coefficient matrix based on the DNF method, obtaining a second feature matrix and a second coefficient matrix. This process is repeated on the second coefficient matrix until the required number of NMF layers is reached. The DNF decomposition is then stopped, and a low-dimensional feature coefficient matrix is output. The DNF method uses three NMF layers. The first NMF layer is used to obtain the relationship between subcarriers of multiple CSIs at each time point. The second NMF layer is used to extract abstract features based on the relationship between subcarriers of multiple CSIs at each time point. Finally, the third NMF layer is used to perform dimensionality reduction on the abstract features.
[0059] Furthermore, to find the temporal amplitude relationship of the CSI signal in each subcarrier, this invention employs Deep NMF to process the CSI signal. Deep NMF is a variant of NMF. The significance of NMF lies in its ability to represent a matrix as a multiplication of a feature matrix and a coefficient matrix. The number of features in the feature matrix is determined by the designer. NMF can obtain a low-dimensional representation of the original matrix, achieving denoising and feature extraction effects. The original data is decomposed into a feature matrix W and a coefficient matrix H through NMF. The NMF expression is X≈W1H1, where the original data is represented as the product of two matrices. W represents the basis vectors of the original matrix, and H represents the combination coefficients of the basis vectors. By combining the coefficients of the basis vectors, a matrix approximating the original matrix X can be obtained. Deep NMF involves further decomposing the obtained coefficient matrix into an NMF matrix. For example, if the coefficient matrix obtained from one NMF operation is H1, further NMF decomposition yields H1≈W2H2, and so on. Multiple NMF decompositions ultimately form Deep NMF, with the final form being X≈W1W2W3...W1. n H n It consists of multiple nonlinear matrix decompositions (NMFs), with each level representing one NMF decomposition. The more levels of decomposition, the stronger its nonlinear expressive power, enabling it to uncover deeper and more abstract features. The product of the decomposed matrices is approximately close to the original matrix X. The error between the two is generally measured using Euclidean distance or KL divergence, and its expression is:
[0060]
[0061] In the above formula, cost represents the error between the two, X represents the original matrix, H represents the coefficient matrix, and W represents the characteristic matrix;
[0062] By continuously updating the cost through iteration, the optimal W and H are finally selected. The DeepNMF method used in this invention employs a 3-layer Deep NMF because the acquired CSI signal is a matrix composed of multiple subcarriers over a period of time when the action occurs. Through Deep NMF decomposition, the first layer obtains the relationship between multiple subcarriers at each time point, i.e., the features between subcarriers in time. The second layer extracts more abstract features based on this. The third layer mainly performs a dimensionality reduction process on the data. Due to the subsequent position encoding problem, the number of features decomposed in each layer of Deep NMF is an even number.
[0063] Furthermore, the use of a 3-layer Deep NMF in this invention will be explained. The coefficient matrix output by Deep NMF serves as the input data for subsequent networks. Regarding the choice of the number of Deep NMF layers, if the network has only one layer, i.e., only simple NMF is performed, the final classification accuracy is 60.6% when the number of features k=2, 63.9% when k=4, and a significant improvement in accuracy of 91.9% when k=6. A single-layer NMF layer already performs well, but a k value of 6 means that the length of each column vector processed by the Self-Attention layer is 6, and the number of parameters in subsequent fully connected layers increases several times. Therefore, another layer of NMF is performed on the coefficient matrix obtained from the first layer. F-decomposition extracts more abstract features. The accuracy is 92.2% when k=2, 92.5% when k=4, and 94.2% when k=6. The accuracy steadily increases with increasing k, reaching 93.9% when k=8, a slight decrease compared to k=6. Each layer of the second decomposition is derived from the previous layer, but the number of parameters is still relatively large when k=6. Therefore, a third layer decomposition is performed. In the third layer, the accuracy changes little with k=2, 4, 6, and 8. Therefore, k=2 is chosen as the final output of the Deep NMF network. The output with k=2 has three times fewer parameters than the output with k=6, simplifying model training and reducing parameter updates, ultimately achieving a classification performance of 95%, as shown in Table 1 below.
[0064] Table 1. Summary of accuracy for the first two K layers of Deep NMF
[0065] K (number of features) 2 4 6 8 Layer1 60.6％ 63.9％ 91.9％ 66％ Layer2 92.2％ 92.5％ 94.2％ 93.9％
[0066] Therefore, this invention uses a 3-layer Deep NMF to perform feature extraction processing on the subcarrier amplitude of CSI.
[0067] S3. Introduce position encoding and use a self-attention mechanism classifier to classify the low-dimensional feature coefficient matrix to obtain the CSI signal data classification results.
[0068] Specifically, a positional encoding is introduced, and the positional encoding is added to a low-dimensional feature coefficient matrix to obtain a summed low-dimensional feature coefficient matrix. This summed low-dimensional feature coefficient matrix is then input into a self-attention mechanism classifier for classification, outputting the CSI signal data classification result. The CSI signal data classification result represents different actions within the acquired CSI signal data. The summed low-dimensional feature coefficient matrix is then input into the self-attention mechanism classifier. The self-attention mechanism classifier includes a Self-Attention layer, a first normalization layer, an MLP network, and a Dropout layer. Based on the Self-Attention layer of the self-attention mechanism classifier, feature extraction is performed on the summed low-dimensional feature coefficient matrix to obtain the temporal relationship of the CSI signal. Based on the first normalization layer of the self-attention mechanism classifier, the temporal relationship of the CSI signal is normalized to obtain the normalized CSI signal. The process involves three steps: First, a self-attention mechanism classifier-based MLP network identifies the normalized CSI signal time relationship, outputting preliminary CSI signal data classification results. Second, a dropout layer based on the self-attention mechanism classifier regularizes the preliminary CSI signal data classification results, yielding the final CSI signal data classification result. The MLP network includes a residual layer, a second normalization layer, and a classification layer. Third, the residual layer of the MLP network maps the normalized CSI signal time relationship, resulting in a mapped CSI signal time relationship. Fourth, the second normalization layer of the MLP network normalizes the mapped CSI signal time relationship, yielding the CSI signal time relationship to be classified. Finally, the classification layer of the MLP network classifies the CSI signal time relationship to be classified, resulting in preliminary CSI signal data classification results.
[0069] Furthermore, this invention employs a 3-layer Deep NMF network, with each layer having 6, 6, and 2 features. The coefficient matrix obtained after Deep NMF processing is fed into the Self-Attention layer. Before being fed into the Self-Attention layer, a positional encoding is added to the original data to add the sequential relationship between each column of the input data. Because the Self-Attention mechanism targets global features but is unaware of the sequential relationship between each column, adding the positional encoding allows the Self-Attention layer to obtain the temporal relationship of the CSI signal. After passing through the Self-Attention layer, new data in the form [q1,q2,…,q] is obtained. nThe output is added to the previous input value to perform a residual operation. This residual operation helps the network get closer to the global optimum during training and parameter updates. After the residual operation, each column is normalized. Normalization helps the data converge faster during gradient calculation and reduces the risk of getting stuck in local optima. The data then passes through an MLP (Multi-Level Processing) layer. The first layer of the MLP also undergoes the same residual and column normalization operations. A Dropout layer is added between the last two layers to prevent overfitting during training, which can lead to poor performance on testing. The last layer is used for classification. The loss function is cross-entropy, and the update function is Adam. Compared to stochastic gradient descent (SGD), Adam can adaptively adjust the learning rate, which can lead to getting stuck in local optima and poor performance.
[0070] In summary, this invention proposes a method combining Deep NMF and Self-Attention to mine the relationships between each subcarrier, the relationships of CSI data at each time point with other time points, and uses this feature for classification. Compared to traditional machine learning methods, this system avoids many manual feature design processes, thus preventing insufficient classification performance due to inadequate manual feature design. Compared to previous deep learning methods, it achieves excellent performance, reaching a classification accuracy of 95%. Among deep learning methods, it has fewer parameters than most CNN and LSTM-based methods, and previous methods employed numerous preprocessing techniques before inputting data into the model. This method avoids this cumbersome process and has fewer parameters, resulting in lower memory consumption. Furthermore, this method is highly interpretable. It allows us to see that the data extracted from each layer in Deep NMF and Self-Attention can be visualized to observe its temporal patterns. The extracted data from each different action has obvious temporal features that can be directly observed through data visualization. Deep NMF can find the variation features of data over its time length, and Self-Attention can find the correlation between each time point and all time points in the entire time series. This correlation is the global feature found by Deep NMF and Self-Attention.
[0071] Compared to previous research, this study focuses more on uncovering the global temporal features of CSI signals. First, it employs Deep Non-negative Matrix Factorization (Deep NMF) to process the raw signal and extract its most prominent features. Next, it uses a self-attention mechanism to find the relationship between a given time step and other time steps. Finally, it performs classification. The basic structure of the self-attention mechanism is that each column vector of the input data has a query(q), key(k), and vector(v). The input data is then multiplied by three matrices W. q W l W v We can see that the Self-Attention mechanism obtains the correlation α through the dot product of q and k. Each column of the data corresponds to an α, for example, a certain column x of the data. i There is a degree of correlation α for each. i , and its α i Multiplying the value of v in other columns yields the attention score between the column and other columns. A higher score indicates a stronger relationship between the two columns. Therefore, the Self-Attention mechanism can obtain global features.
[0072] Reference Figure 2 A CSI action recognition system based on deep nonnegative matrix factorization includes:
[0073] An amplitude calculation module is used to acquire CSI signal data and perform subcarrier amplitude calculation processing on the CSI signal data to obtain the subcarrier amplitude of CSI.
[0074] The feature extraction module is used to extract features from the subcarrier amplitude of the CSI using a deep nonnegative matrix factorization method to obtain a low-dimensional feature coefficient matrix.
[0075] The classification module is used to introduce position encoding and classify the low-dimensional feature coefficient matrix through a self-attention mechanism classifier to obtain the CSI signal data classification result.
[0076] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0077] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.
Claims
1. A CSI action recognition method based on deep nonnegative matrix factorization, characterized in that, Includes the following steps: Collect CSI signal data and perform subcarrier amplitude calculation on the CSI signal data to obtain the CSI subcarrier amplitude; The subcarrier amplitude of the CSI is processed by deep nonnegative matrix factorization to obtain a low-dimensional feature coefficient matrix. Position encoding is introduced, and the low-dimensional feature coefficient matrix is classified using a self-attention mechanism classifier to obtain the CSI signal data classification result; The step of extracting features from the subcarrier amplitude of the CSI using a deep nonnegative matrix factorization method to obtain a low-dimensional feature coefficient matrix specifically includes: Set the number of layers in the depth nonnegative matrix factorization method; The subcarrier amplitude of the CSI is subjected to NMF decomposition based on the aforementioned deep nonnegative matrix decomposition method to obtain the feature matrix and the first coefficient matrix of the first decomposition. Based on the deep nonnegative matrix decomposition method, the first coefficient matrix is subjected to a second NMF decomposition to obtain the feature matrix and the second coefficient matrix of the second decomposition. The NMF decomposition process described above is performed on the second coefficient matrix until the number of NMF decompositions reaches the number of NMF layers, at which point the NMF decomposition stops. The low-dimensional feature coefficient matrix is constructed based on the feature matrix corresponding to each NMF decomposition and the coefficient matrix corresponding to the last NMF decomposition. The step of introducing position encoding and classifying the low-dimensional feature coefficient matrix using a self-attention mechanism classifier to obtain the CSI signal data classification result specifically includes: A positional encoding is introduced, and the positional encoding is added to the low-dimensional feature coefficient matrix to obtain the added low-dimensional feature coefficient matrix. The summed low-dimensional feature coefficient matrix is input into the self-attention mechanism classifier for classification processing, and the CSI signal data classification result is output, which represents the different actions in the collected CSI signal data.
2. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 1, characterized in that, The step of acquiring CSI signal data and performing subcarrier amplitude calculation on the CSI signal data to obtain the CSI subcarrier amplitude specifically includes: CSI signal data is acquired via Bluetooth module; Based on the CSI signal data, the time value of each subcarrier in the CSI signal data is extracted to obtain the complex value of the CSI subcarrier. The real and imaginary parts of the complex value of the CSI subcarrier are successively squared and squared to obtain the preliminary CSI subcarrier amplitude. The subcarrier amplitude of the preliminary CSI is normalized to obtain the subcarrier amplitude of the CSI.
3. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 1, characterized in that, The low-dimensional feature coefficient matrix comprises several feature matrices and one coefficient matrix, and its expression is as follows: In the above formula, Indicates the first The feature matrix of layer decomposition, Indicates the first The coefficient matrix of the layer decomposition, This represents a low-dimensional characteristic coefficient matrix.
4. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 1, characterized in that, The depth nonnegative matrix factorization (NMF) method described herein has three layers, wherein: Based on the first-layer NMF, it is used to obtain the relationship between subcarriers of multiple CSIs at each time point; Based on the second-layer NMF, abstract feature extraction is performed according to the relationship between the subcarriers of multiple CSIs at each time point; The third-layer NMF is used to perform data dimensionality reduction processing on the abstract features.
5. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 1, characterized in that, The loss function of the self-attention mechanism classifier is the cross-entropy loss function, and the update function of the self-attention mechanism classifier is the Adam function.
6. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 1, characterized in that, The step of inputting the summed low-dimensional feature coefficient matrix into the self-attention mechanism classifier for classification processing and outputting the CSI signal data classification result specifically includes: The summed low-dimensional feature coefficient matrix is input into the self-attention mechanism classifier, which includes a Self-Attention layer, a first normalization layer, an MLP network, and a Dropout layer. Based on the Self-Attention layer of the self-attention mechanism classifier, feature extraction processing is performed on the summed low-dimensional feature coefficient matrix to obtain the temporal relationship of the CSI signal. Based on the first normalization layer of the self-attention mechanism classifier, the temporal relationship of the CSI signal is normalized to obtain the normalized CSI signal temporal relationship. Based on the MLP network of the self-attention mechanism classifier, the time relationship of the normalized CSI signal is identified and processed, and a preliminary CSI signal data classification result is output. Based on the Dropout layer of the self-attention mechanism classifier, the preliminary CSI signal data classification result is regularized to obtain the CSI signal data classification result.
7. The CSI action recognition method based on deep nonnegative matrix factorization according to claim 6, characterized in that, The step of the MLP network based on the self-attention mechanism classifier identifying the temporal relationship of the normalized CSI signal and outputting preliminary CSI signal data classification results specifically includes: The MLP network of the self-attention mechanism classifier includes a residual layer, a second normalization layer, and a classification layer; Based on the residual layer of the MLP network, the normalized CSI signal time relationship is mapped to obtain the mapped CSI signal time relationship. Based on the second normalization layer of the MLP network, the mapped CSI signal time relationship is normalized to obtain the CSI signal time relationship to be classified. Based on the classification layer of the MLP network, the temporal relationship of the CSI signal to be classified is classified to obtain preliminary CSI signal data classification results.
8. A CSI action recognition system based on deep nonnegative matrix factorization, characterized in that, The system, applied to the method as described in any one of claims 1-7, includes the following modules: An amplitude calculation module is used to acquire CSI signal data and perform subcarrier amplitude calculation processing on the CSI signal data to obtain the subcarrier amplitude of CSI. The feature extraction module is used to extract features from the subcarrier amplitude of the CSI using a deep nonnegative matrix factorization method to obtain a low-dimensional feature coefficient matrix. The classification module is used to introduce position encoding and classify the low-dimensional feature coefficient matrix through a self-attention mechanism classifier to obtain the CSI signal data classification result.