A Privacy-Preserving Behavior Recognition Method and System Based on CSI Signals in Wi-Fi

Abstract

The present invention proposes a privacy - protected behavior recognition method and system based on CSI signals in Wi - Fi. CSI signals are extracted from Wi - Fi wireless signals, including CSI signals when no one is present and CSI signals when people perform different behaviors. The CSI signals are collected with a unified time series length and converted into I / Q two - path signals reflecting signal amplitude and phase information, and thus a CSI signal data set is obtained. Signal power normalization processing and class labels are performed to construct a training data set and a test data set for a human behavior recognition model. A CSI signal human behavior recognition model with a pyramid structure is constructed based on the self - attention mechanism and convolutional operations to determine the model weights after training is completed. The model parameters of the CSI signal human behavior recognition model are initialized, and the human behavior state is recognized for the CSI signals in the wireless Wi - Fi signals obtained in real - time from the indoor environment. The present invention improves the model recognition accuracy while reducing the model size.

Description

Technical Field

[0001] The present invention relates to wireless signal recognition technology, and particularly to a privacy protection behavior recognition method and system based on CSI signals in Wi-Fi. Background Art

[0002] With the development and popularization of wireless network technology, wireless sensing technology based on wireless signals has become another new type of sensing medium in addition to vision, sensors, light waves, sound waves, etc. Compared with traditional sensing technologies that use precise sensing devices, cameras, and radar detection, etc., which have disadvantages such as high equipment cost, high energy consumption, and poor privacy protection, making it difficult to be widely popularized and applied, non-intrusive wireless sensing technology has the characteristics of low cost, low energy consumption, and high universality.

[0003] The human behavior recognition technology based on Wi-Fi signals has now developed to be applicable to both coarse-grained fall detection and fine-grained scenarios such as respiration, heart rate monitoring, and sleep monitoring. With the aging of China's population structure, the demand for monitoring the safety and health status of the elderly living alone is becoming increasingly urgent, and human behavior recognition based on Wi-Fi signals can protect personal privacy from being violated while having the characteristics of low cost, ease of use, and accuracy. In the field of security and defense, human behavior recognition based on Wi-Fi signals also has a wide range of application scenarios, such as personnel movement monitoring in areas where confidential files are stored, intrusion detection of personnel in military restricted areas, building disaster emergency monitoring, and anti-theft of precious items in museums.

[0004] CSI signals express the orthogonal sub-channel state information of wireless electromagnetic wave signals, combining various factors such as time delay, amplitude attenuation, and phase shift, and describing the amplitude and phase of each sub-carrier in the frequency domain space. Compared with the coarse-grained received signal strength indication (RSSI), the fine-grained CSI signals have higher robustness and stability in sensing and recognition technologies such as human behavior recognition. In existing human behavior recognition methods based on CSI signals in Wi-Fi, traditional machine learning methods based on feature extraction have more complex signal processing processes and computational complexity; deep learning methods based on fully connected, convolutional, or long short-term memory neural networks cannot well learn the long-distance dependencies between signal features, or need to stack more network layers, resulting in problems such as too large model parameter scale, difficult training, and insufficient recognition accuracy. However, the self-attention mechanism can effectively encode the long-distance dependencies within features in a shallow network. As a CSI signal of a temporal feature, there are extensive long-time distance dependencies within its features, so the self-attention mechanism has important application value in the process of human behavior recognition based on CSI signals. Summary of the Invention

[0005] The object of the present invention is to propose a privacy - protected behavior recognition method and system based on CSI signals in Wi - Fi, so as to overcome the shortcomings of lack of privacy protection, low recognition accuracy and low automation degree in existing various human behavior recognition schemes.

[0006] The technical solution for achieving the object of the present invention is: A privacy - protected behavior recognition method based on CSI signals in Wi - Fi, comprising the following steps:

[0007] Step 1, extract CSI signals from Wi - Fi wireless signals, including CSI signals when no one is present and CSI signals when people perform different behaviors. Collect CSI signals with a unified time - series length, and convert them into I / Q two - path signals reflecting signal amplitude and phase information, thereby obtaining a CSI signal data set.

[0008] Step 2, perform signal power normalization processing on the CSI signal data set, and perform class labeling on the CSI signals when no one is present indoors and the CSI signals of different human behaviors, to construct a training data set and a test data set for the human behavior recognition model.

[0009] Step 3, construct a CSI signal human behavior recognition model with a pyramid structure based on the self - attention mechanism and convolutional operations. Use the training data set to train the CSI signal human behavior recognition model, and use the test data set to evaluate the model performance, and finally obtain the model weights after training.

[0010] Step 4, use the model weights after training to initialize the model parameters of the CSI signal human behavior recognition model, and perform human behavior state recognition on the CSI signals in the wireless Wi - Fi signals obtained in real time from the indoor environment.

[0011] Furthermore, it further includes Step 5. After the recognition process is completed, if the human behavior is within the safe range, the system does not make any response; otherwise, different system response levels are triggered according to the danger level of the recognition result, and the corresponding relationship between the recognition result danger level and the system response level is custom - configured by the user according to their own needs.

[0012] Furthermore, in Step 1, obtain CSI signal data when no one is present under different scene locations, weather, and light conditions, and obtain CSI signal data of people of different ages, heights, and genders performing different behaviors under different scene locations, weather, and light.

[0013] Furthermore, in Step 2, the method for performing signal power normalization processing on the CSI signal data set is as follows:

[0014] Let each I / Q time series matrix be H'. First, calculate the average power of each subcarrier time series in H', that is, the mean of the sum of the squares of the amplitudes of each subcarrier time series. Then, calculate the normalized value as the ratio of H' to the square root of the average power. The normalization operation is expressed as:

[0015]

[0016]

[0017]

[0018] where M represents the number of subcarriers of matrix H', N represents the length of the time series of matrix H', I t and Q t respectively represent the in-phase and quadrature values corresponding to the t-th CSI complex value of each subcarrier in matrix H'. |s t | represents the magnitude, sps represents the coefficient whose value is generally 1, and H represents the normalized matrix H'.

[0019] Furthermore, in step 2, perform class labeling on the CSI signals when the room is unoccupied and the CSI signals of different human behaviors. The specific method is as follows:

[0020] Divide the scenario with people into nine different behaviors: falling, dropping, standing, walking, squatting, sitting, standing up, lying, and running. As a control, the unoccupied scenario is regarded as a zero-action scenario. There are a total of ten behavior categories. The CSI signal data for the unoccupied scenario is labeled as category 0; for falling, dropping, standing, walking, squatting, sitting, standing up, lying, and running, they are labeled as category 1, 2, 3, 4, 5, 6, 7, 8, and 9 respectively.

[0021] Furthermore, in step 3, construct a CSI signal human behavior recognition model with a pyramid structure based on the self-attention mechanism and convolutional operations. It consists of a feature extraction network and a classifier. Among them:

[0022] The feature extraction network is divided into four stages. Each stage includes a segment embedding layer, a dynamic position encoding layer, and several Transformer encoder blocks; the classifier consists of a fully connected network layer and a Softmax activation function;

[0023] In the i-th stage, the CSI time series signal features or signal encoding features are first divided into several overlapping time-domain segments in the segment embedding layer and linearly mapped to the feature space. By controlling the step size between the time-domain segments, the overall time-domain length of the signal is reduced to form a feature pyramid structure. Subsequently, the dynamic position encoding layer is used to encode the position relationship between each segment, and then several Transformer encoder blocks are used to extract the global context dependencies between each overlapping time-domain segment;

[0024] Each Transformer encoder block consists of a multi-head self-attention layer and a local feed-forward module. A batch normalization layer is inserted before both the multi-head self-attention layer and the local feed-forward module, and residual connections are used for their inputs and outputs. The multi-head self-attention layer adopts a self-attention mechanism with local feature enhancement, and the local feed-forward module consists of a first linear mapping layer, a depthwise separable convolutional layer, an activation layer, and a second linear layer.

[0025] A privacy protection behavior recognition system based on CSI signals in Wi-Fi, based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi, realizes privacy protection behavior recognition based on CSI signals in Wi-Fi.

[0026] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi, it realizes privacy protection behavior recognition based on CSI signals in Wi-Fi.

[0027] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi, it realizes privacy protection behavior recognition based on CSI signals in Wi-Fi.

[0028] Compared with the prior art, the significant advantages of the present invention are as follows: (1) Using CSI signals in Wi-Fi for human behavior recognition avoids the problems of privacy leakage in vision-based recognition and inconvenience in use in wearable-based methods; (2) The collection of the CSI signal model training dataset takes into account the influence of multiple factors outside of human behavior itself, such as location, age, height, gender, weather, and light, making the model have higher robustness and can be further expanded in practical applications according to this spirit; (3) Using a deep learning method based on self-attention mechanism and convolutional operation with a pyramid structure to construct an identification model, using the I / Q two-channel signals that can reflect the amplitude and phase information of the signal as the input of the behavior recognition model, rather than simply converting the complex-valued representation of the signal into amplitude as the model input, improving the accuracy of the model's human behavior recognition and reducing the model size; (4) After the recognition process is completed, for special situations in the recognition results, automatic emergency response can be supported according to the danger level, improving the practicability, automation level, and emergency response speed of the model; (5) The present invention itself and its design concept are applicable to a variety of application scenarios, that is, it can monitor the behavior safety (falling, dropping, walking, sitting, lying, etc.) of the elderly in families or nursing homes, and can also be expanded for use in scenarios such as anti-theft in museums and intrusion detection of personnel in confidential rooms. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart for creating a behavior recognition model in an embodiment of the present invention.

[0030] Figure 2 It is a flowchart for applying the behavior recognition model in an embodiment of the present invention.

[0031] Figure 3 It is an overall structure diagram of the behavior recognition model in an embodiment of the present invention.

[0032] Figure 4 It is a schematic diagram of the self-attention mechanism with local feature enhancement used in the behavior recognition model in an embodiment of the present invention.

[0033] Figure 5 It is a schematic diagram of the local feed-forward module used in the behavior recognition model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0035] One of the main features of the CSI signal behavior recognition model proposed by the present invention is that while using the self-attention mechanism to encode the long-range dependencies within the CSI time-series signal features, it uses convolutional operations to learn the local context information within the signal features, so that the model has a more comprehensive expression ability for CSI signal features, thereby improving the classification accuracy of human behavior recognition using CSI signals. By constructing a model with a feature pyramid structure, the parameter scale of the recognition model is reduced. In addition, the model is trained using the I / Q two-channel signals in the CSI signal that can reflect amplitude and phase information, rather than only using amplitude information.

[0036] As Figure 1 shown, the privacy protection behavior recognition method based on CSI signals in Wi-Fi mainly includes the following steps: 1. Collect CSI samples; 2. Data preprocessing; 3. Data annotation; 4. Establish a new behavior recognition model; 5. Model training; 6. Apply the trained model for human behavior recognition.

[0037] Step 1, collect CSI samples:

[0038] Currently, the WI-FI technology based on the IEEE 802.11n protocol uses the MIMO-OFDM system. In the 20MHz bandwidth transmission mode, a wireless communication channel is modulated into 64 subcarriers using OFDM technology, and IEEE 802.11n uses 56 of them. The CSI of each subcarrier is a complex number, and the real and imaginary parts of the complex number represent the in-phase / quadrature two-channel signals respectively. For different types of actions under different conditions, the present invention collects CSI signals of WI-FI signals through the Atheros CSI Tool. The CSI of a group of data packets transmitted using m transmit antennas, n receive antennas, and a 20MHz channel bandwidth is a complex matrix of size 56×m×n. The CSI data packet is decomposed into I / Q two-channel signals to obtain a real CSI data packet with a scale size of 56×2×m×n;

[0039] For each action, T CSI data streams are extracted and arranged into a CSI data stream time series matrix X' of size (56×2×m×n)×T according to the sampling time. The second, third, and fourth dimensions of the matrix X' are compressed into one dimension, that is, each training sample corresponds to a time series matrix X with a dimension size of 56×2·mn×T. The set of matrices X of all different types of actions collected is the CSI sample data set;

[0040] The sample collection process consists of two parts, and all CSI signal data streams are collected with a unified time series length. First, collect CSI signal data when there is no one at different scene locations, weather conditions, and lighting. Second, collect CSI signal data of people with different ages, heights, and genders performing different actions at different scene locations, weather conditions, and lighting. Combine different external factors (scene location, weather, lighting, age, height, gender, etc.), and collect several CSI data stream time series samples for each type of action under the conditions of each combination.

[0041] Step 2, Data preprocessing:

[0042] Perform signal power normalization on the CSI data collected in Step 1. Divide the time series matrix X into m×n I / Q time series matrices with a dimension size of 56×2×T according to different channels. Let each I / Q time series matrix be H'. The specific method is to first calculate the average power of each subcarrier time series in H' (i.e., the mean of the sum of the squares of the amplitudes of each subcarrier time series). The normalized value is the ratio of H' to the square root of the average power. The normalization operation can be expressed as:

[0043]

[0044]

[0045]

[0046] where M represents the number of subcarriers of matrix H', N represents the time series length of matrix H', I t , Q t respectively represent the in-phase and quadrature values corresponding to the t-th CSI complex value of each subcarrier in matrix H' (stored in the second dimension of matrix H'), |s t | represents the amplitude size, sps represents the coefficient whose value is generally 1, and H represents the normalized matrix H'.

[0047] Step 3, Data annotation:

[0048] Divide the scenes with people into nine different behaviors: falling, dropping, standing, walking, squatting, sitting, standing up, lying, and running. As a control, the scene without people is regarded as a zero-action scene, for a total of ten behavior categories. Label the CSI signal data of the scene without people as category 0; label falling, dropping, standing, walking, squatting, sitting, standing up, lying, and running as categories 1, 2, 3, 4, 5, 6, 7, 8, and 9 respectively.

[0049] After all CSI time series samples are normalized in step 2 and labeled, the CSI data of various behaviors are divided into a training set and a test set according to a certain ratio, which are used for model training and model performance evaluation of the subsequent behavior recognition model respectively.

[0050] Step 4, establish a new behavior recognition model:

[0051] The overall structure of the behavior recognition model of the present invention is as Figure 3 shown, which consists of a feature extraction network and a classifier. Among them, the feature extraction network is divided into four stages, and each stage includes a segment embedding layer, a dynamic position encoding layer (DPE), and several Transformer encoder blocks; the classifier consists of a fully connected network layer and a Softmax activation function.

[0052] Before describing the model in detail, first introduce Figure 3 the meanings of each part in. The I / Q block represents the model input CSI signal, the Norm block represents the batch normalization (BatchNorm) layer, the Linear block represents the linear mapping layer, the DWConv block represents the depthwise separable convolution layer, the MHSA block represents the multi-head self-attention layer (MHSA), the LFF block represents the local feed-forward module (Local Feed-Forwar, LFF), the AvgPool block represents the average pooling operation in the feature time domain, and the Classifier block represents the classifier.

[0053] The CSI signal normalized in step 2 is the input of the behavior recognition model, and the label annotated for the input sample in step 3 is converted into a one-hot encoding for calculating the value of the model loss function (which can be flexibly selected). Let the CSI signal input to the model be where C 0 represents the number of original signal subcarriers 56 (number of channels) collected in step 1, N 0 represents the number of channels of the I / Q two-way signal, which is 2×m×n, and t 0 is the length of the CSI signal in the time domain. The input signal X 0First, it is injected into a feature extraction network consisting of four different stages to extract the high-dimensional encoded features of the signal (the matrix value after the input signal is operated by the feature extraction network). Then, the high-dimensional encoded features of the signal are injected into a classifier to be mapped to the corresponding categories of the recognition task. Finally, after performing the Softmax normalization operation on the mapping values corresponding to each category and calculating the model loss function value with the one-hot encoding, the weight gradient value can be calculated using the backpropagation algorithm (BP) to optimize the model parameters.

[0054] (1) Feature extraction network

[0055] The feature extraction network consists of four different stages. Each stage includes a segment embedding layer, a dynamic position encoding layer, and several Transformer encoder blocks. At the i-th stage, the input CSI time-series signal features (in the first stage) or signal encoded features (in the second, third, and fourth stages, which are obtained by operating the input CSI time-series signal through the network) X i are first divided into several overlapping time-domain segments in the segment embedding layer and linearly mapped to the feature space (as shown in the Figure 3 Segment Embedding module). Subsequently, the dynamic position encoding layer is used to encode the positional relationship between each segment, and then several Transformer encoder blocks are used to extract the global context dependencies between the overlapping time-domain segments.

[0056] At each stage of feature extraction, the CSI time-series signal features or signal encoded features (collectively referred to as signal features below) are first input into the segment embedding layer, which aims to divide the signal into overlapping segments in the time domain. At the i-th stage, the segment length is denoted as P i , and the stride distance between segments is set to S i . During segment embedding, first, the input features of this stage (for example, in the first stage, the input feature is X 0 , that is, the original CSI signal after normalization in step 2) are divided into T i = T i-1 / S i time-domain overlapping segments (hereinafter referred to as overlapping segments), and then the learnable linear projection (which can be implemented by a fully connected or convolutional layer) is multiplied by each segment and concatenated to obtain the embedding tensor The overlapping segment embedding process can be formulated as:

[0057]

[0058] where is a learnable linear projection, and represents the T i overlapping segments that are divided. When Si When = 2, the time domain length of the feature will be reduced by half compared to the input feature of this stage, thus forming a feature pyramid structure.

[0059] The signal features after segment embedding are injected into the dynamic position encoding layer, and the dynamic position encoding layer is used to learn the position interaction information between overlapping segments, which is defined as:

[0060] DPE(X i ) = DWConv(X i )

[0061] where DWConv represents the depthwise separable convolution operation.

[0062] After passing through the dynamic position encoding layer, the signal features are input into several Transformer encoder blocks (the structure is as shown in Figure 3 Transformer encoder), this part is used to learn the global context dependencies between each overlapping segment, and there are L i Transformer encoder blocks in the i-th stage. Each Transformer encoder block consists of a multi-head self-attention layer (MHSA) and a local feed-forward module (LFF). A batch normalization (BatchNorm) layer is inserted before both MHSA and LFF, and residual connections are used for their inputs and outputs. Formally, a Transformer encoder block can be defined as:

[0063]

[0064]

[0065] where l = 1, 2, …, L i , and respectively represent the input and output feature tensors of the l-th Transformer encoder block in the i-th stage.

[0066] Inside each Transformer encoder block, the input signal features are first batch-normalized and then fed into the MHSA. The MHSA is used to encode the self-correlation between each overlapping segment of the signal features. In the original Self-Attention, the input tensor is respectively linearly mapped (Linear) to obtain the query keys and values And the equation mainly based on the dot product operation is the core operation principle of the self-attention mechanism, which calculates the self-correlation between the internal information of global features. However, the original version lacks local context modeling that is crucial for the fine-grained recognition process. To encode the local context features between adjacent overlapping segments, the present invention adopts a self-attention mechanism enhanced by a Local Contextual Feature Encoding (LCFE) layer (the schematic diagram of which is as shown in Figure 4 ), which uses depthwise separable convolutions with a convolutional kernel size of h×w on K and V to encode their local features, that is, and so as to utilize the information interaction between adjacent overlapping segments to provide a more detailed feature representation. Therefore, the local feature-enhanced self-attention mechanism used in the present invention ( Figure 4 ) can be expressed as:

[0067]

[0068] Finally, several single self-attention heads are connected to form an MHSA block, which can be defined as:

[0069] MHSA(X) = Concat(head 1 , head 2 , …, head h )

[0070] where each head h represents an Attn function.

[0071] After MHSA, the obtained signal features are input into LFF after BatchNorm (the structural diagram of which is as shown in Figure 5 ). LFF consists of a first linear mapping layer (Conv), a depthwise separable convolutional layer (DWConv), an activation layer (GELU), and a second linear layer (Conv), and its operation process can be expressed as:

[0072] X out = Conv(GELU(DWConv(Conv(X in )))) + X in

[0073] where GELU represents the GELU activation function, X in represents the LFF input, and X out represents the LFF output. The first linear mapping layer expands the number of feature channels by E i times, and the second linear mapping layer reduces the number of feature channels to the number of channels when the feature is input into this module.

[0074] (2) Classifier

[0075] The classifier consists of a fully-connected network layer and a Softmax activation function. The fully-connected layer maps the high-dimensional encoded features to the corresponding categories for specific recognition tasks; the Softmax function normalizes the classification results for use in calculating the model loss function.

[0076] Step 5, Model training and performance evaluation:

[0077] The training set data is input in batches into the behavior recognition model established in Step 4 for model training. After training is completed, the test set is used to evaluate the model performance. Finally, the trained model weights are obtained and saved for use in the application of the human behavior recognition model in Step 6.

[0078] Step 6, Apply the trained model for human behavior recognition:

[0079] The application of the human behavior recognition model in the scenario is as Figure 2 shown. During the operation of the system, CSI signals with a time series length not less than that of the training data should be cached. The specific application process is as follows:

[0080] (1) First, initialize the behavior recognition model established in Step 4 using the behavior recognition model weights saved after the training in Step 5;

[0081] (2) Extract CSI signals with the same time series length as the training data from the system CSI cache, that is, recursively retrieve signals from the signal obtained at the current moment to earlier signal sequences, and intercept CSI signals with the same time series length as the training data in Step 1. It should be noted that the acquisition time interval of CSI signals should be flexibly set after weighing the actual application scenario and the device computing performance. The smaller the acquisition time interval of the signal, the more frequent the behavior recognition process;

[0082] (3) Perform preprocessing of signal power normalization on the extracted CSI signals. The specific processing process is the same as that in Step 2;

[0083] (4) Input the preprocessed data into the initialized recognition model for feature extraction and final classification process;

[0084] (5) Obtain the behavior recognition result. The possible classification results are any one of the behavior categories labeled in Step 3, preparing for the subsequent system emergency response based on the recognition result;

[0085] (6) Identification result processing. If the human behavior is within the safe range, the system does not make any response. Otherwise, different system response levels are triggered according to the risk level of the identification result. It should be noted that the user can set multiple system response levels according to the identification result, and the corresponding relationship between the risk level of the identification result and the system response level can be customized by the user in the configuration file according to their own needs;

[0086] (7) Repeat the above processes (2)-(6).

[0087] The present invention also proposes a privacy protection behavior recognition system based on CSI signals in Wi-Fi, which realizes privacy protection behavior recognition based on CSI signals in Wi-Fi based on the above-mentioned privacy protection behavior recognition method based on CSI signals in Wi-Fi.

[0088] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, privacy protection behavior recognition based on CSI signals in Wi-Fi is realized based on the above-mentioned privacy protection behavior recognition method based on CSI signals in Wi-Fi.

[0089] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, privacy protection behavior recognition based on CSI signals in Wi-Fi is realized based on the above-mentioned privacy protection behavior recognition method based on CSI signals in Wi-Fi.

[0090] In summary, the present invention uses CSI signals in Wi-Fi for human behavior recognition, avoiding the problems of privacy leakage in vision-based recognition methods and inconvenience in use in wearable-based methods; uses a deep learning method based on a pyramid structure, self-attention mechanism, and convolutional module to construct a CSI signal human behavior recognition model, reducing the model size while improving the model recognition accuracy; after the recognition process is completed, for the recognition result, when a dangerous situation occurs, automatic emergency response can be supported according to the risk level, improving the practicability, automation degree, and emergency response speed of the model.

[0091] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0092] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A privacy - protected behavior recognition method based on CSI signals in Wi - Fi, characterized in that, it includes the following steps: Step 1, Extract CSI signals from Wi - Fi wireless signals, including CSI signals when there is no one and CSI signals when people perform different behaviors. Collect CSI signals with a unified time series length and convert them into I / Q two - channel signals reflecting signal amplitude and phase information, thereby obtaining a CSI signal data set; Step 2, Perform signal power normalization processing and category marking on the CSI signal data set to construct a training data set and a test data set for the human behavior recognition model; Step 3, Construct a CSI signal human behavior recognition model with a pyramid structure based on the self - attention mechanism and convolution operation. Use the training data set to train the CSI signal human behavior recognition model, and use the test data set to evaluate the model performance, and finally obtain the model weights after training; Step 4, Initialize the model parameters of the CSI signal human behavior recognition model with the model weights after training, and perform human behavior state recognition on the CSI signals in the wireless Wi - Fi signals obtained in real - time from the indoor environment; Step 3, Construct a CSI signal human behavior recognition model with a pyramid structure based on the self - attention mechanism and convolution operation, which consists of a feature extraction network and a classifier. Among them: The feature extraction network is divided into four stages, and each stage includes a segment embedding layer, a dynamic position encoding layer, and several Transformer encoder blocks; the classifier consists of a fully - connected network layer and a Softmax activation function; In the i - th stage, the CSI time - series signal features or signal encoding features are first divided into several overlapping time - domain segments in the segment embedding layer and linearly mapped to the feature space. By controlling the step size between time - domain segments, the overall time - domain length of the signal is reduced to form a feature pyramid structure. Subsequently, the dynamic position encoding layer is used to encode the position relationship between each segment, and then several Transformer encoder blocks are used to extract the global context dependencies between each overlapping time - domain segment; Each Transformer encoder block consists of a multi - head self - attention layer and a local feed - forward module. A batch normalization layer is inserted before the multi - head self - attention layer and a local feed - forward module, and residual connections are used for their input and output; the multi - head self - attention layer adopts a self - attention mechanism for local feature enhancement, and the local feed - forward module consists of a first linear mapping layer, a depth - wise separable convolution layer, an activation layer, and a second linear layer.

2. The privacy - protected behavior recognition method based on CSI signals in Wi - Fi according to claim 1, characterized in that, it further includes Step 5. After the recognition process is completed, if the human behavior is within the safe range, the system does not make any response. Otherwise, different system response levels are triggered according to the risk level of the recognition result, and the corresponding relationship between the recognition result risk level and the system response level is custom - configured by the user according to their own needs.

3. The privacy - protected behavior recognition method based on CSI signals in Wi - Fi according to claim 1, characterized in that, In step 1, CSI signal data when there is no one is obtained under different scene locations, weather, and light conditions, and CSI signal data when people of different ages, heights, and genders perform different behaviors are obtained under different scene locations, weather, and light.

4. The privacy protection behavior recognition method based on CSI signals in Wi-Fi according to claim 1, characterized in that step 2, perform signal power normalization processing on the CSI signal dataset, and the specific method is: Let each I / Q time series matrix be H'. First, calculate the average power of each sub-carrier time series in H', that is, the mean of the sum of the squares of the amplitudes of each sub-carrier time series, and then calculate the normalized value as the ratio of H' to the square root of the average power. The normalization operation is expressed as: where M represents the number of subcarriers of matrix H', N represents the length of the time series of matrix H', I t , Q t respectively represent the in-phase and quadrature values corresponding to the t-th CSI complex value of each subcarrier in matrix H', |s t | represents the magnitude, sps represents the coefficient, whose value is 1, and H represents the normalized matrix H'.

5. The privacy protection behavior recognition method based on CSI signals in Wi-Fi according to claim 1, characterized in that step 2, perform category marking on the CSI signal, and the specific method is: The scene with people is divided into nine different behaviors: falling, dropping, standing, walking, squatting, sitting, getting up, lying, and running. As a control, the scene without people is regarded as a zero-action scene. There are a total of ten behavior categories. For the CSI signal data of the scene without people, it is labeled as category 0; for falling, dropping, standing, walking, squatting, sitting, getting up, lying, and running, they are respectively labeled as category 1, 2, 3, 4, 5, 6, 7, 8, and 9.

6. A privacy protection behavior recognition system based on CSI signals in Wi-Fi, characterized in that Based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi described in any one of claims 1-5, the privacy protection behavior recognition based on CSI signals in Wi-Fi is realized.

7. A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi described in any one of claims 1-5, the privacy protection behavior recognition based on CSI signals in Wi-Fi is realized.

8. A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, based on the privacy protection behavior recognition method based on CSI signals in Wi-Fi described in any one of claims 1-5, the privacy protection behavior recognition based on CSI signals in Wi-Fi is realized.

Images