A lightweight user behavior recognition method based on a multi-level feature extraction algorithm

Abstract

The application discloses a kind of light weight user behavior recognition methods based on multilevel feature extraction algorithm, comprising: collecting CSI original data;Extract first time domain data set;Second time domain data set and first frequency domain data set are obtained by sample feature enhancement strategy;Deep feature extraction is carried out, time domain sample level feature and frequency domain sample level feature are obtained and feature fusion is carried out, obtain first fusion feature, input into bidirectional Mamba network, obtain second fusion feature, input into multi-head attention mechanism to obtain third fusion feature, after global average pooling, input into classifier, output the probability distribution of each action category;Dynamic identification and shield low contribution feature in model training process;The application is through sample feature enhancement strategy, multilevel deep feature representation extraction strategy and model redundant feature light weight strategy, effectively improve user behavior recognition precision, and reduce model complexity and calculation cost.

Description

Technical Field

[0001] This invention relates to the field of human behavior recognition technology, and in particular to a lightweight user behavior recognition method based on a multi-level feature extraction algorithm. Background Technology

[0002] User behavior recognition is a crucial component of artificial intelligence in fields such as smart homes, human-computer interaction, security monitoring, and smart healthcare. In the traditional perception technology stage, researchers primarily rely on cameras or wearable sensors to capture human motion characteristics. Cameras recognize user behavior through visual images, offering high accuracy and intuitiveness, but are susceptible to variations in lighting, viewing angle, and occlusion, and pose privacy risks. Wearable sensors can directly capture human motion data and identify fine-grained behavioral features, but their reliance on user-worn devices presents challenges such as high cost, inconvenience, and ease of forgetting. Therefore, for behavior recognition in daily life, there is an urgent need for a detection method that is wear-free, unaffected by lighting conditions, and low-cost.

[0003] With the development of wireless technology and the widespread adoption of WiFi devices, WiFi has become an ideal device for wireless sensing due to its advantages such as passive sensing, low cost, flexibility, and ease of use. Channel State Information (CSI) in WiFi signals reflects the multipath effects and phase changes of electromagnetic waves during propagation, indirectly characterizing human motion. WiFi devices employ Multiple Input Multiple Output (MIMO) and Orthogonal Frequency Division Multiplexing (OFDM) technologies, enabling CSI to capture high-resolution spatial and temporal characteristics between different carriers and antennas. Compared to traditional sensing methods, CSI sensing requires no additional equipment, does not infringe on privacy, can penetrate obstructions, and offers advantages such as low cost, non-contact operation, and wide coverage.

[0004] However, existing CSI-based behavior recognition models still have several shortcomings: First, deep learning models have large parameters and high computational cost, which is not conducive to real-time deployment of edge devices; second, traditional feature extraction methods have failed to fully explore the complementary characteristics of CSI signals in the time and frequency domains, and have also failed to effectively model the hierarchical correlation information between subcarriers, antenna pairs and the overall sample, resulting in limited feature expression capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight user behavior recognition method based on a multi-level feature extraction algorithm, which solves the problems of large parameters, high computational cost, unfavorable real-time deployment of the above-mentioned CSI-based behavior recognition model, and limited feature representation capability and low user behavior recognition accuracy of traditional feature extraction methods.

[0006] To achieve the above objectives, this invention provides a lightweight user behavior recognition method based on a multi-level feature extraction algorithm, comprising the following steps:

[0007] Step 100: Collect CSI data from multiple users across various action categories to obtain the raw dataset;

[0008] Step 200: Extract the first time-domain dataset from the original dataset;

[0009] Step 300: Process the first time-domain dataset using a sample feature enhancement strategy to obtain the second time-domain dataset and the first frequency-domain dataset;

[0010] Step 400: The deep feature extraction network in the multi-level deep feature representation extraction strategy is used to perform deep feature extraction on the second time domain dataset and the first frequency domain dataset respectively to obtain time domain sample-level features and frequency domain sample-level features.

[0011] Step 500: The time-domain sample-level features and frequency-domain sample-level features are fused using an adaptive weighted fusion method to obtain the first fused feature;

[0012] Step 600: Input the first fused feature into the encoder to obtain the third fused feature;

[0013] Step 700: The third fusion feature is input into the classifier after global average pooling, and the probability distribution of each action category is output.

[0014] Backpropagation is performed using the total loss function to optimize the deep feature extraction network, encoder, classifier, and trainable parameters.

[0015] After training converges, low-contribution features are dynamically identified and masked during training through a model redundancy feature lightweighting strategy.

[0016] After the model is trained, the network recognition accuracy is calculated using the macro-average method. The recognition accuracy is then compared with a set threshold. If the recognition accuracy is greater than the set threshold, the performance is considered to meet the perception requirements, and the final user behavior recognition model is obtained. Otherwise, the performance is considered to not meet the perception requirements. When the performance does not meet the perception requirements, the number of iterations of the loss function is increased and the judgment is repeated until the performance meets the perception requirements.

[0017] Furthermore, the formula for calculating the total loss function is as follows:

[0018] ;

[0019] ;

[0020] in Represents the loss function. For the total loss function, This represents the label of the sample corresponding to the j-th action. This represents the probability of predicting the action as the j-th type. and These are the subcarrier coefficient and the antenna sparsity coefficient, respectively, where F is the number of subcarriers and S is the number of antennas. For the number of action categories, This represents the sequence mask for the i-th subcarrier. This represents the sequence mask for the j-th antenna.

[0021] Furthermore, methods for obtaining a second time-domain dataset and a first frequency-domain dataset by applying sample feature enhancement strategies to the first time-domain dataset include:

[0022] Hampel filtering was applied to the first time-domain dataset to remove outliers that did not conform to the overall trend;

[0023] Butterworth low-pass filter was applied to the first time-domain dataset to remove high-frequency noise and perform smoothing.

[0024] For the first time-domain dataset, linear interpolation is used to fill in missing values ​​based on the average trend of the preceding and following data.

[0025] Perform data regularization on the first time-domain dataset to obtain the second time-domain dataset;

[0026] The obtained second time-domain dataset is processed by short-time Fourier transform to obtain the first frequency-domain dataset.

[0027] Furthermore, the multi-level deep feature representation extraction strategy is specifically a dual-branch input structure, including a time-domain branch and a frequency-domain branch. The two branches are used to process signal features in the time domain and frequency domain, respectively. The time-domain branch receives the second time-domain dataset, and the frequency-domain branch receives the first frequency-domain dataset. The two branches perform feature extraction in parallel.

[0028] Furthermore, both the temporal and frequency domain branches in the multi-level deep feature representation extraction strategy employ deep feature extraction networks. These deep feature extraction networks consist of three separable convolutional layers, specifically:

[0029] The first layer of the deep feature extraction network focuses on subcarrier-level feature extraction. Each subcarrier signal is individually input into the first layer of the deep feature extraction network to achieve feature extraction, and the output is a subcarrier feature sequence.

[0030] The second layer of the deep feature extraction network is used to aggregate subcarrier features belonging to the same transmit-receive antenna pair, and to process the subcarrier feature sequence. Feature extraction, outputting antenna pair features

[0031] The third layer of the deep feature extraction network is oriented sample-level feature extraction, which stitches the antenna pair features into a whole and extracts features from it, outputting time-domain sample-level features and frequency-domain sample-level features.

[0032] A non-linear activation function is used after each layer of the deep feature extraction network.

[0033] Furthermore, through an adaptive weighted fusion method, the time-domain sample-level features and frequency-domain sample-level features are fused to obtain the calculation formula for the first fused feature, which is as follows:

[0034] ;

[0035] in The first fusion feature, and These are the first and second feature fusion parameters, respectively, both of which are trainable parameters. For time-domain sample-level features, These are frequency domain sample-level features.

[0036] Furthermore, the method for inputting the first fusion feature into the encoder to obtain the third fusion feature is as follows: input the first fusion feature into the bidirectional Mamba network to obtain the second fusion feature, and the second fusion feature is used to obtain the third fusion feature through a multi-head attention mechanism.

[0037] Furthermore, the steps to output the probability distribution for each action category are as follows:

[0038] The third fusion feature is input into the classifier after global average pooling. The classifier outputs the scores of each of the M types of actions, and after Softmax activation, it outputs the probability distribution of the action categories.

[0039] Furthermore, a method for dynamically identifying and masking low-contribution features during training, through a model redundancy feature lightweighting strategy, is as follows:

[0040] During training, the trainable scalar corresponding to the subcarrier calculates the subcarrier sequence mask using the Sigmoid function. If the subcarrier sequence mask is 0, it is judged as low contribution, and low contribution subcarrier-level features are automatically suppressed. The trainable scalar corresponding to the antenna calculates the antenna sequence mask using the Sigmoid function. If the antenna sequence mask is 0, it is judged as low contribution, and low contribution antenna-level features are suppressed.

[0041] Therefore, the lightweight user behavior recognition method based on the multi-level feature extraction algorithm described above, as used in this invention, has the following beneficial effects:

[0042] 1. Employ sample feature enhancement strategies to enhance sample features before training. In the time domain, this includes outlier removal, high-frequency noise removal, linear interpolation, and data regularization. In the frequency domain, this includes short-time Fourier transform.

[0043] 2. Multi-level deep feature representations are extracted step by step using subcarrier-level convolutional networks, antenna-level convolutional networks, and sample-level convolutional networks to fully explore the multi-scale feature associations of CSI from local to global perspectives;

[0044] 3. A lightweight bidirectional Mamba module combined with a multi-head attention mechanism is used to further refine and model the features of the fused samples, so as to enhance the temporal correlation and global expressive power of the features;

[0045] 4. A lightweight strategy for redundant features is introduced. Trainable masks are assigned to redundant features at the subcarrier and antenna levels. Low-contribution features are adaptively identified and masked, thereby further enhancing the effectiveness of feature representation and the lightweight performance of the model.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] Figure 1 This is a technical framework diagram of a lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to the present invention.

[0048] Figure 2 This is a flowchart illustrating the model training process of a lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to the present invention. Detailed Implementation

[0049] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0050] Please see Figure 1-2 A lightweight user behavior recognition method based on a multi-level feature extraction algorithm includes the following steps:

[0051] Step 100: Collect CSI data from multiple users across various action categories to obtain the raw dataset. ;

[0052] In an embodiment of the present invention, a transmitting antennas and b receiving antennas are used in the environment, resulting in a total of a·b independent propagation paths. A raw dataset of n users and s types of actions is collected in the environment. , For dataset The corresponding tag set;

[0053] Step 200, from the original dataset Extracting the first time-domain dataset ;

[0054] Step 300: Process the first time-domain dataset using a sample feature enhancement strategy to obtain the second time-domain dataset. and the first frequency domain dataset ;

[0055] The sample feature enhancement strategies for processing the first time domain dataset include outlier removal, high-frequency noise removal, linear interpolation, and data regularization.

[0056] Specifically, in the time-domain data, numerical abrupt changes or outliers may occur due to equipment errors or environmental interference. Therefore, Hampel filtering is first used to remove outliers that do not conform to the overall trend. Considering that human behavior signals are mainly distributed in the low-frequency band, Butterworth low-pass filtering is further used to remove high-frequency noise and smooth the data in order to suppress high-frequency noise and highlight human behavior characteristics. Furthermore, due to equipment acquisition defects or outlier removal, the data may be discontinuous. For missing data, linear interpolation is used to fill in the missing values ​​based on the average trend of the preceding and following data. Finally, differences in the units and numerical ranges of different features can easily lead to the masking of weak features; therefore, data regularization is performed on the samples to balance the influence of each feature. After the above sample feature enhancement strategies, the sample features can be enhanced, compensating for the insufficient feature extraction capabilities of the subsequent lightweight module, resulting in a higher-quality and more continuous second time-domain dataset. ;

[0057] Time-frequency analysis is performed by processing each sample in the second time-domain dataset using short-time Fourier transform.

[0058] ;

[0059] in For the window function, we choose the Hamming window, defined as:

[0060] ;

[0061] Where N is the window length, this transformation yields the first frequency domain dataset. This converts a one-dimensional time-series signal into a two-dimensional time-frequency image representation.

[0062] Step 400: The deep feature extraction network in the multi-level deep feature representation extraction strategy is used to extract features from the second time-domain dataset. and the first frequency domain dataset Perform deep feature extraction to obtain temporal sample-level features. and frequency domain sample-level features ;

[0063] The multi-level deep feature representation extraction strategy is specifically a two-branch input structure, with two branches processing signal features in the time and frequency domains, respectively. In this structure, the time-domain branch receives a second time-domain dataset. The frequency domain branch receives the first frequency domain dataset. Both processes perform feature extraction in parallel. Each branch employs a deep feature extraction network, which consists of a three-layer separable convolutional network, specifically:

[0064] The first layer of the deep feature extraction network focuses on subcarrier-level feature extraction to capture amplitude differences between different subcarriers. Specifically, each subcarrier signal is individually input into the first layer of the deep feature extraction network for feature extraction, with a convolution kernel size of 1. 6, step size 2, number of output channels 64, output subcarrier feature sequence .

[0065] The second layer of the deep feature extraction network focuses on antenna-level feature modeling to learn the spatial correlation between multiple antennas. In an embodiment of the invention, 30 subcarrier features belonging to the same transmit-receive antenna pair are summarized and feature aggregation is performed along the subcarrier dimension. Specifically, a convolutional kernel with a size of 64 is used. 4. A convolutional layer with a stride of 5 and 128 output channels is used to process the subcarrier feature sequence. A convolution operation is performed to extract the local correlation patterns of each antenna pair. After this convolutional layer, six sets of antenna pair features are output, denoted as... .

[0066] The third layer of the deep feature extraction network is oriented towards sample-level feature extraction, capturing temporal changes and dynamic environmental patterns as a whole. In an embodiment of the invention, the features of six antenna pairs are concatenated into a single input, using a convolutional kernel size of 128. 3. A convolutional layer with a stride of 3 and 128 output channels achieves global fusion, outputting temporal sample-level features. .

[0067] In embodiments of the present invention, a ReLU activation function is employed after each separable convolution layer to enhance the nonlinear expressive power of the network and accelerate convergence.

[0068] Similarly, in the frequency domain branch, the same three-layer separable convolutional deep feature extraction network is used to extract features from the frequency domain data layer by layer, ultimately obtaining frequency domain sample-level features. .

[0069] Step 500: Using an adaptive weighted fusion method, the temporal sample-level features are... and frequency domain sample-level features Perform feature fusion to obtain the first fused feature. ;

[0070] First fusion feature The calculation formula is as follows:

[0071] ;

[0072] in The first fusion feature, and These are the first and second feature fusion parameters, both of which are trainable parameters used to dynamically adjust the contribution ratio of time-domain and frequency-domain features in the fusion process, thereby achieving adaptive balance at the feature level.

[0073] Step 600: Input the first fused feature into the encoder to obtain the third fused feature. Specifically, this involves: inputting the first fused feature... The input is fed into a bidirectional Mamba network to obtain the second fused feature. The first fused feature is implemented by two independent selection state space branches: forward Mamba and backward Mamba. Specifically, the state equation and observation equation of the forward branch are as follows:

[0074] ;

[0075] ;

[0076] The state equations and observation equations for the backward branch are:

[0077] ;

[0078] ;

[0079] in , Let these represent the first and second internal representations at time t, respectively. , These represent the outputs of the forward scan block and the reverse scan block at time t, respectively. A, B, and C are derived linearly from the input samples.

[0080] The forward and backward observation outputs are fused using a learnable linear fusion matrix to obtain the second fused feature, calculated as follows:

[0081] ;

[0082] , Let them represent the first linear fusion matrix and the second linear fusion matrix, respectively. This indicates the second fusion feature.

[0083] The second fusion feature is used to obtain the third fusion feature through a multi-head attention mechanism. The specific calculation formula is as follows:

[0084] ;

[0085] in, , , ∈ These are the trainable parameter matrices for the query vector Q, key vector K, and value vector V, respectively. The dimension of the input feature. For projection dimensions;

[0086] Attention weights are calculated based on the query vector Q and key vector K, and the value vector V is weighted and summed using these attention weights to generate the third fusion feature. The calculation formula is as follows:

[0087] ;

[0088] in for function.

[0089] By using a two-way information exchange mechanism, the model's expressive power can be maintained while effectively reducing the parameter size, thus achieving overall model lightweighting.

[0090] Step 700, Third Fusion Feature After global average pooling, the input is fed into the classifier, which outputs the probability distribution of each action category;

[0091] The classifier consists of two fully connected layers, specifically: a third fusion feature. After global average pooling, the input is fed into the first fully connected layer of the classifier, activated by ReLU, and outputs 128-dimensional intermediate features. These intermediate features are then fed into the second fully connected layer of the classifier, mapping them to... 3D space, output The scores for each action class are then processed through Softmax activation to output the probability distribution of the action class. The calculation formula is as follows:

[0092] ;

[0093] in It is the probability of predicting the action as the i-th type. The score for the i-th type of action, For the score of the j-th type of action, For the number of action categories, It is an exponential function.

[0094] By employing a model redundancy feature lightweighting strategy, low-contribution features are dynamically identified and masked during training, thereby achieving feature lightweighting at the subcarrier and antenna levels.

[0095] In an embodiment of the present invention, for subcarrier-level samples, each input subcarrier corresponds to a trainable scalar. A 0 / 1 sequence mask is obtained through Sigmoid. This sequence mask is used to automatically mask low-contribution subcarrier features during training, thereby achieving adaptive selection at the feature level.

[0096] For antenna-level samples, each input antenna corresponds to a trainable scalar. A 0 / 1 sequence mask is obtained through Sigmoid. This sequence mask can adaptively identify and mask low-contribution antenna features during training, thereby further achieving lightweighting and feature optimization in the spatial dimension.

[0097] It should be noted that if the sequence mask of a subcarrier is 0, it is judged as a low contribution; if the sequence mask of a subcarrier is 1, it is judged as a high contribution.

[0098] In embodiments of the present invention, a loss function is used for backpropagation to optimize the deep feature extraction network, encoder, classifier, and trainable parameters. The formula for calculating the loss function is as follows:

[0099] ;

[0100] in Represents the loss function. This represents the label of the sample corresponding to the j-th action. This represents the probability of predicting the action as the j-th type. This represents the number of action categories.

[0101] Furthermore, L1 sparsity regularization is added to the loss function to sparsify the mask, resulting in the total loss function, calculated as follows:

[0102] ;

[0103] For the total loss function, and These are the subcarrier coefficient and the antenna sparsity coefficient, respectively, where F is the number of subcarriers. For the number of antennas, This represents the sequence mask for the i-th subcarrier. This represents the sequence mask for the j-th antenna.

[0104] After training converges, the low-contribution subcarriers and antenna masks are 0, which can be directly masked or pruned to obtain a lightweight model, thus achieving lightweighting at the training level.

[0105] After the model is trained, the network recognition accuracy is calculated using the macro-average method. The recognition accuracy is then compared with a set threshold. If the recognition accuracy is greater than the set threshold, the performance is considered to meet the perception requirements, and the final user behavior recognition model is obtained. Otherwise, the performance is considered to not meet the perception requirements. When the performance does not meet the perception requirements, the number of iterations of the loss function is increased and the judgment is repeated until the performance meets the perception requirements.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm, characterized in that, Includes the following steps: Step 100: Collect CSI data from multiple users across various action categories to obtain the raw dataset; Step 200: Extract the first time-domain dataset from the original dataset; Step 300: Process the first time-domain dataset using a sample feature enhancement strategy to obtain the second time-domain dataset and the first frequency-domain dataset; Step 400: The deep feature extraction network in the multi-level deep feature representation extraction strategy is used to perform deep feature extraction on the second time domain dataset and the first frequency domain dataset respectively to obtain time domain sample-level features and frequency domain sample-level features. Step 500: The time-domain sample-level features and frequency-domain sample-level features are fused using an adaptive weighted fusion method to obtain the first fused feature; Step 600: Input the first fused feature into the encoder to obtain the third fused feature; Step 700: The third fusion feature is input into the classifier after global average pooling, and the probability distribution of each action category is output. Backpropagation is performed using the total loss function to optimize the deep feature extraction network, encoder, classifier, and trainable parameters. After training converges, low-contribution features are dynamically identified and masked during training through a model redundancy feature lightweighting strategy. After the model is trained, the network recognition accuracy is calculated using the macro-average method. The recognition accuracy is then compared with a set threshold. If the recognition accuracy is greater than the set threshold, the performance is considered to meet the perception requirements, and the final user behavior recognition model is obtained. Otherwise, the performance is considered to not meet the perception requirements. When the performance does not meet the perception requirements, the number of iterations of the loss function is increased and the judgment is repeated until the performance meets the perception requirements.

2. The lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The formula for calculating the total loss function is as follows: ； ； in Represents the loss function. For the total loss function, This represents the label of the sample corresponding to the j-th action. This represents the probability of predicting the action as the j-th type. and These are the subcarrier coefficient and the antenna sparsity coefficient, respectively, where F is the number of subcarriers and S is the number of antennas. For the number of action categories, This represents the sequence mask for the i-th subcarrier. This represents the sequence mask for the j-th antenna.

3. The lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, Methods for processing the first time-domain dataset using sample feature enhancement strategies to obtain the second time-domain dataset and the first frequency-domain dataset include: Hampel filtering was applied to the first time-domain dataset to remove outliers that did not conform to the overall trend; Butterworth low-pass filter was applied to the first time-domain dataset to remove high-frequency noise and perform smoothing. For the first time-domain dataset, linear interpolation is used to fill in missing values ​​based on the average trend of the preceding and following data. Perform data regularization on the first time-domain dataset to obtain the second time-domain dataset; The obtained second time-domain dataset is processed by short-time Fourier transform to obtain the first frequency-domain dataset.

4. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The multi-level deep feature representation extraction strategy is specifically a dual-branch input structure, including a time-domain branch and a frequency-domain branch. The two branches are used to process signal features in the time domain and frequency domain, respectively. The time-domain branch receives the second time-domain dataset, and the frequency-domain branch receives the first frequency-domain dataset. The two branches perform feature extraction in parallel.

5. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 4, characterized in that, The temporal and frequency domain branches of the multi-level deep feature representation extraction strategy both employ deep feature extraction networks. These deep feature extraction networks consist of three separable convolutional layers: The first layer of the deep feature extraction network focuses on subcarrier-level feature extraction. Each subcarrier signal is individually input into the first layer of the deep feature extraction network to achieve feature extraction, and the output is a subcarrier feature sequence. The second layer of the deep feature extraction network is used to aggregate subcarrier features belonging to the same transmit-receive antenna pair, and to process the subcarrier feature sequence. Feature extraction, outputting antenna pair features; The third layer of the deep feature extraction network is oriented sample-level feature extraction, which stitches the antenna pair features into a whole and extracts features from it, outputting time-domain sample-level features and frequency-domain sample-level features. A non-linear activation function is used after each layer of the deep feature extraction network.

6. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The adaptive weighted fusion method is used to fuse time-domain sample-level features and frequency-domain sample-level features to obtain the calculation formula for the first fused feature, which is as follows: ； in The first fusion feature, and These are the first and second feature fusion parameters, respectively, both of which are trainable parameters. For time-domain sample-level features, These are frequency domain sample-level features.

7. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The method to obtain the third fusion feature by inputting the first fusion feature into the encoder is as follows: input the first fusion feature into the bidirectional Mamba network to obtain the second fusion feature, and then obtain the third fusion feature through the multi-head attention mechanism.

8. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The steps to output the probability distribution for each action category are as follows: The third fusion feature is input into the classifier after global average pooling. The classifier outputs the scores of each of the M types of actions, and after Softmax activation, it outputs the probability distribution of the action categories.

9. A lightweight user behavior recognition method based on a multi-level feature extraction algorithm according to claim 1, characterized in that, The method for dynamically identifying and masking low-contribution features during training, through a model redundancy feature lightweighting strategy, is as follows: During training, the trainable scalar corresponding to the subcarrier calculates the subcarrier sequence mask using the Sigmoid function. If the subcarrier sequence mask is 0, it is judged as low contribution, and low contribution subcarrier-level features are automatically suppressed. The trainable scalar corresponding to the antenna calculates the antenna sequence mask using the Sigmoid function. If the antenna sequence mask is 0, it is judged as low contribution, and low contribution antenna-level features are suppressed.

Images