Electroencephalogram signal feature extraction method and system based on adaptive time-frequency feature collaboration

Abstract

The invention discloses an electroencephalogram signal feature extraction method and system based on self-adaptive time-frequency feature collaboration, relates to the field of electroencephalogram signal processing, and solves the problems that in existing electroencephalogram signal feature extraction, fusion is carried out in a simple splicing or averaging mode, complementarity and correlation between time domain and frequency domain features are not fully considered, and time domain and frequency domain features are not fully considered. In order to solve the problems of insufficient overall characterization capability and classification performance of signals and the like, the method comprises the following steps: acquiring electroencephalogram signal data, preprocessing the electroencephalogram signal data, constructing a dynamic residual network in combination with dynamic convolution, and training a model based on the preprocessed electroencephalogram signal data; performing a contrast test on the dynamic residual network and other deep learning network models, and designing an adaptive time-frequency collaborative network model to fuse time-domain features and frequency-domain features in the model; and electroencephalogram signal feature extraction based on adaptive time-frequency feature cooperation is completed. The method is also suitable for the field of electroencephalogram signal feature extraction and fusion.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to an EEG signal feature extraction method and system based on adaptive time-frequency feature collaboration. Background Art

[0002] In recent years, electroencephalography (EEG) has been widely used in the research and diagnosis of ASD as a non-invasive, high-temporal-resolution tool. However, EEG signals have the characteristics of nonlinearity, non-stationarity, and high dimensionality. Traditional feature extraction and classification methods have certain limitations when processing complex EEG signals.

[0003] In the feature extraction of EEG signals, firstly, EEG signals are highly complex and dynamic, and the current ability to capture the temporal dynamic change characteristics of EEG signals is insufficient. Secondly, in terms of the fusion and expression of signal features, current research mostly uses simple splicing or averaging to fuse, and fails to fully consider the complementarity and correlation between time domain and frequency domain features, resulting in deficiencies in the overall representation ability and classification performance of the signal. There are certain limitations in achieving coordinated expression and dynamic fusion of time domain and frequency domain at the feature level. Summary of the invention

[0004] In order to solve the problem that the feature extraction of EEG signals in the prior art is mostly carried out by simple splicing or averaging, and the complementarity and correlation between time domain and frequency domain features are not fully considered, resulting in deficiencies in the overall characterization ability and classification performance of the signal, and there are certain limitations in realizing the coordinated expression and dynamic fusion of time domain and frequency domain at the feature level.

[0005] To solve the above technical problems, the present invention is achieved through the following technical solutions: Solution 1: The present invention proposes a method for extracting EEG signal features based on adaptive time-frequency feature collaboration, and the method comprises: S1, obtaining EEG signal data, preprocessing the EEG signal data, and dividing the preprocessed EEG signal data into a training set and a test set; S2, combining dynamic convolution to build a dynamic residual network, and training the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by S1; S3. Compare the dynamic residual network with other deep learning network models, and design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; S4. Based on the above steps, an adaptive time-frequency collaborative network model is constructed to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

[0006] Furthermore, a preferred implementation is provided, in which the preprocessing of the EEG signal data in S1 is achieved by downsampling, filtering, re-referencing, and standardization.

[0007] Furthermore, a preferred implementation is provided, in which the ratio of the training set to the test set in S1 is 8:2.

[0008] Further, a preferred implementation is provided, in which the method for designing an adaptive time-frequency collaborative network model in S3 to fuse the features of the time domain and the frequency domain in the model is: S3.1, linearly project the time domain features and frequency domain features respectively for query, key and value through the pre-projection layer; S3.2, split the time domain features and frequency domain features after linear projection into multiple heads, and calculate the attention weights between the time domain query and the frequency domain key, and between the frequency domain query and the time domain key; S3.3. Add the attention weight parameters time_attention_weight and freq_attention_weight. Use the calculated attention weights to perform weighted summation on the matrix of corresponding values ​​to obtain the fused time domain and frequency domain feature representation.

[0009] Furthermore, a preferred implementation is provided, in which S3.2 also includes a step of normalizing the attention weights using a Softmax function.

[0010] Furthermore, a preferred implementation is provided, wherein adding the attention weight parameter in S3.3 also includes a step of constraining the attention weight parameter using a Sigmoid activation function.

[0011] Further, a preferred implementation is provided, in which the step of constructing an adaptive time-frequency collaborative network model in S4 includes: S4.1. Pass the time domain input time_input through the time domain feature extraction model to obtain the time domain feature time_feature; S4.2. The frequency domain input freq_input is passed through the frequency domain feature extraction model to obtain the frequency domain feature freq_feature; S4.3. The time domain features and frequency domain features are fused using the adaptive cross-domain attention mechanism; S4.4. Calculate the average pooling of the fused feature map in the sequence length dimension to obtain a feature vector of fixed length.

[0012] Solution 2: An EEG signal feature extraction system based on adaptive time-frequency feature collaboration, the system comprising: A preprocessing module, used to obtain EEG signal data, preprocess the EEG signal data, and divide the preprocessed EEG signal data into a training set and a test set; The training module is used to construct a dynamic residual network in combination with dynamic convolution, and train the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by the preprocessing module; The fusion module is used to compare the dynamic residual network with other deep learning network models, and to design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; The feature extraction module is used to construct an adaptive time-frequency collaborative network model based on the above modules to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

[0013] Solution three: A computer device includes a processor and a storage medium, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes any one of the methods described in Solution one.

[0014] Solution 4: A computer storage medium is used to store a computer program. When the computer program is read by a computer, the computer executes any one of the methods described in Solution 1.

[0015] The present invention is beneficial in that: The method for extracting EEG signal features based on adaptive time-frequency feature synergy described in the present invention enhances the ability to capture the temporal dynamic change characteristics of EEG signals by combining dynamic convolution to construct a dynamic residual network. In addition, a feature fusion strategy is designed to achieve the interaction and fusion between time domain features and frequency domain features, thereby improving the accuracy and robustness of classification.

[0016] The present invention is also applicable to the field of EEG signal feature extraction and fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a model architecture diagram of an EEG signal feature extraction method based on adaptive time-frequency feature collaboration as described in implementation mode 1.

[0018] Figure 2 This is a model architecture diagram of a dynamic residual block of an EEG signal feature extraction method based on adaptive time-frequency feature collaboration as described in Implementation Method 1.

[0019] Figure 3 This is a logical diagram of dynamic convolution calculation of an EEG signal feature extraction method based on adaptive time-frequency feature collaboration as described in implementation example eleven.

[0020] Figure 4 This is a structural framework diagram of an adaptive cross-domain attention model for an EEG signal feature extraction method based on adaptive time-frequency feature collaboration as described in implementation example eleven.

[0021] Figure 5 Schematic diagram of the training set and test set loss fitting curves of data set 1 in the EEG signal feature extraction method based on adaptive time-frequency feature collaboration described in implementation mode eleven.

[0022] Figure 6 This is a schematic diagram of the accuracy fitting curve of the training set and the test set of data set 1 in the EEG signal feature extraction method based on adaptive time-frequency feature collaboration described in implementation mode eleven.

[0023] Figure 7 Schematic diagram of the loss fitting curves of the training set and the test set of data set 2 in the method for extracting EEG signal features based on adaptive time-frequency feature collaboration described in Implementation Example 11.

[0024] Figure 8 This is a schematic diagram of the accuracy fitting curve of the training set and the test set of data set 2 in the EEG signal feature extraction method based on adaptive time-frequency feature collaboration described in Implementation Example 11. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of the implementation methods of the present application clearer, the technical solutions in the implementation methods of the present application will be clearly and completely described below in conjunction with the drawings in the implementation methods of the present application. Obviously, the described implementation methods are only part of the implementation methods of the present application, not all of the implementation methods.

[0026] Embodiment 1: This embodiment provides a method for extracting EEG signal features based on adaptive time-frequency feature collaboration, and the method for extracting EEG signal features includes: S1, obtaining EEG signal data, preprocessing the EEG signal data, and dividing the preprocessed EEG signal data into a training set and a test set; S2, combining dynamic convolution to build a dynamic residual network, and training the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by S1; S3. Compare the dynamic residual network with other deep learning network models, and design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; S4. Based on the above steps, an adaptive time-frequency collaborative network model is constructed to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

[0027] Implementation method 2: This implementation method further limits the EEG signal feature extraction based on adaptive time-frequency feature collaboration described in implementation method 1. In S1, the EEG signal data is preprocessed by downsampling, filtering, re-referencing, and standardization.

[0028] Implementation method 3: This implementation method is to implement the EEG signal characteristics based on adaptive time-frequency feature coordination described in implementation method 1. As a further limitation of feature extraction, the ratio of the training set to the test set described in S1 is 8:2.

[0029] Implementation 4: This implementation is a further limitation of the EEG signal feature extraction based on adaptive time-frequency feature collaboration described in Implementation 1. The method for designing an adaptive time-frequency collaborative network model in S3 to fuse the features of the time domain and frequency domain in the model is: S3.1, linearly project the time domain features and frequency domain features respectively for query, key and value through the pre-projection layer; S3.2, split the time domain features and frequency domain features after linear projection into multiple heads, and calculate the attention weights between the time domain query and the frequency domain key, and between the frequency domain query and the time domain key; S3.3. Add the attention weight parameters time_attention_weight and freq_attention_weight. Use the calculated attention weights to perform weighted summation on the matrix of corresponding values ​​to obtain the fused time domain and frequency domain feature representation.

[0030] Implementation method five: This implementation method further limits the EEG signal feature extraction based on adaptive time-frequency feature collaboration described in implementation method four. S3.2 also includes the step of using the Softmax function to normalize the attention weight.

[0031] Implementation method six: This implementation method is a further limitation of the EEG signal feature extraction based on adaptive time-frequency feature collaboration described in implementation method four. Adding the attention weight parameter in S3.3 also includes the step of constraining the attention weight parameter using the Sigmoid activation function.

[0032] Embodiment 7: This embodiment further limits the EEG signal feature extraction based on adaptive time-frequency feature collaboration described in Embodiment 1. In S4, the adaptive time-frequency collaboration network model is constructed, including: S4.1. Pass the time domain input time_input through the time domain feature extraction model to obtain the time domain feature time_feature; S4.2. The frequency domain input freq_input is passed through the frequency domain feature extraction model to obtain the frequency domain feature freq_feature; S4.3. The time domain features and frequency domain features are fused using the adaptive cross-domain attention mechanism; S4.4. Calculate the average pooling of the fused feature map in the sequence length dimension to obtain a feature vector of fixed length.

[0033] Embodiment 8: This embodiment proposes an EEG signal feature extraction system based on adaptive time-frequency feature collaboration, the system comprising: A preprocessing module, used to obtain EEG signal data, preprocess the EEG signal data, and divide the preprocessed EEG signal data into a training set and a test set; The training module is used to construct a dynamic residual network in combination with dynamic convolution, and train the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by the preprocessing module; The fusion module is used to compare the dynamic residual network with other deep learning network models, and to design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; The feature extraction module is used to construct an adaptive time-frequency collaborative network model based on the above modules to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

[0034] Embodiment 9. This embodiment proposes a computer device, including a processor and a storage medium, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method described in any one of embodiments 1 to 7.

[0035] Embodiment 10: This embodiment proposes a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method described in any one of Embodiments 1 to 7.

[0036] Implementation 11. This implementation includes an example, which is used to explain the above implementation. Figures 1 to 8 This embodiment is described. The purpose of this embodiment is to propose a method for extracting features of autism EEG signals based on adaptive time-frequency feature synergy. By extracting time domain features and frequency domain features for fusion, refined feature extraction and efficient fusion of autism EEG signals are achieved, thereby improving the accuracy and robustness of classification.

[0037] like Figure 1As shown, the present invention proposes a method for extracting features of autism EEG signals based on adaptive time-frequency feature collaboration, which specifically includes the following steps: Step 1: Acquire data and preprocess the data to reduce the impact of noise and artifacts.

[0038] Step 2: In order to enhance the ability to capture the time-varying features of EEG signals, we combined dynamic convolution to construct a dynamic residual network. In addition, in order to jointly express the time domain and frequency domain features of EEG signals at the feature level, an adaptive cross-domain attention mechanism, namely ACDA, was designed. Through this improvement, the ability to capture the time-varying features of EEG signals is enhanced while also achieving the joint expression of time domain features and frequency domain features.

[0039] Furthermore, step one includes the following steps: Step 1: Data preprocessing: Use EEGLAB of MATLAB software to preprocess the EEG signals, including: Downsampling: In order to reduce the amount of calculation, the sampling rate of the EEG signal is reduced to 256HZ. Filtering: A Butterworth bandpass filter is used with cutoff frequencies set to 0.5 Hz and 60 Hz to remove low-frequency drift and high-frequency noise. A notch filter is used to suppress power frequency interference, such as 50 Hz or 60 Hz.

[0040] Re-reference: To improve signal quality, enhance comparability across studies, and improve the accuracy of signal analysis, EEG signals were referenced to a whole-brain average.

[0041] Artifact removal: After preprocessing, perform ICAlable to sort out possible physiological artifacts such as electrooculography and electromyography and automatically mark and remove them.

[0042] Standardization: To avoid a certain feature value having too great an impact on the model, Z-Score standardization is used to standardize the preprocessed data.

[0043] Step 2: Divide the preprocessed data into training set and test set in a ratio of 8:2, and process the long time series data into 2-second segments.

[0044] Furthermore, step 2 includes the following steps: Step 2.1: Construct the frequency domain feature extraction model FrequencyDomainModel. The input is the spectrum after SFTF processing. The specific settings are as follows: Conv2d: Performs a convolution operation on the input spectrogram to increase the number of channels to 128.

[0045] BatchNorm2d: Use nn.BatchNorm2d(128) to normalize the feature map after convolution to stabilize the training process of the model.

[0046] Activation='ReLU': Use the ReLU activation function to increase the nonlinear representation capability of the model. This is a nonlinear activation function commonly used in deep learning.

[0047] CBAM: Then combined with CBAM, the Convolutional Block Attention Module attention mechanism, the channel and spatial dimensions of the feature map are enhanced to suppress unimportant features.

[0048] Flattening: Flatten the enhanced feature map into a tensor of shape [batch_size, channels, height *width] for subsequent processing.

[0049] Construct Transformer encoder: Use nn.TransformerEncoder to construct Transformer encoder, where num_layers=2, embed_dim=128, num_heads=4 are defined.

[0050] Step 2: Create a time domain feature extraction model DyTemporalConvNet, which consists of three DyTemporalBlocks. The input is the preprocessed EEG signal time series. Figure 2 and Figure 3 As shown in the figure, the calculation process of dynamic residual block and dynamic convolution is respectively. The specific settings of each dynamic TemporalBlock are: Dynamic_conv1d: Use Dynamic_conv1d with a variable number of convolution kernels, parameter K and temperature parameter temperature to adapt to different input modes.

[0051] BatchNorm2d: Use nn.BatchNorm2d(out_channels) to normalize the convolutional feature map.

[0052] activation='ReLU': The activation function uses ReLU, or rectified linear unit, which is a nonlinear activation function commonly used in deep learning.

[0053] Residual connection: If input_channels is inconsistent with output_channels, it will be adjusted through the linear layer (self.downsample) and a residual connection will be performed to alleviate the gradient vanishing problem.

[0054] Application of the time domain feature extraction model DyTemporalConvNet: The temperature parameter temperature in Dynamic_conv1d controls the smoothness of the softmax function and affects the probability distribution of the convolution kernel selection. The dynamic residual network significantly improves the understanding and processing capabilities of time series signals in the time domain feature extraction model, enhances the ability to capture the temporal changes of EEG signals, and lays a solid foundation for subsequent feature fusion and classification tasks.

[0055] Step 3: Figures 5 to 8 As shown, Figure 5 Schematic diagram of the training set and test set loss fitting curve of data set 1. Figure 6 Schematic diagram of the accuracy fitting curve of the training set and test set of data set 1. Figure 7 Schematic diagram of the training set and test set loss fitting curve of data set 2. Figure 8 Schematic diagram of the accuracy fitting curve of the training set and test set of data set 2.

[0056] In step 3, an adaptive cross-domain attention mechanism, namely ACDA, is constructed to fuse the time domain and frequency domain features. This part specifically includes: nn.Linear: Through the first projection layer, the time domain features and frequency domain features are linearly projected into query, key, and value, respectively.

[0057] Multi-head attention mechanism: Split the projected features into multiple heads to improve the expressiveness of the model.

[0058] attention_weight: Calculates the attention weights between time domain query and frequency domain key, and between frequency domain query and time domain key.

[0059] Normalization and Regularization: The attention weights are normalized using the Softmax function, and Dropout is applied to prevent overfitting.

[0060] Adaptive weighting: Introduce learnable attention weight parameters time_attention_weight and freq_attention_weight to enhance the model's adaptability.

[0061] Weighted summation: Use the calculated attention weight attention_weight to perform weighted summation on the corresponding value matrix to obtain the fused time domain and frequency domain feature representation.

[0062] Learnable balance parameter alpha: Use the Sigmoid activation function to constrain the parameter alpha, and the value range is between (0, 1). The optimal parameter is automatically selected as the model is trained.

[0063] Weighted fusion: Use fused_output = alpha * out_time + (1 - alpha) * out_freq to balance the importance of time domain and frequency domain features and obtain the final fused feature representation of the two features.

[0064] Application of adaptive cross-domain attention mechanism: The adaptive cross-domain attention mechanism plays a bridging role in the entire model. By using the time domain features as the query to focus on the key and value of the frequency domain features, and vice versa, it realizes the joint expression of features, deeply understands the intrinsic connection between the two features of EEG signals, and realizes adaptive feature fusion by dynamically adjusting the importance of features. It enhances the expression and generalization capabilities of the model and improves the accuracy and robustness of the task.

[0065] Step 4: Combining steps 1, 2, and 3, create the class Time-frequency synergy (nn.Module) class to build a complete adaptive time-frequency synergy network model Time-frequency synergy. This part specifically includes: Time domain feature extraction: The time domain input time_input is passed through the time domain feature extraction model to obtain the time domain feature time_feature.

[0066] Frequency domain feature extraction: The frequency domain input freq_input is passed through the frequency domain feature extraction model to obtain the frequency domain feature freq_feature.

[0067] Feature fusion: Use the adaptive cross-domain attention mechanism to fuse the time domain feature time_feature and the frequency domain feature freq_feature.

[0068] Global Average Pooling: Calculates the average pooling of the fused feature map in the sequence length dimension, dim=1 corresponds to the sequence length dimension, and obtains a feature vector of fixed length. self.fc1: The pooled feature vector is input into the fully connected layer self.fc1, and after the ReLU activation function, the hidden layer representation is obtained.

[0069] self.fc2: Finally, the output result of the model is obtained through the fully connected layer self.fc2.

[0070] Application of the adaptive time-frequency collaborative network model: Dynamic convolution is combined in the time domain feature extraction to construct a dynamic residual network, which can adaptively adjust the convolution kernel parameters and enhance the ability to capture the time-varying characteristics of EEG signals. At the same time, by fusing the time domain and frequency domain features, the dynamic changes and complex patterns of EEG signals are deeply revealed. And the time domain features and frequency domain features are effectively jointly represented. This method provides a more accurate and robust solution for the auxiliary diagnosis of EEG signals of autism spectrum disorder ASD.

[0071] Furthermore, step 4 includes the following steps: Step 1: The model of the present invention is trained using two data sets and compared with other deep learning models, achieving excellent performance.

[0072] Step 2: In order to verify the effectiveness of feature extraction of the network model, the performance of the feature extraction method of the present invention is compared with the traditional classifier and the combination of the traditional classifier and TFSNet on two data sets.

[0073] Those skilled in the art will appreciate that the above are only preferred embodiments of the present invention, and the various embodiments of the present disclosure and / or the features described in the claims may be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present disclosure. It is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art may still modify the technical solutions described in the aforementioned embodiments, or perform equivalent substitutions on some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

[0074] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. The method for extracting EEG signal features based on adaptive time-frequency feature synergy is characterized by: The EEG signal feature extraction method comprises: S1, obtaining EEG signal data, preprocessing the EEG signal data, and dividing the preprocessed EEG signal data into a training set and a test set; S2, combining dynamic convolution to build a dynamic residual network, and training the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by S1; S3. Compare the dynamic residual network with other deep learning network models, and design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; S4. Based on the above steps, an adaptive time-frequency collaborative network model is constructed to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

2. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 1, characterized in that: In S1, the EEG signal data is preprocessed by downsampling, filtering, re-referencing, and standardization.

3. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 1, characterized in that: The ratio of the training set to the test set described in S1 is 8:

2.

4. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 1, characterized in that: The method for designing an adaptive time-frequency collaborative network model in S3 to fuse the features of the time domain and frequency domain in the model is: S3.1, linearly project the time domain features and frequency domain features respectively for query, key and value through the pre-projection layer; S3.2, split the time domain features and frequency domain features after linear projection into multiple heads, and calculate the attention weights between the time domain query and the frequency domain key, and between the frequency domain query and the time domain key; S3.

3. Add the attention weight parameters time_attention_weight and freq_attention_weight. Use the calculated attention weights to perform weighted summation on the matrix of corresponding values ​​to obtain the fused time domain and frequency domain feature representation.

5. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 4, characterized in that: S3.2 also includes a step of normalizing the attention weights using a Softmax function.

6. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 4, characterized in that: Adding the attention weight parameter in S3.3 also includes the step of constraining the attention weight parameter using the Sigmoid activation function.

7. The method for extracting EEG signal features based on adaptive time-frequency feature synergy according to claim 1, characterized in that: The adaptive time-frequency collaborative network model constructed in S4 includes: S4.

1. Pass the time domain input time_input through the time domain feature extraction model to obtain the time domain feature time_feature; S4.2, the frequency domain input freq_input is passed through the frequency domain feature extraction model to obtain the frequency domain feature freq_feature, S4.3, use adaptive cross-domain attention mechanism to fuse time domain features and frequency domain features; S4.

4. Calculate the average pooling of the fused feature map in the sequence length dimension to obtain a feature vector of fixed length.

8. The EEG signal feature extraction system based on adaptive time-frequency feature synergy is characterized by: The system comprises: A preprocessing module, used to obtain EEG signal data, preprocess the EEG signal data, and divide the preprocessed EEG signal data into a training set and a test set; The training module is used to construct a dynamic residual network in combination with dynamic convolution, and train the frequency domain feature extraction model and the time domain feature extraction model based on the EEG signal data preprocessed by the preprocessing module; The fusion module is used to compare the dynamic residual network with other deep learning network models, and to design an adaptive time-frequency collaborative network model to fuse the time domain features and frequency domain features in the frequency domain feature extraction model and the time domain feature extraction model; The feature extraction module is used to construct an adaptive time-frequency collaborative network model based on the above modules to complete the EEG signal feature extraction of adaptive time-frequency feature collaboration.

9. A computer device comprising a memory and a processor, characterized in that: A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images