A method and device for motor imagery EEG signal recognition based on deep generalization networks

Abstract

This invention discloses a method and device for recognizing motor imagery EEG signals based on a deep generalization network, belonging to the field of motor imagery EEG decoding technology. The method includes: acquiring multi-class, multi-channel motor imagery EEG signals; preprocessing the motor imagery EEG signals; dividing the preprocessed motor imagery EEG signals into multi-scale frequency bands, and extracting CSP features from each frequency band to obtain multi-scale CSP features; inputting the multi-scale CSP features into a pre-trained self-supervised convolutional autoencoder for feature optimization; and inputting the optimized multi-scale CSP features into a two-dimensional residual convolutional neural network for signal classification and recognition. This invention can improve the recognition effect of motor imagery EEG signals, enhance classification accuracy, and improve model generalization performance.

Description

Technical Field

[0001] This invention relates to the field of motor imagery EEG decoding technology, and in particular to a method and device for motor imagery EEG signal recognition based on deep generalization networks. Background Technology

[0002] With the development of brain-computer interface technology, decoding motor imagery EEG signals has become an important research direction for realizing human-computer interaction, rehabilitation assistance, and neurorehabilitation training. Motor imagery EEG signals refer to the brain electrical activity generated by an individual through imagining a specific motor task without actual limb movement. These signals typically have low signal-to-noise ratios, are non-stationary, and exhibit individual variability, posing a significant challenge to accurate decoding.

[0003] Traditional methods for recognizing motor imagery EEG signals primarily rely on manual feature extraction and shallow classifiers, such as those based on power spectral density, cosmic patterns, or filter bank cosmic patterns. While these methods can extract spatial and frequency domain features of EEG signals to some extent, they are sensitive to individual differences, have weak generalization ability, and struggle to maintain stable decoding performance across different subjects.

[0004] In recent years, deep learning technology has made significant progress in the field of EEG signal processing, especially methods such as convolutional neural networks, recurrent neural networks, and graph neural networks, which have been applied to motor imagery EEG decoding tasks. Deep learning methods can automatically learn complex spatiotemporal features from raw EEG signals, reducing reliance on manual feature design. However, existing deep learning methods still face many challenges in practical applications, such as: training deep models typically requires a large amount of labeled data, while EEG data acquisition is time-consuming and highly susceptible to individual differences among subjects. Significant differences in EEG signals between individuals lead to decreased performance of trained models on new subjects. Therefore, there is an urgent need for a motor imagery EEG signal recognition method that can improve cross-subject generalization ability while maintaining recognition accuracy. Summary of the Invention

[0005] The technical problem to be solved by this invention is: in order to solve the problem that the recognition of motor imagery EEG signals in the prior art is greatly affected by individual differences and has weak generalization ability across subjects, this invention provides a method and device for motor imagery EEG signal recognition based on deep generalization networks.

[0006] The technical solution adopted by this invention to solve its technical problem is: On one hand, the present invention provides a method for recognizing motor imagery EEG signals based on deep generalization networks, comprising: S1. Acquire multi-category, multi-channel motor imagery EEG signals; S2. Preprocess the acquired motor imagery EEG signals; S3. Perform multi-scale frequency band division on the preprocessed motor imagery EEG signal, and extract CSP features for each frequency band to obtain multi-scale CSP features. S4. Input the multi-scale CSP features into a pre-trained self-supervised convolutional autoencoder for feature optimization; S5. Input the optimized multi-scale CSP features into a two-dimensional residual convolutional neural network for signal classification and recognition.

[0007] An end-to-end deep generalization network framework was constructed by sequentially performing preprocessing, multi-scale frequency band segmentation and CSP feature extraction, self-supervised convolutional autoencoder feature optimization, and two-dimensional residual convolutional neural network classification. This method can not only extract multi-scale features from motor imagery EEG signals, but also enhance feature generalization ability through self-supervised pre-training. Finally, high-precision motor imagery classification is achieved using residual networks, significantly improving the robustness and accuracy of motor imagery EEG signal recognition.

[0008] Furthermore, the multi-scale frequency band division involves dividing the motor imagery EEG signal into multiple frequency bands of different widths, including: 2Hz bandwidth (8-10Hz, 10-12Hz); 4Hz bandwidth with 50% overlap (8-12Hz, 10-14Hz, 12-16Hz, 14-18Hz, 16-20Hz, 18-22Hz, 20-24Hz, 22-26Hz, 24-28Hz, 26-30Hz). 4Hz bandwidth with no overlap (4-8Hz, 30-34Hz); 6Hz bandwidth (14-20Hz, 20-26Hz).

[0009] Since the effective features of motor imagery EEG signals are not fixedly distributed in a single frequency band, but may be dispersed in multiple frequency ranges, this step uses different bandwidths and overlapping frequency bands to capture fine frequency information and wide-band overall variation features at the same time, thereby avoiding information omissions caused by single-band feature extraction.

[0010] Furthermore, in step S3, CSP feature extraction is performed on each frequency band to obtain multi-scale CSP features, including: S31. Window each frequency band to obtain the original EEG signal fragments. According to the preset motor imagery EEG signal categories, classify the original EEG signal fragments to form a sample set for each category, and calculate the normalized covariance matrix for each category. S32. Based on the normalized covariance matrices of each category, solve the generalized eigenvalue problem to obtain the projection matrix; S33. Based on the projection matrix, spatial projection is performed on the original EEG signal segments to obtain the projected signal; S34. Calculate the log-normalized variance characteristics of the projected signal; S35. Concatenate the log-normalized variance features of all categories and all frequency bands to obtain multi-scale CSP features.

[0011] Furthermore, the formulas for calculating the normalized covariance matrix of each category are as follows: ; In the formula, frequency band Next category The normalized covariance matrix; frequency band Subcategories The sample set; This is a single raw EEG signal fragment after windowing processing; This is the mean vector of each channel of a single raw EEG signal segment; This is the matrix transpose.

[0012] Furthermore, the formula for calculating the log-normalized variance characteristic is as follows: ; In the formula, frequency band ,category The corresponding log-normalized variance characteristics; A category index for EEG signals related to motor imagery; Frequency band index; This is the projected signal matrix; The number of feature vectors selected for each type of projection; Index for projected components; The first element in the projected signal matrix Time series vectors of projected components; for A dimensional real vector space.

[0013] By dividing the preprocessed motor imagery EEG signals into multi-scale frequency bands and performing CSP feature extraction within each band, discriminative information related to the motor imagery task can be fully extracted across different frequency ranges. Since the effective features of motor imagery EEG signals are not fixedly distributed in a single frequency band but may be dispersed across multiple frequency intervals, this step employs a segmentation method with different bandwidths and overlapping frequency bands. This allows for the simultaneous capture of fine-grained frequency information and wide-band overall variation features, thus avoiding information omissions caused by single-band feature extraction. Furthermore, the CSP method enhances the spatial separability between different motor imagery categories, making the extracted features more class-discriminative. Therefore, step S3 improves the spatial frequency domain representation capability of motor imagery EEG signals, providing a richer, more stable, and discriminative multi-scale feature foundation for subsequent feature optimization and classification.

[0014] Furthermore, step S4 includes: S41. Rearrange the multi-scale CSP features in row priority order to form an original three-dimensional tensor with a size of 28×32×1. S42. Generate a mask tensor with the same size as the original three-dimensional tensor and which follows a Bernoulli distribution; S43. Multiply the original 3D tensor with the mask tensor to obtain the damaged input; S44. The encoder receives the damaged input, first performs convolution operation on the damaged input through the first convolutional layer, and outputs an intermediate feature with a size of 28×28×32. Then, it performs convolution operation on the intermediate feature through the second convolutional layer, and outputs a latent representation with a dimension of 14×16×64. S45. The decoder upsamples the latent representation through transposed convolution, outputs a reconstructed tensor with a size of 28×32×1, and quantifies the difference between the reconstructed tensor and the original three-dimensional tensor through the mean square error loss function.

[0015] Furthermore, the formula for calculating the mean squared error loss function is as follows: ; In the formula, These are the learnable parameters of the encoder; These are the learnable parameters of the decoder; Let the mean squared error loss function be used. The number of EEG signal samples included in a single training batch; For sample index; For the first The original feature tensor of each sample; For the first The reconstructed tensor of each sample.

[0016] By inputting multi-scale CSP features into a pre-trained self-supervised convolutional autoencoder for feature optimization, the robustness and generalization ability of feature representation can be further enhanced. This step, through constructing damaged inputs and performing reconstruction learning, enables the convolutional autoencoder to learn the intrinsic structural relationships in multi-scale CSP features without relying on additional manual annotations, thereby suppressing the influence of noise interference and redundant features. During the reconstruction process, the encoder is forced to extract more stable, compact, and representative latent feature representations, which helps to reduce the impact of individual differences, non-stationarity, and low signal-to-noise ratio in motor imagery EEG signals on recognition results. Therefore, step S4 can improve feature quality, making the optimized features more robust to interference and adaptable across subjects, thereby improving the recognition stability of subsequent classification models.

[0017] Furthermore, step S5 includes: S51. Input the optimized multi-scale CSP features into the first residual block to obtain the first residual connection features; S52. Input the first residual connection feature into the max pooling layer to obtain the spatially downsampled feature; S53. Input the spatially downsampled features into the second residual block to obtain the second residual connection features; S54. Perform global average pooling on the second residual connection feature to obtain a one-dimensional global feature vector. S55. Input the one-dimensional global feature vector into the fully connected layer, output the probability of each preset motor imagery EEG signal category, and take the category with the highest probability as the EEG signal recognition result.

[0018] By inputting optimized multi-scale CSP features into a two-dimensional residual convolutional neural network for signal classification and recognition, local structural information within the optimized features can be further extracted. The convolutional layers in the residual blocks can abstract the features layer by layer, enhancing the model's ability to perceive subtle differences between different motor imagery categories. Residual connections can alleviate gradient vanishing and feature degradation problems that may occur during network deepening, improving the stability of model training. Max pooling and global average pooling operations can reduce feature dimensionality and the number of parameters, reducing the risk of overfitting and improving the model's generalization ability. Finally, the probabilities of various motor imagery EEG signals are output through fully connected layers and the Softmax function, achieving accurate identification of different motor imagery categories. Therefore, step S5 can improve model training stability and recognition reliability while ensuring classification accuracy, further enhancing the accuracy of motor imagery EEG signal classification results.

[0019] Furthermore, the motor imagery EEG signals include left-hand motor imagery EEG signals, right-hand motor imagery EEG signals, double-fist motor imagery EEG signals, and double-foot motor imagery EEG signals.

[0020] On the other hand, the present invention provides an electronic device including a processor and a memory, wherein the memory stores a computer program, and the computer program executes the motion imagery EEG signal recognition method based on deep generalization networks when invoked by a computer.

[0021] The electronic device runs the aforementioned EEG signal recognition method based on the program stored in the memory. With the help of the processor, it realizes multi-scale CSP feature extraction and deep generalization network operation, which can efficiently complete the real-time analysis of motor imagery EEG signals. The device has strong ability to cope with individual differences of subjects and environmental noise interference, and has excellent recognition accuracy and operational stability. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Figure 1 This is a flowchart of the method for recognizing motor imagery EEG signals based on deep generalization networks according to the present invention; Figure 2 This is a diagram illustrating the overall model architecture of the motion imagery EEG signal recognition method based on deep generalization networks of the present invention. Figure 3 This is a schematic diagram of the multi-scale co-space feature extraction network of the present invention; Figure 4 This is a schematic diagram of the pre-trained self-supervised convolutional autoencoder of the present invention; Figure 5 This is a schematic diagram of the two-dimensional residual convolutional neural network of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0025] Example 1: like Figure 1 and 2 As shown, this embodiment provides a method for recognizing motor imagery EEG signals based on deep generalization networks, including: S1. Acquire multi-category, multi-channel motor imagery EEG signals from the subject. In this embodiment, based on the BCI2000 system, EEG signals were acquired at a sampling frequency of 160Hz, recording 64 channels of motor imagery EEG signals. The signal types include four types: left-hand motor imagery EEG signals, right-hand motor imagery EEG signals, double-fist motor imagery EEG signals, and double-foot motor imagery EEG signals.

[0026] S2. Perform routine signal preprocessing on the motor imagery EEG signals, such as bandpass filtering, baseline correction, and artifact removal, and randomly select 80% and 20% of them to divide them into training and test sample sets.

[0027] S3, such as Figure 3 As shown, the preprocessed motor imagery EEG signal is divided into multiple frequency bands at various scales, allowing for parallel processing using multiple frequency band filters of different scales. The resulting frequency bands fall into four categories: 2Hz bandwidth (8-10Hz, 10-12Hz); 4Hz bandwidth with 50% overlap (8-12Hz, 10-14Hz, 12-16Hz, 14-18Hz, 16-20Hz, 18-22Hz, 20-24Hz, 22-26Hz, 24-28Hz, 26-30Hz); 4Hz bandwidth with no overlap (4-8Hz, 30-34Hz); and 6Hz bandwidth (14-20Hz, 20-26Hz).

[0028] For EEG signals related to motor imagery, the discriminative information is not concentrated in a single fixed frequency band, but rather distributed across multiple relevant frequency ranges. The EEG activity most relevant to motor imagery tasks is primarily concentrated in the 8–30 Hz frequency range. Within this range, brain activity exhibits unique characteristics closely related to the motor imagery task at different frequency intervals. Therefore, dividing motor imagery EEG signals into multiple frequency bands of varying widths facilitates the extraction of richer feature information and the acquisition of more refined feature representations.

[0029] Then, CSP features are extracted for each frequency band to obtain multi-scale CSP features, specifically including: S31, For each frequency band Windowing was performed to obtain the original EEG signal fragments. ,in, Represents the number of channels. This represents the number of sampling points within the time window. Based on preset categories of motor imagery EEG signals. (i.e., the four types in step S1) The raw EEG signal fragments are classified to form a sample set for each category. And calculate the normalized covariance matrix for each category, using the following formula: ; In the formula, frequency band Next category The normalized covariance matrix; frequency band Subcategories The sample set; A single raw EEG signal segment after windowing processing (dimension) ); The mean vector of each channel of a single raw EEG signal segment (dimensions) ); This is the matrix transpose.

[0030] S32. Based on the normalized covariance matrices of each category, solve the generalized eigenvalue problem and select the largest eigenvalue. The eigenvalues ​​and the smallest last one eigenvalues ​​are used to obtain the projection matrix. .

[0031] S33. Based on the projection matrix, analyze the raw EEG signal fragments. Perform spatial projection to obtain the projected signal. .

[0032] S34. Calculate the log-normalized variance characteristic of the projected signal. The calculation formula is as follows: ; In the formula, frequency band ,category The corresponding log-normalized variance characteristics; A category index for EEG signals related to motor imagery; Frequency band index; This is the projected signal matrix; The number of feature vectors selected for each type of projection determines the number of components after projection. ; Index for projected components; The first element in the projected signal matrix Time series vectors of projected components; for A dimensional real vector space.

[0033] S35, proceed with each category in turn. and each frequency band Repeat steps S31 to S34 above to obtain log-normalized variance feature vectors for all categories and frequency bands. Then, connect these vectors to form the final multi-scale CSP features.

[0034] S4, such as Figure 4 As shown, multi-scale CSP features are input into a pre-trained self-supervised convolutional autoencoder for feature optimization. Specifically, the steps include: S41, The multi-scale CSP features of each sample are initially represented as: Subsequently, the multi-scale CSP features are rearranged in row-major order to form an original three-dimensional tensor of size 28×32×1. .

[0035] S42. To simulate the situation of feature loss under self-supervised conditions and force the network to automatically repair it, a model is constructed that is identical to the original 3D tensor. Mask tensors of the same size and following a Bernoulli distribution (The percentage of zeroing is 20%).

[0036] S43, convert the original three-dimensional tensor With mask tensor Perform element-wise multiplication to obtain the corrupted input.

[0037] S44. The encoder receives the aforementioned damaged input (size 28×32×1), first expands the damaged input through a first convolutional layer (using a 3×3 convolutional kernel, stride 1, and 32 channels), outputting intermediate features of size 28×28×32. Then, it performs convolution operations on the intermediate features through a second convolutional layer (using a 3×3 convolutional kernel, stride 2, and increasing the number of channels to 64), outputting a latent representation of dimension 14×16×64. .

[0038] S45, The decoder processes the latent representation through transposed convolution. Upsampling is performed, and batch normalization is applied after transposed convolution. The final decoder output is a reconstructed tensor with a size of 28×32×1. Tensor reconstruction is achieved through mean squared error loss function quantization. With the original three-dimensional tensor The difference between the values ​​is used to assess the quality of the reconstruction. The formula for calculating the mean squared error loss function is as follows: ; In the formula, These are the learnable parameters of the encoder; These are the learnable parameters of the decoder; Let the mean squared error loss function be used. The number of EEG signal samples included in a single training batch; For sample index; For the first The original feature tensor of each sample corresponds to the original three-dimensional tensor in step S41. For the first The reconstructed tensor of each sample.

[0039] By minimizing the aforementioned loss function, the network automatically learns its intrinsic structure from unlabeled data in a self-supervised manner, forcing the encoder to extract feature representations with stronger discriminative power, thereby effectively capturing key information relevant to the motion visualization task. After training, the decoder is discarded, and the latent representation output by the encoder is used. These optimized features are used for subsequent classification.

[0040] S5, such as Figure 5 As shown, the optimized multi-scale CSP features are input into a two-dimensional residual convolutional neural network for signal classification and recognition. The number of convolutional layers in the two-dimensional residual convolutional neural network is expanded to four layers to achieve higher-dimensional information representation, and residual connections are used to alleviate the gradient vanishing problem in deep networks, enabling accurate classification of the optimized features.

[0041] Specifically, the steps include the following: S51. Input the optimized multi-scale CSP features into the first residual block to obtain the first residual connectivity features. The first residual block contains two 3×3 convolutional layers (Conv1 and Conv2), which output 64 feature channels. The output of the first convolutional layer, Conv1, is denoted as... The output of the second convolutional layer Conv2 is denoted as Then the first residual connection feature is This is used to alleviate the gradient vanishing problem and enhance feature reuse.

[0042] S52, Connect the first residual to the feature The input is fed into a max pooling layer (2×2, stride 2) to obtain spatially downsampled features, which are then used to extract more robust local features.

[0043] S53. Input the spatially downsampled features into the second residual block to obtain the second residual connectivity features. The second residual block contains two 3×3 convolutional layers (Conv3 and Conv4), which output 128 feature channels. The output of the third convolutional layer, Conv3, is denoted as... The output of the fourth convolutional layer, Conv4, is denoted as Then the second residual connection characteristic is This is used to further alleviate the deep gradient problem.

[0044] S54, Second Residual Connection Feature Global average pooling is performed to obtain a one-dimensional global feature vector, which significantly reduces the number of parameters in the fully connected layer.

[0045] S55. Input the one-dimensional global feature vector into the fully connected layer, followed by the Softmax activation function, and output each preset motor imagery EEG signal category. The probability of the EEG signal is calculated, and the category with the highest probability is taken as the EEG signal recognition result.

[0046] The motor imagery EEG signal recognition method in this embodiment can adaptively capture key feature information and suppress useless features without much human intervention. This method features fast training speed and good representation effect. Experimental verification shows that this embodiment effectively improves the classification accuracy of motor imagery EEG signals. Table 1 shows the recognition accuracy recorded by different methods.

[0047] Table 1 Accuracy of different methods

[0048] Example 2: This embodiment provides a storage medium storing a method for recognizing motor imagery EEG signals based on deep generalization networks. When executed, the program for recognizing motor imagery EEG signals based on deep generalization networks implements the steps of the method described above. Since this storage medium employs all the technical solutions of the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be elaborated upon further here.

[0049] Example 3: This embodiment provides an electronic device, including: a memory and a processor; the memory stores at least one program instruction; the processor loads and executes the at least one program instruction to implement the motor imagery EEG signal recognition method based on deep generalization networks provided in Embodiment 1.

[0050] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for recognizing motor imagery EEG signals based on deep generalization networks, characterized in that, include: S1. Acquire multi-category, multi-channel motor imagery EEG signals; S2. Preprocess the acquired motor imagery EEG signals; S3. Perform multi-scale frequency band division on the preprocessed motor imagery EEG signal, and extract CSP features for each frequency band to obtain multi-scale CSP features. S4. Input the multi-scale CSP features into a pre-trained self-supervised convolutional autoencoder for feature optimization; S5. Input the optimized multi-scale CSP features into a two-dimensional residual convolutional neural network for signal classification and recognition.

2. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 1, characterized in that, The multi-scale frequency band division involves dividing the motor imagery EEG signal into multiple frequency bands of different widths, including: 2Hz bandwidth (8-10Hz, 10-12Hz); 4Hz bandwidth with 50% overlap (8-12Hz, 10-14Hz, 12-16Hz, 14-18Hz, 16-20Hz, 18-22Hz, 20-24Hz, 22-26Hz, 24-28Hz, 26-30Hz). 4Hz bandwidth with no overlap (4-8Hz, 30-34Hz); 6Hz bandwidth (14-20Hz, 20-26Hz).

3. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 1, characterized in that, In step S3, CSP features are extracted for each frequency band to obtain multi-scale CSP features, including: S31. Window each frequency band to obtain the original EEG signal fragments. According to the preset motor imagery EEG signal categories, classify the original EEG signal fragments to form a sample set for each category, and calculate the normalized covariance matrix for each category. S32. Based on the normalized covariance matrices of each category, solve the generalized eigenvalue problem to obtain the projection matrix; S33. Based on the projection matrix, spatial projection is performed on the original EEG signal segments to obtain the projected signal; S34. Calculate the log-normalized variance characteristics of the projected signal; S35. Concatenate the log-normalized variance features of all categories and all frequency bands to obtain multi-scale CSP features.

4. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 3, characterized in that, The formulas for calculating the normalized covariance matrix of each category are as follows: ； In the formula, frequency band Next category The normalized covariance matrix; frequency band Subcategories The sample set; This is a single raw EEG signal fragment after windowing processing; This is the mean vector of each channel of a single raw EEG signal segment; This is the matrix transpose.

5. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 3, characterized in that, The formula for calculating the log-normalized variance feature is as follows: ； In the formula, frequency band ,category The corresponding log-normalized variance characteristics; A category index for EEG signals related to motor imagery; Frequency band index; This is the projected signal matrix; The number of feature vectors selected for each type of projection; Index for projected components; The first element in the projected signal matrix Time series vectors of projected components; for A dimensional real vector space.

6. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 1, characterized in that, Step S4 includes: S41. Rearrange the multi-scale CSP features in row priority order to form an original three-dimensional tensor with a size of 28×32×1. S42. Generate a mask tensor with the same size as the original three-dimensional tensor and which follows a Bernoulli distribution; S43. Multiply the original 3D tensor with the mask tensor to obtain the damaged input; S44. The encoder receives the damaged input, first performs convolution operation on the damaged input through the first convolutional layer, and outputs an intermediate feature with a size of 28×28×32. Then, it performs convolution operation on the intermediate feature through the second convolutional layer, and outputs a latent representation with a dimension of 14×16×64. S45. The decoder upsamples the latent representation through transposed convolution, outputs a reconstructed tensor with a size of 28×32×1, and quantifies the difference between the reconstructed tensor and the original three-dimensional tensor through the mean square error loss function.

7. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 6, characterized in that, The formula for calculating the mean squared error loss function is as follows: ； In the formula, These are the learnable parameters of the encoder; These are the learnable parameters of the decoder; Let the mean squared error loss function be used. The number of EEG signal samples included in a single training batch; For sample index; For the first The original feature tensor of each sample; For the first The reconstructed tensor of each sample.

8. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 1, characterized in that, Step S5 includes: S51. Input the optimized multi-scale CSP features into the first residual block to obtain the first residual connection features; S52. Input the first residual connection feature into the max pooling layer to obtain the spatially downsampled feature; S53. Input the spatially downsampled features into the second residual block to obtain the second residual connection features; S54. Perform global average pooling on the second residual connection feature to obtain a one-dimensional global feature vector. S55. Input the one-dimensional global feature vector into the fully connected layer, output the probability of each preset motor imagery EEG signal category, and take the category with the highest probability as the EEG signal recognition result.

9. The method for recognizing motor imagery EEG signals based on deep generalization networks according to claim 1, characterized in that, The motor imagery EEG signals include left-hand motor imagery EEG signals, right-hand motor imagery EEG signals, double-fist motor imagery EEG signals, and double-foot motor imagery EEG signals.

10. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the computer program, when invoked by a computer, executes the motor imagery EEG signal recognition method based on deep generalization networks as described in claims 1 to 9.

Images