Method and system for identifying motor imagery electroencephalogram signals based on CLSTM-GAN-FBMEEGNet
By using the CLSTM-GAN-FBMEEGNet method in the recognition of motor imagination EEG signal, the problem of insufficient recognition accuracy and stability in the prior art is solved, and higher recognition accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202411429802.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-03-11
AI Technical Summary
The prior art lacks more accurate and reliable methods in the identification of electroencephalogram signals in motor imagination, and there are shortcomings in data augmentation, feature extraction and classifier design, resulting in the need to improve the recognition accuracy and stability.
The recognition method based on CLSTM-GAN-FBMEEGNet is adopted, and the data augmentation of MI-EEG signals is enhanced by convolutional LSTM generation adversarial network, initial feature extraction is performed using group filters, and a multi-task learning method based on category average standard deviation is constructed, and a convolutional neural network designs the FBMEEGNet model for further feature extraction and recognition.
The data enhancement effect and feature extraction of MI-EEG signals are improved, the recognition accuracy and robustness of neural networks are enhanced, and the problem of insufficient recognition accuracy and stability in the prior art is solved.
Smart Images
Figure CN119598305A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of motor imagery electroencephalogram (EEG) signal processing, and in particular to a method and system for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet. Background Art
[0002] With the development of science and technology, artificial intelligence and BCI technology have developed rapidly, and using thoughts to control external devices has become a key direction of scientific and technological development. Brain-computer interface (BCI) aims to establish a normal output channel that does not rely on peripheral nerves and muscles in the brain, and uses a series of electronic devices mainly based on computers to achieve direct communication and control of the brain. Motor imagery EEG (MI-EEG) recognition is a study based on brain-computer interface technology, which aims to realize the recognition and classification of human motor imagery intentions by capturing and analyzing EEG signals. It mainly includes EEG signal acquisition technology, motor imagery task paradigm design, feature extraction and classifier design, and real-time feedback system design. EEG signal acquisition technology can convert EEG signals into digital signals and store them, providing a basis for subsequent processing. The motor imagery task paradigm design enables the experimenter to perform motor imagery through a specific task paradigm to obtain the corresponding EEG signals. Feature extraction and classifier design involve extracting effective features from EEG signals and using machine learning methods for classification and recognition. Finally, the design of the real-time feedback system enables motor imagery EEG signals to recognize and control external devices such as robots or prostheses in real time. At present, the methods of using MI-EEG signals for human motor imagery intention recognition mainly include MI-EEG signal data enhancement, feature extraction and classification design. In terms of MI-EEG signal data enhancement, the three most common methods are generative adversarial networks (GAN), variational autoencoders (VAE) and data reorganization. The enhanced data can be used to improve the generalization ability of the model and improve its performance, but there is still a lot of room for improvement. In terms of MI-EEG signal feature extraction, the common motor imagery EEG signal feature extraction mainly includes time domain analysis, frequency domain analysis, time-frequency analysis and spatial analysis. However, due to the nonlinearity, dynamics and susceptibility to environmental noise and physiological interference of MI-EEG, the accuracy and stability of feature extraction may be reduced, and the accuracy of motor imagery intention recognition still has a lot of room for improvement. In terms of classifier design, relevant research on convolutional neural networks (CNN) has been widely carried out, mainly including convolutional layers, pooling layers, etc., but due to the shortcomings of overfitting, data imbalance and loss of some important features during the training process, the accuracy and generalization ability of the classifier still have a lot of room for improvement.
[0003] In summary, the current existing technology lacks a more accurate and reliable recognition method based on motor imagery EEG signals. Summary of the invention
[0004] 1. Technical issues to be resolved
[0005] In view of the shortcomings of the prior art, the present invention provides a method and system for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet, which solves the problems raised in the above background technology.
[0006] (II) Technical solution
[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:
[0008] A method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet, characterized by comprising:
[0009] Step 1: Acquire motor imagery EEG MI-EEG signal data, and pre-process the MIEEG signal data from multiple channels, wherein the MI-EEG signal data comes from a non-invasive wearable EEG acquisition device;
[0010] Step 2: Construct a convolutional LSTM generative adversarial network (CLSTM-GAN) to perform data enhancement on the MI-EEG signal data;
[0011] Step 3: Perform preliminary feature extraction on the data-enhanced MI-EEG signal through group filters;
[0012] Step 4: Construct a multi-task learning method based on the average standard deviation of the category, and use the group filter feature extraction result as the input network model FBMEEGNet. Input the features initially extracted by the group filter into the FBMEEGNet model to obtain the recognition result of motor imagery.
[0013] Furthermore, the preprocessing of the MI-EEG signal data from multiple channels in step 1 includes:
[0014] Step 101: Use a multi-stage Butterworth filter to filter the MI-EEG signal data (high-pass and low-pass filtering) to obtain EEG data within the relevant range of the motor imagery stage (0.5-70 Hz), thereby effectively eliminating high-frequency noise interference;
[0015] Step 102: resampling the MI-EEG signal data to 160 Hz to reduce the data volume to increase the subsequent operation speed;
[0016] Step 103: Crop the MI-EEG signal data from 0.5s to 5s to further reduce the data volume and remove noise.
[0017] Furthermore, the step 2 of performing data enhancement on the preprocessed MI-EEG signal data includes:
[0018] Step 201: Construct a generator (G(z)) network, including fully connected layers, convolutional layers, LSTM layers, activation layers, and maximum pooling layers. The probability that the generator correctly classifies the actual data is log(D(x)), and increasing log(1-D(G(z))) will make it easier for the discriminator to identify the fake samples generated by the generator.
[0019] Step 202: Construct a discriminator (D(x)) network, including fully connected layers, convolutional layers, batch normalization layers, and activation layers (Tanh, Sigmoid) and other processing.
[0020] Step 203: G and D are in a game-adversarial relationship. During the game, the pseudo samples generated by the generator gradually approach the real data, and the judgment ability of the discriminator gradually increases, eventually reaching a dynamic balance.
[0021] Furthermore, in step 3, preliminary feature extraction is performed on the MI-EEG signal data through a group filter:
[0022] Step 301: construct five 4th-order Butterworth bandpass filters, whose passbands are 0.5-3 Hz, 4-7 Hz, 8-13 Hz, 14-30 Hz, and 30-70 Hz, respectively.
[0023] Step 302: Use the five bandpass filters mentioned above to perform bandpass filtering on the MI-EEG signal in sequence, and perform fast independent component analysis (FastICA) processing. FastICA uses the statistical independence assumption between the components of the observation signal x and the source signal s, which is similar to blind source separation.
[0024] Step 303: performing feature fusion on the features of the five different frequency bands processed by FastICA.
[0025] Furthermore, the multi-task learning method based on the category average standard deviation constructed in step 4 and the network model FBMEEGNet with the group filter feature extraction result as input includes:
[0026] Step 401: Construct a CNN feature processing block, and use a spatial filter with 125 convolution kernels of the same size as the number of EEG channels at the input end to compress the original multi-channel MI-EEG signal into one channel; then use a temporal convolution layer with 256 convolution kernels of size 10 and step size 5 to extract temporal domain features in the time domain; then use another convolution layer to extract and filter effective convolution kernels, and ignore convolution kernels with poor effects, thereby reducing overfitting; finally, use a max-pooling layer to extract the part with the largest and strongest response in the MI-EEG signal.
[0027] Step 402: Construct a residual module to solve the gradient vanishing or gradient exploding problem that may occur during the training of a deep neural network, so that the network can be deeper and easier to train and output recognition results through a fully connected layer and a Sigmoid activation function.
[0028] Step 403: Construct a multi-task learning method based on the category average standard deviation (CASD-MTL), which includes a recognition module and a feature extraction module. The recognition module has the network structure proposed in steps 401 to 402; and the feature extraction module is only used in the model training stage. In terms of network structure, it only lacks the fully connected layer of the output part and its activation function (Sigmoid) compared to the recognition module. The category average standard deviation loss function (CASDLoss) is used as its loss function, and the feature extraction module and the recognition module share network parameters. When training the model, the recognition module is trained first. When the training is close to completion, the feature extraction module intervenes and begins to adjust the neural network parameters to reduce the standard deviation within the class. The features of the same category will be more concentrated in a certain area, which will help reduce the overfitting phenomenon of the model. At the same time, the recognition module will continue to work to ensure the recognition accuracy, so that the model accuracy is maintained while reducing the difference within the class, thereby reducing the overfitting phenomenon.
[0029] A recognition system for a motor imagery EEG signal recognition method based on CLSTM-GAN-FBMEEGNet, characterized by comprising:
[0030] An acquisition module, used to acquire multi-channel motor imagery EEG MI-EEG signal data, wherein the MI-EEG signal data is derived from a non-invasive wearable device;
[0031] A preprocessing module, used for preprocessing the collected motor imagery MI-EEG signal data;
[0032] A data enhancement module, used to perform data enhancement on the preprocessed MI-EEG signal data through a generative adversarial network (CLSTM-GAN) model;
[0033] The group filter feature extraction module uses 5 groups of bandpass filters to perform bandpass filtering on the preprocessed MI-EEG signal data in sequence to obtain multi-band features;
[0034] The FBMEEGNet network classification module, which is constructed based on the multi-task learning method of category mean standard deviation and convolutional neural network, is used to further extract and recognize the MI-EEG signal features after feature extraction to obtain the recognition results of motor imagery.
[0035] (III) Beneficial effects
[0036] Compared with the prior art, the present invention provides a method and system for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet, which has the following beneficial effects:
[0037] (1) CLSTM-GAN data enhancement was performed on the MI-EEG signal to synthesize new samples to increase the diversity and quantity of the dataset, thereby improving the accuracy and robustness of the model. A new group filter scheme was used to extract features from the MI-EEG signal, which improved the effectiveness of feature extraction.
[0038] (2) A multi-task learning method based on the category average standard deviation of CLSTM-GAN-FBMEEGNet was designed, which can improve the effectiveness of the neural network for further feature extraction, thereby improving the recognition accuracy of the neural network. Based on this multi-task learning method, combined with the convolutional neural network, a network model FBMEEGNet with superior performance was designed, which has a high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A schematic flow chart of a method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet provided in an embodiment of the present invention;
[0040] Figure 2 A schematic diagram of the structure of a method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet provided in an embodiment of the present invention;
[0041] Figure 3 A schematic diagram of a process of extracting features using a group filter provided in an embodiment of the present invention;
[0042] Figure 4 A schematic diagram of the convolutional LSTM generative adversarial network (CLSTM-GAN) process provided by an embodiment of the present invention;
[0043] Figure 5A schematic diagram of the generator structure in the CLSTM-GAN provided in an embodiment of the present invention;
[0044] Figure 6 A schematic diagram of the discriminator structure in the CLSTM-GAN provided in an embodiment of the present invention;
[0045] Figure 7 A schematic diagram of the structure of the FBMEEGNet network provided in an embodiment of the present invention;
[0046] Figure 8 A schematic diagram of the structure of a convolution module provided in an embodiment of the present invention;
[0047] Fig. 9 A schematic diagram of the structure of a residual network provided by an embodiment of the present invention;
[0048] Fig.10 A schematic diagram of the structure of the identification system provided by the example of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] Example
[0051] like Figure 1-10 As shown, an embodiment of the present invention proposes a method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet, and the specific steps are as follows:
[0052] Step 1: Reference Figure 2 , obtain motor imagery EEG signals and perform preprocessing.
[0053] Step 101: Obtain motor imagery EEG signals, mainly collecting EEG signals in the parietal and central regions of the head. The MI-EEG signals in the data set are sampled at 250 Hz, and the EEG signals in the left-hand and right-hand motor imagery situations are collected.
[0054] Step 102: Use a multi-stage Butterworth filter to filter the MI-EEG signal data (high-pass and low-pass filtering) to obtain EEG data within the relevant range of the motor imagery stage (0.5-70 Hz), thereby effectively eliminating high-frequency noise interference;
[0055] Step 103: resampling the MI-EEG signal data to 160 Hz to reduce the data volume to increase the subsequent operation speed;
[0056] Step 104: Crop the MI-EEG signal data from 0.5s to 5s to further reduce the data volume and remove noise.
[0057] Step 2: Reference Figure 3 , a convolutional LSTM generative adversarial network (CLSTM-GAN) is constructed to perform data augmentation on MI-EEG. CLSTM-GAN consists of two main parts, called the generator and the discriminator. In the generator, arbitrary random data from a known distribution is provided as input to synthetically generate realistic data. Then, the discriminator determines whether the generated samples are real or fake.
[0058] Step 201: Reference Figure 4 , construct the generator (G(z)) network, including fully connected layers, convolutional layers, LSTM layers, activation layers (Tanh), and maximum pooling layers. The probability that the generator correctly classifies the actual data is log(D(x)), and increasing log(1-D(G(z))) will make it easier for the discriminator to identify the fake samples produced by the generator.
[0059] Step 202: Reference Figure 5 , build the discriminator (D(x)) network, including fully connected layers, convolutional layers, batch normalization layers and activation layers (Tanh, Sigmoid) and other processing.
[0060] Step 203: G and D are in a game-adversarial relationship. During the game, the pseudo samples generated by the generator gradually approach the real data, and the judgment ability of the discriminator gradually increases, eventually reaching a dynamic balance. This process can be expressed as follows:
[0061]
[0062] Among them, x is the real data, which obeys the real data distribution P data (x); z is random noise, following the bright distribution P z (z); G(z) is the pseudo sample generated by inputting random noise z into the generator; D(x) and D(G(z)) are the probabilities that the discriminator judges x and G(z) to be "true", respectively.
[0063] Step 3: Reference Figure 6 , a group filter is used to extract features from the MI-EEG signal.
[0064] Step 301: construct five 4th-order Butterworth bandpass filters, whose passbands are 0.5-3 Hz, 4-7 Hz, 8-13 Hz, 14-30 Hz, and 30-70 Hz, respectively.
[0065] Step 302: Use the five bandpass filters mentioned above to perform bandpass filtering on the MI-EEG signal in sequence, and perform fast independent component analysis (FastICA) processing. FastICA uses the statistical independence assumption between the components of the observation signal x and the source signal s, which is similar to blind source separation.
[0066] x(t)=[x 1 (t),x 2 (t),…,x m (t)] T is the m-dimensional observation signal, s(t)=[s 1 (t),s 2 (t),…,s n (t)] T To generate the observed signal, there are n independent source signals. The observed signal x(t) is a linear mixture of the source signal s(t) through an unknown matrix A, that is, x(t) = As(t). When both s(t) and A are unknown, ICA is used to solve a mixing matrix B so that the output y(t) after x(t) passes through B is the best estimate of s(t), that is, y(t) = Bs(t).
[0067] Step 303: The features of the five different frequency bands processed by FastICA are sequentially concatenated together in a new dimension to form a fused feature.
[0068] Step 4: Reference Figure 7 , construct the FBMEEGNet network, conduct further feature training and obtain the recognition results of motor imagery.
[0069] Step 401: Reference Figure 8 , construct a convolution module, including convolution layer, batch normalization layer, ELU activation layer (ELU), maximum pooling layer and Dropout layer.
[0070] Step 402: Reference Fig. 9 , constructing a residual network (Res-Net) can effectively solve the problem of low recognition rate and slow network convergence after the number of CNN network layers increases to a certain level, that is, the network degradation problem. The residual module can be used as part or multiple parts of CNN. Let the input of CNN be x, the function mapping to be fitted be H(x), and define the residual mapping F(x) = H(x)-x, then the original mapping function H(x) can be expressed as F(x)+x.
[0071] Step 403: A loss function is designed to improve the effectiveness of feature extraction of the neural network in the motor imagery recognition task, which is called the Class-Averaged Standard Deviation Loss Function (CASDLoss), and its formula is:
[0072]
[0073] Where K represents the number of output features of the residual network, k i,j represents the jth output value of the residual network output tensor after flattening in the i-th trial, μ l,j ,μ r,j denotes the average of the sets of features where the jth output value is the left-hand category or the right-hand category in all trials, respectively.
[0074] A multi-task learning method based on category average standard deviation (CASD-MTL) is constructed, which includes a recognition module and a feature extraction module. The recognition module has the network structure proposed in steps 401 to 402; while the feature extraction module is only used in the model training stage. In terms of network structure, it only lacks the fully connected layer of the output part and its activation function (Sigmoid) compared to the recognition module. The category average standard deviation loss function (CASDLoss) is used as its loss function, and the feature extraction module and the recognition module share network parameters. When training the model, the recognition module is trained first. When the training is close to completion, the feature extraction module intervenes and begins to adjust the neural network parameters to reduce the standard deviation within the class. The features of the same category will be more concentrated in a certain area, which will help reduce the overfitting phenomenon of the model. At the same time, the recognition module will continue to work to ensure the recognition accuracy, so that the model accuracy is maintained while reducing the intra-class differences, thereby reducing the overfitting phenomenon.
[0075] A motor imagery EEG signal recognition system based on CLSTM-GAN-FBMEEGNet, reference Fig.10 , characterized in that it includes:
[0076] The acquisition module is used to obtain multi-channel motor imagery EEG MI-EEG signal data, and uses a non-invasive wearable EEG acquisition device to perform step 101 of the embodiment, mainly collecting EEG signals in the parietal lobe and central area of the head, sampling at 250Hz, and collecting EEG signals in two cases of left-hand and right-hand motor imagery.
[0077] The preprocessing module is used to preprocess the collected motor imagery EEG MI-EEG signal data and perform steps 102, 103 and 104 of the embodiment on the collected MI-EEG signal.
[0078] The data enhancement module is used to perform data enhancement of step 2 of the embodiment on the preprocessed MI-EEG signal data through the CLSTM-GAN model.
[0079] The group filter feature extraction module is used to perform multiple bandpass filtering on the MI-EEG signal data after data enhancement, and pass Fast-ICA. That is, the MI-EEG signal after data enhancement is subjected to step 3 of the embodiment.
[0080] FBMEEGNet feature extraction and recognition module, build FBMEEGNet to further extract and recognize the MI-EEG signal features after feature extraction, and perform step 4 of the embodiment on the MI-EEG signal features.
[0081] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and 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 can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet, characterized in that: include: Step 1: Acquire motor imagery EEG MI-EEG signal data, and pre-process the MIEEG signal data from multiple channels, wherein the MI-EEG signal data comes from a non-invasive wearable EEG acquisition device; Step: 2: Construct a convolutional LSTM generative adversarial network (CLSTM-GAN) to perform data enhancement on the MI-EEG signal data; Step 3: Perform preliminary feature extraction on the data-enhanced MI-EEG signal through group filters; Step 4: Construct a multi-task learning method based on the category average standard deviation, and use the group filter feature extraction result as input to the network model FBMEEGNet. Input the features initially extracted by the group filter into the FBMEEGNet model to obtain the recognition result of motor imagery.
2. The method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet according to claim 1, characterized in that: The step 1 of preprocessing the MI-EEG signal data from multiple channels includes: Step 101: Use a multi-stage Butterworth filter to filter the MI-EEG signal data (high-pass and low-pass filtering) to obtain EEG data within the relevant range of the motor imagery stage (0.5-70 Hz), thereby effectively eliminating high-frequency noise interference; Step 102: resampling the MI-EEG signal data to 160 Hz to reduce the data volume to increase the subsequent operation speed; Step 103: Crop the MI-EEG signal data from 0.5s to 5s to further reduce the data volume and remove noise.
3. The method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet according to claim 1, characterized in that: The step 2 of performing data enhancement on the pre-processed MI-EEG signal data includes: Step 201: Construct a generator (G(z)) network, including fully connected layers, convolutional layers, LSTM layers, activation layers (Tanh), and maximum pooling layers. The probability that the generator correctly classifies the actual data is log(D(x)), and increasing log(1-D(G(z))) will make it easier for the discriminator to identify the fake samples generated by the generator. Step 202: Construct a discriminator (D(x)) network, including fully connected layers, convolutional layers, batch normalization layers, and activation layers (Tanh, Sigmoid) and other processing. Step 203: G and D are in a game-adversarial relationship. During the game, the pseudo samples generated by the generator gradually approach the real data, and the judgment ability of the discriminator gradually increases, eventually reaching a dynamic balance.
4. The method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet according to claim 1, characterized in that: In step 3, preliminary feature extraction is performed on the MI-EEG signal data through a group filter: Step 301: construct five 4th-order Butterworth bandpass filters, whose passbands are 0.5-3 Hz, 4-7 Hz, 8-13 Hz, 14-30 Hz, and 30-70 Hz, respectively. Step 302: Use the five bandpass filters mentioned above to perform bandpass filtering on the MI-EEG signal in sequence, and perform fast independent component analysis (FastICA) processing. FastICA uses the statistical independence assumption between the components of the observation signal x and the source signal s, which is similar to blind source separation. Step 303: performing feature fusion on the features of the five different frequency bands processed by FastICA.
5. The method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet according to claim 1, characterized in that: The multi-task learning method based on the category average standard deviation constructed in step 4 and the network model FBMEEGNet with the group filter feature extraction result as input include: Step 401: Construct a CNN feature processing block, and use a spatial filter with 125 convolution kernels of the same size as the number of EEG channels at the input end to compress the original multi-channel MI-EEG signal into one channel; then use a temporal convolution layer with 256 convolution kernels of size 10 and step size 5 to extract temporal domain features in the time domain; then use another convolution layer to extract and filter effective convolution kernels, and ignore convolution kernels with poor effects, thereby reducing overfitting; finally, use a max-pooling layer to extract the part with the largest and strongest response in the MI-EEG signal. Step 402: Construct a residual module to solve the gradient vanishing or gradient exploding problem that may occur during the training of a deep neural network, so that the network can be deeper and easier to train and output recognition results through a fully connected layer and a Sigmoid activation function. Step 403: Construct a multi-task learning (CASD-MTL) method based on the category average standard deviation, which includes a recognition module and a feature extraction module. The recognition module has the network structure proposed in steps 401 to 402; and the feature extraction module is only used in the model training stage. In terms of network structure, it only lacks the fully connected layer and its activation function of the output part compared to the recognition module. The category average standard deviation loss function (CASDLoss) proposed in this patent is used as its loss function, and the feature extraction module shares network parameters with the recognition module. When training the model, the recognition module is trained first. When the training is close to completion, the feature extraction module intervenes and begins to adjust the neural network parameters to reduce the standard deviation within the class. The features of the same category will be more concentrated in a certain area, which will help reduce the overfitting phenomenon of the model. At the same time, the recognition module will continue to work to ensure the recognition accuracy, so that the model accuracy is maintained while reducing the intra-class differences, thereby reducing the overfitting phenomenon.
6. The recognition system of the method for recognizing motor imagery EEG signals based on CLSTM-GAN-FBMEEGNet according to claims 1-5, characterized in that: include: An acquisition module, used to acquire multi-channel motor imagery EEG MI-EEG signal data, wherein the MI-EEG signal data is derived from a non-invasive wearable device; A preprocessing module, used for preprocessing the collected motor imagery MI-EEG signal data; A data enhancement module, used to perform data enhancement on the preprocessed MI-EEG signal data through a generative adversarial network (CLSTM-GAN) model; The group filter feature extraction module uses 5 groups of bandpass filters to perform bandpass filtering on the preprocessed MI-EEG signal data in sequence to obtain multi-band features; Based on the FBMEEGNet network classification module, it is used to further extract and classify the MI-EEG signal features after feature extraction to obtain the recognition results of motor imagery.