Motion Imagery EEG Signal Recognition Method Based on EMD Data Augmentation and Parallel SCN

Abstract

The present invention claims protection for a method for identifying motor imagery electroencephalogram (EEG) signals based on EMD data augmentation and a parallel spatio-temporal convolutional network (SCN). The method includes the steps of: first, preprocessing the original EEG signals, decomposing the preprocessed EEG signals by EMD to obtain intrinsic modes, segmenting the intrinsic modes according to the time dimension, and combining them with non-repeating intrinsic mode segments of another sample with the same label, thereby generating more artificial data that conforms to the characteristics of the original EEG signals to solve the problem of the small amount of EEG signal samples. Then, a parallel spatio-temporal convolutional network is designed. The first layer performs convolution in the time domain, and the second layer performs convolution in the channel domain, which can fully extract the spatio-temporal characteristics of the EEG signals. Considering that the rhythms of motor imagery are mainly distributed in the μ and β rhythms, the μ and β frequency bands of the EEG signals are respectively used as the inputs of the parallel spatio-temporal convolutional network to extract the characteristics of the EEG signals in the time, space, and frequency domains and classify them. The present invention can effectively overcome the problem of low recognition accuracy caused by the small amount of EEG signal data.

Description

Technical Field

[0001] The present invention belongs to the field of motor imagery electroencephalogram (EEG) signal recognition, and particularly relates to a method for recognizing motor imagery EEG signals based on EMD data augmentation and parallel spatio-temporal convolutional network (SCN). Background Art

[0002] Brain-computer interface (BCI) is a human-computer interaction method that directly communicates between the human brain and a computer or external device, and has outstanding application value in the fields of rehabilitation engineering and control. Controlling external devices, such as robotic arms and wheelchairs, through a BCI system can help disabled people recover a certain degree of motor ability; in addition, typing can also be performed through the BCI system, providing a way for patients with speech impairments to communicate with the outside world. However, motor imagery EEG signals have characteristics such as low signal-to-noise ratio, non-linearity, and non-stationarity. Therefore, it is a challenging task to obtain effective features from EEG signals.

[0003] Currently, for feature extraction and recognition in MI-BCI systems, it is mainly divided into traditional methods and deep learning methods. Traditional methods perform feature extraction and recognition separately, while deep learning methods perform feature extraction and recognition together. Traditional feature extraction methods include wavelet transform (WT), power spectral density (PSD), autoregressive model, common spatial patterns (CSP), principal component analysis (PCA), and empirical mode decomposition (EMD), etc. Traditional feature classification methods include k-nearest neighbor (KNN), linear discriminant analysis (LDA), Bayesian classifier, support vector machine (SVM), etc. In traditional methods, feature extraction mainly relies on manually designed extractors, which require a complex parameter tuning process. At the same time, each method is targeted at specific applications, with poor generalization ability and robustness. Moreover, traditional feature classification methods perform poorly when features are complex.

[0004] Compared with traditional feature extraction and recognition methods, deep learning methods have greater advantages. Deep learning models can adaptively process non-linear data, directly extract features from raw data, and can perform multi-classification on EEG signals. However, the method based on convolutional neural network still faces two problems. First, most current networks focus more on extracting spatio-temporal features of motor imagery EEG signals while ignoring frequency domain features, which limits the full extraction of features of motor imagery EEG signals and leads to poor recognition results. In addition, a large amount of training data is required during the training of convolutional neural networks to fully learn the feature distribution of motor imagery EEG signals in order to achieve a high classification accuracy. However, the process of collecting motor imagery EEG signals is complex and the experimental conditions are demanding, making it difficult to construct a high-quality and large-scale motor imagery EEG dataset.

[0005] To address the above problems, the present invention proposes a method for recognizing motor imagery EEG signals based on EMD data augmentation and parallel spatio-temporal convolutional network. To a certain extent, it alleviates the problem of the small sample size of motor imagery EEG signals and extracts features of EEG signals from multiple aspects to enhance the classification performance of the model.

[0006] The inventor's prior application CN111012336A, a method for recognizing motor imagery electroencephalogram based on parallel convolutional neural network with spatio-temporal feature fusion. Taking motor imagery EEG signals as the research object, a new deep network model - parallel convolutional neural network is proposed to extract spatio-temporal features of motor imagery EEG signals. Different from traditional EEG classification algorithms that often discard EEG spatial feature information, through fast Fourier transform, Theta wave (4 - 8Hz), alpha wave (8 - 12Hz) and beta wave (12 - 36Hz) are extracted to generate 2D EEG feature maps. Based on multiple convolutional neural networks, the EEG feature maps are trained to extract spatial features. In addition, a temporal convolutional neural network is used for parallel training to extract temporal features. Finally, based on Softmax, the spatial features and temporal features are fused and classified. Experimental results show that the parallel convolutional neural network has good recognition accuracy and is superior to other latest classification algorithms.

[0007] In the comparative document (CN111012336A), a new deep network model - parallel convolutional neural network is proposed to extract the spatio-temporal features of motor imagery electroencephalogram (EEG) signals. This method uses the fast Fourier transform to convert the original EEG data into an image of 28*28. After processing, there will be a certain loss of the temporal information of the EEG signals. This invention uses a multiple convolutional neural network with four convolutional layers in parallel with a one-dimensional time convolution. The model is relatively deep and has a relatively high risk of overfitting. The feature fusion of this invention uses a fully connected layer and then classifies through Softmax, which may lead to insufficient feature extraction. The present invention uses the preprocessed original EEG signals as input, maximizing the retention of the original features of the EEG signals. It uses a two-layer parallel convolutional network. The two parallel network structures are the same, and they respectively extract the spatio-temporal features of the μ frequency band and β frequency band that are active in motor imagery, reducing the complexity of the model and the risk of overfitting. For the problem of insufficient feature extraction, after feature splicing in the time dimension, a layer of convolution is used to further extract the fused features, enabling the model to fully extract the features of the EEG signals. In addition, the present invention performs data augmentation on the EEG data to improve the generalization and robustness of the model. Summary of the Invention

[0008] The present invention aims to solve the above problems of the prior art. A method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN is proposed. The technical solution of the present invention is as follows:

[0009] A method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN, comprising the following steps:

[0010] Step 1: Collect motor imagery EEG signals and label them according to different motor imagery paradigms. Use a 32-channel EEG cap, which includes the channels of interest (C3, C4, Cz). Labeling means attaching a label to this sample.

[0011] Step 2: Select the EEG signals of the channels of interest for preprocessing operations such as filtering and normalization, and then divide them into a training set and a test set according to a ratio of 4:1.

[0012] Step 3: Decompose the EEG signals of the training set by EMD to obtain the intrinsic modes of each sample channel. Segment the intrinsic modes according to the time dimension and combine them with the non-repeated intrinsic mode segments of another sample with the same label to form new EEG data. Then merge the artificially generated data and the preprocessed training set data into a new training set.

[0013] Step 4: Construct a parallel spatio-temporal convolutional network. The input of the convolutional network is improved. Two-layer convolutional modules are used to extract spatio-temporal features in the μ and β frequency bands respectively for extracting features in different frequency bands of interest. For the two-layer convolutional module, the first layer extracts features in the time dimension, and the second layer extracts features in the channel dimension.

[0014] Step 5: Use the EEG signal training set after data augmentation to train the parallel spatio-temporal convolutional network.

[0015] Step 6: Use the EEG signal test set data to evaluate the performance of the parallel spatio-temporal convolutional network, and select a model with good performance for online classification of EEG signals.

[0016] Further, step 1 specifically includes: using the EEG signal acquisition device BP produced by German Brain Products company to collect left and right motor imagery EEG signals, with a sampling frequency of 250 Hz. A motor imagery paradigm is 10 s, 0 - 2 s is the preparation time, 2 - 6 s is the motor imagery time, and 6 - 10 s is the rest time.

[0017] Further, the preprocessing operations of filtering and standardizing the EEG signals of the selected channels of interest in step 2 specifically include:

[0018] Channel selection: The EEG signal acquisition device BP uses a 32-channel electrode cap. Channels C3, C4, and Cz in the motor area are retained, and the motor imagery EEG data between 2 - 6 s is intercepted according to the acquisition paradigm, and the EEG signals are constructed into a 1000×3 vector.

[0019] Filtering and standardizing: Use a 6th-order Butterworth band-pass filter to filter the EEG signals, with a filtering range of 0.5 Hz - 40 Hz, and then standardize the EEG signals.

[0020] Further, step 3 specifically includes the following steps:

[0021] First, use EMD to decompose the EEG signals. For a single channel, randomly select two data with the same label. After EMD decomposition of the first channel, IMF1 and IMF2 are obtained respectively. IMF1 and IMF2 are equally divided into four segments along the time axis. Select the first and third segments of IMF1, and the second and fourth segments of IMF2 to form new EEG data for the first channel. Similarly, new EEG data for the second and third channels can be generated in the above way, and then the three channels are combined to form a new motor imagery EEG signal data.

[0022] Further, the EMD decomposition of the first channel respectively obtains IMF1 and IMF2, which specifically includes: 1. Drawing the upper and lower envelope lines according to the maximum and minimum extreme points of the original signal; 2. Calculating the mean of the upper and lower envelope lines and drawing the mean envelope line; 3. Subtracting the mean envelope line from the original signal to obtain an intermediate signal; 4. Judging whether the intermediate signal meets the two conditions of the intrinsic mode. If it meets the conditions, the signal is an IMF component; if it does not meet the conditions, based on this signal, repeat steps 1-4; 5. After obtaining the first IMF, subtracting the IMF from the original signal as the new original signal, and then obtaining a new IMF through steps 1-4, and so on to complete the EMD decomposition.

[0023] Further, in step 4, a parallel spatio-temporal convolutional network is constructed, and the input of the convolutional neural network is improved. The EEG signals in the μ band and β band are input into the spatio-temporal convolutional network in parallel so that the model can extract the high-level features in the frequency domain, time domain, and spatial domain of the EEG signals. The network model is specifically as follows:

[0024] When the μ band of 5-13 Hz is adopted and the data shape of the input layer is (20, 1, 1000, 3), the first dimension is the batch of the input data, the second dimension is the newly added channel dimension to meet the two-dimensional convolution, the third dimension is the time dimension of the EEG signal, and the fourth dimension is the number of electrodes of the EEG signal; the time convolution of the first layer uses a convolution kernel size of 60×1 and a stride of 3×1, the channel convolution of the second layer uses a convolution kernel size of 1×3, and the stride is defaulted to 1×1, and then max pooling with a kernel size of 6×1 is performed.

[0025] Further, when the β band of 13-30 Hz is adopted as the input, after passing through the above first-layer convolution and second-layer convolution in the same way, and then through max pooling, the feature vectors extracted by the model are obtained; finally, the two feature vectors are concatenated in the time dimension; the concatenated vector is input into a convolution with a convolution kernel size of 3×1 and a stride of 1×1 to extract the deep spatio-temporal features of the EEG signal; two fully connected layers are adopted later, a Dropout is added between the two layers, and finally two features are output, and then these two features are classified through Softmax.

[0026] Further, in step 5, the parallel spatio-temporal convolutional network is trained using the data-augmented dataset, specifically by setting the data to 20 samples per batch, and training all the data 100 times in total. The optimizer uses the stochastic gradient descent algorithm, and the loss function adopts the cross-entropy loss function.

[0027] Further, the stochastic gradient descent algorithm specifically includes:

[0028] 1. Sample the training set data and make predictions through parallel SCN; 2. Use the cross-entropy loss function to calculate the loss between the predicted value and the true value; 3. Adjust the weights of the network model to minimize the loss.

[0029] The loss function uses the cross-entropy loss function, which specifically includes:

[0030] 1. Normalize the output of the model to calculate the probability belonging to each category; 2. Take the logarithm of the probability corresponding to the label value and then take the negative value to obtain the loss of a sample; 3. Average the loss values of a batch of samples to obtain the average loss of a batch, and then use the stochastic gradient descent algorithm to optimize the loss.

[0031] The advantages and beneficial effects of the present invention are as follows:

[0032] Aiming at the problem of the small amount of electroencephalogram (EEG) signal data samples, in step three, the EMD decomposition method is used to perform data augmentation on the EEG data. The EEG signal is decomposed by EMD to obtain the intrinsic mode functions, and then segmented and recombined on the intrinsic mode functions to generate more artificial data that conform to the characteristics of the original EEG signal, increasing the data samples of the training set and making the trained network model more robust. In addition, in order to fully extract the high-level features in the spatio-temporal frequency domain of the EEG signal, in step four, a parallel spatio-temporal convolutional network is designed. The first layer performs convolution in the time domain, and the second layer performs convolution in the channel domain. Particularly, the network uses the EEG signals of μ and β rhythms as the inputs of the model respectively, and extracts the features of μ and β rhythms of the EEG signal in parallel, reducing the depth of the network and the risk of overfitting; after the feature concatenation, instead of directly passing through the fully connected layer for classification, a layer of convolution is used after the feature concatenation to further extract the fine-grained features of the EEG signal, and then classification is performed through the fully connected layer and Softmax. The present invention can solve the problem of the small amount of EEG signal data and effectively improve the accuracy of motor imagery recognition. Brief Description of the Drawings

[0033] Figure 1 It is a schematic flowchart of the method for classifying motor imagery EEG signals based on EMD data augmentation and parallel spatio-temporal convolutional network provided by the preferred embodiment of the present invention.

[0034] Figure 2 It is a schematic diagram of the implementation steps of EMD data augmentation in an embodiment of the present invention.

[0035] Figure 3 It is a schematic diagram of the network structure of the parallel spatio-temporal convolutional network in an embodiment of the present invention. Detailed Embodiments

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0037] The technical solution of the present invention to solve the above technical problems is:

[0038] A method for recognizing motor imagery electroencephalogram signals based on EMD data augmentation and parallel spatio-temporal convolutional network provided in this embodiment includes the following steps:

[0039] Step 1: Use an electroencephalogram device produced by Brain Products in Germany to collect data. The sampling frequency is 250 Hz. A motor imagery paradigm is 10 s, with a preparation time of 0 - 2 s, a motor imagery time of 2 - 6 s, and a rest time of 6 - 10 s. A total of 300 groups of data are collected, with 150 groups for left and right motor imagery respectively.

[0040] Step 2: First, select the electrodes C3, C4, and Cz in the motor imagery area, and retain the motor imagery electroencephalogram data between 2 - 6 s according to the paradigm of collecting electroencephalogram signals, and construct it into a 1000×3 vector. Then use a 6th-order Butterworth band-pass filter to filter the electroencephalogram signals, with a filtering range of 0.5 Hz - 40 Hz. Finally, standardize the electroencephalogram signals so that the mean of the electroencephalogram signal data is 0 and the variance is 1. The formula for its standardization is as follows:

[0041]

[0042] where z i is the standardized data at time point i, x i is the original data at time point i, u is the average value of single-channel data, and δ is the standard deviation of single-channel data. Finally, divide the data set into a training set and a test set according to 4:1.

[0043] Step 3: Use EMD to decompose the electroencephalogram signals in the training set to obtain the intrinsic modes of each channel. The decomposition result is as follows:

[0044]

[0045] In the formula, it means decomposing the original signal x(t) into N intrinsic modes and a residual component res, removing the residual component and the intrinsic modes greater than 6. If there are less than 6 intrinsic modes, use 0 to fill. The specific steps of EMD data augmentation are as Figure 2As shown in the figure, two samples A and B of the same category are randomly selected. First, the first channel is selected, denoted as A1 and B1. EMD decomposition is performed on each of them. After screening, the intrinsic mode is decomposed into four segments. The first segment (A11) and the third segment (A13) of the intrinsic mode of A1 are taken, and the second segment (B12) and the fourth segment (B14) of the intrinsic mode of B1 are taken. Then, the above four segments are combined into a new intrinsic mode and reconstructed into new motor imagery EEG signal data. The number of artificial data can be generated according to the actual situation.

[0046] Step 4: Construct a parallel spatio-temporal convolutional network, whose structure diagram is as Figure 3 shown. The EEG signals in the μ band and the β band are input into the spatio-temporal convolutional network in parallel. Then, features are extracted through the spatio-temporal convolutional block. Next, the features are concatenated in the time dimension. After that, a deeper spatio-temporal feature is extracted through a convolutional layer. Finally, classification is performed through a fully connected layer and Softmax.

[0047] Taking the input of μ (5 - 13 Hz) as an example, the data shape of the input layer is (20, 1, 1000, 3). The first dimension is the batch of input data, the second dimension is the newly added channel dimension to meet the requirements of two-dimensional convolution, the third dimension is the time dimension of the EEG signal, and the fourth dimension is the number of electrodes of the EEG signal. The time convolution of the first layer sets the input channel to 1, the output channel to 8, the convolution kernel size to 60×1, and the stride to 3×1. Immediately after the first convolutional layer, a BatchNorm layer and a Dropout layer are added to prevent overfitting, and the Relu activation function is used to improve the expression ability of the model. The channel convolution of the second layer sets the input channel to 8 according to the output of the previous layer, the output channel to 16, the convolution kernel size to 1×3, and the stride is defaulted to 1×1. After the second convolution, there are a BatchNorm layer, a Dropout layer, and the Relu activation function in sequence. Then, max pooling with a kernel size of 6×1 is performed, and finally, the feature vector of the μ band is obtained.

[0048] In addition, the β (13 - 30 Hz) band as input also goes through the above first convolution and second convolution, and then max pooling, to obtain the feature vector extracted from the β band. The two feature extraction vectors are concatenated and input into a convolution with a convolution kernel size of 3×1 and a stride of 1×1 to extract the deeper spatio-temporal features of the EEG signal. Then, there are two fully connected layers, and a Dropout is added between the two layers. First, the four-dimensional feature vector is flattened to obtain a two-dimensional feature vector, 100 features are obtained through the first fully connected layer, 2 features are obtained through the second fully connected layer, and then the 2 features are classified through Softmax.

[0049] Step 5: Train the parallel spatio-temporal convolutional network using the data-augmented dataset. Specifically, set the data to 20 samples per batch and train all the data for a total of 100 times. Use the stochastic gradient descent algorithm as the optimizer and the cross-entropy loss function as the loss function. Finally, save the trained model parameters for subsequent evaluation of the network model's performance on the test set.

[0050] The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0051] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, method, commodity or device comprising the element.

[0052] The above embodiments should be understood as being only for illustrative purposes of the present invention and not for limiting the protection scope of the present invention. After reading the content recorded in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for identifying motor imagery electroencephalogram (EEG) signals based on EMD data augmentation and SCN, characterized in that, it includes the following steps: Step 1: Collect motor imagery EEG signals and label them according to different motor imagery paradigms. Use a 32-channel EEG cap, which includes channels of interest (C3, C4, Cz); labeling means assigning a label to this sample. Step 2: Select the EEG signals of the channels of interest for preprocessing operations of filtering and normalization, and then divide them into a training set and a test set according to a ratio of 4:

1. Step 3: Decompose the EEG signals of the training set by EMD to obtain the intrinsic modes of each sample channel. Segment the intrinsic modes according to the time dimension and combine them with non-repeating intrinsic mode segments of another sample with the same label to form new EEG data, and merge the artificially generated data and the preprocessed training set data into a new training set. Step 4: Construct a parallel spatio-temporal convolutional network, improve the input of the convolutional network, and use two-layer convolutional modules to extract spatio-temporal features in the μ and β frequency bands respectively for extracting features in different frequency bands of interest. Among them, for the two-layer convolutional module, the first layer extracts features in the time dimension, and the second layer extracts features in the channel dimension. Step 5: Train the parallel spatio-temporal convolutional network using the EEG signal training set after data augmentation. Step 6: Use the EEG signal test set data to evaluate the performance of the parallel spatio-temporal convolutional network, and select a model with good performance for online classification of EEG signals.

2. The method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 1, characterized in that, Step 1 specifically includes: Using the EEG signal acquisition device BP produced by German Brain Products company to collect left and right motor imagery EEG signals, with a sampling frequency of 250 Hz. One motor imagery paradigm is 10 s, 0 - 2 s is the preparation time, 2 - 6 s is the motor imagery time, and 6 - 10 s is the rest time.

3. The method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 1, characterized in that, The preprocessing operations of filtering and normalization for the EEG signals of the channels of interest selected in Step 2 specifically include: Channel selection: The EEG signal acquisition device BP uses a 32-channel electrode cap, retains the channels C3, C4, and Cz in the motor area, and extracts the motor imagery EEG data between 2 - 6 s according to the acquisition paradigm, and constructs the EEG signals into a 1000×3 vector. Filtering and normalization: Use a 6th-order Butterworth band-pass filter to filter the EEG signals, with a filtering range of 0.5 Hz - 40 Hz, and then normalize the EEG signals.

4. The method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 1, characterized in that, Step 3 specifically includes the following steps: First, use EMD to decompose the EEG signals. For a single channel, randomly select two data with the same label. After EMD decomposition of the first channel, IMF1 and IMF2 are obtained respectively. Divide IMF1 and IMF2 into four equal segments along the time axis. Select the first and third segments of IMF1, and the second and fourth segments of IMF2 to form the new EEG data of the first channel. Similarly, the new EEG data of the second and third channels can be generated in the above way, and then combine the three channels to form a new motor imagery EEG signal data.

5. A method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 4, characterized in that the obtaining of IMF1 and IMF2 respectively after EMD decomposition of the first channel specifically includes:

1. Draw the upper and lower envelope lines according to the maximum and minimum extreme points of the original signal; 2. Calculate the mean of the upper and lower envelope lines and draw the mean envelope line; 3. Subtract the mean envelope line from the original signal to obtain the intermediate signal; 4. Judge whether the intermediate signal meets the two conditions of the intrinsic mode. If it meets the conditions, the signal is an IMF component; if it does not meet the conditions, based on this signal, repeat steps 1-4; 5. After obtaining the first IMF, use the original signal minus the IMF as the new original signal, and then obtain the new IMF through steps 1-4, and so on to complete the EMD decomposition.

6. A method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 4, characterized in that in step 4, a parallel spatio-temporal convolutional network is constructed, and the input of the convolutional neural network is improved. The EEG signals in the μ band and β band are input into the spatio-temporal convolutional network in parallel so that the model can extract the high-level features in the frequency domain, time domain and spatial domain of the EEG signals. The network model is specifically: When using the μ band of 5-13 Hz and the shape of the input layer data is (20, 1, 1000, 3), the first dimension is the batch of input data, the second dimension is the newly added channel dimension to meet the two-dimensional convolution, the third dimension is the time dimension of the EEG signal, and the fourth dimension is the number of electrodes of the EEG signal; the time convolution of the first layer uses a convolution kernel size of 60×1 and a stride of 3×1, the channel convolution of the second layer uses a convolution kernel size of 1×3, and the stride defaults to 1×1, and then max pooling with a kernel size of 6×1 is performed.

7. A method for identifying motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 6, characterized in that when using the β band of 13-30 Hz as the input, after passing through the above first layer of convolution and second layer of convolution, and then through max pooling, the feature vectors extracted by the model are obtained; finally, the two feature vectors are concatenated in the time dimension; The spliced vectors are input into a convolution with a kernel size of 3×1 and a stride of 1×1 to extract the deep spatio-temporal features of the EEG signals; followed by two fully-connected layers, with a Dropout added between the two layers. Finally, two features are output, and then these two features are classified through Softmax.

8. A method for recognizing motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 6, characterized in that in step 5, the parallel spatio-temporal convolution network is trained using the data-augmented dataset, specifically, the data is set to 20 samples per batch, and all data is trained 100 times in total. The optimizer uses the stochastic gradient descent algorithm, and the loss function uses the cross-entropy loss function.

9. A method for recognizing motor imagery EEG signals based on EMD data augmentation and parallel SCN according to claim 8, characterized in that the stochastic gradient descent algorithm specifically includes:

1. Sampling the training set data and making predictions through the parallel SCN; 2. Calculating the loss between the predicted value and the true value using the cross-entropy loss function; 3. Adjusting the weights of the network model to minimize the loss; The loss function uses the cross-entropy loss function, specifically including:

1. Normalizing the output of the model to calculate the probability belonging to each category; 2. Taking the logarithm of the probability corresponding to the label value and then taking the negative value to obtain the loss of a sample; 3. Averaging the loss values of a batch of samples to obtain the average loss of a batch, and then using the stochastic gradient descent algorithm to optimize the loss.

Images