Driving electroencephalogram signal recognition method based on multi-domain asymptotic convolutional neural network

Abstract

The invention discloses a driving electroencephalogram signal recognition method based on a multi-domain asymptotic convolutional neural network, and relates to the field of electroencephalogram intelligent recognition. The method mainly comprises the steps of electroencephalogram (EEG) acquisition, frequency domain feature and spatial feature extraction, and variable asymptotic convolutional neural network modeling, recognition and classification. After the collected EEG signals are preprocessed, frequency domain features are obtained through fast Fourier transform (FFT) processing, and spatial features are obtained through common spatial pattern (CSP) processing. All frequency band features are stacked, an input matrix is constructed and input to the constructed variable asymptotic convolutional neural network model, strategies used in different stages are different, and recognition classification about five driving behaviors is obtained. According to the method, the frequency domain-space features and the variable asymptotic convolutional neural network model are combined, and the global features of the EEG signals are more concerned, so that the model classification capability is remarkably improved, and the EEG signal classification and recognition effect is further improved. Meanwhile, the characteristic of light weight of the model meets the requirement of intelligent driving on real-time performance.

Description

Technical Field

[0001] The present application belongs to the field of intelligent EEG recognition, and specifically relates to a driving EEG signal recognition method based on a multi-domain asymptotic convolutional neural network. Background Art

[0002] Intelligent driving, especially unmanned driving, has gradually become a reality. In order to further improve road safety and reduce the incidence of traffic accidents, driver status and behavior detection technology based on electroencephalogram (EEG) signals is becoming a research hotspot in the field of cutting-edge intelligent driving.

[0003] In the cross-research field of EEG recognition and intelligent driving, there are currently few studies on driving EEG signal recognition, and there are many key issues that need to be addressed. Among them, the low accuracy of driving behavior recognition has become a core bottleneck restricting the development of this field. As a type of feedforward neural network with a deep structure and convolution calculation, convolutional neural network (CNN) has shown excellent performance in many fields such as computer vision and speech recognition, but it has limitations in driving EEG signal recognition. In the process of continuous iterative training, the convolution kernel size of CNN is usually smaller than the input image size, which makes it focus more on local information when extracting features, and its global perception ability of the signal is obviously insufficient, making it difficult to capture the complex patterns and long-range dependencies in driving EEG signals, which in turn makes it difficult to improve the recognition rate of driving behavior.

[0004] In addition, traditional convolutional neural network models used to process EEG signals, such as EEGNet, require a huge amount of labeled data, and the accuracy of recognition can only be guaranteed with the support of a large amount of labeled data. However, EEG data labeling faces high manpower and time costs, and is limited by experimental conditions and the difficulty of recruiting subjects. The sample size is usually extremely limited. Therefore, the sample size of traditional CNN models is usually small, generally with only dozens of subjects, each of whom can only provide hundreds of trials. In the case of limited data, relying solely on the CNN model is prone to overfitting, and the generalization ability in the test set and actual application scenarios drops sharply, seriously affecting the practical value and actual application effect of the model.

[0005] In intelligent driving scenarios, real-time requirements have reached extremely high standards, which makes the weight of time cost in the driving decision-making process far higher than other factors. The processing speed and accuracy of EEG signals directly determine whether the intelligent driving system can respond to complex road conditions in a timely manner, thereby achieving safe and efficient driving control. Traditional EEG signal recognition models generally have problems such as excessive demand for computing resources and low computing efficiency, and cannot meet the strict real-time requirements of intelligent driving. Therefore, a lightweight model is needed to solve the real-time problem.

[0006] Meanwhile, EEG signals belong to complex non-stationary signals. Especially during driving, they are extremely vulnerable to various factors such as the driver's physical condition, fatigue level, and concentration. These factors result in a large amount of noise and irregular fluctuations in EEG signals, which requires more emphasis on the frequency domain and spatial characteristics of EEG signals during the recognition process to extract effective information more accurately and improve the accuracy and reliability of driving EEG signal recognition.

[0007] Therefore, a method for recognizing driving EEG signals based on a multi-domain asymptotic convolutional neural network is needed. Summary of the Invention

[0008] To solve the above problems and improve the recognition rate of driving behavior, an embodiment of this application provides a method for recognizing driving EEG signals based on a multi-domain asymptotic convolutional neural network. The multi-domain includes frequency domain features and spatial features, and the asymptotic includes a variable convolutional layer and an asymptotic size convolutional kernel.

[0009] To achieve the above object, the technical solution of the present invention is as follows: A method for recognizing driving EEG signals based on a multi-domain asymptotic convolutional neural network, the method comprising the following steps:

[0010] Step 1: Extract the EEG signals of the driver during driving through a brain-computer interface to obtain the original EEG data;

[0011] Step 2: Remove power frequency interference, baseline drift, etc. from the EEG signals;

[0012] Step 3: Divide the preprocessed EEG signals into multiple time windows:

[0013] C channels, sampling rate f s , total duration T seconds, divide the signal into windows, each time length L = 2 seconds, and the number of sampling points contained in each segment S = L × f s .

[0014] The data matrix after segmentation is X n ∈R C×S , n = 1, 2, 3…, N;

[0015] Step 4: Convert the signal of each time window into the frequency domain through the fast Fourier transform (FFT):

[0016] F c = FFT(x c ) ∈ C s

[0017] Calculate the power spectral density:

[0018] P c = |F c |2 ∈R S

[0019] The frequency resolution is Δf = f s / S, and the frequency points correspond to k1 is the index frequency;

[0020] Step 5 Integrate the energy of the specific frequency band for each channel to obtain the frequency-domain feature matrix:

[0021] The corresponding frequency index k start , k end , Band energy calculation:

[0022]

[0023] The frequency-domain feature matrix E ∈ R C×B , where B is the number of frequency bands;

[0024] Step 6 Calculate the covariance matrix of the five types of task data. For each class ω ∈ {1, 5}, calculate its covariance matrix:

[0025]

[0026] Solve the generalized eigenvalue decomposition:

[0027] C ω W = λ(C1 + C2 + C3 + C4 + C5)W

[0028] where W ∈ R C×C is the projection matrix, the eigenvalues λ are arranged in descending order, and C1 + C2 + C3 + C4 + C5 is the sum of the covariance matrices of the five types of tasks;

[0029] Step 7 Select the first k2 eigenvectors to form the spatial filter:

[0030]

[0031] Step 8 Perform spatial filtering on the original EEG signal to obtain the eigenvectors:

[0032] The original signal X n , perform spatial filtering on it:

[0033]

[0034] Calculate the logarithmic variance as the feature to obtain the final eigenvector f n ∈R 2k ;

[0035] Step 9 For each frequency band, perform CSP processing (Steps 6 - 8) separately to obtain the eigenvectors;

[0036]

[0037] Step 10 Stack all band features to construct an input matrix:

[0038] I ∈ R 2k×B×1

[0039] Step 11 Construct a variable convolutional layer, select different convolution methods according to different stages of convolution, use regular convolution in the early stage of the convolutional neural network, and use depthwise separable convolution in the later stage.

[0040] Step 12 Construct an asymptotically scaled convolutional kernel, adopt an asymptotically scaling strategy, and use convolutional kernels of different sizes in different stages:

[0041] H early = RegularConviD(I, K n ) ∈ R 2k×B×D

[0042] H later = DepthwiseConviD(I, K n ) ∈ R 2k×B×D

[0043] Where is the convolutional kernel, d n is the convolutional kernel size, D is the number of output channels, and n is the number of stages.

[0044] Use regular convolution in the early stage to directly fuse local spatial information, avoid the fragmented calculation of depthwise separable convolution, and accelerate feature extraction in the early layers. Use depthwise separable convolution in the later stage to balance parameter efficiency and receptive field in the later layers and enhance the weights of key channels. Gradually adjust the convolutional kernel size, and the model learns spatial features of different scales at different training stages, enhancing the generalization ability of the model.

[0045] Using a smaller convolutional kernel size in the early stage can accelerate the training of the network, but usually reduces the accuracy of the model. Research has found that the reduction in accuracy is caused by unbalanced regularization. Therefore, use an asymptotic regularization training strategy. When using a small-size convolutional kernel in the early training, use a weaker regularization strategy, so that the neural network can quickly learn some simple expressions. Then gradually increase the image size while enhancing the regularization method to enhance the accuracy of the model.

[0046] The training sizes and regularization strengths used in different training stages are shown in the following formula:

[0047] Perform the following operations for each stage i (from 0 to M - 1):

[0048]

[0049] where S0 is the initial convolution kernel size, is the initialization regularization parameter, S e is the final image size, is the final regularization parameter, N is the total number of training steps, M is the number of training stages, is the regularization parameter of the i-th stage, R i is the set of all regularization parameters in the i-th stage.

[0050] Step 13 constructs a global average pooling layer and a fully connected layer.

[0051] Step 14 configures an optimizer and a loss function.

[0052] Step 15 defines an early stopping callback function.

[0053] Step 16 trains the model.

[0054] Compared with the prior art, the advantages of the present invention are as follows:

[0055] 1. Improve the model classification ability:

[0056] By combining frequency-domain - spatial features and a variable asymptotic convolutional neural network model, it pays more attention to the global features of EEG signals. When a traditional convolutional neural network is used for driving EEG signal recognition, the convolution kernel size is smaller than the input image size, focusing on local information and having insufficient global perception of the signal, making it difficult to capture complex patterns and long-range dependencies, resulting in a low recognition rate. However, this method can effectively solve this problem, significantly improve the model classification ability, and thus improve the EEG signal classification and recognition effect.

[0057] 2. Meet the real-time requirements:

[0058] The model has the characteristic of being lightweight. Traditional EEG signal recognition models have high requirements for computing resources and low computing efficiency, and cannot meet the stringent real-time requirements of intelligent driving. The model of the present invention can solve the real-time problem, and it has obvious advantages in the intelligent driving scenario. The processing speed and accuracy of EEG signals are directly related to the response ability of the intelligent driving system to complex road conditions, and this model can help achieve safe and efficient driving control.

[0059] 3. Reduce the risk of overfitting:

[0060] Traditional convolutional neural network models used to process electroencephalogram (EEG) signals, such as EEGNet, require a large amount of labeled data. However, the cost of labeling EEG data is high and the sample size is limited. When the data volume is limited, overfitting is likely to occur and the generalization ability decreases. The method adopted in the present invention can effectively reduce the risk of overfitting and enhance the generalization ability of the model by using a variable asymptotic convolutional neural network, combining an asymptotic scaling strategy and different stages of convolutional methods, in the case of limited data volume.

[0061] 4. Adapt to complex non-stationary signals:

[0062] EEG signals belong to complex non-stationary signals and are easily interfered by various factors during driving, with a large amount of noise and irregular fluctuations. This method focuses on the frequency domain and spatial characteristics of EEG signals, and can more accurately extract effective information through the frequency domain-spatial feature extraction step. The frequency domain features are obtained by processing with the fast Fourier transform (FFT), and then the spatial features are obtained by processing with common spatial patterns (CSP), improving the accuracy and reliability of driving EEG signal recognition. Improve the accuracy and reliability of driving EEG signal recognition. Brief Description of the Drawings

[0063] Figure 1 It is a system flowchart of the present invention. Detailed Embodiment

[0064] To deepen the understanding of the present invention, the following further introduces the solution in combination with the embodiments. Example: Refer to Figure 1 A method for recognizing driving EEG signals based on a multi-domain asymptotic convolutional neural network, the method comprising the following steps:

[0065] Step 1, Driving data acquisition:

[0066] Select a 64-channel EEG cap for data acquisition, set the sampling rate to 250 Hz, the bandwidth to 0.5 - 100 Hz, and synchronously record the vehicle data status.

[0067] Experimental paradigm:

[0068] Design a simulated driving task including five types of scenarios: normal driving, emergency deceleration, lane change, vehicle turning, and vehicle acceleration. Each experimental trial lasts for 5 minutes, with a 2-minute break in between.

[0069] Step 2, Data preprocessing:

[0070] Perform power frequency notch filtering (50 Hz) on the data to remove the 50 Hz interference from the mains; perform band-pass filtering, with the frequency set to (0.5 - 45 Hz), to remove interference such as baseline drift and high-frequency myoelectric noise, and retain valuable signal components. Obtain eeg c lean.

[0071] Step 3 Data windowing processing:

[0072] C channels, total duration T seconds, sampling rate f s = 250Hz, divide the signal into N = int(eeg c lean.shape[1] / S) windows, each segment L = 2 (seconds), and each segment contains S = L×f s = 500 sampling points. After segmentation, the data matrix is X n ∈R C×S , n = 1, 2, 3…, N;

[0073] Step 4 Extract frequency-domain features from each time-window data matrix:

[0074] Use the numpy library in python:

[0075] Convert to the frequency domain using the fast Fourier transform:

[0076] Fc = np.fft.rfft(Xn,axis = 2)

[0077] Calculate the power spectral density:

[0078] Pc = np.abs(Fc)**2

[0079] Step 5 Divide the frequency band and calculate the frequency-domain feature matrix:

[0080] Divide the frequency band matrix:

[0081] ('delta', 1, 4),

[0082] ('theta', 4, 8),

[0083] ('alpha', 8, 12),

[0084] ('beta', 12, 30)

[0085] The first elements 'delta', 'theta', 'alpha', and 'beta' are common frequency band names, the second element is the starting frequency of the frequency band (unit: Hz), and the third element is the ending frequency of the frequency band (unit: Hz); B = 4 is the number of frequency bands.

[0086] Calculate the frequency-domain feature matrix:

[0087] f low is the starting frequency of the current frequency band, f high is the ending frequency of the current frequency band;

[0088] Corresponding index frequency:

[0089]

[0090] Frequency domain feature matrix:

[0091]

[0092] Step 6: Calculate the covariance matrix of the five types of task data. For each class ω ∈ {1, 5}, calculate its covariance matrix:

[0093]

[0094] Solve the generalized eigenvalue decomposition:

[0095] C ω W = λ(C1 + C2 + C3 + C4 + C5)W

[0096] where W ∈ R C×C is the projection matrix, the eigenvalues λ are arranged in descending order, and C1 + C2 + C3 + C4 + C5 is the sum of the covariance matrices of the five types of tasks;

[0097] Step 7: Select the first k2 = 3 eigenvectors to form a spatial filter:

[0098]

[0099] Step 8: Perform spatial filtering on the original EEG signal to obtain eigenvectors:

[0100] Original signal X n , perform spatial filtering on it:

[0101]

[0102] Calculate the logarithmic variance as a feature to obtain the final eigenvector 2k3 is the feature dimension of each frequency band;

[0103] Step 9: For each frequency band, perform CSP processing (Steps 6 - 8) respectively to obtain eigenvectors;

[0104]

[0105] Step 10: Stack all frequency band features to construct an input matrix:

[0106] Use the tensorflow library in python

[0107]

[0108] inputs = tf.keras.Input(shape = inputShape) (2k3, B, 1)

[0109] Step 11 Construct a variable asymptotic convolutional neural network type:

[0110] Use the TensorFlow library in Python:

[0111] Step 11.1 Define a variable convolutional layer and implement an asymptotic scaling convolutional kernel and a regularization strategy:

[0112] Define a variable convolutional layer and select different types of convolutional layers according to the training stage. Set a variable stage to represent the current training stage and total_stages to represent the total number of training stages.

[0113] When stage < total_stages / / 3, use the conventional convolutional layer layers.Conv2D; when stage >= total_stages / / 3, use the depthwise separable convolutional layer layers.SeparableConv2D. Expressed in code as follows:

[0114] if stage < total_stages / / 3:

[0115] self.conv = layers.Conv2D(filters, kernel_size, padding='same')

[0116] else:

[0117] self.conv = layers.SeparableConv2D(filters, kernel_size, padding='same')

[0118] Define a callback function ProgressiveTrainingCallback. Assuming the total number of training epochs is N and the entire training process is divided into M stages, then each stage contains epochs of training.

[0119] Dynamic adjustment of the convolutional kernel size: For each stage i (from 0 to M - 1), the convolutional kernel size S i is adjusted according to the following linear interpolation formula:

[0120]

[0121] where S0 is the initial convolutional kernel size and S e is the final convolutional kernel size.

[0122] Dynamic adjustment of the regularization strength: The regularization parameter It is also adjusted according to the linear interpolation formula:

[0123]

[0124] where is the initial regularization parameter is the final regularization parameter, R i is the set of all regularization parameters at the i-th stage.

[0125] At the beginning of each stage, the convolutional kernel size of all layers in the model and the parameters of the Dropout layer are updated. For example, the convolutional kernel size linearly increases from the initial 3×3 to the final 5×5.

[0126] Step 11.2 Global average pooling

[0127] Perform global average pooling operation on the output x of the previous convolutional layer, implemented using tf.keras.layers.GlobalAveragePooling2D():

[0128] x pooled = GlobalAveragePooling2D(x)

[0129] Step 11.3 Fully connected layer

[0130] Input the output x after global average pooling pooled into the fully connected layer. The number of neurons in the fully connected layer is num_classes = 5, a fully connected layer with 5 output units, and the activation function is selected as softmax for multi-classification tasks:

[0131] outputs = Dense(num_classes, activation='softmax')(x pooled )

[0132] Step 12 Configure the optimizer and loss function

[0133] Optimizer: Use the Adam optimizer and pass the learning rate LR into the optimizer:

[0134] optimizer = Adam(learning_rate = LR)

[0135] Loss function: Select categorical_crossentropy as the loss function, which is suitable for multi-classification tasks, assuming the labels are one-hot encoded:

[0136] L = categorical_crossentropy(y true , ypred )

[0137] Among them, y true is the true label, and y pred is the predicted label of the model.

[0138] Evaluation criterion: The accuracy is selected as the index to evaluate the performance of the model.

[0139] Data augmentation: Use the ImageDataGenerator tool in Keras for data augmentation, set rotation_range = 5 (the range of random rotation of the image is ±5 degrees), and width_shift_range = 0.1 (the proportion of random translation of the image in the width direction is ±0.1).

[0140] Step 13 Model training parameter configuration

[0141] Set the initial convolutional kernel size int_size = 3 and the initial regularization parameter ini_dropout = 0.1.

[0142] Batch size: Set the batch size BATCH_SIZE = 32, that is, 32 samples are used for each training.

[0143] Number of training epochs: Set the number of training epochs EPOCHS = 1000, and the model will perform 1000 iterative trainings on the entire training dataset.

[0144] Learning rate: Set the learning rate LR = 1e-4 to control the step size of model parameter update.

[0145] Step 14 Model training

[0146] Use the model.fit() method to train the model

[0147] Set shuffle = True: Before the start of each round of training, the training data will be randomly shuffled so that the model can learn the features of the data more comprehensively.

[0148] Set the weights class_weight = {0: 1.30, 1: 1.89, 2: 1, 3: 1.40, 4: 2.0}. Since the number of samples in different classes varies greatly, a weight is assigned to each class so that the model pays more attention to the classes with fewer samples when calculating the loss.

[0149] Set a list of callback functions callbacks = [early_stopping] to perform some additional operations during training. The early_stopping callback function is used to stop training early when the validation performance of the model no longer improves, avoiding overfitting.

[0150] Set the model.fit() method to return a History object that stores various metrics during training, such as training loss, validation loss, training accuracy, validation accuracy, etc.

[0151] It should be noted that the above embodiments are not used to limit the protection scope of the present invention. Equivalent transformations or substitutions made on the basis of the above technical solutions fall within the protection scope of the claims of the present invention.

Claims

1. A driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network, characterized in that: The following steps are involved: Step 1. Obtain raw EEG data through the brain-computer interface and pre-process the data. Step 2: Divide the preprocessed EEG signal into multiple windows, and convert each window signal into the frequency domain F by fast Fourier transform. c , calculate the power spectral density P c , for each channel, the specific frequency band integrated energy is extracted to extract the frequency domain features of the EEG signal and obtain the frequency domain feature matrix; the frequency domain feature matrix is ​​transformed by co-spatial mode, and for each category ω∈{1,5}, its covariance matrix is ​​calculated, and the generalized eigenvalue decomposition is solved. The k eigenvectors before and after are selected to form a spatial filter, and the original signal is spatially filtered to obtain the spatial vector, and the spatial features of the EEG signal are extracted; Step 3. Stack all the frequency band features and construct the input matrix; construct a variable asymptotic convolutional neural network model, use conventional convolution in the early stage of the convolutional neural network, use depthwise separable convolution in the later stage, and adopt asymptotic scaling strategy, using convolution kernels of different sizes and regularization methods of different strengths at different stages; Step 4. Input the input matrix in step 3 into the variable asymptotic convolutional neural network model for training to obtain the classification of driving behavior and evaluate the accuracy of the classification.

2. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: The raw EEG data preprocessing step in step 1 includes power frequency notching and bandpass filtering, with the frequency set to 0.5-45 Hz to remove baseline drift and high-frequency EMG noise.

3. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: The implementation process of dividing the EEG signal into time windows in step 2 includes: According to the sampling rate f s , with a total duration of T seconds, the signal is divided into windows, each time segment is L = 2 seconds long, and each segment contains S = L × f s .

4. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: The specific process of obtaining the frequency domain feature matrix in step 2 is: For each time window, fast Fourier transform is performed according to the following formula: Fc n [c,k1] is the input signal X n [c,s] The result after Fourier transformation; s is the sampling point index, S is the total number of sampling points, c is the channel index (1≤c≤C), k1 is the frequency index (0≤k1≤S), i is the imaginary unit, and C is the total number of channels; Calculate the power spectral density: Frequency band division: Bands={(delta,1,4),(theta,4,8),(alpha,8,12),(beta,12,30)} Each frequency band consists of a name, a starting frequency f low , end frequency f high composition; For each frequency band, calculate its frequency index in the power spectral density matrix: Frequency domain feature matrix: Indicates the energy of the cth channel in the current frequency band.

5. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: The specific process of extracting the spatial features of the EEG signal in step 2 is: Find the covariance matrix of the five types of task data. For each category ω∈{1,5}, calculate it according to the following formula: c ω is the covariance matrix for each category; The generalized eigenvalue decomposition is performed according to the following formula: C ω In=λ(C1+C2+C3+C4+C5)In where W∈R C×C is the projection matrix, the eigenvalues ​​λ are arranged in descending order, and C1+C2+C3+C4+C5 is the sum of the covariance matrices of the five types of tasks; The spatial filter is constructed according to the following formula: k2 is the number of eigenvectors selected (usually k2=3), are the selected front and back k eigenvectors; Spatial filtering is performed according to the following formula: Where Z n is the signal after spatial filtering, W CSP is the spatial filter matrix, X n is the input EEG signal; The feature vector is extracted according to the following formula: f n =log(var(Z n ,axis=1)) var() indicates variance calculation, axis=1 indicates variance calculation by row, and log() indicates logarithmic transformation.

6. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 5 is characterized in that: The specific process of constructing the input matrix in step 3 is: Perform spatial feature extraction on each frequency band to obtain the frequency band feature vector Stack all band eigenvectors to construct the input matrix I: B is the number of frequency bands, 2k3 is the feature dimension of each frequency band, is the dimension of the construction matrix.

7. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: In step 3, a variable asymptotic convolutional neural network model is constructed. The variable method is to choose different convolution methods at different stages of the convolution of the convolutional neural network. Conventional convolution is used in the early stage of the convolutional neural network, and depth-separable convolution is used in the later stage to obtain the convolution result H early or H later ; H early =RegularConviD(I,K n )∈R 2k×B×D H later =DepthwiseConviD(I,K n )∈R 2k×B×D in is the convolution kernel, d n is the convolution kernel size, D is the number of output channels, n is the number of stages, and I is the input matrix.

8. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 1 is characterized in that: In step 3, a variable asymptotic convolutional neural network model is constructed, where the asymptotic approach is to simultaneously adopt an asymptotic scaling strategy, using convolution kernels of different sizes and regularization methods of different strengths at different stages.

9. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 7 is characterized in that: In step 4, the variable asymptotic convolutional neural network model constructs a global average pooling layer, a fully connected layer with 5 output units, configures the Adam optimizer, the learning rate LR = 1e-4, selects categorical_crossentropy as the loss function, and selects accuracy as the indicator for evaluating model performance; the H obtained in step 3 is early and H later Input to the global average pooling layer for model training.

10. The driving EEG signal recognition method based on multi-domain asymptotic convolutional neural network according to claim 9 is characterized in that: The specific process of using the asymptotic scaling strategy and using different sizes of convolution kernels and different strengths of regularization methods at different stages is: For each stage i (from 0 to M-1), do the following: Where S0 is the initial convolution kernel size, is the initialization regularization parameter, S e is the final image size, is the final regularization parameter, N is the total number of training steps, M is the number of training stages, is the regularization parameter of the i-th stage, R i is the set of all regularization parameters in the i-th stage.

Images