CN110069958B - Electroencephalogram signal rapid identification method of dense deep convolutional neural network - Google Patents

Abstract

The present invention proposes a method for quickly identifying electroencephalogram (EEG) signals by a dense deep convolutional neural network. Combined with the temporal and spatial characteristics of the electroencephalographic signal of motor imagination, a method of feature connection is used in the convolutional neural network to design A Convolutional Neural Network for Motor Imagery EEG Signals. The convolutional neural network designed by the invention can simultaneously extract temporal and spatial features, and connect the outputs of different convolutional layers to each other, thereby reducing the number of weights and achieving the purpose of anti-overfitting and feature reuse. First, the filtered and resampled raw data is input into the dense deep convolutional neural network, then the parameters of each layer of the network are updated through backpropagation and stochastic gradient descent algorithms, and finally the network is tested, and the test data is input into the trained network. The output results are analyzed. Compared with the Shallow ConvNet method proposed in 2017, the present invention improves the signal recognition accuracy and kappa value by 5% and 0.066.

Description

A fast recognition method of EEG signals based on dense deep convolutional neural network

technical field

The invention relates to the rapid identification of original EEG signals, the design of a convolutional neural network suitable for EEG signals, pattern classification and deep learning, and belongs to the technical field of signal processing and pattern recognition.

Background technique

Brain-computer interface (BCI) technology can establish a connection between the human brain and external devices, so as to achieve the purpose of communicating and controlling the external environment without relying on human muscles. The main processing process of BCI technology includes recording brain activity, EEG (Electroencephalogram, EEG) signal processing, signal recognition, and then controlling external devices according to the recognition results. At present, there are many types of EEG signals suitable for BCI, such as P300, steady-state visual evoked potentials, and motor imagery. P300 is an EEG signal induced by a small probability flickering signal that occurs occasionally. The stimulation of visual evoked potential is a flickering picture that appears at a fixed frequency, while the motor imagery signal only requires the subject to imagine and perform the action of a certain part, without the need for real Execution, motor imagery EEG signal has the characteristics of easy acquisition, no external stimulation, and asynchronous communication. Therefore, motor imagery EEG signal has become one of the most widely used EEG signal types.

For the feature extraction of motor imagery EEG signals, the common spatial pattern (CSP) method is commonly used, but the feature extraction effect of this algorithm depends on the frequency bandwidth range specified by the algorithm. On this basis, the proposed filter bank common spatial pattern algorithm (Filter Bank Common Spatial Pattern, FBCSP) designs the filter bank, expands the frequency range, uses the CSP algorithm for each filter, and then uses the output of the filter bank. Selecting the available features in BCIIV datasets 2a, the algorithm achieved a classification accuracy of 68%. However, the FBCSP algorithm still needs prior knowledge to design the range of each frequency bandwidth, and needs to manually extract features before classification. However, brain signals are very complex, and many collected signals have not yet found a clear meaning. Manual feature extraction will cause information loss. .

In recent years, with the rise of deep learning, convolutional neural network (CNN) has received extensive attention from researchers, and has achieved certain application effects in many fields such as image, voice, video, etc. People use convolutional neural networks to automatically extract motor imagery EEG signal features for classification. The weight sharing network structure of the convolutional neural network can not only reduce the number of weights and reduce the complexity of the network model, but also extract temporal and spatial features. In the training process, the neural network relies on the back-propagation algorithm to update the parameters of each convolution kernel, and turns different layers into suitable feature extractors, which can avoid the use of artificially designed feature extractors, thereby extracting more features to improve classification. effect of accuracy. Different from the two-dimensional static pictures, the motor imagery EEG signal is a dynamic time series collected from the three-dimensional brain scalp, and the motor imagery EEG signal has a low signal-to-noise ratio and is easily disturbed by noise unrelated to events, such as the disturbance of electrodes. , light stimulation, subjects' eye movements, etc., which also make it difficult to train convolutional neural networks from the original motor imagery EEG signals, so when designing a convolutional neural network, it is necessary to adjust the structure of the network according to the characteristics of the EEG signal. In 2017, the convolutional neural network structure Shallow ConcNet proposed by Schirrmeister et al. takes the original motor imagery EEG signal as input, and the input original motor imagery signal can be obtained through temporal convolution, spatial convolution, average pooling and other operations. Class Probability, an end-to-end automatic identification method. End-to-end means that the convolutional neural network does not require any prior knowledge, learns features from the original motor imagery EEG signals, and directly obtains the final classification result. Using the Shallow ConvNet structure on the BCIIV2a dataset, the accuracy rate reaches 72%, which is 5% higher than the traditional FBCSP method. It can be seen that the use of convolutional neural networks can significantly improve the recognition accuracy of motor imagery EEG signals. However, the accuracy of the model is still not high, because only two convolutional layers are used, and directly deepening the model will lead to serious overfitting, so it is impossible to extract deeper features. The invention studies how to adjust the connection mode and hyperparameters of the model in combination with the characteristics of the motor imagery EEG signal, so as to deepen the model depth while avoiding aggravating overfitting, and improve the recognition accuracy of the motor imagery EEG signal.

SUMMARY OF THE INVENTION

Based on Shallow ConvNet, the present invention proposes a dense deep convolutional neural network. The accuracy of the present invention is 77% when tested on the BCIIV2a data set. Compared with the 72% accuracy of the Shallow ConvNet network on the BCIIV2a data set, the accuracy of the present invention is still 5% higher. The accuracy of EEG signal recognition has been significantly improved. Compared with Shallow ConvNet, the dense connection method proposed by the present invention connects the input and output feature maps of the intermediate convolution layer as the input of the next layer, so that the feature maps generated by the two convolution layers can be directly transmitted to the next layer, The features of the intermediate layer are fully utilized without increasing the amount of parameters, so the accuracy of the model proposed by the present invention is higher.

The present invention designs a dense deep convolutional neural network to realize the end-to-end identification of the original motor imagery EEG signal, including the following steps:

(1) Perform third-order band-pass 0-40hz filtering on the EEG signal to filter out some noise or other irrelevant components during signal acquisition;

(2) Resampling the filtered signal. The reason is that the length of the data input to the convolutional neural network needs to be consistent, and to ensure the same amount of data under the same time length, it is necessary to sample the data of different sampling frequencies to the same frequency;

(3) Intercept fixed-length events from the above preprocessed data, and obtain their corresponding labels;

(4) Design a dense deep convolutional neural network. Different from the conventional convolutional neural network that adopts continuous convolution and pooling operations, the dense convolutional neural network designed in the present invention connects the input and output feature maps of the two middle convolutional layers. Then as the input of the next layer;

⑸ Train dense deep convolutional neural networks. The squared error function is used as the loss function to calculate the error between the predicted value and the label, and the parameters of each layer of the network are updated through the back-propagation and stochastic gradient descent algorithm. When the accuracy rate converges to a certain value, or the accuracy rate decreases, the training is stopped. ;

(6) Test the dense deep convolutional neural network, the parameters of the network will not change, input the test data and labels, and analyze the output results.

The specific steps of the above step (4) are as follows:

①Data input layer: The format of the input data is a four-dimensional array n×m×990×22, and the meaning of each dimension of the array is the number of samples×the number of feature maps×the number of sampling points×the number of channels;

②Temporal convolution layer: perform convolution operation on the data in the time dimension, the size of the convolution kernel is 11×1, and 25 feature maps are generated;

③Spatial convolution layer: In order to reduce the data dimension, spatial convolution is performed on the data, and all channels are mapped to a feature map. The size of the convolution kernel is 22×1. After convolution, it enters the activation function, and the number of feature maps remains is 25;

④ Pooling layer: the input is the output of the upper layer activation function, the size of the pooling kernel is 3×1, and the step size is 3;

⑤ Feature connection layer: The pooled data first enters two consecutive convolution pooling layers, the first convolution kernel size is 1×1, the number is 200, and the second convolution kernel size is 11×1 , the number is 50, and the 50 output feature maps after activation of the second convolution layer are connected with the 25 input feature maps of the first convolution layer, so there are a total of 75 feature maps in the output;

⑥Pooling layer: the operation is the same as ④;

⑦The last two layers of the network are the fully connected layer and the output layer. The fully connected layer expands the pooled output of the upper layer into one-dimensional data. There are four types of labels, so the output layer has four neurons, and the output result is that the input data belongs to each The probability value of the category.

It should be noted that after each convolution operation of the network, a batch regularization operation is performed on the data and then the activation function is input. Ioffe et al. proved that the use of batch regularization operations in deep neural networks can significantly reduce the number of iterations, while reducing each iteration. required computation time.

The specific operations of the above steps ⑸ are as follows:

① Separate 20% from the training set of the BCIIV2a dataset as the validation set, and the remaining 80% as the training set, while the test set remains unchanged;

②For the first training, first iteratively train on the training set, and then test on the validation set. Stop training when the test accuracy on the validation set no longer changes or the accuracy drops, record the best accuracy on the validation set at this time, and save the model. After many experiments, it is found that the accuracy of the model does not change after the number of iterations of the dataset reaches 1000, so for the BCIIV2a dataset, the maximum number of iterations is set to 1000;

③ In the second training, the training set and the validation set are mixed together as a new training set and tested on the validation set. Continue training on the new training set and test on the original validation set until the accuracy on the validation set is higher than the value recorded in ② and does not change, or stop training after 1000 iterations;

④ After the second training, save the model, input test data and labels, and record the test results.

The advantages of the motor-imagination-oriented EEG signal fast identification method provided by the present invention include:

①The original EEG signal data can be used to train the neural network with simple operations such as filtering and resampling. It is an end-to-end method that does not require time-frequency analysis of the signal or prior knowledge to manually select features. Thereby avoiding information loss;

② Compared with the Shallow ConvNet model, the dense convolutional neural network proposed by the present invention has more layers and can extract deeper features;

③ The present invention proposes a feature connection layer, which connects the input and output of the intermediate convolutional layer and transmits it to the next layer, which not only effectively solves the problem of gradient disappearance, but also supports feature reuse without increasing the number of parameters, which can be effectively suppressed when the amount of data is small. overfitting;

④The model has strong generalization ability. According to different data sets, it can be applied to similar motor imagery EEG signal data sets only by changing the parameters of the input and output layers and fine-tuning some hyperparameters of other layers.

Description of drawings

FIG. 1 is a structural diagram of a dense deep convolutional neural network proposed by the present invention.

FIG. 2 is a flow chart of the data processing, training and testing process of the present invention.

Figure 3 shows the confusion matrix of the classification results of Shallow ConvNet on BCIIV 2a.

Figure 4 is the confusion matrix of the classification result of the network of the present invention on BCIIV 2a.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The invention mainly utilizes the weight sharing of the convolutional neural network and the idea of the local receptive field, and connects the channels of the output feature map of the middle layer to improve the recognition accuracy. Each neuron of the convolutional neural network uses the same convolution kernel when convolving different feature maps, which will greatly reduce the amount of weight parameters. Reuse, so that only very few new feature maps are generated after convolution to achieve the purpose of reducing redundancy. The dense deep convolutional neural network of the present invention uses the stochastic gradient descent method to backpropagate the error, adjust the weight of the convolution kernel, and finally obtain the probability that the input data belongs to each category through the full connection and the linear classification layer.

The convolutional neural network feature extraction and classification method of the present invention specifically includes the following steps:

⑴ Get data. The data of the present invention comes from the data set BCICompetition IV Dataset2a provided by the Berlin BCI Research Group in 2008. The dataset consists of 9 groups, collected from 9 healthy subjects, and divided into two parts: training data and test data. For each subject, perform four types of motor imagery: left hand, right hand, foot, and tongue;

(2) Data preprocessing, the sampling frequency is 100Hz, the third-order Butterworth filter in the filter toolbox of the Scipy library is used, and the band-pass filter is set to 0-40hz to filter out high-frequency signals and some noise;

(3) Design a dense deep convolutional neural network, the network structure can refer to Figure 1. The first layer is the input layer, the input data is 990 × 22, 22 represents 22 channels, and 990 is the number of sampling points. The second layer is the temporal convolution layer, which performs the convolution operation on the data in the time dimension. The size of the convolution kernel is 11×1, and 25 feature maps are generated. The third layer is a spatial convolution layer, and the size of the convolution kernel is 22 × 1. After passing through the Relu activation function, 25 feature maps are obtained. The fourth layer is the pooling layer, the pooling range is 3×1, the pooling method adopts maximum pooling, and the step size is 3. Next is the feature connection layer, firstly a 1×1 convolutional layer with 200 convolution kernels, and then an 11×1 convolutional layer with 50 convolution kernels, which are connected to the pooling layer through feature connections. The last two layers of the network are fully connected and linear classification layers. There are four types of data labels, so there are four output units after full connection, and the output result is the probability value of the input data belonging to each category;

(4) Training dense deep convolutional neural networks;

The network of the present invention adopts the square error cost function as the judgment, and the formula is as follows:

表示第n个样本对应标签的第k维，

表示第n个样本对应网络输出的第k个输出。where N represents the number of samples, c represents the number of sample classes,

represents the kth dimension of the label corresponding to the nth sample,

Indicates that the nth sample corresponds to the kth output of the network output.

The first layer is the data input layer, the input data format is fixed, and there are no parameters that need to be trained.

The second layer is a temporal convolution layer. Each time a batch of samples is input, the input image is convolved with 25 convolution kernels with a size of 11 × 1, a stride of 1 and a random initialization, and 25 feature maps are obtained. This layer has no bias and activation function, and the formula is as follows:

为第l层的第j个特征图，M_j为输入的特征图集合，

是输入的第i种特征图和输出的第j种特征图之间的连接所用的卷积核。in,

is the _jth feature map of the lth layer, Mj is the input feature map set,

is the convolution kernel used for the connection between the input i-th feature map and the output j-th feature map.

The third layer is a spatial convolution layer, the input is the output of the upper layer convolution, the size of the convolution kernel is 22 × 1, and the number of input and output feature maps is 25, but this layer has an activation function and no bias term. The formula is as follows:

为第l层的第j个特征图，M_j为输入的特征图集合，

为本层选用的卷积核，f为激活函数Relu函数，即f(x)＝max(0，x)。in

is the _jth feature map of the lth layer, Mj is the input feature map set,

The convolution kernel selected for this layer, f is the activation function Relu function, that is, f(x)=max(0, x).

The fourth layer is the pooling layer, the size is 3×1, the stride is 3, and the input is 25 feature maps. Pooling does not change the number of feature maps, so 25 output feature maps will also be obtained. The formula is as follows:

为第l-1层的第j个特征图，

为第l层偏置，f是池化层的激活函数，这里无激活函数，所以f＝f(x)。其中down()表示一个下采样函数，这里是将相邻3个像素值取最大一个，所以下采样函数为max()。in,

is the jth feature map of the l-1th layer,

is the bias for the first layer, f is the activation function of the pooling layer, there is no activation function here, so f=f(x). where down() represents a downsampling function, here is to take the largest one of the three adjacent pixel values, so the downsampling function is max().

The fifth layer is the feature connection layer, which includes two convolutional layers. The input of the first convolutional layer is the pooled 25 feature maps, using a 1×1 convolution kernel with a stride of 1. Convolution to generate 200 feature maps, and then activated by the relu function to enter the next convolution layer. The size of the second convolution kernel is 11×1, the stride is 1, the input is 200 feature maps, and the output is 50 feature maps. After relu activation, the 50 output feature maps of the second convolution are combined with the first one. The 25 feature maps of the convolution input are connected as a new output feature map, and the formula is as follows:

x _l =H _l ([x ₀ ,x ₁ ]) (5)

where H _l ( ) represents the feature connection operation, x ₀ is the input of the first convolution of the feature connection layer, and x ₁ is the output of the second convolution of the feature connection layer.

The sixth layer is the pooling layer, the input is the feature map after feature connection, the pooling size is 3×1, the step size is 3, and the operation is the same as the fourth layer.

The seventh layer is the fully connected layer, which expands and fully connects the feature map of the upper layer and converts it into one-dimensional data.

The last layer is the output layer, which outputs the required classification results through the sigmoid activation function.

⑸ Calculation and propagation of errors at each layer;

①For the final output layer, the error between the activation value generated by the network and the actual value can be directly calculated. The formula is as follows:

表示输出层未经过激活函数的权重加权，h_w,b(x)表示输出结果，y表示标准输出，

表示n_l层的第i个输出，

表示求导。where the n _lth layer represents the output layer,

Indicates that the output layer is not weighted by the weight of the activation function, h _w,b (x) represents the output result, y represents the standard output,

represents the ith output of n _l layers,

Indicates guidance.

②For each layer error of l=n ₁ -1,n ₁ -2,n ₁ -3,...,2, the general formula is:

表示每个元素相乘，f′(u^l)表示对该层的输出u^l求导。Where w ^l+1 is the weight of the l+1th layer, δ ^(l+1) is the error calculated by the l+1th layer, the symbol

Represents the multiplication of each element, and f'(u ^l ) represents the derivation of the output u ^l of this layer.

If the lth layer is a convolutional layer, the lower layer of this layer is a pooling layer, and one pixel of the pooling layer corresponds to a 3×1 pixel of the output image of the convolutional layer, then there will be a size mismatch. The pooling layer performs upsampling, the formula is:

表示第l+1层权重，δ^(l+1)为第l+1层计算的误差。where up( ) represents the upsampling operation,

represents the weight of the l+1 layer, and δ ^(l+1) is the error calculated by the l+1 layer.

If the lth layer is a downsampling layer and the l+1th layer is a convolutional layer, the formula is:

Among them, conv2 is the convolution implementation function, and rot180 means flipping the convolution kernel by 180 degrees.

⑹ Calculate the final required partial derivative and update the weight parameters, using the formula:

表示旧的权值，

为新的权值，η是学习率。

表示旧的偏置，

为新的偏置。in

represents the old weight,

is the new weight, η is the learning rate.

represents the old bias,

for the new offset.

①For the convolutional layer, the weight update formula is:

是

在做卷积时，与k_ij做卷积的每一个patch，(u,v)是patch中心，输出特征图中(u,v)位置的值是由输入特征图中(u,v)位置的patch和卷积核k_ij卷积所得的值。in

Yes

When doing convolution, each patch that is convolved with k _ij , (u, v) is the center of the patch, and the value of the (u, v) position in the output feature map is determined by the (u, v) position in the input feature map. The value obtained by convolving the patch and the convolution kernel k _ij .

②For the pooling layer, the weight update formula is:

in

⑺Test the network, add test data and real labels, compare the output results with the real labels, and get the confusion matrix of the output results to analyze the model.

The effect of the present invention can be further explained by the experimental results. The test data of the experiment adopts the official data BCIIV2a of the BCI competition. The dataset has a total of nine subjects, and each subject has a training set and a test set. There are four types of labels for the data, which are the motor imagery of the left hand, right hand, foot, and tongue. Each type of label corresponds to 72 samples of data, so the training set and test set of a subject have 288 samples respectively. Figures 3 and 4 are confusion matrices classified using Shallow ConvNet and the method of the present invention. According to Fig. 3, the accuracy rate of ShallowConvNet is 72%, and the kappa value is 0.632. According to Fig. 4, the accuracy rate of the present invention is 77%, and the kappa value is 0.698. Compared with ShallowConvNet, the method of the present invention has a higher recognition accuracy rate of 5% and a higher kappa value of 0.066 by the dense deep convolutional neural network proposed by the present invention. It can be seen that the present invention significantly improves the classification effect of motor imagery EEG signals.

Claims (3)

1. A motor imagery electroencephalogram signal rapid identification method of a dense deep convolutional neural network is characterized by comprising the following steps:

A) inputting a motor imagery electroencephalogram signal into a densely connected deep convolutional neural network, wherein the densely connected deep convolutional neural network comprises:

as an input layer of the first layer, input data is 990 × 22 data of 22 channels 990 sampling points,

a time convolution layer, which is a second layer, performs a convolution operation of the data in a time dimension with a convolution kernel size of 11 x 1, generates 25 feature maps,

the space convolution layer as the third layer has convolution kernel size of 22 × 1, and is used for obtaining 25 characteristic maps after passing through Relu activation function,

the first pooling layer as the fourth layer had a pooling range of 3X 1, the pooling manner adopted maximum pooling with a step size of 3,

the output of the featured connection layer after the first pooling layer goes to the second pooling layer,

the second layer of the pool is a second layer of the pool,

a fully-connected layer after the second pooling layer,

the linear classification layer behind the full connection layer, the data label has four categories, four output units are arranged after full connection, the output result is the probability value of the input data belonging to each category,

wherein:

the time convolution layer inputs one batch of samples at a time, convolves the input image with 25 convolution kernels of size 11 × 1, step size 1 and random initialization, resulting in 25 feature maps, the layer being free of bias and activation functions, as follows:

wherein,

is the jth feature map of the ith layer, M_jIn order to input a set of feature maps,

is a convolution kernel for the connection between the input ith feature map and the output jth feature map,

the input of the space convolution layer is the output of the time convolution layer, the size of the convolution kernel is 22 multiplied by 1, the number of input and output characteristic graphs is 25, the layer has an activation function and has no offset term, and the formula is as follows:

wherein

Is the first layerj feature maps, M_jIn order to input a set of feature maps,

the convolution kernel is selected for this layer, and f is the activation function Relu, i.e. f (x) max (0, x),

the input of the first pooling layer is 25 characteristic graphs, the pooling does not change the number of the characteristic graphs, and 25 output characteristic graphs are obtained, and the formula is as follows:

wherein,

is the jth characteristic diagram of the l-1 layer,

for layer bias, f is the activation function of the pooling layer, where there is no activation function, so f ═ f (x), where down (·) denotes a down-sampling function, where the neighboring 3 pixel values are taken to be the largest one, so the down-sampling function is max (·),

the featured connection layer includes two convolution layers, wherein: the input of the first convolutional layer is 25 feature maps after pooling of the first pooling layer, convolution kernels with the size of 1 x 1 and the step length of 1 are adopted for convolution to generate 200 feature maps, and then the feature maps are activated through a relu function to enter a second convolutional layer; the convolution kernel size of the second convolution layer is 11 × 1, the step size is 1, the input of the convolution kernel size is 200 feature maps output by the first convolution layer, the output of the convolution kernel size is 50 feature maps, the 50 output feature maps of the second convolution are activated by relu and connected with the 25 feature maps input by the first convolution to serve as new output feature maps, and the formula is as follows:

x_l＝H_l([x₀,x₁]) (5)

wherein H_l(. represents a splicing operation, x)₀For the first convolutionInput, x₁Is the output of the second convolution and,

the input of the second pooling layer is the new output characteristic map after the splicing, the pooling size is 3 x 1, the step length is 3,

the full connection layer expands and fully connects the characteristic diagram output by the second pooling layer, converts the characteristic diagram into one-dimensional data,

the linear classification layer outputs probability values of input data belonging to various categories through a sigmoid activation function,

B) calculating and propagating errors for each layer, including:

calculating the error between the activation value and the actual value generated by each layer of the network directly for the final output layer, wherein the formula is as follows:

wherein the n is_lThe layer represents an output layer of the video stream,

weights, h, representing the output layer without the activation function_w,b(x) Indicating the output result, y indicates the standard output,

the ith output of the output layer is represented,

it is indicated that the derivation is performed,

n for l₁-1,n₁-2,n₁-3, …,2 layer error, general formula:

wherein w^l+1Is the weight of layer l +1, δ^l+1For the error calculated for layer l +1, the notation, indicates each elementMultiplication, f' (u)^l) Representing the output u for that layer^lThe derivation is carried out by the derivation,

when the first layer is a convolutional layer, the lower layer of the layer is a pooling layer, one pixel of the pooling layer corresponds to one pixel (3 × 1) of an output image of the convolutional layer, and in order to eliminate the size mismatch phenomenon, the pooling layer is up-sampled by the formula:

where up (-) denotes an upsample operation,

represents the l +1 th layer weight, δ^(l+1)The error calculated for the l +1 th layer,

when the l-th layer is a down-sampling layer, and the l + 1-th layer is a convolutional layer, the error formula is:

where conv2 is the convolution implementation function, rot180 denotes flipping the convolution kernel 180 degrees,

C) updating the weight parameter by using the formula:

wherein

The weight value of the old is represented,

as a new weight, η is the learning rate,

the old offset is represented by the value of the old offset,

in order to be the new offset,

wherein:

for convolutional layers, the weight update formula is:

wherein

Is that

When making convolution, with k_ijEach patch for convolution, (u, v) is the patch center, and the value of the (u, v) position in the output feature map is determined by the patch of the (u, v) position in the input feature map and the convolution kernel k_ijThe value obtained by the convolution is used,

for the pooling layer, the weight updating formula is as follows:

wherein

2. The method for rapidly identifying motor imagery electroencephalogram signals of the dense deep convolutional neural network of claim 1, wherein the following operations are performed before the step A: carrying out third-order band-pass 0-40hz filtering on the originally acquired motor imagery electroencephalogram signals to acquire signal components of larger frequency bands related to the motor imagery electroencephalogram signals and the motor imagery;

resampling the filtered signal components, and sampling data with different sampling frequencies to the same frequency so as to keep the data length of the input convolutional neural network consistent and ensure that the data volume under the same time length is the same;

and intercepting the events with fixed length from the preprocessed data, and acquiring corresponding labels of the events as motor imagery electroencephalogram signals input into the dense connection type deep convolution neural network.

3. The motor imagery electroencephalogram signal rapid identification method of the dense deep convolutional neural network of claim 1, wherein:

the four categories of data tags are the motor imagery of the left hand, right hand, feet, and tongue, respectively.

Images