CNN Based Multi-class Electroencephalographic Signal Recognition

In recent years,research on brain-computer interface(BCI)systems has become a popular direction and cutting-edge technology in artificial intelligence research and application fields.BCI technology allows humans to control external devices through moving image(MI)signals.It has a good prospect in medical rehabilitation and military technology.The key to this technology is the correct recognition and classification of EEG signals,but many existing feature extraction algorithms The inability to eliminate the effects of individual differences has hindered the widespread application of BCI systems.Based on the existing BCI technology,this study proposes a new processing algorithm through individual discrimination and feature extraction of EEG motion imagination signals.It combines a preprocessing method that eliminates individual differences with a convolutional neural network(CNN)structure.And verified by the BCI competition data set.The work of this paper includes the analysis and screening of EEG signals,algorithm design of data expansion and feature extraction and classification.In the EEG signal preprocessing method,the initial signal filtering or denoising reconstruction method is usually used to remove unnecessary frequency band information and improve the signal signal to noise ratio,but the filtering range and denoising threshold settings are not rigorous,resulting in EEG classification.The algorithm has poor universality and low practical applicability.The extracted features include spatio-temporal domain features and frequency-domain features.The spatio-temporal domain features use the convolution kernel to perform window sliding on the initial signal to extract time-domain and space-domain features.This method is greatly affected by experimental data and has low reliability.The frequency domain feature is a fast Fourier transform of the time domain signal to form a feature vector based on energy values for classification.This method is more complex for multichannel EEG data.In this paper,we use the discrete wavelet transform(DWT)to calculate the energy in eachsub-band to select the best frequency band,and use the power spectral density and convolutional neural network based on visual geometric group network for feature extraction and classification.The experimental framework is described as follows: First,the input EEG signal(filters out the high-frequency part of 50-100Hz)is subjected to a DWT transform with a scale of 1,and the energy ratios of the five sub-bands of the initial frequency band(0-50Hz)are obtained.Filtering reserves the highest proportion of subbands.Secondly,the time-domain signal after frequency band selection is expanded by sliding the time window to shorten the time length of the input sample without changing the original information of the signal.Next,obtain the power spectral density(PSD)map of the expanded samples and use it as the feature map of the training and test sets of the CNN.Finally,the training set PSD feature map is input to the CNN network for training,and the test set feature map is classified to obtain a classification accuracy rate.In terms of performance of this study,we mainly compared the classification accuracy of the imaginary signals of the tongue movements of the same left and right hands and legs.In the classification experiments of EEG signals,the average training accuracy of CNN is 99%,and the average classification accuracy is 96.21%,far exceeding the results obtained by the latest research.In terms of eliminating individual differences,the classification accuracy rate of the nine participants was more than 90%,and the losses were less than 0.35.The algorithm proposed in this paper,based on the calculation of the energy distribution,automatically selects the frequency band for no individual,which eliminates the influence of individual differences on the classification accuracy and improves the universality of the algorithm.Secondly,the PSD feature and the CNN classifier are selected in combination with the energy analysis,which greatly improves the classification accuracy of the four-class EEG signals,reduces training iterations,and improves training speed.It is a new algorithm with high accuracy and strong universality,with less network parameters,and provides a new algorithm for future real-time EEG classification systems.