Noise Benefits In Motor Imagery Electroencephalogram Signals Classification

Noise is always harmful in linear systems and needs to be filtered out.In nonlinear systems,however,noise sometimes helps to enhances the output of the system and optimizes its performance,such phenomenon known as random resonance.Based on stochastic resonance theory framework,this article research the influence of noise on two-class motor imagery EEG signals classification recognition and discuss the existence of stochastic resonance phenomenon.The research theme of this article is noise benefits in motor imagery EEG signals.Through add noise in the EEG signals preprocessing stage and improve the algorithm in the classification recognition phase to improving the accuracy of EEG signals classification.The main research contents of this article are as follows:(1)Adding Gaussian noise to the original EEG signals in motor imagery EEG signals preprocessing phase,and five EEG signals noise-added methods are proposed to explore the influence of noise-added methods on classification system and the optimal noise-added method.Eigenvector extraction and fusion by common spatial pattern and wavelet packet transform,using k-nearest neighbor classifier for EEG signals classification identification.Under the same experimental environment,using nonlinear support vector machine classifier and C4.5 decision tree classifier to avoid contingency.Research shows that the classification accuracy of system depends on Gaussian noise intensity,training sample size and phase of noise-added.Adding appropriate intensity of Gaussian noise to the original EEG signals will improve the system classification accuracy,increasing training sample size will further improve the system classification accuracy,and the highest system classification accuracy is achieved by adding the same intensity of noise to training and test phase while increasing training sample size.Among the five noise-added methods,the most significant noise benefits is achieved in Case5(that is,increasing training sample size and adding the same intensity of noise to training set and test set).(2)A heterogeneous ensemble learning method is proposed in the motor imagery EEG signals classification recognition phase.The base classifiers of heterogeneous ensemble classifier is composed of LIBSVM,Ada Boost and BP neural networks,and second selects the classification results of base classifiers through weighted voting method to obtain the final classification results.Adding Gaussian noise in the signal preprocessing phase and use the optimal noise-added method(that is,Case5).Research shows that the heterogeneous ensemble classifier has better classification performance for system whether the noise is added or not and is robust to the noise,which reflecting the superiority of the heterogeneous ensemble algorithm.In addition,this article also in noise added or not situation compares the difference of system classification accuracy based on different classifiers,and finds that the noise improvement of classification system performance is limited,and stochastic resonance only play an auxiliary role in the optimization of classification system.The more fully trained the classifier itself,the less noise efficiency in the system.(3)The noise-added types in the EEG signals preprocessing phase are expanded from Gaussian noise to generalized Gaussian noise.Three common generalized Gaussian noises are selected from the perspective of the probability density distribution function,namely Laplace,Gaussian and uniform noise.Stochastic resonance phenomenon in classification system is explored under the optimal noise-added method(that is,Case5).Research shows that adding appropriate generalized Gaussian noise to the original EEG signals will improve the system classification accuracy,where system acquires the most noise benefits when adding Laplace noise,Gaussian noise second and minimal uniform noise.The reason is that adding generalized Gaussian noise will improve the sampling value of the original EEG signal,in which adding Laplace noise can almost not change the original EEG signal waveform and almost double promote the original EEG signal sampling value range.While adding Gaussian noise and uniform noise will have impact on the original EEG signal waveform characteristics to a certain degree.