Research On Machine Learning Method For Feature Extraction And Classification Of Four Kinds Of Motion Imagination EEG Signals

In recent years,with the hot development of machine learning,the research field of brain-computer interface has become popular again,and corresponding problems have been exposed.The basic features of EEG signal are non-linearity and nonstationarity,which indicates that different people have different EEG signals.There is no uniform and perfect method for preprocessing and feature extraction.In view of this,this paper presents a new preprocessing scheme based on the electroencephalogram(EEG)signals of 9 subjects from the 2a dataset of BCI Competition IV in 2008.In addition,this paper introduces SSA for the first time to optimize the kernel and penalty coefficients of SVM classifiers.This paper mainly does the following research:(1)In the process of preprocessing,a new set of preprocessing scheme is designed:band-pass filter + sliding time window + overlay averaging.The selection of traditional time slices is usually a fixed time period,which has limitations for different subjects.Firstly,FIR band-pass filtering is applied to the original dataset,then a sliding window is used to select the time period,and then the time period is averaged by overlaying.(2)In the aspect of feature processing,this paper chooses the wavelet packet transformation + common space mode to extract the features of the pre-processed data.The traditional wavelet transform is not very friendly to the high frequency part of the signal.Wavelet packet transformation overcomes this shortcoming very well.It is very sensitive to the characteristics of time-frequency domain.The common space mode can extract the spatial features effectively.This combination of feature extraction takes into account the information in time,frequency and space domains.(3)To verify the effectiveness of the new pre-processing combination,three traditional classifiers are designed: ANN,CNN and SVM.The classification results of the three classifiers show that the pre-processing scheme proposed in this paper is effective and can improve the classification accuracy and Kappa coefficients to a certain extent,especially for EEG signals with poor signal quality.In addition,for the seventh participant in this dataset,the data is excellent,and the average classification accuracy of SVM can reach 86%,which is the highest of all classification accuracy.(4)Although the work(3)has verified the validity of the new preprocessing scheme proposed in this paper,the SVM with the best classification performance among the three classifiers is slightly less than those in other literature.In order to further improve the classification accuracy,this paper introduces SSA by optimizing the SVM kernel coefficient and penalty factor,and presents a new optimization scheme:SSA-SVM.Experiments show that the SSA-SVM classifier improves significantly as a whole,and the highest average classification accuracy reaches 90%.For subjects with poor data quality,the classification effect is also significantly improved.Finally,by comparing the results with other literature with similar feature extraction schemes,the results show that SSA-SVM still has advantages.This study not only provides a new scheme for the preprocessing of EEG signal,but also provides a new idea for the classification of EEG signal.