Research On Classification Algorithm Of Brain Computer Interface Based On Motor Imagery EEG Signal

With the development of the Brain-Computer Interface(BCI)technology,BCI equipment is expected to become a popular commercial product within thousands of households.BCI based on motor imagery EEG signal is considered to be one of the most promising technologies.It has a broad application future in military,medical,commercial,entertainment,learning or other application scenarios.Signal classification algorithm is one of the most important research field in BCI technology.In this field,most of the traditional classification algorithms focus on extracting the spatial energy features of EEG signals,or solving the frequency filters first and then extracting the spatial features of the filtered signals,which fails to extract effective features from the higher dimension of the signal,so the performance of the classification algorithm is limited.Aiming at the problem of EEG signal feature extraction,this paper proposes a Common Frequency Pattern(CFP)algorithm,which can decompose the signal into spatial dimension and frequency dimension at the same time,and extract the spatial-frequency coupling feature which is able to maximize the signal energy difference in spatial distribution and frequency distribution,and effectively improve the performance of the classification algorithm.This research starts with the self-establishment of motor imagery EEG signal database,verifies the signal characteristics and summarizes the research literature of existing classification algorithms.According to the hypothesis that EEG signal frequency band is separable,a feature extraction algorithm based on CFP of motor imagery EEG signal is proposed.The algorithm decomposes the signal in the spatial dimension and frequency dimension,and extracts the spatial-frequency coupling features with a better performance compared to the traditional feature extraction methods.By summarizing the shortcomings of the proposed method,a frequency band selection method based on power spectrum estimation is proposed,which provides a supplementary selection of frequency division for CFP algorithm,and then this paper introduces a robust estimator algorithm to improve the robustness of proposed method.As a result,the average accuracy of the proposed method applied on public dataset for multi-channel classification is 92.24% which is 0.35% higher than the optimal comparison algorithm,84% for three-channel classification which is 2.65% higher than the optimal comparison algorithm.The average multi-channel classification accuracy of 84.60% and three-channel accuracy of 79.37% are achieved on the self-built two-class database,which is3.65% and 2.87% higher than the optimal comparison algorithm,re spectively.The average multi-channel classification accuracy of 63.61% and three-channel accuracy of 57.54% are achieved on the self-built three-class database,which are4.5% and 4.97% higher than the optimal comparison algorithm respectively.The performance advantage of the proposed method in multi-channel,three channel and single channel classification scenarios are verified under a large number of data samples.To sum up,this paper proposed a novel method called CFP algorithm to extract the spatial-frequency coupling features of motor imagery EEG signals,and by further modifying the proposed method,a good and comprehensive classification algorithm of BCI system is obtained.Moreover,the proposed method has a good classification performance in few channel application scenarios,which also provides an algorithm basis for the development and application of portable BCI systems.