A Motor Imagery EEG Feature Extraction Method Based on Energy Principal Component Analysis and Deep Belief Networks

Abstract

The motor imagery electroencephalography (MI-EEG) reflects the subjective motor intention, which has received increasing attention in rehabilitation. How to extract the features of MI-EEG accurately and quickly is the key to its successful application. Based on the analysis and comparison of the existing feature extraction algorithms, a feature extraction method based on principal component analysis (PCA) and deep belief networks (DBN) is proposed, namely PCA-DBN. Firstly, the second-order moment is used to analyze the time-domain of MI-EEG, select the effective time interval. Secondly, PCA is used to analyze the selected time-domain interval and obtain the principal component feature points. Then, feature points are imported into DBN to realize the final feature extraction. Finally, use the softmax classifier to complete task classification. Perform algorithm validation on the BCI Competition II Data set III and BCI Competition IV Data sets 2b, classification accuracies are 96.25% and 91.71%, kappa values are 0.925 and 0.8342. The paired-sample t-test with FDR correction is carried out on the verification results, and the comparison with some better classification algorithms shows that the algorithm has better performance. In the end, this method is used to extract the features of laboratory data, the optimal classification accuracy is 97.69% and kappa value is 0.9538, the validity of the method is further verified.

Keywords

• Deep belief networks
• motor imagery electroencephalogram
• principal component analysis
• second-order moment
• softmax classifier