FEATURE EXTRACTION OF MOTOR IMAGERY EEG SIGNALS BASED ON MULTI-SCALERECURRENCEPLOT AND SDA

When recognizing multi-class motor imagery electoencephalogram (EEG) signals directly using stacked denoising autoencoders (SDA), it is difficult to fully train the weights due to the small sample size, which results in poor classification effect. To overcome this problem, the multi-scale recurrence plot and SDA method are combined to extract features of multi-class motor imagery EEG signals for recognition. Firstly, multi-class motor imagery EEG signals are decomposed into a series of intrinsic mode functions (IMFs) with different scale by synchrosqueezed wavelet transform, and the recurrence plot of each IMF is constructed to form one-level feature data as input samples of SDA. Then, high-level abstract features which can better express category attributes are extracted from multi-scale recurrence plot by SDA. Finally, EEG signals are classified by using Softmax classifier. Four types of motor imagery EEG data of Datasets 2a in BCI Competition IV are used to verify the proposed method. The average classification accuracy is 0.89, which shows that the proposed method has good effectiveness and robustness.