Motor imagery signal classification based on transfer learning

Abstract

The EEG of motor imagery varies greatly according to different subjects and the same subject in different time periods. Traditional machine learning methods can only solve the classification and recognition of the same individual within a short period of time, and the classification and recognition effect also depends on the difference of data sets, with strong individual differences. Many classification methods are unstable and have poor universality. Transfer learning can use knowledge from similar data to enhance the learning process, and use knowledge in related fields to help complete the learning tasks in the target field, so as to change the traditional learning from scratch into accumulated learning and improve learning efficiency. In this paper, the power spectrum characteristics of 8 channels signals related to motor imagery at 7-29hz were extracted, and the motor imagery data were classified and modeled by transfer learning algorithm. Meanwhile, compared with the other two existing classification methods PSD (Power Spectral Density) and CSP (Common Spatial Pattern), the analysis results showed that the classification accuracy of transfer learning (90.9 ± 2.2) was higher than that of traditional PSD+LDA(62.5±11.6) and CSP+SVM (71.3±3.5), which verified the feasibility of transfer learning in motor imagery BCI classification and recognition.