Motor Imagery Multi-classification based on Canonical Correlation Analysis Feature Fusion

• Published in: Chinese Control Conference pp. 3334 - 3339
• Main Authors: Yang, Zhuoyuan, Guo, Sitong, Hong, Yunqi
• Format: Conference Proceeding
• Language: English
• Published: Technical Committee on Control Theory, Chinese Association of Automation 26.07.2021
• Subjects: Brain modeling; Brain-Computer Interface (BCI); Canonical Correlation Analysis (CCA); Common spatial Patterns (CSP); Correlation; Discrete Wavelet Transform (DWT); Electroencephalography; Feature extraction; Feature Fusion; Time-frequency analysis; Training; Wavelet analysis
• ISSN: 1934-1768
• DOI: 10.23919/CCC52363.2021.9549799

The application of discrete wavelet transform (DWT) or common space transform (CSP) extracts the features of motor imaginary (MI) electroencephalogram (EEG) signals. However, these feature extraction algorithms suffer from some drawbacks, such as the low generalization ability of CSP dealing with different subjects and the dependence of time-frequency window of DWT. In this paper, canonial correlation analysis (CCA) is proposed to fuse two features extracted by DWT and CSP so as to eliminate the noise with low correlation. CCA retains not only the advantage of CSP requiring less data in the training set, but also the advantage of DWT having high adaptability to multi-classification problems and nonlinear signals. The experimental results on BCI2005desc-IIIa and BCI2008desc-IIa EEG dataset show that the average accuracy of proposed multi-domain feature fusion model based on CCA is more than 87.50% for four-class MI tasks on BCI2005desc-IIIa data set, which is higher than non-feature-fusion methods, and is 78.07% for four-class MI tasks on BCI2008desc-IIa data set, which is higher than other feature extraction methods.