For the last few years, many feature extraction methods have been proposed to process biological signals. Among these, brain signals have the advantage that they can be obtained even by people with peripheral nervous system damage. Motor imagery electroencephalograms (EEG) are inexpensive to measure, offer a high temporal resolution, and are intuitive. Therefore, these have been receiving a significant amount of attention in various fields, including signal processing, cognitive science, and medicine. The common spatial pattern (CSP) algorithm is a useful method for feature extraction from motor imagery EEG. However, performance degradation occurs in a small-sample setting (SSS) because CSP depends on sample-based covariance. Since the active frequency range is different for each subject, it is also inconvenient to set the frequency range to be different every time. This dissertation proposes the feature extraction method based on a multi frequency to solve these problems.
There are two proposed methods. The first method is based on the Subband Common Spatial Pattern (SBCSP) structure, and the second method is based on the Filter-Bank Common Spatial Pattern (FBCSP) structure. This dissertation combines Regularized Common Spatial Pattern (R-CSP) with SBCSP and FBCSP frameworks, respectively.
The procedure for the first method is as follows. 4?40 Hz band motor imagery EEG signals are divided into nine Subbands, and regularized CSP (R-CSP) is applied to individual Subbands. Fisher’s linear discriminant (FLD) is applied to the features of R-CSP extracted from individual Subbands, and the results obtained through the foregoing are connected for all Subbands to make an FLD score vector. Furthermore, principal component analysis (PCA) is applied to use the FLD score vectors as the inputs of the classifier least square support vector machine.
The procedure for the second method is as follows. Motor imagery EEG is divided by a using filter bank. The R-CSP is applied to the divided EEG. The features are selected based to mutual information based on the individual feature (MIBIF) algorithm. The parameter sets is selected for the ensemble. The features are classified based on selected emsemble method.
The BCI competition III dataset IVa is used to evaluate the performance of the proposed methods. The proposed methods have better classification performance than the previous methods in classifying the motor imagery EEG. In particular, the methods showed better performance in SSS situations. The proposed methods provide guidelines to overcome the difficulties of frequency range selection and performance degradation in SSS situations simultaneously.