Application Research Of Evolutionary Optimization Algorithms In Motor Imagery Based Brain-Computer Interfaces

In motor imagery-based brain computer interface(BCI) systems, common spatial pattern(CSP) algorithm is widely used for extracting discriminative patterns from the EEG signals, and it is successful and effective. The performance of CSP algorithm depends largely on the subject-specific parameters, including the time interval, frequency band and number of electrode channels of EEG signals used for classification. Using subject-specific parameters for the extraction of invariant characteristics specific to each brain state can significantly improve the performance of CSP and classification accuracy of BCI. This thesis uses two evolutionary optimization methods, binary particle swarm optimization(BPSO) and backtracking search optimization(BSA), to select the user specific parameters solely or jointly so that the classification performance of BCI systems can be improved.In order to solve the problem of frequency band optimization, two optimization algorithms combined with CSP have been proposed in this paper. In the first algorithm, the broad band(8-30Hz) EEG was divided into 10 sub-bands with band width 4Hz overlapping 2Hz and BPSO was used to find the best sub-band set. CSP algorithm was applied for spatial filtering and feature extraction on these sub-bands. In the second algorithm, the frequency band width of EEG signal may vary within the range of 8-30 Hz, and the starting point and ending point of the frequency band were randomly selected by BSA. Similarly, CSP algorithm was used for spatial filtering and feature extraction on the selected band. Applied them to the same dataset, the results have shown that two new proposed methods combined with CSP both yield relative better classification accuracies compared with the traditional broad band(8-30Hz) CSP. Compared the two new algorithms, BSA algorithm performed better and it can converge faster than BPSO.Based on the optimization of the frequency band, this paper proposed a method using BSA algorithm for joint optimization frequency band and time segment. In this method, the band width of EEG signal was not fixed, the start frequency and end frequency could be varied within the 8-30 Hz, while the width of the time segment was fixed for 2s, the starting point may vary between 0-2s for the length of time window was 4s. The starting and ending points of frequency band and the starting point of time segment were jointly optimized by BSA. On the selected frequency band and time segment, spatial filtering and feature extraction using CSP. Compared with only optimization of frequency band by BSA, the results of joint optimization performed better.For the importance of channel selection in EEG classification, this paper proposed a method for BCI channel selection based on BSA. In channel optimization, the classification error rate was treated as the objective function, and the number of channels was determined by BSA. Two datasets were used for channel selection in the experiment. BSA method can achieve better classification results and use fewer channels than the case which all channels were used.