Study On Classification Algorithms In Multiclass Motor Imagery Brain-Computer Interfaces Based On Combined Selection Of Time Segment And Frequency Band

Brain-computer interface (BCI) is a direct communication system between the brain and external devices that translates electrical brain signals into commands via computer hardware and software, without using traditional pathway consisting of peripheral nerves and muscle. It is known that the electrical signals of human brain activity are highly variable inter-subject and inter-trials even for the same subject, which makes the practical application of BCI technology still far from the real life. In this context, it is necessary to develop a completely automatic system for EEG signal classification. One of the most complex problems to address in the system design lies in the selection of proper time and frequency intervals in which the EEG signals are filtered before feature extraction. So in this paper we made an intensive study of the problem. We proposed two techniques for the selection of time and frequency intervals of EEG signals to extract discriminative features in BCI application.The first kind of technique combined wavelet packet (WP) decomposition and common spatial pattern (CSP) to select time and frequency intervals. The raw EEG signals were band pass filtered between8and30Hz, then the filtered signals were subject to WP decomposition and reconstruction, and finally the reconstructed signals were spatially filtered by CSP algorithm. The innovation was that an adaptive method was used to select the best subspace coefficients for signal reconstruction, which were the optimal frequency subbands for each specific subject. Three different methods were applied to six datasets recorded during BCI experiments based on motor imagery. The results showed this technique had a superior classification performance, thus verifying the feasibility and validity of the algorithm.The second kind of technique used sliding window to analyse multiclass EEG signals, automatically select the optimal time and frequency intervals, and thus the most informative time and frequency component for individual subject was extracted, leading to best classification result, time segment and frequency band. It was easy to deal with this component by using CSP spatial filter for feature extraction and KNN classifier for classification. The algorithm was respectively applied to nine three-task and four-task datasets from Graz dataset2a of BCI competition IV, and then the best classification result, best time segment and best frequency band were obtained. Two criteria were utilized for evaluate these algorithms. One is cross validation used for single sessions by randomly selecting a part of data for training and the rest data validation. The other is session to session transfer used for two different sessions, one session used for training, and the other for testing. All those classification results not only confirmed the inter-subject and inter-trial variability, but also indicated the sliding window can automatically and adaptively select the proper time segment and frequency band for each subject during different motor imagery task, which played an important part in improving their classification accuracy rate.