The Brain-Computer Interface (BCI) system provides a communication pathway between the brain and computers, and it has recently gained significant attention. One of the most common BCI systems is motor imagery. Currently, for classifying motor imagery-based EEG signals, Common Spatial Patterns (CSP) have been widely used. EEG signals based on motor imagery are often contaminated with noise and artifacts, which significantly reduce the performance of BCI systems. The CSP algorithm has shown promising performance for classifying various types of motor imagery data. However, when the number of trials is limited or the data is noisy, overfitting is likely to occur, hindering the extraction of suitable spatial filters. Another limitation of CSP is that it only extracts spatial-based filters. Furthermore, CSP uses a specific frequency band for filter extraction, which may not be suitable for different individuals. Therefore, a combination of filter banks and regularized CSP methods has been utilized to overcome CSP's sensitivity to noise and overfitting. The CSSP algorithm has also been used to extract both spatial and frequency-based filters simultaneously. In this research, three sections are dedicated to the examination of results. The first section pertains to dataset 2a from the fourth BCI Competition. The selection of suitable frequency bands and time windows for this dataset is crucial, and using a cross-validation approach, the Frequency-Based CSP (FBCSP) method has chosen appropriate frequency bands. Additionally, an optimal time window for each subject's training data has been determined. After preprocessing, the filter bank and the Common Spatial-Spectral Patterns (CSSP) method are combined, and a majority voting approach is employed for classification. The prediction results of six classifiers (SVM-Linear, SVM-RBF, LDA, Naive Bayes, KNN, and PNN) for various parameters are combined, and their average prediction is eva‎luated as the final result. The proposed algorithm achieved an average accuracy of 71.72%. The second section focuses on dataset 2b from the fourth BCI Competition. In this dataset, a combination of the selected filter bank in the frequency range of 8-30 Hz and the majority voting approach based on combining predictions from two SVM-Linear classifiers is used. This proposed method ultimately achieved an average accuracy of 80.96%. Finally, the third section is dedicated to dataset IVa from the third BCI Competition. In this case, a combination of the filter bank and CSSP is used to extract suitable features, and the Extreme Learning Machine (ELM) classifier is employed for classification. The eva‎luation results demonstrated that the proposed method achieved an average accuracy of 85.24%.

مريم ذكري

احسان روحاني , امير اخوان بي تقصير