Motor Imagery based EEG Signal Classification

Abstract

Effective classification of Motor Imagery (MI) tasks based EEG signals is the main hurdle in order to develop online Brain Computer interface (BCI) system. The key part of BCI system is to extract the dominant features from EEG data along with selection of a suitable classifier. In this research thesis, a relatively new approach has been implemented to accurately classify EEG signals that have been extracted from MI. The data-set was obtained from BCI competition-II 2003 named Graz database. Two channels have been selected for preprocessing i.e. C3 and C4. After applying pre-processing techniques feature vector have been extracted. The feature vector consists of bior Wavelet Transform (WT) coefficients, Power Spectral Density (PSD) approximations, average power and aggregated EEG signal. In this study, we have presented a comparison of mostly used classification algorithm with relatively new classification technique i.e. Self-organizing maps (SOM) and Deep Belief Nets (DBN). It has been depicted from measured data that SOM shows a classification of 84.17% on Principal Component Analysis (PCA) implemented reduce dataset. Furthermore, a 2% increase in classification accuracy has been attained by using bi-orthogonal filter banks for wavelet transform instead of Daubechies WT. In Deep Learning, Firstly a weak classifier has been trained using deep belief networks (DBN) after that the concept of boosting has been applied in order to make the classifier strong. The boosting algorithm that has been implemented in this research is ada boosting. Multilayered structure has been used for DBN consisting of hidden units and hidden layers. Furthermore, the performance has also been tested using different hidden units and hidden layers. The experimental results shows that with different hidden layer there is a significant change in classification results but overall performance is better for 15 hidden layers network. The results are compared with different state of the art classification algorithms i.e. Support Vector Machine (SVM) and Self organizing maps (SOM) based classification techniques and DBN shows better results with a recognition error of minimum 6% in the classification performance.