Abstarct: This thesis investigates remove noise and feature extraction as two important fields in signal pro-cessing. First preprocessing is to remove noise from signal for classification it. In this thesis, we study two classical methods and two proposed methods for removing noise. First, noise is removed form signals using classical ICA method, wavelet transform, two proposed methods (Walsh transform and combination method Walsh-ICA). To have a good evaluation, the result of these four meth-ods is evaluated by attention to Mean Square Error (MSE), Percentage Root Mean Square Dif-ference (PRD) and Signal to Noise Ratio (SNR) measure. The result of evaluation used these measures; confirmed Walsh transform and combination method Walsh-ICA had the smallest MSE while having the largest SNR and PRD. Then feature extraction from signal and classification them would be studied. The number of ex-tracted feature is few so feature vector have 22 elements. These features correspond to Walsh transform entropy of channels signal, Walsh transform entropy of whole of signal, Walsh trans-form power of channels signal and Walsh transform power of whole of signal. For evaluating performance of these features, these feature would be extracted using wavelet and Fourier trans-form, too. Also classification baxse on extracted feature from these three methods separately. Then signals are classified using support vector machine (SVM) classification and KNN baxse on extracted feature. The result show that classification using extracted feature of Walsh transform is better than classification using other two transforms. Recognition rate is 42.5 by proposed method and SVM, is 39.0 by KNN method. For another comparing, the result is compared with the obtained result from 4th composition BCI. The result show that classification methods using Walsh transform is the best methods expect one participant’s methods. But advantage of pro-posed method is time complexity. Proposed method consume a few time for testing and training this time is 52 second, that is better than fist method which consume 403 and 640 second.