Research On Classification Algorithms And BCI Based On The Left And Right Motor Imagery

Brain-Computer Interface (BCI) will interpret a person's thinking as series of external control commands, so that people can not rely on neuromuscular pathways to achieve the purpose of communication with the outside world. BCI researches in the field of rehabilitation engineering and in the aspect of theoretical value are of great significance. Although BCI researches have obtained many achievements, however, in the actual application of BCI system, there are still many problems to be solved, such as system's speed and accuracy, continuous online recognition and other issues. Therefore, the study of algorithms with high efficiency and the development of practical on-line BCI system are of great significance.Based on the characteristic ofμrhythm of the brain electrical signals form the right and left hand motor imagery, related feature extraction and pattern recognition algorithms were in-depth research, and some algorithms were made certain improvement to improve the classification accuracy, and the real-time BCI system with online identification function was eventually established. Specifically as follows:(1)The acquisition experiment of brain electrical signals of the left and right hand motor imagery was designed independently, and there were two main subjects who participated in this experiment. The collected data were used for subsequent offline analysis of feature extraction and pattern recognition algorithms.(2) It will greatly affect the accuracy of final classification whether the feature extraction algorithm used can well reflect the characteristic of EEG or not. In this thesis, the dynamic Independent Component Analysis (DICA) based on kurtosis maximization, Auto Regressive (AR) model, Common Spatial Pattern (CSP) and the second moments algorithm were used to extract feature of the collected EEG of the left and right hand motor imagery and they could achieve good results.(3)Pattern recognition algorithms were used to classify feature vector set extracted, its performance also had a great influence on the final classification. Genetic Algorithm(GA) was used to optimize weights and thresholds of some Artificial Neural Networks(ANN) and parameters of support vector machine (SVM),and the classification accuracy was improved by this. In addition, the posteriori probability support vector machine (PPSVM) was used to expend the training set by adding samples with great posterior probability value, thus the expanded training set could improve the accuracy of the identification of the test samples.1) Using visual C++ programming language and on the basis of SCAN 4.3's client software, an online identification system of the left and right hand motor imagery has been developed. DICA based on kurtosis maximization and the second moments of energy were respectively used to extract feature in this system. The time waveform of the mixing matrix which was generated by dynamic ICA algorithm based on kurtosis maximization can promptly and accurately reflect the energy changes ofμrhythm. Both online recognition rates could reach 90% and 95% respectively. The second moment energy algorithm was simple and easy to realize. The application of the dynamic ICA algorithm based on kurtosis maximization was innovative and provided a new idea for the BCI based on the left and right motor imagery.