Brain - Computer Interface P300 Recognition Algorithm

Brain-computer interface (BCI) is a kind of human-computer interface for direct communication between brain and computers or other electric devices, by which people can control electric devices independent of normal peripheral nerve systems and muscle output channels.Many studies on neurology science show that brain can produce special electrical activities (P300 potentials) when the subject is presented with task-relevant stimuli. These electrical activities can be recorded as brain signals by Electroencephalograph. And then, we can extract useful information from these electrical signals by special algorithms. In this way, the thought that the subject wants to express can be inferred easily. Finally, as long as translating the thoughts into device control command, the act of controlling other devices will be realized. Hence, the key problem of BCI system is how to translate P300 signals recorded by Electroencephalograph into device control commands quickly and accurately. However, P300 potential may be hard to be detected within only one repetition, due to the characteristic of variability and low signal-to-noise ratio. Several repetitions are needed to infer the desired information correctly. Hence, the criterions to assess the performance of P300 detection are as follows: the needed number of repetitions when all the information are inferred correctly, the detection accuracy of single repetition, the training speed of the optimal P300 classifier and the detection speed of P300 potentials.Support vector machine (SVM) is a new technique for data classification and pattern recognition. Its basic idea is to map data into a high dimensional space and find a separating hyperplane with the maximal margin. Support vector machine has the good generalization in small samples, nonlinearity and high dimension space. Hence, support vector machine has been successfully applied in the area of EEG recognition. However, the traditional SVM-based P300 detection methods still have the shortcoming of low detection accuracy of single repetition, low training speed of optimal classifier and too many repetitions are needed to infer all the information correctly.To solve these shortcomings of traditional algorithms, we propose some P300 detection algorithms that combine SVM algorithm with F-score method and wavelet decomposition theory. The main contributions of this thesis are given below:(1) To improve the training speed of optimal P300 classifier, a new training selection method is proposed in this thesis. By this new training selection method, we can divide the primal training set into a training set and a validation set. And with this validation set, the optimal parameter of the support vector machine classifiers can be predicted more quickly. The experiment results showed that the parameter estimation speed of our method was one times higher than that of the five-fold cross-validation.(2) To infer all the information correctly with fewer repetitions, we present a new P300 detection algorithm based on F-score feature selection and support vector machine. Using our F-score feature selection method, we can eliminate large number of redundant features. Hence, we can enhance the detection accuracy of SVM-based P300 classifiers. This algorithm was tested with a P300 dataset from Wadsworth Center. And the results showed that this algorithm achieved an accuracy of 100% in P300 detection within only four repetitions.(3) Since the optimal threshold selection procedure of F-score feature selection method is very time-consuming, hence it may slow down the training speed of optimal P300 classifier badly. In order to solve this problem in the case of the P300 detection accuracy is not down, we propose a new P300 extraction algorithm, which is based on wavelet decomposition theory, to improve the training speed of P300 classifier. Since wavelet decomposition method has a good performance in the terms of feature extraction. Hence, by this method we can effectively extract features of P300 potentials and reduce the dimension of input space, and do not need the time-consuming optimal threshold selection procedure.(4) In order to enhance the detection accuracy of single repetition, we present an optimal channels selection algorithm based on F-score method. Using this algorithm, we can effectively eliminate task-irrelevant EEG channels and obtain the optimal task-relevant EEG channels. The experiment results showed that this algorithm achieved a detection accuracy of 83.87% with a single repetition.Finally, we make a conclusion and propose the future research directions in this field.