Rearch On Pattern Recognition Algorithms For Motor-Imagery-Based Brain-Computer Interface

A brain-computer interface (BCI) is a communication system. The messages or commands in the system sends from an individual to the external world not passing through the brain's normal output pathways of nerves system and muscles. The BCI has very important application in medical, artificial intelligence, novel entertainment and so on. Its implement refers to many subjects and frontier technologies. So the study about BCI has both science and application value. It has attracted much interest and been well developed. Signal processing is the key issue for implementing a BCI system as well as improving its reliability and performance, whose essence is the signal pattern recognition. It mainly includes three parts: signal preprocess, feature extraction and classification. The study emphasis of this paper is concentrated on the last two parts and is proved of improving the recognition accuracy of BCI system. Focusing on the main problem of BCI signal processing methods during the development, such as how to enhance classification accuracies and robustness, reduce training time and develop new effective methods for the pattern recognition in asynchronous modes, and so on, our research results and main contribution include:1. Focusing on the EEG characteristic, such as fully noisy, complicated data distribution, high dimensions, relative-un-abound training samples, and so on, two algorithms for the extraction of movement related potentials (MRPs) have been proposed: neighborhood spatial pattern algorithm and adapting spatial pattern algorithm. Based on the idea of manifold learning, the former one constructs the target function to estimate the optimal map matrix, only utilizing the neighborhood and label information, while not depending on the underlying data distribution. The latter estimates the similarity between samples by the corresponding features'distance, and then constructs the target function based on this similar relationship. It adaptively determines the neighbor relationship and the optimal direction via iterations. Its implement also doesn't need any assumption about the underlying data distribution, but proposes an alternative for the similar measurement of EEG data with high dimension. Experimental results show that: the two algorithms all strengthen MRPs feature's robustness, and improve classification accuracy.2. Event-related desynchronization/synchronization (ERD/ERS) and MRPs are two of the most important neurophysiological backgrounds of motor imagine utilized in the feature extraction of BCI pattern recognition. They appear before/after or during the process of motor imagery. Their characteristics are distinct at special temporal area, electrodes and band of frequency, i.e., they are sensitive to time, space and frequency. Based on these, we proposed general temporal-spatial extraction (GTSE) algorithm which can optimize temporal and spatial discriminative information together. The characteristics of ERD/ERS and MRPs are different, so the corresponding GTSE algorithms are different, too. Experimental results show that: the algorithms can catch more essential discriminative information of ERD/ERS and MRPs, and efficiently improve classification accuracy.3. Focusing on reducing training time to meet the requirement in the development of BCI, two semi-supervised learning algorithms for BCI are proposed. One is a learning algorithm based on combining-features. It chooses the most confident unlabeled samples (with their predicted labels) to enlarge the training set via iterations, and then based on the enlarged training set, retrains the parameters of the feature extractor and classifier. And consequently it gains its ends of using the information of large number of unlabeled samples with a few labeled samples to train feature extractor and classifier. The other is a semi-supervised feature extraction algorithm for MRPs. It straightly utilizes the information of labeled and unlabeled samples together to optimize the parameters of feature extraction model. Experimental results show that: based on a few labeled samples, the two semi-supervised learning algorithms can also obtain favorable classification accuracy.4. The pattern recognition of asynchronization BCI system recently is a hot topic. One of its biggest challenges is to discriminate which state the user is during experiment processing: imagining or idle? Because of the diversity of idle state, there are no effective training samples for idle sate. A novel algorithm is proposed for the detection of the idle state. Using the classification accuracy and the within-class scatter of the samples classified correctly in the training set (only with motor imagine samples) as two indexes, the algorithm constructs the target function according to the chosen criterion for the optimal operating point of the receiver operating characteristic curve, and determines the minimal/maximal decision thresholds for classification. To further optimize the classification effect, the algorithm treats the prediction labels by a fuzzy way. Its efficiency was demonstrated by experiment.To sum up, in this paper, we studied feature extraction and classification algorithms for BCI, focusing on the processing challenges including intrinsic nodus for EEG as biomedicine signal and the topics come forth from further development and practical application of BCI technique. The efficiency of the algorithms was all demonstrated by related experiments.