The Study Of Motor-imagery-based Brain-computer Interface Based On Semi-supervised Learning

Brain-computer Interface(BCI) can directly build communication and control between the brain and computer or other electronic equipments without the participation of normal output pathway of the brain(peripheral nerves and muscles). BCI technique can provide new informative communication channels for paralyzed patients and improve the living quality of the patients, and it has great practical value in the medical field, cognitive science, psychology, military, entertainment and wearable intelligent equipment. In the classification of the Electroencephalogram(EEG), the traditional supervised learning needs to collect a large number of labeled EEG data for training classifier, but the costs of obtaining these labeled EEG data are expensive, which needs much time and effort to conduct many trials, and cause subjects’ fatigue easily, thus hinder the development of BCI system. In addition, as time changes, the state of EEG will also change, which makes the difficult of the classification rises further. However, the unlabeled EEG data is easy to get, there will be a lot of unlabeled EEG data wasted if they are not used, therefore, the requirement using the unlabeled EEG data is strong increasingly. Although unsupervised learning uses unlabelled EEG data for training classifier, it is quite easy to cause the decrease of the generalization ability of the model due to the lack information of labeled EEG data. Therefore, it makes sense to apply semi-supervised learning into the classification of EEG data, semi-supervised learning only need to acquire a small amount of labeled EEG data for training classifier, and then use a large number of unlabelled EEG data to train the classifier in auxiliary, thus not only the time used for collecting labeled samples is reduced and the performance of the classification is improved, but also semi-supervised learning itself is an adaptive process, which is helpful for the enhancement of BCI adaptability.For the subjects, the collection of motor imagery EEG makes them feel bored and tired much than other EEG(SSVEP, P300, etc.) because of the longer time for each of motor imagery and the high repeatability. Therefore, in the thesis, the research of motor-imagery-based BCI based on semi-supervised learning has more value. At present, semi-supervised learning is still in its initial stage and has not establish own system, so there are many unsolved problems, for examples, in the semi-supervised learning, how to remove the noise and improve signal-to-noise ratio, how to accurately select parameter, how to build co-training algorithm under the more realistic condition, how to fuse other machine learning methods into semi-supervised learning for further improving classification performance, how to provide the model and framework for the development and application of on-line system. The thesis applies semi-supervised learning into the classification of motor imagery EEG with high user agreement. Aiming at these problems of semi-supervised learning, this thesis puts forward the corresponding solutions. The main research of the thesis includes the following aspects:(1) A self-training based on model selection(STBMS) is proposed, it uses mutual information to select the best parameter pair from a range of parameter combination based on support vector machine(SVM), which solve the problem that small samples can not use cross-validation to select appropriate parameter in self-training of motor-imagery-based BCI. A confidence assessment criteria is propose to choose the samples with high confidence from unlabeled samples and add these samples to training set for retraining and further improving the classification performance of self-training and signal-to-noise ratio. Two evaluation standards are proposed to evaluate feature robustness to the noise and the convergence of the algorithm. The experimental results show the effectiveness of the model selection method and the confidence assessment criteria in improving accuracy. Comparing with the standard SVM self-training, that the re-extraction of features can improve feature robustness to the noise is illustrated. The experimental results also show the effectiveness of feature robustness to the noise and the convergence of algorithm.(2) A co-training based on Modified FLDA(Fisher Linear Discriminant Analysis)(CTBMFLDA) is proposed under the more realistic condition and apply it into the classification of motor-imagery-based BCI. Two classifiers with much difference based on the small sample size are constructed, and the two classifiers respectively choose high confidence samples for the other in order to update itself. Three parameters are proposed to respectively evaluate feature robustness to the noise in the STBFLDA and CTBMFLDA. The experimental results prove that the accuracy obtained by CTBMFLDA is better than self-training based on FLDA(STBFLDA), and also prove that the effective of three parameters evaluating the feature robustness to the noise.(3) Studying three active learning methods( ALNACD, ALSVMactive and ALEBS)used for multi-classification of motor-imagery-based BCI, and exploring sample selection strategy of the three active learning methods: Nearest Average-class Distance(NACD), SVM active learning(SVMactive) and Entropy-based Sampling(EBS). A semi-supervised learning combining with active learning(based on NACD)(SSLCAL) is proposed to combine active learning with semi-supervised learning for solving the problem of multi-classification of motor-imagery-based BCI. The experimental results show that the validity of the three active learning methods and SSLCAL in improving accuracy, and they can use less labeled samples to improve the classification performance of the algorithms.(4) A novel feature extraction method: Segmented Common Spatial Pattern(SCSP), is proposed. By using SCSP as feature extraction method, a batch-mode sequential updating self-training with SCSP(BMSUST-SCSP) is proposed to provide the model and framework for development and application of on-line BCI. The experimental results show that multiple iterations can remove noise and improve signal-to-noise ratio after the arriving of each subset, and that the feature extracted by SCSP algorithm has higher reliability and stronger robustness to the noise than CSP algorithm, finally that the SCSP feature robustness to the noise is on the rise with the increasing of subset in the help of the mutual information.(5) Based on Neuroscan signal acquisition system and the characteristic of motor imagery EEG, EEG acquisition experiment imagining left-right hand movement is designed. Depending on motor imagery EEG of five subjects with independent acquisition, eight different semi-supervised algorithms are used for the classification of the EEG. The experimental results show the advantage of the STBMS, CTBMFLDA, SSLCAL and BMSUST-SCSP proposed in the third chapter to the sixth chapter in improving accuracy, and analyze the classification performance of each algorithm and the factors influencing the classification results.