Abstract: In order to reduce the boring and time-consuming training process and improve the classification accuracy of brain-computer interface system, this paper proposes a self-training algorithm with segmented overlapping common spatial pattern algorithm which applies a semi-supervised learning to motor imagination classification. The new algorithm uses segmented overlapping common spatial pattern as feature extraction method and learns from a small pool of labeled data. In addition, a confidence criterion is proposed to select the most informative samples from the unlabeled samples to improve the classifier. With the help of small labeled data and large unlabeled data, the proposed algorithm achieves better classification performance than self-training algorithm with common spatial pattern and self-training algorithm with filter bank common spatial pattern. The dataset Iva of BCI competition III has been used to demonstrate the validity of the method, and the results show the proposed algorithm can effectively improve the classification accuracy of motor imagination.

Key words: Brain Computer Interface（BCI）, self-training, Segmented overlapping Common Spatial Pattern（SCSP）, confidence criterion

摘要： 为了减少枯燥和耗时的训练进程和提高脑机接口系统的分类率，将半监督学习运用到了运动想象脑电的分类中，提出了一种基于分段重叠共空间模式的自训练算法，将分段重叠共空间模式作为特征提取算法，使用少量标记的数据进行学习，然后使用置信度评估准则从未标记样本中挑选信息量大的样本来提高线性判别分类器的性能。提出的算法在少量标记样本和大量未标记样本的帮助下，能够获得比基于共空间模式作为特征提取的自训练算法和基于滤波带宽共空间模式作为特征提取的自训练算法有更好的分类效果。使用2005 BCI竞赛的数据集Iva来证明算法的有效性，结果表明了提出的算法能有效提高运动想象脑电的分类率。

关键词: 脑机接口, 自训练, 分段重叠共空间模式, 置信度评估准则