基于马尔科夫切换过程的运动想象信号分类

隐马尔科夫模型（HMM）在脑机接口（BCI）领域中已经得到很好的应用，尤其是在运动想象（MI）信号的分类中。但是，很多传统的方法只是利用隐马尔科夫模型描述信号的动态特性，再根据观测数据求得模型参数，然后进行信号分类。由于脑电信号低信噪比、高维数和状态复杂的特点，在研究中先用分层Dirichlet 过程（HDP）描述MI信号，利用HDP自聚类特性，然后使用AR模型描述MI信号的时间特性，最后结合马尔科夫切换过程（MSP）描述MI信号的动态特性，以此来充分地描述MI信号。随后对实验室采集的数据和2003年BCI国际大赛的部分数据，使用HDP AR HMM模型对MI信号分类，获得很好的分类效果，准确率分别是9900%、9200%和7246%。实验结果表明，所提出的方法可以取得更好的运动想象信号分类。

Motor Imagery Signal Classification Based on HDP AR HMM

Hidden Markov model (HMM) is well applied in brain computer interface, especially in the classification of motor imagery(MI) electroencephalogram (EEG) signal. Conventional methods  use HMM to model EEG signal, then use the observed signal under controlled state to estimate the HMM parameters and finally classify the EEG signal through the trained HMMs. However, due to the characteristics of low signal to noise ratio(SNR), high dimensionality and complexity of motor imagery EEG signal, HMM cannot fully describe the dynamic property of motor imagery EEG signal. In this paper, we use hierarchical Dirichlet process (HDP) with self clustering ability to describe MI signals and then use AR/VAR model to highlight the time property of MI signal. Finally we combine them with Markov switching processing (MSP) so that we can get more information of MI signal. In order to verify this method, we tested the algorithm on our in house data and some of the 2003 BCI international competition data sets. High accuracy on classification of MI is obtained.