基于CSP与卷积神经网络算法的多类运动想象脑电信号分类

针对直接利用卷积神经网络(convolutional neural network,CNN)算法对多类运动想象脑电信号分类识别时,因样本量比较少,难以充分训练权值,导致分类效果较差的问题,结合一对多CSP算法与CNN算法对多类运动想象脑电信号进行特征提取与分类。首先,利用CSP算法对多类运动想象脑电信号进行特征提取,形成一维特征数据,作为CNN的输入样本;其次,对传统二维输入样本的CNN结构进行改造,使其适应一维数据的输入样本,对输入样本进行再次特征提取并分类;最后,使用BCI2005desc—Ⅲa的K3b数据进行算法验证;并对不同参数值的确定进行了讨论。算法验证结果表明,单独利用一对多CSP算法得到的分类正确率73%,单独使用CNN算法得到正确率为75%,新算法取得了91.46%的正确率,相比两种原始方法有较大提升。

Classification of EEG signals based on CSP and convolution neural network algorithm

For the classification and recognition of multi motor imagery EEG signals by using CNN algorithm, There is a problem of poor classification result is caused by the small sample size.In this paper, the combination of OVR-CSP algorithm and CNN algorithm were used for feature extraction and classification of multi class motor imagery EEG. Firstly, the CSP algorithm is used to extract the features of multi class EEG signals.The purpose is to reduce the data size of the input sample of CNN and increase the division between the EEG signals.Secondly, the traditional CNN network structure with the two-dimensional input data is modified to adapt to the input data of one-dimensional data.The input samples are extracted and classified again. Finally, the algorithm is verified by the K3b data set of BCI2005desc-IIIa data, and the determination of different parameters is discussed in the paper. Through the simulation, the classification accuracy rate is 73% by using one pair of multi CSP algorithm.the correct rate is 75% Using the CNN algorithm.In this paper, the correctness of the algorithm is 91.46%. Compared with the two original methods, the classification accuracy of this algorithm is greatly improved.