基于半监督学习的脑电信号特征提取及识别

针对有监督学习容易造成未标记样本的浪费和手动特征提取容易导致信息丢失的问题，提出一种基于深层堆叠网络（DSN）的半监督特征学习方法，无监督特征学习的过程由多个受限玻尔兹曼机（RBM）的并行训练完成，将训练得到的参数用于DSN的输入权值初始化，再采用批量模式的梯度下降法进行监督微调。将所提方法用于运动想象脑电信号特征提取及识别，实验结果表明本文方法能够充分利用未标记样本中的隐含信息，有效提取脑电信号特征，识别结果优于共同空间模式（CSP）和深度信念网络（DBN）等算法，该方法可用于提高BCI系统中脑电信号的识别准确率。

Feature extraction and recognition for EEG signals based on semi-supervised learning

A semi-supervised feature learning method based on deep stacking network (DSN) is proposed aiming at resolving the waste of unlabeled samples in supervised learning and the information loss caused by artificial feature extraction. The unsupervised feature learning process is completed by the parallel training of multiple restricted Boltzmann machines (RBMs) and the trained parameters are utilized to initialize the input weights of the deep stacking network. Then the batch-mode gradient descent algorithm is utilized for supervised fine-tuning. The proposed method is applied to classify the motor imagery EEG signals. Experimental results demonstrate that the proposed method can take full advantage of the implicit information in unlabeled samples and efficiently extract EEG features. The average recognition accuracy is superior to other algorithms like common spatial pattern (CSP) and deep belief network (DBN). The proposed method can be applied to improve the recognition accuracy of EEG signals in a BCI system.