Abstract: To solve the problem of features extraction in Brain Computer Interface（BCI） , the paper presents a Sparse De-noising Auto-Encoder（SDAE） based on unsupervised learning theory. This method can learn features of brain electrical signal induced by stimulation and explore the deep features of the raw data. The SDAE, a Sparse Autoencoder（SAE） neural network by adding noise at the preprocessing, can enhance the generalization ability of learning and improve the robustness of the neural network. In the experiments, the multi-channel signals are reassembled firstly, and a sparse feature expression of raw data is built by using the SDAE. Then the Support Vector Machines（SVMs） classify the learned features. Finally, the classification accuracies are compared with those of optimal-single-channel method. The experimental results show that the classification accuracies of SDAE are superior to the optimal- single-channel method, so the SDAE can extract better features, improve the recognition accuracy of the “imitating reading” BCI, thus the method provides a new way of features extraction and classification on the BCI system.

Key words: imitating reading, Brain Computer Interface（BCI）, unsupervised learning, Sparse De-noising Auto-Encoder（SDAE）, Support Vector Machines（SVMs）

摘要： 针对脑-机接口的特征提取问题，提出了一种基于非监督学习的稀疏降噪自编码器，对刺激诱发的脑电信号进行自主学习，构建原始数据的深层特征表达。该编码器引用稀疏自编码神经网络，通过加入噪声，增强其学习的泛化能力，增加了神经网络的鲁棒性。首先对多导联信号进行重新拼接，输入稀疏降噪自编码器，得到原始数据的稀疏特征表达；然后，采用支持向量机将学习到的特征进行分类；最后，同直接使用最优单通道相对比。实验结果为：稀疏降噪自编码器的分类准确率要优于单通道，表明该方法能够更好地学习到特征，并提高了“模拟阅读”脑-机接口的识别正确率，为脑-机接口系统的特征提取和分类提供了新思路。

关键词: 模拟阅读, 脑-机接口, 非监督学习, 稀疏降噪自编码器, 支持向量机