基于不变性特征学习的多主体运动想象脑机接口研究

Besides the neuromuscular system, brain-computer interface (BCI) has provided a substitute way to convey the intention of the brain. For the patients who have lost controllability over the body, BCI technique can recognize the signals sent by the brain and send control order to the external assistive devices which can help the patients in their daily life. For healthy people, BCI technique can greatly improve the experience in multi-media and video games which have promising commercial value. Motor imagery based EEG recognition is a widely studied, and many BCI systems are built based on motor imagery analysis. The EEG signals are recorded when the subject is imaging specific movement, e.g. movement of the tongue, hand or foot. And the motor imagery is classified based on the EEG signals. Multi-subject motor imagery based BCI, which can recognize the motor imagery based on EEG collected from different subjects, is a new branch of BCI. Because of individual differences in anatomy, behaviors and habits, spontaneous low-frequency oscillations and reaction behaviors, the generalization ability among subjects for motor imagery based BCI is weak. To overcome this weakness, the features applied for multi-subject classification model should be robust to the variance of efficient data segments and efficient frequency bands among subjects. Therefore, the adequate features learning method is the key to success and extremely challenging. This project plans to leverage the convolutional neural network (CNN) model to study the invariant deep features learning method for multi-subject motor imagery based BCI. Specifically, the key research points are listed as follows. At first, the features learning model based on the convolutional auto-encoder is studied for motor imagery based EEG. In the second, based on the Difference measurement of feature map in CNN, the invariant features selection method is studied. And finally, combining the model transferring and sample transferring, multi-subject motor imagery recognition method is studied. This project aims to enrich the theory systems of feature extraction and learning method for motor imagery based EEG. With the implementation of this project, it hopes to provide a solid theoretical basis and feasible method guidance for invariant feature learning of multi-subject motor imagery based EEG recognition. Because the multi-subject BCI is capable for new subject without training data, it eases the application difficulty of motor imagery based BCI. In this way, it plays an important role in promoting the development and application of BCI.

多主体运动想象脑机接口能够基于不同主体脑电信号进行运动想象识别，是脑机接口研究的新方向。然而，个体在解剖学、行为习惯和脑自发低频振荡等方面的差异导致脑电信号的有效频带与有效信号段因人而异，极大影响了识别模型的主体间泛化能力。因此，如何提取对有效频带与有效信号段个体差异具有不变性的特征，是多主体运动想象脑机接口成功的关键和难点。本项目拟基于卷积网络结构，研究多主体运动想象脑机接口中的不变性特征学习方法，具体包括三方面内容：1）结合卷积自编码器，研究多主体运动想象脑电信号的特征学习模型；2）基于卷积网络中特征图间差异度量，研究不变性特征选择方法；3）结合模型迁移与样本迁移两种方式，研究多主体运动想象识别模型构建方法。本项目旨在丰富多主体脑电信号特征学习的理论体系和技术手段，研究成果可显著降低脑机接口的应用难度，使新主体即来即用，对脑机接口从实验室走向实际应用具有重要的推动作用和指导意义。

多主体运动想象脑机接口能够基于不同主体脑电信号进行运动想象识别，是脑机接口研究的新方向。然而，个体在解剖学、行为习惯和脑自发低频振荡等方面的差异导致脑电信号的有效频带与有效信号段因人而异，极大影响了识别模型的主体间泛化能力。因此，本项目主要研究了多主体运动想象识别中的脑电信号时频谱预处理方法、脑电信号数据增广方法、脑电信号特征学习方法。首先，提出了一种基于集成支持向量学习的运动想象识别算法，结合了基于事件相关去同步化、事件相关同步化的特征和基于事件相关电位的特征用于运动想象分类，具体地，使用支持向量机算法将样本点与决策边界之间的距离映射到后验概率，集成不同支持向量机分类器输出概率集作为最终预测概率。其次，提出了一种镜卷积神经网络模型，使用集成学习与数据扩增方法提高运动想象脑电图识别精度，训练阶段，基于源脑电通过互换左右侧脑电通道构造镜像脑电扩增训练样本；预测阶段，复制已训练源卷积网络作为镜像卷积网络，集成源卷积网络与镜像卷积网络输出的预测概率，形成最终类别预测。又提出了一种用于多主体运动想象脑机接口的孪生级联softmax卷积神经网络；为了减少个体差异的影响，应用主体识别softmax层和运动想象识别softmax层组成的级联softmax结构，同时完成主体识别和运动想象识别；为了提高运动想象分类精度，提出了一种基于集成学习的孪生网络结构。最后，提出了一种通道丢弃法用于提高运动想象识别卷积网络性能；具体地，训练阶段随机丢弃某脑电通道信号，以增加模型鲁棒性；在测试阶段，所有由通道丢弃处理的脑电都被输入模型，并应用它们的平均输出概率来确定预测结果。本项目的成功实施丰富了多主体脑电信号特征学习的理论体系和技术手段，研究成果可显著降低脑机接口的应用难度，对脑机接口从实验室走向实际应用具有重要的推动作用和指导意义。