基于张量网络的多脑运动想象脑电信号分类

Multi-brain motor imagery EEG signal classification based on TensorNet

为了解决脑机接口中识别率低、稳定性差等问题,提出一种基于张量网络的多脑脑机接口解码方法。首先,用共空间模式提取脑电特征,融合多脑信息,将融合的数据输入张量网络以便捕获时序特征;然后,张量网络把参数表示成高维数组,运用张量分解方法分解神经网络的权重张量,在不影响识别准确率的前提下,削减了神经网络的参数数量。实验结果表明,与单个被试相比,多脑运动想象脑电数据的识别准确率提高了17.2%;与GRU网络相比,张量网络模型的识别准确率有所提高,同时网络中的参数数量更少。

In order to solve the problems of low recognition rate and poor stability in brain-computer interfaces,we propose a multi-brain-computer interface decoding method based on tensor network.Firstly,it uses the common spatial patterns to extract EEG features,fuses multi-brain information and inputs the fused data into a tensor network to capture time series features.Secondly,the tensor network expresses the parameters as high-dimensional arrays,and uses tensor decomposition to decompose the neural network weight tensor,reduces the number of parameters of the neural network without affecting the classification accuracy.Experimental results show that the performance of multi-brain motor imaging EEG data classification is improved by 17.2%compared with the accuracy of a single subject.Compared with the GRU network,the recognition accuracy of the tensor network model is improved,and the number of parameters in the network is reduced.