基于Winger-Ville-CNN-LSTM的运动想象脑电信号识别

Recognition of Motor Imagery EEG Signals Based on Wigner-Ville-CNN-LSTM

摘要: 运动想象脑电信号(MI-EEG)识别是脑机接口系统研究的重要组成部分。为了更好地处理非平稳的脑电信号(EEG)并从中提取特征进行有效分类，本文提出了一种基于Winger-Ville分布(WVD)、二维卷积神经网络(CNN-2D)与长短时记忆网络(LSTM)的MI-EEG分类的模型(WVD-CNN-LSTM)。考虑到线性时频分析方法无法对EEG信号的瞬时功率谱密度进行准确描述，导致时频分析过程中EEG的部分重要非线性信息被丢失，本文使用WVD提取EEG数据的时频特征，构建具有高分辨率的信号时频表示结果。由于LSTM和CNN具有的强大特征提取能力以及对复杂数据的处理能力，为了能够利用两种网络的优势，本文采用将LSTM与二维CNN进行结合得到的CNN-2D + LSTM网络作为分类器。同时使用Softmax层进行分类决策与识别，最终输出MI-EEG的分类识别结果。本文采用来自于BCI竞赛IV公共数据集Data sets 2b进行实验分析，使用本文提出的方法对数据集中的左右手运动想象脑电数据集进行分类识别，并与现有方法进行了对比分析。实验结果表明，本文所提的方法具有较好的分类识别能力。与CSP方法、FBCSP-MIBIF方法和FBCSP-MIRSR方法相比，分类准确率分别提高了7.51%、5.07%和4.16%。

Abstract: Motor imagery electroencephalogram (MI-EEG) recognition constitutes a crucial component of brain-computer interface (BCI) systems. To better process non-stationary EEG signals and extract features for effective classification, this paper proposes a MI-EEG classification model based on the Wigner-Ville Distribution (WVD), two-dimensional convolutional neural network (CNN-2D), and long short-term memory (LSTM) network, termed WVD-CNN-LSTM. Given that linear time-frequency analysis methods cannot accurately characterize the instantaneous power spectral density of EEG signals, leading to the loss of significant nonlinear information during time-frequency analysis, this study employs WVD to extract time-frequency features from EEG data and construct high-resolution time-frequency representations. Leveraging the powerful feature extraction capabilities and complex data processing abilities of LSTM and CNN, and to exploit the advantages of both networks, a hybrid CNN-2D + LSTM network is adopted as the classifier. The Softmax layer is subsequently utilized for classification decision-making and recognition, ultimately yielding the MI-EEG classification results. Experimental analysis was conducted using the BCI Competition IV public dataset Data sets 2b, where the proposed method was applied to classify left-hand and right-hand motor imagery EEG data, and comparative analysis with existing methods was performed. Experimental results demonstrate that the proposed method achieves superior classification performance. Compared with the CSP, FBCSP-MIBIF, and FBCSP-MIRSR methods, the classification accuracy is improved by 7.51%, 5.07%, and 4.16%, respectively.