基于CWGAN?ABiLSTM?FCN的運動想象腦電信號分類

摘" 要： 針對運動想象腦電信號（MI?EEG）樣本數據分布不平衡、時序特征提取時對長距離的依賴和關注度不均衡、局部特征提取難導致的基于MI?EEG的運動意圖識別實時性差、精度低的問題，提出一種融合改進的雙向長短時記憶神經網絡（BiLSTM）和全卷積神經網絡（FCN）的MI?EEG信號分類方法。首先，該方法利用條件生成對抗網絡產生虛假的MI?EEG信號樣本，實現訓練樣本集的有效擴充，解決了數據集過少且各類別數量不平衡的問題；其次，利用雙向自注意力長短時記憶神經網絡和全卷積神經網絡的各自優勢，避免了時序特征提取時對長距離的依賴和關注度不均衡、局部特征提取難以及無法兼顧MI?EEG信號的時?空域特征的問題；在此基礎上，構建融合特征與動作分類標簽間的非線性映射關系，從而提高模型的識別精度。最終將此分類模型與其他的MI?EEG分類模型在測試數據集進行了對比實驗。研究成果表明，該MI?EEG識別模型準確度達到了97%，顯示出較強的泛化能力。

關鍵詞： 運動想象； 腦電信號分類； 生成對抗網絡； 長短時記憶網絡； 全卷積神經網絡； 注意力機制

中圖分類號： TN911?34； TP391" " " " " " " " " " "文獻標識碼： A" " " " " " " " " " 文章編號： 1004?373X（2025）07?0057?08

Motion imaging EEG classification based on CWGAN?ABiLSTM?FCN

WU Shengbiao1， 2， CHENG Xianpeng1， LI Huaning1

（1. School of Mechanical and Electronic Engineering， East China University of Technology， Nanchang 330013， China;

2. Jiangxi Rehabilitation Aids Technology Industry Research Institute， Nanchang 330013， China）

Abstract： In view of the poor real?time performance and low accuracy of motion intention recognition based on MI?EEG （motor imagery EEG）， such as unbalanced distribution of sample data of MI?EEG， imbalance dependence and attention on long distance in the extraction of time?series feature， and the difficulties in extracting the local feature， an MI?EEG signal classification method combining improved bidirectional long short?term memory （BiLSTM） neural network and full convolutional network （FCN） is proposed. In the proposed method， the conditional generative adversarial network （CGAN） is used to generate 1 MI?EEG signal samples， so as to realize effective expansion of the training sample set， which avoids the fact that the data set is excessively small and the number of its categories is unbalanced. By the respective advantages of bidirectional self?attention long short?term memory （LSTM） neural network and FCN， the facts of long?distance dependence and unbalanced attention in time?series feature extraction， difficulties in local feature extraction and inability to take into account the time?space domain features of MI?EEG signals are avoided. On this basis， the nonlinear mapping relationship between fusion features and action classification labels is constructed， so as to improve the recognition accuracy of the model. Finally， this classification model is compared with the other MI?EEG classification models in the test data set. The experimental results show that the accuracy of the proposed MI?EEG classification model reaches 97%， so it has good generalization performance.

Keywords： MI; EEG classification; GAN; LSTM neural network; FCN; attention mechanism

0" 引" 言

腦機接口（Brain?Computer Interface， BCI）作為新興技術，旨在通過記錄大腦活動，將人的意識直接與外部設備進行交互。運動想象（Motor Imagery， MI）是常用的BCI范式之一，通過要求被試者在腦中想象特定的運動，從大腦提取與運動想象相關的腦電信號（Electroencephalography， EEG），MI?EEG便可用于控制外部設備如外骨骼機器人和假肢等。因此，可通過BCI技術幫助具有肢體運動障礙的患者重新獲得一定的肢體運動能力[1?2]。……