电子测量技术ELECTRONICMEASUREMENTTECHNOLOGY第45卷第23期2022年12月DOI:10.19651/j.cnki.emt.2210203基于CNN-LSTM的脑电P300信号检测范方朝1杜欣1谢城壁1刘佳伟1黄涌2(1.北京交通大学电气工程学院北京100091;2.蓝色传感(北京)科技有限公司北京100085)摘要:为提高对无创脑机接口(BCI)中P300脑电信号的检测准确度,本文根据卷积神经网络(CNN)与长短期记忆(LSTM)网络,提出一种CNN-LSTM组合网络模型。卷积网络采取分层结构,同时设计匹配不同特征维度的一维卷积核;长短期记忆网络(LSTM)用来发掘数据时序相互依赖性,学习全局特征的相关性以实现目标分类。试验结果表明,本文提出的模型对于实验诱发出的单试次P300信号,检测准确率达到91.28%,与EEGNet网络和支持向量机算法对比,准确率分别提升2.18%、8.31%。在精确率、召回率、F1分数、AUC值的评价指标下也达到最优性能,具有较强的泛化性能。关键词:脑机接口;P300信号;卷积神经网络;长短期记忆网络中图分类号:TN911.7文献标识码:A国家标准学科分类代码:510.4010AP300signaldetectionalgorithmbasedonCNNandLSTMFanFangzhao1DuXin1XieChengbi1LiuJiawei1HuangYong2(1.SchoolofElectricalEngineering,BeijingJiaotongUniversity,Beijing100091,China;2.BlueSensing(Beijing)TechnologyCo.,Ltd.,Beijing100085,China)Abstract:InordertoimprovethedetectionaccuracyofP300EEGsignalsinnon-invasivebrain-computerinterface(BCI)system,thispaperproposesaCNN-LSTMcombinednetworkmodelbasedonconvolutionalneuralnetwork(CNN)andlongshort-termmemory(LSTM)network.Theconvolutionalnetworkadoptsahierarchicalstructure,anddesignsaone-dimensionalconvolutionkernelthatmatchesdifferentfeaturedimensions;longshort-termmemorynetwork(LSTM)isusedtoexploretheinterdependenceofdatatimeseries,learningCorrelationofglobalfeaturesforobjectclassification.Thetestresultsshowthatthemodelproposedinthispaperhasadetectionaccuracyof91.28%forthesingle-trialP300signalinducedbytheexperiment.ComparedwiththeEEGNetnetworkandthesupportvectormachine(SVM)algorithm,theaccuracyisincreasedby2.18%and8.31%,respectively.ItalsoachievestheoptimalperformanceundertheevaluationindicatorsofPrecision,Recall,F1scoreandAUCvalue,andhasstronggeneralizationperformance.Keywords:brain-computerinterf...
