基于小波包和深度信念网络的脑电特征提取方法  

EEG feature extraction method based on wavelet packet and deep belief network

作　　者：李明爱 张梦 孙炎珺 

机构地区：[1]北京工业大学信息学部自动化学院

出　　处：《电子测量与仪器学报》2018年第1期111-118,共8页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(81471770);北京市自然科学基金(7132021)资助项目

摘　　要：针对运动想象脑电信号(motor imagery electroencephalography,MI-EEG)的时变性、个体差异性等特点,提出一种将小波包变换(wavelet packet transform,WPT)与深度信念网络(deep belief networks,DBN)相结合的脑电特征自动提取方法,记为WD法。首先,利用平均功率谱方法对MI-EEG进行时域分析,选取有效的时序段。其次,使用WPT对有效时域段的各导MI-EEG进行时频分解,并选取与想象任务相关的频段信息重构脑电信号;然后,将各导重构MI-EEG串接,并将其瞬时功率信号输入给DBN模型实现特征自动提取。最后,利用Softmax分类器完成脑电想象任务的模式分类。在DBN模型训练中通过增加Dropout训练技巧来解决因训练数据少等引起的过拟合问题,以提高分类结果。利用BCI标准竞赛数据库进行实验研究,5-折交叉验证法取得了94.06%的分类准确率,证明该方法能够充分利用脑电的神经生理学特点,自适应地提取个性化的深层脑电特征,有利于改善分类效果。Because of the time-varying and subject-specific characteristics of motor imagery electroencephalogram( MI-EEG),an adaptive feature extraction method is proposed based on the wavelet packet transform( WPT) and deep belief networks( DBN),and it is denoted as WD. Firstly,the MI-EEG is analyzed in time domain based on the average power spectrumto select the effective segment of time sequence. Secondly,WPT is applied to decompose each channel of MI-EEG and reconstruct according to the effective timefrequency information related to motor imagery task. Then,all channels of reconstructed MI-EEG are concatenated in series,and their instantaneous power signals are outputted to DBN for adaptive feature extraction. Finally,a Softmax classifier is used to classify MIEEG. In training stage of DBN,Dropout training technique is addedto solve the over-fitting problemresulted from lack of training datato improve the classification results. The experiments are conducted on standard BCI competition dataset,and 5-fold cross validation achieves 94. 06% classification accuracy. The experimental results show that the featuresof MI-EEG can be extracted automatically by WD,andthe neurophysiological characteristics be fully used to improve classification effect.

关 键 词：运动想象脑电信号 深度信念网络 小波包变换 Softmax分类器 DROPOUT 

分 类 号：R318[医药卫生—生物医学工程] TN911.7[医药卫生—基础医学]