摘要
针对传统的人工特征选取需要耗费大量时间和精力的问题,本文在传统卷积神经网络(convolutional neural networks,CNN)模型的基础上,提出了一种基于多尺度卷积核CNN的特征提取与分类方法,并在脑电情绪识别分类上进行了验证。本文首先进行了通道选择方面的研究,其次使用多尺度卷积核CNN模型对提取了微分熵(differential entropy feature,DE)特征的脑电数据进行情绪三分类实验,相比于传统的CNN模型,多尺度卷积核CNN模型在卷积层中采用多个尺度的卷积核,同时从高维度与低维度对脑电信号进行二次特征提取。实验结果表明,预处理数据在33通道的情绪分类平均准确率为89.72%,几乎接近全通道的平均准确率;多尺度卷积核CNN在微分熵特征上的情绪三分类取得了98.19%的平均分类准确率,实验结果证明了该模型的有效性和鲁棒性。
Aiming at the problem that traditional artificial feature selection needs a lot of time and energy,this paper proposes a feature extraction and classification method based on multi-scale convolutional neural networks(CNN)model,which is verified by Electroencephalogram(EEG)emotion classification.Channel selection is studied firstly,and then differential entropy(DE)feature is extracted by using multi-scale convolution kernel CNN model.Compared with the traditional CNN model,the multi-scale convolution kernel CNN model uses multiple scale convolution kernels in the convolution layer.the EEG features are extracted from the high and low dimensions at the same time,Therefore,the classification accuracy can be improved.The experimental results show that the average accuracy of the preprocessing data in 33 channels is 89.72%,which is almost close to the average accuracy of the whole channels;the multi-scale convolution kernel CNN achieves an average classification accuracy of 98.19%for emotion classification on DE features.The experimental results prove the effectiveness and robustness of the model.
作者
戴紫玉
马玉良
高云园
佘青山
孟明
张建海
DAI Ziyu;MA Yuliang;GAO Yunyuan;SHE Qingshan;ZHANG Jianhai(Institute of Intelligent Control and Robotics,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province,Hangzhou Zhejiang 310018,China;School of Computer,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2021年第4期496-503,共8页
Chinese Journal of Sensors and Actuators
基金
国家自然科学基金项目(62071161,61971168,61871427)
浙江省重点研发计划项目(2020C04009)
浙江省教育厅科研项目(Y202044267)。
