基于LSTM的脑电情绪识别模型

Emotion recognition from EEG signals by using LSTM recurrent neural networks

已有研究表明,通过分析人类的脑电信号可以识别出其情绪信息.近年来,机器学习技术的发展为基于脑电信号的情绪识别研究提供了可靠的技术手段.传统的机器学习技术简单地从多个通道的脑电信号中提取特征,然后连接成单个特征向量,但是没有考虑到脑电信号中至关重要的时间动态信息.深度学习技术中的长短时记忆(Long Short-Term Memory,LSTM)网络因其时间上的递归结构,可以很好地解决这个问题.然而,脑电序列通常较长,直接用来训练LSTM模型所需的计算资源非常大且学习到的信息类型单一,而且忽略了许多对情绪识别非常重要的信息,如频域信息和非线性动力学信息.为此提出一种新的基于LSTM的情绪识别模型.脑电信号被分成多个非重叠的信号段,并从每段信号中提取多种时域、频域和非线性动力学特征,这些特征沿时间连接成特征序列并用来训练LSTM分类模型.在DEAP数据集上验证了该模型在愉悦度、唤醒度和喜欢度上的二分类准确率,其中每个情绪维度分为低和高两类.实验结果表明,该模型在愉悦度和喜欢度上的分类准确率均优于已有方法,在唤醒度上的分类准确率仅次于最先进的成果.

It has been demonstrated that it is possible to recognize human emotions by using electroencephalogram(EEG)signals.In recent few years,the development of machine learning technology has provided reliable techniques for EEG based emotion recognition research.Traditional machine learning methods extract features from multi-channel EEG signals in each channel and concatenate them into a single feature vector,ignoring critical temporal dynamic.The Long Short-Term Memory(LSTM)in deep learning technology can solve this defect well due to the recurrent structure.However,directly fitting the LSTM based model using EEG signals consumes huge computer resources and ignores important frequency domain and nonlinear dynamic information.In this paper,we present a new emotion recognition method based on LSTM.Various features in time domain,frequency domain andnonlinear dynamic are extracted from multi-channel EEG signals and constructed into feature sequences which are used to train the LSTM based model.Experiments are carried out on the DEAP benchmarking dataset for valence,arousal and liking classification respectively,and every emotional dimension is divided into two classes(low and high).Experimental results demonstrate that the classification accuracy of the proposed model outperforms the previous methods,with regard to both of valence and liking emotional dimensions,and is also comparable to the most advanced method for arousal classification.