The Research Of EEG Emotion Recognition Based On Attention Mechanism

For a long time,electroencephalogram(EEG)has been widely studied as physiological signal.Due to its high time resolution,wide spatial distrubution and other characteristics,it has been successfully applied in the fields of brain-computer interface,artificial intelligence,and medical health.Recently,the research on EEG emotion recognition is a popular topic at home and abroad,the essence is that the computer can recognize the emotional state of the subject by machine learning algorithms.Currently,due to that deep learning can explore the useful information in the data,it has significantly improved the recognition accuracy in many fields,and has been used by researchers in the study of EEG emotion recognition.There are two types of EEG emotion recognition based on deep learning,first one is using deep learning as a classifier to classify handcrafted features.However,it relies on manual extraction in the process of designing the EEG features and then classifying them by deep learning.The other one is classifying original signals by deep learning methods directly,it relies on the characteristics of signal itself,and can reduce the manual influence in the recognition This paper is going to design a data-driven and end-to-end technique,which utilizes the channel-wise attention mechanism to explore more discriminative spatial features,and untilizes the self attention mechanism to explore more discriminative temporal features,simultaneously.It is an attention based convolutional recurrent neural network,which can directly learn features from EEG signals and classify them.Finally,it can achieve end-to-end emotion recognition.This paper mainly explores the use of the attention mechanism in deep learning to extract more discriminative features in EEG signals,and to realize the recognition of the two dimensions of one emotion.The main innovations are:1.We develop a novel data driven framework named ACRNN to deal with EEG-based emotion recognition.This framework integrates the channel-wise attention mechanism into CNN to explore spatial information,which can take spatial information of multichannel EEG signals by CNN and the importance of different channels by channel-wise attention into consideration.Besides,ACRNN integrates extended self-attention mechanism into RNN to explore temporal information of EEG signals,which can take the different temporal information by LSTM and the intrinsic similarity of each EEG sample by extended self-attention into consideration This paper proposes a novel data-driven deep learning framework named attention-based convolutional recurrent neural network(ACRNN).This framework combines the channel-wiseattention mechanism with a convolutional neural network,and considers the global information extracted by the convolutional neural network and the importance information between the channels.Besides,the self-attention mechanism is combined with the recurrent neural network,and the time information extracted by the recurrent neural network in the EEG signals and the importance information between different samples are considered simultaneously;2.We conduct experiments on two public databases,i.e.DEAP and DREAMER,and experimental results indicate that the average emotion recognition accuracies can achieve 92.74% and 93.14% on valence and arousal classification task of DEAP database,respectively.Besides,the method can achieve high performance with mean accuracy of 97.79%,97.98% and 97.67% on valence,arousal and dominance classification task of DREAMER database,respectively.