EEG Emotion Recognition Based On Multi-kernel Broad Learning Systems

In recent years,the rapid development of machine learning(ML)and information fusion has made it possible for computers to be endowed with the ability to understand,recognize and analyze emotions.Emotion recognition has attracted strong interest from researchers in various fields.Traditional manual feature extraction methods have achieved good results in EEG emotion recognition.With the rise of deep learning,it has been widely used in EEG emotion recognition.Compared with traditional manual feature extraction methods,deep neural networks can learn more robust and deeper features,thereby improving recognition accuracy.However,its width is as important as the depth.Since the introduction of broad learning,it has attracted wide attention.The expansion of width direction makes the network play a more important role,which provides ideas for the research of emotion recognition.In addition,the kernel method,as an effective technique of feature extraction and representation learning,has effectively improved the performance of algorithms,and greatly promoted research in the field of machine learning.Consequently,this paper mainly focuses on the research of EEG emotion recognition based on Multi-kernel broad learning systems.The main contributions are as follows:(1)An EEG emotion recognition algorithm is proposed,which is based on a convolutional neural network and Multi-kernel broad learning system.To extract the effective EEG signal features,this paper adopts the combination of preprocessing and feature extraction.The preprocessing obtains the frequency information in EEG data,and a convolution neural network is designed to further extract the deep spatial features.In addition,the Fourier kernel approximation method in Multi-kernel broad learning system enhances the distinguishability of frequency and spatial information in EEG signals,thereby improving the recognition effect of the broad learning system in EEG emotion.(2)Convolutional neural networks focus more on learning local feature information when extracting EEG signal features,but cannot extract discrete and discontinuous features in the space domain.Therefore,according to its inherent properties and distribution,this paper adopts the graph convolutional network to extract the internal connection of the EEG channel.Regarding the kernel approximation method,considering that a random orthogonal matrix replaces random Gaussian matrix can reduce the error of approximate kernel function,this paper combines orthogonal random feature kernel approximation with broad learning for the first time and proposes an improved Multi-kernel broad learning system.Comparative experiments on commonly used EEG data sets verify the effectiveness of the emotion recognition method based on a Multi-kernel broad learning system.In addition,the improved algorithm proposed in this paper improves the accuracy of EEG emotion recognition.The average accuracy on the SEED dataset is 97.78%,and the average accuracies of the valence and arousal on the DEAP dataset are 96.88% and 97.03%,respectively.The valence,arousal and dominance degrees on the DREAMER dataset are 88.36%,89.34%,and90.09%,respectively.