Electroencephalogram Signal Classification Research Based On Deep Convolutional Neural Network

Electroencephalogram(EEG)signals,which represent electrophysiological activities of the brain,currently bear significant research implications in multiple fields,including cognitive psychology,intelligent healthcare,and human-computer interaction.Emotion recognition and sleep stage classification are two hot topics in EEG research.Under normal circumstances,the recognition of emotions and the classification of sleep stages require manual determination by experts in related fields based on their professional knowledge.Due to the inherent non-stationarity of EEG signals and the variability among different subjects,manual determination methods bear certain subjectivity.Thus,this paper establishes an automatic classification model for emotional and sleep EEG based on the convolutional neural network(CNN)framework.The main research content is as follows:In the task of emotion recognition,this paper proposes an emotion recognition model based on convolutional neural networks and a hybrid attention mechanism.Firstly,an emotion recognition model based on multi-scale convolution and a hybrid attention mechanism is proposed to extract features from different temporal scales,overcoming the limitations of traditional CNNs with a single convolution kernel,and introducing a hybrid attention mechanism to better characterize the channel and spatial features of feature maps.Secondly,considering that most studies neglect the differences between different brain area channels,this paper proposes a regional differential network structure.The aim is to calculate the difference matrix between symmetrical channels and combine depthwise separable convolutions to extract spatial difference features between different channels,thereby further improving network performance.The experimental results on the DEAP dataset show that the model proposed in this paper has classification accuracies of 96.77%and 97.09% for arousal and valence dimensions,respectively.In the task of sleep stage classification,given the existing research on sleep EEG neglecting the influence of sub-bands in different frequency ranges on the staging results,this paper combines the idea of a multi-head self-attention mechanism and proposes an automatic sleep staging model with a frequency band attention mechanism structure.Firstly,the sleep EEG signals are decomposed into various frequency bands using variational mode decomposition.As the penalty factor and modal number of variational mode decomposition significantly affect the decomposition results,this paper optimizes penalty factor and modal number based on the snake optimization algorithm initialized by iterative chaos mapping,to obtain more ideal sub-bands.Then,the frequency band attention mechanism is used to adaptively weight each frequency band,followed by the introduction of the ECAblock network as a feature calibration module to adaptively calibrate channel features,thereby obtaining the final model.Experimental results on the Sleep-edf20 dataset show that the proposed frequency band attention mechanism structure can effectively improve the accuracy of the model.In summary,this paper proposes models for emotion recognition and sleep stage classification based on the CNN framework,which not only has high accuracy and robustness but can also provide important decision-making references and suggestions for workers in the medical field.This represents significant research value and significance.