Classification Of Eeg Signals Based On Band Attention And Decision Fusion

EEG signals classification has recently become a hot topic in the field of pattern recognition,and a variety of different intelligent disease diagnosis methods have been proposed.In order to achieve multi-band analysis of signals and improve the performance of EEG classification,this paper conducts the following specific research from the perspective of band selection and decision fusion:Firstly,this paper proposes a classification method for EEG signals based on BAResNet.A band attention module is added to the classification model to find the key bands out of many different frequency bands and implement differentiated band analysis.It achieves the flexible band attention distribution by calculating the weights of band importance and then scales by the weights to enhance the effects of critical band characteristics,finally implementing the class prediction by using the residual network module.The method is tested on the datasets from the Anding Hospital and the Boston Children's Hospital.The experimental results show that the band attention module can give more attention to the key bands and improve the performance of the method.Secondly,from the perspective of multi-band decision fusion,this paper proposes other two EEG signals classification methods based on decision fusion residual network,including SBDF-ResNet and CMBDF-ResNet.The SBDF-ResNet uses the residual sub-module branches with sharing weights to obtain the different decision according to the different frequency bands of input in and then combines the decision information of the multiple frequency bands to generate the final class prediction value of the samples.Based on the former,the CMBDF-ResNet adds an additional branch for obtaining mixed-band decision information.The branch uses the SE residual module to enhance the impact of key band characteristics and then uses the residual sub-module that do not share weights with other branches to obtain the decision value of the mixed band feature.Finally,the decision information of each branch is fused to finish the sample classification.The experimental results demonstrate that both methods have good performance in the detection of depression and seizure prediction.At the same time,compared with BA-ResNet,the classification performance can also be improved by using these two methods.Finally,a multi-channel EEG signals classification system is built,which extends the functions from three directions: feature extraction,feature classification and auxiliary functions.In the part of the feature extraction,the system adds sub-functions such as band-pass filtering and correlation calculation;in the part of the feature classification,BA-ResNet,SBDF-ResNet,CMBDF-ResNet,and other classification methods are integrated into the system;auxiliary functions include data segmentation,data set partitioning and result evaluation.