Deep Learning Analysis Of Multi-channel EEG And SEMG Signals

Electrophysiological signals are one of the reliable standards for measuring human physiological signs,among which electroencephalography(EEG)signals and surface electromyography(sEMG)signals have been widely studied in China and abroad.Deep learning and transfer learning algorithms,as new techniques,can extract rich feature information from EEG signals and sEMG signals.In this paper,based on deep learning and transfer learning algorithms,we develop computational models for seizure EEG data and sEMG data for various classification tasks to achieve high accuracy seizure prediction and sEMG gesture recognition effects.In seizure prediction research,we propose the channel attention dual-input convolutional neural network(CADCNN).CADCNN effectively uses prior knowledge(shorttime Fourier transform to extract EEG features)to improve the model’s capture of spectral information.By fusing EEG signal and prior knowledge and combining channel attention,the model improves the utilization of EEG signal capability’s temporal,spectral and spatial information.For 17 subjects in the CHB-MIT epilepsy dataset,our proposed method achieves 97.1%,0.029/h,95.6%and 0.917 in sensitivity,false prediction rate,specificity and AUC,respectively,showing better performance and higher prediction accuracy.To achieve high-accuracy sEMG gesture recognition,we develop a Feature ChannelSpatial Attention Multiscale Network(FCSAMnet)to decode sEMG signals.FCSAMnet extends the network width with a multi-branch structure to further superimpose and fuse spatial-temporal feature information.In addition,FCSAMnet pays attention to the channel and spatial significance of the sEMG signal feature maps and enhances the model’s ability to capture the underlying feature information by adding attention to the inter-channel relationships of features and the spatial relationships between features.FCSAMnet presents good classification performance in Ninapro DB2 and CapgMyo DB-a datasets.The models achieve an average classification accuracy of 86.21%,90.77%and 92.53%for 17,8 and 9 gestures in Ninapro DB2,respectively.In CapgMyo DB-a,the model achieves an average classification accuracy of 98.85%for 8 gestures.In addition,we propose a transfer learning model of domain-adaptive adversarial temporal convolutional network(DATCN)for the problem of the variability of data space distribution between different subjects of EEG signals and sEMG signals and the reduced generalization ability of the algorithm on different subjects.The model extracts discriminative domain invariant features from source and target domain data using adversarial temporal convolutional networks and implements domain adversarial training through a gradient inversion layer in network optimization to accomplish the invariance of domain feature migration.In this paper,we design sEMG experiments and collect data from multiple subjects to perform cross-subject migration training,and achieve high accuracy gesture recognition on different subjects.We apply the model on the CHB-MIT epilepsy dataset to verify the algorithm’s effectiveness and achieve the same high precision classification effect.