EEG Signal Classification Based On Attention And Multi-scale Convolutional Neural Networ

The brain-computer interface system based on motor imagery makes communication between human minds and external devices a reality.At present,it has been successfully used in cursor movement,electric wheelchair control,virtual reality,mechanical arm control,etc.The electroencephalogram(EEG)is considered as the input signal source of brain-computer interface system by virtue of its easy acquisition.However,it is still an important core technical problem to realize accurate analysis of brain-computer interface system of motion imagination.Due to the characteristics of low signal-to-noise ratio and large individual differences,it is difficult for the brain-computer interface system of motor imagination to achieve high accuracy on a single subject.At the same time,the powerful learning performance of deep learning injects new research impetus into EEG decoding.However,due to the limitations of collecting brain electrical signals,the scale of data obtained from the same subject is limited.This leads to easy overfitting of convolutional neural networks during training,thus limiting the classification performance of deep learning.In order to increase the number of active brain-computer interface instruction sets.In this paper,a sequential coding paradigm based on motor imagery and speech imagery is designed.Using different combinations of silent Chinese characters and motor imagery.Four types of imagination were obtained: motor imagery;speech imagery;motor imagery followed by speech imagery;speech imagery followed by motor imagery.During the tasks of verbal and motor imagery,subjects exhibit ERD/ERS phenomenon.The separability of EEG signals was proved.We recruited 12 participants for the offline experiment and provided a detailed description of the process of collecting EEG signals.This includes the design of stimulation interface,international standard electrode lead method,signal cycle and frequency,etc.The feasibility of the experimental paradigm was verified by analyzing the characteristics of EEG signal with event correlation spectrum disturbance.It provides a reliable basic for the classification and recognition of four kinds of imaginary data.A convolutional neural network based on attention and multi-scale is designed for the EEG signal collected by the above experimental paradigm.In order to solve the problem that convolution kernel of a single size cannot comprehensively extract time domain features in different frequency bands of subjects.A parallel multi-input cavity convolution and a mixedscale two-dimensional convolution are designed to fully extract the spatio-temporal and frequence domain features of EEG signals.The squeeze excitation module is also added in the network backbone,which can adaptively assign weight to the features with high classification accuracy.Thus improve the performance of the model.Among them,the average accuracy of four types of imaginary data increased from 65.5% to 71.1% of the input of a single scale.The addition of squeeze excitation module improves the effect of EEG decoding by the classifier.To address the overfitting issue caused by a small dataset of EEG signals,we employed a data augmentation algorithm in the time-frequency domain to expand the collected EEG signals.The average accuracy is improved from 71.1% to 75.4% by using the data enhanced samples as input based on attention and multi-scale convolutional neural network.The paired t-test was carried out to verify that the data enhancement algorithm significantly improved the decoding performance of the model.Finally,the open data set BCIIV2 a of brain-computer interface competition was input into the model designed in this paper,and an average classification accuracy of 77.2% was achieved.The accuracy of C2 CM,M3DCNN,CPMixed Net and Wa SFCNN were 74.46%,75.01%,74.6%and 66.9% respectively on the same public contest data set.The experimental results demonstrate the robustness of the proposed network model.