Research On Classification Of Motor Imagery Electroencephalogram Signals Based On Deep Learning

Controlling external devices through mind has become one of the hot topics in the field of artificial intelligence and brain science.Motor Imagery Electroencephalogram(MI EEG)signals are non-stationary,nonlinear and continuous random time series signals.Conventional machine learning methods are not enough to extract its features more completely.Based on the deep learning method and the relevant prior knowledge in the field of brain science,a lightweight feature fusion neural network model was designed to effectively improve the classification accuracy of MI EEG signals.The specific research work is as follows:1.Aiming at the problem that the scale of the pre-training dataset of some advanced neural network models does not match the MI EEG signals dataset,Firstly,the time-frequency domain features of original one-dimensional MI EEG signals were extracted by Continous Wavelet Transform(CWT).Furthermore,the time-frequency image dataset of MI EEG signals was augmented by super-resolution restoration technology.Experimental results demonstrate that compared with conventional data augmentation methods(such as noise injection,gamut transformation,etc.),super resolution restoration has the best performance in suppressing model overfitting and improving classification accuracy when realizing the data augmentation of MI EEG signals time-frequency images,which effectively promotes the combination of deep learning and MI EEG signals classification.2.In order to solve the problem of low classification accuracy of existing algorithms for MI EEG signals,this paper presents an Attention based Lightweight Feature Fusion Network(ALFFN)based on attention mechanism and tensor decomposition.To make full use of the high-level and low-level semantic features,a feature fusion module was seted to the output of each dense residual block,and the improved attention mechanism was used to screen out part of the feature maps that contribute to the classification results.In addition,the model was compressed by low-rank approximation and Singular Value Decomposition(SVD),which greatly reduces the number of model parameters.The amount of parameters of the ALFFN model was compressed by nearly 10 times after lightweight processing.The results of the experiment indicate that the average accuracy of the proposed scheme is 91.58%,and the average Kappa value is 0.881.We carried out T-test which have demonstrated that the performance of ALFFN is not significantly different from original ALFFN before parameters are compressed by nearly 10 times.The research in this paper provides a novel implementation scheme for MI EEG signals classification.The average accuracy of the proposed algorithm is 91.58%,and the average Kappa value is 0.881.The results of T-test demonstrate that the performance of ALFFN is not significantly different from that before compression after the parameters are compressed by nearly 10 times.The research has resulted in a solution of a novel implementation scheme for MI EEG signals classification.