Research On EEG Signal Recognition Method Based On Deep Stacking Network With Adaptive Learning Rate

Human-computer interaction is an important part of artificial intelligence.The interaction between brain and machine is the most challenging task in human-computer interaction.The Brain-Computer Interface communicates the human brain with machines directly.Therefore,it has development potential in many fields.Because it is difficult to obtain EEG signals and they are non-linear and unstable,it is of great significance to extract EEG and apply it to practical applications.In the recent thirty years,scholars in this field have made breakthrough progress through methods such as time-frequency analysis and nonlinear decomposition.In the past decade,the rising of deep learning has made new breakthroughs in feature extraction and recognition of EEG signals.Based on the analysis of several existing common methods,an improved deep neural network to analyze and reconstruct EEG will be discussed and verified.This thesis first analyzes several conventional deep learning methods in EEG processing,including Deep Belief Network,Denoising Auto-encoder,Convolutional Neural Network,etc.Through the analysis and research of the above methods,considering the small scale of EEG data set,deep stacking network is adopted to make an improvement.Aiming at the problem that the vanishing gradient caused by the long training time of the deep stack network,the strategy of adaptive learning rate is introduced into the training algorithm of DSN to accelerate its convergence speed.The restricted Boltzmann machine is pre-trained by semi-supervised method.The learning rate is adaptively determined by performance analysis.Comparative experiments are performed using two sets of data.The result shows that DSN with adaptive learning rate has higher recognition accuracy,faster convergence speed,and can effectively classify motor imagery signals.Due to the low signal-to-noise ratio of EEG data,signal characteristics is difficult to be extracted.This thesis proposes a combinatorial method that using DAE to classify the EEG reconstructed by empirical mode decomposition.In this method,the noise level of DAE is adaptively adjusted based on simulated annealing algorithm,and solve local minima in non-convex networks by adaptive learning rate strategy in training epoch.The corresponding result shows that the network model can achieve faster convergence speed and higher recognition accuracy compared with the method without adaptive noise adjustment.Based on the trained model,an intelligent wheelchair system for EEG recognition is designed and constructed.The experimental result proves that the method can effectively perform left-right hand classification of motor imagery signals.