EEG Feature Extraction And Recognition Based On Deep Stacking Network

Brain-computer interface(BCI)is a kind of human-computer interaction technology which is independent of peripheral nerve and muscle system and has potential application value in many fields such as medical treatment,military and entertainment.The feature extraction and classification of electroencephalogram(EEG)data is the key to realize a BCI system.However,high-dimensional and nonlinear with nonstationary properties,EEG data are challenging to analyze and model.It is easy to loss information applying traditional EEG feature extraction algorithm.Deep learning possesses formidable nonlinear and high-dimensional data processing capability and it allows a machine to be fed with raw data and to automatically discover the representations needed for detection or classification.Recent years have seen it introduced to process EEG signals.This thesis first studies some frequently-used deep learning methods for EEG feature extraction and recognition,including deep belief network(DBN),sparse auto-encoder(SAE)and convolutional neural networks(CNN).Aiming at addressing the problem that it is easy to cause gradient diffusion when back-propagation is applied to fine tune these deep neural networks,deep stacking network(DSN)is chosen as the basic network in this paper.Deep stacking network implements the supervised training of each layer in a deep neural network,thus the gradient diffusion can be effectively avoided.Aiming at the problem that the middle layer of DSN is completely unknown,leading to its learning results cannot be explained,the prior knowledge is introduced to change the initialization of the hidden layer and make it partially visible.Then,the particle swarm optimization(PSO)algorithm is applied to optimize the input weights and a PSO optimized hidden-layer visible deep stacking network is proposed.The proposed method is applied to motor imagery EEG feature extraction and recognition and compared with some typical feature extraction algorithms.Experimental results show that the proposed algorithm obtains higher recognition accuracy than the typical feature extraction algorithms and it is robust against transferring from session to session to a certain extent.A semi-supervised feature learning method is proposed aiming at resolving the waste of unlabeled samples in supervised learning and the information loss caused by conventional feature extraction methods.The unsupervised learning of restricted Boltzmann machine(RBM)is combined with the fine-tuning of DSN.Multiple RBMs are trained independently by unsupervised learning for the parameter initialization of the neural network.Then the batch-mode based gradient descent algorithm is applied to fine tune the network.Experimental results demonstrate that the proposed algorithm has a good effect on the motor imagery EEG recognition.