Research Of EEG Signal Classification Methods Based On Deep Networks

As the increasing attention on the brain science by the government all over the world, the Brain-Computer Interface (BCI), which is of widespread application prospect and theoretical research value, has been a heated interdisciplinary research topic. BCI can offer humans a new way to communicate with external environment. Through this technology, brain can send commands to the machine directly without relying on the peripheral nerve and muscle system. Especially in the field of medical rehabilitation, it provides those atresia patients who have a central nerve injury an important way of communication. The classification of EEG signal is one import sector that concerns the overall performance of the BCI system and also the highlight of this paper. The main innovation and content of this paper is as follow.1. A modified MFCC method is proposed to extract the feature of EEG signal, and after carefully study of the common spatial pattern extraction algorithm based on spatial domain feature extraction and Hilbert-Huang algorithm based on non-steady signal processing, relative experiments is perform to compare the performance of these three methods, results demonstrate this method is more effective.2. The impact of different number of layers, nodes of the network as well as the different number of channels used as input to the classification result is discussed and verified by the experiment, the optimization method for parameters that used to train the deep network specified for EEG signal classification is summarized.3. Three deep architecture, Deep Belief Network, Stacked Auto-encoder and Deep Boltzmann Machine and their performance in EEG signal classification is studied. Besides, this paper also explored the performance of the combination of the three mentioned feature extraction method with different deep networks through experiments.In general, this paper is based on the sufficient study of the domestic and international research work related to brain signal processing and deeply exploration of the mechanism of how the EEG signal generated and the characteristics and significance in the medical field. And then taking the low signal-to-noise ratio, non-stationary signal motor imagery EEG signals as the breakthrough, conduct the research of feature extraction and deep learning classification method with the help of the data acquired from the BCI competition. This paper performs bold innovation and experiment on classification of EEG signal and provides new idea and method to help to develop more practical and efficient BCI system.