The Detection Of Brain Signal Based On Convolutional Neural Networks And Realization Of BCI System

With the development of artificial intelligence and requirement of biomedical systems, many countries and regions have established the research project about brain, Brain-Computer Interface has been one of the most popular research project. BCI system established communication from brain to external equipment. BCI system can divide into two categories based on the method of acquiring evoked Potentials, one is based on motor imagery and the other is based on evoked potentials. The BCI system that based on motor imagery doesn’t need communicate with the outside world, it just needs the subjects imaging or single limb movement to evoke the potentials, but the BCI system that based on evoked potentials must communicate with the outside, it normally needs visual stimulation, the typical evoked potentials is P300 signals and steady-state visually evoked potentials.The important point of BCI system is the evoked potentials acquirement and the feature extraction, traditional algorithms includes Linear Discriminant Analysis, Stepwise Linear Discriminant Analysis and Shrinkage Linear Discriminant Analysis. The algorithms of SWLDA and SKLDA mainly solve the small sample problem, although they can solve the problem in a certain way, but the performance of the algorithms drops in the case of enough samples.With the development of deep learning, it has achieved huge success in the image recognition field, as well as, in other field, such as speech recognition and biomedical signal processing, it has been researched constantly. So in the background, the thesis used the mainly typical algorithm model of deep learning, expand research at the basic of P300 signal, as well as, the thesis designed and realized the BCI system based on SSVEP in the board of OMAP3530. The work of the thesis as follows:1) The thesis used the algorithms of LDA, SWLDA, SKLDA to recognize the P300 signal that from BCI III competition, and then it verified that the algorithms of SWLDA and SKLDA can solve the small sample problem partially but the performance will drop in the condition of enough samples.2) The thesis used the point of local receptive field and weight sharing in the image field and recognize the P300 signal at basic of Convolutional Neural Networks Model. It used 185 characters and did feature extraction and feature representation constantly, and then got the results to compare with other three algorithms. Then it can conclude that the proposed algorithm has better performance and it will do better if there is more samples. Compared with the traditional algorithms, the algorithm has better generalization ability, it has important reference value for the BCI system.3) The thesis designed and realized the BCI system based on SSVEP, it runs in the development platform. Through the progress of LED stimulator designer, the SSVEP signal acquirement and transmission, and the development environment establishment, it has good performance. Before the system working, the subject firstly studied about 30 seconds, the system used sliding windows mechanism to ensure the recognition accuracy above 90%, the system interaction time is within 2 seconds as well, it can meet the practical requirements and bring benefits for the patients with movement disorders.