A Support Vector Neural Network For EEG Signal Classification

With the development of life science,cognitive science and artificial intelligence,brain science research with the recognition of EEG signals as the core is booming.The cognitive research on the brain in the 21 st century is called the brain research century.Brain computer interface(BCI)analyzes EEG signals from the brain and realizes direct communication between people and machines,which can help patients with severe disabilities control external machines or robots to complete the expected tasks.Therefore,the classification method of EEG signals plays an important role in the development of BCI system and technology.The commonly used method in EEG classification is support vector machine(SVM)algorithm.Its essence is a convex quadratic programming(QP)classification problem based on maximizing classification margin.The traditional solution method is a serial numerical solution algorithm,which is limited by real-time and large-scale calculation.Compared with the numerical algorithm,the neural network solver based on neural dynamics has more advantages in real-time and large-scale computing because of its parallelism.At present,there are three methods based on primal dual neural network(PDNN),simplified model of PDNN and variable parameter recurrent neural network(VPRNN),but the above algorithms were only used in solving optimization problems.Therefore,this paper focuses on the design of neural network classifier based on neural dynamics.Based on the convex quadratic programming classification problem of maximizing the classification interval,combined with PDNN,this paper proposes a support vector neural network(SVNN)algorithm to solve it,and applies it to the classification scheme design of EEG signals.This paper proves the global convergence of the proposed algorithm and SVNN It can converge to the optimal solution,and the experimental results show the effectiveness of the algorithm.Then,based on the SVNN model,a simplified support vector neural network(SSVNN)with more simplified structure is proposed,which proves the global convergence of the proposed algorithm.The experimental results on the open data set of motion imagination verify the effectiveness of the proposed algorithm.Finally,in order to obtain better convergence characteristics,this paper proposes a support vector recurrent neural network(SVRNN)of VPRNN solver combined with penalty function,and proves its global convergence and convergence The better classification effect of the algorithm in the public data set verifies the effectiveness and accuracy of the proposed EEG classification scheme.