Imotion Imagination Classification Based On Echo State Network

As an emerging technology,Brain Computer Interface(BCI)has become a research hotspot in the fields of rehabilitation engineering and biomedicine in recent years.The BCI system collects the activity of the cerebral cortex through devices for example electrode caps,and then converts the brain signal(EEG)into a control signal of the external device,and realizes communication with the outside world by controlling the external device.The correct classification of EEG signals is the key factor determining the performance of brain-computer interface.However,the traditional classification methods have complex training algorithms and are easy to fall into local minimum problems,which affect the performance of classification and restricts its application effect.To this end,a new type of neural network-echo state network is introduced,and the classification methods are discussed by using the advantages of simple training algorithm and global optimization.In this paper,the two types of motor imagery EEG signals are studied.Considering the superiority of ESN classification method,aiming at the problem that the feature discrimination of EEG signals is not obvious and the classification recognition rate is not high,the classification recognition rate of EEG signals is improved by optimizing the reserve pool,optimizing the readout part and improving the discrimination of feature vectors.This paper has carried out the following research work:(1)For the problem that the standard echo state network reserve pool is completely randomized and lack of pertinence to classification problems,a classification method based on ESN of internal plasticity reserve pools is proposed to optimize the reserve pools by referring to the known biological mechanism called "Intrinsic Plasticity(IP)".By introducing two parameters of gain and deviation into the neural activation function,the ESN network achieves maximum information transmission and improves classification accuracy.The experimental results show that this method has higher classification accuracy than the standard ESN.(2)Aiming at the problem that ESN readout part adopts linear model and may not be able to simulate the information embedded in the space state of the reserve pool,a bidirectional deep ESN method combining ESN with Multi-Layer Perceptron(MLP)is proposed.By reforming the reserve pool model to complete the feature extraction of time series data,and then by principal component analysis(PCA)method to reduce dimensionality and MLP classification,the classification accuracy can be effectively improved.The experimental results show that this method has higher classification accuracy than the standard ESN.