Research On Implicit Dynamic Learning Network For P300 Signal Recognition

Brain computer interface(BCI)is a system that uses electroencephalogram(EEG)signal recognition technology to express brain intention,and provides patients with severe disability with the ability to communicate with the outside world.As a type of brain computer interface,P300 speller uses P300 signal detection and recognition to complete the character spelling process.The spelling process generally has two recognition steps.The first step is to use the classifier to recognize the collected P300 signal.In this paper,neural network is used as the classifier.The second step is to obtain the corresponding characters by combining the signal recognition results with the character spelling interface.The classifier in step 1 is the key factor to improve the performance of brain computer interface.There are two performance indicators of brain computer interface,one is the accuracy of character recognition,the other is the information transmission rate.In order to improve the above two performances,this paper mainly makes an algorithm improvement and breakthrough for the neural network in step 1.Network parameter updating and optimization is an important step for neural network to realize recognition and classification,and its optimization algorithm determines the recognition performance of the network.Traditional neural networks mostly use gradient descent method to iteratively update the weight parameters along the gradient direction by using the weight gradient information.This method has the problems of slow convergence speed,easy to fall into local minimum and low generalization ability.For this type of neural networks,the performance optimization direction is mainly divided into adaptive learning rate,increasing momentum term and so on.This paper combines the neural dynamic method to design and transform from the algorithm level.Different from the gradient descent method,the neural dynamic method uses the time derivative information of the weight parameter to predict and update the next weight,so that it has the advantages of global convergence,exponential convergence and error accuracy approaching zero.At the same time,due to the parallel computing ability obtained by using the vector value error function,the calculation efficiency of the neural network is improved.Based on the above,this paper designs and proposes an implicit dynamic learning network based on neural dynamic method,which is divided into dynamic convergent differential neural network(DCDNN)and dynamic inverse learning neural network(DILNN).The dynamic convergent differential neural network iteratively optimizes all weight parameters directly through the neural dynamic design formula to improve the recognition accuracy of the network.The dynamic inverse learning neural network only uses the neural dynamics method to update the weight from the input layer to the hidden layer.Combined with the characteristics of the extreme learning machine,the pseudo inverse method is used to solve the weight parameters between the hidden layer and the output layer,which further speeds up the learning speed.In addition,for BCI system,in terms of preprocessing,in addition to the commonly used time window interception,band-pass filtering and training set balance,sliding average filtering is used to reduce the signal dimension on the premise of ensuring all signals work;In the aspect of recognition and classification,the classifier trained by all sample combination data sets is integrated and averaged,and the weak classifier is transformed into a strong classifier to improve the recognition performance of the system.Through the verification of BCI competition II and III public data sets,the proposed system framework has achieved higher character recognition accuracy and information transmission rate than other advanced methods.