A Study On EEG Signal Classification Algorithm Based On Transfer Learning

The brain-computer interface systems can control external devices by interpreting EEG signals and converting them into control commands without relying on human muscle tissue,thereby complete the process of human-computer interaction.This provides a way for patients with dyskinesia or neurological disorders to communicate with the outside world.At present,there are still many problems in the practical application of BCI system.In order to shorten the training time of BCI systems before use and reduce the burden on users,this paper proposes some transfer learning algorithms for reducing the calibration time of BCI systems.The research work is as follows:1)A transfer learning algorithm based on instance weights is proposed from the perspective of instance weights.The algorithm initializes the weights of the source domain data based on the marginal distribution similarity of the data in the source domain and the target domain,and then selects useful weighted source domain samples according to their conditional distribution similarity,adds them to the target domain training set to participate in the model training,and solves solved the problem of too little training data in the target domain2)From the perspective of model parameter sharing,transfer learning algorithms based on parameter sharing is proposed.This chapter considers that EEG signal features are in the form of a matrix rather than a vector.The parameters in time-space feature decomposition logistic regression(FD-LR)can be interpreted as channel weight parameters and time domain feature weight parameters.Based on the characteristics of stable time domain features of P300 EEG signals,the FD-LR model can share time-domain feature parameters.In this chapter,a multi-task FD-LR(MT-FD-LR)algorithm was proposed.Then MT-FD-LR is improved from two aspects.One is combining MT-FD-LR with instance weights method and Maximum Mean Discrepancy(MMD)to extend the multi-task learning model to the transfer learning model;the other is to generalize MT-FD-LR to Bayesian.In the framework,the generalization ability of the model is enhanced.3)The MT-FD-LR algorithm that shares the time domain feature parameters is extended to the convolutional neural network that shares the time domain filter layer,and a transfer learning algorithm based on convolutional neural network is proposed.The algorithm has excellent performance when there are only a few training samples in the target domain.The experimental results show that the proposed algorithms can train reliable classification models with a small amount of training samples.Compared with the traditional machine learning algorithms that do not migrate the knowledge of the source domain,the proposed algorithms can obtain better accuracy.The problem of long training time for BCI system is solved.