Research And Application Of EOG Classification Based On Transfer Learning

With the development of artificial intelligence and Internet of Things,EOG has become the hot research direction of human computer interface and medical rehabilitation,and has broad market prospects in the field of functional assistance and environmental control.The traditional EOG system requires a long-term offline training for the new subject to obtain a classifier model.This process takes a long time and has a pool user experience,which is harmful to the promotion of the EOG system.Therefore,how to cancel the long training phase of new subjects has great research value.The new subjects did not participate in offline training and required direct online experiments.The experimental data of other subjects could be used to train the model,and then the model was applied to new subjects,which is a problem of transfer learning.This study starts with the EOG transfer learning algorithm of single source domain,and then extends to the transfer learning algorithm of multi-source domain.Finally,the EOG system based on transfer learning is designed.For the EOG transfer problem of single source domain,try to use LDA and SVM to transfer directly,the result is that SVM is better than LDA.Then the GFK algorithm is introduced to map the source and target domain samples before use the basic classifier.GFK significantly improves the LDA's result,but there is no obvious improvement to the SVM's,so the GFK algorithm is improved.The experimental results proved that the improved GFK is superior to the original GFK algorithm.For the EOG transfer problem of multi-source domain,the first attempt is to use the aggregation strategy.The experimental results show that the basic classifier most suitable for the aggregation strategy is KNN.After analyzing the TPT algorithm,the TPT algorithm is improved by introducing different personality classifier acquisition methods,different kernel matrix calculation formulas,and different regression algorithms.The results show that the improved TPT algorithm has achieved a greater improvement than the original TPT algorithm.Finally,an EOG system based on transfer learning is designed,and the problem of insufficient local computing power is solved by using the cloud platform.For new subjects,the system does not need training.The system is divided into two phases in the selection of the algorithm.The first phase uses KNN-SVM,and when the cloud in the system receives enough sample samples,it enters the second phase and uses the model generated by TPT algorithm.Experiment confirmed the feasibility of the system without training for new subjects.