Research On Fatigue Detection Algorithm Using EEG Signals Based On Transfer Learning

Fatigue occurs when people are subjected to prolonged or repeated stimulation.As a complex physiological and psychological phenomenon,fatigue is often manifested as the lack of concentration,weakness and so on.While understanding the cause of fatigue,it is also important to find a method that can detect fatigue timely and accurately.At present,among many methods for fatigue detection,EEG-based fatigue detection methods have received extensive attention,and are regarded as the gold standard for fatigue detection.This is because EEG signals are the overall performance of neurons on the scalp,which can accurately reflect the overall state of the brain in time.However,the EEG signals are non-stationarity and vary greatly among different subjects,so that the training data and the test data do not obey the same distribution or in the same feature space,which makes the traditional machine learning models cannot achieve satisfactory performances.The emergence of transfer learning provides a feasible method to solve the kind of problems,because it can use the knowledge that has been learned to help solve related but different task learning problems.This thesis uses transfer learning to research the problem of EEG-based cross-subject fatigue detection,and the contributions are mainly as follows:(1)Firstly,this thesis proposes a method for fatigue detection based on tensor network.When selecting the training set,this thesis uses the Frobenius norm-based similarity metric of tensors to select the training set that is the most similar to the test set.Then,this thesis uses two different balancing methods,namely under-sample majority samples(UMS)and synthesis minority samples(SMS)to balance a public fatigue dataset.In addition,this thesis uses the constructed tensor network to project the training data and the test data into the same feature space.Finally,the support vector machine model is used for classification in the feature space.The offline analysis result on the public fatigue dataset shows that the average accuracy of the fatigue detection method proposed in this thesis has reached 78.7%,which is higher than the traditional machine learning algorithms.(2)Secondly,this thesis also proposes a multi-task learning-based method for fatigue detection.Specifically,a balanced fine-tune set is introduced on the basis of a Bayes-based multi-task learning model.The multi-task learning model is used to learn shared features between different subjects,and the balanced fine-tune set can narrow the differences between shared features and test subjects' features,and the combination of the two steps can achieve cross-subject fatigue detection.The experimental results on a public fatigue dataset show that the average accuracy reaches 77.0% when UMS is used;while when SMS is used,the corresponding average accuracy greatly increases to 83.2%,which is 10.2% higher than domain adaptation transfer learning methods.The two EEG-based cross-subject fatigue detection methods proposed in this thesis have both achieved good results,which make the researches for fatigue detection further.