Research On Federated Learning Algorithm On Driver Drowsiness EEG Detection

Drowsiness leads to the decrease of driver’s alertness and cognitive ability,which has a great impact on road traffic safety.With the continuous development of machine learning,the automatic detection method of driver drowsiness based on Electroencephalogram(EEG)has gradually matured and been widely used.However,the current mainstream drowsiness detection methods mainly improve the classification accuracy by collecting a large number of EEG data,which ignores the privacy involved in the driver drowsiness EEG data and the interpretability of the model.In addition,the data heterogeneity of different clients in the Federated Learning(FL)architecture will lead to slow model convergence and poor generalization performance.Therefore,this thesis aims at the following issues:First of all,this thesis proposes a driver’s drowsiness EEG detection method(TFCG)based on tree Federated Learning(Tree-FL)and interpretable network(CNN-GAP).This method changes the Federated Learning algorithm for the privacy problem involved in the driver drowsiness EEG data.The tree structure of parallel transmission between clients enhances the privacy protection while improving the communication efficiency.In addition,the interpretability module is introduced into the deep learning model architecture to improve the effective decisionmaking level of driver drowsiness classification.Through the Leave-One-Out cross validation experiment on 11 subjects,the evaluation shows that this method has higher average accuracy,F1 score and AUC than the traditional classification method,which are 73.56%,73.26% and 78.23%respectively.Compared with the conventional Federated Learning algorithm,the method better protects the driver’s privacy.Furthermore,this thesis proposes a driver drowsiness EEG detection method(PFCL)based on personalized Federated Learning(Per-FL)and interpretable network(CNN-LSTM).This study improves the deep learning model’s structure of the Federated Learning client on the precedent work in order to further investigate the interpretability of the model.The model extracts EEG features using the Convolutional Neural Network(CNN),and uses the output hidden state of the Long Short-Term Memory(LSTM)network to achieve visualization.In addition,in view of the data heterogeneity of different clients in the Federated Learning architecture,Moreau envelope is introduced as the regularization loss function of the client,which improves the convergence speed of the model.According to the test findings,the technique has an average accuracy of 74.72%,an average F1 score of 75.75%,and an average AUC of 78.39%.To sum up,this thesis proposes two methods of driver drowsiness EEG detection based on TFCG and PFCL respectively,providing a safe and efficient driver drowsiness EEG detection technology for different traffic management platforms,and also providing scientific reference for Federated Learning in the actual scene of driver drowsiness EEG detection.