A Transfer Learning-Based CNN Deep Learning Model for Unfavorable Driving State Recognition

The detection of unfavorable driving states (UDS) of drivers based on electroencephalogram (EEG) measures has received continuous attention from extensive scholars on account of directly reflecting brain neural activity with high temporal resolution and low risk of being deceived. However, the existing EEG-based driver UDS detection methods involve limited exploration of the functional connectivity patterns and interaction relationships within the brain network. Therefore, there is still room for improvement in the accuracy of detection. In this project, we propose three pretrained convolutional neural network (CNN)-based automatic detection frameworks for UDS of drivers with 30-channel EEG signals. The frameworks are investigated by adjusting the learning rate and choosing the optimization solver, etc. Two different conditions of driving experiments are performed, collecting EEG signals from sixteen subjects. The acquired 1-dimensional 30-channel EEG signals are converted into 2-dimensional matrices by the Granger causality (GC) method to form the functional connectivity graphs of the brain (FCGB). Then, the FCGB are fed into pretrained deep learning models that employed transfer learning strategy for feature extraction and judgment of different EEG signal types. Furthermore, we adopt two visualization interpretability techniques, named, activation visualization and gradient-weighted class activation mapping (Grad-CAM) for better visualizing and understanding the predictions of the pretrained models after fine-tuning. The experimental outcomes show that Resnet 18 model yields the highest average recognition accuracy of 90% using the rmsprop optimizer with a learning rate of 1e − 3. The overall outcomes suggest that cooperating of biologically inspired functional connectivity graphs of the brain and pretrained transfer learning algorithms is a prospective approach in reducing the rate of major traffic accidents caused by driver unfavorable driving states.