Classification of EEG Signals from Driving Fatigue by Image-Based Deep Recurrent Neural Networks

Evaluating fatigue driving detection systems has been a major interest of researchers. With the development of deep learning algorithms and multimodal technologies, more and more scholars have begun to look for new representation methods from multi-channel EEG time series data. Here, we propose to convert fatigue and normal state EEG data into a series of topology-preserving multispectral images with accompanying eyelid closure ratio (PERCLOS) labeling input into an image classification-inspired deep recurrent convolutional neural network we train, to find new data-robust representations from image sequences. We used the dataset SEED-VIG published by Shanghai Jiao Tong University in 2017 to help build a fatigue detection system. This dataset contains eye movement signals of 23 subjects and multimodal signals collected by 16 EEG electrodes. We use appropriate frequency bands to characterize fatigue to form color frequency bands, form a three-channel image based on azimuthal equidistant projection and Clough-Tocher interpolation algorithm, and form a two-dimensional image instead of signal as input to the model. The proposed method differs from the eigenvector representation in traditional EEG analysis and aims to preserve the spatial, spectral and temporal structure of the EEG and find the dimensional (frequency) range that best characterizes fatigue. Satisfactory classification results (96%) were obtained on fatigue detection.