Multi-source signal alignment and efficient multi-dimensional feature classification in the application of EEG-based subject-independent drowsiness detection

Abstract

Electroencephalography (EEG)-based driving drowsiness detection in brain computer interface (BCI) systems is an effective way to prevent traffic accidents. Usually a pre-calibration process is required for a specific target because of significant inter-subject variability in EEG signals. Recently transfer learning methods including domain adaptation and deep neural networks are proposed to be applied in subject-independent applications. However, feature extraction in domain adaptation does not consider the multimodal characteristics of EEG signals whereas deep neural networks consume massive computing resources with tremendous parameters. To overcome the limitations, we propose a multi-source signal alignment (MSSA) and multi-dimensional feature classification framework. EEG signals from multiple source subjects are directly aligned with the target subject via one-versus-one minimization of signal covariance matrices. Then the generalized multi-dimensional features are extracted and classified via tensor network (TN). Extensive experiments were conducted in a recently published EEG dataset during a sustained-attention driving task for subject-independent drowsiness detection. Compared with state-of-the-art transfer learning methods, MSSA-TN improves classification accuracy by at least 3.71%, which is promising in developing practical drowsiness detection systems.