A Multi-Model Analysis for Driving Fatigue Detection using EEG Signals

Abstract

Electroencephalographic (EEG) technology's non-invasive, inexpensive, and potable qualities have recently increased interest in EEG-based driving fatigue detection. EEG signals have been one of the most accurate and reliable markers of driver fatigue. Despite this, extracting valuable features from cluttered EEG signals still difficult to detect driving fatigue. This study aims to create a novel real-time methodology for detecting driving fatigue based on EEG signals. The study utilizes the Discrete Wavelet Transform (DWT) to obtain different EEG bands and compute power spectrum density (PSD) and other statistical features over each DWT band for the online detection of mental fatigue. Deep learning, particularly convolutional neural networks (CNN), has demonstrated impressive results in recent years as a method to extract features from EEG signals among various analysis techniques successfully. Although automatic feature extraction and accurate classification are advantages of deep learning, designing the network structure can be challenging and requires a vast amount of prior knowledge. Therefore, we used these features as input to CNN instead of using raw EEG data directly. Classification results of multiple machine learning models such as Support Vector Machine (SVM), k-nearest neighbor (kNN), Linear discriminant analysis (LDA), Decision Tree (DT), and Naive Bayes (NB) classifiers are also explored to obtain an optimum solution of the driver's fatigue evaluation. Two driving fatigue EEG datasets were used as testbeds to denote the effectiveness of five conventional classifiers and CNN. The proposed method reached more than 99% classification accuracy using a kNN and CNN in both datasets. The outcomes confirmed the efficacy of the suggested approach.