EEG driving fatigue detection based on log-Mel spectrogram and convolutional recurrent neural networks

Gao, Dongrui; Tang, Xue; Wan, Manqing; Huang, Guo; Zhang, Yongqing · 2023 · Crossref

DOI: 10.3389/fnins.2023.1136609

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety issue of driver fatigue, which significantly impairs cognitive function and contributes to traffic accidents. While existing detection methods rely on facial expressions or physiological signals, they are often susceptible to environmental factors like lighting. Electroencephalography (EEG) offers a direct measure of mental state but presents challenges in feature extraction due to signal instability and noise. The authors propose a novel deep learning model, LogMel-CRNN, to accurately classify driver states as either alert or fatigued by leveraging raw EEG signals without extensive manual preprocessing. The methodology involves a two-stage feature extraction process followed by a hybrid neural network architecture. First, the raw EEG signal undergoes a Short Time Fourier Transform (STFT) implemented via a one-dimensional convolutional neural network, rather than traditional Fourier methods. This approach allows for batch processing and trainable parameters. The resulting spectrogram is then converted into a log-Mel spectrogram using a Mel filter bank and logarithmic scaling, a technique adapted from speech recognition to capture perceptually relevant time-frequency features. These log-Mel spectrograms serve as input to the classification model, which consists of a six-layer Convolutional Neural Network (CNN) to extract high-level spatial features, followed by a bi-directional Recurrent Neural Network (Bi-RNN) to capture temporal dependencies. The final classification is performed by a fully connected layer with ReLU activation and a softmax function. Experiments were conducted on publicly available EEG datasets to evaluate the model’s performance. The results demonstrate that the proposed LogMel-CRNN method achieves high stability and accuracy in distinguishing between alert and fatigue states. When compared against four existing state-of-the-art methods, the proposed model yielded the best classification results. The integration of log-Mel spectrograms effectively reduced the dimensionality of the input data while preserving discriminative features, thereby accelerating training and improving generalization. The use of a 1D CNN for STFT computation proved more efficient and adaptable than traditional fixed-window transforms. The significance of this work lies in its effective fusion of time-frequency domain analysis with deep learning for EEG-based fatigue detection. By adapting speech signal processing techniques to EEG data, the study provides a robust framework that overcomes the limitations of shallow models and hand-crafted features. The findings suggest that combining CNNs for spatial feature extraction with Bi-RNNs for temporal modeling, preceded by log-Mel spectrogram transformation, offers a superior approach for real-time driver monitoring systems. This method enhances the reliability of fatigue detection, potentially contributing to reduced traffic accidents and improved road safety.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data