Electroencephalogram-Based Multi-Class Driver Fatigue Detection using Power Spectral Density and Lightweight Convolutional Neural Networks

Suprihatiningsih, Wiwit; Romahadi, Dedik; Feleke, Aberham Genetu · 2025 · Crossref

DOI: 10.5614/j.eng.technol.sci.2025.57.4.2

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 public health issue of driver fatigue, a primary contributor to global traffic accidents, by proposing an efficient method for multi-class fatigue detection using electroencephalogram (EEG) signals. While EEG is a reliable indicator of cognitive state, existing methods often rely on complex deep neural networks that require significant computational resources and struggle with feature redundancy. The authors aim to optimize this process by combining Power Spectral Density (PSD) feature extraction with a lightweight Convolutional Neural Network (CNN) to reduce computational cost while maintaining high accuracy in classifying driver states as awake, tired, or drowsy. The research utilized the SEED-VIG dataset, which contains EEG data from 23 participants engaged in driving simulations. Due to data imbalance and completeness issues, the final analysis focused on eight participants, resulting in 7,080 segments labeled based on PERCLOS values. The methodology involved four stages: preprocessing raw EEG signals using band-pass filtering and FastICA for artifact removal; extracting PSD features across five frequency bands (delta, theta, alpha, beta, gamma) using Welch’s method; and applying statistical functions and energy ratios to generate 493 initial features. To handle class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was employed. Feature selection was performed using Chi-square and ReliefF algorithms to identify the most discriminative features. Finally, a lightweight CNN model was trained for 100 epochs to classify the data in both intra-subject and cross-subject scenarios. The results demonstrated that feature optimization significantly impacted model performance. The ReliefF algorithm yielded the highest intra-subject accuracy of 71.01%, outperforming the Chi-square method, which achieved 70.87%. In cross-subject classification, where the model generalizes across different individuals, the proposed lightweight CNN achieved an average accuracy of 69.07%. Detailed performance metrics revealed that the model performed best in detecting the "drowsy" class, with a specificity of 91.26% and an Area Under the Curve (AUC) value near 90%. The "tired" class showed the weakest performance. The study also noted that while the model was robust, individual variations in brain signals caused accuracy fluctuations, particularly in specific subjects like S14. The significance of this work lies in its demonstration that a lightweight CNN, combined with optimized PSD features, can effectively detect multi-class driver fatigue without the heavy computational burden of deeper networks. By reducing feature dimensionality through ranking algorithms like ReliefF, the approach offers a viable solution for real-world Advanced Driver Assistance Systems (ADAS). The findings suggest that while cross-subject generalization remains challenging due to inter-individual variability, the proposed method provides a balanced trade-off between accuracy and efficiency, making it suitable for practical implementation in safety-critical driving environments.

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