3D-STCNN: Spatiotemporal Convolutional Neural Network Based on EEG 3D Features for Detecting Driving Fatigue

Peng, Bo; Gao, Dongrui; Wang, Manqing; Zhang, Yongqing · 2023 · Crossref

DOI: 10.47852/bonviewjdsis3202983

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the critical safety issue of driving fatigue, a leading cause of traffic accidents, by proposing a novel deep learning framework for detecting driver mental states using electroencephalogram (EEG) signals. The authors identify limitations in existing methods, such as reliance on single-feature extraction and unstable model performance across varying fatigue levels. To overcome these challenges, the study introduces a Spatiotemporal Convolutional Neural Network (STCNN) that leverages a newly constructed three-dimensional (3D) feature representation of EEG data, integrating frequency, time, and spatial domain information to enhance classification accuracy and robustness. The methodology involves a comprehensive data collection process and a specialized neural network architecture. The researchers conducted fatigue driving experiments with 14 subjects, collecting EEG signals under three different driving task loads to capture both awake and fatigued states. The raw EEG data, recorded from 31 channels, were segmented into 10-second windows and filtered into five frequency bands: Delta, Theta, Alpha, Beta, and Gamma. Power Spectral Density (PSD) features were extracted and mapped onto a 2D plane based on electrode spatial distribution, then stacked to form 3D feature cubes. The STCNN model consists of two parallel branches: an Attention Time Network (ATNet) and an Attention-based Convolutional Neural Network (ACNN). ATNet processes 1D temporal features using channel and spatial attention modules combined with Bidirectional Long Short-Term Memory (BiLSTM) networks to capture temporal dependencies. ACNN processes the 3D spatiotemporal features using an attention mechanism and a ConvNeXt-based CNN to extract spatial and spectral characteristics. The outputs from both branches are fused for final classification. Experimental results demonstrate the effectiveness of the proposed approach. The STCNN achieved a classification accuracy of 87.55% in distinguishing between awake and fatigued states. The study compared this performance against six competitive methods, showing that the integration of 3D features and the dual-branch attention mechanism significantly improves detection reliability. The attention modules effectively weighted important frequency bands and spatial regions, allowing the model to discard irrelevant information and focus on discriminative features. Furthermore, the extensive experiments validated the model's stability across different levels of fatigue and varying frequency band combinations, addressing the instability issues common in previous models. The significance of this work lies in its contribution to more reliable and accurate non-invasive fatigue detection systems. By effectively combining multidimensional EEG features with advanced deep learning architectures, the proposed STCNN offers a robust solution for real-time monitoring of driver mental states. This approach has strong potential for application in intelligent transportation systems, where timely fatigue warnings can prevent accidents and enhance road safety. The study underscores the value of integrating spatial, temporal, and spectral information in EEG analysis, providing a template for future research in physiological signal processing and driver monitoring technologies.

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