UniTF-ADA: A Unified Time-Frequency Framework with Adaptive Loss and Data Augmentations for Robust Driver Fatigue Recognition.

Saved in:

Bibliographic Details
Title:	UniTF-ADA: A Unified Time-Frequency Framework with Adaptive Loss and Data Augmentations for Robust Driver Fatigue Recognition.
Authors:	Liu, Yan¹ 2798323980@qq.com, Liu, Jie² 17359111@qq.com, Yan, Hong³ yanhong@yku.edu.cn, Zhang, Wenyu⁴ zhangwenyu8518@ustl.edu.cn, Liu, Xiaomeng¹ 1312540696@qq.com, Guo, Xu¹ 1114105794@qq.com
Source:	Engineering Letters. Jun2026, Vol. 34 Issue 6, p2401-2410. 10p.
Subjects:	Time-frequency analysis, Data augmentation, Loss functions (Statistics), Electrophysiology, Statistical reliability, Wakefulness
Abstract:	In fatigue driving detection, conventional methods based on video or behavioral features are easily disturbed by lighting, occlusion, and other environmental factors, making stable and reliable recognition difficult. In contrast, electroencephalogram (EEG) signals directly reflect the driver's neural activity and provide stronger anti-interference capability and higher temporal resolution, and thus have become an important data source for fatigue modeling. However, most existing EEGbased approaches still rely on features from a single domain and therefore lack the ability to jointly characterize rhythmic dynamics and structural rhythm changes. Meanwhile, mainstream public datasets contain relatively few samples and fail to adequately capture inter-individual variability. To address these limitations, this paper proposes a unified architecture, UniTFADA, which fuses multi-dimensional fatigue-related features via time-frequency dual-domain modeling. Considering the limited number of samples, we further design a pseudo-distribution generation strategy to expand the training variation space and improve model robustness. In addition, a dynamic weighted multi-objective loss function is introduced to adaptively balance different discrimination targets, stabilizing training and enhancing cross-subject generalization. Experimental results on two public datasets show that, compared with the Deformer baseline, UniTF-ADA improves ACC and Macro-F1 by 1.38% and 1.69% on Dataset I, and by 8.31% and 9.06% on Dataset II, respectively. [ABSTRACT FROM AUTHOR]
	Copyright of Engineering Letters is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	In fatigue driving detection, conventional methods based on video or behavioral features are easily disturbed by lighting, occlusion, and other environmental factors, making stable and reliable recognition difficult. In contrast, electroencephalogram (EEG) signals directly reflect the driver's neural activity and provide stronger anti-interference capability and higher temporal resolution, and thus have become an important data source for fatigue modeling. However, most existing EEGbased approaches still rely on features from a single domain and therefore lack the ability to jointly characterize rhythmic dynamics and structural rhythm changes. Meanwhile, mainstream public datasets contain relatively few samples and fail to adequately capture inter-individual variability. To address these limitations, this paper proposes a unified architecture, UniTFADA, which fuses multi-dimensional fatigue-related features via time-frequency dual-domain modeling. Considering the limited number of samples, we further design a pseudo-distribution generation strategy to expand the training variation space and improve model robustness. In addition, a dynamic weighted multi-objective loss function is introduced to adaptively balance different discrimination targets, stabilizing training and enhancing cross-subject generalization. Experimental results on two public datasets show that, compared with the Deformer baseline, UniTF-ADA improves ACC and Macro-F1 by 1.38% and 1.69% on Dataset I, and by 8.31% and 9.06% on Dataset II, respectively. [ABSTRACT FROM AUTHOR]
ISSN:	1816093X