Worker-independent multimodal physiological fatigue detection in industrial workers using a domain adaptation network with multi-scale maximum mean discrepancy

Occupational fatigue is a critical safety concern in industrial environments, where substantial inter-worker variability in physiological responses can limit the generalization of conventional fatigue-detection models. This study presents a Domain Adaptation Network (DAN) with Multi-Scale Maximum Mean Discrepancy (MMD) for worker-independent multimodal fatigue classification using physiological and environmental measurements collected from 50 industrial workers during occupational activities. The framework integrates convolutional neural network-based feature extraction, bidirectional long short-term memory (BiLSTM) temporal modeling, multi-head self-attention over the temporal sequence, Gradient Reversal Layer (GRL)-based adversarial learning, and Multi-Scale MMD-based domain alignment to learn representations that are more robust to inter-worker variability. A total of 60,000 windowed instances were generated from the synchronized multimodal recordings, using an 85-feature multimodal representation, and fatigue was classified into Low, Moderate, and High categories. A strictly worker-independent evaluation protocol was adopted, with previously unseen workers reserved for testing. Five models were comparatively evaluated: LightGBM, CNN, CNN+BiLSTM, CNN+BiLSTM+Attention, and the proposed DAN, achieving test accuracies of 84.81%, 87.96%, 96.72%, 96.35%, and 97.08%, respectively. The proposed DAN achieved the highest accuracy by combining temporal multimodal representation learning with adversarial and distribution-alignment mechanisms to address inter-worker variability. SHAP and LIME analyses were further employed to provide complementary global and local interpretability of the proposed model’s predictions. Overall, the findings demonstrate the potential of multimodal temporal representation learning combined with domain adaptation for robust and interpretable worker-independent fatigue assessment in industrial environments.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-13
DOI
https://doi.org/10.1038/s41598-026-75814-4
Primary Topic
Sleep and Work-Related Fatigue
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Worker-independent multimodal physiological fatigue detection in industrial workers using a domain adaptation network with multi-scale maximum mean discrepancy

Sanchita Paul, Rupa Verma
1 links
Scientific Reports
Sleep and Work-Related Fatigue
article

Worker-independent multimodal physiological fatigue detection in industrial workers using a domain adaptation network with multi-scale maximum mean discrepancy

Sanchita Paul, Rupa Verma
article en

Abstract

Occupational fatigue is a critical safety concern in industrial environments, where substantial inter-worker variability in physiological responses can limit the generalization of conventional fatigue-detection models. This study presents a Domain Adaptation Network (DAN) with Multi-Scale Maximum Mean Discrepancy (MMD) for worker-independent multimodal fatigue classification using physiological and environmental measurements collected from 50 industrial workers during occupational activities. The framework integrates convolutional neural network-based feature extraction, bidirectional long short-term memory (BiLSTM) temporal modeling, multi-head self-attention over the temporal sequence, Gradient Reversal Layer (GRL)-based adversarial learning, and Multi-Scale MMD-based domain alignment to learn representations that are more robust to inter-worker variability. A total of 60,000 windowed instances were generated from the synchronized multimodal recordings, using an 85-feature multimodal representation, and fatigue was classified into Low, Moderate, and High categories. A strictly worker-independent evaluation protocol was adopted, with previously unseen workers reserved for testing. Five models were comparatively evaluated: LightGBM, CNN, CNN+BiLSTM, CNN+BiLSTM+Attention, and the proposed DAN, achieving test accuracies of 84.81%, 87.96%, 96.72%, 96.35%, and 97.08%, respectively. The proposed DAN achieved the highest accuracy by combining temporal multimodal representation learning with adversarial and distribution-alignment mechanisms to address inter-worker variability. SHAP and LIME analyses were further employed to provide complementary global and local interpretability of the proposed model’s predictions. Overall, the findings demonstrate the potential of multimodal temporal representation learning combined with domain adaptation for robust and interpretable worker-independent fatigue assessment in industrial environments.

Scientific Reports
Birla Institute of Technology, Mesra (IN)
Arthritis National Research Foundation, National Research Foundation, Science and Engineering Research Board
Good health and well-being
Sleep and Work-Related Fatigue
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon