Industrial Worker Fatigue Detection System: Multimodal Physiological Signals and Attention Deep Learning Model

This article introduces an industrial worker fatigue detection system based on multimodal physiological signals such as EEG, ECG, and GSR, using the TAN (Temporal Attention Network) deep learning architecture to achieve a detection accuracy of 94.9%. It aims to provide an intelligent solution for industrial safety management. The system covers the complete process from signal collection to state prediction, combining wearable devices and AI technology to address the issues of insufficient real-time performance and accuracy in traditional fatigue detection methods.

In high-risk industries such as manufacturing, mining, and transportation, worker fatigue is a major cause of accidents, resulting in billions of dollars in losses and numerous casualties each year. Traditional methods rely on subjective reports or regular breaks, making it difficult to assess physiological states in real time and accurately. With the development of wearable devices and AI technology, real-time detection based on physiological signals has become possible.

Data Collection: Wearable devices are used to collect multimodal signals, including physiological signals (ECG, GSR, EEG, body temperature, HR, SpO2) and environmental signals (noise, dust density), with annotations based on the Multidimensional Fatigue Inventory (MFI) questionnaire.

Processing Flow: 1. Data loading and annotation (mapping worker IDs, filling missing values); 2. Sliding window feature extraction (200-second window, 10-second step size, extracting 22 features including ECG/GSR/EEG); 3. Preprocessing (dividing test sets by worker level, binarizing labels, standardization, sequence construction, setting class weights).

Traditional ML Models: Random Forest, SVM, Logistic Regression (all using class balance strategies).

Data Augmentation: Conditional Generative Adversarial Networks (cGAN) are used to generate synthetic non-fatigue samples to address class imbalance.

Deep Learning Models: Baseline LSTM, TAN v1 (LSTM + Self-Attention), TAN v2 (LSTM + Self-Attention + General Attention), cGAN + LSTM.

Attention Mechanism: Self-attention identifies key time points, while general attention captures relationships between time steps; TAN v1 significantly improves performance through self-attention.

On the independent test set, TAN v1 (LSTM + Self-Attention) achieved an accuracy of 94.91% and an F1 score of 0.9603, which is significantly better than the baseline LSTM (88.59%) and traditional ML models (83-87%).

Key Findings: 1. The attention mechanism improves performance; 2. Self-attention is better than dual attention; 3. cGAN data augmentation is effective; 4. All models have high recall rates (over 90%) and low missed detection rates, making them suitable for safety scenarios.

Application Scenarios: Manufacturing production lines, mining, long-distance transportation, medical monitoring, etc.

Limitations: Large individual differences, environmental interference, privacy issues, and real-time performance to be optimized.

Future Improvements: Transfer learning for personalized calibration, edge computing to enhance real-time performance, multimodal fusion (video signals), and fatigue trend prediction and early warning.

This open-source system demonstrates the application potential of multimodal physiological signals and deep learning in industrial safety. The 94.9% accuracy provides technical support for safety management. The open-source design facilitates reproduction and expansion. With future technological developments, it will protect workers' health and safety in more fields.