By monitoring workers' vital signs and movements in real-time, researchers have developed a powerful system to predict fatigue, offering companies a cutting-edge solution for reducing injuries and improving job performance.
Study: Wearable network for multilevel physical fatigue prediction in manufacturing workers. Image Credit: UNIKYLUCKK / Shutterstock
In a recent study published in the journal PNAS Nexus, researchers explored using multimodal wearable sensors combined with machine learning to measure real-time fatigue among manufacturing workers. Their findings provide important insights into the physical challenges of factory work through monitoring vital signs and motion, with implications for improving work conditions and productivity.
Background
There are high costs of fatigue in the manufacturing industry, with the cost of health-related productivity loss estimated to be $136 billion per year in the United States. High levels of fatigue have also been reported among workers in Sweden, Japan, the European Union, and Canada, with 90% of shift workers reporting regular fatigue, sleepiness, higher injury risk, accidents, and conditions such as chronic fatigue syndrome or musculoskeletal disorders.
Fatigue is difficult to monitor as it has no universal biomarkers. While standard assessment tools focus on physical posture to measure it, they may not fully capture the signs of combined musculoskeletal strain and exhaustion. The oversimplification of fatigue into binary states misses nuances in levels of physical exertion.
Wrist-based wearable devices or lower-limb sensors have been used to record fatigue. However, they are limited by concerns around privacy and rely on the assumption that tasks will involve walking. There is a lack of multimodal and practical sensing systems suitable for factory work environments.
About the Study
In this study, researchers viewed fatigue as a continuous variable, using multimodal sensing to capture kinematic and physiological signs. Fatigue was measured on a scale of 0 to 10 for perceived levels of exertion, which is often used in sports science. A custom asymmetric loss function was also deployed in the machine-learning model to reduce the impact of underprediction errors, providing a deeper understanding.
The participants in the study comprised 43 workers involved in two manufacturing tasks: Composite Sheet Layup and Wire Harnessing. They wore soft wearable sensors that monitored movement and vital signs, including skin temperature and heart rate in real-time and continuously. The sensors were also designed to be skin-compatible, unobtrusive, and flexible.
Researchers developed data analytics and visualization products that can be used to predict fatigue and provide guidance to companies and workers.
Findings
The participants were aged 18-56, and 23.7% were female. They were given two tasks, and data was collected over 18 months. The tasks were designed to mimic real-world manufacturing settings with repeated steps to induce fatigue, including a "Composite Sheet Layup" task, which required workers to place and smooth carbon fiber sheets, and a "Wire Harnessing" task, which involved fastening zip ties around cable systems. Participants also wore weighted vests to increase the exertion they felt.
The models found that different people experience different physiological signs of fatigue, which may also shift over time for the same individual. The results showed that fatigue led to drops in performance scores after repeated rounds of tasks, indicating the negative impacts of fatigue on job performance.
Performance initially improved due to task learning but declined as fatigue increased. Variation in scores between participants also increased as fatigue set in, showing individual differences in fatigue resistance.
Movements of the non-dominant arm, especially in tasks with synchronized movements, were found to be critical in predicting fatigue, along with physiological signs like heart rate. Specifically, maximum heart rate and left arm movement were among the top contributors to fatigue prediction.
Feedback from users working in factory settings regarding the measurement devices suggested that the sensors were generally seen as unobtrusive and that the technology could lead to improvements in working conditions. Surveys also showed positive responses for comfort and ease of use, with participants accepting data tracking and reporting low hindrances to tasks as a result of the sensor.
Conclusions
Fatigue is common among manufacturing workers, leading to higher injury risks, lower productivity, and health problems. Improving ergonomics and mitigating fatigue is critical, but current challenges include the lack of unobtrusive fatigue-sensing methods and reliable biomarkers.
This study collected real-world data from 43 subjects performing two tasks in a manufacturing setting. The goal was to model fatigue as a continuous variable, improving on previous methods that only classified workers as “fatigued” or “not fatigued.”
Fatigue modeling is complex, influenced by noisy self-reports, and varying across individuals and tasks. For example, locomotive features were more important for tasks with synchronized movements. The researchers also adopted a custom asymmetric loss function, which prioritized reducing errors where the model underpredicted the level of fatigue, as this was deemed more critical for practical applications. Performance drops when the model is tested on new individuals not included in the training data, highlighting individual variability.
The system monitors worker fatigue using wearable sensors, providing real-time feedback via a visual dashboard. It was tested successfully in a factory, and workers gave positive feedback about its ease of use and utility.
While the system aims to improve worker safety and empower employees, deploying such systems in workplaces raises ethical and legal concerns. For example, self-reported fatigue scores can sometimes be biased, as workers may underreport fatigue due to concerns over administrative attention. The researchers hope their work will drive discussions on the responsible use of these technologies. The dataset has been made publicly available for future research.
Journal reference:
- Mohapatra, P., Aravind, V., Bisram, M., Lee, Y., Jeong, H., Jinkins, K., Gardner, R., Streamer, J., Bowers, B., Cavuoto, L., Banks, A., Xu, S., Rogers, J., Cao, J., Zhu, Q., & Guo, P. (2024). Wearable network for multilevel physical fatigue prediction in manufacturing workers. PNAS Nexus, 3(10). DOI: 10.1093/pnasnexus/pgae421, https://academic.oup.com/pnasnexus/article/3/10/pgae421/7815440