Automated and Continuous Fatigue Monitoring in Construction Workers Using Forearm EMG and IMU Wearable Sensors and Recurrent Neural Network

Abstract

About 40% of the US construction workforce experiences high-level fatigue, which leads to poor judgment, increased risk of injuries, a decrease in productivity, and a lower quality of work. Therefore, it is essential to monitor fatigue to reduce its adverse effects and prevent long-term health problems. However, since fatigue demonstrates itself in several complex processes, there is no single standard measurement method for fatigue detection. This study aims to develop a system for continuous workers' fatigue monitoring by predicting the aerobic fatigue threshold (AFT) using forearm muscle activity and motion data. The proposed system consists of five modules: Data acquisition, activity recognition, oxygen uptake prediction, maximum aerobic capacity (MAC) estimation, and continuous AFT monitoring. The proposed system was evaluated on the participants performing fourteen scaffold-building activities. The results show that the AFT features have achieved a higher accuracy of 92.31% in assessing the workers' fatigue level compared to heart rate (51.28%) and percentage heart rate reserve (50.43%) features. Moreover, the overall performance of the proposed system on unseen data using average two-min AFT features was 76.74%. The study validates the feasibility of using forearm muscle activity and motion data to workers' fatigue levels continuously.

Keywords: activity recognition; aerobic fatigue threshold; construction labor shortage; fatigue monitoring; muscle activity; oxygen prediction; scaffold building; wearable sensor; work-related musculoskeletal disorders.

Figure 1

Work-related causes of physical fatigue.

Figure 2

Physical fatigue measurement methods.

Figure 3

Proposed fatigue monitoring framework.

Figure 4

The overall architecture of the proposed BiLSTM model for activity recognition and oxygen prediction.

Figure 5

A participant wearing a forearm Myo armband and metabolic analyzer.

Figure 6

Data analysis protocol to evaluate the feasibility and performance of using the aerobic fatigue threshold for fatigue monitoring.

Figure 7

Data analysis protocol to evaluate the feasibility and performance of the proposed fatigue monitoring system.

Figure 8

Average AFT for each activity.

Figure 9

Confusion matrix for decision tree classifier using AFT features.

Figure 10

Actual and predicted AFT for every one second on the unseen dataset.

Figure 11

Actual and predicted AFT for every one-min on the unseen dataset.

Figure 12

Actual and predicted AFT for every two-min on the unseen dataset.

Figure 13

Actual and predicted AFT for each activity on the unseen dataset.

Figure 14

Correlation analysis for actual and predicted AFT values on the unseen dataset for (a) every one second, (b) average of one-min, (c) average of two-min, and (d) average for each activity.

Figure 15

Confusion matrix of fatigue level assessment using predicted AFT features (a) for one-min, (b) for two-min, and (c) for each activity.

Figure 16

Predicted and actual fatigue level for every one-min on the unseen dataset.

Figure 17

Predicted and actual fatigue level for every two-min on the unseen dataset.

Figure 18

Predicted and actual fatigue level for each activity on the unseen dataset.