Although a number of studies attempt to classify human fatigue, most models can only identify fatigue after fatigue has already occurred. In this paper, we propose a novel time series approach to forecasting wearable sensor data and associated fatigue progression during exercise. The proposed framework consists of spatio-temporal attention-based Transformer with an auxiliary critic and a fatigue classifier. The Transformer network is used to analyze the person-independent pattern underlying the past kinematic sequence obtained from wearable sensors and generate short term predictions of the human motion. Adversarial training is employed to regularize the Transformer and improve the time series forecasting performance. A fatigue classifier is used to estimate person-independent fatigue levels based on the forecasted wearable sensor data from the Transformer model. The proposed approach is validated with simulated and real squat datasets which were collected from young healthy participants. The proposed network can accurately forecast a time horizon of up to 80 timesteps for motion signal forecasting and fatigue classification. In terms of fatigue prediction, an accuracy of 83% and a Pearson correlation coefficient of 0.92 were achieved on forecasted motion data with unseen participant data. The experimental results show that our model can predict fatigue progression and outperforms other state-of-the-art techniques, achieving 95% correlation compared to 83% for the best performing baseline method. Successfully predicting fatigue progression can help a patient or athlete monitor and adjust their exercise session to prevent overexertion and fatigue-induced injury.
Keywords: Fatigue prediction; Human motion data analysis; Inertial Measurement Unit (IMU); Machine learning.
Copyright © 2022 Elsevier Ltd. All rights reserved.