Exercise-induced fatigue is one of the critical factors leading to musculoskeletal injuries during sports training and rehabilitation. High levels of fatigue can hinder adaptation during training, inhibit performance and increase risk of injury. Actively monitoring fatigue levels in athletes and rehabilitation patients could provide meaningful feedback needed to adjust training strategy to promote muscle and physical development, therefore preventing injury, and maximizing performance and rehabilitation outcomes. Although a large number of studies attempt to classify human fatigue, the study of automatic movement analysis for fatigue estimation and prediction is still somewhat limited. This thesis aims to develop predictive fatigue monitoring frameworks through machine learning/deep learning techniques.

In this PhD research project, we first focused on monitoring person-dependent fatigue status with continuous feedback. A data-driven approach was investigated to estimate participant-specific fatigue levels based on IMU or force plate data using random forest (RF) and convolutional neural network (CNN) based regression models. This approach was validated in a dataset that includes three different exercises (squats, high knee jacks, and corkscrew toe-touch). A high correlation between the predicted and self-reported fatigue levels were achieved, indicating the ability to monitor fairly small variations of human motion due to fatigue and also the gradual decline in the quality of the movement during exercise execution.

Secondly, we focused on wearable sensor data augmentation and person-independent fatigue estimation. A computer simulation (via OpenSim) based approach was proposed to use collected motion capture data from our existing dataset to generate simulated IMU data associated with different body sizes. Then, a new data-driven model was developed for person-independent fatigue estimation based on experimental data and additional simulated data using current state-of-the-art deep learning techniques, namely, Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), DeepConvLSTM. The findings indicate that deep neural networks can successfully learn the mapping between kinematics data from wearable sensors and self-reported fatigue levels. This demonstrates that simulating IMU data with OpenSim is effective at providing samples for data augmentation and promise for the future use of deep neural networks for rehabilitation in exercise.

Finally, we focused on forecasting future motion to predict future fatigue instead of classifying fatigue status from data that has already been observed. We further developed a novel deep learning-based system that predicts the person-independent kinematic characteristics and associated fatigue status on-line during exercise performance. The proposed framework consists of spatio-temporal attention-based Transformer with an auxiliary critic and a fatigue classifier. In terms of fatigue prediction, a high accuracy and Pearson correlation coefficient were achieved on forecasted motion data with unseen participant data. The proposed approach provided a graduate estimate of fatigue and was validated with simulated and real datasets. The experimental results show that our model can predict fatigue progression and outperforms other state-of-the art techniques. Successfully predicting fatigue progression can help a patient or athlete monitor and adjust their exercise session to prevent overexertion and fatigue-induced injury.