Model-based data augmentation for user-independent fatigue estimation

Objective: User-independent recognition of exercise-induced fatigue from wearable motion data is challenging, due to inter-participant variability. This study aims to develop algorithms that can accurately estimate fatigue during exercise. Methods: A novel approach for wearable sensor data augmentation was used to generate (via OpenSim) a large corpus of simulated wearable human motion data, based on a small corpus of human motion data measured using optical sensors. Simulated data is generated using detailed kinematic modelling with variations based on human anthropometry datasets. Using both the recorded and generated data, we trained three different neural networks (Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), DeepConvLSTM) to perform personindependent fatigue estimation from wearable motion data. Results: The estimation performance increased with the amount of simulated training data. Accuracy and correlation values were higher with the proposed data augmentation method as compared to other general time series augmentation methods (e.g, rotation, jettering, magnitude wrapping) with the same amount of training data. An accuracy of 87% and a Pearson correlation coefficient of 90% were achieved on unseen data when the DeepConvLSTM model was trained with the proposed augmented dataset. Conclusion: The enlarged dataset significantly improves the prediction of inter-individual fatigue. Significance: Appropriate augmentation techniques for biomechanical data can improve model accuracy and reduce the need for expensive data collection.

Automated summary

This study developed algorithms for accurately estimating fatigue during exercise using novel wearable sensor data augmentation methods. Neural networks trained with the augmented dataset showed high accuracy and correlation values, particularly when using the proposed augmentation method.