Model-based data augmentation for user-independent fatigue estimation.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Estimating exercise-induced fatigue from wearable motion data is difficult due to individual differences.
• Developing person-independent fatigue estimation algorithms is crucial for effective monitoring.

Purpose of the Study:

• To develop and validate algorithms for accurate, user-independent estimation of exercise-induced fatigue.
• To investigate the efficacy of novel wearable sensor data augmentation techniques.

Main Methods:

• Generated a large corpus of simulated human motion data using OpenSim and kinematic modeling.
• Trained Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and DeepConvLSTM models on augmented and real motion data.
• Compared the proposed data augmentation method against general time series augmentation techniques.

Main Results:

• The proposed data augmentation method significantly improved fatigue estimation performance.
• Achieved 87% accuracy and 90% Pearson correlation coefficient on unseen data using the DeepConvLSTM model.
• Augmented data yielded higher accuracy and correlation compared to general augmentation methods.

Conclusions:

• Enlarging the training dataset through simulation significantly enhances the prediction of inter-individual fatigue.
• Appropriate augmentation techniques for biomechanical data improve model accuracy.
• This approach reduces the need for extensive and costly data collection.