Fatigue detection in performance sports with multivariate time series analysis: integrated use of wavelet transform and transformer models

Abstract

In performance sports, striking the right balance between training load and adequate rest is crucial to prevent athlete overload, fatigue, and the resulting negative health effects. In this context, the proposed approach integrates multivariate time series data and artificial intelligence techniques to accurately predict athletes' fatigue levels, contributing to better overall health monitoring. This method aims to optimize athletes' performance management while ensuring healthy recovery by providing a more precise and objective assessment compared to traditional monitoring approaches. Time series data collected by IMU devices were transformed into 2D images using wavelet transform techniques (CMT and FCWT) and then trained with transformer models (DeiT3 and Swin). In model training, four different feature sets (A: 'CMT-based DeiT3', B: 'FCWT-based DeiT3', C: 'CMT-based Swin', D: 'FCWT-based Swin') were used to extract the feature sets of the top three best performing models, and a feature fusion technique was applied. Four new sets (A&B, A&C, B&C, A&B&C) were created by feature fusion, and the best performing set of 1536 features (A&B) was next analyzed using feature selection methods (Chi2, mRMR, Relief). As a post-process, the best 100, 500, and 1000 features were selected from this set with Chi2, mRMR, and Relief methods. The highest performance was obtained with the first 500 features selected with the Relief method. These features were classified by the SVM method, and an overall accuracy of 97.25% was achieved. Features selected using the Relief method were reclassified using SVM with cross-validation (k = 5) and achieved an overall accuracy of 97.54%.