Fatigue Detection from 3D Motion Capture Data Using a Bidirectional GRU with Attention

Exercise-induced fatigue can degrade athletic performance and increase injury risk, yet traditional fatigue assessments often rely on subjective measures. This study proposes an objective fatigue recognition approach using high-fidelity motion capture data and deep learning. This study induced both cognitive and physical fatigue in 50 male participants through a dual task (mental challenge followed by intense exercise) and collected three-dimensional lower-limb joint kinematics and kinetics during vertical jumps. A bidirectional Gate Recurrent Unit (GRU) with an attention mechanism (BiGRU + Attention) was trained to classify pre- vs. post-fatigue states. Five-fold cross-validation was employed for within-sample evaluation, and attention weight analysis provided insight into key fatigue-related movement phases. The BiGRU + Attention model achieved superior performance with 92% classification accuracy and an Area Under Curve (AUC) of 96%, significantly outperforming the single-layer GRU baseline (85% accuracy, AUC 92%). It also exhibited higher recall and fewer missed detections of fatigue. The attention mechanism highlighted critical moments (end of countermovement and landing) associated with fatigue-induced biomechanical changes, enhancing model interpretability. This study collects spatial data and biomechanical data during movement, and uses a bidirectional Gate Recurrent Unit (GRU) model with an attention mechanism to distinguish between non-fatigue states and fatigue states involving both physical and psychological aspects, which holds certain pioneering significance in the field of fatigue state identification. This study lays the foundation for real-time fatigue monitoring systems in sports and rehabilitation, enabling timely interventions to prevent performance decline and injury.