Fatigue State Prediction of Athletes Based on Multi-Source Sensors

Real-time monitoring of athlete fatigue status is crucial for preventing sports injuries and optimizing training plans. However, existing approaches often suffer from delayed evaluation and limited capability in multi-modal data fusion. To address these challenges, this paper proposes a multisource sensor-based fatigue prediction framework, ST-GCN-MMF, which integrates long-short-term memory generative adaptive networks (LSTM-GAN) and spatial-temporal graph convolutional networks (ST-GCN) for high-accuracy, low-latency fatigue level classification. To tackle the scarcity of annotated fatigue data for professional athletes, we employ an LSTM-GAN to generate multi-modal sequential signals including acceleration, heart rate, and electromyography (EMG), and enhance physiological plausibility through adversarial training. Furthermore, we construct a dynamic graph structure with sensors as nodes and temporal correlations as edges, incorporating dilated causal convolutions to capture cross-modal dependencies between local muscular fatigue and global physiological load. A focal loss function is also introduced to mitigate the issue of imbalanced fatigue level distribution. Experiments conducted on public datasets (SHL, WESAD) as well as a self-collected athlete dataset demonstrate the superiority of our method over existing approaches. This study provides a scalable, low-latency technical path for intelligent fatigue assessment and offers a novel perspective for multi-modal temporal modeling.