Wearable devices can collect real-time personalized data, including motion, heart rate, and sleep patterns, during sports training. These data are distributed across devices and often contain sensitive information, making traditional centralized training impractical. Federated Learning (FL) offers a promising solution, but existing FL methods in wearable scenarios face challenges such as limited bandwidth, heterogeneous device capabilities, packet loss, and unfair client selection. These factors can reduce model accuracy and compromise training fairness. To address these issues, this study proposes a Fair Communication and Privacy-Preserving Federated Learning model (FPCP-FL). It combines a packet-loss parameter recovery algorithm with a mobility-aware federated aggregation strategy to enable dynamic client selection and parameter restoration, balancing communication efficiency with fairness. Experiments on the Google Speech and Synthetic datasets show that FPCP-FL achieves accuracies of 61.2%, 59.43%, and 58.28% under 10%, 30%, and 50% packet loss, respectively. Compared with Oort, it improves average accuracy by ~ 2.3% and reduces fairness variance by 60%. Under ideal network conditions, 70% accuracy is reached with only 65.9 MB of communication over 69 training rounds. Personalized experimental results further showed that compared with Personalized FL with Moreau Envelopes (pFedMe), pFedMe-FPCP-FL achieved a global accuracy of 71.4% under a 10% packet loss rate, representing a 35.5% improvement. The average accuracy of Fairness-Aware Federated Averaging (q-FedAvg)-FPCP-FL reached 62.79%, with the variance reduced from 1286 to 926 compared with q-FedAvg. The results indicate that FPCP-FL can maintain training accuracy, fairness, and communication efficiency without exposing data, providing a promising approach to personalized modeling in distributed sports training.
Xin et al. (Wed,) studied this question.