Abstract: To address the privacy leakage risks posed by the labeled sensor data collected by wearable fitness action recognition methods and the limitations of traditional centralized model training methods in adapting to new users, this paper proposes a fitness action recognition method based on wearable sports knee sleeve and personalized federated learning. The method achieves both fitness action type and user performance level recognition, enabling users to gain insights into their fitness performance and enhance their exercise outcomes. First, the method treats each user as an independent task and employs a federated approach to meta-train a global embedding network that learns shared representations across users, enabling effective generalization to any user. Then, through an adaptation process, a two-stage fine-tuning of the local classification network is performed on the basis of the global embedding network, generating a personalized model for each user. Finally, through extensive experiments on real-world fitness datasets demonstrate that the proposed method achieves 100% accuracy in fitness action type recognition and 95.94% accuracy in user performance level recognition, significantly outperforming existing state-of-the-art methods. The experimental results indicate that the system not only protects user privacy but also exhibits excellent generalization capability for new users.

摘要： 针对基于可穿戴设备的健身动作识别方法收集标注的传感器数据可能造成隐私泄露风险，以及传统中心化模型训练方法在新用户适应性方面存在局限的问题，本文提出了一种基于可穿戴运动护膝和个性化联邦学习的健身动作识别方法。该方法实现了健身动作类型和用户表现水平识别两项任务，通过帮助用户了解自身健身表现，提高锻炼效果。首先，该方法将每位用户视为独立的任务，以联邦的方式元训练了一个全局嵌入网络学习跨用户的共享表征，此表征可以有效地泛化到任意用户。然后，通过一个适应化流程，在全局嵌入网络的基础上对本地分类网络进行两阶段微调，为每位用户获得个性化模型。最后通过在来自真实世界的健身数据集上进行大量实验，实现了100%的健身动作类型识别准确率和95.94%的用户表现水平识别准确率，显著优于现有先进方法。实验结果表明，本系统在保护用户隐私的同时，还具备良好的新用户泛化能力。