Robust two stages federated learning for sensor based human activity recognition with label noise

Federated learning is widely used for collaborative training of human activity recognition models across multiple devices with limited local data. However, label noise caused by human and time constraints during data annotation is common and severely limits model performance. Existing studies mainly address this through client selection and sample filtering, but still face key limitations: (1) insufficient granularity in client quality evaluation; (2) aggregation methods ignoring data quality differences; (3) client drift under non-IID data distribution. To overcome these challenges of complex label noise and feature drift, this paper proposes LN-FHAR, a two-stage federated learning framework with label noise robustness. This framework effectively mitigates the coupling problem of noise and data heterogeneity by assessing client data quality and designing differentiated training strategies. In the client selection stage, clients are graded based on class-level loss analysis and a Gaussian Mixture Model. In the noise-robust training stage, reliable neighbors are introduced to collaboratively filter clean samples, and prototype regularization is employed to constrain the consistency between local models and global feature representations. Additionally, a data-aware aggregation method is designed, which assigns weights based on both the quality and quantity of client data, reducing the negative impact of noisy clients. Experimental results demonstrate that LN-FHAR has robustness and generalization ability in complex noise environments.