A Hierarchical Federated Learning Framework for Sustainable Deployment in Dynamic-Cluster-Based Fault Diagnosis

The inefficiency of fault diagnosis based on federated learning (FL) is attributed to the variability and diversity exhibited by clients in a dynamic cluster. To tackle this issue, a hierarchical federated learning (HFL) framework is proposed, structured in a three-tiered architecture encompassing global clustering, local federation, and global aggregation phases. First, in the global clustering, a dynamic clustering strategy based on client characteristics is designed to divide all clients into multiple subsets and implement local federation. Second, the contribution of local subsets to global aggregation is quantified using Shapley additive explanation (SHAP), and the model parameters are weighted based on their respective contribution values, which aims to alleviate the negative impact of subset model drift on global aggregation. Third, the loss function in model training of clients is redesigned to incorporate two regularization terms, which serve to reconstruct the relationship between new and old samples. This modification avoids full samples training and reduces computation costs. The validation on inter-shaft bearing fault diagnosis dataset of an aero-engine system demonstrates the efficacy of this study in addressing the security joint modeling of a dynamic cluster. This offers a viable solution for achieving sustainable federated deployment within the industrial dynamic clients cluster.