FedRepu: A Reputation-Based Aggregation Strategy for Federated Learning in Bearing Fault Diagnosis

Abstract

Integrating federated learning in bearing fault diagnosis marks a crucial step in industrial operations, showing great potential for enhancing diagnostic accuracy and data privacy. However, the effectiveness of federated learning is often compromised by varying data quality and quantity among clients. FedRepu, a novel reputation-based aggregation strategy for federated learning in bearing fault diagnosis, is introduced in this article by employing an extended reputation mechanism that scores client trustworthiness based on real-time, historical, and rewarding performance metrics, dynamically adapting model aggregation to quality of contributions. It deals with natural heterogeneity in distributed sources and improves performance uniformly across clients. Comprehensive experiments on the Case Western Reserve University, Paderborn University, and Huazhong University of Science and Technology datasets, including comparison studies of cross-validation tasks and ablation studies, confirm FedRepu's effectiveness in improving model aggregation efficiency despite reputation discrepancies. Sensitivity analysis further demonstrates its robustness to hyperparameter variations, making FedRepu a promising solution for federated learning in bearing fault diagnosis.