15 / 2025-03-19 15:57:26

Distributed Fault Diagnosis with Data-Based Balanced Weighted Federated Learning

Distributed Systems,fault diagnosis,Federated Learning

This paper presents Data-Based Balanced Weighted Federated Learning(DBW-FLA), a data-based balanced weighted federated learning algorithm, designed to address the challenges of data heterogeneity in distributed fault diagnosis for distributed systems. Traditional federated learning approaches, such as Federated Averaging (FedAvg), often underperform in scenarios with imbalanced data distributions across clients, as they uniformly aggregate local models without considering disparities in client data quality or representativeness. To mitigate this, DBW-FLA introduces a novel client weighting mechanism that evaluates the relative balance of data distributions within each client. By calculating a balance metric based on label-wise data proportions and applying normalized logarithmic weighting, the algorithm prioritizes clients with more balanced and representative datasets during global model aggregation. Experiments conducted on the HIT aero-engine benchmark dataset demonstrate the efficacy of DBW-FLA. Compared to independent models and FedAvg, the proposed method achieves superior diagnostic accuracy and faster convergence across clients. Results highlight DBW-FLA’s robustness in handling data silos and its potential to enhance fault diagnosis in distributed systems while preserving data privacy. This work provides a scalable framework for optimizing federated learning in real-world applications with inherent data imbalance.