Federated Learning for Privacy-Preserving Predictive Maintenance in Multi-Site Industrial IoT Networks: Architecture, Convergence Analysis and Benchmark Evaluation
Year: 2025
Type: article
Abstract: Predictive maintenance (PdM) in multi-site industrial settings requires training machine learning models on condition monitoring data that is distributed across facilities operated by competing organisations, making centralised data aggregation infeasible due to privacy, intellectual property, and regulatory constraints. Federated learning (FL) offers a paradigm-shift solution by enabling collaborative model training without raw data exchange: each site trains locally and shares only model gradient updates with a central aggregation server. This paper presents FedPdM, a federated learning architecture tailored for industrial IoT predictive maintenance, incorporating: (i) FedAvg-based global aggregation with adaptive client weighting proportional to local dataset quality; (ii) a non-IID robustness mechanism using Scaffold variance reduction to address heterogeneous failure mode distributions across sites; and (iii) differential privacy noise injection at the gradient level to provide formal ε-DP guarantees. FedPdM is evaluated on the CMAPSS turbofan engine degradation benchmark and a proprietary 12 site industrial compressor dataset, achieving Remaining Useful Life prediction RMSE of 14.3 cycles (CMAPSS FD001 subset) comparable to centralised baseline RMSE of 13.8 cycles while providing (ε=2.1, δ=10⁻⁵) differential privacy at a ommunication overhead of 2.3 MB per round. Convergence is achieved in 38 rounds across 12 heterogeneous clients, confirming practical viability for real industrial multi-site deployment.
Source: Aintelia Science Notes
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7204875146