Federated Learning with Differential Privacy for Predictive Maintenance in Industrial IoT Networks

Abstract

The proliferation of Industrial Internet of Things (IIoT) devices has enabled data
driven predictive maintenance (PdM), yet privacy concerns and data silos across
factories hinder centralized model training. This paper proposes FedDP-PdM, a
federated learning framework incorporating differential privacy (DP) to enable
collaborative failure prediction while preserving sensitive industrial data. Our
approach employs an adaptive clipping mechanism for local model updates and
implements Gaussian noise injection with a dynamic privacy budget allocation
strategy. We introduce a novel client selection algorithm that optimizes for both
model convergence and privacy cost, prioritizing clients with diverse failure
patterns. The framework was evaluated using a digital twin simulation of a
distributed wind turbine network across 12 virtual factories, each containing 50–100
IIoT sensors monitoring vibration, temperature, and acoustic emissions.
Experimental results demonstrate that FedDP-PdM achieves 94.7% prediction
accuracy for bearing failure with a privacy budget of ε=2.0, outperforming non
private federated learning by only 2.1% accuracy reduction while providing formal
privacy guarantees. Comparative analysis shows our method reduces
communication overhead by 38% compared to baseline federated learning and
maintains robustness against membership inference attacks with 89.3% lower
success rate than centralized approaches.