Federated Learning For Privacy-Preserving Predictive Maintenance: A Comparative Simulation Study Across Distributed Industrial Sites

Abstract

Due to its benefits, predictive maintenance (PdM) systems increasingly use machine learning models trained on sensor data collected across multiple production sites. However, centralizing this data to train models raises concerns about data sovereignty laws, competition, and privacy. An alternative is federated learning (FL), where a model is trained collaboratively across multiple devices, but the training data remains where it is collected. In this study, a multi-site industrial sensor simulation composed of five heterogeneous manufacturing facilities is created, and three widely used baseline classifiers—logistic regression, a random forest, and XGBoost—are compared to a Federated Averaging (FedAvg) multilayer perceptron (MLP) with and without a differential-privacy-style noise mechanism applied to local model updates. Over 30 rounds of communication, the federated model achieved a test accuracy of 92.9–93.1% and an area under the receiver operating characteristic curve (AUC) of about 0.84-0.85, which were near those of the centralized baselines that did not send any raw sensor readings to the federated model. Local-Update noise was measured systematically and showed a clear trade-off between perturbation, which increased the apparent accuracy of the failure class due to imbalance, and accuracy measured by AUC, which showed a decrease in discriminative power. This was consistent with the literature on federated learning. These results validate federated learning as a practical approach for privacy-preserving predictive maintenance pipelines and highlight the importance of carefully designing privacy budgets, handling non-independent and identically distributed (non-IID) data, and selecting aggregation strategies before using federated learning in industry.