A Federated Learning Framework for Data-Sovereign Predictive Maintenance in Distributed Smart Manufacturing

Predictive maintenance enables early detection of machine failures and reduces unexpected production downtime. However, conventional approaches typically rely on centralized data collection and model training which introduce challenges related to data sovereignty, communication overhead and data ownership. To address these challenges, this research proposes a collaborative federated learning framework for predictive maintenance that can be deployed in distributed smart manufacturing systems. The proposed data-sovereign federated learning approach allows multiple factories to collaboratively train a machine failure prediction model while maintaining data locality. In the framework, each factory trains a local multilayer perceptron (MLP) model using its own machine operational data, while a central server aggregates local model parameters using the Federated Averaging (FedAvg) algorithm to construct a global predictive model. The proposed framework was evaluated using the publicly available AI4I 2020 predictive maintenance dataset, where multiple factories are simulated by partitioning the dataset into distributed clients. Experimental results show that the federated learning model achieves competitive performance compared to centralized machine learning baselines, attaining an accuracy of 97.17%, precision of 0.6000, recall of 0.5000, and F1-score of 0.5455. These results demonstrate that federated learning can enable effective predictive maintenance while maintaining data sovereignty in distributed manufacturing environments.