1. Introduction
Unplanned downtime in manufacturing is estimated to cost global industry billions annually, and predictive maintenance offers a route to intervene before failures occur. However, aggregating proprietary sensor data across plants owned by different business units is often contractually restricted, motivating a federated approach.
2. Methodology
Vibration and temperature time-series from 620 rotating assets across four plants were segmented into 10-second windows. A stacked LSTM model was trained locally at each site for one epoch per communication round, with parameters aggregated at a central server using FedAvg over 40 rounds. Performance was benchmarked against a centrally pooled training baseline and a per-plant-only baseline.
3. Results
The federated model reached an F1-score of 0.89 for 48-hour-ahead failure prediction, compared with 0.91 for the centralised baseline and 0.76 for per-plant-only training, confirming that federation recovers most of the accuracy lost to data isolation. Communication overhead per round averaged 3.4 MB per client, well within typical factory network budgets.
4. Conclusion
Federated learning offers a practical middle ground between data privacy and model accuracy for cross-plant predictive maintenance. Future work will explore differential privacy noise injection to further strengthen guarantees against gradient inversion attacks.
References
[1] McMahan B. et al., Communication-efficient learning of deep networks from decentralized data, AISTATS, 2017. [2] Li T. et al., Federated learning: Challenges, methods, and future directions, IEEE Signal Processing Magazine, 2020. [3] Lee J. et al., Machine learning-based predictive maintenance, IEEE Access, 2019. [4] Kairouz P. et al., Advances and open problems in federated learning, Foundations and Trends in ML, 2021.