Federated hybrid ARIMAX-LSTM for Collaborative Fan Fault Prognostics: A Cement Plant Case Study

Ensuring the reliability of critical industrial assets is essential in Industry 4.0. However, centralized predictive maintenance (PM) approaches face major challenges related to data privacy and scalability across distributed sites. To address these limitations, we propose a novel fault prognostics framework that integrates a hybrid ARIMAX–Long Short-Term Memory (LSTM) model within a Federated Learning (FL) architecture. The ARIMAX component models linear dependencies and exogenous effects, while the LSTM captures nonlinear residual patterns. Importantly, only the LSTM parameters are collaboratively trained using FL, ensuring that raw operational data remains local. Experiments conducted on real-world data from a cement plant demonstrate strong predictive performance—evidenced by low MSE and MAE values and high R2 scores—along with stable convergence across clients. This work demonstrates the effectiveness of FL for privacy-preserving, scalable, and accurate time-series-based PM in complex industrial settings.