Explainable Federated Ensemble Learning for Industrial Predictive Maintenance

Predictive maintenance (PdM) systems work best with models that perform well both in accuracy and under noisy, imbalanced sensor streams and are suitable for deployment with safety ensured. This paper proposes a novel PdM framework that performs well on both fronts by introducing architecturally modified hybrid CNN-GRU models and principled ensemble fusion. We started from a base CNN-GRU model, and then we designed four innovative models. Another core contribution of this paper is an ensemble architecture. This ensemble model gives better calibration, stability, and recall without sacrificing precision. Decisions of the ensemble model are then rendered via SHAP. We used two industrial datasets; our modified hybrid models and the proposed ensemble model outperform all the baseline models in macro-averaged precision, recall, and F1-score, which demonstrates that our proposed models give superior generalization under sensor noise and class imbalance. After that, a simulated federated learning framework was implemented across the system to enable decentralized data localization and collaborative multi-site training. The core novelty of this paper is the combination of all tailored architectural orchestration to achieve higher accuracy under severe noisy data and with raw-data localization.