Graph neural network-based predictive maintenance algorithm for safe operation of dust removal systems in coal-fired power plants
TL;DR
A novel Graph Neural Network–based Predictive Maintenance (GNN-PdM) algorithm designed to model the spatiotemporal dependencies within dust removal systems for proactive fault prediction and safety optimization is presented.
Abstract
Ensuring the safety and reliability of coal-fired power plants under dynamic operating conditions remains a major industrial challenge. This paper presents a novel Graph Neural Network–based Predictive Maintenance (GNN-PdM) algorithm designed to model the spatiotemporal dependencies within dust removal systems for proactive fault prediction and safety optimization. The proposed framework constructs a dynamic graph representation that captures structural relationships among sensors, actuators, and control nodes, and employs a hybrid message-passing encoder with temporal gating to learn degradation patterns over time. A safety-aware loss function is incorporated to jointly minimize prediction errors and falsealarm probability, enabling a balance between accuracy and operational reliability. Experimental results on real-world plant data show that GNN-PdM achieves a 21.8% reduction in RMSE and a 35.6% decrease in false alarms compared with baseline models, while maintaining a safety reliability index above 96%. These results confirm the algorithm’s ability to enhance predictive precision, ensure emission compliance, and support intelligent decision-making for safe industrial operation. The proposed approach provides a scalable, interpretable, and deployment-ready solution for predictive maintenance in next-generation power systems.
