Energy-efficient graph neural networks for predictive maintenance of interconnected rotating machinery in smart industrial environments

Abstract Predictive maintenance is essential for improving reliability and reducing downtime in smart industrial environments. However, most existing methods either ignore the interdependencies among machines or incur high computational and energy costs, which limits their deployment on edge and fog platforms. This study uniquely proposes an energy-aware graph neural network that jointly optimizes relational learning and energy efficiency for deployment-constrained industrial systems. The framework models interconnected rotating machinery as a dynamic graph, where nodes represent components and edges capture physical and operational dependencies. An adaptive message-passing mechanism with neighborhood sampling and energy-aware training reduces redundant computations while preserving critical system interactions. Experimental evaluation on the CWRU bearing dataset shows that the proposed model achieves 98.4% fault detection accuracy and 98.1% F1-score, with a remaining useful life prediction error of 9.6 h. In addition, the model reduces normalized energy consumption by up to 40% and inference latency by 24% compared with standard graph neural networks. These results demonstrate that the proposed approach provides an effective balance between predictive performance and energy efficiency, making it suitable for real-time predictive maintenance in Industry 4.0 environments.