Energy Propagation Mechanism for Robust Out-of-Distribution Detection in Industrial Fault Diagnosis

Effective fault diagnosis in construction machinery is essential for sustainable and intelligent building objectives. Traditional deep learning methods struggle with out-of-distribution (OOD) data, affecting model reliability. This study proposes an Energy-Driven Out-of-Distribution Detection (ED-ODD) framework using Graph Neural Networks (GNNs) to enhance fault diagnosis accuracy and robustness. By converting sensor signals into graph data with the maximum information coefficient (MIC), the framework captures complex fault relationships. Energy-based models and a label propagation mechanism effectively distinguish in-distribution from OOD data. Experimental validation on a bearing fault simulation platform showed that the ED-ODD framework outperforms traditional OOD detection methods, especially in handling complex machinery faults. The proposed framework offers a reliable and accurate solution for intelligent fault diagnosis, with future work aimed at optimizing GNN applications for broader industrial use.