Detection of Bolt Loosening Using Acoustic Emission Signal and Domain-Generalized Machine Learning Method

Bolted joint structures are critical fastening components across various engineering applications, and the ability to monitor their contact status is crucial for effective structural health monitoring (SHM). The acoustic emission (AE) technique combined with deep learning (DL) methods has been extensively applied in bolt looseness monitoring. Current DL methods assume that the data distribution remains consistent between training and testing datasets. In fact, the surface contact state and the resulting AE signal will be different after each assembly. To address the domain shifts caused by variations in surface contact states and AE signal characteristics across different assemblies, this paper presents a domain-generalized framework using acoustic emission (DGFAE) for bolt looseness diagnosis without requiring prior access to target domain data. The framework integrates a compound loss function capturing the ordinal progression of bolt loosening and employs deep correlation alignment (Deep CORAL) to enhance feature alignment across domains. The effectiveness of the DGFAE method is validated using the “ORION-AE” dataset, with ablation experiments and comparative analysis against other domain generalization (DG) techniques. Compared to state-of-the-art DG methods, superior diagnostic accuracy is achieved under unseen target conditions. Furthermore, a pseudo- DG scenario is explored, where partial healthy samples from the target domain are assumed to be accessible, and the Mixup augmentation technique is integrated to further improve generalization robustness. The diagnostic results confirm that the proposed DGFAE method provides a practical and effective solution for bolt looseness monitoring in real-world engineering settings.