Transferring damage detection knowledge across rotating machines and framed structures: Harnessing domain adaptation and contrastive learning.

The case dependency of vibration-based Structural Damage Detection (SDD) models on their training structure (data) has always been a setback in their application to new structures with no or limited data. Obtaining labeled data for both non-damage and damage conditions is an expensive and time-consuming process, which is nearly impractical for operational buildings. For Fault Detection (FD) methods in rotating machines, the rotating parts (e.g. , ball bearings) are the monitoring targets. The parts used in FD are less complex pieces to deal with in the laboratory environment, and various damage scenarios can be applied to them in a relatively feasible way. From a Deep Learning (DL) perspective, given the similarity of FD and SDD problems, this study aims to leverage Transfer Learning (TL) across rotating machines and framed structures as the source and target domains, respectively. The TL model must be in a zero-shot setting; otherwise, labeled data is needed for the target structure damage condition. This study uses a unique Domain Adaptation algorithm and a Contrastive Learning (CL) approach that is explicitly designed for performing TL between (rotating machines) FD and (framed structure) SDD tasks. A rotating machine benchmark dataset with over 45 data cases is used to test the proposed methodology, and various CL models are trained across its non-damage and damage conditions. The accumulated knowledge is then applied to two SDD benchmarks for framed structures, the Yellow Frame (YF) and Qatar University Grandstand Simulator (QUGS), through a weighted deep ensemble strategy across the CL models. The method is successfully tested on target structures, and two comparisons are made with a competitive zero-shot Autoencoder(AE)-based model, as well as with the supervised performance, which has access to the target framed structures' damage data. The objective is to surpass the AE-based SDD model performance by leveraging the FD knowledge gained from the rotating machine data and evaluating how close it can get to the supervised performance. On average, the mean F1 score across various damage condition cases is 0.963 and 0.989 for the YF and QUGS, respectively, while the corresponding numbers for the AE-based and supervised models are 0.844 and 0.974 for YF and 0.940 and 0.998 for the QUGS, respectively. In conclusion, the methodology presented in this paper opens a new SDD venue, enabling TL operations among rotating machines and framed structures. Data and codes are available at https://github.com/Hesam-92-19/Machine_To_Structure_TL. [ABSTRACT FROM AUTHOR]