Zero-shot transfer learning for structural damage detection using target-to-source structure domain data mapping

Abstract

In the field of structural health monitoring (SHM), transferring damage detection knowledge, e.g., parametric models trained with structural damage across different structures or damage types (namely, domains), is pivotal in addressing the reliance on gathering labeled data from test (target) structures. Recently, powerful Deep Learning models in conjunction with Transfer Learning (TL) have been explored to accommodate the process; however they still show limited generalizability across different types of structures. This paper introduces a novel approach for zero-shot TL in SHM, leveraging autoencoders to facilitate structural damage detection in the presence of limited training data. The proposed approach employs an autoencoder that maps undamaged data from the target domain to the undamaged data of the source domain. This mapping enables a damage detection model, trained exclusively on the source domain, to effectively identify anomalies in the transformed target domain data without requiring additional training or labeled target data. This unique framework allows the autoencoder to effectively capture the underlying structural characteristics by learning to map the target domain to the source domain, thereby facilitating knowledge transfer. Training an autoencoder to map from undamaged data in the target structure to undamaged data in the source structure also transfers knowledge related to damaged data. Once the autoencoder has experienced this mapping, it is leveraged to the source structure to detect any damage in the target structure. To diagnose damage, a trained one-class support vector machine is used on the source structure to identify any anomalies in the target structure. The resulting outcome of two benchmark problems underscores the efficacy of the proposed method in accurately reconstructing source domain data from target domain inputs, thereby demonstrating its potential to enhance structural damage detection using limited training data. Moreover, for the studied cases, the adaptability of the proposed approach to different structural benchmark types demonstrates a strong ability to surpass typical structural similarity requirements in TL-SHM applications. Thus, with further experiments and development, it can support the creation of SHM tools suitable for large-scale adoption.

Topics & Concepts