Physics-Informed Transfer Learning in PBSHM: A Case Study on Experimental Helicopter Blades

Abstract

Data for training Structural Health Monitoring (SHM) systems are often expensive or infeasible to obtain. Population-based SHM, which considers data across a population of structures, presents a potential solution to this issue. However, as differences between structures can lead to differing training and testing distributions, conventional machine learning methods may not generalise between structures. To address this issue, transfer learning (TL) can be used to leverage information across related domains. An important consideration when applying TL is how to asses similarity to identify and extract shared information. In unsupervised TL, a major challenge is that previous data-based metrics are limited to quantifying marginal distribution similarity in the unsupervised setting. This paper proposes utilising the Modal Assurance Criterion (MAC) between modes of healthy structures as a measure of data similarity to identify features that minimise conditional distribution shift. The MAC is incorporated into a feature selection criterion and a TL methodology is proposed. Moreover, the proposed methodology is shown to facilitate label sharing within a heterogeneous population of helicopter blades