On the transfer of damage detectors between structures: An experimental case study

Data-driven frameworks are critical within Structural Health Monitoring (SHM) [1] , as well as related fields, including: condition monitoring [2] , non-destructive testing [3,4] , and prognosis and health management [5] . In general terms, machine learning algorithms are used to detect patterns in previously-recorded data, to infer the condition of the system (operational, environmental, or health) given measurements into the future. Health and/or performance predictions are valuable, as they improve maintenance, decision making and efficiency [6] . The development of machine learning methods for health monitoring of engineering systems is prevalent in the literature Incomplete data ? which fail to represent environmental effects or damage ? are a significant challenge for structural health monitoring (SHM). Population-based frameworks offer one solution by considering that information might be shared, in some sense, between similar structures. In this work, the data from a group of aircraft tailplanes are considered collectively, in a shared (more consistent) latent space. As a result, the measurements from one tailplane enable damage detection in another, utilising various pair-wise comparisons within the population. Specifically, Transfer Component Analysis (TCA) is applied to match the normal condition data from different population members. The resulting nonlinear projection leads to a general representation for the normal condition across the population, which informs damage detection via measures of discordancy. The method is applied to a experimental dataset, based on vibration-based laser vibrometer measurements from three tailplanes. By considering the partial datasets together, consistent damage-sensitive features can be defined, leading to an 87% increase in the true positive rate, compared to conventional SHM. ? 2021 Elsevier Ltd. All rights reserved.