Knowledge Transfer between Oscillators and Real Vibrating Structures to Enrich Dynamic Monitoring Datasets

Abstract

Transfer learning (TL) techniques can be exploited in engineering structures to overcome the main limit of the data-driven approaches for dynamic monitoring, namely the lack of a labelled dataset for some structural configuration of the monitored systems. A variety of methods can be implemented, but those that enable heterogeneous TL based on domain adaptation have proven to be particularly useful, as they allow knowledge to be transferred within a population composed of a wider range of structures. Among them, the kernelized Bayesian transfer learning (KBTL) can be used to improve the knowledge of a less monitored structure exploiting the knowledge of a more monitored one. In this paper the KBTL is assumed to transfer information from an oscillator to a spatial frame and then from a laboratory masonry oscillator to an historical bell tower for which a limited set of dynamic monitoring observations is available.