A Transfer Learning Application to FEM and Monitoring Data for Supporting the Classification of Structural Condition States

Abstract

One of the main problems concerning the field of Structural Health Monitoring (SHM) is the unavailability of data from different structural conditions. This is especially true for civil structures, where the collection of data from different damage states is often infeasible or economically inconvenient, particularly when dealing with architectural heritage structures. In the last few years, this issue has been addressed by using a Transfer Learning (TL) strategy, which allows one to transfer the knowledge obtained from systems where several conditions are known, to different (but related) systems, for which limited data are available. In particular, recent studies have demonstrated the effectiveness of Domain Adaptation techniques, a subcategory of transfer learning, for both homogeneous and heterogeneous populations. By transferring knowledge, these methods improve the classification of different structural conditions. This paper shows results from the application of a domain adaptation technique - Transfer Component Analysis (TCA) - between the monitoring data of a structure and those of its Finite Element Model (FEM). The FEM is a precious resource for this purpose as it allows one to simulate manifold system conditions and obtain the related data without affecting the real structure. The case study considered here is the Sanctuary of Vicoforte, a monumental building from the 17^th century located in Italy, equipped with a permanent static and dynamic monitoring system. The research has shown promising results in distinguishing, via a Relevance Vector Machine (RVM) classification, different environmental conditions affecting the building.