Transfer Learning between Full-Scale Structure Health Monitoring Systems: Application to Oval Masonry Domes

Structural Health Monitoring (SHM) is a powerful instrument used by engineers to analyze the actual condition of the structure under observation. The data that can be obtained with SHM systems, whether static or dynamic, are many and, as a result, are difficult to manage by hand. For this reason, researchers in recent years have increasingly begun to work with artificial intelligence (AI) for data processing and interpretation. AI attempts to replicate the logical flow of a human being but, unlike a human, it can process much more data and much faster. Various algorithms, such as Machine Learning (ML) algorithms and Transfer Learning (TL) algorithms have already been introduced in the processing of monitoring data. However, these algorithms require as input a labelled dataset, this means that at each data is necessary to associate a label that define the structural conditions (e.g., ‘damaged’ or ‘undamaged’), this is the learning process of the algorithm. Summing up it is possible to say that the model cannot identify a damage if it has never seen a damage. The need of labelled data is an important limit of these models because it is not possible to deliberately damage the structure to obtain data that can be labelled as ‘damaged’, especially if the structure under investigation is part of Cultural Heritage. In this view, TL, particularly the branch of Domain Adaptation algorithms, becomes essential. These models share knowledge among different domains, that can represent homogeneous or heterogeneous structures. Since, not much labelled data on the real structure are available, it is possible to obtain them from a calibrated Finite Element Model(FEM) by running several simulations. In this case, the domain obtained from the FEM is called Source Domain, while the domain of the real structure is the Target Domain. However, these two domains cannot be considered as a unique one, indeed no matter how calibrated the model is it will never represent the reality, indeed “all models are wrong, but some are useful” [George Box]. This leads to having to use domain adaptation algorithms, in such a way to reduce the ‘distance’ between the two domains. Another possibility, in the case of lack of labelled data for the structure under observation, is to use information and data coming from other structures, that are somehow similar to the one of interest. In this view, the structure under observation represents the Target Domain, while all the other structures make up the Source Domain, which will share its knowledge with the Target one. In this thesis a TL algorithm is used to understand if a transfer of knowledge is possible between data of the Sanctuary of Vicoforte and those of the Church of Santa Caterina, two architectural heritage structures both characterized by an oval dome, although of different sizes, and with different plan developments. Therefore, these two structures define a Heterogeneous Population. Specifically, the transfer learning algorithm used is a Domain Adaptation approach for binary and multi-class classification, known as Kernelized-Bayesian-Transfer-Learning. In this view, a first application is developed to identify damages in different elements of two churches, but only at the level of their finite element models. Subsequently, the same algorithm is used to analyze data coming from the monitoring systems on the two real structures, but in this case the classes concern the temperature states instead of damage conditions, because fortunately no damaged data are available.