PSI - Issue 62

Valentina Giglioni et al. / Procedia Structural Integrity 62 (2024) 887–894 Giglioni et al. / Structural Integrity Procedia 00 (2019) 000–000

888

2

Keywords: Transfer Learning, Bridge monitoring; Structural Health Monitoring; Domain Adaptation

1. Introduction The increasing number of bridge collapses in recent years has brought to light the need to invest in infrastructure management and develop effective monitoring strategies as integration to visual inspections. To this aim, a vast volume of research on Structural Health Monitoring (SHM) has recently focussed the attention on vibration-based systems to provide real-time damage assessment and ensure condition-based maintenance by analysing changes of structural modal properties (Betti (2013), He et al. (2022)). However, one of the main drawbacks of Operational-Modal Analysis (OMA)-based approaches is represented by the lack of labels, especially when dealing with civil infrastructures, for which the collection of damage-related data may be discontinuous or economically infeasible, because of physical constraints or hardware limitations. This fact translates into a critical issue for conventional ML, hampering the applicability of supervised algorithms. To address such concerns, the framework of Population-Based Structural Health Monitoring (PBSHM) has been recently proposed as a holistic solution in the SHM research community, where the goal would be to exploit the information gained from a member within a population to make inferences on a different unlabelled structure via Transfer Learning (TL) strategies (Worden (2020), Bull et al. (2021), Gardner et al. (2021)). The main motivation is clear and arises from the fact that specific labelled data may only be available for a single individual or a restricted number of systems. In recent years, TL theory for PBSHM is starting to move to the field of bridge SHM. Yano et al. (2023) and Gardner et al. (2022) applied, respectively, kernel-based Domain Adaptation (DA) and Statistic Alignment (SA) to transfer damage labels across two real bridges, but limiting the focus to novelty detection. A DA framework is described in Giglioni et al. (2023) for performing damage classification in a population of real bridges and bridge Finite Element Models (FEM). DA, a feature-based approach belonging to the field of transductive TL (Zhuang et al. (2020)), aims to minimise the distance between a source and a target domain by mapping data distributions into a shared feature space, thereby helping a ML classifier to generalise well. The use of DA does not require any target label, becoming particularly attractive in civil engineering SHM applications. Given that, the usefulness of such approaches to PBSHM is quite clear, since the target domain can be associated to a monitored structure for which labelled data are incomplete or absent. Beyond that, the present paper contributes to the literature by showing a new application of DA from a different perspective, pointing the interest to facilitate and improve damage assessment in a single multi-span bridge. Specifically, if the source domain includes labelled data describing damage in a certain span, the goal would be to identify the same kind of damage affecting a different span via feature transformations, by consequently applying SA and kernel-based DA methods, as suggested in Giglioni et al. (2023). It is worth underlining that one of the main challenges of PBSHM is the difficulty to implement and validate relevant and robust technologies using real datasets from multiple similar structures with different health-states. In this light, the described DA-based procedure is tested using real data from an experimental campaign, accurately presented in Giglioni et al. (2023). Precisely, the position of the intermediate piers of a mock-up bridge is shifted to yield four different bridge configurations, characterised by different layers on the deck surface. A comprehensive dataset is built up by simulating various environmental conditions and damage scenarios. Results show that the information on the effect of structural-stiffness reduction on modal parameters, which is simulated by applying masses on specific portions of the deck, can be successfully transferred across spans, proving that the adopted DA-based method is promising to support bridge SHM when managing massive infrastructures. 2. Domain Adaptation for knowledge transfer across spans The study of TL (Pan and Yang (2009)), defined as a subfield of ML, is motivated by the necessity to transfer the acquired information from a labelled source domain to solve new problems and new tasks in an unknown target domain. As a branch of TL, DA is recommended to be applied in those cases where labelled data are available only in the source domain, with the aim to reduce the distance between the two data distributions by applying proper feature transformations. To provide some basic definitions, a domain = { , ( )} includes a feature space and a marginal