Transfer learning in bridge monitoring: Laboratory study on domain adaptation for population-based SHM of multispan continuous girder bridges

The presence of sufficient labelled data associated to various environmental conditions and damage scenarios often represents a challenge for the applicability of supervised-learning methods when dealing with structural health monitoring of real-scale infrastructures. To address this problem, population-based structural health monitoring has been recently proposed as an attractive solution, with the goal to collect information from similar structures and transfer health-state labels across the population. This paper focusses on the use of a feature-based transfer learning method. A machine-learning model is trained with source labelled data in a transformed features space to afterwards classify the unlabelled target dataset of a different bridge. More specifically, a domain adaptation-based methodology proposing two possible strategies, a single-source or a multi-source approach, is described. Given the difficulties in validating these techniques on real and varied datasets from multiple bridges, this paper presents a physical benchmark for population-based structural health monitoring in civil-engineering applications. The transfer between different configurations of a laboratory-scale bridge model, subjected to multiple experimental tests under changing environmental conditions and to the same pseudo-damage scenarios, is investigated. The results of the experimental campaign demonstrate the possibility of effectively exchanging damage labels to perform novelty detection and damage classification across the population via domain adaptation, improve the identification of specific damage classes, as well as to increase model’s outcome using a multi-source approach, thus overcoming the limitations of conventional machine learning-based methods. Furthermore, this paper provides an open dataset with the physical benchmark-related data, allowing other researchers to test their own algorithms and address the various transfer learning challenges.