Transfer learning and Bayesian calibration addressing data scarcity and uncertainty for structural health monitoring of twin concrete bridges

This paper applies transfer learning in the context of structural health monitoring (SHM) to two almost identical bridges located side-by-side, whose construction dates are separated by almost three decades. The uniqueness of this study is enhanced by the fact that the newer bridge has been reported as damaged for almost one decade, with no monitoring data available from its undamaged condition. To overcome data scarcity and uncertainty in the training of machine learning algorithms, this paper proposes a multidisciplinary framework to reuse monitoring data in the undamaged condition from the older bridge to address damage detection in the new one. A numerical model solved by the finite element method is developed to simulate the undamaged condition of the new bridge. The model is calibrated to account for sources of epistemic uncertainty using Bayesian inference through Markov-Chain Monte Carlo simulations with the Metropolis–Hastings algorithm. During the Bayesian updating process, a global sensitivity analysis using Sobol indices is proposed to identify the main parameters influencing the model’s outputs. The results show that the numerical model is capable of simulating the dynamics of the new bridge in its undamaged condition, and transfer learning through domain adaptation is capable of adapting the data from the old bridge so that it can be reused to train a machine learning algorithm to classify observations from the new bridge, taking into account random uncertainty. This framework provides substantial benefits in addressing data scarcity and uncertainty, model updating, and machine learning challenges in the context of SHM but also reveals some limitations of unsupervised transfer learning.