Transfer Learning to Enhance the Damage Detection Performance in Bridges When Using Numerical Models

Abstract

Classifiers based on machine learning algorithms trained through hybrid strategies have been proposed for structural health monitoring (SHM) of bridges. Hybrid strategies use numerical and monitoring data together to improve the learning process of the algorithms. The numerical models, such as finite-element (FE) models, are used for data augmentation based on the assumption of the existence of limited experimental data sets. However, a numerical model might fail in providing reliable data, as its parameters might not share the same underlying operating conditions observed in real situations. Meanwhile, the concept of transfer learning has evolved in SHM, in particular through domain adaptation techniques. The ability to adapt a classifier built on a well-known labeled data set to a new scenario with an unlabeled data set is an opportunity to transit bridge SHM from research to practice. Therefore, this paper proposes an unsupervised transfer learning approach for bridges with a domain adaptation technique, where classifiers are trained only with labeled data generated from FE models (source domain). Then, unlabeled monitoring data (target domain) are used to test the classification performance. As numerical and monitoring data are related to the same bridge, both domains are assumed to have similar statistical distributions, with slight differences caused by the uncertainties inherent to the FE models. The domain adaptation is performed using a transfer knowledge method called transfer component analysis, which transforms damage-sensitive features from the original space to a new one, called latent space, where the differences between feature distributions are reduced. This approach may increase the use of numerical modeling for long-term monitoring, as it overcomes some of the limitations imposed by the calibration process of FE models. The efficiency of this unsupervised approach is illustrated through the classification performance of classifiers built on source data with and without domain adaptation, and using the benchmark data sets from the Z-24 Bridge as the target data.