Unsupervised transfer learning for structural health monitoring of urban pedestrian bridges

Abstract

Bridge authorities have been reticent to integrate structural health monitoring into their bridge management systems, as they do not have the financial and technical resources to collect long-term monitoring data from every bridge. As bridge authorities normally own huge amount of similar bridges, like the pedestrian ones, the ability to transfer knowledge from one or a small group of well-known bridges to help make more effective decisions in new bridges and environments has gained relevance. In that sense, transfer learning, a subfield of machine learning, offers a novel solution to periodically evaluate the structural condition of all pedestrian bridges using long-term monitoring data from one or more pedestrian bridges. In this paper, the applicability of unsupervised transfer learning is firstly shown on data from numerical models and then on data from two similar pedestrian prestressed concrete bridges. Two domain adaptation techniques are used for transfer learning, where a classifier has access to unlabeled training data (source domain) from a reference bridge (or a small set of reference bridges) and unlabeled monitoring test data (target domain) from another bridge, assuming that both domains are from similar but statistically different distributions. This type of mapping is expected to improve the classification accuracy for the target domain compared to a procedure that does not implement domain adaptation, as a result of reducing distributions mismatch between source and target domains.