ABSTRACT
Machine learning stands out as one of the most widely researched fields in the industry, where it has demonstrated great abilities to model complex system dynamics. This paper takes advantage of one of those models, Bayesian networks (BNs), for analyzing such dynamics in the infrastructure field. Specifically, bridge health monitoring tasks, which are crucial to prevent from dangerous changes within the structure’s behavior and ensure the durability of the bridge. However, such modeling requires to capture a substantial amount of data, which can be time-consuming. Considering this challenge, this work aims to accelerate the deployment of these models through related transferable expertise. This involves leveraging data from similar systems to facilitate the learning process of the network. Specifically, two methods, PC-MAX-TL and DBLogLP, are proposed for estimating the structure and parameters, respectively, of the BN under such settings. PC-MAX-TL is an adaptation of the PC-MAX method (a well-known constrained-based structural learning algorithm) tailored to incorporate auxiliary knowledge, while DBLogLP is based on log-linear pooling principles. To demonstrate the reliability of PC-MAX-TL and DBLogLP, an empirical evaluation will be conducted using data from a real scenario: the San Mamés bridge in Bilbao (Spain). This bridge is composed of six different spans that are evaluated independently, and therefore, can be used to address the data scarcity of a certain target span. Then, the results will be compared against the model estimated just with PC-MAX from two perspectives: graphically, looking at the conditional dependencies found in the BNs; and quantitatively, analyzing (1) the network’s mean log-likelihood, and (2) the log-likelihood variance. Thus, we demonstrate the proposed methodology’s effectiveness, and justify for this bridge that we can accelerate in some days the deployment of the machine learning-based monitoring system.
