Supervised machine learning techniques for predicting multiple damage classes in bridges

The increasing number of bridge collapses has progressively fueled the interest towards the development of monitoring strategies able to ensure real-time damage assessment and preserve structural integrity with correct maintenance. In the context of vibration-based Structural Health Monitoring, recent improvements in sensor technologies and computer science have encouraged the use of Machine Learning algorithms in many engineering fields. In this light, the described methodology proposes to combine Domain Adaptation with supervised learning methods, such as the K-Nearest Neighbors algorithm, in order to correctly assign damage labels to statistically-aligned features. To this aim, natural frequencies gathered during healthy and abnormal conditions are selected as damage-sensitive quantities, bringing a physical meaning and providing information on global structural dynamics. Damage detection results are evaluated and compared before and after Domain Adaptation by employing specific performance metrics. The developed procedure is validated on the Z24 benchmark bridge and the Finite Element Model of the same structure, properly calibrated based on available experimental data. Within the numerical environment, several modal analysis are carried out both assuming pristine conditions and simulating realistic damage scenarios, that involve concrete elastic modulus' reduction for specific bridge elements. The goal of the proposed technique would be to effectively identify and classify different types of damage cases, enabling knowledge transfer among a population of structures and thus representing a prompt engineering decision-support tool to capture damage-induced variations.