Hybrid Training of Supervised Machine Learning Algorithms for Damage Identification in Bridges

Hybrid approaches for training machine learning algorithms to identify damage in bridges rely on the use of both monitoring and numerical data. While monitoring data account for normal operational conditions of the undamaged structure, numerical data are often confined to scenarios that seldom occur in the lifespan of the bridge, like extreme temperature events or damage, although previous research of the authors showed it can also be used to augment the data acquired under regular service. This paper presents a hybrid approach for damage identification and applies it to Z-24 Bridge. To enable the classification of damage, supervised learning algorithms are employed. Unlike unsupervised learning, which relies on unassigned data and is suited for novelty detection, supervised learning uses labeled data corresponding to undamaged and damaged scenarios of the structure, enabling the transition from damage detection and localization to damage type and severity. A hybrid database is constructed using monitoring and numerical data corresponding to undamaged scenarios and numerical data corresponding to damage scenarios. The damage scenarios comprise various degrees of settlement of a bridge pier and a landslide near the same pier. Several common supervised learning algorithms are trained with the hybrid data and a comparison of the results is provided.