A Rapid Machine Learning-Based Damage Detection Technique for Detecting Local Damages in Reinforced Concrete Bridges

Recent earthquakes illustrated that damage to highway bridges, as one of the critical components of transportation systems, could lead to irreparable social and economic losses. Identifying the extent of probable damage to primary and secondary components of the bridges in the shortest time after a seismic event can significantly reduce the losses related to delays in emergency responses. This paper presented a rapid Machine Learning (ML)-based damage detection framework to detect the extent of damage exposed to primary and secondary components of reinforced concrete bridges under earthquake motions. The proposed algorithm used a set of 160 pair of records, originally developed for seismic performance evaluation of highway bridges, to generate a generalized dataset. These records were uniformly scaled to 16 different peak ground acceleration values ranging from 0.05 g to 1.6 g to produce a wide range of ground motion intensities. This study applied damage indicators extracted from the acceleration time histories as the input attributes to the ML algorithm. The acceleration signals were polluted to a maximum level of 10% noise to simulate the field condition. Two different bridge models (i.e. a two-dimensional multi-span concrete box-girder and a three-dimensional multi-span continuous I-girder bridge) were used to validate the proposed technique. A parametric study was implemented to determine the most efficient ML algorithm for determining the level of damage in the bridge's components. Bayesian Optimization (BO) algorithm was conducted to tune the hyperparameters of each ML algorithm. Results indicate that Support Vector Machines (SVMs) are the most accurate learners compared to the Naive Bayes, decision tree, discriminant analysis, and K-nearest neighbor for both primary and secondary components.