Regional Seismic Risk Assessment of Infrastructure Systems through Machine Learning: Active Learning Approach

Abstract

Regional seismic risk assessment involves many infrastructure systems, and it is computationally intensive to conduct an individual simulation of each system. This paper suggests an approach using active learning to select informative samples that help build machine learning models with fewer samples for regional damage assessment. The potential of the approach is demonstrated with (1) failure mode prediction of bridge columns, and (2) regional damage assessment of the California two-span bridge inventory with seat abutments. The active learning approach involves the selection of column attributes or bridge models that are more informative to the creation of machine learning-based decision boundaries. The results reveal that an active learning target model based on 100 bridge samples can achieve a level of accuracy of 80%, which is equivalent to a machine learning model based on 480 bridge samples in the case of damage prediction following an earthquake. With the proposed approach, the computational complexity associated with regional risk assessment of bridge systems with specific attributes can be drastically reduced. The proposed approach also will help plan experimental studies that are more informative for damage assessment.