A lightweight deep learning framework for assessing vehicle-collision damage in double-column bridge piers using structural responses

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Timely and accurate assessment of the damage state for bridge piers under vehicle collision is crucial for efficient bridge operation decision-making. This article presents a damage assessment framework for double-column reinforced concrete (RC) bridge piers subjected to truck collisions, leveraging deep learning techniques. Through experimental validation of numerical algorithms and material constitutive models, high-fidelity finite element models were developed for representative nonbottom and bottom collision scenarios. Based on these models, damage indices were derived from simulated collision responses and correlated with damage states, enabling the construction of an imbalanced dataset reflective of realistic damage-level distributions. The dataset comprises 112 preliminary collision cases and 334 collision cases after guided augmentation, covering both nonbottom and bottom collision scenarios. Using either pier-top displacement or acceleration time-history responses as model inputs, a lightweight physically constrained ordinal regression (P-OR) deep learning model was developed by integrating physical feature constraints with ordinal regression (OR) to map structural responses to postimpact damage levels. The results demonstrate that the proposed model performs excellently on displacement data, with all evaluation metrics exceeding 90% for both the validation and independent test sets, showcasing promising predictive capability within the investigated structural configuration. In contrast, when acceleration responses are used as model inputs, the evaluation metrics generally remained above 70%, indicating moderate yet still feasible assessment capability. The ablation study further demonstrates the effectiveness of integrating physical feature constraints with OR strategy for postcollision damage assessment. This article could provide helpful reference for the intelligent postcollision damage diagnosis of RC bridge piers.