Point Transformer Network-Based Surrogate Model for Spatial Prediction in Bridges

Abstract

Bridges are essential assets of inland transportation infrastructure; however, they are among the most vulnerable elements of these networks due to deterioration caused by aging and the increasing loads to which they are subjected over time. Consequently, maintenance becomes critical to ensure acceptable levels of safety and service. Finite element (FE) models are traditionally used to reliably assess structural health, but their computational expense often prevents their extensive use in routine bridge assessments. To overcome this computational limitation, this paper presents an innovative deep learning-based surrogate model for predicting local displacements in bridge structures. By utilizing point cloud data and transformer neural networks, the model provides fast and accurate predictions of displacements, addressing the limitations of traditional methods. A case study of a historical bridge demonstrates the model’s efficiency. The proposed approach integrates spatial data processing techniques, offering a computationally efficient alternative for bridge health monitoring. Our results show that the model achieves mean absolute errors below 0.0213 mm, drastically reducing the time required for structural analysis.