Deep-Learning-Based Bridge Condition Assessment by Probability Density Distribution Reconstruction of Girder Vertical Deflection and Cable Tension Using Unsupervised Image Transformation Model

Abstract

This study proposes a deep-learning-based condition assessment approach by reconstructing marginal probability density distributions of girder vertical deflection (GVD) and cable tension (CT) using unsupervised image transformation (UNIT) model for a real cable-stayed bridge. 27 and 139 sensors of GVD and CT are distributed along the full bridge, and both the sampling frequencies are 2 Hz. First, vehicle-induced components of GVD and CT are extracted by data pre-processing while the temperature-induced components are removed, a time window of 3 h with a sliding length of 10 min is used to obtain original GVD and CT segments, and the corresponding margin probability density distributions (PDFs) are calculated by kernel density estimation. The training and test sets consist of 2986 and 6236 PDFs, respectively. The proposed UNIT model consists of a series of variational autoencoders (VAEs) and generative adversarial networks (GANs). Both the PDFs of GVD and CT are used as inputs. Then, the proposed UNIT model is updated by solving the mini-max problem of training GANs, in which VAEs act as generative models. Finally, the Wasserstein distance between the predicted and ground-truth PDFs of CT acts as the indicator of bridge condition change. Results show that specific modes of PDF variations induced by SHM system upgrade, cable damage, data anomaly, and traffic jam can be recognized to assess bridge condition.