Deep learning-based estimation of structural component connectivity graphs from bridge point clouds considering instance segmentation errors

Abstract

Connectivity of components that consist of a bridge structural system is effective information for constructing digital twin models applicable to maintenance and performance analysis. This study aims to develop a deep learning-based method generate a connectivity graph representing connections of bridge components from point clouds of a whole bridge. The method consists of two procedures; the instance segmentation of bridge components and the estimation of the connectivity graph within those recognized instances. The verification was conducted by applying the method to a bridge point cloud dataset. While major components of superstructure and substructure were successfully instantiated in case of appropriate training data configurations, the instances of bearings were not appropriately predicted, and false-instance degraded the graph estimation performance. To address this, a training strategy was proposed, which learns connectivity within predicted instances including the false-instances. In the results, it could reduce the mis-connections between components and false-instances achieving improvement in the total performance of component connectivity graph estimation.