Publications / 2024 Proceedings of the 41st ISARC, Lille, France
Every year, accidental damage during excavation leads to numerous disruptions in utility services. These incidents cause not only financial losses but also injuries and fatalities. A major contributing factor is the lack of accurate location data for these utilities of such incidents. The current practice involves a time-consuming coordination process of obtaining utility maps from owners and field surveys, which is often hindered by delays and incomplete records. In response to these challenges, this paper proposes a novel method to predict underground utility lines in situations where records are unavailable or delayed. Our approach leverages visible utility anchor points, such as manholes, and the spatial context provided by nearby ground features like roads. The methodology involves three primary steps: constructing a relational data model of the utility network, transforming this data into graphs, and employing a graph neural network for prediction. This innovative approach demonstrates good performance, achieving a 94.12% ROC AUC score in predicting sewer line connections between manholes. This method automates the inference of utility lines, providing utility owners and excavation contractors a solution for identifying unknown connections and reducing risks from inaccurate information.