Data-Driven Prediction of Excavation Damages
Léon olde Scholtenhuis
Organisation || Kadaster + Agentschap Telecom + KPN
Candidate Eng || Jiarong Li
Start - end || Feb 2022, Feb 2025
Thesis || Download
This project aimed to support the detection of risky excavation operations based on historical datasets of damages in the Netherlands. Jiarong Li developed, using XGBoost, a data-driven model that predicted the likelihood of damage occurring in an excavation polygon. The Dutch agency Kadaster uses such polygons to exchange data on utility locations between network owners and excavator operators.
The machine learning model had a satisfactory performance with an AUC-ROC score of 0.821 and a balanced accuracy of 0.743. The study included not only the design of the prediction algorithm but also a user interface for excavator operators. This interface displays not only the risk but also the features causing the risk. The interface may be integrated with a KLIC viewer.
A remarkable finding in this study was that predictors of excavation damages seem to be proxy features, such as the organisation type of the excavator operator, the number of trees in an excavation location, and the complexity of the excavation polygon. The actual location of utility lines had much less of an influence.
The future task for the industry is to decide how findings from the project may be taken into account when upgrading the KLIC-melding or data exchanges, in order to facilitate the market in running their damage prediction algorithms.
