University of Twente Student Theses


Prediction of damage types during excavations

Kiss, K. (2023) Prediction of damage types during excavations.

[img] PDF
Abstract:The problem addressed in this document is the prediction of individual damage types, for example the breakage of water pipes or the cut of electricity or internet cables, during excavation projects using machine learning techniques. Previous studies have focused on exploring the reasons behind damages or predicting the overall probability of any damage occurring. However, there is limited research conducted on predicting individual damage types. This research is one of the first to utilize machine learning for predicting individual damage types. The unique combined utility registry of excavations and damages of the Netherlands enables this study This prediction. , can be a powerful tool to shorten or even avoid outages. This can be done through either by warning utility providers to the risk or changing the way the excavation is done. The project consists of several main parts. Firstly, the available data is investigated and cleaned to ensure its suitability for modeling purposes. The XGBoost machine learning method was selected, due to its successful track record in similar problem domains. Two approaches are considered for predicting damage types: binary classification and multiclassification. The binary classification approach predicts each damage type individually, treating the selected damage type as one class and all excavations without the given damage type as another class. The multiclassification approach aims to predict the most likely outcome which is either a given damage type or no damage at all. After tuning the models, the following results were obtained for the binary classification approach: usable predictions with adjustable recall and precision were achieved for damage types of internet cable, low voltage electricity, low-pressure gas, and water strikes. To achieve a recall score of around 0.8, which represents the ratio of predicted real damages to all real damages, the precision, the ratio of predicted real damages to all predictions, varies between 0.05 and 0.01. All other damage types have a precision of less than 1% and thus not considered usable. The multi-class model, predicting the most likely outcome, was less successful. This is mainly due to severe class imbalances. In most cases it predominantly assumes no damage will occur, and only when the excavation site is unusually large, it predicts internet cable manage. The most common damage type. The most important features of both binary and multi-class models for damage prediction were the dimensions of the excavation site, related features such as the number of trees around, and the client and excavating companies. The damage types that could be to some extent reliably predicted where the most common one, present almost everywhere. Rarer, most dangerous type such as high-pressure gas or high-voltage electricity, cannot be predicted even with 1 percent precision. This is mainly due to the fact they do occur orders of magnitude less frequent than the common ones like water pipes or internet cables, in all cases with less than 1000 registered cases during the 2019 to 2021 period. To summarize with the available data, it is possible to predict certain damage types in a useful manner, and possible use this prediction to minimize the time and cost of these damages. However, less common damage types cannot be predicted even by extending the data collecting timeframe.
Item Type:Essay (Bachelor)
Faculty:ET: Engineering Technology
Programme:Civil Engineering BSc (56952)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page