Classification of Water Pipe Damage Types Using Random Forest

This study presents the results of a classification of the types of water supply pipe failures using a random forest model consisting of 72 trees. The modeling was done in Statistica software. The classification accuracy was compared with earlier results obtained from single-classification-tree models. The qualitative-dependent variable was the type of failure (corrosion, crack, sealing). The predictors included quantitative variables (diameter, year of construction) as well as qualitative variables (pipe type and material). The choice of 72 trees was made based on an analysis of the misclassification rate (31%) during the training stage. Increasing the number of trees forming the forest did not produce more accurate classification results: for the test set, the accuracy was 82%, 72%, and 37% for corrosion, crack, and sealing failures, respectively. The trees forming the random forest differed in their structure both in terms of the number of split and terminal nodes, as well as in the depth and number of levels of individual trees. The overall classification accuracy for the test set was nearly 66%, which is a better result than in the earlier analyses based on single trees. The proposed approach also aligns with the currently promoted concept of the sustainable operation of critical infrastructure.