General Sewer Deterioration Model Using Random Forest

Collapse of sewers can induce significant damage to roads and buildings, resulting in large economical costs. Therefore, utilities wish to repair or replace the sewers before they collapse. In order to investigate if a sewer needs maintenance or replacement it can be inspected with Closed Circuit Television (CCTV), but as CCTV inspection is very expensive, and hence only a small percentage of the sewers are inspected. This underlines the importance of choosing the correct sewers for inspection and have resulted in development of several deterioration models. However, the best performing existing models are tailored to individual cities and need to be calibrated in order to be generalized to new areas. As the cost for collecting a data set for calibration is high, the utilities could benefit from a sewer deterioration model which generalizes across location. This paper presents a deterioration model based on Random Forest, which is trained on data from 35 utilities spread across the country of Denmark. The model was able to predict the sewer condition with a specificity at 0.80 and a sensitivity at 0.76, which is comparable to the best existing models. This shows that it is possible to make a deterioration model which generalizes across data from different regions, sewers and utilities. This is a significant improvement compared to the current situation where models need to be learned for each new set of data.