Sewer sediment deposition prediction using a two-stage machine learning solution

Sediment accumulation in the sewer is a source of cascading problems if left unattended and untreated, causing pipe failures, blockages, flooding, or odour problems. Good maintenance scheduling reduces dangerous incidents, but it also has financial and human costs. In this paper, we propose a predictive model to support the management of maintenance routines and reduce cost expenditure. The solution is based on an architecture composed of an autoencoder and a feedforward neural network that classifies the future sediment deposition. The autoencoder serves as a feature reduction component that receives the physical properties of a sewer section and reduces them into a smaller number of variables, which compress the most important information, reducing data uncertainty. Afterwards, the feedforward neural network receives this compressed information together with rain and maintenance data, using all of them to classify the sediment deposition in four thresholds: more than 5, 10, 15, and 20% sediment deposition. We use the architecture to train four different classification models, with the best score from the 5% threshold, being 82% accuracy, 70% precision, 76% specificity, and 88% sensitivity. By combining the classifications obtained with the four models, the solution delivers a final indicator that categorizes the deposited sediment into clearly defined ranges.