Integrating dynamic features into machine learning models for predicting sewer network failures: a random forest approach

Sewer blockages and flooding remain persistent challenges in sewer networks worldwide. Water utilities often deploy labour-intensive methods, such as repeated CCTV inspections and jetting, to detect and address these problems. Operatives frequently need to inspect several locations before identifying a significant issue that requires intervention, making this iterative approach a lengthy and costly process. While there have been attempts to enhance the capabilities of inspection technologies, no emerging innovation has achieved widespread adoption. Recently, machine learning (ML) based asset failure prediction modelling has shown promise as a cost-effective alternative, enabling proactive identification of vulnerable pipe sections within sewer networks that can then be the focus for enhanced inspection and intervention. Early ML based predictive models primarily focused on non-dynamic factors, such as physical pipe attributes, while newer approaches have incorporated dynamic variables like rainfall and pipe flow, resulting in significant accuracy improvements. This study examines the impact of integrating knowledge of sediment transport mechanics, driven by parameters derived from network hydraulic models, into a predictive Random Forest (RF) model. The results demonstrate that incorporating dynamic features representing sediment transport capacity and its spatial variation considerably enhances the RF model's predictive power, offering a more reliable tool for identifying and managing sewer defects such as blockages, and flooding in combined sewer networks.