Conference Agenda

More than 13.000 km of Dutch pressure mains are approaching or have already exceeded their expected operational lifetime, highlighting the need for better methods to assess pipeline condition and estimate remaining service life. This research developed a failure probability model for managing pressure mains, integrating physically-based models for mechanisms such as uneven soil subsidence, wall thickness reduction, and pressure fluctuations, with data-driven statistical models for third party excavation damage and defective air release valves. The model calculates failure probabilities based on the evaluation of 144 scenarios, which are combinations of possible future conditions. The application of the model on actual pressure pipelines confirmed its effectiveness in predicting failure risks and identifying dominant failure processes. By accurately assessing failure probabilities and locating vulnerable pipe sections, asset managers can make informed decisions about maintenance, rehabilitation, and inspections, thus optimizing the rehabilitation planning and extending pipeline service life. This model provides valuable insights into managing aging infrastructure.

Sewer networks face significant challenges from blockages caused by fats, oils, grease, tree roots, and non-biodegradable items like wet wipes. Increased flushing of wipes exacerbates blockages, while the hydraulic impacts of wipe accumulation and methods for modelling them in sewer remain underexplored. This study addresses this gap by simulating wipe accumulation in sewer defects under varying flow rates and blockage sizes. Results demonstrated that upstream water levels consistently increased as blockages grew. These hydraulic effects were modelled in the Storm Water Management Model (SWMM) using four methods: adjusting Manning’s roughness coefficient, filling the pipe, modifying the headloss coefficient, and incorporating an orifice. The simulation results quantified the dependency of model parameters on both flow rate and blockage size. This research provides practical guidance for modelling wipe-caused blockages, accounting for reduced sewer capacity and performance. The findings can be incorporated into sewer asset management practices to enhance system efficiency.

Automated sewer defect detection has advanced through deep learning, particularly supervised methods using CCTV images, but based on large annotated datasets. This study proposes a semi-supervised learning (SSL) approach to reduce the dependency on annotations. The method includes two stages: self-supervised pre-training on unlabelled images using SwAV (Swapping Assignments between multiple Views of the same Image), followed by fine-tuning on labelled images for multi-label image classification. Experiments on the Sewer-ML dataset show that both ImageNet-pre-trained models -supervised and SwAV- outperform models trained from scratch on 1.04 million images, achieving higher F1-scores with just 13k labelled samples. The proposed SSL approach achieves 64.22% precision, 66.06% recall, and a 65.13% F1 score, surpassing the fully supervised baseline. Additionally, scaling up the pre-training dataset further enhances performance. These findings underscore the importance of ImageNet initialization and highlight self-supervised learning as an accurate, scalable, and cost-effective alternative to supervised methods, particularly in data-scarce scenarios.

Deterioration models for sewer pipes often rely only on aggregated pipe-level data (pipe condition), without considering individual defects and their evolution. Is-it worth considering individual defects to improve deterioration models? A preliminary answer is to know if it is possible to model the evolution of individual defects. This study presents a methodology for analysing defect transitions in multi-inspected sewer pipes to gain insights into the aging and deterioration processes at the defect level. Using inspection data provided by Berliner Wasserbetriebe, covering 242,920 pipes and nearly 1.9 million observations, incl. defects encoded according to EN 13508-2, defect transitions were analysed across 24,734 inspection pairs. Defects between inspection pairs for each pipe and position are mapped, creating a transition matrix and knowledge graph to highlight defect inter-dependencies. The results reveal plausible transitions, such as gradual surface degradation from increased roughness to missing pipe wall parts, with varying durations, but also transitions that may reflect inspection uncertainties. Future work will incorporate defect severity classes and explore how these insights can enhance machine learning models through feature engineering or domain-informed approaches.

This study explores the application of a machine learning model, specifically a Random Forest classifier, to predict the condition of uninspected pipes using available structural, operational, and environmental data. Originally developed for Berlin, Germany, the model has been adapted and applied to the sewer network of Lausanne, Switzerland. Model performance was evaluated using custom metrics, with results compared to previous applications in Berlin. Despite challenges related to class imbalance, the model demonstrated promising accuracy, supporting its potential as a decision-making tool for inspection prioritization.

The increasing costs of maintaining sewer networks are driven by population growth and the aging of pipes. Predictive maintenance offers a solution by optimizing resource allocation and focusing repairs on the most vulnerable sewer pipes. This study employs Machine Learning (ML) models, particularly Random Forests (RF), to forecast hydraulic failures using citizens’ reporting data and GIS-based pipe parameters. This approach is advantageous for municipalities with accessible data, especially in areas lacking CCTV inspections for structural conditions. While initially utilizing RF models, the study aims to incorporate Graph Neural Networks (GNNs) to exploit spatial connections within the network.