Integrating machine learning failure probabilities and tracer-based infiltration indicators for sewer condition screening and asset management.

Wastewater utilities need transparent decision-support methods to prioritise sewer rehabilitation when CCTV inspection data are incomplete and groundwater infiltration creates operational and environmental burdens. Machine learning (ML) models are being trained on sewer inspection data to perform as modern asset-management tools. However, given the limited availability of inspection data, independent external supporting data that are less resource-intensive to collect would be highly valuable. Tracer-based infiltration estimates are one such indicator. To our knowledge, this is the first study to combine these two techniques and compare ML-based structural failure probability with independent tracer-based evidence of groundwater infiltration under dry-weather conditions. As infiltration cannot be reliably estimated at the pipe level, we aggregated the studied network into 21 operational zones and estimated infiltration severity using dry-weather tracer data, based on electrical conductivity (EC) and chemical oxygen demand (COD), within a binary mixing framework. In parallel, ML-based condition prediction models were developed using an IPCW-weighted machine-learning framework trained on CCTV inspection data, aggregating the resulting predicted failure probabilities to the zone level using length-weighted averaging. Zones were ranked according to ML-predicted failure probability and tracer-based infiltration severity. The agreement between the two rankings was positive but statistically uncertain (Kendall's τ-b = 0.254-0.305; p = 0.108-0.056; concordance index = 0.628-0.655). Overall, the proposed framework provides a practical cross-evidence screening approach for environmental management and sewer asset-management decision-making when comprehensive inspection data are unavailable.