AI-Assisted Predictive NDT for Identifying High-Risk Zones in District Heating Pipelines

Urban district heating pipelines are underground, non-visible infrastructures where failures can lead to ground subsidence, road damage, and significant secondary hazards. Conventional Non-Destructive Testing (NDT) techniques such as thermography and sensor-based inspections face limitations due to weather, accessibility, and restricted monitoring coverage. Meanwhile, actual failure cases represent less than 1% of total pipelines, creating severe data imbalance challenges for AI models. This study proposes a Virtual NDT framework that identifies high-risk pipeline segments using attribute-based K-NN relabeling and machine learning. Structural attributes including diameter, usage type, insulation level, installation year, and burial environment were analysed, and the top 10% of normal segments most similar to failure cases were relabeled as high-risk using K-NN similarity metrics. XGBoost with K=2 achieved the highest performance, yielding an F2-score of 0.921 and AUC of 0.993. Feature importance highlighted installation year and insulation degradation as dominant factors. The proposed Virtual NDT approach compensates for inspection blind spots and enables efficient prioritization f