IPC 2026

Presentation Details:

Probabilistic Crack Fatigue Analysis

When: TBD

This paper shows that pipeline integrity teams often report rare-event failure probabilities (POF) without quantifying how uncertain those estimates are, which can lead to wasted or insufficient simulation effort. Using a precision-controlled workflow tested on a crack-fatigue model, the authors find that Sobol' sampling hits the same precision as Standard Monte Carlo with 75–94% fewer simulations (LHS saves 50–75%) — letting engineers tell real risk differences apart from numerical noise while cutting unnecessary computation in inspection prioritization. 

SCC Susceptibility Modeling to Demonstrate New Guidelines for Machine Learning in Pipeline Integrity

When: TBD

This paper applies newly developed ML best-practice guidelines for pipeline integrity to a real demonstration case: predicting stress corrosion cracking (SCC) susceptibility. Using anonymized data from ~30,000 km of transmission pipeline across six Canadian and US operators (aligned with public soil databases), the team built XGBoost classifiers alongside a simpler industry-standard screening model, carefully splitting data to avoid leakage from spatial autocorrelation and using SHAP values to check for overfitting and improve interpretability. The XGBoost models outperformed the simple classifier, screening SCC-susceptible segments more efficiently (lower false-positive rate at the same recall) — demonstrating both a usable SCC prioritization tool and a worked example of how to apply the ML guidelines correctly.

False Confidence in Machine Learning: Autocorrelation and Its Consequences in Pipeline Integrity

When: TBD

This paper warns that unified pipeline integrity datasets — combining construction records, ILI, CP surveys, GIS, and operations data — are hierarchically clustered and spatially autocorrelated, so the standard practice of randomly splitting data for ML validation leaks information between train and test sets and inflates apparent performance (macro-F1 overstated by ~10 points, with some class-level errors masked by up to 35 points). The authors call this "False Confidence" and propose a fix: validate using a "unit of independence" (e.g., system, region, or inspection run) that matches how the model will actually be deployed, rather than random splits. Using a coating-type imputation case study (XGBoost with Bayesian-tuned hyperparameters), they show the performance gap is real leakage, not chance, and that seam type and pipe grade end up acting as proxies for installation era. The result is a practical, deployment-aligned validation protocol meant to give operators and regulators an honest basis for trusting ML-assisted integrity decisions.

CJ Kowalke

VP, Irth Training

Mike Radich

Account Executive

Will McMillan

Account Executive