Using Machine Learning to Assess Internal Corrosion Constituent Thresholds

Abstract Natural gas pipeline operators often rely on gas and liquid sample analysis as well as bacteria serial dilution to determine if corrosive conditions exist within the pipeline. When corrosive conditions are suspected, additional monitoring methods such as corrosion coupons or probes are employed. One of the challenges faced by operators is determining what constitutes “corrosive conditions” because existing regulations do not define gas or liquid compositions that are corrosive. In this study, over a decade’s worth of sampling and corrosion monitoring data collected by a natural gas pipeline operator are assessed using machine learning in order to determine the parameters with the greatest influence on corrosion rate. This work combines over 2,300 gas tests, liquid samples, solid samples, and bacteria tests in an attempt to determine if the thresholds the company is using are reflective of where corrosion is actually occurring based on over 1,700 coupon analysis results within an operator’s unique system. The results from this study can be used to guide future sampling practices so that effort is spent collecting the most meaningful data and improving the ability to identify corrosive conditions.