Trial and Error — Applications of Machine Learning Models to Pipeline Integrity Data

Abstract For pipeline operators, a thorough understanding of machine learning’s strengths and limitations is key to reaping the benefits of data-informed decision-making and efficient use of resources. This paper conducts an analysis of different machine learning models and data quality assessment tools/analytical methodology, bolstered by case studies, highlighting the critical need to choose a model that best aligns with the specific characteristics of the data to fully leverage these benefits. This paper undertakes an analytical exploration of the strengths and limitations inherent to each model, complemented by various case studies as a proof of concept of leveraging machine learning in the pipeline integrity workflow. The accuracy of Bayesian Networks (BN) and Natural Language Processing (NLP) in predicting text-based data entries from inline inspection feature comments was compared, showcasing one model’s substantial superior performance. Another case study revealed variation in the accuracy of BN in predicting absent information, by comparing predictions derived from a dataset with missing entries, against a modified validation dataset. A final case study highlights the ability of text string comparison using Fuzzy Matching in automated Data Quality Assessment applications, to efficiently match and categorize wall thickness and outside diameter values, portraying significant time saving advantages.