Application of Machine Learning in Pipeline Integrity Reliability

Abstract

Abstract Probabilistic analysis is becoming increasingly adopted by pipeline integrity management practices in recent years. The practice employs reliability engineering methods to address pipeline integrity and safety concerns. At present, the industry is beginning to pair reliability methods with numerical methods to estimate probabilities of failure (PoF) for individual defects, or features, in a pipeline. The effort required for this can be intensive, since it must be repeated on hundreds of thousands of features, which need to be analyzed on a regular basis. This poses a challenge for pipeline reliability engineers, given limited human and computational resources. In the meantime, machine learning applications in many industries have grown significantly due to advancements in algorithms and raw computing power. With massive amounts of raw data available from inline inspection (ILI) tools, and artificial data available through simulation techniques, pipeline integrity reliability becomes a promising field in which to apply machine learning technology to fast-track PoF estimation. Since a large population of reported features have low PoFs and pose low risk to integrity and safety, they can be safely screened out using fast machine learning models to free up engineers who can be dedicated to in-depth analysis of more critical features, which could have a much larger impact on pipeline operational safety. In this paper, two machine learning models are proposed to address the pipeline integrity reliability challenges. The regression model was able to predict features with low PoFs with 99.99% confidence. The classification model was able to conservatively predict PoFs so that no high PoF feature was misclassified as being low PoF, while correctly filtering out 99.6% of the low PoF features. The proposed approach is presented and validated through pipeline integrity simulated case studies.