Overcoming Challenges for Quantitative Risk Modeling Using Machine Learned Data Correlations and Predictive Modeling

Abstract

After 20 years of learnings and successful risk reduction, the pipeline industry is striving to achieve the next step change in risk performance by migrating from relative index-based risk models toward probabilistic approaches across all threats and all pipeline segments for system-wide risk assessment. While the quantification of pipeline risk can be readily supported by in-line inspection (ILI) results coupled with probabilistic limit state modeling for certain threats and pipeline segments where this information is available, where such data is lacking for other pipeline segments or threats, it is necessary to apply a meaningful methodology to establish the probability of failure for all dynamic segments used to quantify risk. The process for establishing probability-based threat assessments for these other segments involves several stages: (1) identify correlations based on ILI historical results (i.e., create the “training” dataset); (2) leverage classification trees to identify statistically relevant data observations of key variables; (3) apply machine learning techniques to develop probabilistic and/or causal models that predict target outcomes from the combinations of key variables; and, (4) apply the relationships established in stages 1, 2 and 3 to all assets lacking ILI results. Although seemingly straightforward, several challenges exist for achieving seamless implementation. This paper will review each process step and provide guidance on preparing and tackling common data intelligence challenges. The paper will also propose strategies for classifying assets, other than indexing, for application in those situations where machine learning does not indicate statistically significant variable correlations or yield strong predictions of target outcomes. Although still early in the understanding of migrating from relative index-based to probabilistic risk algorithms, the value provided in this paper is the sharing of lessons learned regarding “how” to gather the “evidence” necessary to identify statistical dependencies, how to apply data confidence metrics within the decision process, the challenges in data preparation/QC and interpretation techniques, and suggestions for determining the necessary limitations that should be applied to the outcomes toward the objective of quantifying the risk.