Session: NDE-03-01 NDE Reliability - Modeling and Experimental Analysis

Paper Number: 106426

106426 - Using Detection Performance to Assess Outlier Sizing on Truncated Data Sets 

The third edition of API 1163 RP "In-line Inspection Systems Qualification", released in September 2021, includes many improvements over previous editions including updates to Section 8 "System Results Validation", which defines the methodologies to validate ILI run measurement tolerances.  The recommended practice describes three levels of validation, with ’Level 3’ requiring operators estimate ILI tool measurement performance with real-world data measured in validation spools and excavation sites.  It is an advancement over Level 2 which is a statistical significance test of validation measurements against a vendor specification.  Two examples of Level 3 validations are provided in Appendix C, one describes statistical tolerance intervals, the other describes a hierarchical Bayesian model.  This paper uses the second model to estimate measurement performance when a data set is truncated.

In a previous paper, the authors described using the Hausman and Wise algorithm to model sizing performance on truncated data sets.  When the data is truncated, this algorithm corrects for the 'missing' data, improving the results of a linear regression between inspection measurements and validation measurements.  However, the algorithm is a based on several assumptions, which do not hold for some outlier data.  When an outlier under-predicts the validated size, the Hausman and Wise correction overcompensates because it assumes that every outlier represents a larger number of missed anomalies of similar size.  In this paper, we discuss an improvement to the Hausman and Wise algorithm that incorporates the probability of detection of the measurement tool.  The probability of detection informs the truncation algorithm by providing the necessary parameters to determine if outliers are purely sizing errors, or whether they also present a risk that a similar sized anomaly was not detected.  

This paper includes examples of real field measurement data where the improved algorithm is compared to the original algorithm.  In addition, similarly characterized simulation cases are summarized to show edge cases and limits to the algorithm improvement.  Finally, the paper offers further improvements that could be made to the truncation algorithm and comments on a similar application for censored data sets.

Presenting Author: Jason Skow Integral Engineering

Presenting Author Biography: Jason Skow is a Principal Engineer at Integral Engineering, he has a proven track record in a variety of engineering and leadership positions. Jason has over 20 years of experience in the oil & gas industry with a focus on integrity management, data analytics and risk & reliability. He has participated in several collaborative industry research projects through PRCI, PHMSA, API and CGA.

Authors:

Jason Skow Integral Engineering
Joseph W. Krynicki ExxonMobil Technology and Engineering