Abstract
The federal rules governing the operation of natural gas pipelines allow operators to use nondestructive testing (NDT) technologies to verify pipeline material properties provided that these tools are validated against destructive test results and conservatively account for measurement inaccuracy and uncertainty. Any measurement methodology inherently has uncertainty due to both systematic and random effects: systematic errors being those that remain constant during repeated measurements, and random errors being those that vary randomly in repeated measurements. This combined uncertainty not only affects the estimation of material properties, but it may also propagate to downstream analyses, such as during maximum allowable operating pressure (MAOP) reconfirmation. Moreover, this uncertainty might adversely affect the accurate determination of comparable pipe segments when establishing sampling populations.
This paper presents Pacific Gas and Electric Company’s (PG&E) approach for the statistical inference of destructive laboratory values when in situ (field) NDT measurements are collected. This approach has been formulated from an in-house statistical validation of NDT technologies whereby NDT measurement results are compared against laboratory destructive test results. The types of material properties that PG&E has evaluated using NDT technologies include microstructure, hardness, chemical composition, yield strength, and ultimate tensile strength. In total, several thousand NDT measurements across approximately one-hundred pipe features have been evaluated against laboratory test results. The authors describe this statistical analysis approach whereby the lab results are compared to NDT measurements through a regression model to account for systematic errors. Once the regression is performed, the scatter of the data is quantified using a prediction interval to account for the random portion of the uncertainty. Finally, the total uncertainty is quantified and propagated to downstream analyses using a Monte Carlo methodology. Current challenges to this approach are presented, and alternate statistical approaches are described which have the possibility of yielding additional benefits.
Available after the event



