Using odometry drift to match ILI joint boundaries for run comparisons

Abstract

Repeat coincident inline pipeline inspections are used to calculate defect-to-defect corrosion growth rates (CGRs) for accurate reinspection intervals. First, a pair of inspections is aligned at the boundaries between pipe joints. Pattern matching is then used to match individual defects on paired joints. Joint-boundary matching is a correspondence problem that can be described in terms of “drift”, which is the difference in odometry values at corresponding joint boundaries. Here, drift is formalized and characterized. It is used to describe current joint-boundary matching methods — interpolation, zippering, and direct odometry comparison. We use the drift formulation to illustrate why existing methods produce false matches (1) in the presence of poorly documented reroutes, (2) when there are nonlinearities in the odometry from pig speed changes, and (3) from other inspection problems. We develop a new evidence-based intuition for drift as the systematic bias of position-estimating state machines on inspecting pigs. This is followed by deriving a position-specific drift differencing distance metric using the first-order differencing of the drift. We show this value to be trend stationary. The results section compares the four metrics’ abilities to identify common joint-boundary matching challenges. We found that direct matching and interpolation are unreliable. These methods generate difficult-to-detect mistakes because they rely on a nearest-neighbor matching criterion. Zippering and drift differencing work much better. Both rely on pups in the vicinity of pipeline reconfigurations. They both fail when there are no pups or when a reconfiguration is large. Because zippering uses naive pattern-matching, it is subject to aliasing mistakes when non-corresponding joints are in analogous positions. In contrast, drift differencing, which is position-specific, is immune to this type of error. Finally, we show that the assumptions behind the derivation are valid for an example dataset.