312
Friend or Foe – Responsibly applying Machine Learning and AI techniques to In-line inspection for pipeline integrity assessment
Advances in In-line Inspection (ILI) measurements and tool technologies have benefited the detection, characterization and sizing of critical features in pipelines globally. However, the nature of the output or formats of ILI reporting has not changed significantly in 25 years.
The unprecedented abilities of commercial computer graphics and visualization methods to represent physical and imagined objects in a “life-like” manner also apply to ILI features, flaws and the reference pipe. While at the same time, advances in computing, image processing and numerical algorithms have led to the application of Machine Learning (ML) techniques within ILI data processing, which are demonstrating flaw characterization at levels of granularity on the order of variations of the “clean” pipe itself.
ML also offers a unique opportunity to tackle the real-world challenges that existing analysis methods struggle with, such as complex, overlapping or densely packed defect morphologies, providing a means to consider and manage many more essential variables during the interpretation of the corresponding signal responses in these scenarios.
Application of these techniques still has the objective to maximize the predictive accuracy of the true severity of the most challenging defects, which in-turn are those more likely to be outliers and therefore present a greater uncertainty to the risk of failure. The conventions of errors and uncertainty have a distinct role in both perception and in practical evaluations.
Using examples from MFL and other ILI techniques, this paper will build from prior studies of traditional sizing approaches that looked at “How good is good enough?” to consider new dimensions of interest that must be considered by operators in applying ML and AI techniques to ILI:
- accuracies of flaw signal reconstruction vs accuracies of traditional flaw sizing, and
- when, why and how to apply these seemingly “black-box” techniques in safety-critical decision making?
The abilities to generate image-like flaw representations infer perfection but cannot be without an objective scrutiny and criteria for the actual level of accuracy of representation. Analogies can be made to AI generated images, videos as deepfakes that may be influencing subjective human bias as “truth” decisions vs having common objective criteria.
The authors will address these critical topics and expand on proposals as to how ML and AI, advanced flaw characterization and reconstruction can be progressively introduced to the processes of in-line inspection and engineering critical assessments, providing confidence both expected and needed for its adoption by the industry.
Available after the event



