Advanced analytics in North America

Inline inspection (ILI) has been utilised to help manage corrosion threats in pipelines for decades, and throughout that time has been actively developing and maturing. Magnetic flux leakage continues to be the most widely and arguably effective inspection technology employed by pipeline operators to manage corrosion, and while there have been significant advances in sensor density and electronic capabilities, what will not change is that it is an indirect measurement technique. This means that it relies on ‘truth data’ to calibrate or train algorithms to interpret the signals and provide the meaningful data needed to assess pipeline condition.

In parallel, there have been advancements in the quantity and quality of validation or truth data being collected. Historically, a relatively low volume of artificially created or machined defects were used as truth data to calibrate the inspection systems. To-day, high quality and high-resolution depth measurements can be collected by pipeline operators which can provide hundreds or even thousands of measurements of individual corrosion features from a single dig. Combine that across thousands of inspections and there is exponentially more truth data available that can be used to feed data hungry machine learning (ML) algorithms that could step change inspection tool measurements.

Seeing this emerging opportunity and working closely with customers keen to advance inspection system performance, Baker Hughes has established its ‘big data’ library. It is a vast database of high quality, field measured, truth data, aligned together with the highest resolution inspection signals and operational data that is ideally suited to explore these new opportunities arising with advanced data analytics techniques.

As a first application of the data library, Baker Hughes wanted to look closer at the reported tool tolerances of the tools with the aim of removing conservatism in dig programmes while at the same time improving pipeline safety. Why was this important? Well, even a small change in the reported tolerance of reported metal loss can have a significant impact on planned integrity management programmes and knowing when tolerances are more likely greater than expected will lead to earlier warning that pipeline re-pair may be needed.

Currently, and historically, metal loss tolerances are typically re-ported using seven metal loss categories defined by the Pipeline Operators Forum (POF) that subdivide metal loss by length and width (Figure 1). This methodology has provided benefit by allow-ing differentiation in reported tolerance relative to length and width, which can influence MFL sizing accuracy.

The POF-based categorisation does have some limitations though. First, with it being defined as discrete bins, there are potential step changes in reported tolerances as a reported anomaly transitions from one category to the next. This does not represent how the tolerance is truly affected by potentially subtle length and width changes. Secondly, length and width are not the only influencing factors that can affect tool tolerances. Anyone who regularly works with the data will know that specifications are provided with a tolerance and confidence level which means reported results will generally be within specification but sometimes they are not.

Accounting for this adds conservatism to the process of integrity assessment and therefore can ultimately lead to unnecessary costly repairs.

A research project was started in collaboration with TC Energy to explore if modern data analytics methods could be applied to the vast library of known corrosion measurements and associated signals to establish a way to move away from ‘binned’ tolerances based on dimensions with a generalised accuracy, to a specific tolerance that could be predicted based on the signal characteristics of an individual corrosion feature.

Line-specific tolerance prediction model

The initial stages of the research considered two pipelines with over 40 000 defects between them correlated to field laser measurements. Multiple influencing parameters were identified and investigated to evaluate correlations between the actual depth error and the anomaly parameters. By prioritising the influencing factors, four categories of parameters were nominated as input into the model training that had the highest impact on the sizing accuracy. Inputs that had little to no influence were excluded. The categories determined were:

• Predicted anomaly measurements: the predicted length, width and depth output of anomaly sizing models.
• Raw signal characteristics: the parameterisations of the anomaly ILI raw signal, including both the triaxial MFL data and other supplemental raw data.
• Location and interaction to other pipeline fittings and fixtures: the location and interaction of target anomalies to other pipeline fit-tings and fixtures. This includes items like tees, valves, bends, repairs, etc.
• Location and interaction to other defects: the location and interaction of target anomalies to other neighbouring anomalies.

Not surprisingly, the factors mentioned within these categories interact together in complex ways to influence confidence bounds and explain the variable sizing tolerances.

To access the full version of this article, sign up for the magazine (print or digital) here!

Read the article online at: https://www.worldpipelines.com/special-reports/06092023/advanced-analytics-in-north-america/