Abstract:
Pipelines are the primary means to safely and efficiently transport oil or gas at high pressures across long distances. To detect any anomaly or flaw before it has a detrimental effect on the integrity of a line or system, pipeline operators resort to surveys performed 'in-line' using intelligent inspection tools. Analysis of the multi-dimensional data recorded during in-line inspections however requires significant manual effort from experienced subject-matter-experts.
The focus of this work is the development and productionization of an intelligent decision-making system to aid the human experts in the analysis of pipeline inspection data as recorded by Process & Pipeline Services’ UltraScan CDP crack detection tool.
Training and validation of such a system requires high-throughput batch-access to the data recorded by UltraScan CDP, which is a challenge of its own. This work therefore first evaluates and proposes a “Not only SQL” database and novel storage schema for pipeline inspection data.
As the objective of this work is not just to prototype a system, but put it to production use, this work furthermore introduces our bespoke platform to manage the end-to-end lifecycle of machine learning systems.
Leveraging our storage solution and machine learning platform, we then present our novel system to support the analysis of UltraScan CDP’s crack detection data, tuned to minimize the amount of human intervention, whilst ensuring no significant flaw is missed.
As of writing this, no comparable machine intelligence system, lifecycle management platform or storage solution is known.
