Predictive modeling for shrink sleeve failure using machine learning

摘要

In 2007, external corrosion and linear indications were discovered underneath shrink sleeves installed on field girth welds located on a 10-mile section of natural gas transmission pipeline constructed in 1993. Over the next 10 years, 1,615 of 1,971 shrink sleeves were removed. Given the volume of this dataset, the operator established a goal to prioritize the removal of the remaining shrink sleeves in this segment, while establishing a predictive proactive model, to prioritize shrink sleeve removal across the system. The system is expected to contain nearly 20,000 shrink sleeves. To build this model, data was managed in a custom developed SQL database with Microsoft PowerBI, ESRI ArcGIS Server Portals, and machine learning algorithms applied for powerful analytics. We will define the lifecycle of a shrink sleeve to be four stages: 1. Stage 1: Intact 2. Stage 2: Failed (i.e. disbonded) - no corrosion 3. Stage 3: Failed - with corrosion 4. Stage 4: Failed - with corrosion and cracking In concept, Stages 3 and 4 can be detected through In-Line Inspection (ILI), whereas Stages 1 and 2 are believed to be predictable through modelling. The fundamental result of this project delivered a data structure where advanced predictive analytics can be applied to develop models capable of predicting Stage 2, specifically.