Probability of failure feedback loop due to internal corrosion in oil production systems

Abstract

This dissertation aims to provide a methodology for creating a dynamic feedback loop to predict failures due to internal corrosion on high-pressure process piping systems in the oil and gas industry. This has been aided by applying machine learning capabilities to data available in a standard integrity management system. While great advances have been made in integrity management in recent years a significant number of failures still occur due to pipes corroding from the inside out. Although real time or “Online Monitoring” systems are available, they are usually only used on critical piping systems where there are known issues or where the loss of integrity would cause massive disruption to business or loss of life. In a typical facility, such as the one used as the basis of this dissertation, the total length of the pipework can easily exceed 20km. For piping systems not covered by online monitoring the structural integrity is managed through periodic inspection where the remaining wall thickness is measured by ultrasonic thickness (UT) inspection at specific points known as Condition Monitoring Locations (CMLs). This thickness is then compared to the minimum required thickness (Tmin) for that CML. Given that the number of CMLs governed by an integrity management system can easily exceed 100,000 with anywhere between 25,000 to 200,000 UT readings taken on these CMLs being added annually, patterns in data that may have been recognised in smaller sets can go unnoticed. Machine Learning can offer the capability of detecting these patterns in larger datasets. The algorithms developed for this study aim to provide the basic input to screen for different types of corrosion in the entire system by 1) Manually identifying corrosion patterns through inspection 2) Training the system to classify this specific corrosion pattern 3) Training the system to determine corrosion rate 4) Generating the output, i.e. a ranking of all the CMLs in a system by how likely they are to experience the same corrosion pattern 5) Determining the corrosion rate from the trained network