Prediction of Stress Concentration Factor from In-Line Inspection Data Using Convolutional Neural Networks

Abstract

It is becoming increasingly harder to install new oil and gas pipelines, the need for fuel at a reasonable cost is at an all-time high, and the public tolerance for pipeline failure is zero to none. To help existing pipelines maintain these rising energy demands and meet their fitness for service assessments, operators rely on in-line inspection (ILI) data to identify dent defects and calculate their stress concentration factor (SCF). These ILI runs typically identify a large number of dent defects, making it difficult for engineers to process individually. This study developed a Convolutional Neural Network (CNN) trained on 4,667 raw ILI data files containing dent shapes, along with a data pre-processing algorithm to convert ILI data into pseudo-images, to address the need of prioritizing dented pipeline segments based on their SCF. The final CNN model successfully predicted SCFs ranging from 1.04 to 10.69 with an RMSE of 0.418 and R2 of 0.929, therefore showing potential as the framework for a pre-assessment tool used by pipeline operators.