A Case Study Showcasing the Use of Extreme Learning Machine Based on in-line Inspection Data

Oil and gas pipeline failures due to internal corrosion threaten human safety, infrastructural and economic losses, and environmental devastation. In this paper, we propose an innovative method of predicting the internal corrosive depths of oil and gas pipelines using Extreme Learning Machine (ELM) based on historical In-Line Inspection (ILI) data which first undergoes Mutual Information Principal Component Analysis (MIPCA) for feature selection. An in-service pipe section in Qinghai, China, spanning 170 Kilometers with 241 recorded internal corrosion defects is used as a case study to evaluate the efficacy of the proposed model. Features including distance from elbow, distance from upstream girth weld, initial defect depth, pipeline pressure and elevation were identified and used as predictors. The results show that compared to widely used machine learning methods such as Back Propagation (BP) neural networks and Random Forest (RF), the proposed method improves the accuracy and calculation speed of the prediction model than currently used methods because it is highly generalizable. Better predictions help better maintenance plans, thereby minimizing risk and costs of pipeline failures.