Investigation of Pipeline Rupture Detections using Multiple Artificial Intelligence Classifiers

Abstract

Current pipeline monitoring systems have several challenges due to their inability to accurately detect ruptures during transient conditions, delayed response to events, and the inability for conventional or artificial intelligence (AI) based methods to be easily transferred to different pipelines. Many of the detection issues arise with current systems during transients due to the human set thresholds required for their operation. These thresholds are set to balance the accuracy of the system with the time of detection which can lead to many detection issues. There are several challenges associated with AI, one such issue is the need for large data sets for every pipeline to allow proper classifier training. It is a rigorous task to generate a data set which has sufficient data for every pipeline. The second major issue with the application of AI methods is their black-box nature. This black-box nature leaves operator with a lack of explainability behind the alarm. The study proposes a novel rupture detection system which utilizes AI to create a robust monitoring system, which can deal effectively with the challenges seen in current systems. The system utilizes three different classification methods, two using a 2-D Convolutional Neural Network (CNN) and one using Adaptive Neuro Fuzzy Interface System (ANFIS). The classifiers are fused with a risk analysis approach to provide the operator detailed information on the proper response to a release alarm.The system is tested to show its benefits over existing methods using multiple data sets. The two data sets used for system development and testing are, a real operator pipeline and a lab-scale experimental pipeline data set. The lab-scale experimental setup is a pipeline mimicking a section of pipeline between two pump stations. Multiple pump control methods are used to generate a data set with a wide variation of events. With a large varied data set the rupture detection system is trained and shown to have high accuracy. The detection system is tested using the secondary data set to show the transferability and reasoning ability of the proposed detection system.