2020 PCRI Research Exchange Meeting

Pipeline Research Council International meet hears from Rosen on an ML-based digital alternative to in-line inspection. Integral Engineering and SoCalGas on using machine learning to enhance pipeline reliability assessment. Pacific Northwest National Laboratory’s hybrid physics-based/data-driven model fed with PHMSA incident data.

Speaking at the 2020 Research Exchange Meeting of the (PCRI), Christopher De Leon and Michael Smith (ROSEN Group) offered a digital alternative to in-line inspection, aka a predictive analytics solution for monitoring the condition of uninspected pipelines. Using inline inspection data (ILI) from 5,000 pipelines, a trained Bayesian network was used to predict internal corrosion growth rate in an uninspected line, using ILI data from similar inspected pipelines. The approach led to a ‘significant reduction in unnecessary digs compared to traditional modeling’. Potential applications of the technique include input to direct assessment (dig prioritization), decision support for pipeline modifications (piggability) and ILI validation to API 1163 Level 1.

SoCalGas has used machine learning to enhance pipeline reliability assessment as Daryl Bandstra (Integral Engineering) and co-author Mari Shironishi (SoCalGas) explained. The approach addresses the risk of excavation damage, a leading cause of pipeline failure, leveraging earlier PRCI-sponsored research* into fault tree/structural reliability models of mechanical interference from excavation equipment. A machine learning regression model was fed with data on land use and ‘one call’ notifications of upcoming excavations. The model makes location-specific predictions of the rate of excavation notifications along a gas transmission pipeline network. The project showed how ML models can enhance existing risk assessment approaches and how the use of public data sources can improve risk estimates. The authors noted however that ‘an increase in accuracy comes at the cost of decreased interpretability, for many machine learning models’.

* Chen, Q. and Nessim, M. 1999. 'Reliability-based Prevention of Mechanical Damage to Pipelines’. Submitted to Pipeline Research Council International. Catalog PR-244-9729.

Kayte Denslow and other co-authors from the Pacific Northwest National Laboratory reported on DOE/Office of Fossil Energy-supported research into natural gas transmission prognostics with machine learning. The work investigates the use of novel signatures from deployed transmission infrastructure and sensors. Despite advances in inspection technologies, significant incidents continue to occur that undermine the safety and reliability of natural gas transmission pipelines. Insights from past lifecycle data and novel signatures from in-line inspection data can improve the prediction (and enable better avoidance) of incidents. The research sets out a) to develop a ‘health index’ for pipelines, derived from sensor measurements and historical operating data and b) to predict system condition and expected remaining useful life. A hybrid physics-based/data-driven (HBDD) model has been developed and is being fed with PHMSA incident data. So far, the machine learning model has correctly ‘predicted’ the incident year 86% of the time. Looking forward, the researchers plan to deliver ‘PHD’, a commercial-grade pipeline health diagnostic software tool that can integrate an operator’s scada system and provide pipeline health status (situational awareness) along with predictions of when and where pipeline repair/replacement is needed. Download the full (1,100 page) RCRI REX Proceedings.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.