Smarter data-driven pipeline risk assessment keeps the gas flowing
Contact us:

Gas pipeline operators are discovering how a method called Bayesian Networks can, combined with machine learning, address data uncertainties in risk management (Figure 1).1 This approach converts uncertainty into an ally rather than an enemy of pipeline risk managers.
Choosing which data to gather first is complicated. Data may be missing or unreliable. There may also be uncertainty over failure mechanisms such as corrosion, third-party damage and others. In addition, collecting and assessing the validity of data from sensors on pipelines, inspections and front-line operatives is expensive. With safety at the top of the agenda, operators therefore have tough cost-benefit decisions to make when designing data-driven risk management strategies.
When data is missing, traditional risk assessment methodologies require operators to conduct pipeline risk assessments based on worst-case scenarios. “With smart risk assessments based on Bayesian Networks, they do not need to gather all the data in the same way,” said Dr Francois Ayello , principal engineer, risk management, DNV GL - Oil & Gas.
“Bayesian Networks is the logical next step in pipeline risk management because it incorporates data uncertainty and provides justification for how much an operator should invest in collecting data about buried assets,” said Gordon Ye , supervising gas engineer of data, risk & threat, for the Gas Transmission Integrity Management (TIMP) side of US utility Pacific Gas and Electric (PG&E).
“This saves cost and time, while keeping threats below acceptable levels," added Ayello. “Our work applying the methodology for pipeline operators shows they need not wait for all the data to do all risk assessment, and this obviously saves resources that can be spent on another pipeline.”2
Bayesian Networks let risk managers produce a statistical model that can graphically represent a set of variables and their conditional dependencies on each other. Consequently, operators can simultaneously model from cause to effect in a simulation, or from effect to cause for diagnosis. In addition, the machine-learning process is stopped when the cost of data-gathering activities outweighs the benefit to risk predictions.
“It lets the pipeline operator see everything that could happen to the pipeline with its associated probability. After all, nothing is certain, but understanding where the uncertainty comes from improves the decision-making process,” said Ayello. “Quantifying the effects of data uncertainty through Bayesian Network modelling lets operators predict in near real-time the risk of all possible events. Here, uncertainty is not the enemy; it actually drives the decision-making process. With Bayesian Networks, every time data is gathered, future risk predictions become more certain.”
This logic lies at the heart of DNV GL’s Multi-Analytic Risk Visualization (MARV™), a smart risk-assessment approach developed over five years of research. It is now being applied to offshore and onshore pipeline systems.
As Figure 2 illustrates, MARV captures known pipeline data as well as variables that are unknown, such as missing data, and knowledge uncertainty. MARV can then predict all possible futures of a pipeline and show the results visually. The user can perform fully quantitative risk assessments, evaluate life extension strategies, prioritize data gathering, plan mitigative actions, and explain the hidden root-causes of risks.
MARV has been applied in almost every world region and on pipelines operating under varying conditions and differing regulatory regimes such as Australia, Russia and the US.3–5 It has been applied to risks including internal and external pipeline corrosion, stress corrosion cracking, electrical currents, and illegal tapping of oil pipelines.
“Bayesian Networks were practically unused in the pipelines industry just five years ago but have become a growing trend because of the way they can counter uncertainty,” observed Ayello.
DNV GL’s US experts have collaborated with multiple stakeholders including transmission system operators PG&E and Southern California Gas Company (SoCalGas) to customize MARV and apply it to a transmission pipeline network. The pilot study aimed to enhance safety while saving resources.
In addition, SoCalGas independently contracted DNV GL to pilot using MARV for combatting external corrosion and the risk of third-party damage to a natural gas transmission pipeline.6
Paul J Monsour, principal engineer, SoCalGas, said: ”Additional effort is needed to consider and establish relationships between pipeline attributes and data. The MARV pilot helped drive our data-gathering efforts for the pilot project.”
Because MARV places a value on variables that risk managers want to track, it suggests what to look for, and therefore where to look for it in the company, said Monsour: “Some maintenance activities are not run by my group, but are an important part of the pipeline risk. MARV helps to quantify the effect of uncertainty and to understand, for example, how to direct our information-gathering efforts.” (Figure 3)
Regulation will be a key driver for uptake of Bayesian Network approaches in pipeline risk assessment, according to Monsour and PG&E’s Ye.
Monsour sees potential benefits from MARV in dealing with regulatory requirements: “Regulators in the US are moving towards requiring operators to quantify risks rather than considering the relative risks. This is to encourage a better understanding of where to apply resources, and to help quantify the benefits of these investments. MARV could potentially help quantify both risks and benefits.”
While the advantage of the MARV approach to mitigate insufficient data is appealing, uptake of the Bayesian Network methodology will be led by regulators getting operators to invest more in it, said Ye: “Some regulators want to align operators with the same methodology when it comes to assessing pipeline safety risk. They are currently using traditional methodology, hiring experts who have done some plant risk assessment and trying to do that with pipelines. I would like to see more experts vetting the Bayesian Networks methodology and making it more of a standard. The industry also needs more people trained in data science technology and Bayesian Networks.” <7p>
References:
Further reading:
Pipeline evaluation portal
Disclaimer:
DNV GL prides itself on providing accurate information but makes no claims or guarantees about the accuracy, completeness or adequacy of contents in this publication, and disclaims liability for any errors or omissions. The authors’ views here do not necessarily reflect DNV GL’s views.