January 2019, Vol. 246, No. 1


Subsea Bayesian Networks – Integrity and Risk Assessments

By Mahmoud Aboelatta, Chief Inspection Engineer, Wood

If high levels of safety and environmental protection in the subsea oil and gas sector are to be maintained, then understanding the risks involved is more important than ever. That was the view highlighted by the Bureau of Safety and Environmental Enforcement (BSEE) in 2016.  

(photo: Wood)

This is a significant challenge. The economic and regulatory environments mean that it has never been more difficult necessary to optimize inspection, maintenance and repair (IMR) activities.   

As each stage in the asset life cycle has its own challenges and uncertainties, assessing the risks to subsea equipment, understanding barrier health, developing risk mitigation strategies, and calculating inspection intervals in a commercially viable way requires several different approaches.   

Fully subjective qualitative models have a role to play, as do data-driven quantitative models and the semi-quantitative in between. Most operators are familiar with failure mode, effect and criticality analysis (FMECA), risk-based inspection (RBI) and bow-tie analysis.   

However, these techniques are still labor-intensive and tend to lean toward a more qualitative approach. Subject-matter experts (SMEs) must go through the data and make subjective judgments. If the SME is not available, the risk assessment is either delayed or based on potentially flawed assumptions.  

What’s more, the variables and uncertainties of subsea equipment and the changing conditions in which they operate lead to greater caution when it comes to design and assessment. This understandable preference for over-inspection gives operators greater certainty within their asset integrity programmes but comes at a cost that is less and less sustainable.   

Probabilistic Modeling   

More recently, risk assessment techniques that rely on probabilistic risk assessment (PRA) and analysis have been considered. They are less familiar to the oil and gas industry, but could offer significant advantages, particularly when applied in a hybrid approach with other models.  

PRA has its roots in the U.S. nuclear industry of the 1970s. By the early 2000s it was also considered by NASA for landing, take-off and mission control. Simply put, it applies a logical, systematic and comprehensive approach to delivering robust risk assessments. It mixes probabilistic with deterministic models to quantify risk levels and present them in monetary terms.   

PRA really came to the attention of the oil and gas industry in 2016, when the BSEE entered into an inter-agency agreement with NASA to explore its role in evaluating and communicating risk in the offshore sector.   

The statement from BSEE and NASA at the time pointed out that: “Not every situation requires a quantitative approach; however, PRA is appropriate for complex engineering hardware that has critical human interaction and multiple pathways to catastrophic failure.”   

In other words, PRA is highly suitable for the oil and gas industry.   

Bayesian Networks   

One of the most interesting and potentially beneficial forms of PRA is the Bayesian approach and the creation of Bayesian networks. These too have been used extensively by NASA and have been widely deployed in medical and pharmaceutical fields, nuclear risk assessments, and the detection and prevention of payment fraud.   

A Bayesian network shows the probability-based relationship between several events or features. For example, a Bayesian network could show the relationship between wall loss, internal corrosion, erosion, and other or events that are highly predictive of a loss of integrity.   

Given the complexity of interdependent events in the oil and gas sector, the number of events in each individual Bayesian network can be extensive. So, the starting point for a Bayesian network is an initial belief or hypothesis. A Bayesian algorithm evaluates the probability distribution for each event in the network, and uses the results to calculate a consolidated, overall probability of a specified event.   

In our example of showing the relationship between wall loss and loss of integrity, the initial belief is internal corrosion will lead to wall loss by the end of the year. However, because the network also considers all other probable causes, we may find that wall loss will be caused by erosion, and in a much shorter time frame.   

The Bayesian approach can be deployed in almost any scenario. For a new subsea field, where there is no historical data, information from similar fields or industry failure databases such as offshore and onshore reliability data (OREDA) can be used to establish the initial hypothesis.   

For existing subsea fields, information can be sourced from current systems and databases, including but not limited to failure registers, anomaly registers and inspection reports. In both cases, the experience of SMEs also proves to be a useful, if subjective, foundation for the initial hypothesis.   

Once the available data is gathered, a model can be developed according to the depth, quality and specificity of data available. Where the initial hypothesis is poorly understood, then an off-the-shelf-model can be used.   

Then as operations continue, more detail becomes available and can add depth and granularity to the original hypothesis. Inspection data from sensors, continuous monitoring systems, ROVs, divers and in-line inspection (ILI) can all be plugged into the model. The model can then recalibrate itself and update the initial hypothesis to consider the new data and observations.   

In other words, Bayesian networks are self-learning and self-adapting – which is why they are found in artificial intelligence and deep machine learning.   

In Gulf of Mexico  

The Bayesian approach relies on empirical data and a rigorous and repeatable calculation. It is derived from the consistent application of an objective algorithm. As a result, it is less labor-intensive than traditional approaches, less dependent on guesswork and “gut-feelings” and provides a consistent output.  

That’s the theory. To ensure it can be delivered in practice, Wood developed a Bayesian-network to assess multiple threat and damage mechanisms in a deep-water field in the Gulf of Mexico and used it to calculate the consequence and probability of two independent forms of failure: loss of primary containment and loss of production.   

To make sure the model could be used in multiple scenarios, it was designed to be both scalable and modular. Specific threats can therefore be added or removed as circumstances demand. To validate the model, sensitivity analysis was applied to test the underlying mechanism, and its output was compared to that of more traditional approaches and actual experience.   

Initial results proved both the robustness and the validity of the model. The Bayesian approach enabled the operator to predict risks based on observed data, to diagnose subsea fields based on the occurrence of a specific event and to perform risk assessments consistently. The operator was also able to optimize inspection activities, considering both cost fluctuation and equipment condition.  

There are further adjustments to be made. But inherent to the Bayesian approach is the ability to adapt to new data and evolve over time. Having demonstrated the validity of the Bayesian principle and its value to the subsea sector, the Wood approach, much like the model, is constantly learning.   

It is now undergoing its own process of adaptation and adjustment so that it can deliver the same substantial benefits in midstream and downstream scenarios as well as the subsea environment. P&GJ

Author: Mahmoud Aboelatta has a bachelor’s degree in civil engineering from Alexandria University, a master’s degree in asset integrity and over 10 years’ experience in the oil and gas industry, involved in developing, implementing and executing subsea and topsides integrity programs in numerous global positions. He has worked for Wood since 2013.  

Related Articles


{{ error }}
{{ comment.comment.Name }} • {{ comment.timeAgo }}
{{ comment.comment.Text }}