Is There Value in Predicting When and How An Equipment Will Fail? Royal Dutch Shell Thinks So!
See how Royal Dutch Shell plans to use predictive maintenance to improve profitability.
Imagine that you’re Royal Dutch Shell (RDS), a major player in the oil industry. You have an offshore facility capable of producing 200,000 barrels of oil and compressing 100mmscf of associated gas per day. You operate under a zero-flaring policy, such that if this associated gas cannot be compressed, the facility must be shut down. One day, your compressor fails. You shut down the facility. You commission your maintenance engineers to investigate. It takes them two weeks to trace the fault to a particular bearing within the compressor. Incidentally, you’re out of spares. You place a rush-order for this bearing and it takes a week for the bearing to arrive onsite. Your engineers quickly install the new bearing and oil and gas production is restored. The engineers are happy but you aren’t. You aren’t because you know that the 3 weeks of facility down time cost you about $250 million of revenue. Thus, you begin to ask questions: is it possible to predict when next the compressor will fail? Furthermore, is it possible to predict which compressor component will fail so that one can ensure a spare is onsite at the time of failure?
Indeed, the current lack of predictive maintenance costs the oil industry billions of dollars annually. According to a McKinsey 2017 report [1]
, a typical offshore producing platform operates at only 77% availability. Industry-wide, the shortfall comes to about 10 million barrels per day, or $200 billion in annual revenue. Based on RDS’s 2017 annual report [2] in comparison, the company’s plant availability stood at an impressive 91%. Notwithstanding, RDS seeks to achieve 100% plant availability. RDS has thus determined that the answer to the questions posed above is yes, and is proactively exploring the use of predictive maintenance to improve plant availability. According to the Wall Street Journal [3], RDS recently signed a three-year contract to use machine learning technology from C3 IoT and Microsoft Corporation’s Azure to predict when maintenance is needed on compressors, valves and other equipment. While it sounds promising, how will it work in practice?
The first thing a machine learning algorithm requires is data – lots of it. Using the earlier scenario, it will need the compressor’s instrument-generated historical data trends such as vibration, wear metals in lube oil, temperatures etc. It will also need human-generated data on past breakdowns, the failure modes and what the engineers did to restore the compressor to functionality. Fortunately, the former is currently stored (seemingly infinitely) using OSIsoft’s PI system which gathers and stores the data on a real time basis (See OSIsoft’s post on “Shell’s journey to Advanced Analytics” here [4]). Similarly, the latter is stored on RDS’s ERP system, SAP. My recommendation is that the algorithm should be trained using these datasets in an unsupervised manner. This is because maintenance engineers have not always been successful at aggregating the datasets and identifying relevant patterns to reliably predict tentative failure outcomes. Nevertheless, their input on actual failure outcomes will be required to validate the algorithm’s data clusters and patterns, thereby closing the loop on the initial learning process. Once the algorithm is put online, its learning can then be continuously reinforced by comparing its predictions to actual outcomes.
However, for such a data-centric system, RDS will need to focus on data quality. The instrument-generated data is not the concern here, as this source is typically accurate. The human-generated data, on the other hand, is one that will need to be monitored closely. No longer can front-line maintenance technicians be allowed to input generic failure explanations like “failure due to aging equipment” for example. Such inputs will be 
useless to the machine learning algorithm. Hence, RDS will need to significantly invest in technician training, and consider equipping technicians with ERP-linked mobile devices that can enable them input specific information about the compressors as they observe and/or fix them on-the-go.
Furthermore, RDS must ask the hard question of the staff impact in the long run. Imagine the ideal future where the algorithm becomes so accurate that it is enabled to take decisions like initiating the ordering of spare parts in line with its predictions, of what use will RDS’s current staff handling such roles be? What about the reliability specialist whose current role largely consists of optimizing equipment availability based on the equipment’s historical trends; of what use will she be? Finally, even if there is still a place for the specialist, how will a fresh college graduate ever become a specialist in a world ruled by an algorithm’s predictions? This is a pertinent question, given that in RDS’s long term view [5], such algorithms will be able to provide even more nuanced insights such as when equipment will experience performance declines. Perhaps in the new world, everyone may need to become a data scientist to remain relevant.
Word Count: 797
Sources
- Anders B, Trench M, Vermaat Thijs, 2017, “Why oil and gas companies must act on analytics” https://www.mckinsey.com/industries/oil-and-gas/our-insights/why-oil-and-gas-companies-must-act-on-analytics
- Royal Dutch Shell Annual Report, 2017, Page 18 of Strategic Report Sub-Section. https://reports.shell.com/annual-report/2017/servicepages/download-centre.php
- Norton S, 2018, The Wall Street Journal, “Shell Announces Plans to Deploy AI Applications at Scale”
- https://www.osisoft.com/Presentations/Shell-s-journey-to-Advanced-Analytics/
- https://www.shell.us/business-customers/lubricants-for-business/lube-space/the-benefits-of-predictive-maintenance.html
- https://www.shell.com/about-us/major-projects.html
- http://www.offshorepost.com/rigopedia/rigs/the-bonga
- https://www.arcweb.com/blog/predictive-maintenance-or-predictive-operations
I agree that through predictive maintenance, oil companies can generate vast value. However, I would like to bring new perspectives on use of machine learning (ML) at Shell on three dimensions. First, I do not think 100%, as stated by RDS, plant availability is achievable. Platforms have to go through annual or at least bi-annual maintenance and there are some parts that we cannot replace without stopping whole platform. Second, I do not think it is necessarily true that feeding ML algorithm with useful data requires technician training or equipping technicians with ERP linked mobile devices. Even printers at our houses currently can alert us about why it stopped working. I think instead of making the whole algorithm people dependent, I would explore automatic failure reporting systems through sensors. Third, impact of predictive maintenance would be even multiplied if Shell can successfully deploy predictive maintenance to its other assets such as pipelines, refineries and petrochemical plants.
This article does such a great job of explaining predictive maintenance and how machine learning is at the heart of this new phenomenon taking over manufacturing. Kudos Z_A!
I would like to call out three barriers in addition to those covered above that I believe RDS will have to overcome before adopting predictive maintenance:
1. Data Scarcity: you have mentioned data quality but what was surprising to me was that many parts of the Oil and Gas industry are actually a lot less automated than I had thought (esp. in Asia). Digitizing the production end-to-end is the first step to take before predictive maintenance can be a reality.
2. Digital talent attraction: not only will current roles and skills become obsolete as you mention, these new high-in-demand skills will be hard to find and attracting talent to a rig vs. another manufacturing environment will be something companies like Shell will have to consider.
3. Integration of the stack: connecting the Azure interface, with a front end and across digital machines is tricky operationally (although huge advances have been made so far)
Your article is very enlightening about the benefits of predicting maintenance problems in rigs and it’s impressive how RDS has implemented these techniques in their major operations. As I think about this, I can’t help but wonder how an ML algorithm would fare in predicting large scale disasters, especially because such disasters have been few in number so data can be hard to find and might be triggered sometimes by the most random and seemingly insignificant factors. To me, it looks like there’s going to be the need to do some sort of balancing between the signals generated by machines and the intuition of engineers – how this can be done I cannot wait to see!