London Business Conferences Group’s Machine Learning and Artificial Intelligence Congress 2018, held in Houston earlier this year heard from some early adopters of AI/ML in oil and gas. Along with the growing interest in these compute intensive technologies comes a desire to revolutionize onshore Scada systems, replacing incumbents’ expensive proprietary infrastructure with DIY systems that leverage consumer/hacker-style technology (think Raspberry PI) and pumping data straight to the cloud, avoiding again, legacy (and expensive) proprietary wireless and satellite links. So, what ‘disruptive’ startups are engaging in this data revolution? Well let’s start with BP.
Eric Penner presented an ‘intelligent operations’ initiative underway at BP’s Lower 48 Onshore unit. Intelligent operations (IO) is a portmanteau term encompassing new ways of working, using technology to automate routine activities. Analytics-backed logistics maximizes time available for on-site problem solving and ‘makes us a safer, more environmentally responsible operator.’ Here the focus is root cause failure analysis that Penner boldly states can be used to ‘understand our vendors’ businesses better than they do*’. RCFA mandates more data from operations which is where the Raspberry Pi, a $50 DIY computer, is used as a ‘disrupting alternative’ to traditional RTU/Scada systems. The Raspberry Pi underpins an ‘open platform’ that commoditizes endpoint devices and promises the economical automation of all of BP’s wells. The Pi is set to ‘shave millions’ from BP’s automation budget by eliminating costly legacy scada systems, technologists and wireless LANs with ‘consumer-style technology hooked into the cloud!’ Other components of BP’s IO include the FieldBit Hero visual collaboration platform for field services. BP was the first onshore oil and gas company to deploy the augmented reality, smart glasses that empower technicians to solve complex problems on-the-spot. More aligned with the ML/AI theme of the conference is BP’s ‘Arrow’ progressive logistics model and work management solution that uses a BP-developed algorithm to provide a value-optimized route for pumpers, reducing site visits by up to 50%. Arrow automates the task of efficiently deploying resources and eliminates wasted effort.
* This claim merits examination. While an operator may know more about the operating environment of an equipment item, the manufacturer has access to potentially many more data points from multiple clients. The issue of failure data sharing is somewhat fraught.
Johnathan Hottell (Exco Resources) is Raspberry PI enthusiast, although he nuances the new low-cost solutions. Traditional Scada ‘is not going away for real-time control’, the technologies complement each other. Hottell recommends teaming a subject matter expert with a data scientist. One use case is an investigation into well liquid loading. NodeRed’s flow-based programming for the IoT. This runs on a Pi and polls legacy devices. NodeRed can convert and forward messages using many different protocols. Exco’s well liquid lodading ML experiment is described on the Microsoft Azure ML Gallery. A full write-up with more on NodeRed and the PI is available on the Hackster hobbyist website. Ignition Scada also ran as did the Microsoft Azure ML Studio and Cortana AI gallery. Exco’s ML-derived liquid loading classifier is now deployed as a web service. Exco has also used machine vision to check for instance if a flare is still burning or if there is unauthorized vehicular access to remote well site.
* Industrial internet of things.
Mark Reynolds’ Digital Transformation startup is working on optimizing artificial lift with ML. Artificial lift has come a long way since the days of data logging and manual control of pumps to today’s situation where engineering means ‘watching everything all the time’. Reynolds cited Microsoft researcher Tony Hey’s presentation on ‘eScience and the Fourth Paradigm’ aka data-intensive scientific discovery and digital preservation. ‘eScience is the set of tools and technologies that support data federation and collaboration’. Driving the eScience transformation is the enterprise architect, an experienced, interdisciplinary oil and gas engineer and a ‘rock star’ software engineer with a decent understanding of statistics.
Ryan Stalker (Williams) observed that material movement logistics workflows have a tendency to inefficiency. Some are poorly engineered from the start, others get inefficient over time as entropy sets in. ML and AI offer opportunities to optimize workflows by reduce human intervention and automating inventory management, reducing stockpiling and starvation. Williams has worked with Calgary-based Code Expert to add ML smarts to its Maximo materials management system. The solution also leverages the Waze smartphone driving app to optimize equipment and material movement. Waze’s ML-based navigation software adds crowd-sourced information on road hazards, traffic patterns and police alerts.
The main limitation to the application of ML/AI in workflow optimization is the effort involved in digitizing legacy hard copy records which are required for algorithm training, validation, and testing. Stalker warns that ‘broad organizational problems won’t be solved by ML and AI’, although these tools can excel when pointed at specific workflow problems. ML’s usefulness decreases with problem dimensionality. Another potential problem arises from the regulator. In the EU for instance, new General data protection regulation extend to the ‘explainability’ of deep learning. Witness GDPR Article 13, Paragraph 2(f) requiring that people be informed of ‘the existence of automated decision making and [be provided with] … meaningful information about the logic involved’.
Mohammad Evazi (California Resources) had a stab at estimating the ‘size of the prize’ of optimizing rod pump operations. There are around 1,000,000 oil wells worldwide on sucker rod pump with an annual failure rate of 0.2-0.6 per well. The average cost of failure is about $30,000. Do the math! [We did, that is around $12bn/year!]. It makes sense to invest in collecting and understanding your pump-off controller (POC) data.
Today, CRC collects information on load, strokes per minute and casing pressure in Theta’s XSPOC database. But the bulk of historical data required for AI/ML is held in analog-format dynamometer cards. Dynocard data is a health indicator for a rod pump wells. CRC stores over 100,000 cards per day. These are analyzed with a k-Means classifier and translated into time-series visualizations. These enable well diagnostics and failure prediction. CRC has built its Dynocard classification model which is now used for well optimization. CRC Is now working to more accurately determine failure intervals and label the root cause. Card data is now permanently captured in a ‘failure data lake.’
More from London Business Conferences.
© Oil IT Journal - all rights reserved.