By Paul Wiseman
It seemed like a great idea at first. There’s all this data being aggregated from thousands of wells—why doesn’t someone set up some parameters for pumps, tanks, pipelines, and anything else that can exude data—then computers can send alarms to alert managers and field personnel any time there’s a problem!
Very quickly, those personnel learned that a well with one aspect exceeding its parameter—high or low—usually did not indicate an actual problem. The flood of alarms became its own problem. The machines needed to learn how to discern the difference between a problem and not-a-problem.
Richard Link, senior market manager, software solutions for North America-based service company Zedi, says it was a lot like the Aesop’s Fable of the little boy who cried “wolf.” Said Link, “I’ve seen it where they [oil company employees] haven’t even been looking at the alarms because it just becomes ‘cry wolf,’ so they don’t bother looking at it.
“But machine learning is when you take what would be an alarm and you compare that to a document that [records] normalcy of operations and you determine, ‘Is that really an alarm?’ or is it really outside of the true parameters of what’s happening? Or is this an anomaly that has happened before?”
How exactly does the machine “learn” to ignore certain alarms? Link gave an example of how that’s working with a Zedi pipeline leak detection program.
“We’re setting up scenarios that push the system outside of its normal parameters and we’re letting the software learn all of that; learn what’s normal so that it can determine what an alarm is. Then the second step is somebody sitting down and saying, ‘Yes, this is within normal or this is without normal.’”
Link noted that the term “machine learning” sounds like the machine doesn’t need any help, but it actually requires significant expert input from humans, especially at first.
In addition to filtering out alarms that are not relevant, machine learning can factor in previously unnoticed information. If a compressor seems to fail suddenly, technicians can scour the machine’s recorded data, going back a week or two, and sometimes discover a slow change in temperature or pressure that never exceeded parameters yet still may have given clues as to the cause of an impending failure.
Link stressed that machine learning is not designed to replace humans, only to help them become proactive instead of just reactive, and to make everyone more efficient. It also creates new, higher-level jobs, because machine learning requires people to help write programs, to “advise” the machines and to maintain and upgrade computers, communication systems, sensors, and all the other component parts of machine learning.
Further efficiencies can be found in sharing of data, where appropriate. Zedi has one client who receives Zedi’s field data, reformats it, and sends it to a different vendor, who is not a direct competitor with Zedi. Link said Zedi has now taken steps to send the data directly to the other vendor, simplifying and speeding the work and relieving the client of a burden.
Mark Cowan, P.E., is Senior Consulting Automation Engineer for his own firm, Cowan Consulting Services, LLC. For him, at least part of the solution to “wolf syndrome” lies in having enough data to provide context.
For example, when a tank’s high-level sensor goes off, that’s only one part of the puzzle. “You may or may not have time to respond. What you need is the high-level sensor [plus a sensor that shows] how fast the level is rising. That extra context makes a big difference.”
If the fill rate is only one foot per day, this situation does need to be remedied, possibly by a visit from a tank truck, but there is not an imminent, drop-everything spill danger.
As more and more companies add SCADA systems and begin mining data, Cowan says many are Lone Rangers, with custom systems. “There are good platforms out there to build that one, but in terms of just jumping into it and saying, “Here’s what you need, go!” that doesn’t exist yet.”
“Yet” is the key word, because Zedi’s Link says his company and others are being asked to develop programs that can be used across the board, with just a few tweaks. He notes that few companies are big enough (a few majors and larger independents being the exception) to develop and host data management systems on their own. A growing number of companies are more than willing to farm out this development so they can concentrate on their core business.
All three participants in this article agreed with Cowan’s following premise: “It’s delegating all the routine tasks to machines, and a lot of the decision-making and number-crunching to machines,” letting humans make bigger strategic decisions and plan new ways to use this data.
One new idea Cowan listed is in pipeline “peak loading.” A system that monitors the operation of all pumps on a pipeline can make them more efficient by staggering their run times rather than have all run wide open all the time. Software monitors pipeline pressure and activates each pump as the pressure at its entry point is at a low point—making pumps run more efficiently by reducing the back pressure. By reducing the producer’s peak usage this system can also drop per-kwh costs.
Much of the spotlight that shines on technology today is pointed at production and back office software. There are indeed cost savings and production increases to be had from the implementation of these systems. But Weatherford’s Global Director of Well Control Technology, Robert Ziegler, sees more promise for technology on the front end—drilling and completions—than for that on the back end.
He sees technology as a tool for growth more than for cost savings. “Nobody has saved themselves out of a crisis; you only can grow out of it,” he maintains, through the use of technology to make drilling faster, safer, and more efficient.
Ziegler notes that today’s companies are often led by CFOs who understandably are focused on areas they are familiar with—the office and day-to-day operations. But in the bust of the 1980s, which lasted well into the 1990s, he said more companies were led by engineers who focused on improvements in drilling and production in order to survive.
“Managed pressure drilling,” specifically, uses machine learning to monitor and anticipate drilling events, known as kicks, which can slow or halt the process. “If our system does its thing there are no more kicks because any inflow event is detected and countered so that it never develops into a kick that you have to deal with using normal well control measures.”
He told of a company that drilled three appraisal wells in Malaysia using standard drilling procedures. Due to the particularly difficult geology in that unfamiliar region, the three wells took one month each to penetrate to the production zone.
For the fourth well the company switched to managed pressure drilling with machine-learning-based software controls—and the wells were done in one day.
To be sure, Ziegler was quick to point out that this pedal-to-the-metal increase was an outlier—most completions improve by closer to 30 percent than almost 970 percent.
How much improvement is still in the tank? Ziegler referred to a 2016 SPE paper entitled “True Lies: Measuring Drilling and Completion Efficiency,” which concluded that, on the average, drilling efficiency could be improved by 50 percent before reaching technical limits of physics.
For those that fear that technology will replace humans altogether—one gets the image of the 1960s futuristic sitcom “The Jetsons” where title character George Jetson’s job was simply to push a button all day at Spacely Sprockets—Ziegler says there is no reason to fear. Machines will always only make humans better because machines can’t dream, they can’t innovate. What they do best is to handle rote jobs like aggregating and sorting millions of lines of data at blinding speed, which helps humans be proactive in ways they never could before.
For that reason, he is definitely not a believer in the unmanned drilling rig idea that has circulated in the last couple of years.
So where does the main impetus lie for implementing these new systems, which combine speed and safety?
“What we see more and more, especially in the offshore market, is that drillers want to at least have MPD-ready rigs or have a full MPD package as part of their competitive advantage because nowadays it is much more a buyer’s market than it was in the heyday,” he said.
Onshore, he said many of the top level drilling companies want to have a wide variety of services available, with more and more of them also offering MPD packages.
In the end, however, the operator specifies what systems he requires to be in place.
At the same time, many drilling companies who for two years have viewed the depressing sight of forests of parked drilling rigs on their lot may be reluctant to fork over the capex for new equipment with what Ziegler called a “20-year payout.” This is especially risky in view of the unpredictable yet constant up-and-down oil price cycles.
“I think there will be a lot of very interesting new contract ways, where basically the operator will give loans to service companies for major capital expenses for things like that because, otherwise, the service companies won’t do it,” he said.
Technology is no longer the wave of the future alone, although it is that. Cowan maintains that slow adopters will lose out, possibly even ceasing to exist, as competitors rush to become more efficient and more proactive. All three experts see technology as a friend, not an enemy.
And with the next generation of managers flooding into the oil patch with their smartphone apps and social media platforms in hand, the future is happening anyway.
Paul Wiseman is a freelance writer in Midland.