Machine learning, the backbone and main enabler of artificial intelligence (AI) systems, is giving the chemical industry ever greater insights into its operations and maintenance — as the experiences of Nouryon, AVEVA and AspenTech show.
As part of its strategy to pioneer new technologies, Nouryon Industrial Chemicals, Amersfoot, the Netherlands, has signed a framework agreement with Semiotic Labs, Leiden, the Netherlands, to use that firm’s SAM4 self-learning technology to help predict when to maintain and replace pumps and other rotating equipment. SAM4 relies on voltage and current waveform analysis.
AkzoNobel Specialty Chemicals, the predecessor to Nouryon, started working with Semiotic Labs in 2018 after Semiotic was a winner in that year’s Imagine Chemistry Challenge (“Imagine Chemistry Challenge Awards Support to Startups,” a contest the chemical maker ran specifically for startups involved in areas of interest to it.
The new deal follows a successful 6–7-mo. pilot implementation of the technology at Nouryon’s chlorine plant at Ibbenbüren, Germany; that pilot focused on 20 pieces of rotating equipment. Nouryon now is rolling out SAM4 at its seven other chemical sites in northwestern Europe. The company expects to monitor 30% of the rotating equipment at these plants by the end of next year. A further rollout may take place depending on the results.
“We selected Semiotic Labs technology for a number of reasons: first, the sensor isn’t placed on the rotating equipment itself but in an electrical cabinet and can be accessed remotely [Figure 1]; second, it’s easy to install; and, thirdly, I had a good experience with the company previously, when they showed what they could do for the rail infrastructure reliability in the Netherlands. The bottom line is that we need to have knowledge about our assets,” says Marco Waas, director R&D and technology for Nouryon Industrial Chemicals.
During the Ibbenbüren pilot, SAM4 identified three potentially very important issues.
One involved the motor driving a huge conveyor belt transporting salt. Once the initial problem was corrected, SAM4 pinpointed yet another problem: salt deposits falling off the conveyor were increasing friction in the pump, too. “So not only the pump, but also the conveying process itself was an issue,” Waas explains.
The company also found it had installed an inappropriate pump in one location. The unit was oversized, consuming more energy and likely suffering a shorter operating life than a correctly sized pump.
In the third case, SAM4 revealed that a pump was operating at a point near to where cavitation could occur. “We are monitoring this very closely and deciding which type of pump is best to replace it with,” adds Waas.
One benefit Waas didn’t anticipate beforehand is a roughly 15% reduction in pump energy use; the company currently is working to better understand the reasons for this.
“Of course, the more data you have, the better the predictive ability of the sensors to identify important fingerprints in the data. So, we are pooling our data with two other companies who are using the technology on their rotating equipment: Vopak [Rotterdam] and Schiphol Airport [Amsterdam]. The three of us share data to improve our ability to detect potential problems,” he notes.
Figure 1. Sensor mounts in electrical cabinet rather than on rotating equipment. Source: Nouryon Industrial Chemicals.
Tank storage company Vopak trialled the technology on business-critical pumps at its sites in Vlaardingen, the Netherlands, and Singapore last year. It now has signed an agreement with Semiotic Labs to scale up SAM4 deployment to additional terminals while expanding its use at the original pilot sites.
Nouryon Industrial Chemicals also is testing two other predictive maintenance technologies: small inert pipeline sensors known as Pipers (a 2019 Imagine Chemistry winner — see: “Imagine Chemistry: Two Startups Win Novel Contest”) from Ingu Solutions, Calgary, Alberta, Canada, that freely float down a line and collect valuable data about pipeline integrity; and a climbing robot from Invert Robotics, Eindhoven, the Netherlands, which was a 2018 winner.
“We have tested Pipers in a 24-km salt pipeline in Denmark to measure all kinds of parameters, including temperature, pressure and the condition of pipewalls. They are very useful because most of our pipes are underground, so knowing exactly where a problem is developing makes maintenance much easier,” stresses Waas.
Because the 8-kg robot climbs, it can handle duties such as inspecting storage tanks. “Before, you would have to scaffold throughout a tank and use maintenance staff in situations that can be very hazardous. Now, we only need to scaffold to any areas that the robot identifies as having problems, for example cracks,” says Waas. The use of machine learning technology coupled with more data from the new diagnostic tools will make it easier and faster to identify and deal with potential problems, he adds.
A patented predictive algorithm at the heart of the machine learning technology of AVEVA, Cambridge, U.K., plays an integral part of a larger AI strategy aimed at anomaly detection, predictive forecasts, prescriptive guidance and more.
“The company’s predictive analytics system has detected many hundreds of millions of dollars in avoided costs across its global customer base — including several individual ‘catches’ that have been in the tens of millions of dollars each,” notes James H. Chappell, global head of AI and advanced analytics, Houston.
One example is from an industrial gas manufacturer that, prior to a scheduled maintenance outage at a plant, identified a vibration sensor anomaly. Technicians found a cracked impeller in a turbo-compressor. This early catch prevented reactive maintenance and unplanned downtime, saving over $500,000.
“Compressors are an excellent candidate for machine learning to provide early detection of issues, and customers have leveraged our technology on hundreds of these types of equipment,” stresses Chappell.
The company now can generate high-dynamic-range predictive models that can provide a wider range of detection, have increased sensitivity for earlier detection, and can work with grassroots plants or units coming out of major maintenance where data are scarce. They should help users optimize operations and get key guidance for cost-versus-risk decisions.
Over time, Chappell foresees machine learning evolving to handle larger datasets both faster and with fewer false positives. AVEVA already has developed a capability for “grey box” modeling, i.e., combining first principles simulation and AI.
“There are many situations in the chemical industry where first principles algorithms aren’t accurate enough to properly model certain assets and processes, or it becomes too time consuming — and costly — to adequately tune the algorithms,” he explains.
AI technology such as advanced neural networks can simulate those assets or processes better through data-driven learning. AI determines output values for integrating into the overall simulation model. “This provides significant cost saves and reduced deployment time. For the chemical industry in particular, this is especially valuable,” notes Chappell.
The ultimate goal is to have machines make recommendations and humans do the final interpretation based on the information available. To this end, AVEVA currently is infusing AI into software across all areas of its business.
“What’s changing is that the technology is now much more accessible and engineers find it much easier to apply. You do still see little pockets where they see the technology and decide to try it only in one or two very specific situations. However, one-off solutions are not enough. You have to look at plants, whole facilities, other sites, etc. You need this technology across the business,” emphasizes Paige Marie Morse, chemical industry director, AspenTech, Bedford, Mass.
In any project, AspenTech typically looks first at two specific issues: equipment reliability, so it can get some warning before any failures occur to avoid safety/health/environmental impacts; and overall process performance. Hence, many of the company’s tools, including automated machine-learning condition-monitoring Mtell and multivariate analysis tool ProMV focus on these two areas.
Figure 2. Looking at equipment reliability and overall process performance separately blunts benefits achievable. Source: AspenTech.
“However, you must remember that these two issues are interwoven and that you can’t look at them in isolation [Figure 2],” Morse cautions.
The key is to analyze historical data from when a plant was operating well and then look for deviations when it isn’t.
Different sectors of the industry pose disparate challenges, too. For example, batch processors often use the same assets to manufacture a variety of products — some perhaps in high demand and others less so. Most profitable operation depends on maximizing output of high demand products. “We can use simulation here to help these companies get creative with their assets,” she says.
Another imperative is achieving process consistency, either between runs or across multiple lines.
As an example, she cites a batch chemicals manufacturer with three identical process lines running next to each other. The first one always manufactured the same product while the slate of the other two varied. In ten years of operation, the company noticed that, for a few days every year, at a random time, one of the lines not dedicated to the specific product would go off-specification. Analyzing the data with ProMV showed this only happened when one of the other lines made a particular product. The analysis determined the link between the lines that generated quality issues, which then was resolved.
“Digital technology that uses machine learning provides insight on operating anomalies that can create problems with equipment. Mtell uses machine learning to analyze production and equipment data, and learn from previous performance to predict a future breakdown and suggest an alternative solution,” Morse adds.
In another case, this time for commodity chemicals, a European petrochemicals producer used Mtell to develop a data-driven approach to maintenance planning. The new plan enabled the company to eliminate two days of shutdown per year on each piece of equipment, saving $1.8 million/y in downtime costs.
Another issue — running assets safely at less-than-optimum throughput — is one that AspenTech increasingly faces: “This is hugely important in the case of separation towers, for example, where the trays are designed to work best at 90% capacity. If you dial back to 80%, there’s no way they will be as efficient because you get leakage through the trays,” she notes.
“At the same time, sensor technology is getting deeper and cheaper. So, there’s even more data now but our algorithms are getting more efficient at detecting which of this data is garbage. You have to clean up the data so that you are left with the real data sets. We are getting much better at this by working closely with customers to define outlying data and so building AspenTech experience with chemical and refining processes,” Morse stresses.
The company also is improving its AI capabilities with the 2019 strategic acquisitions of Sabisu, Redcar, U.K., and Mnubo, Montreal. Sabisu’s flexible enterprise visualization and workflow technology has been rolled into Aspen Enterprise Insights. Mnubo can assemble and deploy AI-driven Industrial Internet of Things applications quickly at enterprise scale and will help AspenTech’s vision for the next generation of asset optimization technologies that combine deep process expertise with AI and machine learning.