Plants increasingly are turning to machine learning (ML) and artificial intelligence (AI) to recognize precise patterns in sensor data. The technologies ease differentiating between normal and abnormal equipment behavior and also detecting specific patterns that lead up to failures — and, so, enhance capabilities for predictive maintenance.
Meanwhile, advances in sensor technology are reducing both procurement and implementation costs and, thus, spurring greater adoption. “It’s now economically feasible to roll out shadow sensing technology in both greenfield and brownfield applications to capture high fidelity data in volumes that were unachievable several years ago,” notes Jim Chappell, vice president, information solutions for Aveva, Chicago.
The incentive for adopting predictive maintenance is compelling, stresses Mike Brooks, senior director, APM Consulting, Aspen Technology, Bedford, Mass. “A European customer tells us that 15% gross margin losses are attributable to unplanned maintenance. Even best-in-class approaches 4–5% losses.”
He also points to a large automation company’s finding that 63% of scheduled maintenance is unnecessary and such work often causes more damage than if the equipment were left untouched. He cites the experience of a large oil company: out of five automatic tank-gauging systems, the only one giving little trouble at all didn’t get regularly scheduled maintenance and inspection; the other four, subjected to regular planned maintenance, constantly broke down.
“So, the spend in maintenance today is in searching for wear-and-tear conditions while the problem is in process-induced failures caused by operating equipment outside safety and design limits: incorrect setpoints, pumps running dry and cavitating, compressors affected by liquid carryover, and so on,” he emphasizes.
“People talk about the IIoT [Industrial Internet of Things] like it is an initiative. The initiative is always to improve operational excellence. The technology is how you do this. The most important thing you must solve is improving equipment uptime. Mtell, Aspen’s predictive and prescriptive technology, is borne out of operations and maintenance — so stopping the breakdown of machines is right at the heart of what we do,” he adds.
Now in its fifth generation, Mtell uses small pieces of software called anomaly and failure agents. The anomaly agents constantly monitor to detect irregularities while the failure agents recognize a pattern of behavior that leads to the breakdown of a single machine or process.
“They can tell you what will happen, when/why it will happen, and what you can do to avoid or mitigate the outcome. Mtell lets ‘Joe Normal’ solve very complex problems without having to understand how it works,” explains Brooks.
Mtell allows users to scientifically decide on the most appropriate maintenance schedule, he underscores: “This is important because refiners and chemical plant owner/operators tell us they spend 200% more on maintenance than they should but cannot realistically determine which preventive maintenance routines to cut back on.”
Brooks cites the example of refiner and fertilizer manufacturer Borealis, Vienna, which had applied expensive vibration systems and reliability-centered maintenance techniques on the hyper-compressors at one of its refineries but still was losing millions of dollars’ worth of production annually.
Aspen found the root cause of the problem to be issues with packing seals and poppet valves. Such problems now are flagged with 30–50 days’ notice. This has enabled the refinery to avoid unplanned shutdowns; the new technology also has completely eliminated false alarms.
Another example is from Saras, Sarroch, Italy, which operates a 300,000-bbl/d refinery and 575-MW integrated gasification combined cycle plant. As part of the company’s digitalization initiative, it wanted to employ prescriptive maintenance to reduce unplanned downtime.
Aspen used Saras’ existing condition data and maintenance records to build agents for a subset of compressors and pumps. The failure agents accurately predicted two valve events: a high outlet temperature failure with 39 days’ lead time and an instrument failure with 25 days’ lead time that would have led to a valve replacement. They also identified numerous process equipment failures such as oil leaks on pumps. Here, the agents achieved a 91% detection accuracy with 30 days’ lead time, reports Brooks.
The next step for Aspen is culling data from similar process equipment at different sites to find non-variable conditions common to them all. “This way we can solve bigger and more challenging problems. We call this ‘transfer learning’ and expect to do a lot more things like it in the future,” says Brooks.
Rethinking Strategy
To remain competitive, chemical companies must take a more holistic approach to digital technology and look beyond traditional techniques for increasing margin, counsels Chappell.
“Many companies today focus on the gains achievable through process optimization. However, a single critical asset failure in a chemical manufacturing application can wipe out years of savings from process optimization,” he cautions.
Aveva’s answer to this is to combine its process simulators with ML, an approach that enables modeling a wide range of assets and processes — including the dynamic periods when processes are starting up or in transition.
To do this, the company’s Predictive Asset Analytics technology uses a patented algorithm called OPTiCS that applies advanced pattern recognition and ML technology. For systems with lower levels of historical repeatability, high noise or that use process-driven problem solving, Predictive Asset Analytics uses a predictive algorithm plugin called KANN.
“This algorithm allows users to create models that predict future values for signals. It uses artificial neural network technology and allows users to create operational profiles with a specific set of inputs and outputs, testing how the outputs will evolve in the future through data playback and ‘what-if’ analysis,” he explains.
Predictive Asset Analytics learns an asset’s unique operating profile during all loading, ambient and operational process conditions. Existing machinery sensor data are input into the software’s advanced modeling process and compared to real-time operating data to determine subtle deviations from expected equipment behavior and alert if necessary. Once an issue has been identified, the software can assist in root-cause analysis and provide fault diagnostics to help the user understand the problem reason and significance.
“Within the chemicals industry, rotating equipment such as compressors, turbines and pumps have traditionally benefited the most from predictive analytics technology. Some of Aveva’s larger customers have (conservatively) estimated over $34 million savings for individual avoided ‘catches,’ and total avoided costs at over $100 million each,” notes Chappell.
Figure 1. Facility in Abu Dhabi controls and optimizes, including via predictive and prescriptive analytics, operating company’s entire end-to-end value chain. Source: Aveva.
Companies using Aveva’s technology include Covestro, BASF, Air Liquide and Abu Dhabi National Oil Company (ADNOC). The latter firm has installed Aveva’s enterprise visualization and integration technologies, including predictive and prescriptive analytics, at its Paronama digital command center in Abu Dhabi (Figure 1). ADNOC uses the center to monitor, control and optimize the performance of the entire end-to-end value chain across its 16 operating companies.
A major remaining challenge is the cleansing and curating of existing big data to make them suitable for and useful in predictive and condition-based modeling software, he says.
“Over time, we anticipate a shift in maintenance strategies from the traditional cost-centric approach to a more revenue-centric approach. Combining cloud, edge, IIoT and predictive analytics enables companies to develop revenue streams through maintenance-as-a-service programs,” he adds.
Predictive analytics also is advancing rapidly as well, with anomaly detection no longer the state-of-the-art, says Chappell. “To stay cutting-edge, software must predict the future.”
So, Aveva’s ongoing developments focus on leveraging deep learning to forecast the remaining useful life of an asset and, from there, provide prescriptive guidance for maintenance and remediation.
The Value Of Experts
The digital revolution is catalyzing improved outcomes using advanced analytics, stresses Michael Risse, vice president of Seeq, Seattle, citing the company’s rapidly growing list of commodity and specialty chemicals customers as proof. “The drive to gain better insights from existing data is absolutely there, and Seeq is working on spreading awareness.”
Rather than a black box technology that tries on its own to recognize patterns in data, Seeq enables subject matter experts (SMEs) to define what’s important to them or what they’re looking for. To this, they can add context based on experience and other data sources, such as manufacturing execution and laboratory information management systems, and investigate to identify precursors to asset failure or degraded production.
“When SMEs put these efforts into action, in the form of monitoring and analyzing incoming data, they are able to accurately predict failures and prescribe mitigations,” he says.
They can build predictive models that are adaptable over time as plant conditions change — an interactive approach that actively promotes sustainability in the predictive analytics, rather than a one-time optimization effort, Risse adds. The benefits of such an approach include improved asset uptime and proactive rather than unplanned and costly reactive maintenance.
As an example, Risse cites a global analytics program instigated in late 2017 by Abbott Nutrition, Lake Bluff, Ill. One of the pilot studies that Seeq conducted managed to identify the root causes of why some clean-in-place (CIP) runs took longer than others — typically leaking valves and instrumentation failures —highlighting those parts of the CIP process that Abbott should focus on to make further production improvements.
New sensors and more data rarely are the issues these days, Risse is quick to point out. What users want is faster, meaningful insights from the data they already have stored in process historians: “For all the fancy talk about AI and ML, the number one tool for finding insights in process data is the spreadsheet, a 30-year-old innovation. What the spreadsheet lacks in capabilities, it makes up for in terms of accessibility by the SMEs who have experience/expertise in the plant. Whatever the future may hold in terms of AI and ML, it must maintain that level of ad hoc self-service access for SMEs.”
Faster Detection
Meanwhile at Yokogawa, Tokyo, efforts are focused on using its technologies to drill down and better identify and understand individual equipment issues, for example, cavitation in pumps.
“Rather than relying on sound and vibration, we have devised a method to measure cavitation accurately in a different way by monitoring minute pressure changes. This ability to detect very low levels of cavitation has been further enhanced by ML analysis,” notes system products marketing specialist Masaru Kimura.
The system detects the minute pressure fluctuations caused when cavitation bubbles pop (Figure 2) using the oscillation value parameter on the company’s DPharp differential pressure transmitter.
Figure 2. Different pressure transmitter detects data that enable machine learning to predict the onset of cavitation. Source: Yokogawa.
“By using the data detected by this system for ML, it becomes possible to construct a model for detecting signs of cavitation. Usually, the most important barrier in machine modeling analysis and modeling is the collection of labeled-learning data but the cavitation detection system can be used to automatically create it. The data can then be used to construct a prediction detection model of cavitation,” he explains.
This, in turn, does away with the need for operators who are experts in cavitation, and allows equipment changes to be made while minimizing reductions in operating load.
“The system has only just been released and already we have received feedback from field trials showing it can help extend asset life and improve efficiency by reducing or eliminating operating losses,” he adds.
Kimura sees the chemical industry as a very important market because its cavitation issues are more pronounced than those of other sectors. As sensors become more accurate and data collection and processing faster, he also believes other phenomena will become measurable, enabling further reductions in customers’ asset maintenance costs.
Seán Ottewell is Chemical Processing's Editor at Large. You can email him at [email protected].
