Artificial intelligence (AI) and its algorithmic engine, machine learning (ML), have the potential to make revolutionary changes in manufacturing operations, but it’s not a one-sized-fits-all solution. It can also be a complicated and expensive undertaking, where more simple solutions might produce just as good outcomes, or good enough, if budget really matters.
So how does a facility choose when and where to apply advanced analysis and AI/ML technology? AspenTech has long worked in the asset performance management (APM) market and has experience helping its industrial customers deploy AI solutions. Here, experts at Aspen Technology discuss some of the different levels of technology for asset monitoring, and how to choose where to apply your most advanced analysis as you consider AI/ML applications. To learn more about how a paper and pulp producer introduced new levels of predictive maintenance, read this article.
What is the right mix of technology solutions?
Vibration analysis has long been commonplace for asset monitoring in manufacturing. While data scientists can incorporate that data into model predictions, it can also be used for rules-based monitoring—if an asset’s vibration readings go out of a certain range, it triggers a work order. Similarly, with condition-based monitoring, if equipment vibration experiences a certain condition, it will trigger a work order. And that level of monitoring might be sufficient for the operation.
“What you want to have when you think about APM is a wide variety of ways to do this analysis,” says Don Busiek, senior vice president and APM general manager at Aspen Technology. “You really want to apply the right analysis to the right problem.”
Finding the easiest way to solve a problem or predict a failure will make maintenance workers lives easier, not applying deep algorithms and analysis to everything. “When you think about APM, you want to think, do I have a solution that gives me the breadth of different ways to approach a problem, because I want to take the simplest approach,” Busiek says.
Do manufacturers understand their assets as part of a larger system?
Every asset, depending on where the asset sits in the entire infrastructure hierarchy, will have a different maintenance strategy. AI technology can be successful using failure agents, where facilities know they have had historical asset failures. An agent is a self-contained software component that simplified engineering and data science tasks for users.
With this historical data, software can identify patterns in those failures to address them ahead of time. AspenTech software also uses anomaly agents, which detect anomalous behavior in an operation. Deciding which method is best to use is where customer expertise also comes into play. “We deploy different kinds of agents (or different monitoring techniques), such as rules and conditions, advanced first principles, AI/ML models (anomalies or failures), and custom codes written by data scientists,” said Pratibha Pillalamarri, senior product marketing manager at AspenTech. The company works closely with each customer and its experts to understand how it was traditionally addressing maintenance problems and how a new solution can help that process.
“That’s why we have different kinds of technology (rules-based or condition-based or AI/ML-based), so different assets require different treatment,” says Nithiya Parmeswarn, vice president of product management at AspenTech.
The extent to which assets are using sensors and automation should also be considered. It can be the less complicated equipment that are most critical to production. “This requires interaction with the customer to identify what are those critical assets,” Parmeswaran says.
How do you choose the right assets for deeper analysis?
Choosing the right assets for AI/ML deployment is key, Parmeswaran says. The first questions to answer are, which assets are giving your facility the biggest problems, or which failures are most critical to production? Oftentimes, manufacturers might want to focus efforts on their most expensive assets, but that might not be where APM provides the most value because those expensive assets might already have extensive maintenance and upkeep or need less of it. Failure prediction models might not reveal much of a newer, well-maintained asset.
“If I’m just looking at an individual asset, I really can’t understand what I need to do differently, to better maintain it. I need to think about the entire process, the entire line, how my production works. If I’m thinking about it more holistically, I can really understand what I need to do differently, such as adjust that output, adjust the speed, whatever that may be that’s going to make a difference in how I maintain that equipment,” Busiek adds.
How much asset failure data is needed?
To start using AI or ML technology, AspenTech would need some failure information about the asset, which will often come from an enterprise asset management (EAM) system. “That’s where the truth of the asset lies,” Parameswaran says.
There’s not a specific timeline for failure data needed, but the more failure information you have, the better your model is going to be. “Machine learning is all about training,” he says. “These technologies automate what we already know, but it’s automated much faster than a human brain can process.” And it’s not always the volume of data that’s most important; it’s more about quality than quantity.
How should technology incorporate end-user subject matter expertise?
In the case where a facility doesn’t have specific failure data, but it has monitoring information from sensors, this is where different technology like anomaly detection can provides clues and conclusions about asset health.
It also can help understand how closely some of the asset sensors are correlated. “Normally, they are in tandem, like the pressure, the temperature, the regulation, they all have some kind of pattern,” Parameswaran says. Deviations from that typical pattern signal an issue. “We just don’t know what it is because we have not seen it before,” he adds.
This is where the end user subject matter expert can help identify what’s causing the vibration. For example, did the process change? Did the feedstock change? Did the outside ambient temperature change? “When you don’t have previous failure information, but a lot of sensor data, subject matter experts need to come in to further qualify why certain things are happening,” Parmeswaran says.
How can technology help manufacturers walk the fine line between throughput and quality?
APM can also help manufacturers walk the line between product throughput and quality. While manufacturers have production targets based on customer demand and a throughput they need to maintain, they also have quality standards to meet. If a manufacturer has an asset that’s starting to degrade, but it can’t afford to bring the asset down because it needs to hit production targets and quality standards, adjusting process parameters with the help of technology might get more life out older equipment, by sacrificing throughput and maintaining quality. It might extend an asset, for example, from 60 to 90 days, where maybe there’s more room for production interruption or more time to plan for the inevitable. This also must be a discussion between operations and maintenance, where operations dictates those production targets, and maintenance prescribes a plan to fit the goals.
“What AI does is it provides technology where it automates what we already know. It’s not going to figure out anything new, but the thing is, it compresses the time to do those tasks,” Parameswaran says.
What is the optimal time to do maintenance?
APM technology can also incorporate more information to specify exactly when to perform maintenance – not only equipment data, but production goals, scheduling demands, supply issues and more. The more you feed the system, the more it can optimize.
“It’s also telling you, this is the right time based on the output that you need to have based on the equipment that you have, based on how many employees you have, what shifts they’re working, this is the optimal time to do that maintenance,” Busiek says.
Defining business goals: What is a typical project implementation?
For a new APM project, AspenTech says a typical implementation starts with a kickoff meeting, which should include the digital/IT team, maintenance and reliability, and operations. Part of the meeting should identify: what are the business goals behind the project? The goal can’t be simply to use digital technology or machine learning. That might be a final solution, but the goals need to be based on business operations to start. “Sometimes we do a vision workshop, where we’ll understand what their vision is for the future. Where do they want to go? What is it they want to have five years from now?” Busiek says.
All of that is figured into an implementation plan. Depending on a companies’ digital resources, AspenTech will teach it how to set up agents and how to define rules, if necessary. Usually, project managers from both sides meet weekly to check in on the progress and send out status reports to make sure executives teams are aligned on goals as well.
Busiek said coming out of the pandemic, more and more companies do not have the in-house resources to do modeling work or even lower-level analysis, and AspenTech can shoulder that full responsibility too. “But our favorite kind of play is to do it shoulder to shoulder with the customer and then have them maintain the software,” Busiek says.
Another aspect to the implementation is what AspenTech calls an executive cadence, which is a regular checkpoint meeting with the customers’ executive team and AspenTech project leaders to make sure the project is driving the right business value.
AspenTech also finds value in learning from customers about how they are using the software or tweaking certain aspects. “I want to understand how you’re using the software because I want to feed that back into the next version that we come out with,” Busiek says. As they’re building the next release, AspenTech continues to checkpoint in with customers about new features. “Our vision of how we do R&D with our customers is a virtuous circle,” he adds.