Your industrial AI checklist: 10 things you need to get started

Industrial AI is becoming a crucial technology in automation. But for those new to AI technologies, it’s not always clear where to begin.

By Robert Huschka March 5, 2024
Courtesy: CFE Media and Technology

Industrial AI insights

  • Initiating AI integration should focus on solving specific, real-world production issues to ensure practical and beneficial applications.
  • Assessing the value of AI involves analyzing its impact on efficiency, cost savings, and intangible benefits, such as enhancing safety and sustainability, with attention to scalability.
  • The development of AI solutions benefits more from domain expertise within the organization than from data science skills alone, underscoring the need for a deep understanding of the problem area.

Companies can’t escape the drumbeat around artificial intelligence (AI). Everywhere they turn, there’s a TV ad pushing the latest AI software. It seems like every vendor is pushing the latest tool kit. Every day, there’s an article on a new use case for ChatGPT, Bard, etc. The world seems to be screaming: People need AI now!

Artificial intelligence is indeed becoming a crucial technology in automation and industry. But for those new to AI technologies, it’s not always clear where to begin. How is AI be applied to automation applications? Where can it have the most value? How is success measured?

Consider this an AI checklist. Here are 10 things manufacturers should consider when evaluating AI for their own business operations and the steps they should take when introducing AI into their processes:

1. Define the problem that needs to be solved

Don’t use AI just to use AI. Start with a real problem, a real production issue.

“You don’t want to approach AI by saying, ‘I want to use AI for vision processing,” says Jeff Adolf,  enterprise vision specialist at 3M and responsible for introducing AI to many of 3M’s vision-based automation processes.

“For example, you might have a production bottleneck because you don’t have enough trained inspection experts,” Adolf says. “With AI, you can extend the capacity of your current pool of experts by having AI identify parts that can be clearly passed or failed without human interaction. This means your experts can focus more of their time on the parts that require expert inspection.”

When defining the problem, Nick Blum, specialist data engineer at 3M, suggests focusing on processes and searching for relationships between process data and process performance.

“For example, it can take weeks just to wrangle the data together to identify the root cause of a problem on the factory floor. With advanced process analytics based on AI, we can help our engineers find insights to problems in minutes or hours,” Blum says.

When defining the problem, Blum suggests starting by assessing the business impact. “It’s crucial to understand and quantify the business impact of the problem before attempting to solve it. Start by looking at the efficiency of your assets (OEE) to determine where the biggest opportunities lie. Quantify the impact (dollars) associated with improvements in OEE, establish a baseline, then explore potential solutions (AI or other) to the problems. In my experience, this type of top-down approach is far more effective than working from the bottom up. Fundamentally, it’s just basic Six Sigma problem solving methodology.”

2. Determine the real value of an AI solution

While AI can improve automation efficiency and accuracy, it may not be the best solution for an application at this time. For example, a particular AI solution might reduce the number of defects in a line by 50%. However, if the defect rate is already at 1%, the savings of 0.5% defects may not justify the cost of AI infrastructure, never mind the cost of developing the solution and interrupting the line to deploy it.

“What have you really saved if you replace a person with a complex vision system that still requires a person to spot-check its results?” Adolf says.

Instead, look for the biggest opportunities as these have a good chance of providing the most benefit. Then quantify what is trying to be done to improve in terms of a measurable result such as downtime or yield.

For example, 3M implemented AI to a process with a baseline yield of 88%. Over the next few months, the company saw a 7% increase in yield — as well as reduced variability in the line.

But AI doesn’t just save money, and it can be difficult to quantify all the intangible benefits. Sometimes value comes in the form of cost avoidance, improved safety or environmental sustainability. While difficult to quantify, these indirect benefits should be acknowledged.

One of the key factors to consider when determining value is scalability. AI requires research, development, testing, deployment, maintenance and periodic retraining. First, deployment costs are usually more expensive because of the learning curve involved, according to Bernd Raithel, director of product management at Siemens, and it can be difficult to recoup these costs when AI is deployed in only a single location.

“If you have only one machine and one operator,” Raithel says, “AI may not provide the benefit you are looking for.” This is one reason why many small manufacturers don’t find AI viable for their applications. However, for companies that can apply the technology across 50 production lines, AI is a completely different value proposition.

3. Consult the experts

Many people mistakenly think data scientists are going to be the most important people working on an AI system. However, the need for a data scientist has been removed from many applications.

Today’s AI model development tools are highly automated. Companies don’t need to understand the math or neural networks to build a powerful AI system. That’s important to realize. AI is changing fast, and the tools are changing just as quickly.

It’s experts who are the people who truly understand the problem.

“You don’t want a small group defining what everyone else will do. You need to make sure the people who will be using or relying on the AI system day to day are involved from the beginning,” says Siemens’ Raithel.

Integration with domain experts is critical. They are the ones who refine the definition of the problem to solve by determining the difference between a class A product, a class B product and a defect.

“Your experts are the ones who can define what a stable process looks like,” Blum adds.

4. Gather data

Getting the right data in the right place is the foundation of AI.

To build an AI model, a company needs data and context. There is no hard rule for the volume of data required, but it needs to sufficiently represent the operating conditions and capture the uncontrolled sources of variability in the process i.e., temperature, humidity, raw materials, crews, lighting conditions, maintenance, etc. Otherwise, the AI model may drop in accuracy when a variance occurs.

Collect data that is available, even if they don’t yet know how they’ll use it. Consider an AI model that predicts when a production line will stop. If a company already knew which parameters to watch, they would be able to make the prediction manually. But it’s not always clear which parameters are the important ones.

This said, when introducing AI, begin with simple parameters or data that the company already has access to. Siemens’ Raithel points out that they will also have to consider what data is needed to monitor the AI. “You need a way to verify that the AI is working and delivering as you expect it to.”

It’s not always easy to get the data needed. If they have a low defect rate, it can be challenging to collect data that reflect these defects. A manufacturer can also turn to virtual environments for training data if needed. Says Raithel, “With a virtual environmental, you can ‘take’ photorealistic photos and basically create some of the training data you need.”

5. Consider bringing in an AI partner – But not as a first step

There are many, many ways to deploy AI in automation systems – and manufacturers need to consider processing power, connectivity, architecture and overall infrastructure. There’s a lot to figure out with both hardware and software.

“The problem you are trying to solve defines the infrastructure you need,” says Siemens’ Raithel. For example, AI processing can be centralized or implemented at the edge. Manufacturers should only decide what hardware or software they need once they understand the problem. They must understand the solutions available to be able to determine the best way to solve a particular problem.

If manufacturers don’t have experience with AI, it can make sense to bring a partner on board to help understand the options, the immediate payoffs and the long-term benefits to a company.

Adolf does offer this caveat: Don’t start a project by picking a partner.

“Many companies jump to this step immediately. But how can you find the best technology solution without first understanding your problem and what you’re trying to improve?”

6. Start small – “Keep It simple”

Think about the overall problem to solve. Then pick one area to start in. It shouldn’t be the hardest challenge. Start small and pick a problem where there is guaranteed success. Give a chance to gain expertise. Just make sure it’s a problem that can scale.

“Keep it simple,” Raithel says. “Many companies think of starting with a vision-based system. Vision is sexy and often has an excellent ROI. But vision includes lighting, camera selection, angle, reflection, time-of-day, and many, many other factors. It’s complex. Quality defect detection is nothing like telling the difference between a cat or dog.”

3M’s Adolf suggests looking at where the manufacturer can add quickly add value. If they have a yield loss, identify the source of the yield loss and determine what they can do about it. For example, they might be able to improve efficiency by replacing an out-of-date system with a simple AI-based system. Similarly, they could focus on identifying potentially defective parts just before an expensive value-added process.

Another trap is thinking there are proven algorithms that can be just downloaded off the Internet.

“You probably have a part and problem no one has ever seen before,” Raithel says. “Even something as simple as confirming if two connectors are tightly seated is not simple. More than likely, you’re going to have to start from scratch.”

7. Prove AI out in stages

One way to reduce risk is to prove out an AI solution in stages.

For example, one of 3M’s pilot AI applications for vision was to improve an optical film process. The existing system was not performing well, with a huge baseline overmark of 15% — meaning that 15% of parts that should have passed were failed by the inspection system.

“We introduced AI as a software augmentation to the existing equipment and it dropped our baseline overmark to less than 2%,” Adolf says. “Similarly, undermark performance went from 5% to under 0.5%.”

When 3M implemented AI on top of an optical film process, they started by deploying it in parallel to the existing system to verify results.

“We ran the systems with the same parts to evaluate the effectiveness of the AI augmentation until we had confidence in the AI system,” Adolf says.

8. Maintenance and updates — How to care for and feeding an AI system

Updates are an essential part of nearly every AI-based system. When environmental variances (such as lighting) or raw materials (like a component is now a different shade of yellow) change, this can negatively impact performance. If the setup changes, such as a camera is knocked and the angle shifts, this can shift all the current data as well.

Equipment also changes slowly as it ages, a process called drift. For example, a heater might have to work 10% harder to keep up over time.

Systems may need to be updated to accommodate these changes. However, even systems that are robust to change will need to be able to be updated. In particular, as new data is gathered, AI models can be refined, increasing their efficiency, accuracy and confidence in their results.

At Siemens, Raithel works with AI-systems that extend across the factory floor.

“If you’ve got just one machine, then keeping up to date is easy. But production lines tend to have many machines. You need a centralized way to update at scale.” A centralized approach has the added benefit of simplifying operations and enabling monitoring. When the system is easy to update, it’s straightforward to keep performance at optimal levels.

And don’t forget the operators and engineers, Raithel warns. Depending on how the system changes, they need to be kept in the loop — and this takes additional time and effort.

9. Measure results and learn from them

A manufacturer is now up and running. And part of the value of all the historical data is to have a baseline to measure success. Being able to show ROI is important to getting support for the next project.

This said, sometimes the investment is about learning a new technology. “Your first deployment may be a loss,” 3M’s Adolf says, “But it prepares the way for massive savings later at scale.”

In other words, the first deployment might be all about gaining experience while minimizing risk.

10. Rethink what’s possible

One of the most compelling business cases for AI is when it enables a manufacture to do something a person can’t easily do.

Raithel talks of a project related to a bottleneck in their production of printed circuit boards. Before AI technology, every board had to go through an X-ray inspection. To increase production, Siemens would have had to invest another €500,000 euro to buy another X-ray.

“With predictive AI, the line is able to determine that 30% of the boards are good and do not need to be X-rayed,” says Raithel. “The result: a 30% capacity increase in our line.”

This is a great example of how AI can do much more than reduce defects. When used creatively, AI can increase production or efficiency. It also illustrates how AI can perform complex tasks reliably.

Another example of something AI can do that people generally can’t is predictive maintenance. Rather than scheduling downtime at regular intervals to look at a machine and see if anything is wearing out, AI can predict not only when something will break down but what will break down. “You get faster resolution because the technician even knows what parts to bring,” Raithel says.

A good measure of where to use AI is complexity. If there are just five parameters to consider, a person can probably handle it. When there are thousands of parameters, it’s impossible for a person to find a good set of combinations to watch for. That’s where AI can really add value.

At 3M, Adolf says 90% of the vision systems he is adding to production couldn’t have been implement in the past. The technology just couldn’t do it. For example, on a license plate manufacturing line, 3M prints the plate graphics, then adds downstream value such as retroflectivity and weatherproofing.

“Even though the vast majority of defects occur during printing — the first part of the process — it was impossible for a person to inspect a plate at that stage and tell whether it would later fail or not,” Adolf said. “With AI, we can identify defects at the source and catch them before we pay for retroflectivity and weatherproofing. Not only did we increase yield, we reduced the cost of each defect.”

Is AI worth the investment? Siemens’ Raithel thinks so.

“Today, Siemens employs in our Electronics Factory in Amberg about the same number of people as we did in 1990. However, our productivity is 17X what it was back then.”

Raithel attributes this increase to technology, and AI is going to enable the next stage of improvement. “With AI, people can focus where they can provide the most value. Rather than fighting fires, you can put out sparks.”

– The Association for Advancing Automation (A3) is a CFE Media and Technology content partner.

Original content can be found at A3.

Author Bio: Robert Huschka, VP, of education strategies, Association for Advancing Automation (A3).