As AI evolves, manufacturing will face an old data problem
New capabilities in AI technology hold promise for manufacturers, but companies should proceed carefully until issues such as data quality are resolved.
AI has long been an integral technology in manufacturing operations, but technological advances are opening new possibilities for improving efficiency, safety and productivity.
The hype around an AI revolution in manufacturing is a myth, as the industry has used automation and AI techniques for decades, according to the recently published "Industrial AI Insights" report from Honeywell Industrial Automation, a division of Honeywell that provides automation devices, software and services primarily for process manufacturers. However, AI is evolving rapidly, and manufacturers are beginning to derive benefits from cutting-edge technologies such as generative AI.
For the report, Honeywell and Wakefield Research surveyed 1,600 global manufacturing leaders who are currently implementing AI to determine how and why they are using these capabilities.
AI projects are on the rise in manufacturing, even if the technology is not fully understood in the organization, according to the report. Almost every respondent (94%) reported that corporate leadership is "all in" on the technology, although 37% also expressed concern that their executives don't understand AI.
Nevertheless, a large majority of respondents (84%) believe their companies are AI pioneers, and 91% are finding new use cases for AI beyond the initial plans for automating tasks and processes. The respondents also reported that they are getting payback from investments in their AI implementations, including increased efficiency through automation (64%), improved cybersecurity (60%) and real-time data to improve decision making (59%).
There are also indications that these new AI projects are in the early stages, as just 17% of respondents have fully launched initial AI projects, while others are still in the scaling (43%) or prototyping (12%) phases.
Manufacturing tends to use three kinds of AI, according to Jason Urso, CTO at Honeywell Industrial Automation.
The most common form is deterministic AI, which has been available for decades and uses model-based predictive mechanisms to automate processes, Urso said.
Now, however, new forms of AI are coming that can bring new insights and benefits.
One is probabilistic AI, which takes large amounts of sensor data from operations and provides insights into probabilities of what might happen. This can be used to monitor machines and identify when they may need maintenance, he said.
Finally, there's generative AI, which allows workers to "converse" with machines to determine their health and how to resolve problems.
"That's a giant advancement, because now [the machine] is communicating with the human in an almost human way, as if there's an expert behind the scenes conveying this information when really it's your knowledge repository," Urso said.
AI can help by accelerating the time to expertise. It's not displacing people, it's accelerating their expertise by making a GenAI query into a knowledge base that asks about a problem and can suggest ways to resolve it.
Jason UrsoCTO, Honeywell Industrial Automation
One of the biggest benefits of enabling faster and better decision making is helping companies deal with the increasing shortages of workers and skills gaps in younger workers in the manufacturing sector, he said.
"AI can help by accelerating the time to expertise," Urso said. "It's not displacing people, it's accelerating their expertise by making a GenAI query into a knowledge base that asks about a problem and can suggest ways to resolve it."
There are challenges that are slowing fully deployed AI implementations in manufacturing, he said, primarily the quality of data and the ability of users to trust the responses they get from AI.
"We're collecting all the data, but can we trust it," Urso said. "You need to make sure the data set is correct, or else it will diligently provide the wrong response based on the erroneous data. So making sure that there's a thorough process for vetting the data set that's being used to provide the recommendation is a concern."
Despite risks, AI holds promise
New capabilities for AI in manufacturing are promising, but companies must be wary of the potential risks in generative AI, according to a Forrester Research report "Generative AI: What It Means for Smart Manufacturing" published in March.
Benefits include improving efficiencies and effectiveness in the organization through generative AI applications that can help customer service personnel, knowledge workers and service technicians get faster access and better answers from an organization's accumulating data, according to the report. The report’s conclusions were based on interviews with leaders in several industrial companies, including Autodesk, Dassault Systemes, PTC, Schneider Electric and Siemens.
However, generative AI tools are not yet ready for broad adoption across the manufacturing sector, according to Paul Miller, an analyst at Forrester and one of the report's authors.
"All of the excitement around GenAI risks is taking attention away from the good, useful and valuable work that manufacturers have been doing with other forms of AI for many years," Miller said. "That would not be a good outcome for anyone."
Manufacturers need to understand how generative AI differs from other AI technologies, he said. For example, an AI-based computer vision system that checks the paint quality on an automotive assembly line is deterministic or rules-based, as it produces the same output for a specific input.
Generative AI, however, is non-deterministic, which means the same input could produce different outputs, Miller said.
"That fundamental attribute of today's GenAI tools is going to affect where and how you use them in a manufacturing context," he said.
Many manufacturers aren't explicitly aware of the non-deterministic nature of generative AI, but they are aware of some of the risks systems like these can produce such as the technology's ability to hallucinate, Miller said.
"They are understandably worried about a GenAI tool giving poor or incorrect advice, which might lead to equipment damage, downtime or even injury, and they are keen to understand how techniques like retrieval-augmented generation can be used to limit the risks here," he said, referring to a process of using external data to improve a model's output.
But manufacturers are also keen to understand where the technology can deliver business value and are looking for guidance on the best ways to introduce the technology to employees and customers alongside other options, Miller said.
Rather than trying to pick an AI technology -- deterministic or generative -- and apply it to a business problem, Miller said, manufacturers should focus on the business problem that needs to be addressed.
"For example, saying 'I want to reduce the time spent hunting for product documentation in field service use cases' would be a far better starting point than 'I want a GenAI tool for field service,'" he said.
AI in manufacturing has been around for years, but the technology is maturing. If the industry wants to continue to reap the benefits of this technology, it will have to address its long-standing issue with data quality.
AI in manufacturing is changing
The AI revolution is nothing new in manufacturing, agreed Peter Zornio, CTO at Emerson, an industrial automation systems and software firm.
Manufacturing, particularly in process industries, has used digital systems to read and analyze sensor data and software to take corrective actions since the 1970s, Zornio said. This has continually evolved to the current machine learning and generative AI technologies that are coming on now.
"[Manufacturing was doing] AI before AI was cool," he said. "From the mid-1980s through today, we've been constantly applying these numerical mathematical AI techniques to provide ever better control and optimization of what's happening in the manufacturing process."
Manufacturing has always had the data, much of it coming from sensors that capture manufacturing processes as they operate, but the evolution with AI is building the models and improving the analytics, Zornio said.
"You can have a lot of raw data, but if that data is not in the right format or the right context, you can't interrelate the data from this set of applications with that set of applications," Zornio said. "You can't run the optimization and build the model across those multiple data sets, especially when you're trying to span multiple functions inside of a plant."
There are several promising use cases for generative AI in manufacturing, he said. For example, a company can use the technology to design automation systems specifically for each of its facilities.
"GenAI can configure the system automatically by feeding the system drawings of the physical equipment and how they're interconnected, and coming out with an initial configuration of the control system," Zornio said. "That's engineering work that GenAI is able to take over for us."
Generative AI can also be used as a "super assistant" for products by ingesting all the documentation, support interactions, logbooks and application notes to create a product chatbot that has vast knowledge and can respond better when you ask how the product works, he said.
"Everybody who makes a product is looking to build like a super product assistant like that," Zornio said.
Poor data quality holds back AI projects
Having enough data within factories and manufacturing operations has never been the issue; it's the ability to tap into it and use it, according to Jon Sobel, co-founder and CEO of Sight Machine, which provides a data platform for industrial operations.
Data issues have been one of the main reasons why AI projects in manufacturing have hit the wall, Sobel said.
Manufacturing produces about twice as much data as any other business sector, but until now, little has been used for productive purposes he said. One problem is that the data is generated at different times and from various sources that don't interact.
"The real challenge is most of this data has not been used because of how we build applications historically -- you build 50 apps and none of them can talk to each other," Sobel said. "So the paradigm has to change to get the data structured, know enough of what your data is and then start to query."
Once manufacturers have structured and prepared data properly, the new AI advances are the ideal technology to enable this data to be used, he said.
"For the longest time we've talked about the importance of data -- and nobody cared," Sobel said. "AI came along and it makes people think that if they want to do AI, their data has to be right. AI is a catalytic moment in the discussion about having good data."
Jim O'Donnell is a senior news writer for TechTarget Editorial who covers ERP and other enterprise applications.
