ChatGPT and other AI chatbots have become a regular sidekick for office workers seeking a leg up crafting business correspondence or for developers as an accelerant when writing code. Yet in the industrial space, chatbots and their large language model (LLM) technology are still far from ubiquitous, surfacing primarily in pilot projects aimed at boosting operator productivity and problem solving.
As a deep learning model, LLMs employ natural language processing to generate human- like responses to queries and tasks. They are trained on extremely large data sets and are built on neural network technology known as transformer models. In the business sector, LLMs are being rolled out to generate text and visual content, uplevel chatbots for more personalized customer service and interactions, and to empower task automation.
OpenAI CEO Sam Altman has claimed that 92% of Fortune 500 companies are using ChatGPT, but that platform and other LLMs are charging a slower path into the industrial landscape. While there is mounting enthusiasm for LLMs as an anchor for tailored systems in areas like predictive maintenance, computer vision quality inspection, production planning assistants and operator guidance, large-scale implementations are still rare and there remains an array of issues standing in the way of active deployment.
The data obstacle
Perhaps the biggest barrier to widespread LLM use in industry is the current state of industrial data. Most manufacturers are sitting on a mountain of siloed data, much of it time series, that is not properly contextualized, standardized, scalable or even available for LLM use cases and model training. Immature data management processes, the on- going OT/IT divide, and the lack of real understanding of how and where LLMs can make a mark are among the many inhibitors limiting industrial activity to one-off pilot projects and experimentation.
“A lot of initiatives are in the experimentation phase because they haven’t really laid the full foundation from the OT level up,” said Travis Cox, chief technology evangelist for Inductive Automation. “They haven’t made the cultural change [to] where data is fundamental to the business so they can go to the next level. It’s also not a technology problem, it’s more of a people, process problem now.”
The open nature of public LLMs also creates additional data security and customer privacy concerns and complications. To get the desired result, public LLMs need data—lots and lots of data typically fed to the model by the user looking for answers and action. Yet most manufacturers are extremely wary about putting their own proprietary manufacturing, product and consumer data up for consumption by public LLMs for model training in the cloud. Even those manufacturers fully onboard with generative AI’s promise seem partial to building use cases around private LLMs, but they typically lack the resources and AI expertise to navigate a project of this scale and complexity.
“Manufacturers are trying to implement LLMs in certain places today, but it’s not largescale and certainly not everywhere,” said Erik Lindhjem, vice president and general manager of the reliability solutions business at Emerson. “There’s more focus now on how to train private LLMs with priority data to use in different scenarios.”
Industrial virtual assistants dominate
As in the business sector, most of the early industrial use case examples of LLMs are focused on virtual assistants to provide guidance to plant floor operators or to give control engineers a head start writing PLC code. The ease with which GenAI can help synthesize real-time information or provide coding assistance is crucial for today’s manufacturers given current challenges attracting and retaining plant floor talent.
“LLMs can help operators handle complexity as plants try to produce more in a safe environment with less people,” said Claudio Fayad, vice president of technology for Emerson’s process systems and solutions business.
Fayad pointed to Aspen Tech’s Aspen Virtual Advisor (AVA) for DMC3 as an example of putting LLMs to work in this kind of use case. The solution augments operator knowledge with real-time insights into advanced process control hardware, presenting as a chat window where technicians can ask questions that probe constraints on particular assets or to get instant guidance on recommendations for increasing plant floor throughput, Fayad explains. Emerson now owns a 55% stake in Aspen Tech, and AVA will be offered to support Emerson process control hardware.
“We’ve been using help files in neural nets to help operators make decisions, and now LLMs are making that learning capability better,” Fayad said.
Emerson also sees potential of LLMs in plant modernization efforts, serving as a code assist tool as part of its DeltaV Revamp, a cloud-based technology that manages the transition of legacy control applications to the DeltaV distributed control system. LLMs facilitate the process of understanding the old code base and converting it to the language of the modern control platform.
“Just like LLMs can understand Spanish and convert it to English, they can be leveraged to accelerate the conversion of a Honeywell code base to a Delta V code base for rapid systems deployment,” Lindhjem explained.
Other automation providers are testing the power of LLMs for virtual assistants. Aveva’s Industrial AI Assistant, for instance, enables operators to ask questions like: What was my maximum output last month? or why is my compressor less efficient this week? and get answers and context that go far beyond what would be possible sifting through spreadsheets and documents and far faster than a complex data analysis.
Aveva’s patent-pending knowledge linking technology (considered an AI-driven knowledge graph) pulls in time-series data and documents from various, relevant data sources, automatically creating relationships across structured and unstructured data without the need for a data hierarchy or model, according to Jim Chappell, vice president and global head of AI and advanced analytics at Aveva. Data security and AI hallucinations are tackled using an AI orchestrator, which includes guardrails like intent analysis, prompt optimization and response formatter, among other capabilities, while also breaking down questions into individual sub-queries to keep responses more grounded, Chappell explained.
Unlike a traditional approach, which would require familiarity with cloud platforms, sensors, IoT, AI and different types of time-series and engineering data, an LLM-based offering lets engineers facilitate data queries, visualizations and workflows without the complexity. “You can ask questions as a subject matter expert contemporary and don’t need to know the software,” Chappell said. “You just ask questions and it gives you the information you need.”
Siemens and Beckhoff Automation are harnessing the power of LLMs to simplify and accelerate the programming of automated systems and controls. TwinCAT Chat Client leverages ChatGPT to automate code creation, code optimization, code restructuring and code documentation, enabling engineers to write higher quality code faster. It connects to the host cloud of the LLM—in this case, Microsoft Azure for ChatGPT, letting TwinCAT developers ask questions to generate HMI controls and establish links to the PLC. “The LLM is contextualized with our documentation so the code spitting out has the knowledge of best practices, our documentation, APIs,and equipment,” said Brandon Stiffler, software product manager at Beckhoff.
For its part, Siemens’ Industrial Copilot, developed using the Azure OpenAI Service in Microsoft’s Azure Cloud, is connected to Siemens Totally Integrated Automation (TIA) Portal, enabling engineering teams to quickly get help as well as generate basic visualization and structured control language (SCL) code for PLCs. Siemens Industrial Copilot explains SCL code blocks and creates machine and plant visualization in the WinCC Unified visualization platform, reducing the time and effort for PLC programming while minimizing errors.
“Given the lack of skilled labor workers, you need to be able to train quickly on how to program automation hardware, and GenAI really speeds up that process,” said Kristen Quasey, product marketing manager for industrial PCs and Industrial Copilot at Siemens.
Automation vendors are not the only avenue for LLM-enabled tools. Control system integrators such as Northwind Technical Services are taking the opportunity to offer generative AI-based virtual assistant functionality to its customers to address what Matt Lueger, executive vice president, called their biggest problem—the shortage of skilled labor.
“The 30-year veterans who own their knowledgebase are leaving the workforce just as they are putting in more sophisticated systems and more automation,” Lueger said. “They are also having a hard time hiring young technical people to maintain their on-going systems. GenAI could be the tool to help close that gap.”
Northwind developed PlantIQ, a digital expert built on an LLM and trained specifically to understand manufacturing processes through connections to relevant documentation and real-time process data. The accompanying AlarmIQ module has a specific focus on PLC and SCADA system alerts and alarms, ensuring quicker resolution of system faults through delivery of detailed process information, analysis of historical faults and documented control system service tickets. With this technology, new operators are afforded the benefits of domain expertise typically held by a seasoned process engineer, allowing them to come up to speed more quickly than with traditional training.
LLM’s future role
As LLM-enabled virtual assistants evolve, capabilities will go beyond real-time guidance and insights to having the assistant automatically execute a specific task. Take controls system design and configuration, for instance. This is typically a heavy lift for process controls engineers, but LLMs could be used to accelerate the design of HMI graphics as well as the configuration of control systems.
Aveva sees a progression where a user won’t just ask questions of the LLM-enabled virtual assistant, but rather instruct it to perform a specific task like “build me a dashboard,” whether initiated through written commands or conversational interaction.
“We are looking at delivering new types of user experiences for our software,” Chappell explained. “Instead of learning how to use different types of software, you ask the assistant to do it and it will. It provides a great starting point for the average user across different functionalities.”
Taking the time to understand where and how LLMs and GenAI can add business value is crucial to success. It’s also important to frame virtual assistant output as a starting point, not an end point, at least with the current generation of products. “Keeping a human in the loop is critical to build trust in the answers,” said Fayed. “It’s also important to bring in data in a contextualized way so AI can make sense of it. If it’s just numerical data and not relationships, it’s not useful.”