‘Sensing’ plant floor disruption? How GenAI & OT can help

Global disruption in manufacturing brought on by geopolitical and climate-based events, rising technical complexities, and the persistent displacement and reduction of skilled labor, shows no signs of letting up.

And although the era of digital transformation, with its data-driven approach to operations, was viewed as a magic bullet to overcoming some of these challenges, the reality is that turning the vast amounts of information generated by everything from IoT to cyber-physical-systems into actionable insights has proven elusive.

In most plants day-to-day operations consist of a series of activities ranging from commissioning, startups, and shutdowns, to setpoint tuning, inspections, and troubleshooting. All of this is captured under the moniker of operational technology (OT).   

The volumes of information generated from OT typically fall into two distinct categories: OT data includes things like piping and instrument diagrams (P&IDs), mechanical and electrical diagrams, work orders, and more; and OT skills, take the form of diagnostic playbooks that experts follow – what to check first, how to reason about a control loop, which failure modes to rule out, etc.

Understanding and leveraging the complex interconnections between these distinct OT assets has eluded many in industry over the years. Even the advent of large language models (LLMs) has proven challenging. Although generative AI (GenAI) has transformed how we work with text and images, the largest operational gains are still on the plant floor—where pumps, valves, drives, and controllers must run safely and predictably.

One of the only ways to start cracking the code lies in leveraging both OT data and the tacit knowledge of the most experienced plant floor operators.

Building an agent of change

For Hitachi, our heritage in operating equipment and technologies dates to our founding, while our work in analytics, data, and AI goes back decades. Because of this, Hitachi remains one of the few companies that can speak to both sides of the “industrial AI” coin.

Daikin Industries Ltd., which manufactures commercial air conditioning equipment, brought us in to help connect these OT and AI worlds. Specifically, it needed an AI agent that could support equipment failure diagnostics in factories.

We kicked off a trial in April 2025 of an agent designed to understand equipment drawings converted into a knowledge graph, bind OT records, and execute a STAMP (System-Theoretic Accident Model and Processes) / CAST (Causal Analysis based on STAMP) analysis path to propose discriminating checks and corrective actions. 

The solution leveraged historical OT knowledge from Daikin sites as well as new incoming reports. When a failure in equipment occurred during operation, the agent would alert the maintenance tech about the cause and take corrective action.

The trial demonstrated that the AI agent could match or exceed the diagnostic accuracy of average maintenance engineers, even for complex or previously unseen failures. In fact, it logged response times of about 10 seconds with greater than 90% accuracy. This would not only reduce mean time to repair (MTTR) but also help standardize maintenance quality across sites and shifts. By extension, the solution would also help address the skills gap while supporting Daikin’s manufacturing expansion.

The solution, called “AI Agent for Equipment Failure Diagnostics,” worked by first converting Daikin factory equipment drawings into knowledge graphs that GenAI can read. The system then learns the OT data and feeds it into Hitachi’s unique equipment failure cause analysis processes based on STAMP.

Why RAG-only isn’t enough on the plant floor

Classic enterprise AI starts with retrieval augmented generation (RAG), which is useful for surfacing manuals, drawings, and old tickets. But in mission-critical systems, serious failures rarely repeat; after every incident, engineers add countermeasures that change the next signature.

What works in industry is retrieval augmented reasoning: the ability to bring back the right diagram slice or standard operation procedure (SOP) and then reason over the equipment’s graph and skill paths to propose discriminating checks and safe actions – with evidence.

In initial proof-of-concept tests, RAG-based systems could answer questions about known failures and those documented in manuals or past records with reasonable accuracy. However, when faced with similar but not identical failures, or entirely new failure modes, these systems struggled. For example, they might identify a “valve problem” but fail to specify which of dozens of valves is at fault or even misidentify the root cause all together. This is unacceptable in environments where downtime is costly and safety is paramount.

What’s needed is the ability to analyze the control structure, trace the flow of materials or signals, and pinpoint the most likely failure points, even for novel scenarios. This is achieved by combining the knowledge graph (representing the equipment and its interconnections) with encoded skill paths (the step-by-step diagnostic logic used by experts), like our AI agent system for Daikin.

The power of GenAI & OT

In industrial AI, the ethos is simple: move fast —and break nothing. That’s the result of melding GenAI with OT. The real win isn’t a chatbot, but steadier operations, faster troubleshooting, and fewer surprises. And when you add veteran expertise on top of it, you’re not just talking about a vision for the future – but a practical, proven approach that can be deployed today.

By capturing the “playbooks” of your best technicians and embedding them into AI agents, you can ensure that every shift, every site, and every new hire benefits from the collective knowledge of your organization. This is how manufacturing leaders will close the skills gap, reduce downtime, and build truly resilient operations for the next decade. 

For more on Hitachi Research visit: Research & Development: Hitachi

Kentaro Yoshimura, PhD, is Principal Researcher, Mobility & Automation Innovation Center, Research and Development Group, at Hitachi, Ltd. Specializing in software engineering, particularly Software Product Line Engineering (SPLE) and generative AI applications, his work involves developing methods for managing software variability.

---