World disruption in manufacturing introduced on by geopolitical and climate-based occasions, rising technical complexities, and the persistent displacement and discount of expert labor, reveals no indicators of letting up.

And though the period of digital transformation, with its data-driven strategy to operations, was considered as a magic bullet to overcoming a few of these challenges, the fact is that turning the huge quantities of data generated by every part from IoT to cyber-physical-systems into actionable insights has confirmed elusive.

In most crops day-to-day operations encompass a collection of actions starting from commissioning, startups, and shutdowns, to setpoint tuning, inspections, and troubleshooting. All of that is captured below the moniker of operational know-how (OT).   

The volumes of data generated from OT usually fall into two distinct classes: OT knowledge contains issues like piping and instrument diagrams (P&IDs), mechanical and electrical diagrams, work orders, and extra; and OT expertise, take the type of diagnostic playbooks that specialists observe – what to examine first, cause a couple of management loop, which failure modes to rule out, and many others.

Understanding and leveraging the complicated interconnections between these distinct OT property has eluded many in business over time. Even the arrival of enormous language fashions (LLMs) has confirmed difficult. Though generative AI (GenAI) has remodeled how we work with textual content and pictures, the biggest operational positive aspects are nonetheless on the plant flooring—the place pumps, valves, drives, and controllers should run safely and predictably.

One of many solely methods to start out cracking the code lies in leveraging each OT knowledge and the tacit information of essentially the most skilled plant flooring operators.

Constructing an agent of change

For Hitachi, our heritage in working tools and applied sciences dates to our founding, whereas our work in analytics, knowledge, and AI goes again many years. Due to this, Hitachi stays one of many few firms that may converse to each side of the “industrial AI” coin.

Daikin Industries Ltd., which manufactures industrial air con tools, introduced us in to assist join these OT and AI worlds. Particularly, it wanted an AI agent that would help tools failure diagnostics in factories.

We kicked off a trial in April 2025 of an agent designed to grasp tools drawings transformed right into a information graph, bind OT information, and execute a STAMP (System-Theoretic Accident Mannequin and Processes) / CAST (Causal Evaluation primarily based on STAMP) evaluation path to suggest discriminating checks and corrective actions. 

The answer leveraged historic OT information from Daikin websites in addition to new incoming experiences. When a failure in tools occurred throughout operation, the agent would alert the upkeep tech concerning the trigger and take corrective motion.

The trial demonstrated that the AI agent may match or exceed the diagnostic accuracy of common upkeep engineers, even for complicated or beforehand unseen failures. In reality, it logged response occasions of about 10 seconds with higher than 90% accuracy. This is able to not solely scale back imply time to restore (MTTR) but in addition assist standardize upkeep high quality throughout websites and shifts. By extension, the answer would additionally assist tackle the talents hole whereas supporting Daikin’s manufacturing growth.

The answer, referred to as “AI Agent for Gear Failure Diagnostics,” labored by first changing Daikin manufacturing unit tools drawings into information graphs that GenAI can learn. The system then learns the OT knowledge and feeds it into Hitachi’s distinctive tools failure trigger evaluation processes primarily based on STAMP.

Why RAG-only isn’t sufficient on the plant flooring

Traditional enterprise AI begins with retrieval augmented technology (RAG), which is beneficial for surfacing manuals, drawings, and previous tickets. However in mission-critical techniques, severe failures hardly ever repeat; after each incident, engineers add countermeasures that change the subsequent signature.

What works in business is retrieval augmented reasoning: the power to carry again the proper diagram slice or customary operation process (SOP) after which cause over the tools’s graph and ability paths to suggest discriminating checks and protected actions – with proof.

In preliminary proof-of-concept assessments, RAG-based techniques may reply questions on recognized failures and people documented in manuals or previous information with cheap accuracy. Nevertheless, when confronted with related however not an identical failures, or fully new failure modes, these techniques struggled. For instance, they could determine a “valve drawback” however fail to specify which of dozens of valves is at fault and even misidentify the foundation trigger all collectively. That is unacceptable in environments the place downtime is dear and security is paramount.

What’s wanted is the power to research the management construction, hint the stream of supplies or indicators, and pinpoint the almost certainly failure factors, even for novel eventualities. That is achieved by combining the information graph (representing the tools and its interconnections) with encoded ability paths (the step-by-step diagnostic logic utilized by specialists), like our AI agent system for Daikin.

The facility of GenAI & OT

In industrial AI, the ethos is straightforward: transfer quick —and break nothing. That’s the results of melding GenAI with OT. The true win isn’t a chatbot, however steadier operations, quicker troubleshooting, and fewer surprises. And whenever you add veteran experience on prime of it, you’re not simply speaking a couple of imaginative and prescient for the long run – however a sensible, confirmed strategy that may be deployed at present.

By capturing the “playbooks” of your greatest technicians and embedding them into AI brokers, you’ll be able to be sure that each shift, each web site, and each new rent advantages from the collective information of your group. That is how manufacturing leaders will shut the talents hole, scale back downtime, and construct actually resilient operations for the subsequent decade. 

For extra on Hitachi Analysis go to: Analysis & Improvement: Hitachi

Kentaro Yoshimura, PhD, is Principal Researcher, Mobility & Automation Innovation Heart, Analysis and Improvement Group, at Hitachi, Ltd. Specializing in software program engineering, significantly Software program Product Line Engineering (SPLE) and generative AI purposes, his work entails creating strategies for managing software program variability.