Evolution of Industrial AI: From Traditional ML to Multi-Agent Systems

hitachi · Tech · 2024 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: industrial AI and physical-world agent systems

Hitachi needed industrial AI that could augment traditional repair and operations models with generative AI, service manuals, graph RAG, process knowledge, and eventually coordinated virtual and physical agents for mission-critical field work.

Constraints:

Industrial AI must be reliable in the physical world and mission-critical settings.
Historical repair data may be insufficient for new truck models or vendors.
Manuals and industrial processes must be converted into graphs or steps that domain experts can trust.
OT code and physical-world AI require fault tolerance and domain adaptation.
Industrial AI 3.0 requires multiobjective, multiscale reasoning and complex systems-of-systems modeling.

Architecture

Hitachi's industrial AI path starts with traditional repair-classification models using historical truck data, augments them with generative AI that reads service manuals, extracts fault trees into graph structures, grounds answers in manuals, and points toward Industrial AI 3.0 where virtual agents and physical agents support field workers across domains.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  historical-repair-model["Historical repair model"] --> service-manual-parser["Service manual parser"]
  service-manual-parser --> fault-tree-graph["Fault tree graph"]
  fault-tree-graph --> graph-rag-answerer["Graph RAG answerer"]
  graph-rag-answerer --> domain-expert-verification["Domain expert verification"]
  domain-expert-verification --> virtual-field-agent["Virtual field agent"]
  virtual-field-agent --> physical-automation-agent["Physical automation agent"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Historical repair model`	Classifies repair recommendations from driver complaints, fault codes, and past repairs.	`Service manual parser`
`Service manual parser`	Uses generative AI to understand manuals, text, figures, and missing historical cases.	`Fault tree graph`
`Fault tree graph`	Represents faults, causes, components, subcomponents, and actions extracted from manuals.	`Graph RAG answerer`
`Graph RAG answerer`	Queries fault trees and grounds answers back to source manuals.	`Domain expert verification`
`Domain expert verification`	Checks grounded answers and validates process graphs.	`Virtual field agent`
`Virtual field agent`	Provides frontline workers with information at the right time.	`Physical automation agent`
`Physical automation agent`	Robotic or automation support that can fetch tools or interact with equipment.	—

Data flow

A truck or industrial asset produces complaints, fault codes, sensor/event data, and manuals. Historical repair models suggest known fixes; generative AI extracts fault trees from manuals into graphs; graph RAG answers new repair questions with grounding; process knowledge can be converted from human descriptions into graphs; and future virtual/physical agents assist frontline workers.

“when the truck came in what was the driver complaint what was the fault codes that you could download from the truck and what was the correct repair”transcript, stated

“you can generate this store it and then you can some sort of a graph rag over this to give a new query and get the right answer”transcript, stated

Why this architecture

Hitachi combines traditional ML with generative AI because historical data solves known repair patterns while manuals and process knowledge cover new models, vendors, and procedural gaps. The graph representation gives domain experts grounded evidence before AI is operationalized in the physical world.

“sometimes the right answer is not in the historical repair data if it's a new kind of truck for example”transcript, stated

“it also gives them enough confidence in machine learning to be able to operationalize the question”transcript, stated

Models used: Traditional classification ensembles · Generative AI · Graph RAG · Fault tree extraction · Small language models · Virtual agents · Physical agents

Scale

The Penske repair recommendation solution had operated for about 5-7 years.
“the solution has been operational for Last 5 Years maybe 7 6 years”transcript, stated
The system was used at 700 shops.
“it's been used at 700 shops half more than half a million trucks”transcript, stated
More than half a million trucks benefited from the repair recommendation system.
“half more than half a million trucks uh benefit from this”transcript, stated
The talk frames the agentic industrial AI vision as a next-few-years trajectory.
“the world I think that we are approaching towards maybe in the next few years uh is the sort of agentic architectures”transcript, stated

Alternatives considered (roads not taken)

Historical repair data only.

Rejected because: New truck models or vendors may lack enough historical repair examples.

“if it's a new kind of truck for example it's a new make model here they're onboarding a different vendor so then you might not have enough historical data”transcript, stated

Have humans read the whole manual.

Rejected because: Fault trees and graph RAG let users query structured manual knowledge directly.

“you can either read the whole manual but what the Savvy organizations do is they build something called a fault uh fault model”transcript, stated

Build very large language models for industrial domains.

Rejected because: Hitachi did not know how useful they would be, and small domain models may matter more for industrial settings.

“maybe the for industrial domain for particular domains the small language models are what is what is important”transcript, stated

Patterns

`rag-retrieval` (graph RAG manuals)

Hitachi extracts manual fault trees and queries them with graph RAG for grounded repair answers.

“it gives you two things it grounds the answer meaning it says look the answer based on the faultry is here”transcript, stated

`human-in-the-loop-gate` (domain expert confidence)

Domain experts can check where the answer comes from in the manual before operationalizing it.

“it tells you where in the manual you can fetch you can check as a human where the answer is”transcript, stated

`multi-agent-orchestration` (industrial AI 3.0)

The future architecture pairs virtual agents with physical agents to support frontline workers.

“you will have not only virtual agents which we think about when we think about chat GPT but also physical agents”transcript, stated

`fine-tuning-adaptation` (industrial OT code)

Hitachi expects domain-specific fine-tuning or pretraining when generating operational technology code.

“you might need to do your own um uh fine-tuning or additional pre-training of your language model to be able to generate OT code”transcript, stated

Tradeoffs

Augment, not replace, traditional industrial ML.

Chose Historical repair models plus generative AI manuals over A purely generative repair system.

Rationale: Historical models already worked for known patterns, while generative AI adds coverage for manuals and new cases.

“generative AI augments a problem that already existed that was being solved through some traditional AI”transcript, stated

Represent manuals as fault-tree graphs.

Chose Graph extraction and graph RAG over Free-form reading of manuals for each query.

Rationale: Fault-tree graphs provide structured causes, actions, and grounding that experts can inspect.

“once you have uploaded the manual you are able to extract the a fault tree it's just a graph like structure”transcript, stated

Center humans in industrial AI change.

Chose Human process knowledge plus AI over Fully autonomous process replacement.

Rationale: Humans know industrial processes and industries; AI is useful when it taps that knowledge.

“humans will drive the change uh it is the humans who need who know the processes who know their Industries”transcript, stated

Gotchas (what bit them)

Historical data misses new equipment cases.

Root cause: New truck makes, models, or vendors may not appear in prior repairs.

Resolution: Use service manuals and generative AI to augment traditional repair models.

“so at that time you have to revert back to manuals”transcript, stated

Physical-world AI needs stronger guarantees than web content generation.

Root cause: OT code affects physical systems and therefore needs fault tolerance.

Resolution: Use domain adaptation and fault-tolerant production checks for OT code.

“especially if the OT code is being used in the physical world then it needs to be fall tolerant as well”transcript, stated

Time series abstractions can lose time itself.

Root cause: LLM-style sequence treatment may ignore temporal and causal structure in sensor readings.

Resolution: Treat temporality and causality as open industrial AI research problems.

“moment you ignore time you're losing a lot of very valuable information”transcript, stated

Evaluation

Hitachi evaluates industrial AI through operational deployment history, domain expert confidence in grounded manual answers, reliability and fault-tolerance requirements, and whether systems perform valuable mission-critical tasks at low cost.

Metrics: Operational years · Shops deployed · Trucks covered · Grounded manual answer traceability · Domain expert confidence · Reliability · Cost effectiveness · Mission-critical task value

“AI to be reliable cost effective and perform valuable tasks”transcript, stated

“from a domain expert it also gives them enough confidence in machine learning”transcript, stated

Outcomes

The Penske repair recommendation system was a long-running success across hundreds of shops.
Generative AI could augment repair recommendations by extracting fault trees from manuals.
Graph RAG provided grounded answers and manual references for domain experts.
The talk maps a path from traditional/generative industrial AI to virtual and physical multi-agent systems.

The repair system is described as operating for roughly 5-7 years, used in 700 shops, and benefiting more than half a million trucks.

“we solved it sufficiently enough that it was a huge success for pens SK”transcript, stated

“there'll be agents they will be interacting with each other and doing stuff”transcript, stated