Hitachi, Ltd. has developed an AI agent that estimates the causes of manufacturing equipment failures, including unknown ones. In a trial operation for practical use that began in April 2025 as part of a joint demonstration with Daikin Industries, the accuracy of cause estimation improved from 67% with general-purpose AI to over 90%, and the agent generates a response naming the specific suspected part in approximately 10 seconds from the inquiry.
Experienced maintenance staff at Daikin Industries reportedly said the agent is "equivalent to or better than a typical maintenance technician," "clearly superior to a novice," and "at a level that can be used in the field."
Enabling even inexperienced maintenance personnel to perform advanced work is a long-standing challenge at manufacturing sites, made more urgent by the aging and retirement of skilled technicians. Training AI on maintenance records has had limited results because the recorded information is unclear and non-uniform and much knowledge is tacit. Retrieval-augmented generation (RAG) can handle known events in past records and manuals but often fails on events that differ slightly from past cases, and factory failures are typically unknown or new.
Hitachi's approach reproduces the thought process of skilled technicians. It converts design drawings into a knowledge graph format the AI can understand, extracting part connections and hierarchical structures from piping and instrumentation diagrams (P&IDs) and electrical drawings into a knowledge graph or Bill of Materials (BOM) format. Currently, humans interpret the drawing information and provide the AI with structured data on part configurations and connections; the goal for the future is complete automation using AI capable of interpreting drawings. A large language model was trained on the system safety analysis method STAMP/CAST so it can perform systematic root cause analysis.
A paradigm called Retrieval Augmented Reasoning (RAR) searches for appropriate drawings and standard operating procedures (SOPs) while combining the equipment structure data with the analytical skills of an expert, so the system can propose specific inspections and countermeasures supported by evidence. The agent runs on a cloud platform managed by Hitachi.
At the end of 2025, joint verification with Mitsubishi Chemical Corporation also began. The system is part of the digital service HMAX Industry, a suite of solutions that embodies Lumada 3.0.