Hitachi, Ltd.
Circuit-Diagram AI Turns Wiring Into Data
The real advance in drawing AI is not finding components, but turning connections between lines into searchable data. Hitachi, Ltd. further trained an AI using about 80,000 pairs of circuit diagrams and structural data in Mermaid format, raising the accuracy of recognizing connections from about 20% with conventional methods to about 80%. If people can remove extra connections from real diagrams, they may be able to trace the equipment connected to a pump that has stopped at a factory late at night, as well as the inspection procedures, from a terminal.
Diagrams used in factories and infrastructure are not useful simply because their component names can be read. Getting even one connection wrong can lead to mistakes in identifying the scope of a failure and the steps needed to address it. For now, skilled workers trace complex connections by eye, while people verify knowledge held in paper documents and images.
Turning the meaning of lines into data
Vision-language models (LVLMs) are AI systems that describe objects and text in images. But accurately following connections in technical diagrams where thin lines cross is a different challenge. Mermaid, which Hitachi, Ltd. used, is a diagram notation that represents equipment as “nodes” and the relationships connecting them as text. If diagrams can be converted into this structured data, AI can use not only component names but also which pieces of equipment are connected to which in searches and reasoning.
From AI that reads appearances to AI that outputs relationships
Hitachi, Ltd. combined images of circuit diagrams with data describing the same connections in Mermaid format and further trained an LVLM. The idea was to use about 80,000 pairs of circuit diagrams and structural data to make the model output connections between components as a structure, rather than merely infer a plausible answer from an image. According to the company, the accuracy of recognizing connections rose from about 20% with conventional methods to about 80%.
For more complex diagrams, the company combined an AI that views the whole diagram with multiple AI agents that examine components and connections individually. This greatly reduced missed connections, but “hallucinations”—outputting extra lines that do not actually exist—remain. The model is relatively small, and Hitachi, Ltd. says it can be operated in a local data center. Being able to use it without sending confidential diagrams to an external cloud could also be a condition for using it in the field.
Until diagrams become knowledge for the field
If this accuracy can be maintained for piping diagrams, power-distribution diagrams, and plant diagrams as well, diagrams could become maintenance knowledge that can be searched and cross-checked. At a factory late at night, a maintenance worker could select a stopped pump on a terminal, and the equipment connected to it, past inspection records, and procedures to check would be overlaid on the diagram. Hitachi, Ltd. aims to combine this information with design documents, procedure manuals, and inspection histories for use in maintenance and physical AI.
However, the results of interpreting a diagram must be checked against conditions in the field, and incorrect connections must be safely eliminated. Adapting to the diagram formats used at individual sites, and designing the scope of decisions that people should verify, remain challenges.
