Additional training improves the accuracy of connection-state recognition by around 220%, raising the efficiency of maintenance and design work and supporting knowledge transfer
Hitachi has developed training technology that enables generative AI to achieve highly accurate comprehension of electronic circuit diagrams, piping and tubing diagrams, wiring diagrams, and other drawings in extensive use in infrastructure and at industrial worksites. Employing the original technique of using drawing images paired with connection-descriptive text for additional training of generative AI (Figure 1), this technology demonstrated success in raising the accuracy of connection-state recognition of electronic circuit diagrams by around 220% over conventional approaches. It can therefore accurately recognize complex connection relationships between devices, which generative AI had struggled with, facilitating digitalization of field-specific and site-specific knowledge from the huge volume of drawings that can be found at worksites. Applying AI to the extracted knowledge will enable quick referencing of past design cases, trouble response history, and other such information. The technology therefore has strong potential for use at various worksites in transferring the know-how of skilled engineering personnel and enhancing the efficiency of maintenance and design work.
Collaborating with customers and partners to further advance this technology, Hitachi will aim for social implementation of AI solutions in support of DX (Digital Transformation) promotion in the infrastructure and industrial fields.
Figure 1: Overview of the generative AI training technique designed specifically for the comprehension of drawings at industrial worksites
Background and issue addressed
At worksites in the infrastructure and industrial fields, on-site knowledge is stored in the form of electronic circuit diagrams and various other drawings showing wiring or piping and tubing, for example. However, using these types of drawings to train AI can be challenging. Because of the process of converting from paper to digital format, image deterioration, and other factors, it is often difficult for AI to recognize the existence of lines, the direction of arrows, and especially the state of connections between devices. This hinders attempts to transfer know-how from skilled engineering personnel or use past drawings for enhancing the efficiency of design and maintenance work.
Features of the technology developed to solve the issue
To address this issue, Hitachi developed a generative AI training technique specifically for the comprehension of drawings used at industrial worksites. With this technique, “correct answer” text is created and paired with an image of a drawing. The text teaches the generative AI that the device connection states described are what the drawings indicate. These image-text pairs are then used for additional training of the AI.*1 This original technique adapts flexibly to various industrial drawings, including electronic circuit diagrams, wiring diagrams, and piping and tubing diagrams, and the AI is capable of recognizing complex connection relationships among elements at a high accuracy level.
Applying the newly developed technique to reading of electronic circuit diagrams, after additional training, the AI was able to more accurately recognize the existence of lines, the direction of arrows, and other aspects that conventional approaches often struggle with. Testing confirmed an improvement of approximately 220% in recognition accuracy of connection states*2 (Figure 2).
Figure 2: Connection state recognition accuracy before and after additional training
Looking ahead
Hitachi plans to further enhance this technique as one of the technologies for realizing Lumada 3.0 and to accelerate DX promotion in the infrastructure and industrial fields through the digitalization of field knowledge and use of AI. The aim is to develop solutions that contribute to solving problems at worksites. One such example is applying the next-generation AI agent “Frontline Coordinator - Naivy”*3 toward improving the work efficiency and well-being of frontline workers and helping in areas such as transferring the know-how of skilled engineering personnel.
*1 N. Terashita et al., “Can Visual Encoder Learn to See Arrows?”,VisCon 2025 Poster, 2025.
https://openreview.net/forum?id=eLZURVinXI
*2 As measured by the F-values (harmonic mean of precision and reproducibility) indicating how closely the “connected part pairs” output by the generative AI match the correct answers. A value of 0.0 means no matching, and 1.0 indicates full matching.
*3 Hitachi develops "Frontline Coordinator – Naivy" as a next-generation AI agent that helps alleviate the psychological burden on frontline workers and enhance work efficiency: July 3, 2025
For more information, use the inquiry form below to contact the Research & Development Group, Hitachi, Ltd. Please make sure to include the title of the article.
https://www8.hitachi.co.jp/inquiry/hitachi-ltd/hqrd/news/en/form.jsp
