A research group at Advanced Intelligent Systems Laboratories, Toshiba Corporate Research & Development Center announced that it has developed an infrastructure‐document‐understanding AI for increasing maintenance and inspection efficiency. It does so by efficiently and accurately recognizing texts regarding different specialties (specialty data), including inspection results or records accumulated on problems with equipment or facilities, such as factories and plants. The group stated that it aims for proactive high‐quality maintenance by applying the developed AI to infrastructure maintenance sites and other areas. The AI's efficiency will be verified and then it will be put into practical use by 2024. The achievement was announced during the 29th yearly session of the Association for Natural Language Processing (NLP2023), which was held in Okinawa in March.
Two problems that have been recognized in Japan are that facilities and equipment at factories and plants across the country are aging, and the number of maintenance personnel is insufficient. To maintain and inspect obsolete facilities with high accuracy, understanding advanced knowledge and collecting information on previous problems and their countermeasures is necessary. Unfortunately, conventional AI has difficulty recognizing such specialized data.
Therefore, the research group tried developing their AI to efficiently retrieve the knowledge of high‐skilled workers from accumulated specialized data to enhancing maintenance and inspection. Toshiba has been managing the production and maintenance of important facilities and retains an enormous amount of specialized data and this was used in the development of their AI.
By applying this AI, it will be possible to use specialized data, which conventional technology has difficulty recognizing. For instance, when a novice maintenance worker asks the AI a question regarding a specific problem, the worker can obtain some effective countermeasures based on previous data. By accelerating the maintenance of a facility after an incident, workers can shorten the facility downtime. Furthermore, by understanding the frequency trends of various problems, it is expected that the workers can consider replacing certain parts and/or implement preventive maintenance.
As a key aspect, the developed AI can learn both general and specialized terms from a few resources. Conventional general‐purpose language models can learn the meaning of a language depending on the context from a large general text database; however, when applying this to some specialized data, preparing a large number of texts specific to a particular area that have generally not been available in Japanese is necessary. In addition, a large number of calculative resources is frequently required.
By contrast, the AI developed by the group efficiently learns general terms from a general‐purpose language model acquired in a typical manner. At the same time, the AI learns the specialized terms of a particular area from a small number of specialized data in a separate curriculum. Their developed AI can learn specialized terms without forgetting general terms; thus, it can learn using a small data source.
In order to verify performance, the group conducted a language‐analysis test to find expressions concerning issues described in maintenance and inspection records of some electric‐power facilities (information extraction task). The number of calculations for a student model generated using their developed AI was half that of the general‐purpose language model (conventional method), which learns from scratch. The number of texts used in learning was determined to be one percent of that of the conventional method.
As a result, it was confirmed that the AI was able to extract the "phenomenon" describing the status of equipment in trouble and the "measures" taken by maintenance personnel to repair the equipment from maintenance and inspection records with an accuracy rate of approximately 89%, which is higher than that of the conventional method (approximately 86%). It was also verified that the learning time was 5 hours meaning it is possible to reduce the learning time by approximately 97% compared to the conventional method.
The research group stated that they plan to continue their research and development to enhance the maintenance system.
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.