In the maintenance of nuclear power plants, rapid identification of causes is crucial when failures occur. However, the constraints of limited time and available information can lead to the risk of overlooking potential causes, presenting a notable challenge. At the same time, resources such as the Nuclear Information Archives (NUCIA) and internal reports from nuclear maintenance companies contain valuable information on various past failures, and there is a growing expectation for technologies that can accurately search for this information. In response to this need, we developed a technology that constructs a knowledge graph representing causal relationships between failures based on failure reports. This technology enables comprehensive searches of failure information using the knowledge graph and LLM (Large Language Model). In the developed technology, there was a problem where missing links in the knowledge graph led to a decrease in search accuracy. Therefore, we aimed to improve the search accuracy of failure causes by developing a technology that uses active learning to complete these missing links in the knowledge graph. Experimental results showed that the Recall@10 of failure cause searches improved from 0.76 to 0.95.
Odakura et al. (Wed,) studied this question.