In this study, we developed a technology to predict failure phenomena using similarities between parts and between phenomena, with the aim of supporting the utilization of failure information during the design phase. The proposed method applies the concept of collaborative filtering to estimate scores for the blank cells in a parts-phenomena matrix, which organizes the relationships between parts and phenomena. Validation using failure data from the nuclear domain confirmed that the method can achieve an average precision and recall of over 0.6. This suggests that potential failures can be identified even from limited failure information.
SHIMIZU et al. (Wed,) studied this question.