Performance Comparison of String Similarity Algorithms for Component Classification in a Failure Databook

Purpose: A failure databook provides failure rate information to assess system reliability. Such databooks have been developed and adopted in industries across several countries. Recently, a BR/Korean-style failure rate databook, based on the field failure data of weapon systems, has been developed and distributed to defense industries. A critical aspect of a databook's development is the categorization of items, which is a time-intensive process. This study aims to support the classification of new items by evaluating the accuracy of five string similarity algorithms.BR/Methods: The similarity algorithms considered in this study include the Jaro‒Winkler similarity, longest common subsequence (LCS) similarity, N-gram similarity, Levenshtein similarity, and Hamming similarity. To assess the accuracy of these algorithms, verification items were extracted through stratified sampling from the Korean failure rate databook, with the remaining items designated as reference data. Simulations were conducted by treating the verification items as new entries and evaluating how accurately each algorithm recommended the optimal major classification for these items.BR/Results: The simulation results revealed that the Jaro‒Winkler similarity algorithm achieves the highest overall recommendation accuracy, followed by the LCS and N-gram similarity algorithms. Furthermore, as the number of recommended categories increases, the difference in accuracy between the Jaro‒Winkler and LCS algorithms becomes negligible.BR/Conclusion: The findings of this study can be applied to automate item classification during the revision of failure rate databooks. Additionally, collecting supplementary data beyond item names could enable the development of advanced item classification recommendation techniques using various big data analysis methods.