To save labor and improve reliability of railway vehicle maintenance, we have been studying application of Artificial Intelligence (AI) technology to visual inspections of rolling stock equipment. This paper describes how we acquired appearance images for visual inspections, created training data and trained AI models. We built a practical AI development environment to find out what affects the judgment accuracy of AI models and how to improve accuracy through consistent creation and evaluation of AI models. Specifically, we worked on visual inspections of brake shoes for conventional trains and rubber tires for Maglev vehicle given importance of those parts. Our research established a procedure to develop highly precise AI models to distinguish conditions of railway vehicle underfloor equipment.
抄録全体を表示