Failure Cause Analysis of Automobile Parts Using Machine Learning

Abstract

We propose a method for analyzing the cause of automobile parts failure from the vehicle data and maintenance data of a car maintenance contractor using machine learning. There are two types of automobile parts: ones that need to be replaced periodically, such as rubber wiper blades, and ones that do not usually need to be replaced, such as the engine. Some of the ones not expected to repace are expensive, and they cause rarely. If we know the user who will break the expensive parts, the contract company can up the contract fee. The car users will not interrupt their business by breaking their cars if we predict the failures and repace them preventively. We learn a model that predicts whether a component replaces in the next maintenance or not. We use gradient boosting and analyze the cause of the failures by explaining the learned model using SHAP. Because the replacements of the expensive components occur rarely, the classification classes are imbalanced. We also propose a method for oversampling imbalanced data by using the characteristics of gradient boosting. We then confirm the effectiveness using the actual data.