Field Reliability Estimation of Agricultural Tractors Based on Warranty Data

HighlightsA LightGBM regression model for predicting tractor usage rates was established based on warranty data and considering agricultural tractors’ usage context (region and season) and was then interpreted using SHAP.The field reliability of tractors was estimated based on the usage of failed and unfailed tractors, after unfailed tractors’ usage was imputed using the LightGBM regression model.The proposed methodology was validated by predicting warranty claims using estimated reliability parameters.The proposed methodology was demonstrated using warranty data from a tractor manufacturing company in China.Abstract. Warranty data provide a valuable source of information for assessing the reliability of products in operation (called the field reliability). However, warranty data consist of failure information only. The unavailability of usage data for unfailed products makes it difficult to estimate the reliability of durable products such as agricultural tractors, for which usage is a greater concern than age for reliability analysis. Several studies have proposed methods to address this problem, but they did not include information on the usage context. This study proposes a methodology to estimate the field reliability of agricultural tractors from warranty data considering the tractors’ usage context. First, by taking features representing tractors’ usage context as the input, a usage rate regression model was established using a light gradient boosting machine (LightGBM). The usage of unfailed tractors was then generated. Finally, parametric estimates of the tractors’ reliability were determined based on the usage of failed and unfailed tractors. By interpreting the LightGBM model using SHapley Additive exPlanations (SHAP), it was found that tractors that were used more days in October and April had higher predicted usage rates. To validate the effectiveness of the proposed methodology, the estimated reliability parameters were used to predict the warranty claims of six types of tractors. The results showed that the proposed methodology performed the best in four cases and close to the best in two other cases when compared with two other baseline methods. The proposed methodology was demonstrated using warranty data from an agricultural tractor manufacturing company in China and can be applied to improve understanding of tractor reliability. Keywords: Field reliability, LightGBM, SHAP, Usage context, Warranty data.