Transfer learning applied to bivariate forecasting on product warranty data

Abstract

The reliability and resource management of products for warranty is important. Furthermore, the number of failures of a
product over time of use and level of expenditure can assume different distributions. Approaches with parametric models
bring good results when there is a normal distribution, and the application of Deep Learning (DL) is very promising. We
show a new methodology for the application of DL models with transfer learning to bivariate forecasts of repair rates in
products that are under warranty. The solution was applied to data from an American company, recorded from 2015 to
2022, of 12 different types of parts from 69 different types of cars. An evaluation of the absolute error of the forecasts was
performed for each combination of part, car and model year. Tests showed that the model performed well in predicting
data for 70 months in service and 70,000 miles, using data from cars with at least 15 months in service and 1,000 miles
as input. It was also concluded that the solution is robust for cases of incomplete data and distributions far from the
normal distribution.