Automobile warranty data predictive models for interpreting engineering design and process changes.

Abstract

Beginning in the late 1980's manufacturers increased the basic automobile warranty coverage region to 5 year or 50,000 mile and beyond. These larger coverage regions resulted in extensive field failure data that did not match predictions using the log-log technique of predicting R(t), repairs per thousand vehicles. The high level of prediction error inhibited the manufacturer from relying on the log-log predictive model to make inferences regarding engineering design and process changes. This thesis used a set of 1991 model luxury cars as a case study. The manufacturer provided actual option content, component design and warranty claim data that kept the research practical. Focusing warranty data analysis on one observation per vehicle sold, rather than warranty claims, provided a structure that accommodated vehicle specific explanatory variables. This structure supported fitting random variables to component lifetimes and Poisson and stochastic process models to the number of claims per vehicle. Warranty claims represent a mixture of: manufacturing defects, assembly defects, usage related failures, and appearance items. The two-dimensional vehicle life (time and mileage) and claims before coverage (predelivery) complicated predictions. This thesis developed a mixture distribution for component lifetimes that incorporated predelivery claims and distinguished between usage related failures and manufacturing or assembly defects. By projecting a vehicle into the time domain across the usage function, the relationship between time and mileage, this thesis developed a time based definition of population size. This thesis developed a non-homogenous Poisson process predictive model for the number of warranty claims using the population size as input. The technique fit the empirical hazard function of time to first claim using data available at 6 months in service to predict claims through a specified date. This thesis fit the Weibull, a Weibull and uniform mixture, and linear hazard to the luxury car warranty data to validate the prediction strategy. Subsystem specific stochastic process models developed by this method provided very accurate predictions of future warranty claims.