Predictive maintenance with multi-target classification models

Unexpected failures occurring in new cars during the warranty period increase the warranty costs of car manufacturers along with harming their brand reputation. A predictive maintenance strategy can reduce the amount of such costly incidents by suggesting the driver to schedule a visit to the dealer once the failure probability within certain time period exceeds a pre-defined threshold. The condition of each subsystem in a car can be monitored onboard vehicle telematics systems, which become increasingly available in modern cars. In this paper, we apply a multi-target probability estimation algorithm (M-IFN) to an integrated database of sensor measurements and warranty claims with the purpose of predicting the probability and the timing of a failure in a given subsystem. The multi-target algorithm performance is compared to a single-target probability estimation algorithm (IFN) and reliability modeling based on Weibull analysis.