Comparison of binary classifiers for data-driven prognosis of jet engines health

Abstract

A reliable prognosis is crucial to manage asset health and predict maintenance needs of large civil jet engines, which in turn contribute to enhanced aircraft airworthiness, longer time on wing and optimized lifecycle costs. With the accumulation of large amount of data over the last decade, one can relate the number of components serviced during a maintenance visit to the history of parameters inside and outside the engine (temperatures, pressure, shaft rotation speeds, vibration levels, etc.). While established statistical models had been developed for small samples, more recent computer-intensive statistical techniques from the field of Machine Learning (ML) can handle more complex datasets. In particular, binary classifiers constitute an attractive option to predict the probability of servicing the components of a given jet engine at the next maintenance visit. This paper demonstrates the validity of such data-driven methods on an industrial case study involving failures of thousands of compressor blades in aeronautical turbomachines. The prediction accuracy obtained with the ML techniques presents a significant improvement over the state-of-the-art. Moreover, the performance of six binary classifiers with different characteristics - logistic regression, support vector machines, classification trees, random forests, gradient boosted trees and neural networks - was compared according to four qualitative and quantitative criteria. Results show that there is no clear winner, although ensemble models based on trees (random forests and boosted trees) offer a good overall compromise while neural networks offer the best absolute performance. In the industrial world, the business objectives, the environment in which the models are deployed and the users’ skills should dictate the choice of the most adequate statistical technique.