Cross-Working-Condition Remaining Useful Life Prediction of Aircraft Engines Via Domain-Invariant Features and Domain-Related Features

Abstract

Cross-working-condition remaining useful life (RUL) prediction of aircraft engines is particularly challenging when data in the target domain is unlabeled. Previous studies that have addressed this unsupervised domain adaption (UDA) problem have shown limitations in that they only use domain-invariant (working-condition-invariant) features for cross-domain predictions and ignore the potential contribution of domain-related (working-condition-related) features. To this end, we propose a UDA framework based on both domain-invariant features and domain-related features. In the first stage, considering that extracted domain-invariant features are subject to the contamination of noise or domain-related features, we construct two separate feature generators, i.e., domain-invariant feature generator and domain-related feature generator. Besides, we introduce the orthogonality loss to reduce interference between the feature spaces of two feature generators. In the second stage, a domain discriminator and a domain classifier further process the generated domain-invariant features and domain-related features separately to allow the exploration of the potential contribution of domain-related features for inferring RULs under scenarios of working-condition shift. Then, the cross-domain RUL representations are learned by a designed RUL predictor through domain-invariant features and domain-related features. Experiments performed on 12 cross-working-condition RUL predictions of aircraft engines validate the superiority of the proposed method, reducing the root mean square error (RMSE) by 9.3%.

Keywords: remaining useful life prediction, domain adaption, prognostics and health management, working-condition shift, aircraft engines