CycleGAN-Based Data Augmentation for Enhanced Remaining
Useful Life Prediction under Unsupervised Domain Adaptation
Dorian Joubaud
1
, Evgeny Zotov
1
, Oguz Bektas
1
, Sylvain Kubler
1
, Yves LeTraon
1
1
Interdisciplinary Centre for Security, Reliability and Trust (SnT), University of Luxembourg, Luxembourg
dorian.joubaud@uni.lu, evgeny.zotov@uni.lu, oguz.bektas@uni.lu,
sylvain.kubler@uni.lu, yves.letraon@uni.lu
ABSTRACT
Predictive maintenance is crucial for enhancing operational
efficiency and reducing costs in Prognostics and Health Man-
agement (PHM). One of the key tasks in predictive mainte-
nance is the estimation of Remaining Useful Life (RUL) of
machinery. In practice, the data for different machines is not
always accessible in sufficient quantity or quality, therefore
the machine learning models trained on machines in one do-
main often perform poorly when applied to other domains due
to covariate shifts. As a solution, Domain Adaptation (DA)
aims to tackle domain shifts by extracting domain-invariant
features. However, traditional methods often fail to ade-
quately address the complexity and variability of real-world
data. We propose to address this challenge, using a Wasser-
stein CycleGAN with Gradient Penalty (W-CycleGAN-GP)
to learn mappings between domains and generate augmented
data in the target domain from data in the source domain. We
use our approach to generate realistic augmented data that
bridge domain gap coupled with recent work on adversarial-
based and correlation alignment-based DA models to improve
the performance of RUL prediction models in target domains
without having access to labeled data. The experimental re-
sults on the C-MAPSS dataset demonstrate a significant im-
provement in the RUL prediction score and accuracy within
the target domain.
1. INTRODUCTION
The demand for reliability in complex systems has led to
significant advances in Prognostics and Health Management
(PHM). In this context, accurately estimating the Remaining
Useful Life (RUL) of systems and their components is essen-
tial for robust predictive maintenance strategies. RUL predic-
tion enables maintenance decisions that improve operational
efficiency and reduce costs.
Dorian Joubaud et al. This is an open-access article distributed under the
terms of the Creative Commons Attribution 3.0 United States License, which
permits unrestricted use, distribution, and reproduction in any medium, pro-
vided the original author and source are credited.
Machine learning models have been widely adopted for RUL
prediction due to their ability to learn complex patterns from
historical data. Data-driven models, in particular, offer ad-
vanced frameworks for RUL estimation. They can learn from
historical data and identify patterns and features associated
with system degradation. These models can effectively han-
dle non-linear dynamic systems and provide high-accuracy
predictions. However, there are still challenges persisting
such as multidimensional data and the need for extensive
datasets in different domains. These require further develop-
ments that can provide advanced ML techniques to improve
prediction accuracy and system reliability.Also, in practice,
the data available for different machines is often not accessi-
ble in sufficient quantity or quality. Additionally, there can
be significant variations between the operating conditions or
failure modes of different machines, even when they are of
the same type. These variations lead to covariate shifts, where
the distribution of the data in the source domain (where the
model is trained) differs from that in the target domain (where
the model is applied). As a result, machine learning models
trained on data from one domain often perform poorly when
applied to other domains.
Domain Adaptation (DA) techniques have been developed
to address the challenge of covariate shifts by extracting
domain-invariant features. The goal of DA is to learn a rep-
resentation that is robust to changes in the data distribution
between the source and target domains. The adversarial train-
ing enable the model to learn representations that are indis-
tinguishable between the source and target domains. Despite
progress in DA research, traditional methods often do not ad-
equately address the complexity and variability of real-world
data, particularly in the context of RUL prediction.
To address these challenges, we propose applying DA for data
augmentation. Using a Wasserstein CycleGAN with Gradient
Penalty (W-CycleGAN-GP), our objective is to learn map-
pings between the domains and generate augmented data in
the target domain from data in the source domain. Taking ad-
vantage of recent advances in adversarial DA models in RUL,
1
