A Design Science Approach Comparing Ensemble Learning and
Artificial Neural Networks for Uncertainty-Aware Helicopter
Turbine Engines Health Monitoring
Victor Henrique Alves Ribeiro
1
, Gilberto Reynoso-Meza
2
1,2
Pontif
´
ıcia Universidade Cat
´
olica do Paran
´
a, Programa de P
´
os-Graduac¸
˜
ao em Engenharia de Produc¸
˜
ao e Sistemas,
Curitiba, Paran
´
a, 80.215-901, Brazil
victor.hribeiro@pucpr.br
g.reynosomeza@pucpr.br
ABSTRACT
This work presents the development of an uncertainty-aware
health monitoring system for helicopter turbine engines, fo-
cusing on improving operational availability and reducing
maintenance costs We address the critical issue of uncertainty
quantification in data-driven fault detection and prognostics,
essential for increasing system reliability. The project follows
an iterative development cycle, incorporating multiple tech-
niques for data processing, such as polynomial feature gener-
ation and data cleansing, and model development, including
ensemble learning and artificial neural networks. Evaluation
is performed using K-fold and group-fold cross-validation.
The final solution consists of a cascaded ensemble learning
model combining bagged linear regression built on polyno-
mial features and random forest. This model demonstrates
robust performance, achieving a test score of 0.955719 and a
validation score of 0.886953, showcasing the effectiveness of
uncertainty-aware machine learning methods in health moni-
toring systems.
1. INTRODUCTION
To increase operational availability of helicopters, reduce the
required number of maintenance activities and increase the
inspection interval period, it is important to implement con-
dition based maintenance systems, which are based on health
and usage monitoring (Banks et al., 2011). In this context,
fault detection and prognostics are important tasks in Systems
Health Management, which improve system safety while re-
ducing operating and maintenance costs (Berri, Dalla Vedova,
& Mainini, 2019; Ribeiro & Reynoso-Meza, 2018). The field
has taken huge advantage of using data-driven solutions for
such tasks for a long time (Schwabacher & Goebel, 2007).
Victor Ribeiro et al. This is an open-access article distributed under the terms
of the Creative Commons Attribution 3.0 United States License, which per-
mits unrestricted use, distribution, and reproduction in any medium, provided
the original author and source are credited.
However, there are still many problems that difficult the de-
ployment of such solutions in practice. One of such issues
is the lack of methods to estimate the uncertainty of the pre-
dictions, which aim to increase the reliability of such systems
(Zio, 2022).
Uncertainty quantification is the process of characterizing the
proximity between predictions and observations (Ghanem,
Higdon, Owhadi, et al., 2017). Recent studies have em-
ployed different uncertainty-aware machine learning methods
to fault detection and health monitoring, such as using Monte-
Carlo dropout in Artificial Neural Networks (Das, Gjorgiev,
& Sansavini, 2024), predicting output distribution functions
with deep learning (Yao, Han, Yu, & Xie, 2024), and building
uncertainty-aware ensemble models (Kafunah, Ali, & Bres-
lin, 2023).
Given the number of possible techniques to build prognostics
and health management systems, as well as the complexity,
risks, and timespan associated with the product development,
it is important to define a life cycle for the project stages and
milestones (Hu, Miao, Si, Pan, & Zio, 2022). In this sense,
the design science research methodology (Peffers, Tuunanen,
Rothenberger, & Chatterjee, 2007) has shown to be a valuable
tool when developing machine learning models (Pumplun,
Peters, Gawlitza, & Buxmann, 2023; Del Mar-Raave, Bahs¸i,
Mr
ˇ
si
´
c, & Hausknecht, 2021; Duque, Godinho, Moreira, &
Vasconcelos, 2024).
This work employs an iterative development cycle to build
an uncertainty-aware health monitoring system for helicopter
turbine engines. We compare different methods for data pro-
cessing, such as building polynomial features and data cleans-
ing, model development, such as ensemble learning and ar-
tificial neural networks, and evaluation techniques, such as
K-fold cross-validation and group-fold cross-validation. The
final proposed solution comprises a cascaded ensemble learn-
ing model using bagged linear regression and random forest,
1
