Using Explainable Artificial Intelligence to Interpret Remaining
Useful Life Estimation with Gated Recurrent Unit
Marcia L. Baptista
1
, Madhav Mishra
2
, Elsa Henriques
3
, and Helmut Prendinger
4
1
NOVA Information Management School (NOVA IMS),
Universidade Nova de Lisboa, Campus de Campolide,
1070-312 Lisboa, Portugal
m.baptista@novaims.unl.pt
2
RISE Research Institutes of Sweden, M
¨
olndal,SE-431 53, Sweden
madhav.mishra@ri.se
3
University of Lisbon, Instituto Superior Tecnico,Lisbon, Portugal
elsa.h@tecnico.ulisboa.pt
4
National Institute of Informatics,Tokyo, Japan
helmut@nii.ac.jp
ABSTRACT
In engineering, prognostics can be defined as the estimation
of the remaining useful life of a system given current and
past health conditions. This field has drawn attention from
research, industry, and government as this kind of technol-
ogy can help improve efficiency and lower the costs of main-
tenance in a variety of technical applications. An approach
to prognostics that has gained increasing attention is the use
of data-driven methods. These methods typically use pat-
tern recognition and machine learning to estimate the resid-
ual life of equipment based on historical data. Despite their
promising results, a major disadvantage is that it is difficult
to interpret this kind of methodologies, that is, to understand
why a certain prediction of remaining useful life was made
at a certain point in time. Nevertheless, the interpretability of
these models could facilitate the use of data-driven prognos-
tics in different domains such as aeronautics, manufacturing,
and energy, areas where certification is critical. To help ad-
dress this issue, we use Local Interpretable Model-agnostic
Explanations (LIME) from the field of eXplainable Artificial
Intelligence (XAI) to analyze the prognostics of a Gated Re-
current Unit (GRU) on the C-MAPSS data. We select the
GRU as this is a deep learning model that a) has an explicit
temporal dimension and b) has shown promising results in
the field of prognostics and c) is of simplified nature com-
Marcia L. Baptista 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.
pared to other recurrent networks. Our results suggest that it
is possible to infer the feature importance for the GRU both
globally (for the entire model) and locally (for a given RUL
prediction) with LIME.
1. INTRODUCTION
Improved reliability is one of the key drivers of the develop-
ment of more efficient maintenance strategies (J. Lee, Hol-
gado, Kao, & Macchi, 2014). The vision here is to have
machines that can monitor themselves and alert the operator
ahead of time of future maintenance needs to maximize func-
tion time and avert failure. The framework behind this vision
is that of Reliability Centered Maintenance (RCM) (NASA,
RCM, 2008), a discipline that aims to propose tools and prac-
tices to better monitor, predict and understand the behavior
of physical assets (Moubray, 2001). Major goals are to im-
prove safety, availability, reduce logistics and maintenance
costs, and to drive customer satisfaction and loyalty. Impor-
tantly, successful adoption of RCM aims to provide a greater
understanding of the nature of the risk that is being managed.
For a given physical asset, the outcome of an RCM program
is the implementation of an appropriate maintenance strat-
egy (NASA, RCM, 2008, pp. 3-1). A strategy that many
industries have followed for years due to its simplicity and
generality is the preventive or time-based maintenance. In
Time-Based Maintenance (TBM), repair and replacement are
based on simple measures of the expected life of the equip-
ment, such as calendar or usage time (NASA, RCM, 2008, pp.
1
