Article
Contrastive Learning for Fault Detection and Diagnostics in the
Context of Changing Operating Conditions and Novel
Fault Types
Katharina Rombach, Gabriel Michau and Olga Fink *
Citation: Rombach, K.; Michau, G.;
Fink, O. Contrastive Learning for
Fault Detection and Diagnostics in
the Context of Changing Operating
Conditions and Novel Fault Types.
Sensors 2021, 21, 3550. https://
doi.org/10.3390/s21103550
Academic Editors: Kim Phuc Tran,
Athanasios Rakitzis and Khanh T. P.
Nguyen
Received: 2 April 2021
Accepted: 14 May 2021
Published: 20 May 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Swiss Federal Institute of Technology, ETH Zürich, 8092 Zürich, Switzerland; rombachk@ethz.ch (K.R.);
gmichau@ethz.ch (G.M.)
* Correspondence: ofink@ethz.ch
Abstract:
Reliable fault detection and diagnostics are crucial in order to ensure efficient operations
in industrial assets. Data-driven solutions have shown great potential in various fields but pose
many challenges in Prognostics and Health Management (PHM) applications: Changing external
in-service factors and operating conditions cause variations in the condition monitoring (CM) data
resulting in false alarms. Furthermore, novel types of faults can also cause variations in CM data.
Since faults occur rarely in complex safety critical systems, a training dataset typically does not
cover all possible fault types. To enable the detection of novel fault types, the models need to be
sensitive to novel variations. Simultaneously, to decrease the false alarm rate, invariance to variations
in CM data caused by changing operating conditions is required. We propose contrastive learning
for the task of fault detection and diagnostics in the context of changing operating conditions and
novel fault types. In particular, we evaluate how a feature representation trained by the triplet loss
is suited to fault detection and diagnostics under the aforementioned conditions. We showcase
that classification and clustering based on the learned feature representations are (1) invariant to
changing operating conditions while also being (2) suited to the detection of novel fault types. Our
evaluation is conducted on the bearing benchmark dataset provided by the Case Western Reserve
University (CWRU).
Keywords: contrastive learning; triplet loss; fault diagnostics; fault detection
1. Introduction
Modern industrial processes are increasingly subject to oversight by condition monitor-
ing (CM) devices. The recorded data opens up the possibility of data-driven maintenance
models [
1
]. Purely data-driven solutions are especially interesting with regard to complex
assets for which model-based approaches are limited or do not exist. Recent successes in
deep learning have demonstrated the potential of data-driven solutions [
2
,
3
]. However,
for the task of fault detection and diagnostics, particular challenges arise when applying
deep learning to CM data from an industrial asset.
Complex industrial assets are often subject to a variety of operating conditions as
well as external (e.g., environmental) factors that strongly influence the acquired data.
Changing ambient temperature, for example, might affect the roughness of the asset, which
could then be sensed by accelerometer measurements resulting in changes of the signals.
The ambient temperature is therefore a factor that causes variations in the data but cannot
be controlled. This means that a complete training dataset that is recorded in summer
will deviate from the data experienced in the winter season. Predicting or foreseeing all
of these influential factors is not always possible as some factors of variations are simply
not known or cannot be controlled. Even if all future operating conditions are completely
controllable and known (e.g., defined in the specifications of a working environment),
the multitude of possible combinations makes it often infeasible to collect a dataset with
Sensors 2021, 21, 3550. https://doi.org/10.3390/s21103550 https://www.mdpi.com/journal/sensors
