Development of a methodology for diagnosing faults in bearings
operating under variable operating conditions based on
self-supervised learning
Racquel Knust Domingues
1
, J
´
ulio A. Cordioli
2
, and Danilo Silva
3
1,2,3
Federal University of Santa Catarina, Florian
´
opolis, Santa Catarina, 88040-900, Brazil
racquel.knust@lva.ufsc.br
julio.cordioli@ufsc.br
danilo.silva@ufsc.br
1. PR OBLEM CONTEXTUALIZATION
Predictive maintenance is crucial for ensuring the efficiency
and availability of industrial assets by analyzing their current
state to predict failures and enable timely corrective actions.
Among the existing industrial assets, rotating elements are
commonly used, leading to widespread utilization of bear-
ings, as they are essential for reducing friction in rotary mo-
tion (Lei, 2016). Despite the many existing methods for de-
tecting and diagnosing faults in these elements, the increas-
ing complexity of systems due to technological advancements
has led to a greater diversity of operating conditions, requir-
ing new diagnostic methods.
Fault diagnosis methods can be classified into physical
model-based, prior knowledge-based, and data-based meth-
ods (X. Zhang, Zhao, & Lin, 2021). The first two rely on
a deep understanding of the element’s physical behavior and
can become complex and challenging to implement. In con-
trast, data-based methods have proven efficient by extracting
useful maintenance information directly from the measured
data of the asset’s internal parameters.
Data-based methods consist of three main steps (Mushtaq, Is-
lam, & Sohaib, 2021). The first step concerns data acquisi-
tion. In this stage, an internal parameter of the asset, vibration
will be considered in thus study, must be measured and stored
for later analysis. In the second step, the process of attribute
extraction and selection occurs. Here, various techniques can
be used, and the goal is to extract representations and metrics
that make it possible to distinguish between data from differ-
ent conditions. The final step involves analyzing the attributes
obtained in the previous step, allowing for the determination
of the asset’s current condition and establishing a solid foun-
dation for maintenance decision-making.
Racquel Domingues 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.
Traditionally, the steps of attribute extraction and selection, as
well as state classification, were performed manually, relying
on the expertise of analysts knowledgeable about the behav-
ior of the machine element. With the development of machine
learning algorithms, this process began to be automated, mak-
ing fault diagnosis methods increasingly robust (S. Zhang,
Zhang, Wang, & Habetler, 2020). Initially, traditional shal-
low learning algorithms, such as SVM, were applied only in
the classification step, which still required manual extraction
and selection of attributes from the data processing. Subse-
quently, with the development and use of deep learning meth-
ods, the attribute extraction process also became automatic,
as these methods have the capability to automatically extract
the most useful representations for the task to which they are
applied.
Deep learning algorithms applied to fault diagnosis typically
rely on supervised learning, which requires labeled data. The
performance of these algorithms improves with the num-
ber of labeled samples (Long, Chen, Yang, Huang, & Li,
2023). However, labeling large datasets can be costly and
impractical due to diverse operating conditions. To address
these challenges, researchers are exploring self-supervised
learning (Chowdhury, Rosenthal, Waring, & Umeton, 2021),
which leverages unlabeled data to learn useful representations
from artificial labels created from the data itself. These rep-
resentations can then be transferred to specific fault diagnosis
tasks, allowing the model to achieve high performance with
fewer labeled samples.
2. THEORETICAL BACKGROUN D
To learn useful representations of the data within self-
supervised learning, a pretext task must be defined, and
deep learning models are used to solve it. Additionally,
this methodology involves the automatic creation of la-
bels for unlabeled data according to the defined pretext
task (Morningstar et al., 2024). A pretext task can be de-
1