IEEE Systems, Man and Cybernetics Magazine - January 2023 - 19

and replacement of faulty bearings
are necessary. In this view, the
prediction of bearing RUL can
effectively enable condition-based
and predictive maintenance to
save maintenance costs and reduce
safety hazards [3].
The popularization and applicaWe
propose a DPNN
to predict the RUL
of bearings.
tion of automation and artificial
intelligence are expected to greatly facilitate the operation
and management of equipment [4]. To ensure safety
in production, degradation prediction becomes very
important. By monitoring the health state of bearings, it
is possible to output maintenance decisions on whether
bearings need to be replaced. Many scholars have made
great progress in the areas of degradation and life
prediction. In particular, datadriven
modeling and analysis
have been extensively proposed
to achieve bearing prediction [5].
Shao and Nezu proposed a progression-based
prediction of
remaining life to predict the RUL
[6]. Among artificial NNs (ANNs),
Gebraeel et al. [7] used a feedforward
NN (FFNN) to predict the degradation of machines.
Pan et al. used the Enhanced Online Sequential Learning
Fuzzy Neural Network (EOSL-FNN) to predict the condition
of bearings [8]. Vachtsevanos proposed the prediction
of crack propagation in rolling bearings by a
recurrent wavelet NN (RWNN) [9].
Deep learning, which consists of multiple levels of
complex learning structures, has recently developed
into a successful new branch of machine learning [10],
with significant progress in many applications [11].
Still, now much work must be done related to the application
of deep learning NNs for the prediction of bearing
RUL.
We propose a DPNN to predict the RUL of bearings.
We use vibration characteristics obtained from the
data collected from the bearings' run to failure. We
extract the features from the complete data by wavelet
packet decomposition (WPD). Then, data from the
extracted features are selected for training and testing.
We compare the proposed DPNN with the LS-SVM
and LSTM.
The DPNN overcomes the weaknesses of standard
ANNs with advantages in ease of implementation and
small training set requirements. Furthermore, extracting
effect ive features from the new signals can
improve the speed and accuracy of the prediction [12].
Some researchers have extracted vibration data features
by short-time Fourier transform (STFT) or Wigner-Ville
distribution. In this article, we use WPD,
which is more effective than the STFT and Wigner-
Ville distribution.
In the " Proposed Prediction Methodology for Bearings "
section, we illustrate the RUL prediction model based on
the DPNN band WPD. In the " The Experiment " section, we
compare its accuracy with the LS-SVM. Finally, the " Conclusion "
section gives precise conclusions.
The Proposed Prediction Methodology
for Bearings
WPD
The extraction of characteristic features is of great significance
to the discovery of abnormal vibrations in bearings
due to defects and degradation [13]. To effectively
characterize the nonstationary and sensitive characteristics
of vibration signals for predicting bearings' RUL, this
article extracts the energy features by WPD, as shown in
January 2023 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 19
IEEE Systems, Man and Cybernetics Magazine - January 2023

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2023
