Page 108 - IJOCTA-15-4

Zhang et al. / IJOCTA, Vol.15, No.4, pp.649-669 (2025)
be used for the exact identification of fault fea- ignoring the interrelationship between parame-
tures of rolling bearings. ters. Researchers have explored intelligent opti-
mization algorithms, such as the grey wolf op-
Traditional signal processing methods, such timizer (GWO), 18 the grasshopper optimization
6
as the Fourier transform and wavelet transform, 7 algorithm, 19 and cuckoo search, 20 to determine
are effective in extracting frequency-domain char- optimal values for K and α. Among these, golden
acteristics of fault signals. However, these meth- jackal optimization (GJO) 21 has demonstrated
ods have limitations when dealing with nonlin- faster convergence and a stronger ability to avoid
ear, non-stationary signals and lack adaptability. 8 local optima, making it more suitable for opti-
Consequently, researchers have developed new mizing VMD parameters for better decomposition
signal processing methods, such as empirical outcomes.
9
mode decomposition (EMD) and ensemble EMD However, despite the effectiveness of signal de-
(EEMD), 10 and have achieved certain successes. composition and reconstruction, capturing fault
For example, Zhou et al. 11 employed EMD to characteristics under complex operating condi-
filter noise from signals and selected sensitive tions remains challenging. Entropy measures,
intrinsic mode functions (IMFs) for reconstruc- such as multiscale fuzzy entropy (MFE), effec-
tion, achieving fault diagnosis of the bearing’s tively quantify signal regularity and detect subtle
inner race. Yao et al. 12 applied EMD with an changes, offering advantages like reduced sensitiv-
auto-regression model (AR) spectrum analysis for ity to parameter variations and high convergence
bearing fault diagnosis. Likewise, Qin et al. 13 accuracy. 22 This makes MFE capable of distin-
developed dynamic models for different bearing guishing different fault types. Therefore, in the
faults and extracted joint time–frequency entropy field of rotating machinery fault diagnosis, it is
features from EMD-decomposed signals for fault often used as a signal feature for fault pattern
classification using machine learning techniques. recognition. 1
Zhang et al. 14 proposed a new method using To further enhance diagnostic accuracy, effec-
EEMD and box dimension analysis, where de- tive classification models are needed for different
noised signals are decomposed, and relevant fea- types of bearing faults. The back-propagation
tures are used in a probabilistic neural network for (BP) neural network, 23 a classic machine learn-
fault identification. Similarly, Damine et al. 15 ap- ing algorithm, is widely applied in fault classifica-
plied a kurtosis-based median absolute deviation tion. It minimizes error by adjusting weights and
method to directly identify sensitive IMF com- biases and mapping input data nonlinearly. How-
ponents, demonstrating the effectiveness of this ever, traditional BP networks are prone to local
approach for bearing fault detection. minima and sensitive to initial parameters, reduc-
ing classification performance. To overcome this,
The performance of these methods in pro-
a hybrid optimization approach combining GWO
cessing nonlinear, non-stationary signals has also
and simulated annealing, termed the golden sine
been established. EMD suffers from mode-
subtraction-average-based optimizer (GSABO), is
mixing issues, while EEMD, though able to miti-
introduced. GSABO effectively optimizes BP
gate these issues, introduces white noise that re-
neural network parameters, avoiding local minima
quires extensive averaging, impacting decompo- and improving classification accuracy and stabil-
sition accuracy. 16 To address these issues, varia- ity.
tional mode decomposition (VMD), 17 the adap-
Aiming at the challenge of effectively extract-
tive signal decomposition method, has arisen as
ing early fault characteristics of rolling bearings
a prominent technique. VMD decomposes com-
and achieving reliable diagnosis results with small
plex signals into a series of IMFs while effectively
samples, this paper proposes a novel fault di-
avoiding mode mixing and providing good fre-
agnosis method. The signal feature information
quency resolution. Since its introduction, VMD
is extracted using VMD combined with MFE,
has been widely applied in rolling bearing fault
while the fault-diagnosis result is achieved using
diagnosis.
a GSABO–BP neural network. The major contri-
butions of this paper, compared to existing meth-
The choice of the decomposition number (K)
ods, are the following:
and the quadratic penalty parameter (α) has a
strong effect on the decomposition performance (i) The GJO algorithm is employed to jointly
of VMD, which significantly influences the re- optimize the two key decomposition pa-
sults. Relying on empirical or prior knowledge rameters of VMD—K and α. This ap-
to set these parameters may lead to inaccurate proach effectively eliminates the reliance
decomposition, reducing VMD’s efficiency and on manual parameter selection, which
650