Systems Science & Control Engineering,
Год журнала:
2024,
Номер
12(1)
Опубликована: Апрель 30, 2024
With
the
rapid
development
of
artificial
intelligence,
deep
learning
is
considered
a
promising
technique
for
intelligent
fault
diagnosis
using
large
amounts
data
in
various
industrial
fields.
Under
such
circumstances,
imbalanced
datasets
real
world
will
not
only
hinder
further
classification
models,
but
also
degrade
performance
existing
models.
To
overcome
this
limitation,
paper
proposes
novel
Mel-Frequency
Cepstral
Coefficent-based
Generative
Adversarial
Network
(MFCC-GAN)
to
augment
high-quality
small
class
data.
Specifically,
MFCC
first
used
capture
time-
and
frequency-domain
features
signals
as
priori
information,
which
then
fed
into
generative
model.
The
temporal
structural
energy
contained
prior
information
can
provide
effective
guidance
process
model
map
Gaussian
distribution
real-world
distribution.
Moreover,
contrastive
loss
introduced
refine
discriminative
generated
signals,
aiming
improve
distinguishability
among
different
health
states.
Experimental
results
show
that
MFCC-GAN
algorithm
improves
quality
fidelity
compared
other
state-of-the-art
algorithms.
International Journal of Network Dynamics and Intelligence,
Год журнала:
2023,
Номер
unknown, С. 100017 - 100017
Опубликована: Дек. 21, 2023
Article
An
Improved
Generative
Adversarial
Network
with
Feature
Filtering
for
Imbalanced
Data
Jun
Dou
1,
and
Yan
Song
2,*
1
Department
of
Systems
Science,
University
Shanghai
Science
Technology,
200093,
China
2
Control
Engineering,
*
Correspondence:
[email protected];Tel.:+86-21-55271299;
fax:+86-21-55271299
Received:
7
October
2023
Accepted:
31
Published:
21
December
Abstract:
adversarial
network
(GAN)
is
an
overwhelming
yet
promising
method
to
address
the
data
imbalance
problem.
However,
most
existing
GANs
that
are
usually
inspired
by
computer
vision
techniques
have
not
taken
significance
redundancy
features
into
consideration
delicately,
probably
producing
rough
samples
overlapping
incorrectness.
To
this
problem,
a
novel
GAN
called
improved
feature
filtering
(IGAN-FF)
proposed,
which
establishes
new
loss
function
model
training
replacing
traditional
Euclidean
distance
Mahalanobis
taking
ℓ1,2-norm
regularization
term
consideration.
The
remarkable
merits
proposed
IGAN-FF
can
be
highlighted
as
follows:
1)
utilization
make
fair
evaluation
different
attributes
without
neglecting
any
trivial/small-scale
but
significant
ones.
In
addition,
it
mitigate
disturbance
caused
correlation
between
features;
2)
embedding
contributes
greatly
guaranteeing
sparsity
well
helps
reduce
risk
overfitting.
Finally,
empirical
experiments
on
16
well-known
imbalanced
datasets
demonstrate
our
performs
better
at
metrics
than
other
11
state-of-the-art
methods.
Measurement Science and Technology,
Год журнала:
2024,
Номер
35(4), С. 046125 - 046125
Опубликована: Янв. 23, 2024
Abstract
Water
supply
pipeline
leakage
not
only
wastes
resources
but
also
causes
dangerous
accidents.
Therefore,
detecting
the
state
of
pipelines
is
a
critical
task.
With
expansion
scale
water
pipeline,
amount
data
collected
by
leak
detection
system
gradually
increasing.
Moreover,
there
an
imbalance
sample
in
data.
This
makes
performance
traditional
methods
deteriorate.
To
solve
above
issues,
this
paper
proposes
intelligent
method
based
on
support
vector
weighted
twin-bound
machine
(SV-WTBSVM).
Noise
negatively
affects
classifier.
eliminate
effect
noise,
hybrid
denoising
algorithm
improved
complete
ensemble
empirical
mode
decomposition
with
adaptive
noise
(ICEEMDAN)
used
for
to
filter
out
Twin
bound
(TBSVM)
classical
classification
that
has
been
widely
leakage.
decrease
accuracy
caused
imbalance,
SV-WTBSVM
oversamples
minority
class
samples
distance
density
and
integrally
undersamples
majority
obtain
balanced
sample.
Since
often
have
multiple
working
conditions,
binary
cannot
meet
requirement,
combines
‘one-to-one’
strategy
address
multi-classification
problem.
Finally,
experiments
verified
retains
advantages
fast
training
speed
simple
operation
TBSVM
improves
generalization
ability
when
dealing
imbalanced
