Engineering Research Express,
Journal Year:
2024,
Volume and Issue:
6(4), P. 045430 - 045430
Published: Dec. 1, 2024
Abstract
Due
to
the
influences
of
sensor
faults,
communication
lines,
and
human
factors,
it
is
difficult
collect
label
fault
data
in
large
quantities,
resulting
imbalance
between
normal
data,
data.
Those
kinds
imbalances
violate
assumption
relatively
balanced
distribution
most
traditional
diagnosis
methods.
Associated
with
those
trends,
some
imbalanced
methods
have
been
put
forward.
However,
only
consider
that
proportion
various
samples
remains
unchanged,
is,
rate
stable.
In
actual
manufacturing
processes,
industrial
flows
are
fast,
continuous,
dynamically
changing.
The
rates
all
often
change
continuously,
showing
dynamic
characteristic.
To
solve
this
problem,
a
novel
adaptive
cost-sensitive
convolution
neural
network
based
framework
designed
for
processes.
More
specifically,
new
convolutional
firstly
constructed
by
coordinating
cross
entropy
loss
function
specific
cost
sensitive
index,
which
performance
indicators
comprehensively
considered.
Subsequently,
time
factor
reasonably
introduced
make
model
pay
more
attention
identification
flow,
aiming
at
improving
performance.
Finally,
sufficient
simulation
experiments
conducted
typical
process,
hot
rolling
demonstrate
superiority
proposed
compared
classical
algorithms.
Automation,
Journal Year:
2025,
Volume and Issue:
6(2), P. 14 - 14
Published: March 30, 2025
Rotating
machines
predominantly
operate
under
healthy
conditions,
leading
to
a
limited
availability
of
fault
data
and
significant
class
imbalance
in
diagnostic
datasets.
These
challenges
hinder
the
development
deployment
diagnosis
methods
based
on
deep
learning
practice.
Considering
these
issues,
novel
hierarchical
adaptive
wavelet-guided
adversarial
network
with
physics-informed
regularization
(HAWAN-PIR)
is
proposed.
First,
wavelet-based
severity
score
used
quantify
within
Second,
HAWAN-PIR
generates
synthetic
time
domain
via
multiscale
wavelet
decomposition
represents
first
attempt
embed
incorporate
relevant
knowledge.
The
quality
then
evaluated
comprehensive
synthesis
index.
Furthermore,
scale-aware
dynamic
mixing
algorithm
proposed
optimally
integrate
real
data.
Finally,
one-dimensional
convolutional
neural
(1-D
CNN)
employed
for
extracting
features
classifying
faults.
effectiveness
method
validated
through
two
case
studies:
motor
bearings
planetary
gearboxes.
results
show
that
can
synthesize
high-quality
fake
improve
accuracy
1-D
CNN
by
17%
bearing
15%
gearbox
case.
Confidence
calibration
in
classification
models,
a
technique
to
achieve
accurate
posterior
probability
estimation
for
results,
is
crucial
assessing
the
likelihood
of
correct
decisions
real-world
applications.
Class
imbalance
data,
which
biases
learning
model
and
subsequently
skews
probabilities
model,
makes
confidence
more
challenging.
Especially
often
important
minority
classes
with
high
uncertainty,
complex
necessary.
Unlike
previous
surveys
that
typically
separately
investigate
or
class
imbalance,
this
paper
comprehensively
investigates
methods
deep
learning-based
models
under
imbalance.
Firstly,
problem
data
outlined.
Secondly,
novel
exploratory
analysis
regarding
impact
on
carried
out,
can
explain
some
experimental
findings
existing
studies.
Then,
conducts
comprehensive
review
57
state-of-the-art
divides
these
into
six
groups
according
method
differences,
systematically
compares
seven
properties
evaluate
their
superiority.
Subsequently,
commonly
used
emerging
evaluation
field
are
summarized.
Finally,
we
discuss
several
promising
research
directions
may
serve
as
guideline
future
