2021 China Automation Congress (CAC),
Год журнала:
2023,
Номер
unknown, С. 7372 - 7377
Опубликована: Ноя. 17, 2023
Semi-supervised
learning
can
effectively
utilize
limited
labeled
data
and
large
amounts
of
unlabeled
to
achieve
fault
diagnosis
on
process
industries.
This
paper
proposes
a
novel
cosine
distance
based
semi-supervised
using
discriminant
graph
convolutional
networks
(CD-GCN)
at
node-level
Firstly,
the
CD-GCN
method
uses
information
pull
training
sample
features
different
classes
farther
away
from
each
other.
Secondly,
replaces
Euclidean
with
Cosine
as
metric
in
original
samples
space
feature
space.
With
information,
better
considers
spatial
structure
improve
whole
by
dual
effects
convolution
nearest
neighboring
these
moving
features.
Finally,
real
industrial
simulation
case
is
carried
out
verify
performance
proposed
method.
Compared
other
related
classic
methods,
results
show
that
achieves
best
diagnostic
accuracy.
ACS Omega,
Год журнала:
2024,
Номер
9(5), С. 5954 - 5965
Опубликована: Янв. 24, 2024
Quality
variables
play
a
pivotal
role
in
monitoring
the
performance
of
chemical
production
systems.
However,
certain
critical
quality
cannot
be
measured
online
through
instruments.
In
such
scenarios,
using
soft
sensors
becomes
imperative
to
enable
real-time
measurements,
accurately
reflecting
system's
operational
status.
The
development
high-performance
requires
abundantly
labeled
samples.
Nevertheless,
prolonged
periods
and
substantial
costs
associated
with
acquiring
variable
data
pose
challenges
obtaining
sufficient
Therefore,
this
paper
proposes
regression
generative
adversarial
network
generate
virtual
proposed
method
considers
mapping
relationship
between
auxiliary
target
while
learning
distribution.
Moreover,
importance-weighted
autoencoder
is
introduced
enhance
training
stability
model.
samples,
selected
by
similarity
measurement
algorithm,
are
incorporated
into
set.
This
inclusion
addresses
diminished
predictive
when
samples
insufficient.
sensor
employed
anaerobic
digestion
process
serves
as
case
study
illustrate
efficacy
method.
Experimental
results
validate
that
generated
exhibit
greater
proximity
actual
compared
those
other
methods.
Furthermore,
integrating
long
short-term
memory-based
yields
21.03%
reduction
root-mean-square
error
original
set
alone.
Journal of Vibroengineering,
Год журнала:
2025,
Номер
27(1), С. 93 - 107
Опубликована: Янв. 22, 2025
It
usually
affects
the
accuracy
and
reliability
of
deep
learning
based
intelligent
diagnosis
methods
under
condition
insufficient
samples.
Existing
for
handling
samples
often
have
problems
such
as
requiring
rich
expert
experience
or
consuming
a
lot
time.
To
solve
above
problems,
rolling
bearing
fault
method
on
multi-scale
long-term
short-term
memory
network
(MSLSTM)
transfer
is
proposed,
which
mainly
consists
an
improved
named
MSLSTM
learning.
By
introducing
convolution
operation
into
traditional
LSTM
to
improve
its
drawback
that
only
extracts
single
type
feature
information,
leads
poor
diagnostic
performance
in
noisy
environments.
Besides,
pooling
layer
global
average
are
replaced
with
avoid
problem
information
loss.
Subsequently,
combined
learning,
fine
tunes
model
parameters
using
small
amount
target
domain
data.
Feasibility
proposed
verified
through
two
kinds
experiments.
The
has
stronger
extraction
ability
training
efficiency
compared
other
models.
IEEE Transactions on Industrial Informatics,
Год журнала:
2023,
Номер
20(2), С. 2378 - 2386
Опубликована: Июль 3, 2023
Soft
sensing
using
the
neural
network
technique
has
been
increasingly
applied
to
industrial
processes.
Recently,
security
and
robustness
of
network-based
soft
sensors
have
become
primary
concerns.
In
addition,
current
studies
indicated
that
networks
are
vulnerable
adversarial
attacks.
other
words,
small
perturbations
imposed
on
input
can
lead
significant
deviations
in
output.
If
a
sensor
for
key
process
variables
is
attacked,
considerable
damage
may
be
brought
This
article
focuses
attack
methods
sensors.
Considering
characteristics
sensors,
this
proposes
two
new
methods.
The
first
method,
called
mirror
output
(MOA),
subtle
method
flips
curve
change
direction
outputs.
second
translation
MOA
(TMOA),
easy
make
operators
misoperate.
TMOA
translates
while
flipping
achieve
purpose
changing
conditions.
effectiveness
demonstrated
an
case
study
sulfur
recovery
unit
process.
Simulation
results
show
attacked
by
both
proposed
provide
basis
defending
against
attacks,
thereby
enhancing
