Industrial & Engineering Chemistry Research,
Journal Year:
2024,
Volume and Issue:
63(44), P. 19051 - 19062
Published: Oct. 22, 2024
In
the
field
of
industrial
production,
precise
and
timely
implementation
fault
diagnosis
methods
is
crucial
for
improving
product
quality,
enhancing
operational
safety,
reducing
downtime,
minimizing
losses.
Recent
studies
have
shown
that
most
CNN-based
models
are
more
suitable
handling
Euclidean
data
such
as
images
or
videos
but
not
dealing
with
non-Euclidean
sensor
data.
practical
scenarios,
chemical
process
imbalanced
patterns
may
lead
data-driven
to
assign
different
attentions
patterns.
The
SMOTE
algorithm
commonly
used
generate
new
data,
it
often
tends
overfit
when
there
very
few
nearest
neighbor
samples.
To
address
these
issues,
we
designed
an
efficient
model
named
KRGAT.
fully
utilize
spatial
structural
information
on
employed
graph
attention
networks
(GATs),
which
well-suited
Additionally,
introduced
top-k
loss
method
select
hard
samples,
thereby
increasing
weight
Furthermore,
improved
DropMessage
enhance
model's
accuracy
robustness.
Experimental
results
demonstrate
our
outperforms
baseline
under
both
balanced
conditions.
Measurement Science and Technology,
Journal Year:
2025,
Volume and Issue:
36(3), P. 036203 - 036203
Published: Feb. 10, 2025
Abstract
Data
based
fault
diagnosis
technologies
are
important
measures
to
improve
the
operation
safety,
stability,
and
reliability
of
manufacturing
processes,
which
key
entry
points
innovation
powers
promote
intelligent
as
well
efficiency.
Class-balanced
datasets
often
used
for
modeling
by
traditional
data
methods.
However,
in
practical
engineering
applications,
processes
produce
multiple
classifications
imbalance
data,
bringing
great
challenges
promotions
applications
classical
To
this
end,
an
framework
is
proposed
issues
on
intraclass
interclass
have
been
specially
focused.
Specifically,
considering
non-independently
identically
distribution
characteristics
among
different
low
recognition
rates
minority
samples,
a
new
cost
sensitive
convolutional
neural
network
constructed
base
classifier
coordinating
cross
entropy
loss
function
with
specific
index.
Subsequently,
federated
learning
aggregation
algorithm
designed
optimize
participation
weights
local
classifiers
purpose
cooperating
model
generalization
performance.
Finally,
validity
demonstrated
typical
hot
rolling
process
forms
data.
The
simulation
results
show
that
superior
performance
can
be
achieved
compared
some
comparative
algorithms
each
scenario.
