Polymer Composites,
Journal Year:
2024,
Volume and Issue:
unknown
Published: Sept. 13, 2024
Abstract
Artificial
intelligence
(AI)
techniques
are
increasingly
used
for
structural
health
monitoring
(SHM)
of
polymer
composite
structures.
However,
to
be
confident
in
the
trustworthiness
AI
models,
models
must
reliable,
interpretable,
and
explainable.
The
use
explainable
artificial
(XAI)
is
critical
ensure
that
model
transparent
decision‐making
process
predictions
it
provides
can
trusted
understood
by
users.
existing
SHM
methods
structures
lack
explainability
transparency,
therefore
reliable
damage
detection.
Therefore,
an
interpretable
deep
learning
based
on
vision
transformer
(X‐ViT)
proposed
composites,
leading
improved
repair
planning,
maintenance,
performance.
approach
has
been
validated
carbon
fiber
reinforced
polymers
(CFRP)
composites
with
multiple
states.
X‐ViT
exhibited
better
detection
performance
compared
popular
methods.
Moreover,
effectively
highlighted
area
interest
related
prediction
each
condition
through
patch
attention
aggregation
process,
emphasizing
their
influence
process.
Consequently,
integrating
ViT‐based
deep‐learning
into
provided
diagnostics
along
increased
transparency
reliability.
Highlights
Autonomous
using
model.
Explainable
highlighting
region
attention.
Comparison
state
art
Mechanical Systems and Signal Processing,
Journal Year:
2024,
Volume and Issue:
211, P. 111194 - 111194
Published: Feb. 2, 2024
Compared
to
the
single-source
domain
adaptation
fault
diagnosis
methods,
multi-source
methods
can
not
only
take
advantage
of
rich
and
diverse
diagnostic
information
multiple
source
domains
but
also
draw
on
feature
alignment
setting
reduce
discrepancy.
However,
forcing
distributions
is
challenging
may
lead
negative
transfer.
Meanwhile,
labeled
data
are
often
scarce
difficult
collect
in
actual
production,
which
be
mitigated
by
information,
performance
model
degraded
large
differences.
To
tackle
above
issues,
a
attribute
transfer
network
proposed
unified
deep
achieve
cross-domain
diagnosis.
In
transferable
attributes
learning
section,
we
adopt
an
attention
mechanism
loss
function
extract
latent
from
information.
features
apply
local
maximum
mean
discrepancy
metric
adjust
category
distribution
target
domains.
Then,
intra-class
compactness
pseudo-labeling
strategies
utilized
further
obtain
richer
representations.
Finally,
propose
knowledge
fusion
module
fuse
results
classifiers
yield
more
reliable
result.
Extensive
experiments
three
different
datasets
show
superiority
our
method
compared
state-of-the-art
(SOTA)
comparing
indicators
various
aspects.
