Conformal clustering for predictive maintenance: A case study at a commercial gas terminal

This poster presents an approach
for
predictive
maintenance
comprising
various
machine
learning
methods
such
as
conformal
clustering
and
probabilis6c
predic6on
using
K-­‐Nearest
Neighbours.
A
gas
terminal
has
a
component
the
field
gas
compressor
(FGC)
to
compresses
gas
to
the
required
pressure.
FGC
breaks
very
frequently
so
the
task
is
to
predict
its
failure.
It
contains
sensors
which
produce
data
on
183
aJributes.
We
show
that
using
a
combina6on
of
dimensionality
reduc6on
(t-­‐SNE)
and
the
conformal
clustering
algorithm
it
is
possible
to
produce
clusters
and
iden'fy
anomalies
which
lie
outside
these
regions.
We
also
suggest
an
approach
for
iden6fying
which
aJributes
are
the
most
likely
cause
of
failure
in
a
given
cluster.
Furthermore
we
predict
likelihood
of
failure
of
the
FGC
within
a
24-­‐hour
period
using
a
K-­‐Nearest
Neighbours
classifier
on
the
most
sta6s6cally
significant
aJributes.
We
find
that
when
making
predictions
with
high
confidence
the
classifier
is
highly
accurate.