The computational cost of ML models
is different in the training and deployment
phase. In this case, the model is
expected to be updated monthly so that
training costs are not a problem. Training
such a model takes about a half an
hour with one GTX 2080 GPU. The
model consists of an encoder
pz | x
and a decoder px |,z
i^h
{^h which are
parameterized by an LSTM network of
parameters, i and {, respectively. The
LSTM-based VAE is trained to minimize
the following loss function:
T
|||
lE logpx z
=+
/
t
=1
Dq zx pz
~|
zq zx h 6
i^ tt
i tt
^
KL^^ hhh
{ ^ tt |
h@
"
DEVELOPMENT OF THE
TWIN SHIP PROVIDES A
UNIFIED SOLUTION THAT
SAVES DIFFERENT TYPES
OF DATA OF CRITICAL MACHINERY
IN A DATABASE.
„
at time ;tht 1^
h
^
and ct 1where
the first term in the loss function is the expected negative
log likelihood, which can be replaced by the meansquare
error between the inputs x and reconstructed inputs xt .