Information Technology Reference
In-Depth Information
Table 5.2
Confusion matrix for experiments at defect orientation level
Homoge-
neous
Hole
X
axis
Hole
Y
axis
Crack
XY
plane
Crack
ZY
plane
Crack
XZ
plane
Multiple-
defects
Homogeneous
1
0
0
0
0
0
0
Hole X axis
0
0.72
0
0.04x
0.12
0.1
0.02
Hole Y axis
0
0
0.75
0.07
0.11
0.06
0.01
Crack XY plane
0
0
0.04
0.84
0.01
0.05
0.06
Crack ZY plane
0
0.00
0.02
0
0.9
0.08
0
Crack XZ plane
0
0
0.01
0
0.1
0.89
0
Multiple-
defects
0.2
0
0
0.09
0.06
0.07
0.58
data is reduced to the problem of modelling the one-dimensional data density that is
associated to each possible source in every ICA member of ICAMM. Hence, in the
same way that ICA is a valuable tool for blind source separation by reducing higher-
order dependencies or by directly imposing statistical independence among the
sources, ICAMM is also a valuable tool for classification since each ICA component
can be associated to a different class of the global data model.
A new ICAMM approach has been proposed for non-destructive testing using
impact-echo. The material under evaluation is modelled as a linear system that
describes the wave propagation phenomenon of the impact-echo. A compressed
and representative pattern of the spectra for the multichannel setup of the impact-
echo has been obtained using ICAMM. This modelling allowed the spectra dif-
ferences in the resonance modes to be discerned for different kinds of defective
materials in several simulations and lab experiments.
We have demonstrated the feasibility of the proposed procedure for extracting
patterns of different kinds of defects from impact-echo signal spectra both in
simulations and in experiments. Note that there was only one piece of material for
a kind of defect in a certain location in the bulk, and it was not in the training
stage; therefore, the classifier had to assign the right class using the patterns of
pieces of the same class in other locations. The results could be used to implement
the proposed method in industrial applications of quality evaluation of materials.
In these applications, the database collected within a reasonable time could have
samples that are similar to the tested piece, significantly improving the classifi-
cation results.
General results show that a classifier based on a mixture of independent
components is a suitable technique for the impact-echo problem even in complex
levels of classifications in which up to 12 classes of homogeneous, single-defect
and multiple-defect materials are discerned. The underlying ICA mixture models
that were learned seem to be related to the shape and location of the defects. This
is a promising area of application for ICA mixture modelling.
Search WWH ::
Custom Search