Introduction - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 21

mixture problem. A new model is formulated assuming the impact-echo overall
scenario as a multiple-input-multiple-output linear time invariant system
(MIMO-LTI) [ 87 ]. Classification of an extensive set of materials is performed at
different levels of detail depending on the knowledge about the defects. It is
demonstrated that the mass spectra from impact-echo testing fit ICAMM and a
kind of defect signature is registered in the ICA mixture parameters. This application
represents the first contribution of the application of ICAMM to NDT [ 88 ].
1.3.3.2 Chronological Cataloguing of Archaeological Ceramics
In this application, a new hardware-software system is proposed for the charac-
terization of archaeological ceramics. The NDT methodology is based on ultra-
sounds. Mixca algorithm is used to classify ceramic shards using frequency and
temporal features extracted from the ultrasound signal. The archaeological prob-
lem researched consisted of chronological determination of the ceramic shards.
The results of classification of pieces from different deposits and ages demon-
strated the best performance for Mixca even over classic methods. Mixca with a
small number of labelled data (semi-supervised training) obtained higher classi-
fication accuracy than supervised methods. This is very interesting since it could
be used for handling the expert uncertainty in labelling the pieces. The method has
been patented and offers an alternative that could complement or replace the
destructive, costly, and time-consuming techniques that are currently used.
1.3.3.3 Analysis of Restoration in Historical Buildings
The problem here consists of distinguishing between consolidated and non-con-
solidated zones in a restored wall of a historical building. The ultrasound signal
measured is modelled with an ICA model. The mixture model considers a part of
the signal from material backscattering sources and another part of the signal from
sinusoidal interferences. It was possible to separate these parts and obtain
improved images of the inside of the wall. The non-consolidated zone in the wall
was separated from the consolidated zone using the sources extracted by ICA. In
addition, the model was used for a second application consisting of the estimation
of the thickness of the material layers in a historic building wall. The sources to
separate were material backscattering and interferences due to instrumental noise.
These applications are new in the field of ICA.
1.3.3.4 Diagnosis of Sleep Arousals in EEG Signals
The problem in this application is the detection of micro arousals that occur during
sleep at night due to apnea. Modelling of dynamic changes in the ICA mixtures by
including temporal dependencies in classification is proposed (that we called