Conclusions - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 175

mass spectra from impact-echo testing fit ICAMM, and we also showed the
feasibility of this modelling to contribute to NDT applications.
• The accurate classification of archaeological ceramic fragments is quite a
challenging application. Very often, shards provide no evidence (decoration,
shape, etc.) about their origin, and they are obtained from deposits that are
removed from their original manufacturing location. Therefore, we devel-
oped an ultrasound-based NDT method by applying ICAMM (implemented
in the Mixca algorithm) for chronological cataloguing of archaeological
ceramics. We applied Mixca in order to take advantage of its capabilities to
deal with the arbitrary forms of the underlying probability densities in the
feature vector space (even using small size training samples) and to take
advantage of the possibility to incorporate partial labelling of the training
samples, which can model the uncertainty of an expert in shard labelling.
The features extracted from the ultrasonic signals were: centroid frequency;
maximum frequency; bandwidth; maximum frequency amplitude; parameters
A and b corresponding to an exponential model of the signal attenuation
x ð t Þ¼ Ae bt ; total signal power; propagation velocity; centroid frequency
evaluated at time t 0 ; and higher-order statistics (time reversibility and third-
order autocovariance). Mixca was tested using different variants depending
on the embedded ICA algorithm (non-parametric, TDSEP, JADE, FastIca).
The results demonstrated classification accuracy for shards from the fol-
lowing periods: the Bronze Age, Iberian, Roman, and the Middle Ages (from
deposits in Eastern Spain). Physical interpretation of the results was provided
considering various analyses: morphological and physiochemical character-
ization, ceramic composition and processing, and ultrasound propagation.
The new method, which included a prototype, was patented. It is a real
alternative to complement or replace destructive, costly, and time-consuming
techniques,

which are currently used by archaeologists for ceramic

characterization.
• Novel applications in restoration of historical buildings were introduced. Two
problems were approached: diagnosis of the consolidation status, and
detection of layers in a wall of a heritage building. The ultrasound signals
measured in auscultation of the wall were modelled as superposition of
backscattering from the material microstructure plus sinusoidal waves. The
sources recovered by Mixca (configured to estimate an ICA) allowed sinu-
soidal contributions to be accurately separated from the backscattering of the
wall. The sinusoidal sources corresponded to non-consolidated zones and
interferences. Therefore, enhanced B-Scans (images of the wall) were
achieved using ICA as preprocessor to eliminate or separate the sinusoidal
sources.
• Non-parametric ICA was introduced in webmining. The goal was to discover
student learning styles from a huge historical web-log database of the use of
facilities at a virtual campus. An ICA model was proposed assuming that the
underlying independent sources that generated the web log data were