Application of ICAMM to Impact-Echo Testing - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 121

Fig. 5.9 Scatterplot of the
experimental data and
estimated parameters of the
ICA mixture for Level ii of
classification: kind of defect.
The classification space
corresponds to the first three
spectra features of the data set
vectors

specific defect, every specific deployment of the sensors, and every specific impact
location. Since the deployment of sensors and impact location have been kept
constant, we have considered grouping different defects having a common char-
acteristic (like orientation, size and/or shape) in the same ICA class of an overall
ICAMM model. Hence, it is assumed that by keeping one or more of the defect
characteristics constant, a common ICA model could be estimated, because A and
s would not vary very much for the different defects of a class. This assumption is
required to define categories that describe, in more or less detail, the inner state of
the specimen from the multichannel impact-echo signals.

5.3.4 Results

In this section, we compare the results of Mixca with MLP and LDA. Figure 5.11
shows a summary of the classification results for the classifiers LDA, MLP and
Mixca using JADE (Mixca-JADE) and the non-parametric kernel-based density
estimation (Mixca) for the ICA parameter updating, for both simulations and
experiments (the accuracy is the mean percentage of success in classification). In
general, the curves show that the classification accuracy decreases when the
number of classes increases in more detailed levels of classification. For experi-
ments, the best results were obtained by non-parametric Mixca; for simulations,
the best results were obtained by LDA and non-parametric Mixca.
The experiment results were better than simulation results since the number of
specimens and models and the implementation to obtain the signals were different.
The experiments of impact-echo involved some randomness in their execution
since the force injected in the impact excitation and the positions of the sensors
(which can vary from piece to piece) are manually controlled. These variables
yield repetitions of the experiments with their corresponding signal spectra. These
spectra separated class regions better than the Gaussian noise that was used to
obtain

replicates of the simulated model signals (since data densities are