Material Identification of Particles in Space-Borne Electronic Equipments Based on Principal Component Analysis - Intelligent Computing for Sustainable Energy and Environment - page 466

4.2 SVM Design
The conventional classifier for binary problems cannot be used to deal with
multi-class classifications directly. In this research, the binary tree (BT) method
was adopted to identify the varieties of particle materials. The SVM-BT is able
to employ a coarse-to-fine strategy that makes coarse classes easy to differenti-
ate. According to the above analysis, using one SVM (SVM1) is first achieved
classification with a high accuracy on two coarse classes, namely the metallic ma-
terial (tin and aluminum) and nonmetallic material (wire), and then go through
the next level of classification on tin and aluminum classes using another SVM
(SVM2), as shown in Fig.5.
Fig. 5. Multi-class SVM design for particle material identification
4.3 Classification Result
In this study, particle impact signals obtained in the laboratory with the afore-
mentioned three types of materials are used, and the data are split into one
training data set and one test data set. The training data contains 250 sam-
ples of class 1 (wire), 250 samples of class 2 (tin) and 250 samples of class 3
(aluminum) from 0.5mg to 10mg particle impact signals. These data were all
randomly selected to provide enough information for the training model.
In order to obtain the input parameters, especially the first N principal com-
ponents which has most significant effect on the classification accuracy, SVM2 in
Fig. 5 was chosen as the object to be optimized due to the diculty in accurately
distinguishing between tin and aluminum. Thus, by analyzing the classification
performance of SVM2 with different number of input vectors after PCA, the
number of support vectors (SVs), the different margins, and the error rate are
shown in Table 1. It was found that N =6( R N =0 . 978) is the best choice that
will produce the SVM model with better generalization performance.
The classification results of the SVM classifiers on the test data are presented
in Table 2. With the first six principal components features, the accuracy rate of
classifying the wire particle class is 94.8%, 90.8% for the aluminum and 90.0%
for the tin. As listed in table 3, it is also found that the accuracy rate with nine
features directly is 91.2%, 82.0% and 84.0% for wire, aluminum and tin particles,
respectively. This indicates that the method to combine the PCA is more effective
than directly using SVMs. Statistically, misjudgements exist between a small
mass (0.5-2 mg) wire particle and a large mass (8-10 mg) tin or aluminum
particle, and between tin and aluminum in the same mass. This overlapping
patterns are caused by different size of the particles.