Building SVM and PNN optimal classifiers based on GA-PLS algorithm and the application in infrared spectrum

Abstract

Olefin infrared spectrum is a comprehensive presentation of its feature data. If the structures are different, it will certainly lead to feature differences between the spectrums. In this paper, with olefin cis/trans IR spectrums in OMNIC IR database as research objects, we have designed four types of classifiers based on support vector machine SVM and probabilistic neural network PNN upon Fisher ratios and genetic algorithm and partial least square GA-PLS, respectively, so as to select the optimal classifiers to apply into other databases. The results show that: all the optimal classifiers based on SVM and PNN are designed with GA-PLS algorithm; and when the corresponding feature sets include the 70 features and 50 features, respectively, selected by GA-PLS, the classifiers are optimal. Upon case verification, it is found that: SVM-GA-PLS classifier is more suitable for the prediction to olefin cis-structure and PNN-GA-PLS is more suitable for the prediction to olefin trans-structure.