Neural net and genetic algorithms for spectral pattern recognition

Abstract

Spectral pattern recognition (SPR) methods are among the most powerful tools currently available for noriinvasively examhiin the spectroscopic and other chemical data for environmental analysis and monitoring. Using spectral data, these systems have found a variety of applications in chemometric systems such as gas chromatography, fluorescence spectroscopy, etc. An advantage of SPR approaches is that they make no a priori assumption regarding the structure of spectra. However, a majority o these systems rely on humanjudgment for parameter selection and classification. We considered a SPR problem as a composite of five subproblems: pattern acquisition, feature extraction, feature selection, knowledge organization, and pattern classification. One ofthe basic issues in SPR approaches is to determine and measure the features useful for successful classification. Selection of features that contain the most discriminatory information is important because the cost of pattern classification is directly related to the number offeatures used for classification. Various features present in a pattern and a large variety of classification algorithms could be used. A spectral pattern classification system combining the above components and multivariate decisiontheoretic approaches for classification is developed. It is shown how such a system can be used for large data analysis, warehousing, and interpretation. In a preliminary test, the system was used to classif' synchronous UV-vis fluorescence spectra ofrelatively similar petroleum oils with reasonable success.