Pattern recognition/expert system for mass spectra of volatile toxic and other organic compounds

Abstract A pattern recognition/expert system for classifying, estimating the molecular weights and identifying the low resolution mass spectra of toxic and other volatile organic compounds was developed with a rule-building expert shell program. The system was designed for low concentration spectra by using high intensity and high Shannon information content masses. A sequential filter design was used with an initial classification module followed by molecular weight and base peak filters and identification modules. Ternary encoded reference spectra of 74 target and 31 other compounds were used to establish the five target and the one unknown classes. The target classes were nonhalobenzenes; chlorobenzenes; bromo- and bromochloroalkanes/alkenes; mono- and dichloroalkanes/alkenes; and tri-, tetra- and pentachloroalkanes/alkenes. A molecular weight estimator provided molecular weight estimates even for unclassified spectra. The resulting system was extensively tested with reference spectra of 31 potential air pollutants, 100 randomly selected compounds, 100 GC-MS field spectra and with 54 000 NIST spectra. Classification and identification accuracies for training, test, random and field spectra were 100 and 100; 94 and 100; 76 and 100; and 90 and 96%, respectively. The median absolute deviations of the estimated molecular weights from the true values for the same sets were 0, 7, 7 and 1.