The feasibility of utilizing small, easily trained artificial neural networks for mass spectral data processing is demonstrated. The identification of individual components in a mixture of volatile organic compounds is accomplished by filtering their time-sampled mass spectra. Specifically, the chemical mass spectra of benzene, toluene, ethyl benzene and xylene are used as inputs to two different types of feed-forward neural networks. The performance of a network trained by back-propagation (BP) gradient descent is compared to the performance of a probabilistic neural network (PNN) trained with a single pass through a training data set. The results of the comparison are that the PNN network shows better noise rejection and fewer false-positive identifications than does the BP network. |