On the performance evaluation of hybrid-trained neural classifier for the detection of hazardous vapours using responses from SAW sensors array

A neural classifier has been designed by a new two-phase hybrid training algorithm introduced by us for classification of hazardous vapours. The neural network is trained using genetic algorithm in initial phase. This is followed by a second phase of backpropagation training that uses weight matrix determined by first phase for initialization. For establishing the superior performance of our classifier, published data from polymer-coated surface-acoustic wave (SAW) sensors array exposed to varying concentration of each of nine vapours belonging to two different classes have been used. Vapours of class I are toxic vapours of interest in ambient air that contains common interferents (class II vapours) at much higher concentration. Performance of the classifier is evaluated by reducing dimensionality of resulting data matrix from 4 to 1 by taking a different set of sensors. We show that as the dimension is reduced, the gas identification problem becomes harder for backpropagation. Whereas the same set of problems when solved using a genetic algorithm with heuristic switch over to backpropagation as a training paradigm, significantly better results are obtained in predicting class and type of test vapours.