Coupling fast variable selection methods to neural network-based classifiers: Application to multisensor systems

We address the problem of variable selection within the pattern recognition engine of multisensor systems. This problem arises when there is high dimensionality in the data used for training (e.g. because a high number of sensors are used, or because many features are extracted from the response of each sensor, or both). Different variable selection techniques (including deterministic and stochastic methods) have been coupled with neural network-based classifiers. The usefulness of each technique implemented is benchmarked by evaluating its performance in terms of three objective parameters: the success rate in classification, the dimensionality of the final set of variables used for training and the time needed to complete the variable selection procedure. The database consisted of 96 measurements of ammonia, acetone and o-xylene vapours and their binary mixtures gathered using a metal oxide gas sensor array. Each measurement was described by 120 variables (12 sensors and 10 parameters each). A new strategy for variable selection, which is based on a two-step approach, is introduced that leads to the building of parsimonious classification models based on either the fuzzy ARTMAP or the probabilistic neural networks. For example, a 91.66% success rate in the simultaneous identification and quantification of the species and their mixtures was obtained using nine input variables only (out of the 120 available). This process of variable selection was conducted in two-step, i.e. a coarse selection based on a variance criterion followed by a simulated annealing process. This two-step variable selection took about 10 min to complete in a Pentium 4 PC platform.