Data-Dimensionality Reduction Using Information-Theoretic Stepwise Feature Selector

Abstract

A novel information-theoretic stepwise feature selector (ITSFS) is designed to reduce the dimension of diesel engine data. This data consist of 43 sensor measurements acquired from diesel engines that are either in a healthy state or in one of seven different fault states. Using ITSFS, the minimum number of sensors from a pool of 43 sensors is selected so that eight states of the engine can be classified with reasonable accuracy. Various classifiers are trained and tested for fault classification accuracy using the field data before and after dimension reduction by ITSFS. The process of dimension reduction and classification is repeated using other existing dimension reduction techniques such as simulated annealing and regression subset selection. The classification accuracies from these techniques are compared with those obtained by data reduced by the proposed feature selector.