Model-Based Feature Selection Based on Radial Basis Functions and Information Measures

In this paper the development of a new embedded feature selection method is presented, based on a Radial-Basis-Function Neural-Fuzzy modelling structure. The proposed method is created to find the relative importance of features in a given dataset (or process in general), with special focus on manufacturing processes. The proposed approach evaluates the impact/importance of processes features by using information theoretic measures to measure the correlation between the process features and the modelling performance. Crucially, the proposed method acts during the training of the process model; hence it is an embedded method, achieving the modelling/classification task in parallel to the feature selection task. The latter is achieved by taking advantage of the information in the output layer of the Neural Fuzzy structure; in the presented case this is a TSK-type polynomial function. Two information measures are evaluated in this work, both based on information entropy: mutual information, and cross-sample entropy. The proposed methodology is tested against two popular datasets in the literature (IRIS - plant data, AirFoil - manufacturing/design data), and one more case study relevant to manufacturing - the heat treatment of steel. Results show the good and reliable performance of the developed modelling structure, on par with existing published work, as well as the good performance of the feature selection task in terms of correctly identifying important process features.