Authors: Nicole Challita; Mohamad Ali Khalil; Pierre Beauseroy
Addresses: Azm Center for Research and Biotechnology, EDST, Lebanese University, El-Mitein Street, Tripoli, Lebanon; UMR ICD/LM2S, University of technology of Troyes, 12 rue Marie-Curie, CS42060, 10004 Troyes, France ' Azm Center for Research and Biotechnology, EDST, Lebanese University, El-Mitein Street, Tripoli, Lebanon ' UMR ICD/LM2S, University of technology of Troyes, 12 rue Marie-Curie, CS42060, 10004 Troyes, France
Abstract: One of the most advanced forms of industrial maintenance is predictive maintenance. Indeed, the present analysis of the behaviour of a material helps to predict future behaviour. So as the diagnosis of faults in rotating machines is an important subject in order to increase their productivity and reliability, the choice of features to be used for classification and diagnosis constitutes a crucial point. The use of all the possible features will cause an increase in the computational cost and it will even lead to the increase of the classification error because of the existence of redundant and non-significant features. In this context, we are interested in presenting different methods of feature selection and proposing a new approach that tends to select the best features among existing ones and perform the classification-identification using the selected features. A study of the proposed method stability is also provided.
Keywords: feature selection; dimension reduction; feature selection methods; LASSO; Elastic Net; Relief; performance; stability; classification error; important features.
International Journal of Computer Aided Engineering and Technology, 2018 Vol.10 No.3, pp.320 - 339
Received: 04 Jan 2016
Accepted: 28 Mar 2016
Published online: 09 Mar 2018 *