Feature Selection and Categorization to Design Reliable Fault Detection Systems

Abstract

In this work, we will develop a fault detection system which is identified as a classification task. The classes are the nominal or malfunctioning state. To develop a decision system it is important to select among the data collected by the supervision system, only those carrying relevant information related to the decision task. There are two objectives presented in this paper, the first one is to use data mining techniques to improve fault detection tasks. For this purpose, feature selection algorithms are applied before a classifier to select which measures are needed for a fault detection system. The second objective is to use STRASS (STrong Relevant Algorithm of Subset Selection), which gives a useful feature categorization: strong relevant features, weak relevant and/or redundant ones. This feature categorization permits to design reliable fault detection system. The algorithm is tested on real benchmarks in medical diagnosis and fault detection. Our results indicate that a small number of measures can accomplish and perform the classification task and shown our algorithm ability to detect the correlated features. Furthermore, the proposed feature selection and categorization permits to design reliable and efficient fault detection system.