1 Single and Ensemble Fault Classifiers Based on Features Selected by Multi-Objective Genetic Algorithms Enrico Zio, Piero Baraldi, Giulio Gola and Nicola Pedroni Politecnico di Milano, Dipartimento di Ingegneria Nucleare, Via Ponzio 34/3, 20133 Milan, Italy Abstract: The problem of identifying faults in systems and processes can be formulated as a problem of partitioning objects (i.e., the measured data pat- terns representing the symptoms) into classes (i.e., the types of faults causing the symptoms). In this view, two main steps need to be carried out in order to effectively perform the fault identification: i) the selection of the features carry- ing information relevant for the identification; ii) the classification of the mea- sured data patterns of features into the different fault types. In this work, the two tasks are tackled by combining a multi-objective genetic algorithm search with a Fuzzy K-Nearest Neighbors classification. Two different approaches to the development of the fault classification model are considered: a single classi- fier based on a feature subset chosen a posteriori on the Pareto-front identified by the multi-objective genetic search and an ensemble of classifiers, each one built on a different feature subset taken from the genetic algorithm population at convergence. Examples of application of the proposed approaches are given with reference to two different industrial processes: the classification of simu- lated nuclear transients in the feedwater system of a Boiling Water Reactor and of multiple faults in rotating machinery. Keywords and phrases: Fault classification, Feature Selection, Multi-objective genetic algorithms, Fuzzy K-Nearest Neighbors, Ensemble, Diversity 1.1 Introduction In this paper, the issue of fault diagnosis in complex engineering systems and processes is framed as a pattern classification problem. The basis for the clas- 1