A new methodology for improving the performance of fault diagnosis systems (FDS) has been proposed by combining both supervised and unsupervised learning methods. Within this framework, different techniques have been applied, such as Independent Component Analysis (ICA) as feature extraction method, Gaussian Mixture Models (GMM) with Bayesian Information Criterion (BIC) for unsupervised clustering and Support Vector Machines (SVM) for the classification steps. Since supervised learning may fail to fully discriminate some individual faults, the algorithm presented allows the unsupervised grouping of some critical faults (classes) having a diagnosis performance below a threshold defined by the user. Next, an additional classification step provides practical information for decision-making in terms of the quantitative confidence on the occurrence of one fault (or some) among a known reduced subset. The methodology presented was assessed on the Tennessee Eastman Process (TEP) benchmark. The whole set of the TEP faults was considered and an improved diagnosis performance was obtained for all of them, including those faults (3, 9 and 15) whose diagnosis had hardly been addressed previously. These results demonstrate the enhanced capability of this method and the promising potential for the diagnosis of industrial applications. (C) 2009 Elsevier Ltd. All rights reserved.