Nonlinear Analysis 71 (2009) e2665–e2673 Contents lists available at ScienceDirect Nonlinear Analysis journal homepage: www.elsevier.com/locate/na Real-time fault diagnosis of nonlinear systems ✩ Daniel F. Leite a, ∗ , Michel B. Hell a , Pyramo Costa Jr. b , Fernando Gomide a a Faculty of Electrical and Computer Engineering, University of Campinas, Brazil b Graduate Program in Electrical Engineering, Catholic University of Minas Gerais, Brazil a r t i c l e i n f o Keywords: Fault diagnosis Electrical machine Artificial neural network Real-time system a b s t r a c t This paper concerns the development of a real-time fault detection and diagnosis system for a class of electrical machines. Changes in the system dynamics due to a fault are detected using nonlinear models, namely,nonlinear functions of the measurable variables. At the core of the fault detection and diagnosis system are artificial neural networks and a new neural network structure designed to capture temporal information in the input data. Difficulties such as voltage unbalance, measurement noise, and variable loads, commonly found in practice, are overcome by the system addressed in this paper. Because false alarms are significantly reduced and the system is robust to parameter variations, high detection and diagnosis performance are achieved during both, learning and testing phases. Experimental results using actual data are included to show the effectiveness of the real- time fault detection system developed. © 2009 Elsevier Ltd. All rights reserved. 1. Introduction Fault diagnosis of fast nonlinear dynamic systems in real-time is a major challenge. Learning approaches to fault detect and diagnosis give a wealth of methodologies to detect, identify, and accommodate faults in nonlinear dynamical systems in real-time. The main idea behind the learning approach is to monitor and approximate abnormal behaviors of dynamical systems using real-time approximation structures such as neural networks or adaptive nonlinear models. When faults occu in a system, its dynamics change and the function modeling the faulty behavior through a real-time approximator can be used as an estimate of the corresponding nonlinear fault function, thereby providing a natural framework for fault detectio identification and accommodation. From a generalviewpoint, learning approaches for fault diagnosis falls in the category of model-based redundancy methods. For a detailed overview of the different approaches for fault detection and diagnosis, and the aspects of analytic redundancy methods see the survey papers of Gertler [1], Isermann [2], and Patton [3]. This paper focuses on fault diagnosis of induction motors. Induction motors are highly nonlinear electrical machines and in most practical and industrial contexts, they are subject to numerous and distinct types of faults and adverse operating conditions. Currently, there is no general theory that provides universal solutions for nonlinear fault diagnosis problems an no methodology is unanimously accepted as the most effective and efficient to solve fault diagnosis of induction motors. ✩ This work was presented at the 2008 World Congress of Nonlinear Analysts, Orlando, USA. ∗ Corresponding address: Faculty of Electrical and Computer Engineering, University of Campinas, Av. Albert Einstein,400,13083-970 Campinas, Sao Paulo, Brazil. Tel.: +55 19 35213706 (Secretary); fax: +55 19 35213845. E-mail addresses: danfl7@dca.fee.unicamp.br, danfl7@yahoo.com.br (D.F. Leite), mbhell@dca.fee.unicamp.br (M.B. Hell), pyramo@pucminas.br (P. Costa Jr.), gomide@dca.fee.unicamp.br (F. Gomide). 0362-546X/$ – see front matter ©2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.na.2009.06.037