Robust Fault Detection of Nonlinear Systems using Local Linear Neuro-Fuzzy Techniques with Application to a Gas turbine Engine Hasan Abbasi Nozari*, Mahdi Aliyari Shooredeli**, Silvio Simani*** *Department of Mechatronics, Faculty of Engineering, Azad University, Science and Research branch, Iran, Tehran (Tel: +989376414624; e-mail: H.Abbasi@ srbiau.ac.ir). **Khaje nasir Toosi University of Technology, Faculty of Electrical Engineering, Iran, Tehran (e-mail:aliyari@eetd.kntu.ac.ir) *** Department of Engineering, University of Ferrara,Via Sargat, IE-44122 Ferrara(FE), Italy (e-mail: silvio.simani@unife.it)} Abstract: This study proposed a model-based robust fault detection (RFD) method using soft computing techniques. Robust detection of the possible incipient faults of an industrial gas turbine engine in steady- state conditions is mainly centered. For residual generation a bank of Multi-Layer perceptron (MLP) models, is used, Moreover, in fault detection phase, a passive approach based on Modellling Error Model (MEM) is employed to achieve robustness and threshold adaptation, and toward this purpose, Local Linear Neuro-Fuzzy (LLNF) model is exploited to construct error model to generate uncertainty interval upon the system output in order to make decision whether or not a fault occurred. This model is trained using the Locally Linear Model Tree (LOLIMOT) algorithm which is an incremental tree-structure algorithm, In order to show the effectiveness of proposed RFD method, it was tested on a single-shaft industrial gas turbine prototype model and has been evaluated using non-linear simulations, based on the gas turbine data. Keywords: fault detection, neural network, gas turbine engine, local linear neuro-fuzzy local linear model tree (LOLIMOT), system identification. 1. INTRODUCTION Nowadays reliability is one of the crucial issues in automatic system design and has received great attention during last two decades. Due to manufacturing defects, erosion-corrosion and tear, and other kind of performance deteriorations in system’s components, and in order to prevent major collapses in plant, system shutdowns, “early” diagnosis of faults is an important factor. Among different fault diagnosis approaches model based methods are still a wide open area of research. In order to make model-based fault detection(FD) algorithms more applicable to real industrial systems, neural networks, fuzzy sets or their combination (neuro-fuzzy) can be considered (Patan et al., 2008). A FD method must be effectively developed to cope with un- wanted and uncontrolled effects such as disturbance, noise, uncertainty of the model, etc. which could dramatically decrease the reliability of fault detection. Robustness could be included in fault diagnosis procedure via active and passive approaches (Patan et al., 2008). Active methods usually leads to define suitable performance index and optimize it with the objective of achieving most sensitivity to fault and most robustness to disturbance, noise, etc. A variety of active robust fault diagnosis methods, of course, with application to linear/linearized systems, are proposed in the literature such as unknown input observer (Chen et al., 1996), robust parity equation (Gertler, 1998), H  (Frank and Ding, 1994), H  (Jaimoukha et al., 2006). The main drawback of above active methods is that they are not applicable in real industrial applications, because some realized hypothesis, which are not possible in practical environment, are taken in to account in enhancing of robustness to fault diagnosis such as: prior knowledge of disturbance and noise acting on the system is always available, and the model of the system is accurate enough to describe the plants dynamics. Another approach for RFD is passive approach that which is usually based on the adaptive threshold computed for the residual by propagation of uncertainty to residual. Passive approaches tackle to RFD problem despite of model uncertainty, and that is the main reason makes passive approach more suitable for experimental applications than active one. are ideas which were proposed in order to drive adaptive threshold for nonlinear systems using soft computing techniques. Fuzzy logic was used to describe threshold changes (Sauter et al., 1993; Schneider, 1993). GMDH neural networks were also used for threshold adaptation by estimating of model uncertainty in order to perform robust fault diagnosis (Patan, 2008). Model error modeling (MEM) can be used as passive approach in RFD. Robust fault diagnosis using MEM was performed successfully on dynamic systems (Patan et al., 2008). In addition to importance of robustness extension to fault detection procedure, FD method must also tackle to detection of incipient faults. Since in industrial applications, it is commonplace for most of the faults to develop slowly over a long period of time, these type of faults are hardly detectable immediately by a simple inspection of output signals, hence, a proposed FDI method must be developed effectively so that Proceedings of the 8th ACD 2010 European Workshop on Advanced Control and Diagnosis Department of Engineering, University of Ferrara, Ferrara, Italy 18-19 November, 2010 Regular Paper 356