Model-based fault detection and isolation of a steam generator using neuro-fuzzy networks Roozbeh Razavi-Far a,Ã , Hadi Davilu a , Vasile Palade b , Caro Lucas c a Department of Nuclear Engineering, Amirkabir University of Technology, P.O. Box 15875-4413 Tehran, Iran b Computing Laboratory, Oxford University, Wolfson Building, Parks Road, Oxford OX13QD, UK c Center of Excellence on Control and Intelligent Processing, Department of Electrical and Computer Engineering, University of Tehran, P.O. Box 14395-1515 Tehran, Iran article info Article history: Received 8 June 2008 Received in revised form 23 April 2009 Accepted 26 April 2009 Communicated by T. Heskes Available online 10 May 2009 Keywords: Fault detection Fault isolation Neuro fuzzy networks Locally linear neuro fuzzy model Locally linear model tree (LOLIMOT) algorithm Steam generator abstract This paper presents a neuro-fuzzy (NF) networks based scheme for fault detection and isolation (FDI) of a U-tube steam generator (UTSG) in a nuclear power plant. Two types of NF networks are used. A NF based learning and adaptation of Takagi–Sugeno (TS) fuzzy models is used for residual generation, while for residual evaluation a NF network for Mamdani models is used. The NF network for Takagi–Sugeno models is trained with data collected from a full scale UTSG simulator and is used for generating residuals in the fault detection step. A locally linear neuro-fuzzy (LLNF) model is used in the identification of the steam generator. This model is trained using the locally linear model tree (LOLIMOT) algorithm. In the fault isolation part, genetic algorithms are employed to train a Mamdani type NF network, which is used to classify the residuals and take the appropriate decision regarding the actual behavior of the process. Furthermore, a qualitative description of faults is then extracted from the fuzzy rules obtained from the Mamdani NF network. Experimental results presented in the final part of the paper confirm the effectiveness of this approach. & 2009 Elsevier B.V. All rights reserved. 1. Introduction Large-scale systems, such as nuclear power plants (NPPs), are increasingly required to provide safe and reliable operation for long periods of time. Unfortunately, system components are subject to manufacturing defects, interactions with the environ- ment, wear and tear, and other causes of performance degrada- tions. For these safety-critical systems, the problem of detecting the occurrence of faults is of paramount importance due to their disastrous consequences. An early detection of faults can prevent major plant breakdowns, system shutdown or cata- strophes involving human fatalities and material damage, as well as maintain the optimal functioning of the system. ‘‘A fault represents an unpermitted deviation of at least one characteristic property or parameter of the system from the acceptable, usual or standard condition, tending to degrade overall system perfor- mance’’ [23]. ‘‘A fault diagnosis system is an operator decision support system that is used to detect faults and diagnose their location and significance in a system’’ [4]. Fault diagnosis can be performed by means of a three step algorithm. Firstly, one or several signals reflecting faults in the process behavior are generated. These signals are called residuals. In the second step, the residuals are evaluated. A decision is made using these residuals, in order to determine the time and the location of possible faults. Finally, the nature and likely cause of the faults are analyzed by the relations between the symptoms and their physical causes. For the latter stage of fault diagnosis, classification or inference methods, including fault-symptom trees, fuzzy rules or neural approaches can be used [12]. Classification approaches to fault diagnosis were presented in many previous works. Previous methods to identify Nuclear Power Plant transients were based on time-series data of various transient signals. Various diagnostic methods were worked out [6,26]. However, such techniques need a large amount of fault data to extract the features of individual faults, or expert knowledge about the system and its misbehavior. Most common methods of fault detection are using mathema- tical and signal processing models. Signal processing [9] or feature-based approaches [7] are suitable for fault detection. These approaches usually avoid system modeling. Signals may be studied either by using time-domain or frequency-domain methods or with more sophisticated methods like time-frequency or wavelet analysis [15]. The difficulty common to all these approaches is how to ensure that a change in some value is due to a particular fault. Contemporary diagnostics of processes is mainly based on process models and, in order to detect the occurrence of a fault, a model of the normal process behavior is needed [2]. Fig. 1 shows the general structure of model-based fault detection and ARTICLE IN PRESS Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/neucom Neurocomputing 0925-2312/$ - see front matter & 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.neucom.2009.04.004 Ã Corresponding author. E-mail address: razavi_roozbeh@cic.aut.ac.ir (R. Razavi-Far). Neurocomputing 72 (2009) 2939–2951