Elsevier

Neurocomputing

Volume 91, 15 August 2012, Pages 29-47
Neurocomputing

Model-based robust fault detection and isolation of an industrial gas turbine prototype using soft computing techniques

https://doi.org/10.1016/j.neucom.2012.02.014Get rights and content

Abstract

This study proposes a model-based robust fault detection and isolation (RFDI) method with hybrid structure. Robust detection and isolation of the realistic faults of an industrial gas turbine in steady-state conditions is mainly considered. For residual generation, a bank of time-delay multilayer perceptron (MLP) models is used, and in fault detection step, a passive approach based on model error modelling is employed to achieve threshold adaptation. To do so, local linear neuro-fuzzy (LLNF) modelling is utilised for constructing error-model to generate uncertainty interval upon the system output in order to make decision whether a fault occurred or not. This model is trained using local linear model tree (LOLIMOT) which is a progressive tree-construction algorithm. Simple thresholding is also used along with adaptive thresholding in fault detection phase for comparative purposes. Besides, another MLP neural network is utilised to isolate the faults. In order to show the effectiveness of proposed RFDI method, it was tested on a single-shaft industrial gas turbine prototype model and has been evaluated based on the gas turbine data. A brief comparative study with the related works done on this gas turbine benchmark is also provided to show the pros and cons of the presented RFDI method.

Introduction

Nowadays, reliability has become one of the crucial issues in control system design and received great attention during last two decades. An intelligent diagnostic is one of the essential components of any modern mechatronic system. Due to manufacturing defects, erosion and corrosion, and other kinds of performance deterioration in system components, and in order to prevent major collapses in plant and system shutdowns, “early” diagnosis of faults is an important factor [26]. Model-based fault detection and isolation is a two-phase sequential algorithm: one or several signals so-called “residual” are generated in order to characterize each fault, then in second phase, time and location of the possible faults are determined by analysing of the residual signals.

There are different kinds of models to achieve the model-based idea: the analytical and data-based models [2], [3], [9], [11], [12], [28]. Analytical model-based approaches mainly consider the systems described by linear models. When there are no mathematical models of the system or the complexity of a dynamic system and number of operating points increase, the modelling problem will be very difficult to solve, thus the analytical models cannot be applied or cannot produce satisfactory results.

In order to make fault detection and isolation algorithms more applicable to real industrial systems, neural networks, fuzzy sets or their combination (neuro-fuzzy) can be considered [18].

FDI methods must be developed to cope effectively with uncontrolled effects including disturbance, noise and model uncertainty, etc which could dramatically decrease the reliability of the fault detection. Robustness could be taken into account in a fault diagnosis procedure through active and passive approaches [18]. Active methods usually lead to define a suitable performance index to be optimised with the objective of achieving higher sensitivity to faults and more robustness to disturbance, noise, and so on. A variety of active robust fault diagnosis methods with application to linear systems are proposed in the literature such as unknown input observer [4], robust parity equations [7], H2 [6], H2 in time domain [24], H [5], H [10], and mixed H/H [8]. The main drawback of the active methods is that they are not applicable in real industrial applications, because some realized hypotheses which are not possible in practical environments are taken into account for enhancing the robustness to fault diagnosis including prior knowledge of disturbance and noise acting on the system which are always available, and the model of the system is accurate enough to describe the dynamics of the plant.

Another approach for realizing the RFDI is passive one which is usually based on the adaptive threshold computed for the residual by propagation of uncertainty to the residual. Passive approaches tackle RFDI problem in spite of model uncertainty and that is the main reason which makes the passive approach more appropriate for experimental applications than the active one. There are ideas proposed in order to extract adaptive threshold for nonlinear systems using soft computing techniques. Fuzzy logic was used to describe threshold changes in order to generate adaptive thresholds [22], [23]. Group method of data handling (GMDH) neural networks were also used for threshold adaptation by estimating the model uncertainty in order to perform robust fault diagnosis [18].

In addition to the importance of robustness extension to fault diagnosis procedure, the FDI method must also cope with detection and isolation of incipient faults. Since in industrial applications it is commonplace for most of the faults to develop very slowly over a long period of time, these types of fault are hardly detectable promptly by a simple inspection of output signals. Therefore, a proposed FDI method must be developed to detect such faults efficiently.

General conceptual structure of model-based fault detection and isolation is depicted in Fig. 1. Fault indicative residual(s) is/are obtained by comparing the observed conditions and nominal behaviour of the process. Any deviation of the generated residual(s) from the pre-settled threshold(s) yields fault alarm. Detectable deviations of the residual(s) R yield to analytical symptom(s) S [1] and subsequently suitable decision on the relation between symptoms and faults are made for isolation purposes.

This study is concerned with the use of MLP neural networks applied to FDI in nonlinear dynamic systems on the basis of multiple modelling. Besides, LLNF models trained with LOLIMOT algorithm are used to build up error-model in order to generate adaptive thresholds. Due to computational complexity of the MEM approach, the capability of LOLIMOT algorithm in fast training and evaluating of the LLNF model with any given network topology leads to less computational cost compared with using classical or neural network models. In the present work, the proposed RFDI method is validated with an IGTP model developed at ABB-ALSTOM power, United Kingdom.

The rest of this paper is organised as follows. In Section 2 a brief overview of the gas turbine under consideration and its realistic faulty scenarios are presented. Section 3 introduces the proposed RFDI method briefly. Residual generation via MLP models is described in Section 4. Fault detection using simple thresholding and also MEM-based robust fault detection using LLNF models trained by LOLIMOT algorithm are given in Section 5. How to deal with the fault isolation as a classification problem using MLP neural network is discussed in Section 6. Simulation results obtained by proposed RFDI method are included in Section 7. Moreover, a brief comparison with other proposed FDI methods on the considered gas turbine benchmark is presented in the last part of this section as well. Finally, the main concluding remarks are drawn in Section 8.

Section snippets

System and fauly scenarios description

In our industrial gas turbine of interest, the air flows via an intake duct to the compressor and the high pressure air from the compressor is heated in combustion chambers and expanded through a single stage compressor turbine. A butterfly valve provides a means of generating a back pressure on the compressor turbine (there is no power turbine present in the model). Cooling air is bled from the compressor outlet to cool the turbine stator and rotor. A governor regulates the combustor fuel flow

Proposed robust FDI method

The proposed hybrid RFDI scheme is shown in Fig. 7 and split into two parts: residual and adaptive threshold generation, and decision making for residual evaluation. The detection of faults is implemented by modelling the normal behaviour of monitored industrial gas turbine using MLP networks as nonlinear observers. Hence the residuals are generated by comparing the corresponding predicted and system outputs. In change detection block, detectable deflections of the residuals from their preset

Residual generation via MLP model

The residual signals are generated based on a comparison between the measurements coming from plant full scale simulator and predicted signals given by the MLP models. The residual are thus calculated as follows:Ri(k)=yi(k)yˆi(k)where yi(k) and yˆi(k) are process measurements and predictions, respectively. It is apparent that the residual signals should have near zero behaviour in absence of faults otherwise, meaningful deviation from zero. The fact that residual should only reflect fault

Fault detection

The main objective of fault detection is early detection of faults occurring in the system to allow actions such as reconfiguration, maintenance, repair or other operations. On the other hand, there is always modelling uncertainty due to un-modelled disturbances, noise and so on. Although modelling uncertainty may not be crucial to the process behaviour, but may blur fault detection by increasing false alarms. Hence, in a fault detection problem there is always a compromise between false alarms

Fault isolation

Basically isolation of faults aims at associating each pattern of the residual vector to one of the pre-specified classes of faulty condition. Hence, the decision about which fault has taken place as well as its corresponding location is made in this stage. In order to isolate the faulty patterns of the residuals, the classification capability of a two layer perceptron is used. The topology of an MLP neural network with two hidden layers suitable for fault classification is provided in Fig. 11.

Simulation results

In the case of NN-based system identification, the crucial factor is the number of neurons. Large number of neurons caused complexity in computations and also over parameterisation problem. Thus, a small and reasonable neuron number is desirable. Neuron numbers are determined using the MSE curves. The number of the process outputs determines the number of MLP networks. As four outputs are taken into account for the FDI of the gas turbine under consideration, the relevant number of the MLP

Conclusions

A multiple model-based robust fault detection and isolation approach has been proposed and tested on single-shaft industrial gas turbine working on different operating points. Both approximation and classification capabilities of the MLP neural network are employed to perform an FDI task, and moreover, a lateral algorithm on the basis of LLNF modelling has been also presented to add robustness to fault detection stage. The combination of these soft computing techniques leads to an effective and

Hasan Abbasi Nozari was born in 1985 in Sari, Iran. He received his B.Sc. degree in Computer Engineering in 2007 from the Mazandaran University of Science and Technology, and M.Sc. degree in Mechatronics Engineering in 2010 from the Islamic Azad University Science and Research branch of Tehran. His research interests include neuro-fuzzy systems, neural networks, swarm optimizations and their applications to model based fault detection and isolation of dynamic processes, fault tolerant control

References (28)

  • J. Gertler

    Fault Detection and Diagnosis in Engineering Systems

    (1998)
  • M. Hou, R.J. Patton, 1996. An LMI approach to H−/H∞ fault detection observers, in: Proceedings of UKACC International...
  • R. Iserman

    Fault Diagnosis Systems. An Introduction From Fault Detection to Fault Tolerance

    (2006)
  • J. Korbicz et al.

    Fault Diagnosis. Models, Artificial Intelligence, Applications

    (2004)
  • Cited by (56)

    • Fault detection and isolation of gas turbine using series–parallel NARX model

      2022, ISA Transactions
      Citation Excerpt :

      Due to the features of many real nonlinear dynamic systems, improving model-based fault detection approaches for these systems is important. The existence of disturbance, noise, and uncertainty in actual systems affect the obtained residuals and decrease the accuracy of the model-based fault detection and thus reduce the reliability of these methods [9,10]. This problem has led to the promotion of robustness in fault detection approaches for various nonlinear processes.

    • Nonlinear robust fault diagnosis of power plant gas turbine using Monte Carlo-based adaptive threshold approach

      2020, ISA Transactions
      Citation Excerpt :

      Hence, it can be claimed that the MEM method would only help to increase the accuracy of the nominal model, and robustness against uncertainties and disturbances are the same as a simple threshold method. In 2012, Abbasi et al. have used the MEM method for robust fault detection in industrial gas turbines [29]. In 2015, Rahime and Meskin used adaptive sliding mode observer for sensor faults diagnosis in gas turbines.

    View all citing articles on Scopus

    Hasan Abbasi Nozari was born in 1985 in Sari, Iran. He received his B.Sc. degree in Computer Engineering in 2007 from the Mazandaran University of Science and Technology, and M.Sc. degree in Mechatronics Engineering in 2010 from the Islamic Azad University Science and Research branch of Tehran. His research interests include neuro-fuzzy systems, neural networks, swarm optimizations and their applications to model based fault detection and isolation of dynamic processes, fault tolerant control systems, intelligent control, and system identification.

    Mahdi Aliyari Shoorehdeli received his B.Sc. degree in Electrical Engineering in 2001, M.Sc. degree in Control System Design in 2003, Ph.D. degree in Control System Design in 2008 from K. N. Toosi University of Technology, Tehran, Iran. He is currently an Assistant Professor in the Department of Mechatronics Engineering of the K. N. Toosi University of Technology. His research interests include Fault Detection and Identification of Mechatronics Systems, Mechatronics Systems Identification, Intelligent Systems, Fuzzy, Neural Networks, Multi-Objective Optimisation, Hybrid Control based on Classic and Intelligent systems.

    Silvio Simani was born in Ferrara on April 21st, 1971. He received the “Laurea” degree (cum laude) in Electrical Engineering in June, 1996 from the Department of Engineering at the Università degli Studi di Ferrara (Italy). In February, 2000 he was awarded the Ph.D. in “Information Science: Automatic Control” at the Department of Engineering of the University of Ferrara and Modena (Italy). Since 1999 he has been an IEEE Valued Member and Research Associate at the Dipartimento di Ingegneria of the Università di Ferrara. Since July 2000 he is a member of the Technical Committee SAFEPROCESS. Since February 2002 he is Assistant Professor at the Department of Engineering of the University of Ferrara. He received the nomination of IEEE Senior Member in December 2006. His research interests include fault diagnosis of dynamic processes, system modelling and identification, and the interaction issues between identification and fault diagnosis. He is a reviewer of many international journals and author of more than 80 international journal and conference papers, as well as one book on these topics.

    Hamed Dehghan Banadaki received his B.Sc. degree in Computer Engineering from Islamic Azad university of Meybod, and M.Sc. degree in Mechatronics Engineering from Islamic Azad University, Science and Research branch of Tehran, Iran. His main research interests are identification and modelling, robust fault detection and diagnosis, industrial control, intelligent control, intelligent systems and multi variable optimisation.

    View full text