Elsevier

Neurocomputing

Volume 74, Issue 18, November 2011, Pages 3865-3876
Neurocomputing

Self adaptive growing neural network classifier for faults detection and diagnosis

https://doi.org/10.1016/j.neucom.2011.08.001Get rights and content

Abstract

Fault detection and diagnosis have gained widespread industrial interest in machine monitoring due to their potential advantage that results from reducing maintenance costs, improving productivity and increasing machine availability. This article develops an adaptive intelligent technique based on artificial neural networks combined with advanced signal processing methods for systematic detection and diagnosis of faults in industrial systems based on a classification method. It uses discrete wavelet transform and training techniques based on locating and adjusting the Gaussian neurons in activation zones of training data. The learning (1) provides minimization in the number of neurons depending on cost error function and other stopping criterions; (2) offers rapid training and testing processes; (3) provides accuracy in classification as confirmed by the results on real signals. The method is applied to classify mechanical faults of rotary elements and to detect and isolate disturbances for a chemical process. Obtained results are analyzed, explained and compared with various methods that have been widely investigated for fault diagnosis.

Introduction

The reliability, safety and availability of industrial plants play an important role during their operational use because industrial installations and control algorithms are becoming more and more sophisticated. In the case of simple technical systems, human inspection was enough but the increased complexity of industrial systems and the high level of process quality, reliability and safety requirements force the automation of diagnosis in order to make it possible to determine the place, reason and time of the fault precisely [1], [2].

A fault can be defined as a non-permitted deviation of a characteristic property of the process, which will cause a certain level of deterioration in the performance of the process. The deviation can be caused by temporary or permanent physical changes in the system [3]. Obviously, it is desirable to detect and diagnose faults as soon as possible after their occurrence. Therefore designing an intelligent real time system for Fault Detection and Diagnosis (FDD) is receiving considerable attention both from industry and academia. Artificial Neural Networks (ANNs) can be used for that purpose [4–6]. ANNs have the ability to approximate non-linear relations and to determine flexible decision regions in both continuous and discrete forms. Their learning and interpolation capabilities have led to several successful implementations over various processes. ANNs have been shown to be particularly effective in handling some complexities commonly found in data [7], [8]. In addition, they can provide quick detection and isolation, which refers to the ability to discriminate between different failures [9]. However, neural networks can give rise to models that can over fit to the training samples. In addition if network expands or topology changes (for example, if a new type of faults appears), ANN has to be totally retrained [10], [11].

The aim of this article is to develop an ANN based classifier ready to use for fault detection and diagnosis starting from temporal signals measurement and processing. The main contribution is to propose an adaptation algorithm that is able to add nodes and to change the network structure according to supervised learning. The main idea of our contribution is initiated from this background. More precisely, how can we build a network classifier with nodes to be activated in specific subspaces for the different classes of fault? In this article, we introduce Self Adaptive Growing Neural Network (SAGNN) to respond this question. SAGNN is based on automatic structure building and parameter tuning procedure. The concept behind this modular non-parametric classifier is to divide the pattern space of faults in a hierarchical way into a number of smaller sub-spaces depending on the activation zones of clustered parameters. For each type of faults, in a particular sub-space, a special diagnosis agent is trained. Then in order to improve the performance of the diagnosis system, SAGNN is combined with Discrete Wavelet Transform (DWT) that is helpful to extract significant features in the measured signals. Finally, SAGNN act as a classifier that maps each sample with a class of fault.

The paper is organized as follows: Section 2 introduces FDD for industrial systems and presents the principle of SAGNN. Features extraction using DWT is presented in Section 3. Section 4 details SAGNN classifier. The applications to the fault classification of a real mechanical system and to the disturbance detection and diagnosis of the Tennessee Eastman Chemical Process (TECP) simulator are presented in Section 5. Moreover different ANNs classification methods are displayed and performances are compared. Some conclusions are drawn in the last section concerning the advantages and limitations of the proposed approach.

Section snippets

Context

Most manufacturing processes involve many correlated variables. A quick and correct fault diagnosis system helps to avoid product quality problems and facilitates preventive maintenance [12]. The fault diagnosis system should perform two tasks, namely fault detection and fault diagnosis [13]. The purpose of the former is to determine that a fault has occurred in the system. To achieve this goal, all the available information from the system should be collected and processed to detect any change

Wavelet decomposition and feature extraction

In any fault diagnosis problems, feature extraction from the raw signals is a very critical step. Because these extracted features cannot only characterize the information relevant to machine conditions, but also affects the final diagnosis results. In our previous work [41], we have shown that features extracted from time- frequency domain increase the classification accuracy. For this reason, DWT is used to generate features for classification.

Self adaptive growing neural network

This section proposes a self adaptive growing neural network (SAGNN) classification technique for detecting, classifying and diagnosing faults or disturbances in industrial machines [47]. It is composed of three main stages:

  • 1-

    The features extraction stage, which consists of different parameters extraction from measured signals and selection of the more significant ones.

  • 2-

    The learning stage that adapts the parameters of the neurons and eventually create additive nodes in several hidden layers, each

Applications

The proposed method is applied on a mechanical system and on the Tennessee Eastman Challenge Process (TECP) that is a simulator of chemical reactor. The classification performances are first evaluated and compared with other classification techniques. DWT is used in the preprocessing stage of feature extraction because in our previous work [54], we have shown that classification approaches are improved in time-frequency domain using wavelet transform. Then detection and diagnosis issues are

Conclusion

In this paper, a novel approach to pattern classification is propounded for tackling fault diagnosis tasks. The non-parametric SAGNN classifier based on the concept of Neural Network with supervised training process is put forward for fault detection and diagnosis purposes. SAGNN classifier optimizes the number of nodes, minimizes the processing time for calculation and generates hidden sub-spaces for every output class. The combination between DWT and SAGNN improves the performances of the

Acknowledgements

The authors thank the CEDRE, the CNRS in Lebanon and the Region Haute-Normandie in France for their support.

Mustapha Barakat was born in Rob-Talaten, Lebanon, in 1985. He received the master (I) degree in general physics from the faculty of sciences of Lebanese University in 2007. He obtained master (II) degree in MMM (Mechatronics, Materials and Modelization) from the faculty of engineering of Lebanese University incorporation with Versailles University (France) in 2008. He is currently a third year Ph.D. in automatic and signal processing in doctoral school of Lebanese University and Le Havre

References (55)

  • M.R. Dellomo

    Helicopter gearbox fault detection: a neural network based approach

    J. Vib. Acoust.

    (1999)
  • L.B. Jack, Applications of Artificial Intelligence in Machine Condition Monitoring, Ph.D. Thesis, Department of...
  • V. Palade et al.

    Computational Intelligence in Fault Diagnosis

    (2006)
  • V. Venkatasubramanian et al.

    A neural network methodology for process fault diagnosis

    AIChE J.

    (1989)
  • V. Venkatasubramanian, G.M. Stanley, Integration of process monitoring, diagnosis and control: issues and emerging...
  • A. Arthur

    Signal Processing Applications of Wavelets

    (2003)
  • K. Watanabe et al.

    Diagnosis of multiple simultaneous fault via hierarchical artificial neural networks

    AIChE J.

    (1994)
  • R. Eslamloueyan et al.

    Multiple simultaneous fault diagnosis via hierarchical and single artificial networks

    Sci. Iranica: Trans. Chem. Environ. Eng.

    (2003)
  • G. Ge et al.

    Fault diagnosis using support vector machine with an application in sheet metal stamping operations

    Mech. Syst. Signal Process.

    (2004)
  • S. Simani et al.

    Fault diagnosis in power plant using neural network

    Inf. Sci.

    (2000)
  • M. Basseville, Model-based statistical signal processing and decision theoretic approaches to monitoring, in:...
  • M.J. Corinthios

    A fast Fourier transform for high-speed signal processing

    IEEE Trans. Comput.

    (1971)
  • M. Portnoff

    Time-frequency representation of digital signals and system based on short-time Fourier analysis

    IEEE Trans. Acoust. Speech Signal Process. ASSP

    (1980)
  • S. Lesecq, A. Barraud, Arcing fault detection using wavelet transform, In: Proceedings of the 5th IFAC Symposium on...
  • Z.L. Gaing

    Wavelet-based neural network for power disturbance recognition and classification

    IEEE Trans. Power Delivery

    (2004)
  • Y.J. Kim, J.M. Kim, S.H. Lee, T.U. Jung, Detection of shorted-turns in the rotor winding of cylindrical synchronous...
  • W.A. Wilkinson et al.

    Discrete wavelet analysis of power system transients

    IEEE Trans. Power Delivery

    (1996)
  • Cited by (81)

    View all citing articles on Scopus

    Mustapha Barakat was born in Rob-Talaten, Lebanon, in 1985. He received the master (I) degree in general physics from the faculty of sciences of Lebanese University in 2007. He obtained master (II) degree in MMM (Mechatronics, Materials and Modelization) from the faculty of engineering of Lebanese University incorporation with Versailles University (France) in 2008. He is currently a third year Ph.D. in automatic and signal processing in doctoral school of Lebanese University and Le Havre University (France). His research interests include fault detection, fault diagnosis, neural networks, signal processing and classification approaches.

    Fabrice Druaux received the B.S. degree in physics and mathematics in 1976 the M.S. in physics in 1981 and the Ph.D. degree in physics from University of Rouen (France) in 1986. Since 1988 he is a Lecturer with the department of Electronic, Electro-technology and Automatic at the Faculty of Sciences and Technology of Le Havre (France). Between 1988 and 1999 he has been with the L.A.CO.S. (Laboratoire d'Analyse et de COmmande des Systèmes) working on dynamical neural network for pattern recognition and classification. Since 1999 he has been with the G.R.E.A.H. (Electric and Automatic Engineering Research Group). His current research interests include modelling, control and fault detection using dynamical neural network. The principal applications are electro-technical processes such as motors and wind generators.

    Dimitri Lefebvre is graduated from the Ecole Centrale of Lille (France) in 1992. He received the the Ph.D. degree in Automatic Control and Computer Science from University of Sciences and Technologies, Lille in 1994, and the HAB. degree from University of Franche Comté , Belfort, France in 2000. Since 2001 he is Professor at Institut of Technology and Faculty of Sciences, University Le Havre, France. He is with the G.R.E.A.H. (Electric and Automatic Engineering Research Group). His current research interests include learning processes, adaptive control, fault detection and diagnosis and applications to electrical engineering.

    Mohamad Khalil was born in Akkar Atika, in Lebanon in 1973. He obtained an engineering degree in electrical and electricity from the Lebanese University, faculty of engineering, Tripoli, Lebanon in 1995. He received the DEA in biomedical engineering from the University of Technology of Compiegne (UTC) in France in 1996. He received his Ph.D. from the University of Technology of Troyes in France in 1999. He received his HDR (Habilitation a diriger des recherches) from UTC in 2006. He is currently researcher at Lebanese University. He is director of the Azm center for research in biotechnology at the doctoral school of sciences and technology at the Lebanese university. His current interests are the signal and image processing problems: detection, classification, analysis, representation and modeling of non stationary signals, with application to biomedical signals and images.

    Oussama Mustapha was born in Houla, in Lebanon in 1965. He obtained an engineering degree in electrical and telecommunication from the Beirut Arab University, faculty of engineering, Beirut, Lebanon in 1990. He received the DEA in Modeling and simulation engineering from the Agence Universitaire de la Francophonie (AUF) in Lebanon in 2003. He received his Ph.D. from the University of Le Havre in France in 2008. He is currently researcher at Islamic University. His current interests are the signal processing problems: detection, classification, analysis, representation and modeling of non stationary signals, with application to biomedical signals.

    View full text