Elsevier

Expert Systems with Applications

Volume 84, 30 October 2017, Pages 200-219
Expert Systems with Applications

Fault Detection and Diagnosis in dynamic systems using Weightless Neural Networks

https://doi.org/10.1016/j.eswa.2017.05.020Get rights and content

Highlights

  • Weightless Neural Networks for Fault Detection and Diagnosis in dynamic systems.

  • Several methods for data pre-processing (input patterns mapping) are evaluated.

  • Attributes selection per class to pre-processing multivariate data.

  • The importance of counting and bleaching in classification is shown.

  • Successful Fault Detection and Diagnosis systems with accuracies around 99%.

Abstract

This work examines Fault Detection and Diagnosis (FDD) based on Weightless Neural Networks (WNN) with applications in univariate and multivariate dynamic systems. WNN use neurons based on RAM (Random Access Memory) devices. These networks use fast and flexible learning algorithms, which provide accurate and consistent results, without the need for residual generation or network retraining, and therefore they have great potential use for pattern recognition and classification (Ludermir, Carvalho, Braga, de Souto, 1999). The proposed system firstly executes the selection of attributes (in the multivariable case) and does the time series mapping of the data. In the intermediate stage, the WNN performs the detection and diagnosis per class. The network outputs are then passed through a clustering filter in the final stage of the system, if a diagnosis per fault groups is necessary. The system was tested with two case studies: one was an actual application for the temperature monitoring of a sales gas compressor in a natural gas processing unit; and the other one uses simulated data for an industrial plant, known in the literature as “Tennessee Eastman Process”. The results show the efficiency of the proposed systems for FDD with classification accuracies of up to 98.78% and 99.47% for the respective applications.

Introduction

Both current technological advances and the increasing demand for more productive and reliable industrial processes have resulted in more complex automation with greater data availability. As a result, there is an increasing need for more efficient supervision and control systems, especially for Fault Detection and Diagnosis (FDD) in dynamic environments. In a production system, a fault is an abnormal operating condition caused by factors such as design errors, installation errors, misuse or the effects of natural degradation. The availability of mechanisms for early and reliable detection of faults decreases the risks of malfunction or unscheduled shutdowns of the system. Consequently, it increases equipment reliability and avoids material losses, environmental accidents and harm to workers (Blázquez and Miguel, 2005, Chiang et al., 2001, Romano and Kinnaert, 2006, Yang and Liu, 1998).

Considering the a priori knowledge used for their development, FDD systems are designed using process models or historical data. Such approaches can be subdivided into quantitative and qualitative methods (Venkatasubramanian et al., 2003, Venkatasubramanian et al., 2003). Systems based on historical data can extract relevant features from the data in order to map the relationships and the existing limits between the considered classes or groups. In this group, some quantitative methods can be highlighted: statistical classifiers (Luo et al., 2011, Ma et al., 2010, Soares and Galvão, 2010); neural networks (Lau et al., 2013, Leite et al., 2012, Sartori, 2012, Zarei, 2012); fuzzy logic (Andrade, 2012, Andrade et al., 2011, Lau et al., 2013, Leite et al., 2012, Li et al., 2013, Silva et al., 2012); Principal Component Analysis (PCA) (Barragan et al., 2016, Jiang et al., 2013, Lau et al., 2013); method of partial least squares (Zhang, Zhou, Qin, & Chai, 2010); wavelet transforms (Barragan et al., 2016). Among the qualitative methods, expert systems (Saravanan et al., 2009, Wang et al., 2012, Zadeh, 2008) and qualitative trend analysis (Maurya, Rengaswamy, & Venkatasubramanian, 2007) are noteworthy. In the model-based approaches, the actual behavior of the monitored system is compared to the response obtained by a representative model of the process. The result of this supervised comparison is a residual vector used to detect the presence of faults. In this group the following quantitative methods can be highlighted: state and output observers (Chetouani, 2008, Kalman, 1960); parity space and equations (Beckerle et al., 2012, Blesa et al., 2014, Zakharov et al., 2013, Zhong et al., 2015); extended Kalman filter (Kalman, 1960, Patwardhan and Shah, 2006); support vector machine (Zhang, Zhou, Guo, Zou, & Huang, 2012; Deng et al., 2011, Duan et al., 2016, Park et al., 2011); and parameter identification and estimation methods (Johansson et al., 2006, Pouliezos et al., 1989). In the qualitative approach, the following methods stand out: fault trees (Nguyen and Lee, 2008, Simões Filho, 2006); qualitative simulation (Berleant, 1991); qualitative process theory (Venkatasubramanian, Rengaswamy, & Kavuri, 2003); and Bayesian networks and other Bayesian reasoning extensions, such as signed directed graphs and evidence theory (Ji et al., 2015, Luo et al., 2012, Xiao et al., 2014). Serdio et al., 2014, Serdio et al., 2015) have developed approaches that combine historical data, automatic model extraction and residual generation, which seem useful for fault detection and isolation with abrupt and incipient faults.

Many studies have shown the positive advantages of statistically based and artificial intelligence based techniques for solving FDD problems, especially hybrid systems involving fuzzy logic, genetic algorithms and mainly neural networks (Chen and Chen, 2011, Ghate and Dudul, 2010, Mendonça et al., 2009, Niaki and Abbasi, 2005, Sartori et al., 2012, Sharma et al., 2015, White and Lakany, 2008, Wu, 2011, Zarei, 2012). The noise tolerance shown by artificial neural networks (Özyurt & Kandel, 1996) and the ability of fuzzy logic to deal with information inaccuracies, ambiguities and uncertainties (Leite et al., 2012) mean that neuro-fuzzy hybrid systems are widely used (Hell et al., 2008, Lau et al., 2013). These techniques are well suited to non-linear systems because they do not require explicit mathematical models (Angelov and Yager, 2012, Bartyś et al., 2006, Bocaniala and da Costa, 2006, Li et al., 2013, Lo et al., 2009, Ma et al., 2010, Rigatos and Zhang, 2009). According to Sartori, Amaro, Arduini, Souza Júnior, and Embiruçu (2016), most applications of FDD systems are in power generating and distribution units, pieces of equipment in process industries (reactors, columns, sensors and actuators) and motors and bearings. This study lists neural networks, fuzzy logic, principal component analysis, Kalman filter, support vector machines, genetic algorithms and expert systems as the most commonly used techniques in descending order.

In general, such approaches are well suited to the detection and diagnosis of abrupt faults. In the context of incipient faults in dynamical systems, the FDD problem is more complex and the studies and research found in the literature have yet to find appropriate solutions for problems such as: multivariate problems and a diversity in the number of classes; nonlinear processes with incipient faults present in two or more considered classes; difficulty in obtaining historical data with real-world applications. The main contribution of this paper is to propose a detection and diagnosis system for incipient faults in dynamic systems, without the need to use a mathematical model or residual calculations and with a low false alarm rate. The proposed system is based on Weightless Neural Networks [WNN, an acronysm also used for Wavelet Neural Networks (Lei, He, & Zi, 2011)], initially proposed by Aleksander (1967). WNN are digital models based on Random Access Memory (RAM) devices. Unlike conventional neural models, learning happens in memories inserted into the neuron, in the form of truth tables. Compared to weighted models, WNN have the advantages of a diversity of use of these memories, such as similarity with the conventional digital systems; fast and flexible learning algorithms; accuracy and consistency in the results, no need for generating residuals and network retraining; and above all, great potential for pattern recognition and classification.

The applications of WNN for pattern recognition and classification problems can be found in several areas, including: digit and fingerprint recognition (Bandeira et al., 2009, Conti et al., 2009, Grieco et al., 2010); faces and facial features recognition (Araújo, 2011, Sirlantzis et al., 2009, Subhashini and Nagarajan, 2014); robot navigation (McElroy and Howells, 2011, Nurmaini et al., 2009); data stream clustering (Cardoso, Lima, de Gregório, Gama, & França, 2011); and time series forecasting (De Souza et al., 2010, Mpofu, 2006). However, no work addressing the problems of FDD with the use of WNN was found in the literature. The papers presented by De Gregorio and Giordano (2014) and Cardoso, De Gregorio, Lima, Gama, and França (2012) approach this context, but do not deal with FDD problems. De Gregorio and Giordano (2014) used WNN for the problem of detecting changes in the vision field of a camera. The proposed system, called CwisarDH, uses a discriminator for each coverage point of the video with the color concept. Cardoso et al. (2012) presented the StreamWiSARD system for flow data grouping with sliding-window. The system consists of WiSARD discriminators as primary units and is able to define high quality clusters, restricted to a small number of microclusters. The absence of FDD systems based on WNN, and the potential of these networks for pattern recognition and classification, are the main factors that motivated this work.

In order to validate the developed systems two case studies were considered. The first one uses a univariate database of temperature sensors at a Natural Gas Processing Unit (NGPU) sales gas compressor from Petrobras, located in Pojuca, Bahia, Brazil (Andrade, 2012, Andrade et al., 2011, Sartori, 2012). In the second case study, multivariate data from an industrial plant simulator considered a benchmark in Fault Detection and Diagnosis problems, known as Tennessee Eastman Process (TEP) was used (Downs and Vogel, 1993, Ricker, 1995). In the multivariate application an algorithm of attributes selection per class proposed by Vale, Neto, and Canuto (2010) was used, with modifications proposed in this paper.

This paper is organized as follows: Section 2 presents the fundamentals of WNN; Section 3 describes the FDD-WiSARD (Wilkes-Stonham-Aleksander Recognition Device) system and the method used for training and testing; and in Section 4 the two case studies are used and their respective results are presented. Finally, Section 5 presents concluding remarks.

Section snippets

Weightless Neural Networks

Aleksander (1967), one of the pioneers in the field of ANN (Artificial Neural Networks), proposed an entirely digital neuron/node model based on RAM devices. The model gave rise to a class of neurons based on Boolean logic, also known as weightless neurons or RAM-based neurons. Neural networks made up of these neurons are called Weightless Neural Networks (WNN). The learning in these networks consists of changes in the memory contents in truth tables inserted in the RAM. A weightless node

System structure

The system proposed in this work performs fault detection as a classification problem in which an input pattern x = [a1, a2, …, an], defined as a n-dimensional vector where ai are the measured variables (attributes), is considered a normal operation or fault state. The diagnosis, on the other hand, consists of the partition of the fault state into subclass labels defined by the different types of faults considered. For this, the attributes a1, a2, …, an should contain information that enables the

Natural gas processing unit

For this case study a univariate database with temperature measurements obtained from an industrial plant was used. Temperature sensors are coupled to a Sales Gas Compressor (SGC) in a Natural Gas Processing Unit (NGPU) at Petrobras, located in the town of Pojuca, Bahia state, Brazil (Andrade, 2012, Andrade et al., 2011, Sartori, 2012). This compressor works with a minimum load of 50% and its functional structure includes an electric motor, a crankcase, five compression cylinders, two intake

Conclusions

This work addresses the problem of Fault Detection and Diagnosis in dynamic systems with the use of weightless neural networks. The DW network was used as reference for the development of the proposed systems and these systems were tested and validated with univariate data, representing a real problem (historical data from a NGPU, focusing on the temperature monitoring of its sales gas compressor), and with simulated multivariate data from the so-called TEP benchmark problem. With the results

Acknowledgments

The authors acknowledge CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior), CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico) and FAPESB (Fundação de Amparo à Pesquisa do Estado da Bahia) for their financial support, and thank Jaime Freire de Souza for his collaboration in the development of the implemented software systems.

References (94)

  • T. Fawcett

    An Introduction to ROC Analysis

    Pattern Recognition Letters

    (2006)
  • V.N. Ghate et al.

    Optimal MLP neural network classifier for fault detection of three phase induction motor

    Expert Systems with Applications

    (2010)
  • B.P. Grieco et al.

    Producing pattern examples from mental images

    Neurocomputing

    (2010)
  • A. Johansson et al.

    Dynamic threshold generators for robust fault detection in linear systems with parameter uncertainty

    Automatica

    (2006)
  • C.K. Lau et al.

    Fault diagnosis of Tennessee Eastman process with multi-scale PCA and ANFIS

    Chemometrics and Intelligent Laboratory Systems

    (2013)
  • Y. Lei et al.

    EEMD method and WNN for fault diagnosis of locomotive roller bearings

    Expert Systems with Applications

    (2011)
  • H. Luo et al.

    A SVDD approach of fuzzy classification for analog circuit fault diagnosis with FWT as preprocessor

    Expert Systems with Applications

    (2011)
  • H. Luo et al.

    Agent oriented intelligent fault diagnosis system using evidence theory

    Expert Systems with Applications

    (2012)
  • M.R. Maurya et al.

    Fault diagnosis using dynamic trend analysis: A review and recent developments

    Engineering Applications of Artificial Intelligence

    (2007)
  • L.F. Mendonça et al.

    An architecture for fault detection and isolation based on fuzzy methods

    Expert Systems with Applications

    (2009)
  • B. Özyurt et al.

    A hybrid hierarchical neural network-fuzzy expert system approach to chemical process fault diagnosis

    Fuzzy Sets and Systems

    (1996)
  • J. Park et al.

    Spline regression based feature extraction for semiconductor process fault detection using support vector machine

    Expert Systems with Applications

    (2011)
  • S.C. Patwardhan et al.

    From data to diagnosis and control using generalized orthonormal basis filters

    Journal of Process Control

    (2006)
  • N.L. Ricker

    Optimal steady-state operation of the Tennessee Eastman challenge process

    Computers and Chemical Engineering

    (1995)
  • G. Rigatos et al.

    Fuzzy model validation using the local statistical approach

    Fuzzy Sets and Systems

    (2009)
  • N. Saravanan et al.

    Vibration-based fault diagnosis of spur bevel gear box using fuzzy technique

    Expert Systems with Applications

    (2009)
  • F. Serdio et al.

    Residual-based fault detection using soft computing techniques for condition monitoring at rolling mills

    Information Sciences

    (2014)
  • F. Serdio et al.

    Fuzzy fault isolation using gradient information and quality criteria from system identification models

    Information Sciences

    (2015)
  • S.V. Stehman

    Selecting and interpreting measures of thematic classification accuracy

    Remote Sensing of Environment

    (1997)
  • V. Venkatasubramanian et al.

    A review of process fault detection and diagnosis: Part II: Qualitative models and search strategies

    Computers and Chemical Engineering

    (2003)
  • V. Venkatasubramanian et al.

    A review of process fault detection and diagnosis: Part I: Quantitative model-based methods

    Computers and Chemical Engineering

    (2003)
  • W. Wang et al.

    An evolving neuro-fuzzy technique for system state forecasting

    Neurocomputing

    (2012)
  • C.J. White et al.

    A fuzzy inference system for fault detection and isolation: Application to a fluid system

    Expert Systems with Applications

    (2008)
  • Q. Wu

    Car assembly line fault diagnosis model based on triangular fuzzy Gaussian wavelet kernel support vector classifier machine and genetic algorithm

    Expert Systems with Applications

    (2011)
  • F. Xiao et al.

    Bayesian network based FDD strategy for variable air volume terminals

    Automation in Construction

    (2014)
  • L.A. Zadeh

    Is there a need for fuzzy logic?

    Information Sciences

    (2008)
  • A. Zakharov et al.

    Fault detection and diagnosis approach based on nonlinear parity equations and its application to leakages and blockages in the drying section of a board machine

    Journal of Process Control

    (2013)
  • J. Zarei

    Induction motors bearing fault detection using pattern recognition techniques

    Expert Systems with Applications

    (2012)
  • X. Zhang et al.

    Vibrant fault diagnosis for hydroelectric generator units with a new combination of rough sets and support vector machine

    Expert Systems with Applications

    (2012)
  • M. Zhong et al.

    Parity space-based fault detection for linear discrete time-varying systems with unknown input

    Automatica

    (2015)
  • I. Aleksander

    Adaptive systems of logic networks and binary memories

  • I. Aleksander et al.

    WiSARD: A radical step forward in image recognition

    Sensor Review

    (1984)
  • V.E. Andrade

    Sistema de detecção e diagnóstico de falhas em sensores de um compressor de gás natural utilizando lógica fuzzy tipo-2 (in Portuguese)

    (2012)
  • V.E. Andrade et al.

    An interval type-2 fuzzy logic approach for instrument fault detection and diagnosis

    Proceedings of the World Congress on Engineering

    (2011)
  • P. Angelov et al.

    A new type of simplified fuzzy rule-based system

    International Journal of General Systems

    (2012)
  • L.A. Araújo

    RWiSARDUm modelo de rede neural sem peso para reconhecimento e classificação de imagens em escala de cinza (in Portuguese)

    (2011)
  • L.C. Bandeira et al.

    NC-WiSARD: Uma interpretação Booleana da arquitetura neocognitron (in Portuguese)

    Anais do IX congresso brasileiro de redes neurais e inteligência computacional

    (2009)
  • Cited by (42)

    • A multigroup fault detection and diagnosis framework for large-scale industrial systems using nonlinear multivariate analysis

      2022, Expert Systems with Applications
      Citation Excerpt :

      This drawback makes it unsuitable for nonlinear systems. Some fault detection methods have been developed to cope with nonlinear properties of industrial systems (Harkat, Mansouri, Nounou, & Nounou, 2019; José, Oliveira, Pontes, & Marcelo, 2017; Pilario, Shafiee, Cao, Lao, & Yang, 2020). Most of them used the kernel-based multivariate analysis methods for data modeling (Deng, Tian, Chen, & Harris, 2017; Fezai, Mansouri, Taouali, Harkat, & Bouguila, 2018), such as kernel PCA and kernel CCA.

    • Nonlinear process monitoring based on decentralized generalized regression neural networks

      2020, Expert Systems with Applications
      Citation Excerpt :

      Fortunately, the calculations involved in using the m GRNN models can be implemented in parallel, and the required computational time does not increase significantly in real applications. To examine the effectiveness and superiority of the proposed method, the methodology was applied to two nonlinear processes: a numerical nonlinear example and the Tennessee Eastman (TE) benchmark process (Downs & Vogel 1993; José, Oliveira, Pontes, & Marcelo, 2017). Comparisons with other state-of-the-art nonlinear process monitoring techniques, such as the KPCA, EKPCA, CPM, NCA, and AAELM, are considered.

    • Iterated feature selection algorithms with layered recurrent neural network for software fault prediction

      2019, Expert Systems with Applications
      Citation Excerpt :

      Machine learning (ML) is one of the data-driven soft computing techniques that has been widely used to predict the faulty modules in software projects. Representative ML methods include Logistic Regression (Li & Henry, 1993), Support Vector Machine (Xing, Guo, & Lyu, 2005), Naive Bayes (Alshayeb & Li, 2003), Neural Network (Oliveira, Pontes, Sartori, & Embiruçu, 2017), and Ensemble classifiers (Rathore & Kumar, 2017b; 2017c). Malhotra in Malhotra (2015) conducted a systematic literature review on several ML techniques that were used in SFP.

    • Fault detection and diagnosis based on particle filters combined with interactive multiple-model estimation in dynamic process systems

      2019, ISA Transactions
      Citation Excerpt :

      Many different FDD approaches have been developed in literature, such as state and parameter estimation [1,2], hybrid diagnosis [3], ensemble empirical mode decomposition [4], soft computing techniques and multiple-model approaches [5–8]. The machine learning methods such as fuzzy systems and neural network are widely used in FDD [9–17]. Many studies have shown the positive advantages of machine learning based and artificial intelligence based techniques for solving FDD problems, especially hybrid systems involving fuzzy logic, genetic algorithms and mainly neural networks.

    View all citing articles on Scopus
    View full text