Monitoring Approach Using Recurrent Radial Basis Function Neural Networks and Neuro-Fuzzy Systems Daniel Racoceanu1,2 Science Department University of Franche-Comté 25000 Besançon, France E-mail :
[email protected]
Abstract – Multiple reconfiguration and complexity of modern production systems lead to design intelligent monitoring aid systems. The use of artificial intelligence techniques in order to exploit their learning and human experience modeling seems very promising. In this paper, we propose a new monitoring aid system composed by a dynamic neural network detection tool and a neuro-fuzzy diagnosis tool. Learning capabilities due to the neural structure permit us to update the monitoring aid system. The neuro-fuzzy network provides an abductive diagnosis. Moreover it takes into account the uncertainties on the maintenance knowledge by giving a fuzzy characterization of each cause.
I. INTRODUCTION The improvement of the complexity of real production systems in a hard concurrent marketing context encourages the managers to give more importance to the maintenance functions. The industrial monitoring, which is one of the most significant of them, is divided into two tasks: the failure detection, and the failure diagnosis (failure localization and failure causes identification) (How et al., 1999; Pencolé, 2002; Tromp, 2000). More the system is complex, more the monitoring is difficult. An efficient monitoring system must be easy to improve due to system reconfiguration and experts / operators experiences feedbacks. The heterogeneity of maintenance and production information is taken into account for the creation of our monitoring system. Another important consideration is given by the actual tendency to decentralize the control (so the monitoring) using distributed systems. We have seen now new very ambitious e-maintenance concepts3. Moreover, the migration of the intelligence towards the down stages of a control system is already a technological 1
Laboratoire l’Automatique de Besançon (control system laboratory), France – UMR CNRS 6596 2 IPAL (Image Processing Application Lab) – Singapour – FRE CNRS 2339 3 PROTEUS European Project - a generic platform for e-maintenance: http://www.proteus-iteaproject.com
Noureddine Zerhouni1 Control System Department Ecole Nationale de Mécanique et Microtechniques 25000 Besançon, France E-mail :
[email protected]
reality. The use of artificial intelligent techniques seems also very adapted. The monitoring method presented in this paper concerns both monitoring phases: the fault detection and the fault diagnosis. Learning capabilities due to the neural structure permit us to update the monitoring aid system. The neurofuzzy network provides an abductive diagnosis. Moreover it takes into account the uncertainties on the maintenance knowledge by giving a fuzzy characterization of each cause. The following paragraph presents some consideration about monitoring and especially the fundamental criteria used to choose detection and diagnosis tools. In a second part, we briefly describe the two tools with the initialization and configuration of both of them. Finally we conclude by presenting our comment on this approach and the future works. II. CONSIDERATIONS ABOUT MONITORING Generally, the monitoring function is composed of two phases: the detection and the diagnosis. The diagnosis is often the most expensive and complicate step, due to the complex relations between component and possible failure propagation in the system. A. The diagnosis problem In (Dubuisson, 2001) we can find a definition of diagnosis like a problem of pattern recognition. This definition is valid in diagnosis cases in which the modes are well identified, and the associated faults origins and localization well known. In those cases, pattern recognition methods are quite efficient. There are other approaches of diagnosis. Peng, (Peng, 1990) defines diagnosis like having the goal to explain the occurrence of a symptom i.e. to go back to the cause using knowledge on the considered system. This definition will be adopted in our study. Indeed, this approach is common to many authors (Grosclaude, 2000), (Bouchon-Meunier, 2003) and better put in evidence the whole real challenge of the diagnosis.
Close to this definition, diagnosis, as monitoring, takes into account numeric and symbolic data. Moreover, to make a diagnosis possible, the diagnosis needs causal knowledge on the system. Indeed, a default is easily described by the relation between its causes and its effects. The inference used to allowing to “go back to the causes” is called Abductive Inference. Given the complexity of diagnosis problems, many methods have been developed using different tools. We present a brief overview of the most efficient of it, focusing on those based on Artificial Intelligent (AI) techniques. B. Monitoring methods classification The monitoring method can be classified in two categories (Dash et al., 2000): monitoring methodologies based on the existence of a system model, and those not based on this model existence. The first methodologies are the automatic techniques based on the difference between system’s model parameters and the equipments one (Combacau, 1991). The major inconvenience of these techniques is inherent on the difficulty of obtaining the formal model of the complexes equipments. The second methodologies are not sensitive to this problem. These techniques are divided in two categories: the signal processing techniques – “low level” techniques, using sensor signals in order to elaborate simple alarms, without any information about their signification – and the artificial intelligence technique (called also symbolic) – generally used for communication with experts, like decision aid tools and having the ability to interpret (association to a failure mode) and to diagnose faults. We choose to work on symbolic methods which seem to be more adapted to deal with numeric and symbolic data for diagnosis. C. Symbolic (AI) methods for diagnosis The classification proposed here joins the classifications in (Basseville, 1996) and (Aghasaryan, 1998). It gives three classes of symbolic monitoring methods: the methods based on behavioral models, the recognition methods and the methods based on explicative models. - The methods using behavioral models are characterized by the ability to simulate the behavior of the system. They are generally based on tools like finite state machines (Aghasaryan, 1998), (Sampath, 1996) or (Behavioral) Petri Nets (Anglano, 1994). - Recognition methods (like patterns recognition systems and rules based systems) work in two phases: learning and recognition. Most of these methods are based on the Artificial Intelligence with
particular tools such as case based reasoning (CBR), neural networks (NN) and fuzzy logic (FL). - Methods using explicative models are based on models which give a representation of a causal analysis of the system to diagnose. We find here causal graphs (Brusoni, 1995, 1997) (Grosclaude, 2000), contextual graphs and stochastic, partial stochastic Petri nets (Aghasaryan, 1997, 1998), (Tromp, 2000), (Fabre, 2001), and fuzzy logic (Bouchon-Meunier, 2003), (Mellouli, 2000). The next sub-section presents some comment about each symbolic monitoring category. 1) Methods based on behavioural models Finite state machines and Petri nets are well adapted tools for building detection mechanisms when the normal operation of the system is described by these formalisms. On the other hand, their use in diagnosis is still limited. For the automata, the main difficulties are the significant size of the state space. This leads to memory and execution speed problems for the diagnosis. 2) Pattern recognition methods These methods assume that no model is available to describe the relations between causes and effects. Only the knowledge relying on the human expertise consolidated by a solid feedback is considered. The AI applications (CBR, NN, FL, …) used are numerous and for some, the results are overall satisfactory. However, most of these methods carry out a classification by pattern recognition. The diagnosis thus amounts to identify an operating mode of the process which reflects the state of breakdown. In this direction, the diagnosis carried out does not make it possible to identify the causes of the dysfunction unless if they are explicitly described in the identified mode or case as in the CBR. For the other tools, the applications are connected more with "intelligent detection", for which the output of the system of diagnosis is carrying information on the state of the system, but does not give the causes of them. In a monitoring architecture, these tools seem thus better suited for the detection modules. 3) Methods based on explicative models These methods are mainly based on the representation of the relations between the various states of breakdowns and their possibly observable effects. They thus rely on a major analysis of the system, so as to have sufficient knowledge on these relations of cause to effect. Some models allow using an abductive approach which consists in going back to the causes of the breakdowns starting from the observations corresponding to the symptoms.
Monitoring methods
Methods without process model
Methods with model
Fonctionnign and material modelling Industry FMECA FT
Physical modeling Automatic
Symbolic methods Artificial Intelligence
Threshold test Redundancy Test of the mean Parametric estimation
Industrial current methods
Statistic tools Signal Processing
Test of the variance
Alarmes Simple simples alarms
Methods based on behavioral models (modeling/simulation)
Pattern recognition methods (learning/recognition)
Finite state automata
Expert systems
Causal graphs
Behavior Petri nets
Statistic tools
Contextual graphs
CBR
Stochastic Petri nets
Neural networks
Fuzzy logic
Explicatif models methods (Causal analysis of faults)
Fuzzy logic Diagnosis aid Intelligent alarms
Figure 1. Monitoring methods classification. For all these category of methods, four important points can resume the suitable properties for a diagnosis tool: - For the beginning, the first step in order to make a diagnosis system is the model acquisition. - As we have seen, making a diagnosis uses the system knowledge. Thus explicative models seem to be the most adapted to express the causal knowledge of a system, which are essential to carry out a diagnosis. - In the industrial maintenance practice, this knowledge – based on human expertise - is often uncertain. Fuzzy logic is the best tool to express and take into account these uncertainties. - Moreover, a diagnostic tool has to be robust and must use a generic method. At last whatever tool we use, results have to be validated by an expert. III. THE MONITORING SYSTEM DESCRIPTION Following the previous analysis, we have chosen two different tools for the dynamic monitoring and the diagnosis. A. The dynamic detection tool. The input of the detection system is given by sensors data. These data are treated dynamically. The output gives the operating mode (symptom) of the supervised system. The detection tool is based on the recurrent RBF neural network (Zemouri et al., 2000, 2003). The RRBF neural network considers the time as an internal representation. The dynamic aspect is obtained using an additional selfconnection of input neurons with a sigmoid activation function. These looped neurons are a special case of the
Locally Recurrent Globally Feedforward architecture, called local output feed back. The RRBF neural network can thus memorize a part of the historic of the input signal. The following figure shows the principle of this architecture: w11 Input
I1 w22
I2 w33 Output Neurons
I3 Sigmoid Function
Radial Basis Function
Fig. 2. RRBF neural network Inputs are connected to the sensors information through SCADA. The RRBF Outputs will be defined in the initialization phase. The parameters of the input neurons are deducted from the dynamic of the sensor signal. The output of the neuron i is defined by the next function:
with Ii the input, wi the self-connection weight and ki the parameter of the sigmoid. The RRBF neural network outputs are given by the failure modes of FMECA. The learning algorithm of the detection tool is based on DDA algorithm (Berthold and Diamond, 1995) modified following three point: - any prototype can validate a training pattern, - a maximum standard deviation is added, and
-
conflicting prototypes will be adjust only when there is an introduction of a new prototype. The hidden layer is composed of units with a selective response for some range of the input variables. Each unit has an overall response function, a Gaussian:
x is the input, r is the center of the radial basis function (RBF) and σ determines the standard deviation. B. The diagnosis tool. The input of the diagnosis system will be the degree of membership of each operating mode given by the detection. We find also external qualitative or quantitative inputs like information given by operators to improve diagnosis. The output gives a list of possible causes ordered by degree of credibility, and as complementary information: the degree of severity. These degrees help the maintenance manager to evaluate and plan the maintenance actions. During the process, the dynamic detection tool scans continuously the system. When a failure or degradation occurs, an alarm is raised and the diagnosis tool starts. According to the information provided by the detection tool, the diagnosis tool proposes to the operator the possible causes of the symptom as well as the fuzzy interpretation of these causes. This point of view enables us to predict a possible failure. The requirements of the diagnosis aid system are: - easy to use; - capacity of interfacing with industrial data acquisition tools (SCADA – Supervisory Control And Data Acquisition, Ethernet, . . . ); - ability to use an incomplete database and knowledge base; - possibility to take into account new knowledge; - ability to identify false alarms; - interfacing with industrial tools for maintenance management (CMMS, FMECA, Fault Tree); - possibility to interface with Human-Machine Interface (HMI) on PDA, laptop, . . . ; - integration of the tool in an industrial platform. The transformation of the fault tree creates the diagnosis tool. We have associated to each type of gate a neuro-fuzzy architecture. The next figure shows the basic principle of this transformation:
Figure 3. Transformation of AND and OR gates into a neuro-fuzzy architecture. In this figure, we see two types of neurons: - one with linear activation function:
-
and one with sigmoid activation function :
with xi the output and Ii the input and αi the factor representing the frequency of the fault; Concerning the basic events (according to the definition of a fault tree event), we introduce an additional transformation: Each basic event is transformed into a neuron with linear activation function. The Tool Expert extracts data from FMECA and introduces these data in the diagnosis tool. The extraction step consists in considering the frequency and severity of the faults in the FMECA: Then, we introduce the frequency and severity into the diagnosis tool. The severity value is important additional information, to use by the maintenance operator. For each basic event of the fault tree, corresponding to a cause of the extracted FMECA, we extract the value of frequency and transform it to obtain the corresponding α. For the upper sigmoid neurons, each factor α is the maximum of factors α below. For example, in the next figure (two OR gates), α c = max (α a, α b).
Figure 4. Example of α learning. When a link can be established between an operating mode and a state in the fault tree, an input is added on the corresponding neuron. This new input is more important than the propagation of information in the diagnosis tool. C. Monitoring tool functioning principles Before using the neuro-fuzzy system, two steps are necessary: the first one is the configuration where data are collected and extracted to create the tools and the second one is the initialization where data extracted are learned by the tools. The presented procedure allows to do those two steps. In use, the tools are in detection state where the dynamic neural network determines in which mode the system works, with an associated membership degree. When a failure or degradation occurs, the detection tool raises an alert. When a diagnosis is requested, the diagnosis tool uses data from the detection tool to give possible localizations and causes of the problem, classified by credibility and severity degrees. During the monitoring, the maintenance manager can improve the tools by configuration and/or model updating.
The neuro-fuzzy network provides an abductive diagnosis. Moreover it takes into account the uncertainties of the maintenance knowledge by giving a fuzzy characterization of each cause. So, localization and identification of the fault causes are implicitly given by the events in the fault tree. Learning capabilities due to the neural structure permit us to update the configuration and the model, according to the new events given by the CMMS. IV. CONCLUSION In this article, we present a artificial intelligence based integrated monitoring aid system divided into two parts: a dynamic neural network detection tool and a neuro-fuzzy diagnosis tool. The detection is an intelligent and dynamical one due to the RRBR neural network. This provides to identify false alarms and to emit alerts even before the failure threshold is exceeded. The neuro-fuzzy system permits us to keep alive a natural modeling of the human experience, using on-line learning and the abductive inference. Some improvement and a prototype of this tool are in course now. Further work will investigate methods to improve the online learning of the monitoring aid system. V. REFERENCES Aghasaryan, A., Boubour, R., Fabre, E., Jard, C., Benveniste, A (1997), A Petri Net Approach to fault detection and diagnosis in distributed systems, Armen., Internal Publication n°1117 IRISA, France Aghasaryan, A. (1998), Formalisme HMM pour les réseaux de Pétri partiellement stochastiques : Application au diagnostic de pannes dans les systèmes répartis, PhD University of Rennes1, France Anglano, C., Luigi Portinale, L. (1994), B-W Analysis : a Backward Reachability Analysis for Diagnostic Problem Solving Suitable to Parallele Implementation, Proceeding of the 15th International Conference on Application and Theory of Petri Nets, Zaragoza Spain. Basseville, M., Cordier, M-O. (1996), Surveillance et diagnostic de systèmes dynamiques : approches complémentaires du traitement du signal et de l’intelligence artificielle, Report INRIA n°2861. Berthold M. R., Diamond J. (1995), « Boosting the Performance of RBF Networks with Dynamic Decay Adjustment », Advances in Neural Information Processing Systems, Gerald Tesauro, David S. Touretzky, and Todd K. Leen editors, vol. 7, p. 521528, MIT Press, Cambridge, MA. Bouchon-Meunier B., Marsala C. (2003), Logique floue, principes, aide à la décision, Ed. Hermes, Paris. Brusoni, V., Console, L., Terenziani, P., Theseider Dupré, D. (1995), Characterizing temporal abductive diagnosis, , Proc. of International Workshop on Principles of Diagnosis pp. 34-40.
Brusoni, V., Console, L., Terenziani, P., Theseider Dupré, D. (1997), An Efficient Algorithm for Temporal Abduction, Lecture notes in Artificial Intelligence 1321 pp. 195-206. Combacau M. (1991), Commande et surveillance des systèmes à événements discrets complexes : application aux ateliers flexibles, PdD, University Paul Sabatier of Toulouse, France. Dash S., Venkatasubramanian V. (2000), « Challenges in the industrial applications of fault diagnostic systems », Proceedings of the conference on Process Systems Engineering Comput. & Chem. Engng24 (2-7), Keystone, Colorado, p. 785-791. Dubuisson, B. (2001), Diagnostic, Intelligence Artificielle et reconnaissance de formes, Ed. Hermes, Paris. How, W.Y., M. Khalid and S.A.F. Syed Zain. (1999) Transformer Fault Diagnosis Using Fuzzy Logic Interpretations. Conference Proceeding of Instrument Asia, Technical Symposium, Singapore. Fabre, E., Benveniste, A., Jard, C., Smith, M. (2001), Diagnosis of distributed discrete event systems, a net unfolding approach, internal publication IRISA, France. Grosclaude, I., Quiniou, R. (2000), Dealing with Interacting Faults in Temporal Abductive Diagnosis, Proc. of the 8th International Workshop on Principles of Diagnosis. Mellouli, N., Bouchon-Meunier, B. (2000), Fuzzy Approaches to Abductive Inference, Conference on Non-Monotonic reasoning, Berckenridge, Colorado Pencolé, Y. (2002). Diagnostic décentralisé de systèmes à événements discrets: application aux réseaux de télécommunications. PhD thesis, University of Rennes 1, France. Peng Y., Reggia J. A. (1990), Abductive Inference Models For Diagnostic Problem-Solving Springer-Verlag, New York. Sampath M., Sengupta R., Lafortune S., Sinnamohideen K., and Teneketzis D. C. (1996), Failure Diagnosis Using Discrete Event Models, IEEE Transactions On Control Systems Technology, Vol. 4, No. 2. Tromp L. (2000), Surveillance et diagnostic de systèmes industriels complexes : une approche hybride numérique/symbolique, PhD. University of Rennes 1, Zemouri R., Racoceanu D., Zerhouni N. (2000), « Application of the dynamic RBF network in a monitoring problem of the production systems » , 15e IFAC World Congress on Automatic Control, Barcelona, Spain. Zemouri R. (2003), Contribution à la surveillance des systèmes de production à l’aide des réseaux de neurones dynamiques :Application à la emaintenancePhD University of Franche-Comté, France
