Download presentation

Presentation is loading. Please wait.

Published byJamar Ufford Modified over 2 years ago

1
VIBRATORY DIAGNOSIS BY THE ARTIFICIAL NEURAL NETWORKS: APPLICATION TO A STEAM TURBINE 1 S. BENAMMAR, D. BENAZZOUZ, M.A MESSARI, F. YAHIAOUI Laboratory LMSS – University of M’hamed Bougara/Boumerdes, 35000, Algeria. 1 Abstract This article concerns the application of the vibratory diagnosis with artificial neural networks to a steam turbine. In this study we propose a fault detection and localization (isolation) system at real-time by using the artificial neural networks (RNA). The network inputs are the numerical values obtained from four sensors of relative vibrations for four bearings. The network outputs represent the detection and the localization of vibratory defects to the real time in four bearings of the turbine. Key words: Failure, Diagnosis, artificial neural networks, Industrial systems, Steam Turbine, Vibration. I-INTRODUCTION The first role of the industrial diagnosis is to increase the availability of the industrial installations to reduce the direct and indirect maintenance costs of the production equipments. The direct costs of this maintenance are the ones relative to the diverse spare parts, the work hand, etc. On the other hand, the indirect costs are essentially due to the loss of income engendered by a stop of production [1] [2]. The increase of the time to repair influences negatively on the indirect maintenance costs. The proposed diagnosis system has for role, to minimize the time of repair by the detection and the fast localization of the failing points, by facilitating to the maintenance operators to intervene as fast as possible. An important characteristic, that’s our system has the possibility of locating several failing points at the same time, for example: an increase in the vibratory level in the four bearings of turbine. The data vectors for the training are intervals limited by two values, (minimum and maximum) witch represent the beach of good performance, affected to “1” and outside of this beach it’s the faulty operation which is affected to “-1”. The training algorithm used for the network is the Levenberg-Marquardt algorithm, the choice of this algorithm that is because it gave a fast training compared to the other algorithms of gradient decent [3] [4]. The programming was completely made by MATLAB 7.5. Bibliographical references [1] Basseville, M. et M-O. Cordier (1996). « Surveillance et diagnostic de systèmes dynamiques: approche complémentaire du traitement de signal et de l'intelligence artificielle », Rapport INRIA, N°2861. [2] Nicolas Palluat, Daniel Racoceanu, Noureddine Zerhouni « Utilisation des réseaux de neurones temporels pour le pronostic et la surveillance dynamique ». Etude comparative de trois réseaux de neurones récurrents. RSTI - RIA. Volume 19 – n° 6/2005, pages 911 à 948. France Article [3] Mohamed Bouamar, Mohamed Ladjal « Système multicapteur utilisant les réseaux de neurones artificiels pour la surveillance des eaux potables » LASS, Laboratoire d’Analyse des Signaux et Systèmes, Université de M’sila, Algérie 4th International Conference: Sciences of Electronic, Technologies of Information and Telecommunications March 2007 – TUNISIA. [4] Manolis I. A. Lourakis (February, 2005), “A Brief Description of the Levenberg-Marquardt Algorithm”, Implemened Institute of Computer Science Foundation for Research and Technology - Hellas (FORTH) Vassilika Vouton, P.O. Box 1385, GR Heraklion, Crete, GREECE. [5] Tarik ALANI, « Réseaux de Neurones Tutorial en Matlab » Département Informatique ESIEE-Paris, Novembre [6] Jawad KARIM, « Surveillance, Diagnostic et Pronostic en Temps Réel de Systèmes Hybrides » : Application à des Bancs d'Essais CERTIA. LAAS-CNRS Groupe DISCO 7, avenue du Colonel Roche Toulouse, France Article CONCLUSION The choice of the architecture MLP for all the treated monitoring systems, returns to its big use in the monitoring field from one part, and from another part, to its speed and its capacity of training by the use of the Levenberg-Marquardt algorithm. The monitoring system suggested facilitates to the maintenance specialists to find the failure origin and to repair this one as soon as possible, all that is added in butts of minimization of the maintenance costs in first part and to increase the production rate in second part. Among the essential advantages of the artificial neural network used in this part: We can first of all supervise several points at the same time without using for each point or parameter a network suitable for him by avoiding the obstruction of the monitoring signals in the control room. The learning and test speed due to the use of the Levenberg-Marquardt algorithm such as this speed allows the network to react in the real time with the data obtained from the sensors. This type of the neural network allows us to realize a big stage of the diagnosis, which is the localization of defects (fault isolation). The application system is a steam turbine of the electrical production thermal plant (SONELGAZ) of Cap-Djinet at Boumerdes. The turbine transforms the thermal energy contained in the vapor coming from the boiler into a rotation movement of the tree. Mechanical work obtained is used to actuate the alternator. It is composed of three bodies, HP body (high pressure), MP body (average pressure) and BP body (low pressure). It has a power and a nominal number of rotation of 176 MW and 3000 rpm respectively. The line of tree rests on four bearings, each one of these bearings thus carries two relative vibration sensors, it is the total of eight (08) sensors on all the line of tree, but for our case, we chose four sensors only for reason of simplification. The maximum value of relative vibrations that can be supported by the system is 120 μm. The following figure represents the supervision sensors site in the bearings turbine. See Fig. 1 Fig. 1: Sensor vibrations site in the turbine bearings C1, C2, C3 and C4 are respectively the supervision sensors of the tree relative vibrations compared to landings 1, 2, 3 and 4. III. CALCULATION OF THE OPTIMAL NEURAL NETWORK ARCHITECTURE II- PRESENTATION OF THE APPLICATION SYSTEM Fig. 2: Suggested network structure Fig. 3: General flowchart for the choice of an optimal architecture Fig. 4: Training of the neural network Fig. 5: Output network simulation Fig. 6: Unacceptable value of test Fig.7: Test with random intervals of values Fig. 8: Suggested network performances

Similar presentations

© 2016 SlidePlayer.com Inc.

All rights reserved.

Ads by Google