Retour au site du LAAS-CNRS

Abstract

Complex industrial processes invest a lot of money in sensors and automation devices to monitor and supervise the process in order to guarantee the production quality and the plant and operators safety. Fault detection is one of the multiple tasks of process monitoring and it critically depends on the sensors that measure the significant process variables. Nevertheless, most of the work on fault detection and diagnosis found in literature place more emphasis on developing procedures to perform diagnosis given a set of sensors, and less on determining the actual location of sensors for efficient identification of faults. A methodology based on learning and classification techniques and on the information quantity measured by the Entropy concept, is proposed in order to address the problem of sensor location for fault identification. The proposed methodology has been applied to a continuous intensified reactor, the 'open plate reactor (OPR)', developed by Alfa Laval and studied at the Laboratory of Chemical Engineering of Toulouse. The different steps of the methodology are explained through its application to the carrying out of an exothermic reaction.

Mots-Clés / Keywords

Abstract

Complex industrial processes demand significant financial investment in sensors and automation devices to monitor and supervise the process in order to guarantee the production quality and the plant and operators safety. Fault detection is one of the multiple tasks of process monitoring and it critically depends on the sensors that measure the significant process variables. Nevertheless, most of the work on fault detection and diagnosis found in literature place more emphasis on developing procedures to perform diagnosis given a set of sensors, and less on determining the actual location of sensors for efficient identification of faults. A methodology based on learning and classification techniques and the information quantity measure, by the entropy concept, is proposed in order to address the problem of sensor location for fault identification. The proposed methodology has been applied to a new concept of intensification reactor, the Open Plate Reactor, developed by Alfa Laval and the Laboratory of Chemical Engineering located at Toulouse.

Mots-Clés / Keywords

Abstract

The principal objective of this work is the identification and the location of sensors on a complex chemical plant needed for online process situation monitoring, fault detection and diagnosis of malfunctions. This identification is based on the use of a classification technique and a measure of the quantity of information provided by the process variables, the entropy. Any classification method providing an interpretable description of the classes describing the process situations can be applied. In this work, the LAMDA (Learning Algorithm for Multivariate Data Analysis) classification method was employed for the design of the support tool. LAMDA combines Fuzzy Logic concepts, such as the adequacy of an element to a class, and the neural model representation. It allows, without changing of algorithm, to carry out classifications using a supervised (directed) or unsupervised (automatic) learning stage. The illustration of such a methodology is shown on a classical chemical plant: the propylene glycol production plant. This chemical process is composed of a mixer, a chemical reactor (CSTR) and a rectification column. This plant has been designed and simulated (dynamic simulation) using the well-known HYSYS simulation package. This simulation model has been used to generate scenarios of the various faults and malfunctions generally encountered in this type of plant. In particular, faults affecting the production quality have been simulated. After a short presentation of the most popular classification methods and the Entropy concept, the steps for the development of the proposed support tool are explained. This methodology is then applied to the example of the propylene glycol production plant. The present results highlight the contribution of both the methodology to select the right sensors and the classification technique to the design of a behavioral model used for monitoring and fault detection.

Mots-Clés / Keywords

Mots-Clés / Keywords

Mots-Clés / Keywords

Abstract

This paper presents a monitoring methodology to identify complex systems faults. This methodology combines the production of meaningful error signals (residuals) obtained by comparison between the model outputs and the system outputs, with a posterior fuzzy classification. In a first off-line phase (learning) the classification method characterises each fault. In the recognition phase, the classification method identifies the faults. The chose classification method permits to characterize faults non included in the learning data. This monitoring process avoids the problem of defining thresholds for faults isolation. The residuals analysis and not the system variables themselves, permit us to separate fault recognition from system operation point influence. The paper describes the proposed methodology using a benchmark of a two interconnected tanks system.

Mots-Clés / Keywords

Mots-Clés / Keywords

Abstract

PrODHyS is a dynamic hybrid simulation environment which provides common and reusable object oriented components designed for the development and the management of dynamic simulation of industrial systems. In this work, we present a method for the fault detection based on the comparison the real system and the reference model evolution generated by the extended Kalman filter. The reference model is simulated by PrODHyS.

Résumé

Le travail présenté propose une méthodologie de classification par apprentissage qui permet lidentification des états fonctionnels sur une unité de coagulation impliquée dans le traitement des eaux de surface. La supervision et le diagnostic de ce procédé ont été réalisés en utilisant la méthode de classification LAMDA (Learning Algorithm for Multivariate Data Analysis). Cette méthodologie dapprentissage et dexpertise permet dexploiter et dagréger toutes les informations provenant du procédé et de son environnement ainsi que les connaissances de lexpert. Létude montre quil est possible dajouter aux informations issues des capteurs classiques (température, matières en suspension, pH, conductivité, oxygène dissous), la valeur de la dose de coagulant calculée par un capteur logiciel développé dans une étude antérieure afin daffiner le diagnostic. Le site dapplication choisi pour lidentification des états fonctionnels est la station de production deau potable Rocade de la ville de Marrakech, Maroc.

Mots-Clés / Keywords