AUTOMATIC DESIGN OF DETECTION TESTS IN COMPLEX DYNAMIC SYSTEMS Stéphane Ploix, Matthieu Désinde, Samir Touaf Laboratoire d’Automatique de Grenoble, INPG, UJF, UMR 5528, BP 46, F-38402 Saint Martin d’Hères Cedex, France Phone: 33 4 76 82 62 44, Fax: 33 4 76 82 63 88 Abstract: In complex industrial plants, there are usually lots of sensors and the modelling of the plant leads to lots of mathematical relations. Before using classical tools for fault detection, the first problem to solve is: what sensors and mathematical relations have to be selected for the design of a detection test such as a state observer or a parity space based detection algorithms. This paper presents a general method for automatically selecting relevant sensors and relations that may be used for the design of the different detection tests. This method, which is based on a structural analysis of the process, provides all the testable subsystems and permits the selection of the most interesting detection tests regarding detectability and diagnosticabillity criteria. Copyright © 2005 IFAC Keywords: Dynamic Systems, Fault detection, Fault diagnosis, Detection algorithms design 1. INTRODUCTION In the scientific literature, there are two main streams for the design of detection tests. The first one, mostly used by researchers coming from the Fault Detection and Isolation community (Patton et al., 1989), relies on global models of systems to be diagnosed. It is often called structured or robust approach because it aims at projecting residuals in different spaces in order to discriminate the different faults that may occur. Another stream comes from the Artificial Intelligence community (Reiter, 1987). It relies on component based approaches. The principle is to model the different components of the system and to combine these models in order to perform detection tests. The FDI approach mainly deals with dynamic systems whereas the DX approach mainly focus on static system. Detailed studies about the comparison between these approaches have been achieved by the IMALAIA French research group (Cordier et al., 2000) and by the BRIDGE action of the European network of excellence called MONET (http://monet.aber.ac.uk: 8080/monet). The MAGIC European project (Köppen-Seliger et al., 2002) has shown that bridge approaches between the 2 communities (Nyberg and Krysander, 2003; Ploix et al., 2003), taking advantage both of detection tools for dynamic system and of formal reasoning for fault isolation are very suitable for industrial plants (Garcia-Beltram et al., 2003). These bridge approaches are, on one hand component based, i.e. each component state is individually modelled, and in the other hand, they cope with sophisticated detection tests. If the systems are simple, the design of detection tests may be easily handled but if the system becomes a little bit more complex, this task becomes unachievable. For instance, a bioprocess modelled by 34 mathematical relations has led to 736 possible detection tests. It can obviously not be solved by hand. A detection test has to be distinguished from its support, called Testable Sub System (TSS) in (Ploix et al, 2003). A TSS gathers the constraints used for its design and the hypotheses the test relies on. A given set of constraints may be combined in many different ways and, providing that the set is testable, leads to many different detection tests but all of them