Copyright © IFAC 12th Triennial World Congress. Sydney. Australja. 1993 GENERATION OF OPTIMAL STRUCTURED RESIDUALS IN THE PARITY SPACE M. Staroswleckl, J.P. Cassar and V. Cocquempot LAIL URA 1440D. llniversiti des Sciences et Technologies tU Lille. 59655 Villeneuve d'Ascq Cedex, France ABSTRACT This paper proposes an approach for the improvement of the performances of Failure Detection and Isolation (FDI) sys- tems. Structural speci ficat ions of the expected results lead to optimize the robustness of the system with respect to the modelling uncertainties and unknown inputs. Most of the solutions of this optimization problem have been developed through the transformation of an observer - based or a parity space approach. We show in this paper how the parity space approach allows the di rect generation of residuals which respond to the given speci fications. KEYWORDS: Fault detection and isolation; muIticrite- ria optimization; robustness; sensitivity. I - INTRODUCTION Modern indutrial plants become more and more complex and as a consequence more and more sensitive to failures. Component failures may degrade the overall system performances and have serious consequences on the security. Thus, it becomes important to detect quickly and accurately any occurence of a failure and to identify its nature. This is the aim of the Failure Detection and Isolation (FOI) systems, whose design becomes a necessary complement to the design of the control algorithms . The classical FDI system scheme [11 makes appear a characterization and a decision step. Using a model based method [21[31[41. the first step generates a set of residuals which express the di fference between the information provided by the system and that delivered by its model in normal operation. Then, the residuals characterize the system's operating mode: 535 close to zero in normal operation, diffe- rent from zero in faulty situations. Whatever the approach used to generate the residuals from a linear state space re- presentation, their values will be func- tions of the system's failures, but also of the measurement noises, unknown inputs and modelling errors. The aim of the decision procedure is to decide, from the residuals values, whe- ther the no fault hypothesis may be for- mulated or not. In order to insure a minimal false alarm rate, the residuals have to be robust whith respect to unknown inputs and modelling uncertainties. In order to in- sure better detection conditions, they have to be sensitive whith respect to fai- lures [2][5J. Isolation considerations lead to add structural constraints [6][7] on each resi- dual in order to insure that each faulty element can be identified from the set of residuals. The structural specifications associate the detection conditions of two sets of residual deviations causes [8][9]: the first set includes the elements whose failures must be detected and the second set the modelling errors and the elements which must weakly influence the results. For a given residual, the robustness concept is thus extended to failures whose occurence must not affect its value. Classical approaches to robustness lead to create residuals which are insensitive to the unknown inputs. This insensitivity is geometricaly traduced by orthogonality conditions which are obtained by the de- sign of specific observers (unknown in- put observer, or eigenvalue assignement methods [10]) or by structural approaches [Ill. In all cases, that leads to a decrease of the number of design parameters, since the orthogonality condition creates links between them . For some cases, no solution can be found, or the sensitivity of the obtained residuals to failures may be low.