Qualitative Multi-faults Diagnosis Based on Automated Planning I: Theory and Modelling He-xuan Hu, Anne-lise Gehin, and Mireille Bayart Laboratoire d’Automatique, Génie Informatique & Signal, UPRESA CNRS 8021, Université des Sciences et Technologies de Lille. Address: Bat. Poly-Tech Lille, Cité Scientifique, 59655 Villeneuve D’Ascq Cedex, France. Email: {He-Xuan.Hu@eleves.polytech-lille.fr, Anne-Lise.Gehin@polytech-lille.fr, mireille.bayart@univ-lille1.fr} Abstract: This paper is intended to present a flexible qualitative framework of multi-faults diagnosis based on the first principles theory of Reiter (1987). We extend his consistency-based approach to deal with the dynamic and continuous systems and offer a necessary assumption and a formal demonstration. Multi-faults diagnosis is a partially observable problem because there is usually not enough information about faults. STRIPS, a classic technique of automated planning, is chosen to build the system model. It provides the reasoning ability for the multi-faults diagnosis when diagnosis is formalized as reasoning from effects to causes with causal knowledge. Keywords: Qualitative, First Principles, Multi-faults Diagnosis, Automated Planning. 1. INTRODUCTION The multi-faults diagnosis is an increasingly active research domain. Reiter (1987) proposed a consistency based approach for multi-faults diagnosis. De Kleer and Williams (1987) have independently implemented General Diagnostic Engine (GDE) which also has the similar ideas. A lot of good practical works have been developed from their theory, Malik et al. (1996), Williams and Nayak (1996) and Ligęza and Kościelny (2008). There is a review about all these diagnostic notions in Lucas (1998). We don’t list the references here again. There are two main components (Venkatasubramanian et al., 2003) in the diagnostic algorithms: (1) the type of knowledge and (2) the type of diagnostic search strategy. Diagnostic search strategy is usually a very strong function of the knowledge representation scheme which in turn is largely influenced by the kind of a priori knowledge available. In Reiter’s theory, the knowledge of system is represented in logic language. The search strategy is generally based on these logic formulae (e.g. Chittaro and Ranon, 2004). The logic formulae describe the static characteristics of system very well, but they are limited in describing the dynamic characteristics of system, for example, the regulation goal always maintains a water level at a set-point in tank. In control domain, automata have been largely applied in the discrete event systems (DES) to describe the behavior of dynamic system (Cassandras and Lafortune, 2007). The variables in a state of automata can be represented by logic atoms. But the DES builds the system model through composing the various pre-designed component models with controller models by parallel composition. The diagnosis is performed by a large diagnoser into which all possible diagnoses are embedded (Sampath et al., 1995). It is obvious that it is neither flexible nor practical to list all possible diagnoses in advance, especially considering the combinations of multi-faults. Dvorak (Dvorak and Kuipers, 1991) and Ng (1991) use the theory of qualitative simulation (Kuipers, 1986), a semi-quantitative method called “qualitative physics”, to extend the Reiter’s theory to deal with the dynamic and continuous system. Qualitative physics does not require detailed information. An order of magnitude information about the normal operating values of process parameters and variables is often sufficient. The advantage is that it does not depend on an accurate mathematical model. But qualitative physics predicts qualitative behavior by using qualitative differential equations (QDEs) that are an abstraction of the ordinary differential equations (ODEs) that represent the state of the system. It has the problem of convergence, divergence and fluctuation in the simulation. The Signed directed graph (Travé-Massuyès et al., 1997) has been the most widely used form of qualitative based model methods for process fault diagnosis. It is constructed to represent the cause-effect relations among the dynamic process variables. It allows tracking of faults, elimination of unwanted causes and identification of fault resources very efficiently. But it is assumed non-deviating sensor readings indicate the absence of any possible faults. If there is no redundant information about the sensor readings, the SDG can not well deal with the sensor faults. In this paper, we propose “automata” as the form of model and “STRIPS” (STanford Research Institute Problem Solver) (Fikes and Nilsson, 1971), a classic technique of automated planning (Ghallab et al., 2004), as the means of building model. This allows us to have a method of modeling that can describe the dynamic and continuous system, that does not depend on an accurate mathematical model and that has an efficient cause-effect diagnostic search strategy. Moreover,