Model Based Diagnosis of Dynamic Systems

Abstract

Automated fault diagnosis in complex systems quickly detects and isolates component failures. To this end, system models can be effectively and efficiently exploited by mapping deviating measurement values onto a functional description of the system. Functional relations between measurements are then used in a candidate generation stage to find all sets of possible deviating component parameters that explain the measurement values. Typically, this initial set of hypothesized faults contains a large number of candidates and to isolate the true fault, future system behavior is predicted from the system model for each of the possible faults. Continued monitoring of these predictions then prunes the initial set of fault hypotheses till the true fault is identified. This interaction of candidate generation, prediction, and monitoring is referred to as model based diagnosis. Given an underlying modeling formalism that is applicable across physical domains (e.g., electrical, mechanical), the generic diagnosis algorithms can be combined with a model of a specific system to quickly generate a diagnosis engine. If abrupt faults occur, fault detection and isolation (FDI) can be efficiently performed based on transients in system behavior. This requires a dynamic model of the system. To support compositionality, topological models are well suited for diagnosis. Furthermore, because diagnosis is basically a search process, to keep computational complexity low, it is important for diagnosis models to incorporate as many constraints as possible. All of these requirements are elegantly addressed by the bond graph modeling formalism which is based on energy exchange, power, between system components. Therefore, it captures phenomena in a number of physical domains, e.g., as in electro-mechanical systems. Moreover, bond graphs allow for algorithmic assignment of causality to derive a set of dependency relations of the system. This paper shows how the temporal aspects of these relations can be systematically captured by a temporal causal graph. The use of bond graphs inherently requires conservation of energy and continuity of power to be satisfied, and, therefore, the derived temporal causal graph provides a well constrained model. We show how this graph results in efficient diagnosis of a complex, nonlinear, model of a secondary cooling system in fast breeder reactors.