Towards Active Diagnosis of Hybrid Systems leveraging Multimodel Identification and a Markov Decision Process

Abstract

Active diagnosis is defined as the association of fault detection and isolation algorithms with the execution of control plans that optimize fault research performance. This paper addresses active diagnosis of hybrid systems. It proposes to associate a diagnosis method based on multimodel identification and a framework for optimal conditional planning relying on a Markov decision process (MDP). The multimodel diagnosis algorithm identifies the most probable fault by measuring a distance between residual vectors generated from the test system and a set of REFERENCE fault models. Moreover a criterion called the correct diagnosis rate (CDR) is set up to evaluate the accuracy of the diagnosis results depending on the applied operation sequence. Conditional planning is formulated as a MDP, which is a model mixing a discrete structure and probabilistic variables. It is based on a reward function weighing diagnosis accuracy and the cost of actions and the optimal conditional plan is characterized thanks to the recursive Bellman function. An application to a diesel engine airpath model is presented so as to illustrate the diagnosis and planning methods in practice.