Fault Diagnosis of Hybrid Systems with Dynamic Bayesian Networks and Hybrid Possible Conficts

Abstract

Hybrid systems are very important in our society, we can find them in many engineering fields. They can develop a task by themselves or they can interact with people so they have to work in a nominal and safe state. Model-based Diagnosis (MBD) is a diagnosis branch that bases its decisions in models. This dissertation is placed in the MBD framework with Artificial Intelligence techniques, which is known as DX community. The kind of hybrid systems we focus on have a continuous behaviour commanded by discrete events. There are several works already done in the diagnosis of hybrid systems field. Most of them need to pre-enumerate all the possible modes in the system even if they are never visited during the process. To solve that problem, some authors have presented the Hybrid Bond Graph (HBG) modeling technique, that is an extension of Bond Graphs. HBGs do not need to enumerate all the system modes, they are built as the system visits them at run time. Regarding the faults that can appear in a hybrid system, they can be divided in two main groups: (1) Discrete faults, and (2) parametric or continuous faults. The discrete faults are related to the hybrid nature of the systems while the parametric or continuous faults appear as faults in the system parameters or in the sensors. Both types af faults have not been considered in a unified diagnosis architecture for hybrid systems. The diagnosis process can be divided in three main stages: Fault Detection, Fault Isolation and Fault Identification. Computing the set of Possible Conflicts (PCs) is a compilation technique used in MBD of continuous systems. They provide a decomposition of a system in subsystems with minimal analytical redundancy that makes the isolation process more efficient. They can be used for fault detection and isolation tasks by means of the Fault Signature Matrix (FSM). The FSM is a matrix that relates the different parameters (fault candidates) in a system and the PCs where they are used. PCs provide a computational model that can be implemented by means of different tools. In this dissertation, Dynamic Bayesian Networks (DBNs) have been chosen to implement the PCs and they can be used during the three diagnosis stages. DBNs main problem is the computational burden for inference tasks, this is partially solved using approximate inference like Particle Filter (PF) Algorithms. This dissertation proposes to derive minimal DBNs from PCs which reduces the complexity of the DBNs improving their performance. The main contribution of this dissertation is the Hybrid Possible Conflicts (HPCs) formalism, PCs have been extended for hybrid systems. Related to that, the dissertation can provide the second important contribution: a diagnosis architecture for hybrid systems integrating discrete and parametric faults. This architecture is based on Hybrid Possible Conflicts and DBNs derived from them. The architecture uses DBNs along the three diagnosis stages. The diagnosis architecture has been also proposed for continuous systems fault diagnosis as well as a method to derive minimal DBNs from Possible Conflicts. Finally, several simulation systems have been used to test each contribution. The systems come from different fields: Connected tanks (hydraulic), twelfth order electrical circuit (electric), and the ROS, a more complex hybrid system from the aerospace field. The results obtained have been satisfactory with all of them.