Multi-level reasoning and diagnosis for complex continuous-valued systems

Over the last several years, a number of different paradigms have been applied to design of diagnosis systems: (i) traditional approaches that use test programs, fault dictionaries and decision tree representations, (ii) conventional expert system approaches that use associational knowledge derived from human experts, and (iii) the model-based approaches that reason from structural, behavioral, and functional knowledge. Different approaches have their advantages and disadvantages. Heuristic or associational knowledge based approaches link tests, measurements, and observations to faults and are usually fast in identifying a specific set of encoded faults, but often incomplete and inconsistent in terms of covering the set of all possible faults the system may exhibit. On the other hand, model-based approaches are more complete and robust, but sometimes suffer from high computational cost. The success and limitations of these approaches have inspired the development of diagnostic systems that employ knowledge in different forms, and multiple levels of detail. This is also backed up by observations that human experts involved in diagnosis of complex systems use a mix of different kinds of knowledge and strategies for problem solving tasks. This dissertation explores ideas and mechanisms that improve the effectiveness and efficiency of diagnostic problem solving of complex engineering systems. More specifically, a hybrid diagnostic system called MDS has been designed and implemented that combines the associational approach to diagnosis with model-based diagnosis techniques. The diagnostic system consists of two major subsystems: an associational module which is based on heuristic knowledge gained from manuals and expert mechanics, and a model-based diagnosis (MBD) module that incorporates modules for diagnostic reasoning with schematic, functional, and behavioral knowledge. A global controller is designed to coordinate the activities between the two diagnostic subsystems (associational and MBD). Substantial research effort has been focused on the development of model-based diagnosis techniques, and a primary contribution of this thesis has been the development of schemes for building parsimonious models of complex, continuous-valued systems operating in multiple domains (e.g., electricity, thermodynamics, and fluid mechanics) that focus on diagnosis, and efficient algorithms for performing component-based diagnosis with the constructed models. A bond graph-based modeling scheme has been developed that generates equations relating measurable parameters to components of the system. Efficient diagnostic algorithms have also been developed that identify the possible faulty components by performing causal analysis on the set of generated equations. An entropy-based measurement selection algorithm has also been developed that effectively discriminates among possible candidates.