A Comparative Evaluation of Models and Algorithms in Model-Based Fault Diagnosis

To increase safety and reliability, many devices are equipped with a diagnostic system. The diagnostic system is a software program that uses a model and observations of the system to create a diagnosis. Diagnosis is often a computational intensive activity. To build a good diagnostic system one has to make trade-offs between the required computing power and diagnostic accuracy. This thesis compares the diagnostic performance of various diagnostic techniques, and their required computing power. In particular, it studies the use of the following diagnostic techniques: Kalman filtering, hybrid mode estimation, and discrete reasoning. Kalman filtering reduces noise in a signal. Hybrid mode estimation uses multiple Kalman filters to find the current state of the system. Discrete reasoning uses discrete instead of continuous variables, for efficiency and modeling convenience. The diagnostic systems used for the tests were created using the Lydia toolset. Lydia is a system specification language. The model-based Lydia tools operate on this specification to perform various, diagnosis related, tasks. This thesis also describes the extension of the Lydia toolset with a software program that performs Kalman filtering and hybrid mode estimation. A simulation model of the Space Shuttle Main Engines was used as a case study on which to test diagnostic systems. Diagnostic systems were tested on their ability to detect failures of the rocket engine, and on the amount of CPU time they required to produce the diagnosis. The results of the test are as follows: The optimal variants of the diagnostic systems under study used about the same amount of CPU time. When the sensor noise is negligible, diagnostic systems that use discretization and discrete reasoning have the highest performance. When there is more sensor noise, the use of continuous domain techniques becomes necessary. Hybrid mode estimation, gives the highest average accuracy, however occasionally outputs spurious fault diagnoses. If this must be avoided, then discrete diagnosis has to be applied on the basis of Kalman filtered sensor readings. It was also necessary to apply an additional filtering of the diagnostic output, in order to decrease the number of false alarms.