Methods of Robust Active Fault Detection and Isolation in Thermal Fluid Systems

Advancements in cyber-physical system architectures and technologies have increased system uncertainty. The resulting lack of robustness in system diagnostics, specifically the chosen methods of fault detection and isolation (FDI), has garnered research interest in aerospace, automotive, chemical, defense, electronics, and energy industries. Of the existing FDI methods, model-based methods have drawn attention due to their ability to create analytical redundancy via predictions of system responses using physics-based or empirically-derived equations. Additionally, the integration of model-based FDI and optimization has been studied to further reduce the occurrences of nondetections, false alarms, and misdiagnoses during maintenance. One benefit of conducting system diagnostics during maintenance is the absence of normal operating requirements, allowing for greater manipulation of the system operating conditions to obtain more accurate information about the presence of fault(s). The system operating conditions that maximize the information accuracy were found by formulating and solving constrained mathematical programs. Three thermal-fluid systems were studied to illustrate the benefit of the model-based FDI methods of this dissertation in various applications. First, a mathematical program was formulated and solved for the system inputs that provided unique system responses for each studied fault scenario. The solution to this program was a diagnostic test capable of generating a map of unique system responses for all fault scenarios, providing complete FDI. Next, a semi-infinite program was formulated and solved for the system inputs that maximized FDI effectiveness at the worst-case realization of uncertainty, while adhering to safety-critical constraints. The corresponding program was solved using an algorithm capable of providing global solutions, resulting in a diagnostic test that guaranteed system safety over the entire uncertainty domain. Finally, inferential sensors were explored to improve system diagnostics. Symbolic regression was implemented using a genetic programming procedure to design inferential sensors that functionally combine the system inputs and outputs to improve the information of a system diagnostic test. The designed inferential sensors provided more accurate information with respect to faults and reduced the impact of uncertainty than their original sensor counterparts. The diagnostic capability of the inferential sensors was also compared to traditional fault classification methods and it was found that the inferential sensors were equivalent to, and in some cases outperformed, these methods. Lastly, the explainable inferential sensors were applied to an information theory metric, the Fisher Information-derived Bayesian D-optimality criterion, to validate their capability of reducing the impact of uncertainty and distinguishing fault parameters of interest. It was proven that the designed inferential sensors increased the D-optimality criterion when compared to the original output information.