Integrating diagnostic knowledge with nonlinear estimation for comprehensive fault diagnosis

Faults lead to loss of productivity and can eventually lead to loss of human lives. Therefore, fault diagnosis is a critical procedure for increased reliability and safety requirement. There are a variety of diagnostic techniques studied in literature. Most of these diagnostic methods either focus on parametric faults assuming reliable sensor information or address the various issues in sensor failure without considering other variations in process. The real diagnostic problem is one in which identification of various process and sensor failures is addressed by a single diagnostic system. In this thesis, a framework is presented to address the comprehensive diagnostic problem.; The proposed solution strategy is a hypothesis driven framework which integrates diagnostic techniques with estimation methods. The diagnostic module comes up with explanations for the process abnormality. The hypothesis sets are constructed from the diagnosis module results and are ordered based on apriori information. The estimation module in conjunction with statistical testing is used to evaluate the hypothesis. Various diagnostic techniques based on nonlinear observers, dynamic time warping of qualitative trends and signed directed graphs have been studied. Various estimation methods like Recursive Nonlinear Dynamic Data Reconciliation (RNDDR), Unscented RNDDR and Two-Step Optimization have been explored. A specific implementation of the proposed comprehensive framework for diagnosis of parameter and sensor faults in a Continuous Stirred Tank Reactor (CSTR) has been presented. This implementation uses signed directed graph model of the process for incipient fault diagnosis based on the initial response of the process. The hypothesis sets are evaluated based on principle of parsimony. The Two-Step estimator has been implemented for rapid estimation of the relevant parameters in the hypothesis.