Model-based sensor placement for component condition monitoring and fault detection in an integrated gasification combined cycle plant
MetadataShow full item record
The goal of this thesis is to establish a comprehensive methodology to determine the type, location, and cost/number of sensors required for component condition monitoring and fault detection in fossil energy systems. The root cause for productivity losses and shut-downs are called “faults”. Through a transformative two-tier framework, the thesis seeks to develop a model-based sensor placement methodology that addresses: (i) sensor placement for fault detection based on tractable models that are developed from the system-level dynamic model, (ii) identification of precise locations for component condition monitoring based on distributed component-level models. With these objectives in mind, algorithms are developed for maximizing the effectiveness of the sensor network for system-level fault detection and component-level condition monitoring. Integrating the networks identified at each level, the 2-tier sensor network is obtained. The algorithms are developed for and tested on a high fidelity model of the gasification island of an integrated gasification combined cycle (IGCC) plant. For the identification of the system-level fault detection network, the problem is approached through qualitative reasoning using a cause and effect model (graph model) to identify an optimal cost network. This well-studied approach avoids solving thousands of partial differential equations in an optimization loop and reduces the problem to solving an integer linear programming problem. The classical algorithms for fault detection using graph models are enhanced through the use of numerical simulations and introduction of quantitative features. For a condition monitoring network, whether equipment should be considered at a component-level or a system-level depends upon the criticality of the process equipment, its likeliness to fail, and the level of resolution desired for any specific failure. Because of the presence of a higher fidelity model at the component-level, a sensor network can be designed to monitor the spatial profile of the states and estimate fault severity levels. The state estimation is performed using an extended Kalman filter (EKF) that estimates the states, including faults, at the component level. A genetic algorithm is used in consort with the EKF to find the optimal measurement model for highest accuracy of the estimates with a fixed number/budget of the sensors. The identified measurement model represents the optimal sensor network for component condition monitoring. In an IGCC plant, besides the gasifier, the sour water gas shift (WGS) reactor plays an important role. Yet, it is one of the equipment with a high likelihood of failure because of the harsh conditions that it is subjected to. In view of this, we have considered condition monitoring of the sour WGS reactor at the component-level, while a detailed plant-wide model of gasification island (including sour WGS reactor and the SELEXOL process) is considered for fault detection at the system-level. SELEXOL process is a unit of acid gas removal process in an IGCC for removing carbon dioxide and other impurities. Finally, the developed algorithms unify the two levels and identify an optimal sensor network that maximizes the effectiveness of the overall system-level fault diagnosis and component-level condition monitoring in gasification island. Measurement and model uncertainties are naturally handled in the solution approach while sensor failure probabilities and failure occurrence probabilities can be easily included, if required. In addition, the same algorithms developed in this thesis can be further enhanced to be used in retrofits, where the objectives could be upgrade (addition of more sensors) and/or relocation of existing sensors.