| Summary: | Thesis: Ph. D., Massachusetts Institute of Technology, Department of Chemical Engineering, 2016. === Cataloged from PDF version of thesis. === Includes bibliographical references (pages 121-126). === In the pharmaceutical industry there has recently been much interest in design spaces: sets of critical process parameters (CPPs) which guarantee that critical quality attributes (CQAs) of a manufacturing process are within specifications. For continuous pharmaceutical processes, design spaces are usually calculated by assuming steady state operation and approximating the mapping between CPPs and CQAs using a Taylor series. The full design space can then be calculated using a plantwide approach or a unit-by-unit approach. Common inner approximations of the design space (e.g. hyper-rectangles) can result in significant conservatism, especially when a unit-by-unit approach is employed. Because control loops tend to have a linearizing effect on processes, design spaces for closed-loop processes can often be calculated using low-order Taylor series approximations, resulting in simpler expressions for the full design space (e.g. polytopes). Control loops also tend to enlarge design spaces, sometimes by more than an order of magnitude. Unfortunately, disturbances, noise, and uncertainties will prevent real processes from ever reaching "steady state". Therefore, design spaces calculated at steady state cannot be used to guarantee quality specifications. In fact, because design spaces fail to take into account any process dynamics, constraining a controller to work within a design space may result in failure to meet quality specifications, significant degradation of controller performance, and input jitter. As a substitute for design space, robust model predictive control (RMPC) is a promising technology for dynamically guaranteeing constraint satisfaction on process outputs. Although many RMPC algorithms have been proposed in the literature, the computational cost of these algorithms tends to be a strong function of the state vector size. This is problematic for continuous pharmaceutical processes, which are typically high- or infinite-dimensional. However, input-output models (e.g. finite step response models) can integrated with traditional RMPC strategies to robustly control high-dimensional systems. Although RMPC can be used to counteract the presence of disturbances, uncertainty, and measurement noise, faults also present a threat to quality constraint satisfaction of continuous pharmaceutical processes. Active fault diagnosis of hybrid systems is particularly difficult due to the explosion of mode combinations with prediction horizon. Fortunately, the set of input sequences which do not guarantee diagnosis can be outer bounded offline as a function of a parameterized initial condition set. This enables an algorithm for guaranteed active fault diagnosis of hybrid systems which can be implemented quickly online. === by Lucas Charles Foguth. === Ph. D.
|
|---|
