Multiple model predictive control: A novel algorithm applied to biomedical and industrial systems

Biomedical and chemical processes are multivariable in nature with highly nonlinear systems, making it difficult to obtain on-line measurements for many of the outputs that are to be regulated. For example, chemical reactors rarely have sensors to measure product concentrations or extent of reaction. Likewise, biomedical data such as heart contractility, or blood drug concentration are not available to a critical care physician. To make matters worse, there are real constraints on inputs in both industrial and biomedical systems. In the case of drug infusion, if the controller does not adhere to limits on the quantity of a drug or the rate of its infusion, severe toxic effects and possible death can occur.; A model-based controller that can handle multivariable systems with transport delays and input constraints is necessary for the regulation of these systems. Model Predictive Control is an on-line optimization technique which naturally and optimally handles multivariable processes with any set of dynamics and can explicitly impose input constraints in its calculations. The difficulty with Model Predictive Control for biomedical and chemical processes is in obtaining quality models that can predict responses of the systems at different operating regimes. The approach developed here uses a bank of models to help describe the range of dynamics that exist in most biomedical and chemical processes. How well each model predicts the responses exhibited by the system determines what weighting the model gets. A weighted sum of all the models provides a prediction model to the Model Predictive Controller. Multiple Model Predictive Control has been used successfully in canine experiments to regulate mean arterial pressure and cardiac output as well as in simulations of chemical industrial processes. Being a relatively new approach, there are many issues investigated in this thesis, such as determining the number and type of models in the bank, changing tuning parameters to maximize controller performance under different operating regimes, and designing a model bank to reject as many disturbances as possible.