Using mathematical models in knowledge-based control systems

Knowledge based systems (KBS's) are valuable tools for developing control applications in the process industries. The goal of these applications is to automate many of the tasks that are currently performed by human operators. This work presents systematic methodologies for two problems related to process control: (1) process fault diagnosis and (2) supervisory control.; The approaches developed in this work assume that there is some mathematical representation of the process. Models are an important source of information for solving problems related to process operation. Most models however are only estimates of the actual process behavior. This fact must be considered when relying on this source of information. The methodologies developed provide direct means for dealing with specific types of model uncertainty.; Model-based information is difficult to incorporate into traditional KBS architectures in a fashion that can be easily maintained and understood. In this work, the models are represented in their explicit form to alleviate these problems.; A fault diagnosis methodology is developed that uses model equations as evidence. Equations that are violated indicate that there is a possible fault in the process. The pattern of failed equations is used to isolate the most likely fault conditions. Equation residuals are rated using fuzzy sets to give an indication of the degree of equation violation. This allows for uncertainty in the model equations, as well as the tolerances on the equations.; The supervisory control problem is concerned with the optimal distribution of hydrogen in a refinery. This problem is solved using a KBS that employs a numerical optimization routine to reach its conclusions. The system offers specific advice to operating personnel regarding the best distribution of hydrogen gas. Uncertainty in the optimization model is accounted for by allowing the constraints to be represented in terms of fuzzy sets instead of sharp limitations.; This work offers systematic methodologies for dealing with model-based information and uncertainty as applied to these two specific control problems. These approaches are useful in building advanced control systems that are capable of helping operators solve many problems related to process operation.