Automated rule-based dynamic modeling and controller design

Date of Completion

January 1997


Engineering, Chemical




Classical design of automatic process controllers follows a specific procedure that can be organized into three steps. The first step consists of forcing the process with a change or sequence of changes in the manipulated variable and recording the manipulated and measured process variable data as the process responds. The second step is the regression of a low order linear dynamic model to this manipulated-to-measured process variable data. The third step involves using the resulting dynamic model to complete the controller design. The objectives of this work are to develop guidelines for generation of dynamic data with good quality and to develop methods for automating the last two steps of the controller design procedure. The objective of the guidelines is to help the control designer perform experiments with the appropriate conditions. To establish the guidelines, several studies were conducted to demonstrate the effect of design variables such as manipulated variable wave form, rate at which data is collected (sample rate) and signal-to-noise ratio on the quality of the dynamic data and the shape of the criterion (SSE) space. The methods developed to automate the last two steps employs the Levenberg-Marquardt algorithm and two rule-based expert systems to estimate, in step 2, the process parameters. For step 3, the methodology uses the resulting model parameters and several tuning rules from the literature or derived as part of the work in the design of PID controllers. The two rule-based expert systems are based upon the results found in the studies performed to understand the effect of the design variables mentioned above on the SSE space. The significance of this work are: (1) the development of guidelines for generation of proper dynamic data, (2) a novel fashion to study the impact of design variables on the estimation of model parameters, (3) a methodology to help the control practitioners in the industry or students in the classroom in performing parameter estimation and controller design, and (4) new tuning rules for FOPDT and FOPDT/INT models. Finally, the implementation of these methods as computer software permits transfer of this technology to industrial practitioners. ^