Pattern recognition for the diagnosis of process excitation and controller performance in model-based adaptive process control

Date of Completion

January 1993


Engineering, Chemical|Engineering, System Science




The use of pattern recognition for model-based adaptive control is presented. Vector quantizing neural networks (VQNs) analyze patterns displayed in recent histories of the process input, output and controller error. The results of these analyses lead to diagnoses and adaptive actions. Process excitation diagnostics are made for online model regression. Performance feedback diagnostics are made for updating both process model and controller tuning parameters.^ Traditional control algorithms do not account for the nonlinearity and nonstationarity of chemical processes so they are unable to track changing process characters. The effort they exert toward keeping the process at set point becomes too great or too small and controller performance degrades. Performance feedback adaptive algorithms update the controller parameters in response to observed controller performance. This tracks the changing process character and maintains desired performance. Model-based adaptive algorithms update a process model used in controller design. These are widely applicable to any model-based controller.^ Presented here are pattern-based approaches to performance feedback and model-based adaptive control. These tools are designed and demonstrated individually using VQNs to perform the pattern recognition task.^ One pattern-based tool analyzes process input and controller error response patterns after changes in set point and unmeasured disturbances. This leads to diagnoses concerning controller performance and disturbance character. These diagnoses determine appropriate adaptive actions for updating both model and controller tuning parameters.^ A second pattern-based tool continuously analyzes the recent histories of the process input and output to determine if dynamic trends exist. The results of these analyses are used to diagnose if input-output data contain sufficient dynamics for online model regression. Once determined valid, the new model is used to redesign the controller.^ These pattern-based tools are combined into a unified adaptive framework. A supervisory adaptation logic allows the adaptive mechanisms to work together to maintain both model accuracy and desired controller performance. This framework is implemented as a piggy-back adaptive module and investigated as a viable commercial product. Demonstrations of this work employ IMC tuned PI with prediction, GPC, and DMC control of both simulated and bench-scale processes. ^