Adaptive process control using pattern-based performance feedback

Abstract

Two adaptive control strategies using pattern-based performance feedback are studied. Both strategies are based on an analysis of patterns exhibited in the recent history of the controller error and focus on adaptation opportunities that arise as a result of sustained set-point changes. The strategies consider a two-parameter adaptation where model gain and time constant are updated after each pattern analysis. The first method uses a vector quantizing neural network for the pattern analysis task. The network is constructed using exemplar controller error patterns developed by deliberately mismatching parameters of a PI controller's internal model against those of an ‘ideal’ simulated system. Once on-line, the network then receives controller error response patterns from the actual application and outputs controller model parameter updates. The second method uses a rule-based pattern analysis to determine features of the controller error response pattern. These pattern features are evaluated relative to user-specified desired features, and rules and procedures again produce PI controller model parameter updates. Both methods are compared in setpoint tracking demonstrations to determine qualitative robustness for non-ideal situations such as measurement noise, constraint of the manipulated variable, model order mismatch and unmeasured oscillatory disturbances. Results reveal that rule-based feature analysis has the benefit of being time independent but the disadvantage of not being able to handle non-ideal situations without a cumbersome rule base. Although the neural network approach requires specification of a time window for pattern analysis, the method proves to be easier to implement and more robust when confronted with non-ideal operating situations.