Supervisory adaptive control: Monitoring, diagnostics and model uncertainty

Many of the problems prohibiting more widespread application of adaptive control can be traced to model identification difficulties and model uncertainty resulting from mismatches in plant/model order, stochastic and deterministic disturbances/faults, and the loss of excitation due to the closed-loop nature of feedback control that is common to many adaptive control problems. A potential solution is to use a supervisory level built around the adaptive controller, referred to as Supervisory Adaptive Control (SAC). The objective of this research is to advance the area of SAC by developing a new set of generic supervisory monitoring and diagnostic tools that more effectively deal with model uncertainty issues.; The SAC tools developed in this thesis include: (1) A computationally efficient, numerically stable algorithm for simultaneous identification of model order and model parameters. The algorithm, which results from a new order-recursive interpretation of the Inverse QR Decomposition in least squares, can be applied to a wide variety of modeling problems. (2) An algorithm, based on a Generalized Likelihood Ratio Test, for detecting additive disturbances/faults in complex linear systems. The algorithm combines fault detection, estimation and isolation and has application to statistical process control of autocorrelated processes. (3) A new formulation and interpretation of caution and probing in stochastic adaptive control. A qualitative and quantitative measure of the effects of caution and probing on model accuracy, as it relates to controller performance, is introduced, allowing elements of the optimal dual controller to be efficiently incorporated into an SAC structure. (4) A new method, referred to as Directional Updating, of avoiding parameter drift and bursting in adaptive control. Using a novel decomposition of the Kalman gain vector in Recursive Least Squares (RLS), new insight is gained into the mechanisms by which model accuracy degrades when excitation is lost due to closed-loop identification. A direct measure of the instantaneous level of excitation was derived and used to monitor and appropriately modify the RLS algorithm, effectively eliminating deterioration of model accuracy in the absence of adequate excitation. These SAC tools make adaptive control a more practical solution to many control problems.