Detection and diagnosis of model-plant mismatch in multivariable model-based control schemes

The extent of approximation in modelling a given process, characterized by the model-plant mismatch (MPM), amongst other factors, critically determines the performance of a model-based control scheme. It is necessary therefore to carry out model maintenance and correction on a regular basis. However, a complete re-identification is usually a costly exercise. Therefore, it is highly desirable to precisely determine the specific elements that are in mismatch and re-identify only those parameters. In the recent times, the plant-model ratio (PMR) was proposed as an effective metric for diagnosing MPM in single-input single-output (SISO) systems from closed loop data. The PMR facilitates unique detection of mismatch in gain, dynamics and delay. A straightforward application of PMR to multivariable closed-loop systems is challenging primarily due to the confounding effects of other inputs and loop-to-loop interactions under closed-loop conditions. Furthermore, the metric requires high-frequency excitation for identification of delay mismatch. In this work, we first present a method to overcome the latter requirement using Hilbert transform relation and partial cross-spectral densities. Subsequently, we present the key contribution of this work, that of generalizing the PMR approach to multivariable control systems. Two threshold-based hypothesis tests are presented for diagnosing mismatch in gain and dynamics. Three simulation case studies are presented to demonstrate the efficacy of the proposed method.