Data driven model mismatch detection based on statistical band of Markov parameters

Abstract

In modern process industries, model accuracy is critical for safe operation as well as control performance of process industrial systems. It is in practical intractable to locate the corresponding system models that contain significant mismatches and sometimes re-identification may be needed to implement to hundreds of process loops. As such, a novel data driven methodology is proposed to detect the model-plant mismatches. Subspace approach and moving window scheme are integrated to estimate the Markov parameters of the process models. Then, the statistical bands of the Markov parameters are calculated using routine operation data. Thus, the model mismatch can be detected by evaluating the bias between the band of the normal case and that of the monitored case. The mismatch models can be isolated, which facilitates the decision when and where to take the re-identification. The proposed method avoids extra efforts and costs caused by full-scale experiments to the process. Simulations on a distillation process is employed to demonstrate the efficiency of the proposed approach.