Outlier Sensitivity of the Minimum Variance Control Performance Assessment

Abstract

Minimum variance (MinVar) control performance assessment (CPA) constitutes one of the most common approaches to the control quality estimation. There are dozens of versions of this method, enriched with practical implementations. However, it should be remembered that the method relies on the same assumptions as the minimum variance control. It is essential that considered disturbance is an independent random sequence. This paper addresses the situations, when loop noise has non-Gaussian properties and is characterized by outliers exhibiting fat-tailed distribution. Sensitivity analysis of minimum variance method against the outliers is conducted using commonly used PID control benchmarks. It is shown that CPA using minimum variance may be significantly biased in non-Gaussian situations, which are very frequent in the industrial reality.