Study on outlier robustness of minimum variance control performance assessment

Abstract

Summary Minimum variance (MinVar) method for control performance assessment constitutes one of the most common approaches to the control quality estimation. There are dozens of versions, enriched with numerous reported industrial implementations. MinVar methodology uses the idea of minimum variance, which has been introduced by Kalman. Therefore, it should be remembered that MinVar concept relies on the same assumptions as an idea of the minimum variance control. Among other assumptions, it is essential that the modeled disturbance is an independent random sequence. This article addresses scenarios, when loop noise exhibits non‐Gaussian properties and is characterized by outliers having fat‐tailed distribution. Sensitivity analysis of minimum variance method against such disturbances is evaluated using commonly used PID SISO control benchmarks. It is shown that MinVar method may be significantly biased in such non‐Gaussian situations, which are very frequent in the industrial reality. Reasons of the biased performance are traced and respective solutions are proposed.

Maximum Correntropy Square-Root Cubature Kalman Filter with State Estimation for Distributed Drive Electric Vehicles

“…The process noise and measurement noise in most current studies on state estimation are assumed to satisfy the Gaussian distribution, while in practice, measurement data may be disturbed by various environmental factors, resulting in significant deviations between individual measurement data and other data, i.e., outliers, whose nearby noise has a heavytailed characteristic, which is a general non-Gaussian phenomenon [14,15]. In a nonlinear non-Gaussian environment, the minimum mean square error (MMSE) on the KF shows high sensitivity, which degrades the performance of the KF significantly [16][17][18].…”