Comparison of Covariance Estimation using Autocovariance LS Method and Adaptive SRUKF

State estimators are used to reconstruct current plant states based on information received from plant sensors and the use of a mathematical model. The typically applied Kalman filter derivatives require knowledge about the noise statistics affecting system states and measurements. These are often unknown and inaccurate parameterization may lead to decreased filter performance or even filter divergence. In this paper, a comparison between two covariance estimation methods is presented. The offline time-varying autocovariance Least-Square (LS) method is compared to the online adaptive Square-Root Unscented Kalman Filter (SRUKF). Both methods are evaluated in simulations and experiments using a pendubot w.r.t. robustness against random covariance initializations and state estimation accuracy. The results show that with both methods the filter performance can remarkably be improved.