Subspace-based noise covariance estimation for Kalman filter in virtual sensing applications
In newly published work in MSSP (Elsevier), we explore a direct approach to estimate the covariance of possibly correlated process and measurement noises, from actual datasets based on subspace identification. We show that the subspace-based method outperforms the established autocovariance least-squares scheme and provides a good initial guess on the noise covariance, even for the case where the assumed system model is subjected to modeling errors.
If you have worked with Bayesian filtering, then you have likely suffered the curse of the influence of the process and measurement noise sources. The covariance matrices of these noise sources are usually treated as tuning parameters and often adjusted in a heuristic manner to fine-tune the predictions of the system's response (state).
💡 In newly published work in MSSP (Elsevier), we explore a direct approach to estimate the covariance of possibly correlated process and measurement noises, from actual datasets based on subspace identification. We show that the subspace-based method outperforms the established autocovariance least-squares scheme and provides a good initial guess on the noise covariance, even for the case where the assumed system model is subjected to modeling errors. We validate the proposed scheme on a laboratory experiment, where we demonstrate that estimation in virtual sensor positions (locations assumed unmeasured) well match the reference signals:
external page https://www.sciencedirect.com/science/article/pii/S0888327024006708?via%3Dihub
This work is led by Szymon Greś (Aarhus University and SMM Group) and Michael Döhler (Inria and Université Gustave Eiffel) in collaboration with Vasilis Dertimanis and Eleni Chatzi at the Chair of Structural Mechanics and Monitoring at ETH Zurich.