Consider a sensor for which we are designing a Kalman filter for improved accuracy or sensor fusion. Let us restrict this discussion to the goal using the filter to improve the accuracy of the aposteriori state estimate and avoid addressing sensor fusion. Further, let us assume the state and output to be identical and one-dimensional for simplicity.
The designer of the Kalman filter assumes the knowledge of the system dynamics and output channels including apriori knowledge of the variance (or squared standard deviation or squared intensity) of the Gaussian white noise appearing in the process dynamics and observations. While the well known methods to determine the noise variances are as follows,
- the standard industry approach to determine the noise intensity is by tuning the value in the Kalman gain computation to improve closed-loop performance,
- the auto-correlation least squares method which uses closed-loop data,
it is not clear in the literature if there are any algorithms which use open-loop data for determining the noise covariance.
How is the intensity of the Gaussian white noise model of a sensor determined experimentally using open-loop data?