The SVSF modification to the Kalman filter is an interesting approach. It works well for nonlinear models if the bounds are well understood. You can find more details in
An additional problem arises if you do not know well the noise statistics. In this case, the SVSF extended Kalman filter may produces large and even unacceptable errors. One of the solutions is the extended UFIR filter.
Yes actually it is very strange that I have seen that not only SVSF kalman but also SVSF has unaaceptable error. But strange is that when I apply this iterative method to a clean data points, the noise is created but why? Also when SVSF is applied to unclean laboratory signals, the clean signal is shifted original signal. why? what should be changed to solve the problem?
The problem is not only with noise. First, you must be sure that your state model fits the process. So, you need to investigate how much states has your process. The number of the states is typically equal to the degree of the relevant differential equation. If it fits exactly than there would be no shift (bias). This can be an answer to your questions.
Thank you but the problem of SVSF is for the signals of simulation that have transient oscillations so that this method works on transient part but produces error for the rest of signal in such a way the estimation has to be undertaken separately in these cases.