Sensorless control of IPMSM.
I want to compare the Extended or Unscented Kalman Filter and Extended Luenberger Observer to estimate my position of an IPMSM machine.
Can anybody say something, which i should try to implement in my control, first?
Which is the better solution?
Maybe somebody has tested it already, i look forward to hear from you :)
May be , if you are modelling in a noisy environment, yo can try Kalman filter while if you are modelling in a deterministic environment, you can try Luenberger observer .
We have a paper comparing Kalman filters and observers which might be helpful.
https://www.researchgate.net/publication/281363385_Dynamic_State_Estimation_under_Cyber_Attacks_A_Comparative_Study_of_Kalman_Filters_and_Observers
Article Comparing Kalman Filters and Observers for Power System Dyna...
Dear friends, thank you very much for your answers and especially Mr. Qi for his paper! :)
If somebody else wants to join the discussion and has a meaning or a paper about this topic, feel free to answer :)
For me is important the running time (performance) of these both and the robustness against estimation failures of the input parameters.
For example it could be that one method reacts better on a Torque step, but the other method reacts better if the Torque has only small changes (ramp-like).
My environment will be noisy I think, because of electromagnetic fields which can influence the measurement, but if a Luenberger Observer is also robust and has a better performance, it can be the better solution.
By the way, if somebody knows a good method to test the robustness of a system in an analytical way, let me know ;)
Hello. Your question contains an incorrect assumption. The Luenberger observer is a state estimation technique for a deterministic system with known control inputs. The Kalman filter is a state estimation technique for stochastic linear systems, i.e., ones in which the state is a Gaussian random process. As far as failures are concerned, it depends what the failure is and how it affects the system or measurement dynamics. Increasing the process noise in a Kalman filter can allow it to adapt to some instances of mismodelling, as long as they are not too outrageous. I would not call either filter robust. There is no notion of robustness built into either one.
My opinion is that you should tested your system using all of kind of parameter identification like following ARX, ARMAX, FIR, State-space, Multi-variable and choosing one which correspondence at best way to your IPMSM . Then introduce first Kalman filter in the system and after that the Luenberger Observer and check the work of system in a sense of performances and robustness to failure . When you test with Kalman filter state estimation technique inputs should be stochastic Gaussian random process, but with the Luenberger observer state estimation technique inputs should be deterministic known control inputs.After that you should compare all that results for your IPMSM machine and choose one which give the best results you need.
Thank you very very much for your answers :)
I think, I have a better view on this topic than before. So it is not simple to know, which method could be better and the robustness is not really given (only a bit mismodeling like Graham W. Pulham says). And like Vesna Brasic says, i have to test both.
There is another point/problem, somebody before me has implemented a parameter estimator, so he needs omega for estimation and I need the parameters to calculate omega/position angle ;)
So I think, it is very tricky. If I feedback omega, the system maybe unstable.
Another way are Lookup-Tables, I have asked another question about that, if somebody in this circle knows anything about parameter identification of Ld(id,Iq),Lq(Id,Iq) and PsiPM(Id,Iq), i would be very very happy, if the person let me know :):)
Kind regards and thanks to all :)
P.S.: Opinions of new participants who want to join the discussion are welcome :)
THe stabilization of the all system under control with the estimator (Kalman Filter) is not possible ,but it works and gives all the estimated states and parameters,while the use of the extended Luenberger estimator gives a possiblity to synthesise the stability of the all system under control (control based observer).
My response to this question would be more to the application similarities of the two. EXTENDED versions of both kalman and luenberger are suited for use in uncertain or unknown systems where modeling assumptions are mostly the cause of this uncertainty.
However, the basic notion of the luenberger observer is that it is ideal for deterministic or well identified parameter systems. Such systems have few or little room for parameter variation during operation considering their structure and operating environments. however where uncertainty and randomness in events becomes more pronounced in any system, kalman filter should be more considered.
Implementation-wise, kalman filters were mainly introduced for the space program where we know quite well the highly uncertain environment existing in space, therefore the need for some accomodation of stochastic algorithms for system states that could be rapidly and widely varying. if your system is not one that is going to be experiencing rapidly changing parameters, maybe kalman filter is an over-kill...just use the luenberger observer instead.
can anyone also speak in terms of ease of implementation of both algorithms?
and:
Article Multi-site modeling and prediction of annual and monthly pre...
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs