Hi everyone,
I have wondered about choosing rightly cycle time for measurement and sampling time for control speed of DC motor. I have incremental encoder and I have implemented Kalman Filter for estimate speed which is used in feedback loop.
Am I right that cycle time of Kalman Filter algorithm (generally speaking measurement) should be at least 2 times lower than sampling time of speed controller? What about setting the same cycle time for both measurement and controller algorithm?

Is it any difference in case of using Kalman Filter or direct measurement ?
Thank you very much for any help and suggestions!

 

The sampling time for a Kalman filter is usually the same as for your closed loop control. I/we don't use Kalman filters for motion control because Kalman filters rely on knowing the process and measurement noise. Fat chance our customers would know that.
We use Luenberger observers. The K gains for the Luenberger observer determine how quickly the observer will correct for errors. In the case of our motion controller you choose a frequency that appears to work.

No matter which method you are using, H-inf, Kalman, Luenberger observer, alpha-beta-gamma filter, the most important thing is the model that determine the transition matrix. If the transition matrix is bad then nothing works well.
Therefore the key to making any of these method work is system identification. System identification is necessary to compute the open loop model of whatever you are trying to control.

If you have a good model then a Luenberger observer will work well.
Peter Ponders PID - Observers/Estimators - YouTube

Kalman filters are ideal if you know the process and measurement noise but in most industrial applications process and measurement noise are not known. In which case a Luenberger observer or alpha-beta-gamma filter will do.

 

The sampling time for a Kalman filter is usually the same as for your closed loop control. I/we don't use Kalman filters for motion control because Kalman filters rely on knowing the process and measurement noise. Fat chance our customers would know that.
We use Luenberger observers. The K gains for the Luenberger observer determine how quickly the observer will correct for errors. In the case of our motion controller you choose a frequency that appears to work.

No matter which method you are using, H-inf, Kalman, Luenberger observer, alpha-beta-gamma filter, the most important thing is the model that determine the transition matrix. If the transition matrix is bad then nothing works well.
Therefore the key to making any of these method work is system identification. System identification is necessary to compute the open loop model of whatever you are trying to control.

If you have a good model then a Luenberger observer will work well.
Peter Ponders PID - Observers/Estimators - YouTube

Kalman filters are ideal if you know the process and measurement noise but in most industrial applications process and measurement noise are not known. In which case a Luenberger observer or alpha-beta-gamma filter will do.

Thank you! What do you think about rule that for example position controller should 5x times slower than speed controller? Speed controller should be of course faster than position controller, but do you think that it should be at least 5 times faster?
And as I understand it is desirable that kalman filter kalman filter would have also 5 x times higher bandwith than controller. Am I right?

 

Who's rule? Where do you see such rules? if you have a separate feedback for velocity and position then you might try a cascaded loop but usually there is only a position feedback for position. 1 millisecond loop times is OK for position and velocity. However if you are doing testing and need to generate a 100 Hz sine wave then 1 millisecond updates only allow for 10 updates per sine wave. That is a bare minimum. Then it would be better to increase the sample rate to 4 KHz or 250 microseconds between updates so there would be 40 updated per sine wave. The minimum frequency of updates is dependent on what you are trying to control. There is a down side to sampling at a high frequency and that is that the feed back resolution needs to be finer to reduce quantizing error.