Materials: [ Cód.: SensorTSTKlmVsDifFinAPPROX.mlx ] [ PDF ]
This video concludes the case study and Matlab examples from the videos [
But, well, 99% of the students are told that if we have a continuous process of order 2, it is discretized and a Kalman filter of order 2 is made, not of order 3 as we did in the videos just mentioned above. This video shows that, indeed, if the sampling period is small enough (if the signal to be acquired has a very small bandwidth compared to the Nyquist frequency of ), a digital filter of the same order as the process to be observed also approximates well enough the performance of the continuous-time Kalman filter, justifying why all the details of the augmented models in the previous case study videos are omitted in presentations for undergraduate students on these topics.
The last section of the video discusses the fact that the ‘averaging’ time on the data acquisition card before having a valid sample may be s, but if we are controlling an oven that takes hours to heat up, perhaps the sampling period of a temperature control PID is 2 minutes. Therefore, the measurement noise variance will depend on that first interval that we call (acquisition time) and not on the oven PID sampling period . Obviously, the theoretical optimum will be both and (with ) tending to zero; however, as we have seen in our Matlab case study, we can approximate the idealized continuous-time situation reasonably well with approximate digital processing that has much less computational load requirements.
*Link to my [ whole collection] of videos in English. Link to larger [ Colección completa] in Spanish.
Colección completa en ESPA
Previous Measurement, optimal filtering (2/3): discrete-time vs continuous-time Kalman filters, example (Matlab)
Next Normas de se
Full collection in ENGLISH [View all]:
Previous Measurement, optimal filtering (2/3): discrete-time vs continuous-time Kalman filters, example (Matlab)
Next Bayesian Optimization motivation (1/4): problem statement and model classes