1
$\begingroup$

I'm currently trying to reduce the noise in/smooth sampled signal strength data. The context is that I collected data from a transmitting device using bluetooth - the available data is basically a series of timestamps and the receiving signal strength.

The problem: The signal strength fluctuates quite strongly, which is why I want to smooth my data. Actually the data is not equally spaced - some signals get lost, the transmitter has some quirks leading to irregular periods in which signals are sent etc. - but I could aggregate them on a fixed time-period, if that's necessary / recommended.

I read some articles but are still not sure how to approach this. What kind of filter would probably lead to the best results? Kalman, Savitzky-Golay, adaptive, exponential...many to choose from and not enough experience to do so.

What would you recommend?

Thanks for your time!

| improve this question | |
$\endgroup$
0
$\begingroup$

Between these I only have experience with Kalman filters. They don't need equally spaced measurements but you can make it a lot easier with it. Missing measurements are really easy to cope with.

If you want a simple solution and you think that the most likely outcome of your next measurement is the same as the last one (there are no 'dynamics') you could also interpolate your missing values and use a linear filter.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Well, there is a certain dynamic involved - the transmitter is attached to a person moving around freely. I never know in which direction though. I looked at Kalman Filters but struggle to formulate a mathematical model depicting a person/transmitter moving around while only knowing the signal strength/estimated distance from transmitter to receiver, but not the transmitter's actual position. Do you have a tip regarding that? $\endgroup$ – justonemorething Jun 4 '15 at 21:45
  • $\begingroup$ Standard Kalman filters assume some sort of random walk which won't be ideal for tracking people. If you go this way I suggest using multiple-model Kalman filters (IMM for example) so you can catch some sudden 'model changes' from stopped to moving or whatever. Evolution matrix can be only position/speed, and the measurement noise you will know better how to model, make the variance vary with distance or something. First Google result for IMM Kalman filter is already a paper about human tracking citeseerx.ist.psu.edu/viewdoc/… $\endgroup$ – gsmafra Jun 4 '15 at 22:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.