Replies to This Discussion

DeveloperPermalink Reply by Ryan Beall on June 1, 2010 at 8:17pm

You are spot on, I'm not really following it in the code either. The ISR and the way it is laid out makes it hard to follow but i'm seeing what you are seeing. Doug? Jose?

• ▶ Reply

Permalink Reply by Lew Payne on June 1, 2010 at 11:23pm

See this post for real-world details as applied to MEMS sensors.

• ▶ Reply

Permalink Reply by hopslink on June 2, 2010 at 1:32am

Dave,
The method you describe is necessary to increase resolution correctly if you work with binary integer data. The ArduIMU code sidesteps that problem by converting accumulated data to a floating point number before averaging (see the cast in the code).

Both methods increase bit width (necessary for extra precision) and both are correct, just select the method that works best for your data. ArduIMU does subsequent floating point operations on data so converting to floats early is logical.

• ▶ Reply

DeveloperPermalink Reply by Ryan Beall on June 2, 2010 at 12:31pm

Is there any documentation or explination for this? I can grasp it, I just like to have solid foundations for my many "black majic assumptions" when it comes to the wonderful math and theory of digital signal processing!

Thanks!

• ▶ Reply

Permalink Reply by hopslink on June 2, 2010 at 4:28pm

For oversampling and averaging, the effective increase in resolution comes from the assumption that your signal has additive noise spread evenly across a far wider bandwidth than your signal. Using a lowpass filter you select your signal and any noise that the filter passes from the rest of the spectrum. Noise power from your filter passband is spread over the spectrum (Fsample/2). Oversampling has the effect of spreading noise over a greater bandwidth, effectively improving your signal to noise ratio. This is a naive explanation and includes assumptions about signal and noise (but these hold or can be made true in many situations). For a solid foundation any text on DSP should cover this in detail.

In order to exploit increased resolution (SNR) in a binary system you need more data bits to represent the signal. The answer I gave above was directed more toward the mechanics of increasing bit width, since that seemed to be the drive of the original question.

• ▶ Reply

Permalink Reply by Jonathan Lussier on June 2, 2010 at 2:00pm

probably the method of bit shifting while an integer is quite a bit faster then doing FP Add / divide...

• ▶ Reply

DeveloperPermalink Reply by Ryan Beall on June 2, 2010 at 3:19pm

Yeah, yet another reason to do everything in integer format as much as possible.

• ▶ Reply

Permalink Reply by hopslink on June 2, 2010 at 4:51pm

Yes, but it also limits you to integer bit width increases. So you need to oversample 4x for 1 bit, 16x for 2 bits etc. If you can't get enough samples in a time period you can't make the gains.

Using fp you can average any number of samples together and each gives a fractional increase in effective resolution. So if you oversample 14x you can get an effective increase of 1.7 bits resolution (value plucked from ether for illustration, there is a formula but not in my head at this time of night :P) vs 1 bit for the integer shifting technique.

• ▶ Reply

DeveloperPermalink Reply by Ryan Beall on June 2, 2010 at 5:21pm

awesome, I like it!

• ▶ Reply

Permalink Reply by skystalker on August 12, 2011 at 1:06am

en..I think the floating operation is more suitable for the DSP

• ▶ Reply

DeveloperPermalink Reply by Ryan Beall on June 2, 2010 at 12:40pm

Another great read

http://www.mdpi.com/1424-8220/9/5/3767/pdf

• ▶ Reply

Permalink Reply by skystalker on August 12, 2011 at 1:31am

here is another post of the oversample technique from microchip with the document and sourecode

http://www.microchip.com/stellent/idcplg?IdcService=SS_GET_PAGE&...

• ▶ Reply