[sdiy] ADC filtering (was Re: A/D converter with good accuracy for a reasonable price?)

[Tom Wiltshire]

On 6 May 2011, at 00:43, Scott Gravenhorst wrote:

> jon schatz <jon at divisionbyzero.com> wrote:
>> On 5/5/2011 3:02 PM, Scott Gravenhorst wrote:
>> 
>>> If you are averaging (for example) four ADC readings, you collect them,
>>> add them together and divide by four (or right shift by 2 bits).
>> 
>> you can also do some exponential filtering:
>> 
>> newResult = ( prevResult * (n-1) + curAdcReading) / n;
>> 
>> which has the advantage of giving you a new result every cycle instead 
>> of every n cycles (which the above averaging would do). it also means 
>> that you'll never get a change greater than 1/n * resolution which can 
>> be a good thing for reducing noise but a bad thing if you want to track 
>> sharp changes.
> 
> Yeah, I was thinking that too after I hit send.  The filter is a single pole lowpass
> IIR filter and is analogous to a simple RC lowpass filter.  The higher the value of n,
> the lower the "bandwidth" of the filter, that is more suppression of noise at higher
> frequencies.  So it really depends on what you need from the sample stream.  I think
> that for reading the value off of pots, the exponential filter method can be nicely
> advantageous, and as Jon points out, there is a loss of slew rate, but for pots, how
> fast can you turn them anyway?  
> 
> -- ScottG


+1 for averaging as a useful technique for microcontroller on-chip ADCs.

I like to store the past values (say the last 16) and a running total. Then you can get a higher resolution averaged value (e.g. with reduced noise and no abrupt changes) by simply taking off the oldest input value and adding on the new input value.

If you had a 12-bit input, you could add 16 of them together and get a 16-bit output. It won't be 16-bit accurate, but it'll never vary by more than 1/16th of the value. This is the "loss of slew rate" already mentioned, and it's often pretty handy. It's much better than upshifting the raw 12-bit ADC value to 16-bit since this also multiplies any jitter and errors, which is often the alternative. It's a very basic, dirty sort of digital filtering, but it works usefully well, and if you don't have one-cycle MACs for real filters it might be the best you get (one add, one subtract, maybe a shift if you don't need 16-bits).

HTH,
Tom