Bob Lunn
09/24/97 02:24 PM


> Anybody got a good averaging algorithm?

     Lawrence, you've had a number of good responses
     (especially the suggestion to use a median filter,
     though this depends critically on your sample rate,
     spike width, etc.), but not much directly addressing
     your original question.

     First, keep in mind that because you have a continuous
     data stream you are actually interested in filtering
     rather than averaging.

     Casually, we often talk about 'rolling' or 'running'
     _averages_ when we really mean an IIR (infinite impulse
     response) filter.

     Take the simplest case:

          val = (val + sample) / 2

     This requires only the calculated (running) value, and
     the latest sample.  You can also see why this is called
     an INFINITE impulse response filter.  A very large
     sample can effect a large number (potentially infinite
     number) of subsequently calculated values.  The response
     of the filter to an impulse can extend indefinitely.

     Take the next simplest case:

          val = (val + old_sample + sample) / 3
          old_sample = sample

     Take the next simplest case:

          val = (val + oldest_sample + old_sample + sample ) / 4
          oldest_sample = old_sample
          old_sample = sample

     You get the picture...

     Now consider FIR (finite impulse response) filters.

     Take the simplest case:

          val = (old_sample + sample) / 2
          old_sample = sample

     You can see the difference.  The calculated value
     depends ONLY on the two latest samples.  A very large
     sample can effect only the current value and the next
     calculated value.  After that, the large sample (spike)
     no longer influences the calculated value.  The response
     of the filter to an impulse extends finitely.

     Take the next simplest case:

          val = (oldest_sample + old_sample + sample ) / 3
          oldest_sample = old_sample
          old_sample = sample

     Again, you get the picture...

     Now, in all the preceeding examples each sample is
     assigned the same 'weight'.  If you really want to
     get into filtering you need to assign different
     weights to the samples and from this derive different
     filter characteristics.

___Bob