Morgan Olsson wrote:

> Gee...  I mean:
>
> If injected > median, trash lowest
> If injected < median, trash highest

Ahh! That's the key point I was missing. I was assuming that the
'trashing' would occur on the opposite ends.

The only kind of signal that could cause this type of filter to 'fail'
would be one oscillating with an amplitude larger than the current
extremes. In which case, you can safely make the argument that the
filter is not 'failing', but actually filtering out the noise.

But OTOH, in some applications you might have to ensure that this
doesn't happen. For example, I could envision a situation where an
incoming signal is very steady and the filter is filled with this
quiescent value. Then the signal slowly changes its value while at the
same time some 'noise source' is introduced. The DC/median component
which we wish to capture is changing. However, because of the
oscillations due to the noise, the filter's median value remains
unchanged. (Each oscillation pushes the previous extreme off of the
sample stack.)

In your case (temperature monitoring), this scenario wouldn't occur. So
the filter (or I should say the MOMA - the Morgan Olson Median
Algorithm) would work beautifully. In my case, the samples were all over
the place (lots of high frequency noise). It really boils down to: Know
your signals and know your noises!

Scott