[time-nuts] Re: GPSDO Software & Performance

[Poul-Henning Kamp]

--------
Pluess, Tobias writes:

> However in the other plot "gpsdo-outliers" we can see that there are huge
> outliers in the phase error. At these times, we had huge thunderstorms here
> in Switzerland at my place and I believe that the GPS signal quality was so
> bad at these times that the GPS time pulse did not work reliably. (what
> other reasons could there be for these 1PPS timing outliers?)

Thunderstorms, in particular in regions with high ground resistance
(= mountains) causes electrostatic fields which can cause trouble.

One of the lesser known kinds of this trouble is induced voltages
in high impedance CMOS inputs.

Therefore: make sure that all unused CMOS pins are either tied
to something with a suitably low resistance.

On microcontrollers one can configure unused pins as outputs and drive
them low.

> Obviously, the "magic" of a GPSDO control algorithm is in the filtering of
> these outliers! so far, my control code (src/cntl.c on Github) does not
> seem to filter these outliers good enough

In general median filters are probably the most robust way to deal with
outliers, but if you do them the naiive way, they cause a usually
unacceptable time-delay.

When the outliers are less than half of samples, and always wrong,
running the median filter in "gate mode" works well.

Since the output from the median filter only defines a "gate" which
input signals must pass, it causes no time-delay in the control
loop.

Something like:

        NMEDIAN = 300
        GATE = 1e-3

        median_filter = [None]

	def new_measurement(x):
            median = sorted(median_filter)[len(median_filter)//2]
            if median is None or median - GATE < x < median + GATE:
                ok_measurement(x)
            if len(median_filter) >= NMEDIAN:
                median_filter.pop(0)
            median_filter.append(x)

This will reliably discard any sample which are more than 1e-3 from
the median of the previous 300 samples.

If the outliers are sometimes true, ie, where step-changes are
to be expected and sensibly handled, median filters are not
so swell and Kalman filters may be indicated.


-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.