[time-nuts] GPSDO & Crystal Aging
magnus at rubidium.dyndns.org
Sat Apr 12 18:25:16 EDT 2014
On 11/04/14 21:38, Chris Albertson wrote:
> Look at what NTP does to select "good" clocks when it has many to choose
> from. It does not simply average them.
> It looks at the noise in each one and then sees which clocks have
> overlapping error bars. It assumes that all good clocks have the same time
> within limits of their precision. Then from the good clocks there is a
> second level weeding out process then finally it does a weighted average of
> the remainders where I think those with less jitter get more weight.
> It would not be impossible to do this with 10MHz oscillators. Certainly a
> simple average is not a good idea as a broken unit can pull the entire
> average way down. I think you'd have to check reasonableness first and
> eliminate outliers I think today you might simply digitize the signals
> and figure out which were best using software.
> In short the output is "ensemble time" (not "average time") but there is a
> careful selection of who is allowed to be member of the ensemble.
NTP uses the ensamble clock style that Dave Allan developed for the NBS
AT time-scale and originally programmed on a PDP-7. Applying this type
of phase-comparison, estimate stability, weighing and updating ensamble
stability should indeed be possible to do. You need three or more
clocks, but one of these can be the GPS when you have it.
Jim Gray pointed out that it is important to watch your data. At
NBS/NIST they started to see some 1/f⁴ noise on one of their standards.
They could not figure it out. Turned out that the cleaning-lady was
pushing the standard over the floor once a week in order to clean under
it. This systematic "noise" where not in their standard model, but they
A frequency jump on the crystal oscillator in a control-loop will be
tracked in eventually, so it will look more like a phase-spike than a
frequency jump. Atomic clock FLLs will however not track in the full
phase difference, at least not guaranteed to do.
> I used a joke last week to explain to a class why we don't use averages,
> with no other qualifications. The joke is "Bill Gates walks into a bar....
> What's the average net worth of everyone in the bar? Maybe $250 million."
> My point was that it is hard to describe a population that is not Gaussian
> distributed. "Stuck" and jumping crystals are not Gaussian. You'd have to
> detect the misbehaving devices.
More information about the time-nuts