[time-nuts] ADEV ScaleFactor correction

Magnus Danielson magnus at rubidium.dyndns.org
Fri Jan 17 17:50:06 EST 2014


On 16/01/14 07:54, WarrenS wrote:
>
> ADEV is a great tool for measuring random noise but not so good for systematic errors like ageing drift.
> I have seen on the better oscillator's that the typical ADEV rise with time is often the effect of changing temperature and/or ageing drift.
>
> I did an experiment to determine what scale factor correction is needed for determining an oscillator's linear freq ageing drift rate over time using ADEV / MDEV data.
> I used a good HP10811 osc whose drift rate when left on continuously is well below 1e-10 / day with an ADEV / MDEV of around 1e-12 between 0.1 to 100 sec.
> This osc has not been powered up for a few months, so it's initial turn on drift rate was very high at ~1.3 e-8 / day.
> By plotting both its actual drift rate change and it's varying ADEV as the unit is restabilizing, I found that the oscillator's actual drift rate was equal to 1.4 times the ADEV value.
>
> That is with an ADEV value of 1e-10 at 1000 seconds, the oscillator's actual frequency ageing drift rate at that time was 1.4 e-10 per 1000 seconds.
> The 1.4 scale factor correction worked in this case from below 100 seconds to greater than 10K seconds.
> Of course this scale factor only applies when the oscillator's drift rate is constant and is the major error source for the given ADEV time period and data run.
> The same scale factor also works using MDEV data, because in this case, the MDEV values are the same as the ADEV values at time periods when ageing drift is the major error source.
>
> Attached are two TimeLab plots that I used to find the scale factor correction, showing the ageing drift rate and the changing ADEV values as this oscillator  restabilizes.

The linear frequency drift causes an ADEV linear ramp of D*tau/sqrt(2) 
as found here:

http://en.wikipedia.org/wiki/Allan_variance#Linear_response

For raw measures being uncompensated of oscillator systematic drift, the 
drift will indeed overshadow the upper end of the ADEV/MDEV plot.
It turns out that below the linear frequency drift is higher terms, so 
just removing the linear drift does not completely remove the systematic 
components. ADEV was only means for the noise components, so systematic 
components should be separated and analysed separately. Their confidence 
bounds is way different.

I try to let oscillators sit powered up as long as possible to reduce 
the drift from my measurements.

Cheers,
Magnus


More information about the time-nuts mailing list