[time-nuts] Low-Cost Rubidium Performance
magnus at rubidium.dyndns.org
Thu Feb 9 16:42:38 UTC 2012
On 02/09/2012 04:10 PM, Jim Lux wrote:
> On 2/9/12 4:51 AM, WarrenS wrote:
>> ADEV is for random freq variation not easily measured by other means.
>> Temperature fluctuations do not cause random freq changes and the
>> temperature's effect should be removed if one wants accurate long term
>> ADEV numbers.
>> Even daily diurnal cycles due to temperature can have major negative
>> effect on ADEV numbers as low as 2000 to 3000 seconds,
>> and if there is an Heater or AC cycling, then any ADEV numbers about a
>> few hundred seconds can be due to TempCoeff, which should not be
>> measured with ADEV or included in ADEV plots.
>> This is much the same as a single outlier data point that can screw up
>> the whole ADEV plot and make it pretty much meaningless and unrepeatable.
>> Ditto for linear ageing, Should be remove first if one wants true ADEV
> Interesting point you make here. The rising ADEV at 100-1000 second-ish
> tau in a system that should be better is a classic sign (at least around
> here) that temperature effects are showing up.
I regularly see the building AC at 900-1000 s for instance.
> However, how could one remove that effect from the raw data? And isn't
> the measurement of the "system", which includes the environmental effects.
ADEV and friends is there to handle random sources, where as this is a
> I suppose you could run your widget in a temperature controlled chamber,
> get those numbers. Then run it in a less controlled benchtop
> environment, and get those numbers, and claim that the difference is
> But at some point, what you're interested is the performance of the
> system in the environment in which it will be used. If you need good
> ADEV performance at the 1000 second tau, then you need an oven, a vacuum
> bottle, or a better design that's less environment sensitive.
You could also build active systematic effect predictors to lower this
By doing proper logging of key environmental effects, build a model for
how the dominant variations will systematically affect the signal and
then remove that from the measurements you get a better random jitter
Frequency drift of an oscillator is one such systematic effect. If it
where linear, processing it with ADEV would cause a sqrt(2) scale error.
Also, it would not give you a good prediction since usually you follow a
A*ln(B*t+1) curve which isn't matching the requirement, so you will only
get first degree compensation of that with HDEV style measures.
Temperature variations is tricky to say the least.
When you have random and systematic effects, separate them and estimate
them separately and then build a combined prediction from these models.
Random jitter and deterministic jitter are two such aspects. Same
applies at longer taus as well.
More information about the time-nuts