quartz drift rates, linear or log
Thanks for all the comments on this thread. Here is the first set of replies:
Don Latham:
Interesting, Tom. I don't think I see any of those pesky grain boundary
shifts or readjustments in the lattice structure? If I remember, these
can cause instant shifts in frequency that do not heal?
Don
In this experiment I was more interested in long-term frequency. I think the frequency jumps you speak of may be at a finer level than I was measuring. They are easier to see if you use a TimePod and continuously collect data every second. They are harder to spot if you use a frequency counter and collect data once an hour.
That said, you can see that CH03 exhibits more spikes than any of the others and it will be set aside for a closer look. The nice thing about this stage of my time nut hobby is that I am no longer looking for the best oscillator. Now oscillators that are weird are interesting to me.
Attila:
These look like textbook examples of random walk frequency modulation.
As this is a random process it is not surprising that they look different
for each oscillator.
Yes and no. Look again at:
http://leapsecond.com/pages/tbolt/TBolt-10day-fit0-e10.gif
Many of the plots differ in appearance and it's clearly not always random walk. Each oscillator's drift rate, short-term noise (hours) and long-term noise (days) contributes to the appearance of the plot.
But, when you apply a 10-day linear fit to each DUT, you get a new plot. Attached. Also at:
http://leapsecond.com/pages/tbolt/TBolt-10day-fit1-e10.gif
And then, yes, the dominant noise is very much random walk FM. Still, there is a variety in performance: some more noisy than others. This shows up nicely in the HDEV plots or drift removed ADEV plots. Those are important because when an OCXO is used in a GPSDO, the linear-fitted residual data is much more relevant than the raw data.
To put it another way, the first plot above makes you think some of these OCXO will make a far better GPSDO than others. But the second plot shows they may all be quite similar. The 2nd order term of the loop enables this.
Attila Kinali:
Do you know what caused the frequency jump of CH18?
Good eye. I think that was me doing EFC testing. I can check my notes. It's not the TBolt or OCXO.
Bob Stewart:
So, are you measuring OCXO stability or EFC stability?
Bob
I measured the 10 MHz coming out of the BNC connector. The TBolt's are free-running (no GPS, no disciplining) and DAC/EFC is forced to 0 volts to reduce any impact it may have.
I have also done EFC testing on each unit; that's a separate report for later. In this thread I mostly just wanted to show examples of log and linear drift and to convey that long-term (days, weeks) logarithmic trends are effectively linear trends over shorter-term (hours, days).
Tom Miller:
Just out of curiosity, what is the age of each of these Tbolts? (i.e. date codes?)
These TBolt's are all from the original 2008 TAPR group buy. The Trimble date codes range from 2002 to 2005. These are US-units, not Chinese eBay imports.
Charles Steinmetz:
When the purpose is correcting a GPSDO local oscillator during holdover,
it depends on how long one expects to trust the corrected frequency.
Practical realities make it pointless to trust corrections longer than a
day or so, if that long. At that scale, these all look pretty linear.
I agree. See especially the new linear-drift-removed plot I posted for Attila above (and attached).
Jim Lux:
I think this is a general rule: when they make precision oscillators
(USOs) for spacecraft, they make a fairly large batch (a couple dozen)
to some intermediate level of assembly, and then they run them for a
while and watch them, and from that, they pick the "good" ones. (where
good is somewhat mission dependent).
Right. Many of us time nuts have done the same with 10811A oscillators. You just keep buying them here and there over the years, comparing them, and keep the unusually good one that's down in the low -13's.
/tvb
Hi Tom,
Fascinating when you've done that linear fit - many of the plots now
look very similar, suggesting environmental conditions? From that it would
now be nice to log temperature, pressure, humidity, (& mains voltage?), and
see if there is any correlation there.
Wonderful to see plots of such a large group - well done!
Peter
On 13 November 2016 at 13:05, Tom Van Baak tvb@leapsecond.com wrote:
...
But, when you apply a 10-day linear fit to each DUT, you get a new plot.
Attached. Also at:
Tom et all
While our instinct based on some "pre- knowledge" of the aging and drift
processes is to try and fit these to linear or logarithmic curves there
is a third possibility . That is, in fact the aging is not exactly
either and may be better represented in fact be some kind of polynomial
curve. The fact that there may be more than one drift and aging process
at play here would also fit this hypothesis. Ii makes my head hurt to
think about how one would derive the polynomial. Following this thread
further ( and not to discourage your endeavor) but the entire history of
each Xtal may be more of a factor than we initially surmised as well.
Each xtal is at a different point in its journey through TIME and its
history may have as much or more to do with how it behaves at this point
in TIME than we can characterize
Statistically it is also advisable to throw out (from the curve fitting
exercise anyway) unusual units that are clearly not like all the other
kids since they are clearly marching to a different drummer and for the
purpose of this exercise are adding to the NOISE of the analysis 8^)
Dave
Hi Tom,
Fascinating when you've done that linear fit - many of the plots now look very similar,
suggesting environmental conditions? From that it would now be nice to log temperature,
pressure, humidity, (& mains voltage?), and see if there is any correlation there.
Wonderful to see plots of such a large group - well done!
Peter
That's correct. Yes, I am also logging environmental parameters. Stay tuned for that. Also, you may have noticed the "thermal events" that I deliberately caused in the lab once in a while to make environmental correlations more obvious.
The end goal is not only to extract the approximate tempco of each unit through correlation but also to post-process the data to partially back out the temperature component and produce a set of second residual plots. So you go from raw data, to measuring drift, to removing drift, to measuring tempco, to removing tempco. At this point you get closer to revealing the intrinsic performance of the oscillator.
I did the same for "Clock B" last year:
http://leapsecond.com/pend/clockb/2015-tvb-Greenwich-ClockB-ppt.pdf
http://leapsecond.com/pend/clockb/
In fact you don't really need to go through the plotting and fitting steps. There are standard techniques to fit N-dimensional models to data. So a frequency and temperature time series goes in -- and best fit linear drift and temperature coefficient estimates comes out. This will be automated, but for now I like the plots because they are more educational. Also the eye is extremely good at spotting interesting or unforeseen things that math and statistics are blind to.
/tvb
Hi
On Nov 13, 2016, at 9:34 AM, Artek Manuals Manuals@ArtekManuals.com wrote:
Tom et all
While our instinct based on some "pre- knowledge" of the aging and drift processes is to try and fit these to linear or logarithmic curves there is a third possibility . That is, in fact the aging is not exactly either and may be better represented in fact be some kind of polynomial curve. The fact that there may be more than one drift and aging process at play here would also fit this hypothesis. Ii makes my head hurt to think about how one would derive the polynomial. Following this thread further ( and not to discourage your endeavor) but the entire history of each Xtal may be more of a factor than we initially surmised as well. Each xtal is at a different point in its journey through TIME and its history may have as much or more to do with how it behaves at this point in TIME than we can characterize
Having fit a few (quite a few) OCXO’s and TCXO's to various curves … the log curve in 55310 is about as good as any you can use. Polynomial curves will (in general) give you a real mess … The most common outcome is that your time interval is to short / your data to noisy / your aging to low to get a real fit with a good confidence estimate. The statement “as good as” should be taken in that context. The only real way to validate the fit is to go ahead and run the parts for another month or year to see what happens over various time periods. Because of the inevitable noise in the data, the curve fit is a bit tricky….
Bob
Statistically it is also advisable to throw out (from the curve fitting exercise anyway) unusual units that are clearly not like all the other kids since they are clearly marching to a different drummer and for the purpose of this exercise are adding to the NOISE of the analysis 8^)
Dave
--
Dave
Manuals@ArtekManuals.com
www.ArtekManuals.com
time-nuts mailing list -- time-nuts@febo.com
To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts
and follow the instructions there.