We can also use the series directly for the Allan Deviation estimator of choice in either time or frequency form. Thus, the lack of zero dead time does not necesserilly prohibits the use, but care in setting up the signals and I/O can curcumvent the problem.

Depending on the system, just loosing signal or when the expected holdover breaks some estimated frequency or time deviation limit should qualify as the reason. The reasons for those decissions should be logged so that post-mortum analysis can be done. > Spoofing the time from a remote location seems impossible. Or is it just > difficult?

> Considering that Allan deviation estimation has problem of its own, this > scale error is not significant. What you do need to check is that the > relationship between intended gain and injection gain is sufficiently > different.

At this point a rough estimate is good enough, is the GPS bouncing around at the 100 meter level or the 1m level? I tried using an old Garmin 12 GPS with NMEA only NTP driver and got about 10X worse then your result. Basically the "12" was useless.

If this is an off the air signal, it's noise is probably not good enough for you to estimate it accurately enough in that time frame. If this is FSK, then modulation theory gets in the way a bit. 150KHz instantaneous shifts generate a lot of energy at the shift point. You also need to worry about dynamic range in addition to signal to noise.

Appendix B of that doc says that you need to be able to measure phase to 0.1 degree for state estimation, stability monitoring, control, and relaying. > > B. There is no phase adjustment for the generator. Phase angle with respect > to the grid is determined by whether you are giving or taking power with > respect to the grid.

Hi If: 1) You are after better than 1.0 x 10^-13 accuracy 2) You are getting 1 to 9x10^-9 at one second ADEV out of your GPS 3) You have a telecom Rb with 1 to 5x10^-10 temp coef over a 70C delta 4) Your Rb self heats 20 to 30C in still air Here’s some math: You will need at least 10,000 seconds to get a single frequency estimate and likely 100,000 seconds

Using a phase detector, just differentiate the phase data to get a frequency difference estimate, however phase-wrapping usually upset the result. In digital processing avoiding the wrapping upsetting the result is relatively trivial.

Indeed yor estimation is correct. The Q-factor of the tiny tuning-fork crystals @32.768 kHz is typically between 20,000 and 50,000.

Plot the estimated variant and also plot the difference. As Jim Barnes used to teach, always check the whiteness of matching residues! Cheers, Magnus On 11/23/2016 09:58 PM, Lars Walenius wrote: > Bob, > I have to ask about the B-term. In the paper that Scott started this with I see that B was 4.45.