DMTD Tutorial

[5065AGuru]

OK, Follows (in several editions!) the theory of and how to operate a DMTD system to get the best results out of it.



Once hooked up as shown and you have the PC showing the counts we can start to take data.
In the following example the RF inputs are 5MHz, the offset is 1Hz, and the counter is an 8 digit counter with a 10ns resolution. Also I am taking data with HyperTerminal and using Ulrich Bangert's PLOTTER program.

On the PC screen you should see the counts occurring once per second and the value will be either incrementing or decrementing.

0.xxxxxxxx

Adjust the fine frequency of the DUT or the REF  to slow the rate of change and get into the low to midrange of the scale.

Now capture the data to a file for lets say 15 Minutes.

Then open plotter and make sure you input 1.00 into the Tau-0/s box
Select file and open and select the file you just saved via HyperTerminal.

Here is an example:


This is a 4K+ Second file of a HP5065A Rubidium versus an FTS 1200 Quartz.

This shows how the phase changed over time.
Full scale is 200nS and the LSB value is 2X10-15th.

Now if you want to plot the Allan Deviation you will need to remove the 2 wraps shown.
To do this click on "all steps"
Then select "I have no idea" and OK.

You should now see the full 4K+ plot without wraps.
One more step before you can compute the AD.
Click on Series on the top tool bar and then enter a name (I just enter a single letter lets say C), and the formula.
For this example the formula is 2.0e-15*x1 .
the 2.0e-15 represents the LSB value for the particular counter used.
You should now see a black line added to the plot and the box to the right now shows an added plot called C.
Deselect the Column 1 red plot and the black plot should expand to fit the screen.
Now click on Time Stability Statistics in the toolbar, select data is phase, use powers of 2, and normal Allan deviation reselecting Time Stability Statistics each time.
Finally select Time Stability Statistics one more time and select process.
You should now have the Allan Deviation plot! Shown is the plot for the examples data.

*******************************************************************************
One thing about the all steps, it does not do the best job!
I forgot this as I always manually remove them!
Instead of clicking on all steps do this.
Click on show single marker, then using the cursor position the marker so that it covers the left step.
Now click single step. You should see most of the step gone. Now using the cursor click on and out outline the remaining portion of the step. This zooms in. Position the single marker between the two vertical traces or on the single trace. If between two click +- 1 points until you are back to 1 line. Once there click single step again. Keep repeating until nothing happens when you click single step. Then click once on the +- 1 point button. Now click un-zoom on the tool bar. The right step will still be there. repeat the procedure to eliminate that step and un-zoom.
Now finally you can click on series and compute it.
Attached is a plot of the manually step removed Allan deviation so you can see the difference!
Blue is remove all (auto), and red is single step Manual)
Whew.
********************************************************************************


More follows over a couple days!

[5065AGuru]

Now let's examine the DMTD method of precise time interval
and Allan deviation measurements.
In this example we will talk about 5MHz oscillators and a 5.000001 or
4.999999 MHz offset oscillator. (OSO)
The DMTD system is composed of dual mixers with a common L.O. to both mixers.
The other inputs to the mixers are the RF from your reference oscillator (REF)
and the RF from the oscillator being tested (DUT)
The OSO noise will cancel out (with some caveats) allowing a better noise
floor versus a single channel mixer scheme.
The dual mixer filters the beat signals, triggering on the zero crossings,
and ends up with two square waves with a 1 Second period. These are input to
the start and stop channel of a time interval counter to measure the time
difference.
Mixing the RF frequencies down to 1Hz multiplies the resolution
greatly.
Here is an illustration. 1.00000000 is an example of a full scale count in
the 8 digit counter I use with my dual mixers.
This actually represents 200ns
1Hz is 2X10-7 so the 1 in the example represents 2x10-7. As you can see
the LSB therefore is 2X1-15 or 2 fs! (Theoretically!) If you are using a
53131 53132 or SR620 you will have an even more ridiculous resolution!
Thes bottom digits are in the noise and don't contibute to the plots.
All the counters however will work fine, and the software (Ulrich's PLOTTER
or John's Timelab), both free, will plot properly once you input the proper
scaling factor.
Some info on other ofsets and frequencies:

Inputs         offset             full scale      LSB 8 digit  LSB 10 digit  LSB 11 digit

5MHz              1Hz        1.x = 200ns   2X10-15      2X10-16         2X10-17
5MHz             10Hz        0.1x = 20ns   2X10-15      2X10-16         2X10-17
10MHz      1Hz        1.x = 100ns   1X10-15      1X10-16         1X10-17
10MHz     10Hz        0.1x = 10ns   1X10-15      1X10-16       1X10-17

So you can see that your data files will be 10X larger for 10Hz offsets (10 count/Sec)
and the wraps will occur more often as the full scale counter value is 10X less.
You are however able to get data for .1Sec Tau.

You can also see that an 8 digit counter is more than adequate!
We are working on the design of an 8 digit counter to go with the simple dual
mixer board and hope to have some data soon.

Cheers,

Corby

[5065AGuru]

Here are some more tidbits!

For 10Mhz with 1 or 10Hz offsets a good 10811 is fine, just keep it powered up all the time!
For 5Mhz at 1Hz there are many candidates.
Try to use the most stable one you can!

You will find in some explainations of DMTD a variable delay inserted into one of
the RF channels.
This is because the instability of the offset oscillator is perfectly eliminated only
if the two RF signals are exactly in phase! This is not realistic when making
realworld measurements, but is feasable when making noise floor tests.
As you get further from that point more noise gets through.
This is usually not important except for oscillators that have very good stability
at a Tau of 1 and below. I usually get around this by running two plots.
For instance I run one plot against my 5065A not worrying about the offset.
Then I run a plot against my FTS 1200, starting the plot at a low offset.
(by adjusting the DUT or REF frequency you can get the plot to start at the low
offset without needing a delay line.) This gives me a plot at high Tau to leverage
the excellent stability of the 5065A, and a plot against the FTS 1200 to use it's
good stability for the lower Taus.
I can then marry these plots to give a composite plot.

Cheers,

Corby

[5065AGuru]

One little mentioned DMTD parameter is Bandwidth.
As bandwidth increases the noise level increases.
This will effect the baseline noise of the instrument.
My NBS 106D has variable switched BW:

1Hz
3Hz
10Hz
30Hz
100Hz
300Hz
1KHz

This allowed optimum noise levels to be used.
The manual shows some plots illustrating this.
https://nvlpubs.nist.gov/nistpubs/Legacy/IR/nbsir75-827.pdf


Most others you will find have a fixed BW.
The simple dual mixer has a fixed BW of about 35Hz to allow use at 1Hz or 10Hz offsets.
Not that important unless comparing noise floors between different DMTD designs but good to keep in mind.

I'll post some details on the noise floor measurements later when I get some actual data together to share.

Cheers,

Corby

[notfaded1]

Thanks for this Corby.  This is exactly the kind of guide that you just can't find anywhere really minus reading papers and some old manuals.  I noticed in some old NIST lab pictures they had two different DMTD systems on a rack and right next to them was the SR620 used with them as the TIC.  I decided to add the picture here of them with their noise floor ADEVs.  It seems like with timelab or some other software we could get close to output like the TSC 5120A.  Two of the DMTD systems are on the rack in center on bottom two shelves.

Bill

[5065AGuru]

Here is some info on noise floor measurement. Once your system is running and the noise floor is tested once, you most likely will never need to do it again!

The noise floor of the dual mixer determines how well you can determine the performance of an
oscillator (DUT) compared to your reference (REF.
Also your REF stability has to be better than or at least equal to your DUT.
For example if your REF into the dual mixer has an AD at a Tau of 1 Second of 1X10-12th
then trying to measure a DUT that is capable of say 5X10-13th would not work!
In the case of both DUT and REF having 5X10-13th performance then your noise floor
should also be lower than or equal to 5X10-13th.
Remember this is at a Tau of 1 Sec. the noise floor drops as the Tau increases!
Since the offset oscillator noise only cancels out when the DUT and REF are exactly in
phase (coincidence) the way we measure the noise floor is to input a single oscillator
into both ports at the same time and induce coincidence.
A simple tee connector can be used but I prefer to input the oscillator
into a broadband resistive power splitter. Then the splitter outputs are applied to the input
ports. Then by swapping the two ports if needed and also inserting small delays into one or both
channels you get the count as observed on your PC to read the smallest possible value.
The PIX shows a typical setup and the SMA male to female adaptor mentioned.
For my 8 channel counter (with 5Mhz RF and 1Hz offset) it's easily possible to get down to a
count of say 15,000 and with time and effort to get a bit below 3000. For the 15,000 count
example this equates to an time interval of 50 ps. The delay through one of the
SMA male to female adaptors runs about 120 ps. Once you are happy with your delay you capture
a run of data say for 2 hours. The count will wander a bit due to temperature variations
but hopefully will not under-range. Then take the data file produced and run an Allan deviation
plot with it. This is your noise floor.

For a dual mixer the noise floor should really be called the effective noise floor and
defined by the ability of your oscillators to stay stable long enough for the measurement.
If you were able to get the count to be at 250 the noise floor might look very good.
However once you go back to the real world and input real oscillators it's a different story!
If you have two Oscilloquartz 8607 option 8 BVA oscillators, you might be able to get that
low of an offset and have it stay in that range long enough to get a useful count!
Remember it's only the lowest Taus that are effected by poor coincidence. Most measurements
on real oscillators can be run further from coincidence and still be effective.
In reply #2 I describe how to work around this to cover the low and high Taus.

Cheers,

Corby

[notfaded1]

So you're basically adding those M-F connectors to get the timing exactly right then?  To increase the delay you keep adding another piece is that basically it?  I've seen some delay boxes where if you open up the box the only thing inside is a coil of coax cable at a certain length.  Since one connector adds possibly 150ps I assume that's why you have so many using them on both sides until the time difference is as small as possible from the split signal on the TIC?  I'm really glad that it looks like I'll finally be able to totally figure this DMTD thing out in practice.  Reading about it and actually doing it are two different things.   

Can you explain the 8 channel counter (with 5Mhz RF and 1Hz offset) part?  What does the offset have to do with the noise floor test?  Don't you essentially want the exact same delay through the two inputs?  Or are you referring to the 1Hz offset on the middle input port where your offset oscillator goes on the DMTD?

Bill

[5065AGuru]

Bill,

For the 5MHz inputs with 1Hz offset part, I was using count values I measured using those inputs just to give an idea of what you would see and what the counts were worth.
For other RF inputs and offsets the counts will not have the same value as the example but since you still try to use RF connectors/cables to get the smallest delay it really does not matter what the absolute values are. Remember this noise floor test is seldom needed again once you are up and running!

I'll be adding some real world data soon of a nice Quartz oscillator and give a few more tips on how best to run your plots.

Cheers,

Corby

[5065AGuru]

Here is an example of a plot between an HP 5065A and a nice Quartz oscillator that has only been on for a day or so, and using PLOTTER as the processing program.
I usually set the REF and DUT into the Dual mixer inputs that causes aging to slow the phase difference over time from the initial value. (If it's going the wrong way just reverse the inputs)
The initial value is set by adjusting the DUT or REF frequency to get the counts in the low end of the range on the counter and slowly increasing. This allows the you to stay in the lower counts for the longest time.

In the 1st phase plot you can see the phase counts rise slowly and then turn around and accelerate as the aging kicks in.
This was an overnight plot (Seconds of time on lower horizontal scale) and as you can see the counts wrap eight times.
Since this is a plot of a Quartz you really don't need the data after the first wrap so we will remove the remainder.
You will see a blue marker that you position just to the left of the 1st wrap.
Then click on delete right.
The 2nd phase plot shows the remaining data that you want.
Then the scaling factors needs to be input. (in this example 2.0E-15th)
Click on Series and the compute new series box appears. Input a name (C in this example) and the scaling factor then OK.
Then deselect the Column 1 red box. This leaves the black plot of the new series. (the third phase plot)
Then click on Time Stability Statistics and select the values you want. You have to click on Time Stability Statistics before each change. Once happy click on process. (4th phase plot)
Now the (in this example) Allan Deviation plot appears. (last plot)
TimeLab can also be used.

Cheers,

Corby

[notfaded1]

Good Morning Corby-

So you're essentially selecting just the first part of the samples in the first graph before the wraps started?  Basically the 0-10000 part correct?  I did also download Ulrich's plotter since you used for these examples.  I'm more familiar with timelab although it looks like you can do some things with Ulrich's plotter that timelab doesn't as easily or even make available?

The ADEV for that oscillator looks pretty good on the DMTD considering it hasn't been on long!

Regards,

Bill

[5065AGuru]

Bill,

Yes! I just usually lop off the stuff after the 1st wrap for Quartz although if the oscillator has been on a week or so you might want to lop it off at the 2nd wrap and then eliminate the single remaining step to get a longer data run. Not so important for Quartz but for a Rubidium or Cesium a good long run is OK, although you may not get many wraps as they drift must less!
TimeLab has many goodies and I do like it but I'm much more familiar with PLOTTER. You will get the same results with either.
One caveat not mentioned is that when comparing a Quartz reference against another Quartz reference you need to be careful.
For the shorter TAUs Quartz on Quartz is great and I usually do a plot against my 5065A for the mid to long TAU then a companion plot against my good Quartz reference for the shorter TAU.
Now imagine you have two 10811 that have been running for a year or so and you decide to test them against each other. Once you match their frequencies to get minimum drift of the Dual Mixer counts you start taking data. Now much later you take the data and plot the AD. You might be pleasantly surprised when you see excellent data for the shorter TAU (which are real!) but you also see really good AD at the larger TAU. (which might not be real!)
Usually Quartz will start turning up on the AD plot due to aging. What can be happening is that both your oscillators are aging at the same or similar rate! This gives the erroneous impression of excellent long term stability! To see what they are actually doing at the longer TAUS you would need to plot against a good Rubidium.


Cheers,

Corby

[edpalmer42]

Hi Corby,

I have a few questions and comments on your use of Plotter.

Is there a reason that you used Allan Deviation rather than Overlapping Allan Deviation?  According to Wikipedia, the overlapping version has been accepted as the preferred version by the IEEE, ITU-T, and ETSI.  I think I read somewhere that these days, if someone says "Allan Deviation", it's generally assumed that they mean "Overlapping Allan Deviation".

I suggest that you upgrade to a newer version of Plotter.  More recent versions include Total Deviation which modifies Overlapping Allan Deviation to produce higher values of Deviation than Overlapping Allan for a given data set.  e.g. for a given data set, Overlapping Allan might give you Deviation values up to Tau = 100 K sec. while Total Deviation gives values up to Tau = 200 K sec. 

It should be pointed out that Plotter has an extensive suite of data editing tools.  You can automatically or manually unwrap datasets like your example.  Sometimes they work better than others, but they can save a glitchy, hard-to-repeat dataset.  You can also edit data points automatically or manually to identify and remove outliers.

If you don't like the way a chart looks, click on 'Chart Editor' and explore various ways to change the appearance of the graph.  Once you get it to your liking, you can save the data and/or the chart with the 'Export' option under the Chart Editor.

Unfortunately, the author of Plotter passed away a few years ago.  I believe the source code still exists, but it's not publically available.  It's also written in Borland Delphi.  Combine these and it's unlikely that there will ever be another version of Plotter.

For anyone who's interested, Plotter is available here:  http://www.ulrich-bangert.de/html/downloads.html

[5065AGuru]

Ed,

Hi. I'm aware of the other plot options but since I have tested hundreds of oscillators and instruments using "regular" Allan Deviation that's what I use.
So when I test an HP5065A to see if it meets my expected performance I have in my mind a good idea if it compares to others I have tested.
I might just try a newer version to see if the remove all steps function works better than the one I have.
I find MUCH better results removing them via single step one at a time.
As for longer TAUS, most of the stuff I test does not need to get much over 1 to 10K Seconds.

Cheers,

Corby

[notfaded1]

Hi Corby,

I have a few questions and comments on your use of Plotter.

Is there a reason that you used Allan Deviation rather than Overlapping Allan Deviation?  According to Wikipedia, the overlapping version has been accepted as the preferred version by the IEEE, ITU-T, and ETSI.  I think I read somewhere that these days, if someone says "Allan Deviation", it's generally assumed that they mean "Overlapping Allan Deviation".

I suggest that you upgrade to a newer version of Plotter.  More recent versions include Total Deviation which modifies Overlapping Allan Deviation to produce higher values of Deviation than Overlapping Allan for a given data set.  e.g. for a given data set, Overlapping Allan might give you Deviation values up to Tau = 100 K sec. while Total Deviation gives values up to Tau = 200 K sec. 

It should be pointed out that Plotter has an extensive suite of data editing tools.  You can automatically or manually unwrap datasets like your example.  Sometimes they work better than others, but they can save a glitchy, hard-to-repeat dataset.  You can also edit data points automatically or manually to identify and remove outliers.

If you don't like the way a chart looks, click on 'Chart Editor' and explore various ways to change the appearance of the graph.  Once you get it to your liking, you can save the data and/or the chart with the 'Export' option under the Chart Editor.

Unfortunately, the author of Plotter passed away a few years ago.  I believe the source code still exists, but it's not publically available.  It's also written in Borland Delphi.  Combine these and it's unlikely that there will ever be another version of Plotter.

For anyone who's interested, Plotter is available here:  http://www.ulrich-bangert.de/html/downloads.html

Thanks for the info Ed.  This makes me think of another question now... does anyone know what type of ADEV timelab gives?  Does it give the overlapping or total ADEV?  I'd be curious to know.

I totally get what Corby is saying.  He wants apples to apples comparisons and switching his main method would rule out true comparisons.  Thanks to many here this is all starting to sink in!  I like the ADEV graphs Corby makes where he uses the Quartz for the shorter time and the Rb for the longer time measurements and combines them into one graph where they overlap.

Best Regards,

Bill

[thinkfat]

TimeLab plots overlapping ADEV.

[5065AGuru]

OK,

Here I'll try to illustrate what I mean when I talk about keeping the counts in the lower region to get the best results at low TAU.

The noise contributed by the offset oscillator will only cancel out exactly when the DUT and REF are in phase. As the units drift and the counts depart from the lower region the noise from the offset oscillator will start to degrade the measurement!

First I'll show a plot showing the performance of 2 different offset oscillators.
One is an FTS 1200 and the other a Piezo clone of the 10811.


Then a set of plots showing the baseline plots for low/mid/and full-scale counts using the FTS 1200 and then the Piezo as the  offset oscillator.

As you can see both oscillators perform well in the lower region and show worsening performance as the  count region goes up.
The 1200 does give better results than the Piezo when the counts rise.

The Piezo looks a little better when in coincidence but I think I managed to get a better phase match for that particular plot.

Also includes a noise plot.

With care you can take a long measurement disregarding where you are in regards to scale and then take a shorter plot after getting the counts into the lower region.
On a long plot there may be wraps where the counts are in the lower region so just select those portions and then plot your AD.
************************************************************************************
Well as usual when you try to teach something you end up learning yourself!
I never took the time over the many years to quantify the degradation versus distance from coincidence.
I knew it existed and always made the effort to minimize the effect.
Anyway I just finished another set of noise floors using the NBS 106B with its built in delay line to try and pin down where it starts hurting the results.

Here is a chart showing what I found.

% of full scale delay         AD @1Sec
******************************
0                                    1.30-13th
.01                                 1.57-13th
.1                                   6.58-13th
1                                    6.82-13th
2                                    6.78-13th
3                                    6.62-13th
4                                    6.85-13th
50                                  7.07-13th
100                                8.75-13th   

As you can see staying in the lower .1% will allow testing to low -13ths, as you go above that to 50% you can only test in the 7X10-13th range.

Cheers,

Corby

[5065AGuru]

Update to chart with more data points!

% of full scale delay         AD @1Sec
******************************
0                                    1.30-13th
.01                                 1.57-13th
.02                                 2.66-13th
.03                                 3.84-13th
.04                                 5.19-13th
.05                                 5.61-13th
.06                                 6.02-13th
.07                                 5.59-13th
.08                                 6.35-13th
.1                                   6.58-13th
1                                    6.82-13th
2                                    6.78-13th
3                                    6.62-13th
4                                    6.85-13th
50                                  7.07-13th
100                                8.75-13th   

[edpalmer42]

If your REF and DUT sources were both 1 Hz different from the Offset, would a long data run show a 1 Hz signal on top of the ADEV graph as they swept through the different phase values?  And could you say that the actual ADEV value was at the lowest level of this signal?

[tkamiya]

I'm sorry, I am totally confused. 

Say I'm working with 10MHz input and 1Hz offset, what is the multiplication factor I need to use?  1 second difference represents 10E-7, right?  so 1 x 10^7?  I am using TimeLab.  This whole concept is escaping me.

[thinkfat]

What I'm not quite getting is the sharp increase in noise after roughly 1000 seconds. Why doesn't it slowly degrade, why the sharp rise? It's one magnitude more all of a sudden.

[tkamiya]

I can only imagine the graph you are looking at.  If it is like ones I've seen, it is slowly degrading....  it's just that it's a logarithmic scale so graph looks the way you see it. 

I'm having more basic problem...  HELP!  I'm being out smarted by a device I made myself! 

[chuckb]

What I'm not quite getting is the sharp increase in noise after roughly 1000 seconds. Why doesn't it slowly degrade, why the sharp rise? It's one magnitude more all of a sudden.

Just a thought, do you have a phase wrap happening at 1000 sec? That software data correction may not be getting handled correctly.
Good Luck!

[edpalmer42]

What I'm not quite getting is the sharp increase in noise after roughly 1000 seconds. Why doesn't it slowly degrade, why the sharp rise? It's one magnitude more all of a sudden.

Based on the notes that Corby added to the graph, I think he flipped a switch to change the phase offset from near zero to near full.

I can only imagine the graph you are looking at.  If it is like ones I've seen, it is slowly degrading....  it's just that it's a logarithmic scale so graph looks the way you see it.

He's looking at 1200lhnoise.png which is 5 or 6 messages above yours!

I'm having more basic problem...  HELP!  I'm being out smarted by a device I made myself!

Just repeat both of Corby's tests from early in the 'DMTD Board' thread.  Your results should be similar to his - particularly for the noise floor at Tau = 1 sec.  If your scaling factor is wrong, you'll be off by one or more orders of magnitude.  Now make sure that the scaling factor makes sense.  I've never used a DMTD but I believe your comment is correct.  Basically, scaling factor = output frequency / input frequency.  Depending on how the program deals with the scaling factor, you might have to invert that.

[cncjerry]

I'm sorry, I am totally confused. 

Say I'm working with 10MHz input and 1Hz offset, what is the multiplication factor I need to use?  1 second difference represents 10E-7, right?  so 1 x 10^7?  I am using TimeLab.  This whole concept is escaping me.

If you are bringing the data directly into timelab with a 10Mhz DUT and a 1 second offset, then multiply (in the timelab box) by 1e-7.  The formula is 1/(F/Fo) where F is the frequency being measured and Fo is the offset frequency.  So with a 10hz offset you would use 1e-6.

[5065AGuru]

Thinkfat,

At 1000 Sec I switched the phase relationship instantly from close to zero to full scale.

If you could slowly change the phase from 0 to full scale over say a thousand Sec. then you would see the gradual increase in the noise as you departed from zero, but in a noise floor setup that is not feasible.
That's why I made the chart showing the AD degradation at different points.

All,

Here is a nice explanation about resolution and multiplier from Tom VanBaak:

"The resolution is simply the TI measurement times the DMTD "multiplier".

The "multiplier" is based on DUT/REF vs. LO, so in your case it's 1 Hz / 10 MHz or 1e-7. That remains the same regardless of what make/model counter you use to measure the period. It only changes if you alter the DUT/REF frequency or the LO frequency.


So that's the multiplier part of the equation. For the counter part, let's look at 3 different counters.


A 53131A is a counter which outputs numerical TI readings with actual units. So that's why you see "s" and "us" in the reading. The units of a time interval measurement are always seconds. By contrast when a 53131A measures frequency the numbers have Hz or MHz units. Either way, there are units. Units are important, even critical. The units for a voltage measurement are in volts, etc. This is just normal science; all measurements have units.


Your purpose built  counter, appears to bypass the use of a decimal point and bypass the use of scientific units. Instead it outputs a raw integer count of 10 ns clocks, which is then often a very large number, in the tens of millions. To use this device, one has to manually divide these super large integers by 10^8 in order to get proper time interval measurements that are in units of seconds. So for your DMTD, and this counter, you use the 7+8 = 15 scale factor.


The third, like the 53131, outputs properly scaled measurements without the user having to add fudge factors. The units are already in seconds. That's why you see numbers in the modest range of about 0.00  to 1.00 s. This means, the only scaling factor you need to apply is the DMTD multiplier, which is 1e-7.

The other thing that the it does is output measurements without resorting to commas, which some software, like Plotter or Stable32 or Excel don't properly recognize.

Bottom line, when using it, the only number you have to scale by is the DMTD multiplier, which in your case is 1e-7."

Cheers,

Corby

[5065AGuru]

Here is some info on the "dead-time" problem experienced when using the HP 53131/132/181A or the SR620.

I did utilize the 131A and SR620 the first couple years of DMTD use, and was able to work around this problem.

Since then I have been using purpose built counters that do not have this problem.

The first phase plot using a 131A is a of a 5065A vs a 5061B and is over 8000 Seconds long.

If you follow the phase as it slowly increases on the plot you will see the slope change (it doubles) as it approaches the wrap.

This is caused when the counter cannot finish its RS232 transmission before the next start pulse arrives.
This results in starting on every other pulse.

I isolated one section for the 2nd plot shown so you can get a better look.

If you just delete the data to the right of the point where the slope changes (which includes the wrap) you end up with a clean plot that can be processed.

We are still working on the companion counter to the simple dual mixer and it does not have this problem!

Cheers,

Corby

[5065AGuru]

Made some checks at low and high offset on the Simple Dual Mixer board.
Main conclusion is that as long as you have a decent offset oscillator you don't need to worry
too much about the noise being lower in the lower counts! On my NBS106B with its better noise floor you do see that effect!
First is a phase plot between a good FTS 1200 and a very nice HP 10811 showing the AD at low and high count sections.
The next is an AD plot of the last 7500 Sec (after the 2nd wrap)
Not too shabby!
In the next couple weeks I'll be running some tests at a lower bandwidth to see if a lower noise floor can be reached. Will share the results.

[tkamiya]

Trying to do a base line noise test and getting some weird results.

Adev graph moves wildly in diminishing oscillation pattern.  In phase shift view, I can see phase is shifting very periodically.  Looking at raw data, I can confirm this is not an artifact.  By process of elimination, so far, I found coax between TIC and DMTD is picking up the same noise.  TICC itself isn't the problem because I see the same thing on HP53132A.  What I don't know is what it is that is causing this interference.  It could be AC mains and it could be other 10MHz signals that are near by. 

I plan to move necessary parts into a different room (living room?) away from lab and see if I can reproduce it.  Last resort will be run everything off battery and do it in middle of my back yard. 

Questions for those who are already working on one.  What kind of power supply do you use for DMTD?  Internal or external?  Switching or linear?  3 terminal regulators or something more complex?

[5065AGuru]

Hi,

Here is more data on how the quality of the offset oscillator affects the performance and also
another illustration of the importance of keeping close to coincidence.
I took a well performing oscillator and injected calibrated amounts of noise and measured
how much the STS was degraded by several fixed amounts.

AD at 1Sec for indicated noise levels
 
8.65X10-13th          Zero
3.10X10-12th          62.5
5.66X10-12th          125
9.72X10-12th          250
1.36X10-11th          375
2.26X10-11th          500

Then I setup to measure the simple dual mixer noise floor with this oscillator as the offset
oscillator.

I measured the noise floor at close to coincidence and also at close to full scale at the 6
levels of noise shown.

The plot shows all 12 300 Sec. measurements with the 6 close to coincidence on the left.
As you can see close to coincidence the noise floor stays fairly constant at all noise levels.
At the midpoint we transition to the 6 close to full scale measurements.
With no noise the first section is not too bad but as the noise is increased you can clearly
see the noise levels rise up.

This should help you select what oscillator to use as your offset oscillator.

I'm currently running some oscillator to oscillator plots with the 6 noise levels and will
post those results soon.

Cheers,

Corby

[5065AGuru]

OK, here is a plot that shows the AD degradation as the offset oscillator noise increases, all plots close to coincidence.

Cheers,

Corby

[5065AGuru]

Taka,

Removing wraps automatically even in PLOTTER is problematic. I find it best to use the single step function. Although for most Quartz plots I usually just use the data up to the first wrap and ignore the rest. In this example I'd delete all the wraps just leaving the parabola. Then I'd do my AD plot on that

Here is an example of removing wraps via the single step function. (Using PLOTTER)

1st is the raw phase plot with 3 wraps of a Quartz vs a 5065A.
As I noted before you see I adjusted to start the plot with the count slowly climbing.
As the oscillator ages you will see the plot make a parabola and then wrap.

2nd is after I deleted the 3rd step.

3rd is after positioning the single marker over the 1st step and clicking on single step. As you can see it did not remove it completely.

4th is after drawing a rectangle around the area of the step. (position cursor, left click and hold while you draw the rectangle.) You now have a zoomed area around the step.

5th shows after you position the marker in the middle of the step as shown. Click on delete +- points until the step is clean and then click on single step again.

6th Shows another remaining residual step. Zoom as required and keep clicking on single step until you get no change.

7th show the final step that does not respond to the single step click. Make sure you position the marker in the middle of this final step.

8th shows the result after clicking once on the delete +- points and then clicking UnZoom.

The first wrap is now gone.

Repeat for the next wrap.

The last plot shows the unwrapped phase plot and can now be used to perform the AD plot.

At this point you can also save the unwrapped file to allow you to import it into TimeLab if desired.

Cheers,

Corby

[tkamiya]

When you say "coincident" and "coincidence" what do you mean, please?

[edpalmer42]

Another trick you can use with Plotter is to use the 'Drag Y' function to distort the data to improve the unwrapping and then remove the distortion to get a perfect unwrap.

Let's say you're looking at Corby's seventh picture above.  You've removed the step but there's still a bit of a glitch.  Instead of deleting points, click on the 'Drag Y' button.  Now, all the points are replaced by big squares.  You want the graph on the right side of the blue line to shift down so click and drag a point on the right higher.  Now do a 'Single step'.  The distortion forces the graph to drop.  Too much or too little movement?  Just adjust and repeat the 'Single step'.  Once you're satisfied, click and drag your distorted point back where it should be.

Yes, it's tedious, but sometimes it's necessary.  I've found that sometimes removing points doesn't give a perfect result.  This gives you another option.  Also, if your datafile is timestamped, Plotter won't close up the graph when you remove points.  It just jumps over the gap.  This trick retains all your data points so there are no gaps.

[tkamiya]

I will have to try PLOTTER.  I have been using TimeLab pretty much exclusively.  It's ability to graph as it records is a big plus.  Perhaps my expectation was wrong.  I wasn't aware this much data editing is necessary to get clean graphs.  I haven't heard anyone talk about it extensively like this before either.

[edpalmer42]

As the quality of your measurements and, therefore, of your data improves, your need to edit data drops, but never disappears.  I haven't used Plotter for a few years because Timelab is just too convenient.  If you've got hundreds of points per wrap, Timelab unwraps perfectly.  The only data editing needed is to remove an occasional noise glitch.  Timelab easily handles this.

But it would be really nice if Timelab included Total Deviation.  Just saying..... 

[5065AGuru]

Ed,

When you are at seventh plot and performing a single step results in no change then PLOTTER has calculated that a single delete +- points will remove the exact amount to result in no remaining "bit of a glitch" I have never seen any remaining glitch after following this procedure.

I can't remark on TimeLab unwrapping as I'm only an occasional user.

Cheers,

Corby

[tkamiya]

I've been wondering about this.

Some SMA connectors are rated for 200 to 500 screw/unscrew cycles only, if rated at all.  500 is for Amphenol's $60/piece type.  (wow....  expensive!)  I've seen 200 for common ones.  I personally have damaged threads few times with far, far, less.  As I understand it, it was originally designed for internal component connection, not an external interface.  Yet, I've seen images of NIST lab using them for latter.  I wonder about this from time to time.  I do not like BNC as I can see readings change just pushing on the connector lightly.

Has anyone in DMTD world or really anyone had their measurement error because of this?  I used N connector for this reason.  I can see myself doing 10 ocxo tests at a time and easily exceeding this rather low rating. 

Corby:
I've just aligned my second best 11801 for 10MHz + 10Hz duty.  I can't wait to get the new TIC!  I've suspended my effort to use my other counters for now.  I feel like chasing an ever moving target.

Ed:
It is odd that I'm always having wrapping problems and you don't have it or rarely.  As I understand it, most counters around year 2000 and before has this dead-time issue.   What kind of counters do you have?  I used TICC, 53132A and 5335A.

[edpalmer42]

To clarify:

The glitch remains after the *first* step removal.  Further step removals are unnecessary with my trick.  It also doesn't delete any data points - if that's important for that particular measurement.

[edpalmer42]

Taka:

I usually use a Fluke PM6681 counter.  I don't know if it has a deadtime or not - the specs are a little obscure on that point.  One thing I always do is make sure that I'm not using any averaging.  One time interval - one measurement.  Now that I think about it, I seem to remember that averaging does confuse Timelab because the averaging smears the wrap over many samples.

[tkamiya]

Does averaging even a concern in time interval measurements like we are doing?  The pulse/signal we are giving to the counter is "start" and "stop", so that there isn't anything to take average of....  As I understand it, in frequency measurements, modern reciprocal counters make hundreds of period measurements and give you the best guess answer.

I have to remind myself...  do not buy any more counters...  except for ProTic.  I already have enough counters to completely fill a full height rack.

[edpalmer42]

Does averaging even a concern in time interval measurements like we are doing?  The pulse/signal we are giving to the counter is "start" and "stop", so that there isn't anything to take average of....  As I understand it, in frequency measurements, modern reciprocal counters make hundreds of period measurements and give you the best guess answer.

If your measurement reporting interval is long enough that you could make multiple measurements your counter might do that.  If there's a wrap during a measurement interval, the counter will average the results and report a value somewhere between 0 and the maximum.  But, when I look at Corby1.txt, I don't see those intermediate values around the wraps so it doesn't look like averaging is part of your problem.

However, I do see a very significant problem in that file!

Code: [Select]

-0.957360687806 TI(A->B)
-0.963215360909 TI(A->B)
-0.969065112713 TI(A->B)
-0.000101747664 TI(A->B)
-0.005955872314 TI(A->B)
-0.011832818799 TI(A->B)
What happened to -0.975, -0.981, -0.987, and -0.993?  The data points are spaced ~ 0.006 sec. apart except for a 30 ms. gap around the wrap.  I looked at a few of the other wraps and found the same anomaly.  Care to guess what happens when I add the missing values to the file?  Spoiler - the first glitch at 238 sec. almost disappears.  If I added another digit or two I expect the glitch would completely disappear.  See the attached Frequency Difference graph.

Are you still using the HP 5335A?

[thinkfat]

That is interesting. That series was recorded with the TICC.

[tkamiya]

Thanks for finding that.  I actually tried to find the missing data, but I couldn't.

So far, I have tried TICC, HP53132A, and HP5335.  The last I tried was HP5335.  They all had similar issues except the last one being less periodic than former.  Earlier recordings were also problematic as DUT and REF frequencies were way off.  What's interesting is, in log series of data, there were only few missing.  I would've expect it'd be all over the place.  Currently, I'm not doing anything pending Corby's counter's arrival.

[5065AGuru]

Taka,

I use a stainless male to female SMA "connector saver" adaptor on the mixer inputs.
That way the wear and tear is on the sacrificial adaptor.

Cheers,

Corby

[tkamiya]

I changed the offset oscillator after noticing the original one (11801-6011) taking few cycles to cross the wrapping point.  My guess is, instability?/phase noise? in oscillator caused it to fluctuate around the wrapping that it took few cycles to complete the transition.  This may have caused algorithm in TL to become confused.  New one is from a discarded HP5071A.  (11801-60219)  Still using HP5335 with 10Hz offset.

I observed frequency difference graph looks a lot cleaner.  Original phase graph shows 15 wrappings in 20 minutes and unwrapped shows a straight line.  So far, I have observed 10 wrappingd and they all unwrapped correctly using TimeLab.  WIth the original offset oscillator, it never did.  Adev looks reasonable and as the data accumulates, there are less "whipping" effect of the graph.  There is this odd jump that looks like pulse in frequency difference graph.  I don't know what this is about.

Still watching...

[thinkfat]

Still the same issue, still not properly unwrapped. You need to switch the phase measurement to display the "linear residual" to see the jumps there. "Frequency difference" does show them.

PS: @tkamiya which firmware version are you running on the TAPR TICC? I can imagine a lot of things, but the TICC loosing measurements is not among those.

[tkamiya]

Where is Linear Residual?  I don't see such option.

[thinkfat]

Where is Linear Residual?  I don't see such option.

In the "Trace" menu, the option is called "Show linear phase/frequency residual".

[tkamiya]

This is interesting....

[thinkfat]

This is interesting....

Not unexpected, this is why you see the jumps in the frequency difference graph. They must come from somewhere and a frequency jump must have a corresponding phase change.

PS: While we're at it, "Ctrl-c" in TL will create a screenshot of the currently displayed graph and store it in the clipboard. You can "paste" it into Explorer to save it and attach it to a posting. Must be way easier than taking actual "screenshots" with a camera and downloading the photos to post them.

[tkamiya]

Actually no....  in my setup, it's rather cumbersome to do it that way.

I see the jumps in exported frequency data.  From 67.3 to 67.8 second, I see frequency going back and forth.  Oddly enough, there is no such jumps in phase data, unwrapped or otherwise.

[thinkfat]

Actually no....  in my setup, it's rather cumbersome to do it that way.

I see the jumps in exported frequency data.  From 67.3 to 67.8 second, I see frequency going back and forth.  Oddly enough, there is no such jumps in phase data, unwrapped or otherwise.

What I find quite odd is the number of jumps in the unwrapped phase data is about double the wraps you have in the original phase data.

[edpalmer42]

This is interesting....

Not unexpected, this is why you see the jumps in the frequency difference graph. They must come from somewhere and a frequency jump must have a corresponding phase change.

PS: While we're at it, "Ctrl-c" in TL will create a screenshot of the currently displayed graph and store it in the clipboard. You can "paste" it into Explorer to save it and attach it to a posting. Must be way easier than taking actual "screenshots" with a camera and downloading the photos to post them.

Or just use File -> Save image or .TIM file -> Save as type -> Image files (*.PNG)

I refuse to try and figure out those blurry photos.

[tkamiya]

Sure.  If you don't like the images, by all means....  I just wanted to give an idea of what I'm seeing.

Concerning number of shifts, that may be because the slope is a lot steeper this time.  I didn't do as good of a job as before having frequencies converge.  What I really need to find out is where this is coming from: hardware side or software side.  I would like to dump the counter data to text file first but being GPIB, I don't have a way to do it.  I'll keep playing.  There got to be a reason.  At least it's consistent though, so I have a hope I will eventually find a solution.

[5065AGuru]

Taka,

Here are some examples of how to set up PLOTTER to use data taken with the ProTIC counter.


DUT and REF       Offset         *************PLOTTER SETTINGS***************************
Frequency            used          Nominal frequency      Tau-0/s      Formula to compute new series with
----------------------------------------------------------------------------------------------------------------
10MHz                 1Hz           1.00000000000E+07      1.0                           1.0e-7*x1
10Mhz                10Hz           1.00000000000E+07        .1                           1.0e-6*x1
5MHz                  1Hz            1.00000000000E+07      1.0                           2.0e-7*x1
5MHz                 10Hz           1.00000000000E+07        .1                           2.0e-6*x1

You load the data file into PLOTTER, Make sure your PLOTTER settings are correct, you
click on Series and then "Compute new series from existing".
Put a name or just a single character into the top line and Input the Formula above into the bottom line.
Before clicking on OK I usually highlight and copy the formula. That way on subsequent measurements I don't have to type it in again, just paste it in.
After you click OK you will see another column labeled with you name or character in the upper right
of the screen. Click on the red Column 1 box to eliminate it.
The black phase plot can now be manipulated if required to remove wraps or to select just a portion to plot. Once you are happy click on Time Stability Statistics and select how you want to plot. For instance select data is phase, use powers of 2, and Normal Allan Deviation.
Once your selections have a check mark click Process on the bottom line.
You will now have the AD plot of the selected data.

Cheers,

Corby