Lars DIY GPSDO with Arduino and 1ns resolution TIC

[iMo]

First we try to have the cleanest Vref possible, would using it for the 74HCxx not make it dirty again?

You want to have the cleanest and most stable PWM DAC possible. Thus one possible approach is to detach the PWM drivers from the MCU (the major source of noise, imho) and feed them with an external clean voltage.
Vref in my picture above could be a separate one, not related to OCXO's Vref.

@dbldutch: happy to see it worked..
Intersect - I see it at 700secs with MADEV.. 

[Dbldutch]

UPDATED with finished chart

I was able to very gradually increase the TC value, and I'm now running at 700 seconds with a damping factor of 3.0 with a solid lock.
 
Below is the MDEV report that I started with 700s and a damping of 3.0. The filter is applied. Unfortunately, even 18hrs is not enough to bottom-out. Bummer. But, it looks like I'm on the right track.

What I still don't get is the hump in the beginning (from 0 to 11 seconds). I don't see that in any of the other charts people have made. Can somebody explain this?

In earlier tests, I was probably too fast switching from a TC level to the next, or took steps that were too large. From what I gathered, you need to have enough data fed to the PI controller so it is stable enough before you can take the next step.

Some of you have commented on automating this procedure. It would be nice if you could explain in detail what you have done. Most importantly are not only the TC step size going up, and on what condition(s), but also what to do when you loose the lock, or are moving towards loosing the lock. This would be a nice addition to the code Lars wrote.

PS the title in the graph says t4, but I forgot to update the text before I started a new one. The test was run with a TC of 700.

[thinkfat]

I don't have a particular implementation in mind, but it seems natural to detect a stable lock by looking at the integrator state for a longer period of time. If you find it's staying close to 0 for many TC, it will not hurt to increase it. If its value increases continually it is an indication that the loop cannot keep the lock and a shorter TC should help.

[Dbldutch]

Because I have two broken Bliley OCXO’s, both without the 10MHz output, I contemplated opening them up and see if I could possibly fix them.

Not that I had high hopes of fixing them, but I was interested to find out what is in the little cans as well.

Luckily, I didn’t need to get destructive, it had been done before:

https://www.eevblog.com/forum/metrology/teardown-of-bliley-ocxo-module/25/

After seeing the complexity of the circuit and difficulty in opening the can, I’m pretty sure my efforts would be fruitless.
They will get a good and final resting place.

[Dbldutch]

I finally got the replacement for a Bliley OCXO that I ordered after I had 3 dead ones.
This is a NV47M1008, the M indicating it has a CMOS square wave output. Makes it easier for this application.

The circuit I published earlier has been modified and I implemented the suggestion from imo to separate the DAC PWM signals from the Arduino circuit with a pair of HC14's, to make it cleaner.

Because I don't have a better voltage reference in my stash that can drive a few mA, I used an LM4040 4.096V Zener shunt. At this moment I'm unsure if it's better to also power the VC circuit from this reference. Currently, because the Bliley has a 4V reference voltage, I'm using that. I can be convinced that it's better (temp co-efficiency) to have both circuits powered by the same reference voltage. BTW, I tried to use the Bliley V-ref to power the HC14 to no avail, as can be expected.

Here is the circuit I'm using for this OCXO. It can be easily changed for other types and brands.

With the current resistors for the DAC and VC offset, my gain is 231, a little on the low side. The ideal value should be around 500. After a few days of burn-in, I will raise the value of the 68K to 82K to reduce the frequency span, which is 3Hz at the moment.

Tests are running, I'll report them later.

[metrologist]

Now the fun begins. I'm amazed that you guys achieve such high TC. It would work for a while but always lost lock within a week. Maybe the CTI OCXO is just too drifty (but is that what Miti uses?). I never got around to setting up some MTI OCXO's I had because they are 12V units and would tune 0-10V. A lot of interesting ideas to think about now...

[bingo600]

implemented the suggestion from imo to separate the DAC PWM signals from the Arduino circuit with a pair of HC14's, to make it cleaner.

I see most timenuts uses HC04 AC04 as buffers , they say it's cleaner than than HC14.
Prob. the Schmitt that adds hysteresis.

Ie.
https://www.tapr.org/kits_tadd-3.html
http://www.tapr.org/~n8ur/TADD-3_Manual.pdf

/Bingo

[iMo]

Experienced timenuts would advice you not to put logic gates, which work with different clock domains, into the same package.. It may come to "pulling", "beating", etc. 
I wished someone measured those effects..

[Dbldutch]

I'm struggling with my 4th Bliley, a model NVA47M1008 OCXO with serial number 285 0603.

I see massive jumps making this specimen practically unfit for this application. I'm not showing this, but the system recovers eventually.
This happens every couple of hours.

I've seen others reporting oscillator jumps, Lars included, unfortunately, they did not provide much information.

Not sure if this is a typical case, please chime in if you know more.

Another Bliley hits the dust.

[metrologist]

Would that be on the order of a phase inversion? I seem to recall something about A/B/C mode operation of the oscillators, but not that they would switch modes during operation. I've had some power up in 'B' mode, 10.7 or 10.9MHz.

[Leo Bodnar]

I'm struggling with my 4th Bliley, a model NVA47M1008 OCXO with serial number 285 0603.

I had massive injection locking problems with cheap Blileys.  They were 100MHz but internal design was similar to 10MHz.  Watch out.

I am curious as to why you are using internal Vref for the tuning voltage base?  You are loading already very simple internals of the OCXO. Vref is generated internally by nothing more than a cheap 4.0V LDO which is then followed by an emitter follower (or capacitance multiplier if you will) and designed to be used as a crutch for a tuning trimpot in cases when there is no electronics around the OCXO.  Tuning voltage is not ratiometric to Vref, so you just need a good stable low noise voltage - you probably already have it on your board elsewhere.

What I still don't get is the hump in the beginning (from 0 to 11 seconds). I don't see that in any of the other charts people have made. Can somebody explain this?

Do you have any running averaging going on in your data?  Averaging will artificially improve ADEV or MDEV for shorter tau and will look like a hump.
Cheers
Leo

[Miti]

I'm struggling with my 4th Bliley, a model NVA47M1008 OCXO with serial number 285 0603.

I see massive jumps making this specimen practically unfit for this application. I'm not showing this, but the system recovers eventually.
This happens every couple of hours.

I've seen others reporting oscillator jumps, Lars included, unfortunately, they did not provide much information.

Now that you mentioned , that may be the issue with my Lars GPSDO. Something “jumps” randomly and I didn’t figure out what is it. It may stay locked for months, then loose lock few times a week. Every time it shows a big jump in ns. Details in some posts not to far above. I’m using a Bliley NV... something from Ebay

[Dbldutch]

Hi Leo,

Thank you for chiming in.

I used the Bliley Vref in the hope it would provide a better thermal "connection" to the frequency setting circuit because the Bliley Vref circuit is inside the ovenized enclosure.
With all the data we collected, it's probably better to not use the Vref output on any OCXO at all and use a separate Vref circuit that can power the complete DAC circuit.

I tried an LM4040 4.096V Zener to power the DAC with the HC14 circuit, but there is still too much jitter.

Maybe a bit overkill, but I'm now using a 5V reference, the REF02EZ and I switched to my CTI OCXO again.

I use the Lars program to collect the DAC data, so there is no extra filtering afaik. I have seen some small humps in other charts, but none as big as mine. I'm not sure that bigger is better. 

[thinkfat]

Could it be a bug in the software?Some integer overflow not handled correctly?

[Dbldutch]

After switching back to my CTI OC5SC25 OCXO, I ran a new measurement with the modified PWM DAC setup to determine the optimum TC value.
This test setup is explained in Lars's document as option 2 on page 12. I used a TC of 4 and a damping of 3.0.

The results are astonishingly different. 

Please help me to explain this vast difference.

Not very scientific, but this has changed between the two test setups:
1. The first test ran with live data, the second is an import from Excel made from the Arduino serial monitor.
2. With the first version I used a current boost circuit to reduce the temperature of the 7805 voltage regulator. With the second setup, I only use a 7805 regulator. (see the circuit diagram in the post just above)
3. For the second setup, the PWM/DAC circuit was moved from the Arduino board to the OCXO board and I added 100nF caps in parallel to the 4.7uF Tantalum caps. The PWM/DAC outputs (D3 and D11) are now connected through 20cm Dupont patch cables. In the previous version, these cables connected the output of the DAC circuit to the oscillator board. All connections between the two boards are made with these cables.
4. The second version uses the 74HC14 per imo's suggestion to reduce DAC noise from the Arduino supply.
5. For the second version, a REF02 5V precision voltage reference was added to supply the complete PWM/DAC circuit.
6. The gain for the first setup was 455 with an f delta of 1.4Hz, the gain for the second setup is 303 with an f delta of 1.1Hz.
7. Note that both setups already had the 10MOhm removal modification implemented. See my earlier posts.
8. The dip in the OSXO temp in the beginning is the dropping of the ambient office temperature for the night.
9. The OCXO board is covered during the tests by a plastic container to reduce draft effects.

My questions:
1. The mysterious (for me) hump in the old setup is gone. Why?
1. Are the hardware changes plausible cause for these results?
2. Where do the wriggles/oscillation come from in the new setup?
3. Why are the ADEV and MDEV charts so different for the new setup? (both have the drift line filter applied)

[iMo]

Your MADEV's sigma at tau=1000secs is now 1000x lower than in previous measurement.. That's suspicious.. 

The oscillation - it could be you now operate in a sigma region where more finetuning is required.

I would boost the REF02's output current with a PNP and 220ohm resistor. It should not have any impact on noise and stability of the REF02's output. Add a 10uF tantalum (parallel 100n) at the output. For example:

[thinkfat]

If you're talking about the hump at tau = 4s, that would be the effect of your control loop TC combined with the settling time of the PWM DAC.
If it's missing in the second setup, the settling time of the PWM DAC must have dramatically changed. It's now completely masking your loop TC.

[iMo]

I would also calculate the rms values of your above measurements (ns, DAC, ns-diff) - good for comparing the configurations. In your excel:
=STDEV.P(C:C)
where the ie. C is the column with your data.

[iMo]

I do not see a significant change in the PWM DAC settling time with the HC14 buffer - about 2.5 seconds in dbldutch's wiring.

My improvement N3.: I would wire a choke (10uH-100uH) in series with the EFC (at the OCXO's EFC input), best made of a small toroid. It will dampen 10+MHz leakage off the OCXO into the DAC.

[iMo]

1. The first test ran with live data, the second is an import from Excel made from the Arduino serial monitor.

Doublecheck your TimeLab settings. It could be you made a typo in the "Numeric Field" multiplier setting..

[Dbldutch]

Hi all, thank you for the inputs. Interesting puzzle eh?

Here are some answers:

@thinkfat:
The most significant change has been the move of the PWM/DAC circuit to the OCXO board. Previously, I "transported" the very sensitive output with two un-shielded 20 cm wires to the OCXO board. Now the digital outputs are send across the wire and the whole (sensitive) circuit is close to the OCXO. That might explain the hump removal?

@imo
The standard deviations are :
ns 6.946
DAC 373.856
OCXO temp 0.226
diff_ns 9.552
filtx10 44.760

I have attached the Timelab settings for both measurements. The only obvious difference is in the Numeric Field. This value is calculated according to Lars's document on page 12 as 1E-9/gain. I already listed my gains. (Using e12 or E12 makes no difference)
I don't see a glaringly obvious explanation.

According to my reading of the datasheet, the REF02 should be able to deliver 10mA, and that should be more than adequate. I will try your suggestion.
I'm not sure I have a choke that I can use. I'm not very experienced with them, so I will have to research and learn and see if I have something in my stash that I can use or modify.

If (big IF) the modifications are indeed the cause of the major changes, my next step would be to try a "real" DAC with a reference, built in or not, like many of you suggested or already use.

I'm now running a test with a TC of 500 and damping at 3.0 and the prelim. results are strange as well.

[iMo]

I think the MADEV values below 1e-12 at tau=1000 are pretty unrealistic numbers, unless your OCXO is a fake and it is rubidium or cesium inside ..
The new PWM DAC cannot introduce such a huge difference, imho.
There is an issue somewhere..

The 5V/10mA could be ok for driving a 10k pot trimmer, but you have got there 4x HC14 inverters, which pull some current when in transition 0/1/0.

PS: the bottom "peaks" of the madev are spaced in 60secs period, it seems..

PPS: could you post your raw data from the last measurements, plz?

[Dbldutch]

@imo,

Here is the zipped Excel file with the latest numbers of entries of the current test with TC 500 and damping at 3.0.

[iMo]

What I see off your data:
1. raw DAC data, 20.3 units RMS
2. madev with DAC*3.3e-12, freq difference, all trends subtracted
3. freq diff
4. setting

[iMo]

You have imported 3rd data column off the file, that are the "seconds" of the time information. Therefore the peaks with 60seconds period in your MADEV.
You have to import the 6th data column off your file.