开云体育

N6LF has been trying to calibrate his NanoVNA-H4 through a wideband balun for common-mode current suppression when measuring an open wire line ground probe. He has had perplexing results. As a test he tried paralleling two 50-ohm loads, calibrating the VNA with that load, and then measuring it. He did not disconnect/reconnect the load and he measured it immediately after calibrating. He got the results depicted in the plots.

Resistance is off 1 part in 5000 and reactance has an upward creep at LF of more than that. Spikes occur. Are these results reasonable?

Brian

The spike at the extreme left could be WWVB at 60 kHz. Have no idea what
the spike at 1.8 MHz might be.

Dave - W?LEV

--
Dave - W?LEV

That's an interesting idea, Dave. The LF spike is indeed at 60 kHz. However, N6LF is in Oregon. I would not expect WWVB to be that strong there. When connecting the VNA to a ground probe, occasionally he sees a spike from a 50 kW local AM station at 1.12 MHz.

My main concern is the resistive offset from 50 ohms and the climb in reactance at LF. I would think the NanoVNA has plenty of accuracy to nullify any such aberrations, but I don't know the details of its internal resolution or processing.

Brian

On 2/6/23 11:36 AM, Brian Beezley wrote:

N6LF has been trying to calibrate his NanoVNA-H4 through a wideband balun for common-mode current suppression when measuring an open wire line ground probe. He has had perplexing results. As a test he tried paralleling two 50-ohm loads, calibrating the VNA with that load, and then measuring it. He did not disconnect/reconnect the load and he measured it immediately after calibrating. He got the results depicted in the plots.
Resistance is off 1 part in 5000 and reactance has an upward creep at LF of more than that. Spikes occur. Are these results reasonable?

1 in 5000 is actually pretty good. That's 0.02% - I doubt the load is known that accurately over that bandwidth.

Similarly, the reactance of the load is probably not zero. 0.1 ohms at 50 kHz is about 0.3 uH. That's the inductance of about a foot of wire. And with the falling reactance with rising frequency, maybe there's a parasitic L and C. (or an offset in the calibration)

On 2/6/23 1:33 PM, Brian Beezley wrote:

That's an interesting idea, Dave. The LF spike is indeed at 60 kHz. However, N6LF is in Oregon. I would not expect WWVB to be that strong there. When connecting the VNA to a ground probe, occasionally he sees a spike from a 50 kW local AM station at 1.12 MHz.
My main concern is the resistive offset from 50 ohms and the climb in reactance at LF. I would think the NanoVNA has plenty of accuracy to nullify any such aberrations, but I don't know the details of its internal resolution or processing.

The measurements are made by integrating over 1 millisecond, after a I/Q conversion to 5kHz. It's a 16 bit ADC, but the noise floor isn't quite that good, and there's system noise too.


I started a formal analysis last summer but never finished.

1 part in 5000 is ~70 dB (20 log10(5000) = 74 dB)

The trace noise is bigger than that.

Thanks, Jim. I'll read the analysis.

I didn't realize the ADC was 16 bits. That's a lot. Now I feel even less forgiving about the observed performance.

I don't understand why the possible imperfections in the load you postulate matter. Shouldn't calibration nullify them?

Brian

On 2/6/23 2:15 PM, Brian Beezley wrote:

Thanks, Jim. I'll read the analysis.
I didn't realize the ADC was 16 bits. That's a lot. Now I feel even less forgiving about the observed performance.

The actual limitation is the SNR into the detectors. That's driven more by circuit noise within the 5 kHz measurement bandwidth.

In round numbers, each measurement is about 50-60 dB SNR. 60 dB SNR is 0.1% uncertainty.

But the calibration combines multiple measurements (O,S,L), so the uncertainty of the result is bigger. That is, if you combine two 0.1% measurements, the result is 0.2% uncertainty (or, alternately 0.14% if your RSS)

The cal actually combines more than that.

I don't understand why the possible imperfections in the load you postulate matter. Shouldn't calibration nullify them?

You were measuring something you expected to be perfectly flat. But is it flat to 0.02%?

Or are you calibrating with the load, and then remeasuring it and getting a different answer?

On Mon, Feb 6, 2023 at 02:38 PM, Jim Lux wrote:

Or are you calibrating with the load, and then remeasuring it and getting a
different answer?

Yes. N6LF calibrated with a 25-ohm load and then immediately measured it. That's why I don't understand the result. The resistance offset looks like a numerical bias issue. I don't know what to make of the LF reactance rise.

Brian

On 2/6/23 2:45 PM, Brian Beezley wrote:

On Mon, Feb 6, 2023 at 02:38 PM, Jim Lux wrote:

Or are you calibrating with the load, and then remeasuring it and getting a
different answer?

Yes. N6LF calibrated with a 25-ohm load and then immediately measured it. That's why I don't understand the result. The resistance offset looks like a numerical bias issue. I don't know what to make of the LF reactance rise.
Brian

That is weird. The reflection coefficient (raw) is going to be (50-25)/(50+25) or 0.333 That's not getting anywhere near the noise floor.

It would be interesting to see what he gets if he calibrates with 50 ohms and measures 25. Is there the same frequency dependent variation?

On Mon, Feb 6, 2023 at 03:45 PM, Jim Lux wrote:

It would be interesting to see what he gets if he calibrates with 50 ohms and
measures 25. Is there the same frequency dependent variation?

Here it is. Same resistance offset (looks like same percentage, too). The reactance slope now extends throughout the frequency span.

Brian

I put less emphasis on the 25-ohm measurement after calibrating with 50 ohms because then the load accuracy matters. I think he used a pair of highly accurate HP loads, but who knows the exact values. Also disconnecting the cal load and connecting the test load invites SMA issues. That's why the original data I presented impressed me so. Cal with xxx, then read xxx without disconnecting anything. Results should be flat to within the data and algorithm resolution. I wonder if the divide by 16 you noted in your uncertainty analysis is coming into play.

Brian

Sorry to ask, but is he resetting the calibration each time before
calibrating? Missing that requirement was my problem last time I had a
similar issue.

Stan, here's Rudy's answer to your question:

"Yes, I always reset before a new calibration and then check afterwards with Smith Chart OSL."

Brian

Larry, AE5CZ, sent me files from a DeepVNA 101 attached to a 12" ground probe. The first three plots are with the probe rods in air. You do this to determine the compensation for the capacitance between rods due to probe dielectric. The dim region of the S11 curves is where |S11| > 1 and extends to 18 MHz. Note the effect on calculated resistance, which is the real part of 50(1+S11)/(1-S11). Measurement noise is minimal. This issue did not seem to affect ground measurements, shown in the last plot. They show no artifacts or noise, and are what you might expect from New Mexico ground.

Brian

This shows what calculated resistance looks like when not restricted to positive numbers. |S11| > 1 has real consequences. For now the best strategy seems to be to multiply the real and imaginary parts by 1/|S11| when |S11| > 1. This always yields positive resistance values. At one time I used a complicated AGC scheme that reread the VNA file and multiplied all points by 1/max, where max was the largest value of |S11| that was > 1 during the first pass. Clipping seemed to work just as well and really simplified the code.

Brian