Building a 22 GHz network analyzer for under $1000

[gf]

The software takes the raw S11 data, windows it, and then does an inverse FFT to get the impulse respose from the S11 data. The resulting signal is complex, while the impulse response should be real, so the developer used the absolute value function to make it real. This is the wrong way to do it (for starters, a short and open would look exactly the same, since their S11 values are just negatives of each other, so if you take the absolute value you will get the same result).

It is not "wrong". There are two modes for VNA-based TDR, bandpass mode and lowpass mode. What this software does is bandpass mode. It still enables "time to fault" measurements w/o requiring low-frequency measurements down to DC.

The correct way is to take the S11 data and append the conjugate of the measured data. Basically, for real signals, the FFT must be symmetric (in that the negative frequencies must be the complex conjugate of the positive frequencies). After adding terms to the end so that the negative frequencies and positive frequencies are conjugates of each other, if you take the inverse FFT, the signal is real, and you get the correct impulse response.

What you describe is lowpass mode. Untorfunately, it requires equally spaced frequency points from 0Hz (DC) to f_max, while many VNAs have a lower frequency limit.

From Brian and CMT, they use chirp-Z transform:
https://www.eevblog.com/forum/rf-microwave/vna-for-cable-characterization/msg5656049/#msg5656049

There is still nothing wrong with doing a IFFT in order to calculate the full time axis (possibly with zero-padding - after applying the window function - in order to get a higher (interpolated) resolution for the resulting time domain points).

Consider the ICZT rather as a tool to calculate an arbitrary "zoom-IFFT" with interpolation. For example, if you want to zoom into a small range of the full time-axis, or if you want an arbitrary spacing for the resulting time domain points (independent of the frequency points).

[EggertEnjoyer123]

The software takes the raw S11 data, windows it, and then does an inverse FFT to get the impulse respose from the S11 data. The resulting signal is complex, while the impulse response should be real, so the developer used the absolute value function to make it real. This is the wrong way to do it (for starters, a short and open would look exactly the same, since their S11 values are just negatives of each other, so if you take the absolute value you will get the same result).

It is not "wrong". There are two modes for VNA-based TDR, bandpass mode and lowpass mode. What this software does is bandpass mode. It still enables "time to fault" measurements w/o requiring low-frequency measurements down to DC.

The correct way is to take the S11 data and append the conjugate of the measured data. Basically, for real signals, the FFT must be symmetric (in that the negative frequencies must be the complex conjugate of the positive frequencies). After adding terms to the end so that the negative frequencies and positive frequencies are conjugates of each other, if you take the inverse FFT, the signal is real, and you get the correct impulse response.

What you describe is lowpass mode. Untorfunately, it requires equally spaced frequency points from 0Hz (DC) to f_max, while many VNAs have a lower frequency limit.

From Brian and CMT, they use chirp-Z transform:
https://www.eevblog.com/forum/rf-microwave/vna-for-cable-characterization/msg5656049/#msg5656049

There is still nothing wrong with doing a IFFT in order to calculate the full time axis (possibly with zero-padding - after applying the window function - in order to get a higher (interpolated) resolution for the resulting time domain points).

Consider the ICZT rather as a tool to calculate an arbitrary "zoom-IFFT" with interpolation. For example, if you want to zoom into a small range of the full time-axis, or if you want an arbitrary spacing for the resulting time domain points (independent of the frequency points).

It seems like you are right, but their previous implementation was still wrong, because they convolved the absolute value of the impulse response with the step function to try to get the step response, which doesn't work (you just get a stairstep up).

I modified the nanovna-saver code again so you can select between bandpass and lowpass modes. I think the chirp Z transform is probably going to be too hard for me to implement, plus all the processing is done on a computer anyways which can easily do million point IFFTs in microseconds so there is no advantage to using it. (I'll have to admit that this is the exact kind of thinking that makes software slow nowadays)

[gf]

It seems like you are right, but their previous implementation was still wrong, because they convolved the absolute value of the impulse response with the step function to try to get the step response, which doesn't work (you just get a stairstep up).

IMO, only impulse response (magnitude) makes sense in conjunction with bandpass mode, and step response should not be offered in the GUI. I think other VNAs don't offer bandpass step response either.

EDIT:

So basically I see the following three choices:

- bandpass impulse response (magnitude)
- lowpass impulse response (real)
- lowpass step response (real)

[EggertEnjoyer123]

I implemented bandpass TDR and importing S1P files for calibration in software. (Unfortunately, I don't have the loads that I sent my friend yet, so I can't get any accurate measurements)

I also built a resistive splitter out of standard 18 ohm 0201 resistors. They are all soldered upside down to reduce parasitics. The S parameters look good up to 21 GHz (I need to get better amplifiers so I can go higher in frequency).

After swapping out the old splitter with the new resistive splitter, I have 60dB of isolation for S21. I think I can improve it even more by fixing my DSP. I also measured a Beatty standard and a through, which were made by JLCPCB. I think I need to make the 25 ohm section longer on the Beatty. Also, it seems like my soldering is bad for one of the connectors on the through. (The reversed measurements were taken by flipping the board and measuring again). The theoretical reflection for the Beatty is 0.33.

[EggertEnjoyer123]

I found another source of calibration loads.

The loads here seem to be good up to 26.5 GHz at least (my friend can test them up to 40 GHz but it will take a while): https://www.aliexpress.us/item/3256807233299645.html
I bought two of the <1.15 VSWR ones, and I got the following response (looks like the seller might be honest).

The seller also offers 18GHz loads with VSWR <1.06 here: https://www.aliexpress.us/item/3256807162155931.html. I suspect someone is getting paid to measure and bin these. I haven't tested them yet, but I might buy a few once I get my loads back from my friends (so I can finally have a NIST traceable calibration set of dubious quality).

The S88 load is from a calibration set that my friend has at his lab.

[joeqsmith]

...
The seller also offers 18GHz loads with VSWR <1.06 here: https://www.aliexpress.us/item/3256807162155931.html. I suspect someone is getting paid to measure and bin these. I haven't tested them yet, but I might buy a few once I get my loads back from my friends (so I can finally have a NIST traceable calibration set of dubious quality).
...

If that is true (and would not surprise me), this is a very good deal.  Owners of the low cost VNAs looking for a better load while not spending some major cash should be very interested in your find.

[EggertEnjoyer123]

I got the measurements and it seems like the seller is honest. The measurement is different because a different cal kit was used (my friend said that someone else calibrated the VNA in the previous measurement so maybe they did it wrong)

This is better than the last loads I got, where someone just shoved a 50 ohm resistor into a SMA.

[prutser]

From which of the loads you mentioned is this last plot ?

[EggertEnjoyer123]

From which of the loads you mentioned is this last plot ?

https://www.aliexpress.us/item/3256807233299645.html

This one

My friend says they're good to 40 GHz as claimed.

[EggertEnjoyer123]

I implemented windowing in my software. My code now uses a Hanning window on the ADC data before correlating it with the -1, 0, 1, 0, 1, ... sequence (to basically filter out signals at fs/4). This helps eliminate all the DC frequencies (the DC level changes every time the frequency of the oscillators changes, and as a result there is a lot of unwanted signals at low frequencies). According to my Python simulation, with the original filter there are some frequencies near DC which only experience -60dB attenuation, but with the Hanning window all frequencies near DC are basically eliminated. (The cost is obviously that the 12kHz "passband" is slightly wider, but it's not that big of an issue).

Doing this seems to have eliminated a lot of the randomness with S21 measurements, and my isolation is now about -75dB with calibration (compared to my old results in post #53). Without calibration, it is still around -60dB (I probably need better shielding than aluminum foil). As you can see from the filter measurements, I am now able to measure a lower S21 (around -70 instead of -60dB). I'll probably have to redesign everything with better shielding to do better (there's only so much you can do with calibration).

[EggertEnjoyer123]

I also spent the last month looking at how to make a good resistive bridge, which should let me go down to 10 MHz. The balun seems to be the hardest part.

I took apart a 10 MHz to 20 GHz mixer module to see how it worked. (To be fair it was already broken when I got it). They are using twisted pair instead of coaxial cable, along with a very small ferrite. The wire is 0.08mm diameter. I bought wire of the same diameter, but I got 50 ohms impedance with only 3 wires instead of 4. I tried making a balun using my twisted pair (more like triple) but I wasn't able to get it to work at all, so I'm probably doing something wrong. Maybe they are using special ferrite or something? I'm going to try again with 0.35mm microcoax from Aliexpress, which should hopefully be easier to work with. Also, twisted pair isn't TEM so I'm not sure how it would work in place of coaxial cable (which is normally used in these baluns). Hopefully someone else can explain (I suspect that the ferrite is probably special).

From the image you can see the baluns, and the white objects are two diode rings and two resistors (there is one hiding behind the rings). The mixer is triple balanced, and the part number is T3-20GLCTG. The four wire "twisted pair" goes to the IF, and the other two are the LO and RF. There are five baluns: the two ferrites on top of the diodes have a twisted pair going through each hole (the LO and RF each connect to two baluns), and the IF balun has two wires connected to ground, and two wires connected to the IF output.

[EggertEnjoyer123]

I'm doing some measurements while I wait for parts to try to make a lower frequency directional bridge.

Here are the measurements of a bias tee made with some random chinese 100nF 0201 capacitor and a BLM15GG471 0402 ferrite bead from Murata. The loss seems to be at most 2dB up to 23 GHz. I measured a through and the bias tee and I got the results attached. This ferrite is recommended in the ADL9005 and MAAM-011100 datasheets.

I wanted to see if there were any resonances (which could be the reason why my output power drops by a lot after 21 GHz). It seems like there are no resonances, so the issue is probably with the connectors I used after the amplifiers. Hopefully fixing that issue in a later revision will stop the VNA from drifting while the temperature stabilizes (measurements past 21 GHz only become stable if you wait for the PLL and amplifiers to heat up because the power going to the mixer changes, which affects the conversion loss when the power is low).

[EggertEnjoyer123]

I made some clones of the original mixer, and I am getting decent results up to 19 GHz. However, it seems like there is a huge resonance of some sort above that frequency which completely ruins my results. I am using 4 BAT24-02LS diodes (the cheapest I found was about $50 for 255 on winsource).

I accidentally broke the PCB, so I don't have any good pictures of it. But I have the gerber files for the board, as well as an image I took of the back side of the board before I broke it. The capacitors are 22pF 01005.

To test the conversion loss, I replaced a mixer in my VNA setup with my design. The VNA is uncalibrated, and the difference between the brown and blue line is the difference in conversion loss between the commercial mixer and my mixer. As you can see, my mixer is slightly lossier up to 19 GHz (which is partially explained by the fact that it is twice as large as the commercial design). However, after 19 GHz there is a huge drop in the mixer conversion loss, and I'm not exactly sure what is causing it. Changing the capacitors seems to have no effect. Touching the board also has no noticeable effect unless I touch the small wires connecting the diodes to the IF output. If I touch that, then the conversion loss drops even more after 19 GHz, so I suspect that I might have an unintended antenna somewhere with the pickoff for the IF. The wire which biases one end of the diode ring to ground seems to have no effect, and nothing happens when I touch it. In the commercial mixer, the wires are 5 mm long, but in my mixer, the wires are 8 mm long (so that could be a factor). Also I have no microwave absorber other than my finger.

Also please do not order the gerber files without modifying the design! The copper gap between the vias is too small and it will bridge if you try to get it manufactured. I had to cut the bridges open with a knife. I'm using 0.6 mm FR4 from JLCPCB.

In a later revision I'm going to add ferrites so the balun can work down to DC. My IF is 12 kHz (while the commercial mixer has to support up to 1 GHz IF) so I can get away with it.

[EggertEnjoyer123]

Here's the commercial mixers for comparison.

Adding some copper tape to make the IF pickup lines shorter did nothing. It seems like the issue might actually be with the capacitors, because swapping the LO and RF results in a much higher insertion loss (~5dB) at all frequencies, and the only difference between the LO and RF baluns is that one has capacitors and the other doesn't. Maybe SMD capacitors are just bad relative to the beam lead capacitors used in their design.

[arlo_g]

Thanks for posting the excellent photos of mixers.
I had expected that Marki was using twisted strips of PCB (parallel strip transmission line) threaded through ferrite beads in their ultra-broadband mixers, but evidently not in the one that you autopsied. 

I seem to remember that Darrell Harmon had posted on his website, or maybe on Twitter, that he got excellent broadband performance from inexpensive multi-layer ceramic caps, but I can't find it again.  It might be worth trying multi-layer (SMT) caps soldered on edge, too, as that can smear out the abrupt impedance transitions in the cap structure.

Does anyone have any bright ideas for how to implement flex/rigid PCB on the cheap from flex PCB's at home?  A lot of the commercial microwave mixer modules seem to be fabricated in thin suspended substrate held to a frame of metal (or maybe rigid PCB).

[EggertEnjoyer123]

Thanks for posting the excellent photos of mixers.
I had expected that Marki was using twisted strips of PCB (parallel strip transmission line) threaded through ferrite beads in their ultra-broadband mixers, but evidently not in the one that you autopsied. 

I seem to remember that Darrell Harmon had posted on his website, or maybe on Twitter, that he got excellent broadband performance from inexpensive multi-layer ceramic caps, but I can't find it again.  It might be worth trying multi-layer (SMT) caps soldered on edge, too, as that can smear out the abrupt impedance transitions in the cap structure.

Does anyone have any bright ideas for how to implement flex/rigid PCB on the cheap from flex PCB's at home?  A lot of the commercial microwave mixer modules seem to be fabricated in thin suspended substrate held to a frame of metal (or maybe rigid PCB).

I already looked into the flex PCBs. They are too thin and a 50 ohm microstrip trace would be impossibly small on them (at least JLCPCB's flex is too thin). I'm probably sticking with 0.5mm FR4 which is extremely cheap, and probably not that lossy (if I make everything thinner so it's more in line with the commercial mixers). The cheapest way I can think of is 3D printing a case and covering it with copper tape. Certain plastics might be lossy enough but you can always add antistatic foam to kill resonances.

I'm also going to switch my cheapo Chinese 0201 capacitors for better ones. (Maybe also use reversed 0204 capacitors for the mixer, since those have a lot less impedance). The CL03A line of 0201 capacitors have much lower impedance at 6 GHz (around 2 ohms for the larger ones in the microfarad range) which extrapolates to about 7 ohms at 20 GHz.

Also Harmon did measure the 0402 ferrite bead that I'm using right now. Unfortunately he doesn't say how he measured it. I can probably use the bisection method to verify his measurements with a better test setup for SMDs. The plot is attached below. Coincidentally he has been working on a 20 GHz VNA for over 10 years at this point, but there aren't any details about it.

[Gerhard_dk4xp]

I was in the team in 2003-5 that designed these 10 GB/s XFP fiber optic transceivers
at Infernion fiber optics. Note the thick wire bracket over the flex cable next to the
solder joint. That clamp was the breakthrough. The laser and the optical receiver
must swimm because the FO cable connector requires that. It took a huge amount of
HFSS simulations to get the ribbon work and produce nice eye patterns at 10 GB/s.


I tried JLCPCB's el cheapo 4 layer process (not even the impedance controlled option)
for my LMX2594 synthesizer. On a 100*100 mm2 minimum board is enough place for
4 synths and some micro strip test structures.

JLCpcb's impedance calculator says that 50 Ohm is 11.5 mil wide on top when
layer 2 is GND. The FR4/Cu  below gives mechanical strength. When tested with
an Agilent 18 GHz TDR it turned out that the impedance calculator was right on spot.
I needed another iteration for the SMA connectors. (Rosenberger) 
The center pad was much to large.

regards, Gerhard

[EggertEnjoyer123]

Here's my latest mixer design. Still about 5-6dB short of the commercial mixer at high frequencies. (I am comparing the mixers by doing an uncalibrated S21 measurement and swapping the mixer out. You can ignore the S11 plot and only look at the S21 plot. The blue line is my mixer and the brown line is the commercial mixer.)

Again, diodes are BAT24-02 and everything is done on 2 layer JLCPCB 0.6mm FR4.

One thing I find weird is that if I swap the LO and RF of my mixer (the LO side has the DC block capacitors and the RF part has a short to ground), there is a very sharp resonance at 16 GHz and my insertion loss spikes to 30 dB. If I put a 3dB attenuator on the LO side, it fixes the issue somewhat. The resonance frequency is unaffected by what signal generator I use, so it is very likely related to something on my mixer board. Interestingly nothing happens at 16GHz when the LO and RF aren't swapped. The commercial mixer also has worse performance when LO and RF are swapped, but not by much.

Design files are attached. Perhaps my mixer would get closer to the commercial mixer if I used a higher LO power (I think more power is lost across my substrate + capacitor). I also measured the mixer the right way around, but with a 3dB attenuator on the LO port. The results are exactly what you expect (last picture).

[EggertEnjoyer123]

I also measured the loss of my microstrip balun, which isn't great because FR4 was used.

[ftg]

Exiting progress!

Just be aware that JLCPCB does not guarantee the FR4 material their for 2-layer boards.
So the Er of the material used can change between batches.
For 4-layer the FR4 material should stay the same between batches.

[prutser]

Never considered to use JLPC PTFE for mixer PCB ?
It's only 2 layer and more expensive than 4L FR4, but still able to get 5 PCBs for approx $60,-

[EggertEnjoyer123]

Using a 0204 reversed style capacitor seems to get me a few dB improvement at high frequencies. (The part number is LLL153C80J104ME01E)

I'm within 5dB of the commercial mixer now. I can probably get a few more dBs by optimizing the connectors.

[EggertEnjoyer123]

I am getting a weird resonance after swapping out the solder bridges with a ferrite bead (BLM15GG471). The good thing is that the mixer now works down to 10 MHz (and probably lower). Conversion loss is about 8-9 dB at 10 MHz. I'm not sure if everything is still double balanced since the baluns don't work at low frequency, but I think that is easily fixable with ferrites.

The commercial mixers also seem to exhibit the weird resonance effect (see bottom image). It looks like one of the commercial mixers is a lot worse than the other in my test. I think the only reasonable explanation is that my LO power is insufficient at high frequencies (since my original board has bad design and my amplifier is not designed for these frequencies). If you go to post 19 you can see that the output of my PLL + amplifier could go down to 0 dBm (my spectrum analyzer is also +/- 3.5 dB). This means the diodes aren't being driven hard enough, and look capacitive. The biasing inductors could resonate with them. When the LO is driven hard enough, the diodes look more like shorts than capacitors so we don't see the effect.

I might have to upgrade the amplifiers in my next revision. I have a couple RF amplifier MMICs laying around that I could probably use (or I could buy the Mini Circuits LVA-273PN+). +15dBm is probably enough to drive all my mixers without breaking them. The Macom amplifier part (MAAM-011100) is good enough to get to about 20 GHz with no issues, but afterwards the amplitude drops significantly. The Macom part is about 4 times cheaper than the Mini Circuits part though.

[rloc]

Good day!

Hopefully someone else can explain (I suspect that the ferrite is probably special).

Minicircuits recommends using a high permeability ferrite. I would guess that Marki used something similar to B62152A0015X030. Quote from one patent (dual output autotransformer from 30 MHz to 3000 MHz or greater):

has a binocular core, preferably made from a high permeability ferrite material. The preferred core dimensions are approximately 0.096″ high×0.080″ wide. The wire used for the windings is two strand twisted 34 gauge and three strand twisted 36 gauge, both preferably 7 twists per inch

The length of the LO/RF twisted pairs should be slightly shorter than a quarter wavelength of the mid-range, adjusted for propagation speed. The characteristic impedance of the LO/RF twisted pairs should be about 100 ohms, since they are connected in parallel on the port side.

two resistors (there is one hiding behind the rings

Can you show in more detail the location of these resistors, what points they are connected to and the values?

[EggertEnjoyer123]

I tried messing around with OpenEMS

It seems like you can make a decent resistive splitter at 40 GHz. The "resistors" (white) are modeled by a thin sheet of material with finite conductivity, and some alumina on top (adding the alumina did not make a noticeable difference). Each resistor models 2 36 ohm resistors (01005) soldered upside down.

I'm also trying to design a resistive bridge to replace the directional coupler in my design. So far my design works to maybe 15 GHz or so before my return loss suffers.