Improving the resolution bandwidth of the SA

While trying to do some analysis of narrow band signals it became obvious the current design of the spectrum analyzer has two limitations.
Below picture shows both of them.

The signal to analyze is at 575kHz. The resolution filter is clearly too wide, about 30kHz at -3dB, and the staircase patter shows scanning is done in 10kHz steps caused by the minimum frequency steps of the first mixer LO, a ADF4351. Making very narrow RBW filters is a considerable effort and an FFT could be an alternative.
So it was time to go test if a mixed mode SA can work. At high spans the SA works with the log detector but as soon as the minimum frequency step is below 10kHz the log detector is no longer used. Instead a third mixer is used  to convert to an IF of 50kHz. This is fed into the PC line-in of a good audio card at 192kHz sampling rate and analyzed using a 1024 point FFT. As the usable buckets of the FFT are limited due to the RBW, now acting as the 3 IF filter, multiple FFT's, spaced 10kHz apart, are stitched together. The FFT bucket width is about 100Hz, About 300 times better compared to the RBW filter used above,
The result is a nice sharp signal due to the flattop window function applied. The scan is 1000 points wide. Measurement speed is considerably faster as instead of 1000 steps the ADF4351 has to step only 10 times. It takes about twice the time for the audio samples to be collected compared to the stabilization of the log detector so in total still 50 times faster



There are still many thing to improve or test such as:
- The dynamic range of the audio input should be in the order of 110dB (24 bit audio card). This needs to be validated together with the behavior of the third mixer
- The frequency calibration and peak labeling needs to be improved for this much higher resolution.
- The noise floor shows a repetitive pattern so something is still wrong in the signal path
- Instead of averaging or duplication of FFT buckets to match the required resolution a better approach is probably to have an adaptive FFT length. A 10k FFT will result in 10Hz RBW resolution (and run 10 times slower) and a 128 bin FFT (for 1kHz resolution) will be much faster
So many things to try!

Mirrors and spurs in Spectrum Analyzers

While analyzing the quality of a signal generator the SA shows a number of components next to the base frequency at 6.18MHz.. 
There are multiple causes for these components. The first obvious are harmonics generated either by the generator or internally in the SA.
A second cause is the generation of unwanted mixer products from the various LO's in the SA.
A third cause are mirrors where the quality of the IF filters is insufficient to suppress the opposite mixer output.
A real life example is this measurement




Which of the signals are real?
A common way to reduce spurs and mirrors is to wobble the intermediate frequencies of the SA and use exponential averaging to smear the energy of the unwanted signal over a wider range.
As you can see enabling this form of spur reduction does have some impact. The signals a 46MHz and 10MHz are almost gone




The IMD2 and IMD3 measurements at 12.33MHz and 18.48MHz remain at -42dB and -53dB but how to be sure these are from signal generator and not generated in the SA?
The simplest way to check is to enable some attenuation. Adding -10dB again changes the picture. The noise floor moves up 10dB. 
Most harmonics did go down as is reflected in the IMD2 and IMD3 measurement so the SA did generate most of the harmonics.




A further increase of the attenuation does not change the IMD2 and IMD3 so we can be fairly sure we are now seeing the real content of the signal from the signal generator.
The peak at 42.83MHz should be at 43.26 (=7*6.18MHz) to be a harmonic (its actually the small peak to the right). In fact it is not from the signal generator but from the PC keyboard laying in top of the coax

Building a Spectrum Analyzer resolution filter

The narrowest resolution filter of my spectrum analyzer did not perform as expected to I decided to build a new filter.
As I did not want to buy many crystals and go through all the difficult sorting, matching and calculations I decided to go for some crystal filters rather cheaply available on ebay.




In contrast with a receiver a SA resolution filter should not be as steep as possible otherwise you may miss some signal easily when you are using a too large frequency range
The NDK 10F7.5A looked suitable so I bought some. Measuring them on my VNA they al seem to be on the same center frequency (10.7MHz) which is nice!
The input impedance is, according to the datasheet, 1.5kOhm/5pF so using the online matching calculator the matching circuit should be something like this.


The 5pF of the crystal should be subtracted from the calculated value of C1 to get the actual C1
In order to confirm the matching circuit I mounted one of the filters on my universal test jig, connected the VNA and connected a tunable inductor and capacitor of about the correct value.



This simple setup lets you tune all components till you get the right performance.
After some fiddling the polar input impedance chart looked like this


Tuning could still be a bit better but the filter loss is very acceptable


As I had more of these filters the obvious next step is to use more then one. Two connected directly in series with the impedance matching at the input and output of the whole filter I got a rather disappointing result.
Way to wide, not symmetrical and too much loss



But then I remembered about connecting a small capacitor to ground in the middle of the filter



And using this ancient variable capacitor I was able to tune the filter




Adding a third filter stage and tuning for minimum loss created a somewhat wider but certainly steeper filter. I can not yet get rid of the pass band ripple but have not yet tuned the impedance matching capacitors and the input/output impedance is still a bit too high.

The benefits of switchable attenuation for spectrum analyzer measurements

During measurements there may be certain spurs that do not have an obvious cause. Are they caused by limitations of the SA? Or are they present in the input signal?
An example is this two tone measurement of the input IIP3 of a mixer


The many spurs below -70dB are cause by bad shielding of the two signal generators. Without these connected the noise at about -100dB is without spurs
The SA automatically finds the peaks and calculates the input IP3 in two independent ways, the results should be equal but there is some difference.
Left IIP3 is calculatec at +9dB where right IIP3 is calculated at +7.7dB
But can we be sure the IIP3 of the mixer is indeed around +8dB?
The simplest way to know is the add attenuation before and after the mixer.
Attenuation after the mixer did not change anything (as it should) but -10dB attenuation before the mixer resulted in a very different picture.
The measured levels are increase by the level of attenuation to keep the displayed levels equal so the noise floor moves up about 10dB

The results (15dB improvement of IIP3)  is not entirely what was expected as every dB reduction of the input signal level should  increase the IIP3 with one dB.
There is still more to investigate and learn.

Phase noise and the choice of the first IF in a spectrum analyzer

Many of you may have heard about "phase noise" but do you thoroughly understand what this is all about.
I also was not aware of the relevance before I started measuring the performance of the my home build spectrum analyzer.

The 10.7MHz resolution filter (third IF filter) I'm using  has a -50dB width of about 60KHz and a -90dB width of 100kHz when measured on a VNA.
When sweeping this third IF filter in the SA while using a first IF at 2.6GHz a very different filter picture appears

The staircase at the center is caused by the discrete steps of the fractional PLL used for the sweep
From 20kHz offset and -40dB down there are side skirts and even a shoulder at 150kHz from center. (the peak at 120kHz is leakage) where neither the side skirts or the shoulders are visible on the VNA.
These skirts and shoulders are caused by the phase noise of the LO's. Not all energy is in the single intended output frequency but there is also noise generated that reduces when farther away from the intended frequency.
A standard way to measure this phase noise is to remove the first mixer and use the first LO as test signal and scan this LO and use a log frequency scale as can be seen in below plot

What you see is a upper side band scan, the lower side band scan should and actually does looks the same (apart from the small leakage peak)
The horizontal scale is the frequency in MHz from the the LO frequency. The sweep of the first LO is still done lineair so the lower frequencies have less measurement points compare to the higher frequencies.
The first point at 0.01MHz is the full LO signal normalized at 0dB and the first point with offset is at 0.1MHz away from the LO. You can not see the resolution filter details (as can be seen in the first picture in this post) as there are insufficient points in this scan at low frequencies but the noise fall-of when further away from the LO signal is clearly visible till about 10MHz where the phase noise goes below the SA noise floor of -105dB
A practical implication of this phase noise is when you have a strong (0dB) signal 300kHz away from a weak signal the noise floor of the SA will increase from -105dB (right part of scan) to -80dB so the sensitivity of the SA is reduced in the near presence of strong signals.
Do keep in mind that what you see is actually a result of 3 LO's  (first IF at 2.6GHz, second IF at 110Mhz and third IF at 10.7MHz)  so you can not simply attribute all to one LO but the bandwidth of the first and second IF will impact contribution of the second and third LO. This still needs more investigation.

Now what has this to do with the choice of the first IF of the SA? 
Phase noise is caused by noise in the steering of the VCO in the PLL of the LO. If you have a high first IF you need a high output frequency from the LO, in this case of a ADF4351 and a first IF at 2.6GHz  no output dividers are being used. When using a lower first IF frequency (say 110MHz) and using the same ADF4351 the output divider will be 16 and this will reduce the phase noise.
To check this I measured the close phase noise when a first IF of 110MHz is used and you get below picture (first IF at 110MHz and second IF at 10.7MHz and no third IF)

The shoulders have moved down with about 25dB, still not as good as the VNA measurement.
The VNA measurement was done at 10.7Mhz so phase noise is expected to have less impact in the VNA measurement
 
But what about the far out phase noise?
As you can see in below graph the far out phase noise also has gone down.

The divide by 16 of the output of the PLL has increased the steepness of the fall-of of the phase noise.
The phase noise at an offset of 300kHz is at the level of the noise floor so a strong (0dB) signal 300kHz way from a weak signal will have no impact on the noise floor of the SA when using the first IF of 110MHz instead of 2.6GHz

All this implies when building your own SA you should not blindly go for the highest possible first IF. You have to understand the impact of the PLL in the LO's you use and their phase noise and the output divider in relation to the selected IF frequencies. In general having a high first IF will introduce more phase noise and this makes your SA less sensitive in the presence of strong signals. Its probably better to choose the first IF low enough for most measurements and use a down converter for the odd measurement where you have to go higher.

Testing the full chain of the SA

Now most of the modules are in, I combined everything into the complete signal chain.
.

The +10dB after the ADL5801 should not be needed but the conversion gain of the ADL5801 is -8dB i.s.o +0.8dB. I still don't understand why. 
The Arduino Zero controls the ADF4351's and the 12bit ADC measures the output of the AD8307 which gives a 120dB range and 0.05dB resolution

The settings are (speed is actually in 0.1ms units):

The noise level, 1dB CP and IIP3 performance of the whole chain has been verified by doing the performance measurement including two tone IP3 performance of each stage starting with the last stage and working towards the input
The SW allows any mixer to have any role and all frequencies of all filters can be changed. Unused LO's can be used as signal generators or tracking generator. Even after building the total chain it is stil possible to sweep each filter

Scanning 0 till 1Ghz without input signal gives

The many spurs around 100MHz are from the ADF4351 as it runs with R=10 from a 25MHz reference.
Spur performance of the black ADF4351 eBay module is not acceptable. I still need to investigate the loop filter and improve supply decoupling
The 433MHz spur is not yet explained. At some point I had a terrible spur at 42.2MHz. After some testing I found it to be caused the keyboard of the PC. Shifting the keyboard to the other side of the bench did remove the spur.
The signal around 950MHz is leakage from the mobile base station close to where I live.
The absolute signal level is calibrated using a calibrated generator (which was checked using a scope with a 50 ohm input) up till 12MHz but I still need to check the sensitivity at higher frequencies. 
Scanning again 0-1GHz after connecting a 20cm antenna to the input of the SA gives:

Below 250MHz its AM, FM, DAB and some analog TV broadcast. around 550MHz, 800MHz and 950MHz there are mobile base stations and digital broadcast signals.

Connecting two old analog signal generators with a passive -20dB combiner to the low-pass filter gives the two-tone test

Do I read this correct as an IIP3 of around +10dBm? At least changing the signal strength of one or both signals always results in the IIP3 of around +10dBm and that would be consistent with the performance of the weakest part in the chain, the ADE-25MH.
With the current spur performance of the ADF4351 the active mixers seem to be the best choice, they have more noise but their LO rejection is better so unwanted mixing with the LO spurs (specially with R=1) will not obscures the measurement.

I see the following next steps:
- Investigate why the ADL5801 has so little output, solving this allows me to remove the +10dB amplifier and will motivate me to buy another ADL5801 module to replace the ADE-25MH
- Solving the spur problem of the ADF4351 will allow me to use passive mixers everywhere, I have mixers with an IIP3 at +13dBm, less then the active mixers but with much better noise performance so in total more spur free dynamic range. I ordered a green ADF4351 eBay module that uses a 10MHz reference with possibly a better layout around the loop filter
- Find the cause of the 433MHz spur. 
- The 15kHz RBW filter as seen in the two-tone test has way too much BW below -50dB. The phase noise of the ADF4351 is much better (was verified using two ADF4351 modules 5kHz apart and a mixer to produce audio into an audio spectrum analyzer) I probably need to improve shielding

Verifying the IIP3 of a mixer

Measuring harmonic products with a Spectrum Analyzer (SA) requires understanding of the artifacts created by the SA. One of the well known artifacts is intermodulation distortion create by the non-lineair mixing of two signals in the first mixing of the SA. The dominant products are IMD2 (second order) en IMD3 (third order). IMD2 increases with 2dB for every 1 dB signal increase and IMD3 with 3dB for every 1 dB signal increase.

The key number for a mixer is the third order Intercept Point (IP3) where (theoretical) the signal and the harmonic products become equal in strength. The first mixer of my home build SA is supposed to have a input IP3 level of about 10dBm.

To measure this we need two independent signal generators generating equal streng signal with somewhat different frequencies, then combine the two signals and use a switchable attenuator to set this as input signal level for the SA (see above picture for " Key 3rd order terms" )

You need two measurements with different input levels that both show the wanted signals and the  unwanted harmonic products

First two signals at -30dBm, resolution filter is set to 15kHz (still not computer controlled)

The input signals are at -30dBm and the IMD3 products are at -77dBm, so -47dB below.

Now increase the input level to -13dBm

The signals are at -13dBm and the IMD3 at -36dBm, so -23dB below. So 17dB signal increase gives about 41dB IMD3 increase. 

Theoretical an increase of input signal of again 17dB would bring the input at +4dBM and IMD3 at +5dBm but the compression point is around 0dBm so this can not be directly measured

This is 5 dB below the spec of the mixer but not too bad. The current LO drive level of the first mixer may be a bit at the low side.
More measurements to do!