Continuing with experiments and spur measurements, I found that closing
the lid on the little filter box does seem to reduce the LPF's
effectiveness at the higher frequencies, but leaving it open reduces
effectiveness at the lower. I can sculpt it to a taller structure if
necessary, which would give more clearance to the choke, and allow for
some microwave absorbing material. Yes Askild, I did manage to squeeze a
little strip into the existing can, but only near the output end -
there's no room above the parts, especially the choke, so pinching
anything there would likely spoil the whole thing. For now, I have the
lid closed, and the absorber strip at the output area. The hinge of
course has already broken, but I just tack some solder gobs over it as
needed.

What I discovered next, however, may mean I won't have to improve this
box anyway. Doing a spur review, I found that the remaining significant
ones were the 70 MHz, and all but one of the ones between 1260 and 1820
MHz. Above and below, everything else was in the noise floor. I had
gradually worked the 70 MHz down some with shielding and such, but a
little remained even all closed up.

I started thinking again about possible resonance of the cable from the
mixer to the LPF. The length is in the right ballpark to aggravate the
problem spur range. The only reason for the length was to get the
desired three turns on the CM choke, so one option was to give up most
of the choke value and go one turn, with a short, straight connect to
the LPF, which would force any resonances way upward (but maybe they'd
just show up elsewhere, if there's not enough loss at the higher
frequencies). The other option was to revisit padding the input of the
FPF, or the diplexer again.

After thinking back over previous experiments with these, I recalled
that I was really only looking for noise flatness then, and hadn't even
gotten to detailed spur measurements. I also recalled that the original
diplexer setup did interact some at the top of the noise band - I chose
a 5th order 50 MHz .05 dB Chebyshev response, but the real parts made it
something a little different. I realize now that I should have just
disconnected it and left it in place - unfortunately, I took it all out
during my last cleanup and consolidation round.

So, not wanting to change too much around, and only guessing about the
cable situation, I figured on trying something simple and quick to
diplex out the upper stuff, to suppress possible resonances. I chose
somewhat arbitrarily a 3rd order Chebyshev around 140 MHz, which is
where the upper image lies, and the choke is about 50 nH, and I just
happened to have another of the same part I had measured around this
value and used in the 300 MHz LPF. So, in went a 22 pF/50 nH/22 pF/51 R
HPF, and out went the reflections.  All the bad spurs are in the noise
floor in a broadband view, but can still be found with narrow band spot
checks, around -95 to -100 dBm. The net reduction from previous "feels"
like maybe 6 dB, which I think corresponds with cutting VSWR in half.
Interestingly, the 70 MHz is now virtually gone too.

So, it looks like my first instinct to have a diplexer was right, but I
didn't study it deeply enough, and my assumption about its effect on the
spurs being small was wrong. Now that I can see some results, I can set
the HPF a little lower to at least terminate the entire upper image
(about 115-165 MHz), but not so low that it interferes the LPF response.
Another interesting thing is although the upper image is the second
biggest after the desired output signal, it has never shown above the
noise floor since I installed my new LPF to replace the commercial one.

Ed

Yes, that transformer sure looks burned out. It's hard to tell how big
it is from the pictures, but my impression is that it looks kind of
skimpy to run a FRK Rb plus whatever else is going on like a GPS RX and
uP system, and maybe battery charging too. You can easily estimate the
VA rating by measuring the dimensions and comparing to standard
transformer frame sizes. Generally, the VA rating should be at least
twice the total raw (not the regulated output values) DC power produced,
with conventional rectification and filtering. This can be exceeded for
a while, say during warmup of the Rb, as long as it goes back to normal
in a reasonable amount of time. It's mostly about temperature rise - if
you have good cooling, you can get more out of it.

Transformers are pretty tough, so having one burn out in normal service
calls for some investigation of why it happened, before you risk taking
out a replacement too.

Regardless of the VA rating that should be used, you're probably stuck
with using the same size and style as the original, just to fit it
mechanically. If it's plenty big enough VA-wise, then all's well. If
it's marginal, you can at least add enhanced protection to avoid another
burnout.

Regarding DC supply voltages, the main one will be something around 24 V
for the Rb. I would guess that the DC-DC converter on the supply board
makes +5 V (or 3.3 or whatever) for the brain and GPS RX, and the 78M12
makes +12 V for the analog, and that there are no negative supplies -
unless there's more to the supply system that's not shown. Since
external 24 VDC can supposedly run the whole thing, I don't think you'd
have to worry about making any of the voltages from the AC transformer
except for the 24 V, even though it appears to have a multi-tapped
winding. I didn't see anything in the OP about whether the thing works
with just external DC, so this should be confirmed.

There's a lot more circuitry on the board than seems necessary just for
power, so it may be worthwhile to reverse engineer it a bit - especially
the four big transistors and U3 and U4, which looks like two identical
functions of some sort. Maybe extra voltage regulation, or maybe 1 PPS
amplifiers?

Once you do figure everything out and get a fresh transformer, note that
the original is banded to reduce magnetic emission. It appears to have
both the copper strip around the bobbin zone, and the steel (or
sometimes mu-metal) band around the core, but not the third thing
commonly done, which is insulating the core mounting. It will function
without these, but may interfere with the Rb unit, especially if it's
nearby. You won't find these features in run of the mill OEM replacement
transformers, so you'd have to specify them, or add them yourself. If
you get a transformer with same dimensions as original, you can
transplant these pieces from the old one.

Ed

This may give some idea of how fast things can happen when the OCXO is
subject to drafts. I have this dual GPSDO box that usually is open for
experimenting, and have a setup comparing one of the 10 MHz outs to my
portable Rb reference. The 10 GHz multiplied output from the Rb is
indicated on a microwave counter, using the GPSDO as reference. This
gives 1 mHz resolution on the 10 Mhz signals at the 1 Hz counter
resolution limit. It normally reads 10 GHz "exact" +/- 1 Hz when things
are stable, or up to maybe up to 2 Hz when garage ambient is changing. I
just turn the counter on whenever I'm in the mood to take a look.

The upper GPSDO board is exposed, so I can just put a finger on the case
of the small (about 1" x 1.5") OCXO for a few seconds. Almost
immediately, the counter shows several Hz change, which gradually
recovers, with some over- and under-shoot. During all this, the OCXO is
changing, and the GPSDO is trying to fix it.

Having a bigger OCXO with more thermal mass and insulation, and having
more protection from fast ambient changes can help a lot. As others have
said, you don't want to overdo it - the oven heating system must be kept
working under all conditions, but it's OK to make it not have to work
too hard.

An extreme example of a bad thermal situation is in the beloved HP8566.
I have often lamented about the poor placement of its internal OCXO,
which is right in the main air plenum that feeds the fan cooling air to
the whole instrument. The OCXO is subject immediately to any change in
ambient, and its heater has to work very hard. I'm convinced that this
is the cause of most OCXO failures in the 8566. I've had to refurbish a
number of these. The typical failure I've encountered is that the foam
insulation deteriorates from the high heat flux needed, and the
chemicals from the foam cause the oven setpoint adjustment pot wiper
contact to fail. An easy way to spot this problem is to gently shake the
OCXO - if you can hear and feel the guts clunking around inside, then
it's due for repair.

At an opposite extreme, in my "Z3801A in a HP5065A carcass" project, I
substantially isolate the OCXO from ambient. It's already a double-oven
style, and I further enclosed it in a mu-metal box (made from a CRT
shield). The OCXO is suspended on rubber vibration mounts, inside the
box, and has a thin (~1/4") layer of non-woven fiber insulation on all
sides between it and the box. The insulation has very little R-value,
but suppresses turbulence and convection flow inside. The Z3801A guts
are arranged specially to fit and occupy about two thirds of the cabinet
volume, and this section is largely sealed off from the outside and from
the right side battery compartment. A small fan runs at very low speed
to gently circulate the air inside the compartment, and the plentiful
amount of cabinet skin easily dissipates the total power. The same type
of insulation is also placed under and atop the main board in the
DAC/EFC circuit area, to slow down thermal changes there. The EFC's SMB
connector set will also be shrouded with an insulating tube, to reduce
thermal voltage. I even changed the nearest board mounting post to
plastic, to reduce effects of thermal conduction and ground current in
the vicinity.

All of this does not protect from ambient, but only the rate of change.
It's more or less a constant temperature rise type deal, assuming
constant power dissipation when everything's stable - and not too much
wind or draftiness on the whole cabinet.

Ed

Erik, I'd really recommend that you use a real, "solid" ground reference
on the instrumentation side, with +/- large (12-20 V) supplies, as
others have suggested.

Your most recent setup diagram indicates that you're relying on the
"differential" input of the audio PC card etc analyzer to allow for the
"floating" common of the analysis circuit. Do you know what the
common-mode rejection characteristics are? A true differential input
would have two coax lines entering a symmetric differential to
single-ended conversion stage at the front end. I doubt that the PC card
actually has this, but maybe some form of DC/LF isolation from the local
input common to chassis ground.

The PC likely has lots of SMPS noise in common-mode form, which probably
can be ignored for audio (the SMPS frequencies are almost always quite
far above audio). As long as the interference signals aren't too big to
upset the LNA operation by say, rectification in various junctions
(especially the front end), it should be OK. You will also have in-band
line frequency and harmonics present in the common-mode signal, but
these should be easier to deal with by virtue of whatever LF CMRR the
sound card does have at lower frequencies.

Now consider the analysis circuit environment, where you have apparently
zero intentional bypassing capacitance from the floating measurement
common to chassis/earth ground. Here, the only bypass caps effectively
are C1 at the REF buffer's input (which will only aggravate the
situation), and the small capacitance between the ports of the mixers. I
believe you have some bypassing at points in the other portion of the
circuit - the PLL for the reference - but I don't know what that looks
like now. So, just looking at this section, I'd say you need some
serious bypassing to ground, for the RF signals from the mixers, and the
common-mode signals in and out of the audio analyzer, DUT, and REF.

I recall there were some recent discussions about rail-splitting and
such, but I didn't look closely. I thought surely someone would have
mentioned the simple way to rail-split with an opamp, into a large
capacitive load, but maybe not.

Without resorting to a more desirable ground-referenced, +/- supply
scenario, you can add significant bypass capacitance from the signal
common to ground, with slight change to the buffer circuit.

1. Add a resistor between the opamp's output and the load, which is
   signal common. The current demand appears small, so maybe around a
   couple to few hundred ohms should do.

2. Add a resistor in series with the (sense line) inverting input. This
   can be in the many k ohms range, depending the opamp's bias current.

3. Add a small capacitor between the opamp's output and inverting input
   to stabilize it.

4. Add the bypass cap.

This setup just isolates the opamp from the capacitive load, with the
LF/DC regulated by the opamp, and the HF shunted by the bypass cap.

I'm guessing that once you get good bypassing here, the LNA will work
much better, and you should see the difference with the lower noise
opamp. The reason is that any opamp has limited CMRR, so improving the
bypassing makes the "CM" part smaller. This is also another reason to
operate opamp inputs at or near ground. Actually, the best CM
improvement can be provided by running in inverting mode, so both inputs
are always at ground. Non-inverting modes require the inputs to move,
depending on the signal. In your LNA, the CM input signal range is not
too bad, due to the high gain. The trick is to keep the overall CM - the
operating common level wrt ground and the power supplies - constant and
noise-free.

Regarding microphonics, since you mentioned tapping the housing, it
sounds like you have "canned it up," which is a good thing. Assuming the
REF and DUT are external, so not involved, the audible is coming from
the analysis circuit only, right? That's not too surprising since it's a
high gain system. It could be related to individual component
microphonics, but I'd guess it's an RF effect. The whole thing is awash
in the 2f signal and harmonics from the mixer, and to a lesser extent
the DUT frequency signal that leaks through, so mechanical dimension
changes or movements in the can, board, wiring etc, can change the EM
pattern inside, giving tiny, noticeable phase shifts - after all, that's
what it's for.

Ed

I forgot to mention that you should also consider possible effects from
the RF present, on the LNA. This can be more significant than SMPS
frequencies getting where they don't belong, especially since the RF is
intentionally right at the opamp's input. Your LPF only reduces, and
does not eliminate, the 2F and harmonics, so there can be significant RF
present on the LNA circuit.

A simplistic view is that the RF is far beyond the opamp's GBW or closed
loop gain and should have no response, but it's not at all beyond
upsetting or altering the operation. This can result in extra DC offsets
and noise due to RF rectification in the input circuits, which only
remain "linear" at frequencies where the output and feedback can keep up
with the input.

This can be fixed if necessary, by adding extra RF filtering,
particularly some built to low-pass at a higher cutoff frequency well
above the analysis frequency, and well below the expected f and 2f.

For instance, in your circuit it looks like L1 is 1 mH, with 100 nF
caps, which ideally cuts off quite low. However, 1 mH is a pretty big
choke, and will tend to have a lot of inter-winding capacitance (and
high resistance - don't forget to include it in noise), making it less
effective at the higher frequencies. Adding an LC section in front of
it, but set up for something in the MHz region, will give much greater
rejection of the f and 2f, due to having more appropriate smaller L and C.

Anyway, if it works fine as is, then no problem, but it's something to
be aware of if you get strange effects down the road.

Ed