Re: [Discuss-gnuradio] A longish-term drift study of the USRP/DBS_RX combination

[mike revnell]

The astronomy club at the local college (New Mexico Tech) has a little 
two element interferometer that I can get access to. I'm supposed to be 
working on some software to point the antennas now. It has two 10 ft 
dishes with about a 200 foot baseline. I bought two dbsrx boards with my 
USRP with the idea of maybe hooking it up one day and have been giving a 
little thought about what is needed to get fringes. Getting some kind of 
image out of it is another thing entirely.

Using it as a transit instrument (point the antennas south at the 
transit elevation of the source and letting the source drift through the 
field) should be relatively straightforward. Getting fringes while 
tracking the source across the sky is more involved.

The receivers presently installed form an adding interferometer which is 
nice and simple and don't require stable LOs and all that. The signals 
from the antennas are combined at the receiver input and the whole mess 
is downconverted together. Interpreting the data is simpler as well. But 
a real correlator is more interesting.

In looking at the schematics for the USRP and DBRX it looks like all the 
LOs are derived from the same reference oscillator. This is a 
requirement for correlating at baseband and would be a problem to solve 
if using more than one USRP. Mine looks like it has a footprint for an 
SMA to apply an external reference but it is occupied by a piggyback 
that has a little crystal oscillator on it. It seems to be a fix for a 
problem on the main USRP board.

Ultimately performance will probably be limited by phase noise in the 
reference osc and synthesizers in the assorted downconverters before the 
ADC. No one is expecting to reach the stability and sensitivity of the 
VLA though :-)

On a strong source with a transit instrument it should be possible to 
use short integration times and record AC fringes at the correlator 
output. This is the simplest case. Just FFT both basebands multiply one 
FFT output by the complex conjugate of the other and take the magnitude. 
The inegration time is the length of the FFT input sequence. Could 
average or filter by frequency bins or integrate over the spectrum and 
work with that.

Normal practice is to tune the LOs in the antennas to move the fringes 
for a give piece of sky down to DC. This allows integrating for longer 
times using hardware accumulators. This can be done, as in the VLBA 
correlator, at the FFT input. The frequency translating FIR block should 
work nicely but it may be necessary to work out some way of controlling 
the relative phases of the chains in the two antennas. LOs in radio 
interferometers need to be continuous phase. This means that if you move 
from a source, to a calibrator say, and move back you need to be able to 
put the LO phase back to where it would have been if you had continued 
to track the source. This has a profound impact on the design of the LO 
synthesizers.

To track a source across the sky things get even more interesting. When 
signals from a piece of the sky don't arrive at the correlator at the 
same time things start to decorrelate. With an FX correlator this is by 
a factor of how much the sequences at the input to the FFT don't 
completely overlap. For a single baseline that's only a couple hundred 
feet this won't be a problem. The delay correction in this case can be 
done at the FFT output before correlation. A single sample delay change 
at the input of the FFT is one full turn of phase in the output.
This is also done in the VLBA correlator.

Interestingly, the largest correlators that I'm aware of that are being 
designed and built are XF architectures.
The correlator for the ALMA project being built at NRAO in 
Charlottesville VA, and the WIDAR correlator for the EVLA project being 
built in Penticton, BC Canada. A smaller version of the WIDAR design is 
to be used in the EMERLIN project at Manchester, UK. The reason for this 
is that FX correlators tend to be what people who architect these things 
call copper dominated. It takes a lot of interconnects to get the 
results from the frequency analyzers (usually hardware FFT engines) to 
the core correlator. XF correlators tend to be what's referred to as 
silicon dominated, there is a lot of processing in the correlator. Right 
now silicon is cheaper than copper.

Ayway, enough of my noise for now.

---------------------
BTW, I'm the correlator group leader at NRAO socorro. We maintain the 
VLA and EVLA correlators and various other things as well.

Marcus Leech wrote:

> Now that I understand a bit more of the internals, it's really 
> starting to gell for me.
>
> Some things I'd like to add:
>
>       o interference stopping
>
>       o recording the FFT results, without having to define a whole 
> new chain (I'm at 50-60% CPU already!!)
>          [I might pull the same trick here that I did with 
> ra_stripchartsink--define a "calibration" function that
>            can record data as a side-effect.  I'd simply make an 
> ra_fftsink that does this].
>
>       o Possible integration of PyFits (although whether it makes 
> sense to do it in real time remains to be seen)
>
>       o Deriving the total power from the FFT results, eliminating the 
> parallel total-power chain
>
> It would be interesting and cool at some point to demonstrate an FX 
> correlator using a pair or trio (or quad, or pentad :-)
>  of USRPs with appropriate front-ends and antennae.  In an FX 
> correlator (which is a style that's currently in vogue)
>  you compute the FFT before computing the various cross products.  It 
> has more manageable data structure than
>  X-F correlators for RA use.  You can see that you can martial the 
> various FFT bins off to different CPUs for
>  further processing in an FX design, and thus control, to some extent, 
> the herculean data flow that occurs in
>  these RA systems...