[time-nuts] fast switching quiet synthesizer

Jim Lux jimlux at earthlink.net
Wed Oct 8 10:48:20 EDT 2014

On 10/8/14, 7:21 AM, Bob Camp wrote:
> Hi
> On Oct 8, 2014, at 3:52 AM, Hal Murray <hmurray at megapathdsl.net> wrote:
>> bruce at ko4bb.com said:
>>> Kratos   (www.kratosepd.com) do fast switching synthesiser subsystems that
>>> can be locked to a reference..
>> What does "fast switching" mean in the context of a DDS?
> Not much at all. Your DDS clock will be something like 1 GHz and it will “switch frequency (actually it switches phase)” on a clock edge.  You probably can’t load a billion updates a second through the update port. People do indeed load millions of updates a second.The DDS is quite happy to “follow” the updates. It’s one of the ways to get a clean(er) 10 MHz reference signal out of a DDS on a GPSDO.
>> What does the spectrum of a DDS look like if I switch back and forth between
>> 2 frequencies at 1 KHz?
> Since it is phase coherent FSK with instantaneous phase change, it’s pretty much a textbook spectrum. Lots of energy at the switch points. It’s likely closer to the textbook than anything you can get out of a conventional (non DDS) signal generator.
>> Or what question should I be asking?
> Is that waveform (and it’s spurs) OK in the application or not. Only Jim knows for sure …

It's a Stepped frequency CW radar..you make amplitude and phase 
measurements at a series of frequencies, then do an inverse FFT to 
transform to the time/range domain.. just like a VNA doing time domain 
measurements.  It's a flavor of FMCW radar like used in radar 
altimeters. The system is basically a homodyne (think cops radar speed 
gun), so you're detecting the received signal by mixing it with the 
transmitted signal and looking mostly at the phase.

  Close in phase noise tends to cancel since ranges of 1000 meters are 
only 6 microsecond delay.  Likewise, close in spurs tend not to have a 
lot of effect.  I've used a bluetooth transmitter as the source for a 
very short range (few meters) experiment.  The PSK modulation on the BT 
cancels itself out.

The real issue is that you want as much "time on target" as you can get. 
  So if you're stepping every 1 millisecond, and you have to wait 300 
microseconds to settle, you're really only measuring for 700 
microseconds, so you're giving up SNR (assuming fixed Tx power, Rx noise 
figure, etc.)

There is an alternate architecture we want to experiment with where the 
receive LO is offset from the transmit frequency by a small amount: 
this shifts the detected signal away from baseband, so you don't have to 
deal with DC offsets in the I/Q detection, and in fact, you can use a 
single mixer instead of a quadrature demodulator.  These systems have 
outputs in audio frequencies, so going to offset rather than baseband 
also means you can use one ADC instead of two.  That would make the 
eventual system smaller, lighter, cheaper, etc.  (with the increased 
complexity of the offset oscillator and mixers).

And finally, the ultimate range resolution of a stepped frequency radar 
is determined by the number of steps. But you need a certain time at 
each frequency to make the measurement (SNR driven, mostly), and if you 
want to deal with a target that is moving (aren't most), you need to 
have the entire "step sequence" be fast enough that the target hasn't 
moved "too much" in that interval.  So the ability to do multiple 
frequencies at the same time is desirable (if you have 4 transmitters, 
and 4 receivers, you can have 4 times the number of "steps/second")

There's an interesting tradeoff among all the various system parameters: 
how accurate is the step frequency (small errors are like noise and 
spread the target out), how many steps (sets the number of resolution 
bins), what's the bandwidth (sets the size of a bin), what is the 
ambiguity range (if you have 3 MHz steps, and 100 steps, then you have 
0.5 meter resolution, but the bins repeat every 50 meters, so a target 
at 105 meters looks like it's at 5 meters).

You can also do cool stuff like run two sequences simultaneously, one 
with big steps (good range resolution, but short ambiguity) and one with 
small steps (large unambiguous range), or non-uniform sequences (e.g. 
spacing the frequencies with log steps).

And, of course, there's no particular reason you need to step in any 
particular order, or even that the frequencies be exactly the same on 
each "group" of steps, so you can use pseudo random schemes to do code 
division multiple access.

One case where good phase noise (and low Adev) is important is when 
you're doing it bistatic: you don't have that sample of the transmitted 
signal to use as the LO for the receiver.  Now, the transmitter and 
receiver oscillator noise is important, and if you're doing the usual 
baseband detection, the close in noise is in band for the signal you're 
detecting.   That's where the offset frequency approach comes in handy. 
  But still, you're ultimately making a phase measurement between 
transmitted and received signal, so the two separated sources have to be 
phase coherent (or at least "known" phase offset)

More information about the time-nuts mailing list