Flex Education: Noise Floor in Pan vs Slice

On the Alpha list, there was a great discussion that I thought would be a valuable read for the entire group.  I asked permission and wanted to put the flow here.   I hope you enjoy reading it as much as I did.  I removed the senders' email address to protect their Inboxes from spam.


On Jan 9, 2015, at 11:44 AM, XXXXX wrote:

I did some testing on v1.3.8 on my 6500 to try to determine why the dBm noise level in a Slice receiver is often greater than the dBm noise level of the Panadapter supporting that Slice receiver.
On 10meters using Xvtr with no antenna connected and RF gain=0 the Pan’s noise floor was a nominal -116dBm at about 620 kHz wide.  Zooming was kept constant in the following tests. 
  Now I thought that any Slice Rx created in that Pan, at it’s widest BW, would have a noise floor < -116dBm because the Slice’s bw is always < the Pan’s bw.  But that’s not what I found.

SliceRx BW Rx dBm meter
   175Hz -121.5
   350Hz -120
   700Hz -116.5
 1400Hz -113
 2800Hz -110
 5600Hz -107
11200Hz -104

As I halve the Slice Rx  BW from 11200Hz, the dBm meter decreases 3dB.  Fine business.  But how can it’s dBm level (-107, -104 ) be larger than the Pan’s noise floor of -116dBm??  

Now, as I further halve the Slice BW below 700 Hz,  the dBm meter drops approx.  in 3dB steps  below -116dBm which is expected.  So is the dBm level in the Slice passband the power level of the noise there only?  

 If I now reduce the Pan’s BW by zooming In (+), the Pan’s noise floor drops as expected. So is the graphically displayed  dBm level of the Panadapter , the power level of the noise of the Pan’s “passband” only? 

However, the Slice’s noise floor dBm remained the same when I zoomed  the Pan (good thing too ;-}.  Does that mean there is no connection between the Slice’s passband dBm power level and the surrounding Panadapter’s “passband" dBm power level ? 

How does one measure the noise floor dBm level of a quiet band?  Without including bandwidth, that question seems meaningless, yet we hams seem to have an intuitive idea of what a “quiet band” is.

 Please clarify this Steve and TNX.    

On Jan 10, 2015, at 7:35 AM, Graham Haddock <****> wrote:

As you have figured out, when you are measuring noise, knowing the bandwidth is a necessary part of the measurement.

A signal has a specific level, as long as the measurement bandwidth is wide enough to enclose the entire signal, and all the modulation sidebands.

Noise has an energy level per unit bandwidth, and if you double the measurement bandwidth, then you double the energy inside the measurement bandwidth and the measured level goes up 3 dB.

The panadaptor is the graphical plot of the signal strength in an extremely large number of very narrow receivers. Each receiver is called a "bin" from the mathematics of the process used to generate them. So, think of the panadaptor display as the graphical plot of the levels in each of several thousand receivers (or bins) side by side on different equally spaced frequencies.  The measurement (bin) bandwidth is on the order of 1.5 Hz to 10 Hz, depending on the zoom level of the panadaptor.

As you zoom in and out, the bin bandwidth goes up and down, but it is not directly connected to the slice receiver bandwidth, since the panadaptor is an entirely different receiver structure in the software, and makes its own decisions as where to change the bin or measurement bandwidth.

If you assume a measurement bandwidth down in the range of single digit Hertz, then the levels you are being presented will make sense.  Depending on the zoom level, you are dealing with 2000 to 4000 receivers (bins) and individual bin bandwidths of 1.5 to 10 Hz.

It is hard to fool Mother Nature.  No laws of physics were broken, mutilated, or killed in the operation of the Panadaptor.  :-)

--- Graham / KE9H

On Jan 11, 2015, at 9:01 AM, Stephen Hicks, N5AC  wrote:

​Just seeing this.  I see Graham has made a good response that tells you most everything here.  I can't resist this topic and will add a few more details below:​

​As Graham mentions, the panadapter is actually a collection of receivers that slice up the band.  We use it to measure the energy in a sliver of the band and then display that.  We are currently using a fixed number of bins (or slivers) of the band at 4096.  Incidentally, all of the bins are not usable -- please see the side not at the end "Why can't I use all the bins?"  We typically use about 60% of the bins, but this changes at certain zoom levels.  

So when you zoom to a width on ​the screen that is 100kHz wide, you can figure that our bin WIDTH is roughly 100kHz / 0.6 / 4096 =~ 40Hz.  This is not exact as we employ a complicated algorithm underneath to give you the feel of continuous zooming.  There are a million ways to build a panadapter.  I could give a 3-hour talk on all of the methods that could be used and what the trade-offs of each method are!  But this number will get you close to the actual bin width.  I can also say that the exact bin widths are all a multiple of a power of two times 5.8594Hz -- in other words 5.85Hz, 11.72Hz, 23.44Hz, 46.88Hz, etc.  So the 40Hz we just found is actually either 23.44 or 46.88Hz exactly.

Anyway, let me use a cooking analogy to try to help you understand noise measurement.  Lets say that all the noise you can see in the bandwidth of your ADC can be represented by a 1/4c (cup) of water.  This would be the equivalent of taking the bits directly out of the ADC, converting them to ONE BIN and taking a wideband noise measurement.  Let's say that we will be using a 1c measure for a bin.  When you take the 1/4c of water and pour it into the 1c measure, it fills it about 1/4 of the cup and we have a high noise level.  Let's say our signal is a Hershey's Kiss and it is setting in the cup currently, but the water line is ABOVE the top of the kiss.  There is a signal there, but you can't see it (pretend that water is opaque or use milk if you're trying this experiment at home ;-) ).

If we run a 10-point FFT to look at the spectrum, we take that same 1/4c of water and divide it across ten 1c measures.  Now, each cup has 1/40c of water and we have "lowered the noise floor"  Now the Hershey's Kiss sticks above the water line.  The more bins we use like this, the further out the kiss sticks and the more we can tell the true nature of the Hershey's kiss (he has bad keying sidebands, it seems).

If you imagine 4096 of these cups, when you zoom in and out using the bottom of the spectrum (zoom slow), we are just hiding and showing you more or less of the bins, but the bins are the same size, hence the noise floor stays the same.  But occasionally, you hit the edge of the bins we show you and we switch from bins of one size to bins or twice or half that size and the noise floor jumps by 3dB.

* * * * *

If you are measuring noise, it behaves as I have described above.  If you are measuring a monotonic single tone signal, however, it behaves differently.  Let's talk about how and then discuss why.  Let's say that you stick a signal generator into your radio and set it for -40dBm.  You will see a sharp spike in the panadapter with the noise many dB below on the sides.  If we zoom way out where the bins are several kHz wide (remember from above if we zoom to, say 7MHz, our bin size might be 3kHz.  We take a measurement in the radio by looking at the panadapter and see that the signal is at -40dBm.  As we zoom in, though, it stays there.  One click and our bin size is now 1.5kHz, but the signal is the same ... why?  Imagine that Hershey's Kiss again and as we change the water level in the 1c measure in minuscule amounts, we can still see and measure the top of the Hershey's Kiss and can tell exactly how tall it is.  If your signal fits entirely in the bin and it rises our of the noise 10dB or more, as you adjust the noise around it by changing the bins size, the signal is still the dominant energy.  Back to the kitchen -- imaging that we are changing microliters of water in the cup and we are measuring the WEIGHT of the volume in that 1c measure.  The Hershey's Kiss will always dominate because it weighs so much more.  In the case of the radio with noise, a -40dBm signal "weighs" 80dB or 100,000,000x more than the noise typically does.  This is like adding a gnat to the weight of a car and expecting a change in the scale measuring the car!  

The point of all of this last paragraph is to show than when you are measuring a signal, the readings in the bins (and the receiver) show the same number as you adjust the bin width.  But when you are measuring the noise, it is billions of little signals all on various frequencies and as you split the bin, half of those signals go to one bin and half go to the other and so the level goes down by 3dB (half)!

 
Now, as I further halve the Slice BW below 700 Hz,  the dBm meter drops approx.  in 3dB steps  below -116dBm which is expected.  So is the dBm level in the Slice passband the power level of the noise there only?  

Yes -- the slice receiver is a separate receiver from the 4096 receivers in the panadapter.  It's level is completely independent, BUT they have the same source material that they derive their ​data from.

If I now reduce the Pan’s BW by zooming In (+), the Pan’s noise floor drops as expected. So is the graphically displayed  dBm level of the Panadapter , the power level of the noise of the Pan’s “passband” only? 

Well, the passband of the bin --​

However, the Slice’s noise floor dBm remained the same when I zoomed  the Pan (good thing too ;-}.  Does that mean there is no connection between the Slice’s passband dBm power level and the surrounding Panadapter’s “passband" dBm power level ? 

YES!​

How does one measure the noise floor dBm level of a quiet band?  Without including bandwidth, that question seems meaningless, yet we hams seem to have an intuitive idea of what a “quiet band” is.

Hams always use the term noise floor, but what they mean is "noise floor in a 500Hz bandwidth."  If you want to report "the noise floor" to a ham, set your slice receiver to a 500Hz filter and then take the measurement.  This is what we do in the lab and it's what the ARRL does.​

Hope this helps.

​73,
Steve​



On Jan 11, 2015, at 10:41 AM, Stephen Hicks, N5AC wrote:

oh forgot the sidetone...Why can't I use all the bins?

In a sampled system, energy that appears above the sampling frequency aliases into the first Nyqvist zone.  For example, if you sample at 100MHz, a signal at 75MHz will appear at 25MHz (mirrored about half the sampling frequency) just like it was generated at 25MHz.  You can read a lot about this on the Internet so I won't go into it in detail here.  In the panadapter, we can zoom in to a frequency range by halving the sampling rate and filtering out anything that would alias into the band.  This process is typically called "decimation" although it has nothing to do with the pure definition of decimation (dividing by ten).  When we do this, we must filter out everything that we do not want to alias into our sampling area.  

Let's use our example from above: we are sampling at 100MHz and our first Nyqvist zone is 0-50MHz.  Signals at 55MHz will show up at 45MHz.  Signals at 51MHz will show up at 49MHz.  This is aliasing.  To prevent this, we "pick" which zone we want to look at - let's say Zone 1 which would be 0-50MHz.  The way we pick is to filter before we sample, eliminating any signals from 50-100MHz so that we are sure that what we see is actually in the 0-50MHz region.  We do this with a real hardware filter which is called  an anti-aliasing filter.  Because filters have transition bands, we will still see aliasing in the transition band.  For example, a signal at 51MHz won't be cut to zero, our filter might reduce it by 10dB there, 20dB at 53MHz, 30dB at 70MHz, etc.  So a signal that is -40dBm at 53MHz will still show up as a signal of -40dBm - 20dBm = -60dBm at it's mirror frequency of 47MHz.  For this reason, we don't have access to all the Nyquist zone.  A radio designer would generally claim something like 0-40MHz instead of 0-50MHz and he would ensure that signals from 50-60MHz might alias, but ones from 60+ MHz would be down sufficiently at 40MHz and below as to not be visible.

OK so in DSP, we do all the same things, but they are not physical hardware filters, they are math.  The cost of a DSP filter is 1) processing effort, 2) latency, 3) power & heat which comes from (1).  To avoid trying to do something Herculean that costs lost of latency, power and money we build modest filters in DSP that have acceptable roll-off.  Often times these filters make for large transition bands, or areas where we can't count on no interference from outside the filter.  In this case, we just discard anything in that range.  And this is what happens in the panadapter -- we filter, but discard anything in the transition band.  For the panadapter, we use a couple different kinds of filters -- one of which discards 40% of the collected data and the other discards 18% of the collected data.  So we are clear, we never show you any data from the transition bands -- the software in the radio always hides this so you only see what really exists on the bands.

You might wonder why the pandapter in the radio is so good knowing what I've just told you -- we use a 4096 point FFT whereas PowerSDR did an 8192 point filter and we even throw away up to 40% of the data (although when you zoom way in we use much more of the data).  How it possible that the panadapter is so much better than PowerSDR?  The thing about PowerSDR is that it does only one FFT at the sampling rate you have selected.  So let's say that you are at 192kHz sampling rate in PowerSDR.  The bin size will always be 23.43Hz.  When you zoom in, you are just spreading the bins available over more pixels, but you get no better resolution...  Since we can see down to 1.46Hz bins in SmartSDR on a FLEX-6700, we have a 12dB visual advantage over PowerSDR.  The panadapter literally sees 12dB further into the noise.  By always adjusting the sampling rate in SmartSDR, we can keep showing you better and better data in the panadapter and also allow you to see wide bandwidths when you want to.

Steve

On Jan 16, 2015, at 10:35 PM, Stephen Hicks, N5AC wrote:

On Sun, Jan 11, 2015 at 7:59 PM,wrote:

Are representations in the waterfall exactly the same resolution (detail) as in the panadapter as you describe here?

​They are currently derived from the same FFT, but they have a lot of time integration which reduces noise variance and reveals more in the noise floor.  This only gets you so far, though.  To see further, you do have to reduce the bin size.  We can already see 25dB into the noise floor which is around where JT65 can hear.  The time averaging probably gets you another few dB.  Generally you can see anything you can work on JT65.  If you could see further into the noise floor, I'm not sure how you'd work those folks -- I guess you could do long averaging, etc. using a special mode.

​Steve​