Jerry points out a real weakness of many people's approach to FFT analysis -
windowing. When I was doing vibration analysis for Kodak we used a B&K 2032
audio spectrum analyzer, which is of course an FFT analyzer. Later we moved
up to their multiple channel in, multiple channel out FFT analyzer (whose
model number I don't recall). I was able to borrow the 2032 a couple times
and put on demonstrations at our local ham club showing filter shapes and
the like.
As a full time analyst (at that time) I was sent to school several times to
learn FFT analysis. One of the first things we learned was about windowing
of the signal. As Jerry says, the results are useless without this. And if
you do not select the right window, the right sampling rate, the right
number of averages, and a host of other parameters, you will get garbage.
However, often the garbage looks reasonable - its just wrong. :) This was
fun.
We also used software to post-process the signals. My favorite was the well
known MeScope software, written by Mark Richardson, of which I was a beta
tester. Recently I discovered that this software is being used today to
analyze the vibration signatures of wind turbines, as part of a predictive
maintenance program. I ought to call Mark and say Hi and ask how that
application is working out.
I agree that I doubt very much that there is enough computing horsepower in
the Orion to take this brute force approach.
I'd would STILL like to hear from Ten Tec who could lay the whole argument
to rest with an authoritative response.
Oh well.
73 de Gary, AA2IZ
----- Original Message -----
From: "Dr. Gerald N. Johnson" <geraldj@storm.weather.net>
To: "Discussion of Ten-Tec Equipment" <tentec@contesting.com>
Sent: Saturday, December 09, 2006 12:13 AM
Subject: Re: [TenTec] Noise Reduction Setting
> On Fri, 2006-12-08 at 17:58 -0600, Merle Bone wrote:
> > Jerry said:
> > "While the effect on the output spectrum is that of a bandpass filter, I
> > don't think NR works that way. I think it works on correlation of the
> > pass band with a time delayed copy of the pass band."
>
> --------------------------------------------------------------------------
-----------------------
> > Jerry, I am curious as to why you believe this? Are there other amateur
> > transceivers - ICOM or Yaesu - where you have seen the technique you
> > described used before?
> > Thanks & 73,
> > Merle - W0EWM
>
> Primarily because the books I've read on noise reduction tend to
> emphasize autocorrelation as the technique with the main variations
> being the selection of the time delay and judging the results to adjust
> that time delay. There are several books in the ISU library on noise
> reduction, maybe a shelf full, I've not read them all, but I might if I
> get a start on my own receiver design, providing I believe that noise
> reduction is a needed thing the DSP can do better for me than my own
> ears can do with practice. I know my Timewave DSP won't find a weaker
> signal than I can detect by ear, but it will improve the S/N of any
> signal I can detect to make copy take less effort.
>
> I propose a test to compare how a narrow filter might work and how
> autocorrelation might work. Look for some weak multiple tone signals,
> perhaps a digital multiple tone signal or many weak signals in a DX pile
> up spread over the receiver bandwidth. Attenuate those signals to be in
> the noise. Apply the NR. Autocorrelation has a chance of enhancing all
> the CW signals, while a single adaptive filter can only do ONE. It takes
> as many parallel adaptive filters as you have separate signals to go the
> filter approach and I don't think there's enough DSP in your receiver to
> do that while there can be an autocorrelation time delay that enhances
> all the tones. It may be long because the tones are not coherent with
> each other and the autocorrelation delay has to find the delay that
> amounts to full cycles for all the tones at the same time. This test
> might work better with two or three than with 50 signals.
>
> Another adaptive filter technique could be doing a FFT of the signal and
> reconstructing the signal peaks in the output. I don't think a radio DSP
> chip has the compute power for that. And if the sample interval isn't
> long enough (latency) and the window function isn't well chosen the FFT
> can introduce MORE noise than was in the original signal. There's this
> little detail that the math is based on a continuous signal from time =
> - infinity to + infinity but the sample interval is much smaller. If the
> amplitude and the slope of the beginning and end samples of the interval
> don't match thats introducing a step function into the computation which
> shows up as broad band noise. Many who apply the FFT to signal analysis
> miss that detail and get useless results.
>
> If the FFT sample interval is an even second, but there's a tone of 50.5
> Hz, it will show up at a greatly reduced amplitude compared to a tone of
> the same input amplitude but on 50.00000 Hz. That's great on paper, but
> a real problem with real radio signal data.
>
> --
> 73, Jerry, K0CQ,
> All content copyright Dr. Gerald N. Johnson, electrical engineer
>
> _______________________________________________
> TenTec mailing list
> TenTec@contesting.com
> http://lists.contesting.com/mailman/listinfo/tentec
>
>
_______________________________________________
TenTec mailing list
TenTec@contesting.com
http://lists.contesting.com/mailman/listinfo/tentec
|