Prev: designing a multiband filter
Next: Floating-point to Fixed-point in (Sharc) DSP
From: kevin on 4 Jun 2010 00:43
On 2 juin, 18:46, "FatScouser" <john.hague(a)n_o_s_p_a_m.truebit.co.uk>
wrote:

>Could someone recommend the best spectrum analysis software package - or
>method - for dealing with signals having only partial waves? I've been
>using SigView's FFT function on such signals, but the results are
>unexpected. I've tried applying filters/windows, varying the sample rate,
>and other things, but to no avail...

I don't know exactly what you're trying to do, but I presume it has to
do with estimating a low frequency signal of very limited extent in
the time domain (a few cycles or less than a cycle).

Is the signal sinusoidal? Is it stationary? Or perhaps you have to
estimate those kinds of things, too?

Ideally, for a stationary sinusoid, you'd want to improve your
resolution in the frequency domain by using the minimum sample rate
and maximizing the number of samples (bin spacing in the frequency
domain = sample rate/N, and this is also your effective 'bin width').
The inverse of the bin spacing is your sampling interval – the length
of time over which you collect your samples.

Without a long enough sampling interval, you could have a problem
trying to estimate things. There are ways of interpolating between
FFT outputs to estimate the parameters of a sinusoid, and they usually
require that you meet certain conditions. There are other techniques
that can yield excellent results but they require that you apply
windowing to the data, and it's tough to do if you have too few cycles
of the waveform in your sampling interval.

That's because windows tend to 'squish' the data at the beginning and
end, and you can get some bad results if you don't have a least a few
cycles of the waveform in the sampling interval (and I believe that
'squish' is the proper technical term, except perhaps in Canada, where
'squoosh, eh?' is acceptable, but only between 4 and 7 PM …. on
Thursdays).

Perhaps if your signal is indeed a sinusoid, and you have little
noise, you can square it (r[n] = r[n]*r[n] ) and subtract out the DC
value. Your 'difference' frequency will be 0, and your 'sum'
frequency will be double. For instance, if you started with one
second of data from a 1.25 Hz input, you'd generate a 2.5 Hz result.
You might even iterate on that to get enough cycles to then use a
windowing technique for estimation, but if your signal is not really a
sinusoid, or if you have too much noise, this approach isn't very
good.

Perhaps you could try zero crossing or other methods? I don't really
know if they would be applicable in your case.

Do you actually have specific requirements? (sinusoidal signal of a
certain frequency or frequency range, stable over so many seconds,
must estimate frequency within x%, must do so within y seconds after
data capture, etc.).

The more you can bound and constrain and explain your problem, the
easier it may be for others to suggest solutions.

Kevin McGee
From: FatScouser on 4 Jun 2010 05:44
>On Jun 2, 5:20=A0pm, Jerry Avins <j...(a)ieee.org> wrote:
>> On 6/2/2010 12:46 PM, FatScouser wrote:
>>
>> > Hi,
>>
>> > Could someone recommend the best spectrum analysis software package -
o=
>r
>> > method - for dealing with signals having only partial waves?
>>
>> =A0 =A0...
>>
>> What is a partial wave?
>>
>> Jerry
>> --
>> Engineering is the art of making what you want from things you can get.
>
>My 1st thought is "partial wave expansion" in physics. This has to do
>with the scattering of a plane wave by a particle where the resulting
>scattered wave is expressed as a sum of spherical waves each with its
>own amplitude and phase shift. But I'm quite sure the OP didn't want
>this version. LOL.
>
>Clay
>
>

Sounds like a lot of uncertainty to me, Heisenberg flavour. Does my signal
have partial waves, or maybe partial particles, or both? We are all partial
waves/particles after all.

<ahem>


I'm looking at harminv - thanks a lot for the recommendation above.

http://ab-initio.mit.edu/wiki/index.php/Harminv

Our signal comprises finite waves that damp away quickly. I didn't
appreciate that when I tried using SigView and FFT. 'Harmonic inversion'
seems to tick the boxes, so am looking at getting harminv to work now...
From: illywhacker on 4 Jun 2010 05:57
On Jun 4, 11:44 am, "FatScouser"
<john.hague(a)n_o_s_p_a_m.truebit.co.uk> wrote:

> Our signal comprises finite waves that damp away quickly.

You need a model!

illywhacker;
From: FatScouser on 4 Jun 2010 07:32
Hi again,

Many thanks for all the posts so far - especially the amusing ones.

Very sorry for not giving any information on our particular application: it
was stupid to expect good answers in response to vague questions.

We have some truck suspension, ride-height, and other chassis data (0-20
Hz) from which we need to extract frequency spectrum information. That's
the problem.

Rather than using FFT, which doesn't seem to handle the very few numbers of
wavelengths due to all the damping going on, I'm looking at the harminv
module developed by Prof. Steven G. Johnson (who posts on this forum). That
module assumes the signal is composed of a finite number of decaying
sinusoids - which sounds like it should fit our purposes, but am not 100%
sure, so any confidence from you experts would be great!!!

Thanks a lot, John
From: illywhacker on 4 Jun 2010 08:08
On Jun 4, 1:32 pm, "FatScouser" <john.hague(a)n_o_s_p_a_m.truebit.co.uk>
wrote:

> We have some truck suspension, ride-height, and other chassis data (0-20
> Hz) from which we need to extract frequency spectrum information. That's
> the problem.
>
> Rather than using FFT, which doesn't seem to handle the very few numbers of
> wavelengths due to all the damping going on, I'm looking at the harminv
> module developed by Prof. Steven G. Johnson (who posts on this forum). That
> module assumes the signal is composed of a finite number of decaying
> sinusoids - which sounds like it should fit our purposes, but am not 100%
> sure, so any confidence from you experts would be great!!!

Dear John,

With all due respect, what you want frequency spectrum information
*for*? Surely it is not of interest in itself.

I do not know the harminv package, but if you are basically dealing
with superimposed damped harmonic oscillators, then perhaps it is the
thing. There must be many models of truck behaviour out there: why not
start with one of those?

illywhacker;
First  |  Prev  |  Next  |  Last
Pages: 1 2 3 4
Prev: designing a multiband filter
Next: Floating-point to Fixed-point in (Sharc) DSP