Zoom FFT

Hi all!
I have fd=65KHz. I do fft with window Hann 4096 points. I need  detect
frequency
12kHz...20kHz with resolution 2 Hz.  What must I do? I understand that
i must mix  input signal with 12 kHz sine. How?

Piter Wolf wrote:
> Hi all!
> I have fd=65KHz. I do fft with window Hann 4096 points. I need  detect
> frequency
> 12kHz...20kHz with resolution 2 Hz.  What must I do? I understand that
> i must mix  input signal with 12 kHz sine. How?

You need to set up a different FFT length. If you by "fd" means
"sampling frequency", you will need an FFT length of at least
32500 points. You basically need a frequency bin width that is
smaller than the desired frequency tolerance.

Rune

Yes, i know, but  I hav't set up diffferent FFT length, i must do it on
4096 points.
 I know algorithm: bandpass fiter->downshift frequency
->filter->decemation->fft. I don't know how do downshift :((

Sorry, previous post is my, my sister log in and i don't log off

"Rune Allnor" <allnor@tele.ntnu.no> wrote in 
news:1121712804.369670.159280@z14g2000cwz.googlegroups.com:

> 
> 
> Piter Wolf wrote:
>> Hi all!
>> I have fd=65KHz. I do fft with window Hann 4096 points. I need  detect
>> frequency
>> 12kHz...20kHz with resolution 2 Hz.  What must I do? I understand that
>> i must mix  input signal with 12 kHz sine. How?
> 
> You need to set up a different FFT length. If you by "fd" means
> "sampling frequency", you will need an FFT length of at least
> 32500 points. You basically need a frequency bin width that is
> smaller than the desired frequency tolerance.
> 
> Rune
> 

No he doesn't need 32500 points since he is interested in 12k to 20k = 8k 
range.

A 4096 point FFT will give him 2 Hz bins

Here is a block diagram of zoom fft

http://www.numerix-dsp.com/zoomfft.html

Basically you want to multiple by a 16 k cosine and sine to translate the 
signal to DC.





-- 
Al Clark
Danville Signal Processing, Inc.
--------------------------------------------------------------------
Purveyors of Fine DSP Hardware and other Cool Stuff
Available at http://www.danvillesignal.com

If your FFT is fixed at 4096 points you could also try to
estimate the true peak position using parabolic interpolation
around the peak in the calculated spectrum. I have had
acceptable results using this approach in the past.

Hope this helps,

Dave.

Ok, I multiplie witn sine and cosine, what filter i need? 8 k or 4 k?
and I shall have spectrum on 4k, but no 8k. I don't understand:((

"Piter Wolf" <kuzmina@smtp.ru> wrote in message
news:1121760742.785368.45030@f14g2000cwb.googlegroups.com...
> Ok, I multiplie witn sine and cosine, what filter i need? 8 k or 4 k?
> and I shall have spectrum on 4k, but no 8k. I don't understand:((

You need two 4kHz LPF's, one on the cosine'd and one on sine'd signal.

You downconvert so zero hz is in the middle of the passband (ie, mix with
16kHz so 16kHz -> 0Hz).

If you only take one of the two mixed quadrature signals, you cannot tell if
the signal is in the 4kHz above or below 16kHz.

By using both the quadrature signals, you can tell if it's a negative or
positive frequency relative to the 16kHz and the FFT will correctly show
this, albeit shifted round by 2048 bins.

Cheers, Howard

my C code spectrum :
 win_out_r- mixed signals 1 after window -4096 points
win_out_i -mixed signals 1 after window - points
r_output - real after fft 4096 points
i_output -imag after fft 4096 points
.
.
.

cfft4096( win_out_r,win_out_i ,r_output, i_output );
 for( i=0 ; i<n_samp; i++ )
                    spectr[i] = (r_output[i] * r_output[i] + i_output
[i] *i_output [i])*2/n_samp;
 .
 .
 .
Standart procedure, i have symmetrical spectrum. Real spectrum have
2048 points
how using         quadrature signals ?

Hi Piter


>                     spectr[i] = (r_output[i] * r_output[i] + i_output
> [i] *i_output [i])*2/n_samp;

This won't work for what you're trying to do! For the power spectrum of a
quadrature signal you need to do this:

    spectr[i] = ((r_output[i] + i_output [i])/n_samp)^2

I know it may look wrong, but you need to look at
http://www.dspguru.com/info/tutor/QuadSignals.pdf. This document fairly hard
work if your math is rusty (like my 20+ year old degree level math) but the
description is as good as I've seen.

> Standart procedure, i have symmetrical spectrum. Real spectrum have
> 2048 points
> how using         quadrature signals ?

You will get a 4096 point non-symetrical spectrum with my correction above.

Cheers, Howard