Comments on: Signal Filtering with Python http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/ A collection of thoughts in technological degradation Wed, 08 Feb 2012 09:12:28 -0700 http://wordpress.org/?v=2.8.6 hourly 1 By: Anonymous http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/comment-page-1/#comment-16801 Anonymous Wed, 07 Sep 2011 16:43:04 +0000 http://www.SWHarden.com/blog/?p=657#comment-16801 Thanks for the code snippet. I've got a few suggestions for you, though: 1) When I use the line "bp=fft[:]", modifying bp still alters fft. I've used "bp = fft.copy()" instead. 2) When zeroing out FFT terms, you need to make the zeros symmetric around the mid-point of the FFT spectrum. The inverse FFT will have a non-negligible imaginary component if the frequency spectrum isn't symmetric, even though the input data is strictly real. 3) To zero out fft coefficients, it's easier to write bp[10:-10] = 0 4) To normalize the data, add the line bp *= real(fft.dot(fft))/real(bp.dot(bp)). fft.dot(fft) gives the total power in the signal, so this redistributes the power in the stop band more or less equally to the pass band. 5) It's perhaps better to change the variable name 'fft' to something else so that it doesn't conflict with numpy.fft.fft. If you're working in iPython, numpy.fft.fft is automagically imported into the global namespace, so trying to use the fft function after running your script could cause a few headaches. Thanks for the code snippet. I’ve got a few suggestions for you, though:
1) When I use the line “bp=fft[:]“, modifying bp still alters fft. I’ve used “bp = fft.copy()” instead.
2) When zeroing out FFT terms, you need to make the zeros symmetric around the mid-point of the FFT spectrum. The inverse FFT will have a non-negligible imaginary component if the frequency spectrum isn’t symmetric, even though the input data is strictly real.
3) To zero out fft coefficients, it’s easier to write bp[10:-10] = 0
4) To normalize the data, add the line bp *= real(fft.dot(fft))/real(bp.dot(bp)). fft.dot(fft) gives the total power in the signal, so this redistributes the power in the stop band more or less equally to the pass band.
5) It’s perhaps better to change the variable name ‘fft’ to something else so that it doesn’t conflict with numpy.fft.fft. If you’re working in iPython, numpy.fft.fft is automagically imported into the global namespace, so trying to use the fft function after running your script could cause a few headaches.

]]> By: Ken http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/comment-page-1/#comment-16373 Ken Thu, 21 Jul 2011 21:17:44 +0000 http://www.SWHarden.com/blog/?p=657#comment-16373 Be careful with this method of "filtering." This kind of direct manipulation of the spectrum results in time-aliasing when you take the inverse FFT. The result is not truly the original signal with its high frequencies filtered out, because the time-aliasing leaves the signal littered with artifacts. The easiest way to do it properly would be to do a lowpass filter in the time-domain. You can do a similar filter in the frequency domain (sort of like you're doing), but the time-domain data and a filter impulse response needs to be properly zero-padded before you FFT them to avoid circular convolution (which results in time-aliasing). Be careful with this method of “filtering.” This kind of direct manipulation of the spectrum results in time-aliasing when you take the inverse FFT. The result is not truly the original signal with its high frequencies filtered out, because the time-aliasing leaves the signal littered with artifacts.

The easiest way to do it properly would be to do a lowpass filter in the time-domain.

You can do a similar filter in the frequency domain (sort of like you’re doing), but the time-domain data and a filter impulse response needs to be properly zero-padded before you FFT them to avoid circular convolution (which results in time-aliasing).

]]> By: Danny http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/comment-page-1/#comment-11535 Danny Mon, 22 Nov 2010 15:13:16 +0000 http://www.SWHarden.com/blog/?p=657#comment-11535 Thanks for the article, sure it'll come in handy cleaning some of my (very noisy) data. Thanks for the article, sure it’ll come in handy cleaning some of my (very noisy) data.

]]> By: Kamil http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/comment-page-1/#comment-10388 Kamil Thu, 19 Aug 2010 22:13:43 +0000 http://www.SWHarden.com/blog/?p=657#comment-10388 Thank you, great article! Thank you, great article!

]]> By: Jphn http://www.SWHarden.com/blog/2009-01-21-signal-filtering-with-python/comment-page-1/#comment-9700 Jphn Mon, 17 May 2010 11:39:32 +0000 http://www.SWHarden.com/blog/?p=657#comment-9700 "This python file requires that test.wav (~700kb) (an actual ECG recording of my heartbeat) be saved in the same folder." => false Thanks for the code! “This python file requires that test.wav (~700kb) (an actual ECG recording of my heartbeat) be saved in the same folder.”
=> false

Thanks for the code!

]]>