SY said:
You think so? Then it might happen also in FFTs of real
measurements. Still, the behaviour depends on the ratio
between the signal period and the simulation step length,
not the number of samples, although the latter of course
also has an effect.
At least this shows that understanding the mathematical
proof of something does not necessarily imply understanding
the concept being proved. Or perhaps it shows that I didn't
understand the proof after all?
🙂
It does indeed happen in real measurements. It's something that people doing FT spectroscopy in general are painfully aware of; that's one of the reasons that windowing is used so heavily, despite the consequences on resolving line shapes.
Hi,
An FFT only works for harmonic signals. To get it working the number of periods MUST EXACTLY fit in the FFT window. Thus if you have a sine signal and a 32k FFT, the number of sine waves must exactly fit in the 32k samples to get it working.
To circumvent this there is indeed a trick called windowing. A well-known window is a cosine function or a sin^2 function applied over the FFT length. This has the side effect of generating side bands in the frequency plot which are not present in the original signal. Many people have investigated various types of window functions to minimise these side band effects. One of the best are those of Blackman-Harris and Rife-Vince. But be aware that those windows “widens” the measurement bandwidth app. 2 to 4 times.
An FFT only works for harmonic signals. To get it working the number of periods MUST EXACTLY fit in the FFT window. Thus if you have a sine signal and a 32k FFT, the number of sine waves must exactly fit in the 32k samples to get it working.
To circumvent this there is indeed a trick called windowing. A well-known window is a cosine function or a sin^2 function applied over the FFT length. This has the side effect of generating side bands in the frequency plot which are not present in the original signal. Many people have investigated various types of window functions to minimise these side band effects. One of the best are those of Blackman-Harris and Rife-Vince. But be aware that those windows “widens” the measurement bandwidth app. 2 to 4 times.
The problem when using a spice simulator without is as mentioned by PJOTR that there is no window function... Without this, you will add a DC component to the simulation. and then you get a result like the 1001Hz you have shown.
Sonny
Sonny
Sometime ago I used a program like this with my soundcard - a Soundmaker 32x. I was testing the output of my preamp when suddenly the card stopped functionning. I attributed this to a too high voltage on the input. I looked at it but could not repair it.
So I believe that a system for protecting the soundcard is a must unless you are willing to buy a new one once in a while. But then you could buy a scope for the same money, with way better accuracy...
Miguel
So I believe that a system for protecting the soundcard is a must unless you are willing to buy a new one once in a while. But then you could buy a scope for the same money, with way better accuracy...
Miguel
True Miguel,
An oscilloscope is an indispensable tool for the serious DIY’s involved in electronics as will do a good signal generator on the workbench. But such RTA’s are very useful too. Adjusting the bias of a class-AB amp is really a snap with a RTA + sine generator. Just look at the harmonics. This is way more accurate and faster than doing it with an oscilloscope. If you are lucky and you own a DSO you are luckier if you have a FFT function build in. But even then you are limited to the 8bit resolution of the scope. Not sufficient for serious work on amplifiers.
For building your own amps such harmonic patterns will give you useful information about if you make progress or not with your design.
For the speaker builder? Well …. err, I can’t live without it anymore. Measuring filter responses and the combined acoustical response is of great value for checking what you are doing.
But I agree most soundcards will die if you feed them with more than 5V_pp. So a breakout box with protection is mandatory.
An oscilloscope is an indispensable tool for the serious DIY’s involved in electronics as will do a good signal generator on the workbench. But such RTA’s are very useful too. Adjusting the bias of a class-AB amp is really a snap with a RTA + sine generator. Just look at the harmonics. This is way more accurate and faster than doing it with an oscilloscope. If you are lucky and you own a DSO you are luckier if you have a FFT function build in. But even then you are limited to the 8bit resolution of the scope. Not sufficient for serious work on amplifiers.
For building your own amps such harmonic patterns will give you useful information about if you make progress or not with your design.
For the speaker builder? Well …. err, I can’t live without it anymore. Measuring filter responses and the combined acoustical response is of great value for checking what you are doing.
But I agree most soundcards will die if you feed them with more than 5V_pp. So a breakout box with protection is mandatory.
Sonny and Pjotr,
thanks, I think I have a better practical understanding of FFT
now. I realized that I was actually wrong in what I said about
my experiments. I didn't actually adjust the simulation length
when switching from 1 kHz to 1001 Hz, but kept it at 20ms, which
of course is not 20 cycles in the latter case. I made some new
experiments now, using the more correct, although not exact,
value 19.98 ms and the results look much better. The artifacts
are still there, but the sloping of the spectrum almost goes
away. Adding windowing seems to give a nice floor again. This
probably also explains why I didn't notice much difference
with any windowing function previously, that is, I have mostly
used simulation length and signal that match up to an integer
number of cycles.
Then, the conclusion would be that windowing would be useful
in a real-time analyzer, like the one I have tested.
thanks, I think I have a better practical understanding of FFT
now. I realized that I was actually wrong in what I said about
my experiments. I didn't actually adjust the simulation length
when switching from 1 kHz to 1001 Hz, but kept it at 20ms, which
of course is not 20 cycles in the latter case. I made some new
experiments now, using the more correct, although not exact,
value 19.98 ms and the results look much better. The artifacts
are still there, but the sloping of the spectrum almost goes
away. Adding windowing seems to give a nice floor again. This
probably also explains why I didn't notice much difference
with any windowing function previously, that is, I have mostly
used simulation length and signal that match up to an integer
number of cycles.
Then, the conclusion would be that windowing would be useful
in a real-time analyzer, like the one I have tested.
- Status
- Not open for further replies.