Products > Test Equipment
New Rigol HDO1000 12-bit DSO - BUGs
ebastler:
--- Quote from: Martin72 on January 02, 2024, 07:36:26 am ---
--- Quote ---4. As found and fixed by TurboTom, Rigol's Flattop window is broken. With Tom's fixed weights, FlatTop is my go-to window function.
--- End quote ---
Afaik flattop window was not the only one or did you check all of them ?
--- End quote ---
I did look at the spectra generated with other window functions, and they look much more plausible. There is, however, a slight asymmetry in the main peak when you zoom in a lot, with all the window functions. That's probably not real, so yes, there might be (lesser) deviations in other window functions too.
As a consequence of setting the sampling parameters via the regular time base and memory depth controls, the FFT always works on record lengths which are not a power of two, apparently. I wonder whether the implementation handles that cleanly?
gf:
--- Quote from: ebastler on January 02, 2024, 08:54:35 am ---
--- Quote from: Martin72 on January 02, 2024, 07:36:26 am ---Afaik flattop window was not the only one or did you check all of them ?
--- End quote ---
I did look at the spectra generated with other window functions, and they look much more plausible. There is, however, a slight asymmetry in the main peak when you zoom in a lot, with all the window functions. That's probably not real, so yes, there might be (lesser) deviations in other window functions too.
--- End quote ---
The left sides of these images definitively do not look correct. Seems to be rather a general problem, not window function specific. Maybe related to the interpolation upon zoom-in?
--- Quote ---As a consequence of setting the sampling parameters via the regular time base and memory depth controls, the FFT always works on record lengths which are not a power of two, apparently. I wonder whether the implementation handles that cleanly?
--- End quote ---
There do exist FFT algorithms for non-power-of-2 sizes. All "1-2-5 sizes" can be decomposed into combinations of radix-2 and radix-5.
ebastler:
--- Quote from: gf on January 02, 2024, 10:08:23 am ---The left sides of these images definitively do not look correct. Seems to be rather a general problem, not window function specific. Maybe related to the interpolation upon zoom-in?
--- End quote ---
Ah, thanks for the link. I remembered that Martin had done FFT measurements with the DHO800, but could not find them yesterday.
I just repeated with the same settings on the DHO1074, and with an input signal at 1 kHz - 1/2 delta_f. As one would expect things look pretty much the same. I had seen that distortion in yesterday's experiments; it's the "slight asymmetry" of the peak shape I referred to above. Certainly an artefact, but not as problematic as the massive peak broadening and ripple produced by the incorrect flat-top window.
EDIT: I don't think the distorted peak shapes are due to interpolation artefacts. For the FFT, simple linear interpolation between data points is used; while there is no "dot" mode, the individual straight line segments are visible when zoomed in. The asymmetry is on a scale of several delta_f steps.
Maybe, if TurboTom finds himself with too much time on his hands, he could look into those other window functions too? :)
--- Quote ---There do exist FFT algorithms for non-power-of-2 sizes. All "1-2-5 sizes" can be decomposed into combinations of radix-2 and radix-5.
--- End quote ---
Sure, I know they exist. My question was whether Rigol got them right. ;)
That might be another reason for FFT artefacts, besides deviations in the window functions.
EDIT 2: Looking at Martins DHO800 screenshots again, they do show more reasonable "RBW" values. The displayed "RBW" is actually the frequency step in Hz in all those screenshots. On the DHO1074, I get displays of "9.999m" or such. And as mentioned, if I switch the time base up and down again, the RBW often does not come back to where it was before, so there is no clear correlation between delta_f and RBW at all. Looks like they at least fixed that in the DHO800, so there is hope for the larger models... If they furthermore added the "Hz" unit and called it "Step" instead of RBW, I'd be happy. ::)
gf:
--- Quote from: ebastler on January 02, 2024, 10:49:28 am ---I don't think the distorted peak shapes are due to interpolation artefacts. For the FFT, simple linear interpolation between data points is used; while there is no "dot" mode, the individual straight line segments are visible when zoomed in. The asymmetry is on a scale of several delta_f steps.
--- End quote ---
The Siglent screenshots I've seen did indeed interpolate linearly between frequency points. But do you also see straight line segments on the DHO?
With linear interpolation between frequency points, you would not see the side lobes of Rectangular, Hann, Blackman, Flattop (and some others) windows, because the frequency points outside the main lobe happen to coincide with zeros of the frequency response of these window functions - and linear interpolation would give zero anywhere in between, too. But in various posted DHO screenshots, I do see side lobes. This leads me to believe that the DHOs do not interpolate linearly.
EDIT: I have to correct: We still can see contributions of side lobes, of course, if the signal frequency is not an integer multiple of delta_f.
ebastler:
--- Quote from: gf on January 02, 2024, 11:49:03 am ---The Siglent screenshots I've seen did indeed interpolate linearly between frequency points. But do you also see straight line segments on the DHO?
[...]
We still can see contributions of side lobes, of course, if the signal frequency is not an integer multiple of delta_f.
--- End quote ---
Yes, the interpolation clearly shows linear segments -- see the attached screenshot.
In the attached, I set the signal frequency to 1 kHz, a full multiple of delta_f, and there still is an asymmetry. Also, strangely there never seems to be an FFT data point right at 1 kHz; they are always a bit displaced to the left and right, and asymmetrically from the 1 kHz mark. (Assuming that the kinks in the curve denote the data points.) Is that a consequence of the non-power-of-two FFT? EDIT: I don't see how that would be the case though, with an even 10 MHz sampling rate.
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version