Why is there such a big difference between the two diagrams on the Rigol HD4000? I think the curve of the Rigol is very smooth in the second diagram.
I agree. Something looks off here.
The "noise" on the noise spectral density is just a function of data length and chosen frequency resolution. The same file can be analyzed to get smoother plot at high frequency or "noisier" plot that goes to lower frequency. You can see that I combined two different settings in the original plot. For comparison to Siglent I used even coarser frequency resolution for Rigol to match the frequency range of Siglent while still using all the 250M of Rigol's data. For Lecroy the file I have is only 100k points, that is why it has even more fluctuations.
The general property of the noise spectral density is that the average level does not depend on the number of points, sampling rate (if no aliasing) or on the frequency resolution. This is not the case with FFT power spectrum since the noise power in each bin depends on the width of the bin. On the other hand when looking at discrete peaks, it is better to use power spectrum.
I have asked the poster earlier , but I did not get the answer.
1. What is the sample rate?
2. Do you apply FFT for all 250Msamples or you use signal stacking tecniques
3. Do you apply any FFT window function -e.g. Hamming, Tukey, Blackman
The sampling rate can be found as twice the maximum frequency for each curve, I did not cut off the edge of the plot.
I mentioned that I used pwelch that breaks data into overlapping segments, and then averages resulting FFTs
The default window in pwelch is Hamming. It does not matter what is used because there are no strong peaks, I also checked rectangular window and Kaiser window. I subtract the mean to eliminate the spectral leakage from DC offset.