Let’s look at a few more signals posted to sigidwiki.com. Just for fun.
There isn’t much of a lesson here. Perhaps just that the cyclostationarity of real-world signals is often (1) strong and (2) complicated. The complexity is relative to that for typical signal models used in the engineering (My Papers [25,26,28,30,44] and references therein) and machine-learning literature–there we see a strong focus on textbook QAM and PSK.
(Is ‘machine-learning-based modulation recognition’ engineering? So far I’ve seen a lot of admissions of hyper-parameter selection by trial-and-error, and a superhuman ability to ignore any details about the inputs to the machines, so it looks to me like it lands in the category of tinkering. Not that tinkering doesn’t have its place, it’s just not quite engineering. Right?)
I suppose another lesson or conclusion here is that the cyclostationarity property isn’t all that special. It isn’t particularly rare and, like power spectra, comes in a lot of shapes and sizes. In fact, you have to work hard to design a communication signal that is simultaneously stationary and demodulation-friendly.
So I grabbed some signals from sigidwiki.com and subjected them to blind cyclostationary signal processing, like I did for the DMR signal. These signals did not have interesting visual temporal structure, unlike the DMR signal, so instead of plotting the time-domain samples along with the spectral correlation surfaces, I plotted the cyclic domain profile. Also, I neglected to plot the conjugate spectral correlation surface. (I have to leave something for you to do.)
You can find the data files here. Also be sure to download read_binary.m from that same Downloads page for reading the files into MATLAB. Or use it as a model to write your own simple data converter so you can work in C or python or whatever you like.