I'm evaluating a cyclostationary approach to signal classification. My first step in doing so was generating signals and embedding them into a few specified parameter sets of the Watterson channel model. This seemed to be a good fit at the time. A colleague of mine proposed a new idea to me:
Embed/Add true positive digital signals generated by myself into broadband files recorded from an antenna?! Since I saw reason in his notion, I did as told. Using standard FFT-based signal processing I embedded digitally modulated signals into the broadband data files.
I subsequently found my true positive rate to be significantly less favorable compared to the Watterson channels tests. So his proposal, to me seemed to gap the difference from simulation to experimental study on a real HF channel. Since I now want to present my results to a scientific audience, I started musing over the influence of using real antenna data, with artificially embedded signals, as a test set:
The true positive signals on the one hand, acquired additive noise by embedding them into the real scenarios. But on the other hand, by not transmitting the signals, there is no significant influence resulting from fading, multipath or spread distortion to my scenarios.
The cyclostationary approach predominantely requires a conservation of the structure of the time signal. Therefore I assume, the relative strength of the detection didn't change through the additive noise? The much more lively real data HF channel just introduces a lot of distorted similar signals which might be mistaken to be true positive signals by my algorithm?
How would you think does my choice of using real data with added true positives compares to real scenarios with truely transmitted true positives?