RTLSDR Flow Graph
I’ve now created a flow graph making use of Sebastian’s signal separation blocks & GUI, along with his signal extractor, to obtain a specific signal. This complex signal is then fed into the CNN model.
Using Sebastian’s Blocks/GUI together with the classifier
You can see below an FM Signal (BBC Radio 1), which is detected by the signal detection blocks, the classifier is then able to produce outputs, on what it believes the modulation scheme to be.
You can see currently the outputs from the classifier aren’t consistently producing an output of WBFM, for an FM signal (the majority of the outputs claim the signal is GFSK), I am looking into issues as to why this is the case.
I am currently having difficulties getting the CNN to work with real world signals, as well as with detecting synthetic signals, of symbol
rates it wasn’t explicitly trained on.
Creating and saving the CNN as a TensorFlow graph
You can now generate the CNN using TFLean and have the output saved in TensorFlow’s format
The current code for generating the CNN model are in the repository below
Currently working on
Attempting to solve issues with the classifier, for the classification of both synthetic and real world signals.
Also I am just working out how to obtain the dimensions of a dynamic Tensor, so that I can automatically attempt to coerce the input to the Tensor’s format.
- I will look into producing a 2D plot of the FAM output of gr-specest, then a 3D plot of it
- Cyclostationary visualisation, to show cyclostationary features across the spectrum in a 3D manner, making use of the FAM output from gr-specest and 3D graphing tools from QT
- Command line spectrum scanning tool, to enable you to easily scan a large range of the spectrum for user specified modulation schemes
- Last week of SOCIS, code cleanup