I want to see if I can make my Software Defined Radio, SDR, to classify unknown radio signals with the help of an artificial neural network. That is, my SDR outputs a sequence of complex numbers (IQ-data), which I want to use to determine if the receieved signal is, for instance, FM or AM modulated. This approach was used in a paper (https://arxiv.org/pdf/1712.04578.pdf) and they created and used a freely downloadable dataset (https://github.com/sofwerx/deepsig_datasets/blob/master/README.md).
Being new to both SDR and Deep Learning I have now tried for a couple of months to create an LSTM network, train it on the dataset and then use it for classification, but have sadly failed. I have concluded that it is likely due to the fact that I do not seem to understand how the dataset is structured. I have not been able to find any documentation concerning this dataset (other than a text file with a list of the modulation forms used in the dataset) though. My hope is that someone on this forum has some prior experience to share about how to use it.
The dataset (DEEPSIG DATASET: RADIOML 2016.10A) is split in three matrixes, X: 2x1024x2555904 cells, Y: 24x2555904 cells and Z: 1x2555904 cells. My belief has up until now been that "X" contains the complex time series, i.e. one row for the real component and one row for the imaginary component both in a sequence of 1024 samples. The "Y" to contain the corresponding 24 classes, whereas the "Z" somehow contain the signal to noise level for each signal.
After training and deploying the network in Matlab I found out that it does not even manage to classify the local FM radio station. I then looked at the data I had used for training and plotted one of (what I thought to be) the FM sample sequences from the dataset in the complex plane, but I did not get the "circle" I had expected from a frequency or phase modulated signal. I am thus totally lost, the dataset has been used for a high class scientific paper so the problem lies with me, and my lack of understanding. I apologize for the long and still very unconcise question, but I don't want to infer any of my own misconceptions into the query. Thanks for any help!