Training a CNN on just the raw I/Q values with only 128 timestamps might be the reason we are only getting ~50% accuracy. Decomposing signals into frequency bins and training on that might improve performance. Example pipeline:
I/Q -> window (Hann) -> FFT -> power -> log/normalize -> 1D CNN
Training a CNN on just the raw I/Q values with only 128 timestamps might be the reason we are only getting ~50% accuracy. Decomposing signals into frequency bins and training on that might improve performance. Example pipeline:
I/Q -> window (Hann) -> FFT -> power -> log/normalize -> 1D CNN