- Pytorch (with Numpy, Scipy etc.)
- Matlab (only for .m file)
- Google Colaboratory (current working envoronment)
Passband_Demodulation.ipynb QPSK demodulator in passband. The symbol phase is the regression output. Note that the loss function, circ_mse_loss, is a customized one. The phase_step is used to solve the problem raised when the carrier frequency are not integer multiple of the symbol rate.
sig_gen5_new_timing.m is used for data generation
CNN_for_symbol_timing.ipynb uses a 2D CNN as a classifier to estimate the best timing offset. It works with passband signal, so no complex value is involved. May run into few errors due to some dependency lost. Will be fix later (I hope).
CFO_recovery_training.ipynb trains the CNN with synthesized data. The frequency recovery module is basicly a PLL, but the phase error detector is replaced with the CNN. One advantage of using synthesized data is that in every epoch of training, a set of new traing data can be generated, so the overfitting is not an issue here.
CFO_recovery_for_sc80.ipynb tests the pretrained CNN with sea trial data.
Note that it works in basedband, but only the signal phase components are considered, and the signal amplitudes are ignored. This is only resonable when a perfect timing has been done and no visable ISI in the received signal.
equalization.ipynb is obsolete. The traing signal is synthesized using Numpy only, so no Matlab is required.
Currently a fractionally spaced, decision feedback equalizer has been built for a multipath transmission, however the performance is not as good as expected.
Demodulation in passband with phase compensation.- Integrate CFO estimation/compensation.
- Integrate DFE for multipath