Recurrent Neural Network-Based Single-Input/Multi-Output Demodulator for Cochannel Signals

In this letter, a data-driven single-input/multi-output (SIMO) demodulator is proposed, to demodulate concurrently cochannel signals. The SIMO demodulation is formed as a signal-wise sequence labelling problem, which we propose to solve by multi-layer recurrent neural networks (RNN). Numerical results validated the low bit-error rates of the proposed demodulator against cochannel signals of varied power ratios. Comparing with existing model-based and data-driven SIMO demodulators, the proposed demodulator provided superior BER performance, especially when cochannel signals were of distinct powers. More importantly, the proposed demodulator avoids one major deficiency of existing model-based schemes, which require the symbol rates of cochannel signals to be strictly identical. Meanwhile, the offline-trained demodulator generalized well in varied open set tests.