Signal Detection Method at the OFDM Receiver Based on Conditional GAN

Deep learning is an effective approach for signal detection which is a challenging issue in wireless communication systems. Some DNN deep learning methods have shortcomings, like complex structures, many training parameters, and converging difficulty, which may cause information loss and affect accuracy and reliability if the received signals pass through the continuous layers of the neural networks. To overcome those, in our paper, the signal detection part is replaced with a conditional generative adversarial (cGAN) network at the OFDM receiver, and we preprocess the received signals using LS and ZF algorithms based on model-driven as the initial input of the neural network. And the cGAN model introduces pilots as the condition to counter-train the two deep learning networks for generating signals closer to being under real channels. In addition, with an adaptive loss function (GAN Loss), the cGAN model has a certain corrective effect on optimizing the neural network for recovering data. The simulation results illustrate that the cGAN model performs better than the existing DNN models under various signal-to-noise ratios (SNR), especially facing low SNRs and short pilot sequences, that is, it has better robustness for restoring effectively the transmitted signals under real channels.