Improved Conditional GAN-Based Channel Estimation for MIMO-OFDM Systems

For massive multiple-input multiple-output (MIMO) systems, convolutional neural network (CNN)-based channel estimation methods have achieved remarkable success. However, the channel estimation accuracy still needs to be improved, and the loss functions for channel estimation are not well-designed in conventional deep learning (DL)-based approaches. In this paper, we propose an improved conditional generative adversarial network (cGAN)-based scheme to improve both the training stability and the channel estimation performance of cGANs. Specifically, two CNNs in the proposed cGAN are trained adversarially to generate an adversarial loss that compensates for the information loss during training. To estimate channels more accurately, we replace the received signals with the least-squares (LS) channel estimates and feed them into the improved cGAN as conditional inputs. Moreover, to improve the performance and the training stability of cGANs, we propose a self-preserving loss by coupling the mapping from the LS channel estimates to the true channels with its inverse mapping. We also introduce a cycle-consistency loss to further optimize the improved cGAN by exploiting the implicit data augmentation effect. These two loss functions and the proposed cGAN are jointly trained to improve both the training stability and the performance of cGANs. Simulation results show that the proposed improved cGAN enhances the channel estimation performance and achieves strong robustness.