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

Convolutional neural network (CNN) based channel estimation approaches have achieved remarkable success in massive multiple-input multiple-output (MIMO) systems. However, the channel estimation performance even needs to be improved. Meanwhile, the loss functions are not well designed in traditional deep learning (DL) based schemes for the problem of channel estimation. To improve the channel estimation accuracy, we propose an improved conditional generative adversarial network (GAN) based channel estimation method in this paper. Specifically, we first design two novel loss functions to better train the proposed conditional GAN model. The training stability of GANs is improved by coupling the mapping from the received signals to the channel matrices with its inverse mapping. Then, by jointly training these two loss functions and the proposed conditional GAN, we develop an improved conditional GAN to estimate channels more accurately. Simulation results demonstrate that the proposed improved conditional GAN based method improves the estimation performance and shows great robustness in MIMO systems.