Low-complexity mmWave massive MIMO channel estimation using denoising autoencoder

Abstract

Application of deep learning algorithms for wireless channel estimation has been extensively studied due to the robustness of deep networks to recognize patterns and features as well as mimic any transformation function so long as the appropriate size and structure of the network is configured. These deep learning based estimators are known to achieve high estimation accuracy, and in general, perform better than conventional estimators. For massive MIMO systems, another concern alongside estimation accuracy is pilot length reduction to allow more time for useful data transmission. In this paper, a denoising autoencoder network is used alongside a fully connected network to predict the uplink channel between a user and a base station (BS) in a point-to-point massive MIMO communication system. The goal of the denoising autoencoder scheme is to effectively exploit information redundancy in the channel matrix. By doing so, it is possible to minimize the pilot overhead while maintaining a tolerable estimation accuracy. As a result, the spectral efficiency of the communication system is increased especially at high mobility. Since denoising autoencoders reduce the dimension of the input vector space to a latent representation, fully connected network can be easily incorporated to exploit temporal correlation at low dimension.