Deep Learning-Based Joint Pilot Design and Channel Estimation for OFDM Systems

Abstract

We propose a non-uniform joint pilot design and channel estimation (JPDCE) scheme in closed-loop orthogonal frequency-division multiplexing (OFDM) systems. Specifically, the encoder of the proposed JPDCE scheme is composed of two sub-networks which are used for pilot location assignment and pilot power allocation, respectively. To ensure accurate learning of the pilots, the OFDM block consisting of intelligently customized layers is designed to constrain the output of the encoder. Furthermore, the decoder is used to learn the channel state information (CSI) based on the output of OFDM layers by minimizing the mean square error (MSE) of the channel estimation. The numerical results indicate that the JPDCE scheme considerably outperforms the traditional methods as well as two state-of-the-art deep learning (DL)-based ones, demonstrating its excellent ability to learn the statistical characteristics of the wireless channel. In addition, we demonstrate that the JPDCE scheme shows excellent robustness against various channel distortion and interference: when (i) there are limited pilots, (ii) the cyclic prefix (CP) is removed, and/or (iii) clipping noise is introduced.