An Autoencoder Based Channel Estimate Denoising Network (CEDNet) for OFDM Systems

In this paper, we propose a novel data-driven low-complex denoising autoencoder (DAE) based CEDNet for channel estimation in OFDM systems. The proposed deep learning (DL) model estimates channel state information (CSI) response at unknown resource elements (REs) with the help of few known pilots. The CEDNet model resides between channel estimator and equalizer block, where it refines estimation accuracy of conventional estimator and interpolates refined channel response over the whole resource grid. Through rigorous simulations, we show that addition of novel CEDNet module helps to achieve considerable improvement in bit-error rate (BER) performance compared to widely used least squared (LS) method. Also, it attains signal-to-noise ratio (SNR) improvement of 3dB or more from LS under varied user speed and 3GPP compliant channel models. Furthermore, its BER performance is equivalent to the existing well-known DL based approaches, at 80% reduction in number of trainable parameters and 90% reduced computational complexity. This indicates CEDNet is a lightweight DL model that can be practically realized with lesser training data compared to the existing architectures.