Deep Learning Based Channel Estimation with Flexible Delay and Doppler Networks for 5G NR

Abstract

In this paper, we propose a deep learning based algorithm for downlink channel estimation for 5G new radio. The channel estimation block in the downlink plays an important role in mobile device performance. Here, we explore the usage of convolutional neural network (CNN) and compare its performance with conventional and other known CNN based channel estimation methods. The novelty of this method is that, it separates the channel estimation method into three logical blocks. First, obtaining channel coefficients from pilots and filtering for noise reduction. Second, interpolating the channel in frequency domain using the noise filtered channel estimate on pilots. Third, interpolating the channel in time domain across orthogonal frequency-division multiplexing (OFDM) symbols. This architectures allows considerable flexibility in handling various combinations of delay spread (DS) and Doppler spread (DoS) that are possible in practical scenarios. This in turn helps in reducing the complexity and storage requirements. The proposed channel estimation method showed a performance gain of 4 dB in NMSE compared to linear minimum mean square error for clustered delay line channel model at signal to noise ratio of 30 dB, and is robust to mismatches in parameters like DS and DoS because of estimation errors.