Deep Learning-Based Data-Assisted Channel Estimation and Detection

We introduce a novel structure empowered by deep learning models, accompanied by a thorough training methodology, for enhancing channel estimation and data detection in multiple input multiple output (MIMO) orthogonal frequency division multiplexing (OFDM) systems. Central to our approach is the incorporation of a Denoising Block, which comprises three meticulously designed deep neural networks (DNNs) tasked with accurately extracting noiseless embeddings from the received signal. Alongside, we develop the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Correctness Classifier</i>, a classification algorithm adept at distinguishing correctly detected data by leveraging the denoised received signal. By selectively utilizing these identified data symbols as additional pilot signals, we augment the available pilot signals for channel estimation. Our <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Denoising Block</i> also enables direct data detection, rendering the system well-suited for low-latency applications. To enable model training, we propose a hybrid likelihood objective of the detected symbols. We analytically derive the gradients with respect to the hybrid likelihood, enabling us to successfully complete the training phase. When compared to other conventional methods, experiments and simulations show that the proposed data-aided channel estimator significantly lowers the mean-squared-error (MSE) of the estimation and thus improves data detection performance. Github repository link is https://github.com/hamidreza-hashempour/5g-dataaided-channel-estimate.