Knowledge-Driven Machine Learning-based Channel Estimation in Massive MIMO System

Accurate channel state information is critical for the massive multiple-input multiple-output(MIMO) system. However, due to the fundamental issue of pilot contamination, most existing works have to use additional channel information or complex machine learning network to design channel estimator. In this paper, we propose a knowledge-driven machine learning (KDML)-based channel estimator, which has a simple network structure and low training overhead. Firstly, we analyze the channel estimation problem under both uncorrelated and correlated channel models and convert it into the well-studied image denoising problem. Then, without any prior information about channel, we take advantage of the traditional channel estimation algorithms to construct the knowledge module in KDML, which maps the input data into object space without any training. After that, we use the denoising convolutional neural network (DnCNN) to build the learning module in KDML, where the residual learning accelerates the training process neural network. Finally, we obtain the estimation of the latent noise from the polluted channel matrix and then calculate the output of the proposed channel estimator. Simulation results demonstrate that the proposed channel estimator outperforms the traditional channel estimator in terms of the normalized mean-squared error (NMSE) without any statistical information about the channels. Besides, the proposed estimator also significantly better than the conventional machine learning-based estimator under the same network depth.