Joint Blind Estimation and Equalization Method Based on Deep Learning for Fast Fading Channels

In this paper, we propose a hybrid architecture based on deep learning to perform joint blind channel estimation and equalization on fast fading channels. The architecture uses a combination of Extended Kalman Filters (EKF) and neural network in order to predict the process noise covariance matrix at each symbol time. First, we present the system model with the state-space representation. Next, the proposed architecture, the training method and the dataset used are detailed. Then, the performance between a classical blind EKF equalizer structure and the proposed solution is compared with simulations. We expose the obtain results and show the contribution of deep learning in this context. Our architecture, the Smart Extended Kalman Filter (SEKF) can track fast time-varying channels and equalize at high SNR better than a classical blind EKF equalizer structure in a context of random walk channel impulse response (CIR) evolution.