A Deep Neural Network with Residual Skip Connections for Channel Estimation

Abstract

When using orthogonal frequency-division multiplexing (OFDM) for communication in channels with high resistance to inter-symbol interference (ISI), conventional channel estimation methods and neural networks with few layers perform poorly. To address this issue, the residual channel estimation network (ResCENet) is a deep neural network (DNN) containing residual skip connections that is suggested for channel estimation. ResCENet differs from previously investigated simple DNNs in that it is deeper and capable of detecting the link in sequential OFDM symbols, as well as performing estimation and demodulation tasks end-to-end. Specifically, ResCENet consists of convolutional neural networks (CNNs), bidirectional recurrent neural networks (Bi-RNNs), fully connected neural networks (FCNNs). To prevent the degradation problem caused by too many layers, residual skip connections were added, which can increase the number of layers in DNNs. Some regularization measures were also added. ResCENet receives a complex number-based digital signal and can directly restore the transmitted digital symbols. The simulation results show that ResCENet can achieve excellent channel estimation with only 4 out of 64 symbols as pilots, outperforming the conventional methods, as well as simple FCNNs.