Deep learning-based channel estimation for the OFDM system

Abstract

This dissertation introduces a joint implementation of channel estimation and signal detection functions in Orthogonal Frequency Division Multiplexing (OFDM) systems using Deep Learning (DL) methods. Different from the traditional modular communication system, this method uses an end-to-end network instead of the original complex channel estimation and signal detection module. The network can implicitly estimate the channel state information and recover the received signal to original binary data directly, which simplifies the structure of the receiver. The experimental results show that the channel estimation method based on DL has stronger adaptability to the extreme situations when the number of pilots is insufficient as well as the wireless channels are complicated by serious distortion and interference. Even under ideal conditions, the DL method also has the performance not inferior to minimum mean square error channel estimation, which is very close to the ideal bit error rate curve. This result fully proves the superiority of deep learning methods in the field of communication. In addition, this dissertation also uses a weight pruning method to compress the trained model. This method can increase the sparsity of the model while keeping the accuracy of the model unchanged, thereby reducing the storage capacity of the model. Index Terms: OFDM, channel estimation, DL, end-to-end network, weight pruning