Deep learning-based receiver for orthogonal frequency-division multiplexing (OFDM) system

Abstract

In this dissertation, we build a deep learning (DL)-based orthogonal frequency division multiplexing (OFDM) receiver. This receiver replaces the traditional two processes, channel estimation and signal detection with one network deployment process, which can reduce the system complexity. DL method, rather than first calculate the channel state information (CSI) then recover the detected symbols, allows the received signal to be decoded. This approach can be divided into two parts, offline training and online deployment. Firstly, we train a network using data generated on multi-path Rayleigh fading channel. Then, this trained network is used in online deployment to recover the received data symbols directly. In this dissertation, two different kinds of DL network, DNN and LSTM are designed. The simulation is run on MATLAB platform and the simulation results demonstrate that both DNN and LSTM methods can achieve BER performance as good as minimum mean-square error (MMSE) method does when pilot information is sufficient. And in deficient pilot circumstance, DL method demonstrates better robustness. Further more, LSTM can learn channel statics quicker than DNN and can achieve better BER performance than DNN when SNR is low. In conclusion, combining traditional communication system with DL method has shown its advantages and is very promising.