Deep learning based encoding and decoding schemes for performance enhancement of mobile communication systems

Abstract

In this thesis, we try to improve the performance of mobile communication systems by designing encoder and decoder with deep neural networks. Denoising deep autoencoder which is a unsupervised generative model in deep learning is applied to wireless communication systems. First, we focus on orthogonal frequency division multiplexing (OFDM) system where single symbol is allocated to a single resource. Then, we focus on sparse code multiple access (SCMA) system where multiple symbols are allocated to a single resource. High peak-to-average power ratio (PAPR) has been one of the major drawback of OFDM system. We propose a novel PAPR reduction scheme, PAPR reducing network (PRNet), which is based on the deep autoencoder architecture. In the PRNet, the constellation mapping and demapping of symbols on each subcarrier is determined adaptively through deep learning technique such that both bit error rate (BER) and PAPR of OFDM system can be jointly minimized. Through simulations, we show that the proposed scheme outperforms the conventional schemes in terms of BER and PAPR. We showed that the probability that PAPR is higher than 3 dB is less than 0.1 % and at the same time the BER is improved compared to conventional schemes. Secondly, SCMA is one of the promising multiple access schemes that can satisfy the criteria of 5G wireless communication system in terms of spectral efficiency and massive connection. Designing the SCMA codebook is a difficult work because of non-orthogonality and multi-dimension traits of SCMA system. We proposed a novel codebook generation method for SCMA system in the perspective of bit error rate (BER) based on deep autoencoder structure. The codebook generator is built with multiple deep neural networks (DNNs) and trained to generate BER minimizing codebook. Also, a DNN based decoder is proposed which is shown to require much less computation time than conventional MPA detector. The proposed codebook is shown to outperform conventional SCMA codebook upto 3 dB SNR. Lastly, we proposed a deep spread multiplexing (DSM) which is the generalization of SCMA. DSM is able to deal with any number of signal streams and resources. Also, effective methods to train the DNN based encoder and decoder are studied. Through simulations, we showed that the proposed DSM outperforms conventional SCMA upto 2 dB SNR.