Deep Learning-based End-to-End Physical Layer Design of Wireless Communications

Abstract

The end-to-end physical layer design of wireless communications based on deep learning algorithms is considered a promising approach since it aims to jointly optimize the transmitter and receiver to adapt to the channel conditions. This end-to-end design is based on the concept of the autoencoder which is a type of neural network (NN) that learns to reconstruct its input at the output. For practical scenarios, the over-the-air (OTA) channel and hardware effects such as carrier frequency offset (CFO) cannot be modeled accurately by mathematical expressions in the training process. This challenge could be overcome by deploying a two-phase training process. Moreover, timing synchronization is essential for continuous transmission schemes in practical systems. In drone communications, the system encounters time-varying Doppler shift, for which different scenarios are considered depending on the variation rate. Training a single model to deal with different scenarios efficiently can be a challenging problem in this research area.

In this thesis, first, the implementation and finding the optimum hyper-parameter setting for the basic autoencoder model are investigated. Then, the two-phase training approach is validated using a simulated channel in GNU Radio. Furthermore, the training process of the offset estimator for timing synchronization in continuous transmission schemes is elucidated. Next, the implementation and hyper-parameter setting of the autoencoder model in the orthogonal frequency division multiplexing (OFDM) scheme with the cyclic prefix (CP) are investigated. At last, the OFDM-Autoencoder is studied for drone communication links encountering time-varying CFO. The effect of the time-varying CFO on the received samples is modeled in different scenarios depending on the variation rate. A training approach is proposed by which the OFDM-Autoencoder can handle time-varying CFO efficiently in all Doppler shift scenarios using the same structure.

The study of hyper-parameter setting in the basic autoencoder model indicates that improper batch size degrades the performance close to 1 dB for a BLER of 1e−4. In the investigation of the two-phase training process, it has been shown that inaccurate modeling of the real channel in phase I has a significant impact on the performance of the initially trained model which can efficiently improve after the fine-tuning process. While the fine-tuning gain is negligible for low Eb/N0 range, in high Eb/N0 range, it achieves almost 1 dB gain for a BLER of 1e−3. The OFDM-Autoencoder can compensate perfectly for the time invariant CFO by employing CFO compensation NN in the time domain; such that it achieves BLER performance identical to the no-CFO scenario for most of the SNR range. Furthermore, the proposed training approach enables time-varying CFO compensation of OFDM-Autoencoder in almost all Doppler scenarios with the same structure. In the case of high-rate variation, the OFDM-Autoencoder can achieve almost 4 dB performance gain for a BLER of 1e−2 compared to a conventional system.