Deep Learning for Wireless Signal Parameter Predictions
This dissertation presents our researches investigating a highly accurate methodology for Deep Learning (DL) prediction of the fundamental set of wireless signal parameters, including channel profile, Doppler shift, and signal-to-noise ratio (SNR), necessary and complete for the optimal reception in today's state-of-the-art wireless communication systems. The methodology consists of a hybrid convolutional neural network and long-short-term-memory (CNN-LSTM) model designed to learn features spatially in every input and temporally across inputs, and two prediction accuracy enhancement techniques: input diversity to empower multiple inputs per prediction, and binary prediction to reduce prediction uncertainty. The inherent advantage of DL prediction is realized, to enhance spectral efficiency, via the utilization of random payload inputs (i.e., non-data-aided), in contrast to pilots of known data needed for theoretical derivation of many traditional estimation techniques. The methodology has been robustly validated with prediction accuracy at ~%95 via simulations of 5G and LTE communication systems with comprehensive real-world environment parameterizations. Four technical papers have been successfully produced, presented, and published. With lessons learned and future research directions, the methodology paves the way for a universal DL prediction methodology in communications and other domains.