Deep Learning Based Prediction of Doppler Shift for Mobile Communications

This paper presents a Deep Learning (DL) based methodology to predict the Doppler shift in mobile communications. MATLAB simulations conducted on single-carrier, LTE and 5G systems, in the Doppler shift range of [0, 50, …, 550] Hz, attain a prediction accuracy at ~95%, a normalized mean squared error (nMSE) between [10
⁻²
, 10
⁻³
] and a prediction latency equal to the duration of a sequence of five inputs (e.g. frames, subframes and slots, respectively). Simulations are performed on random payloads (e.g., non data aided - NDA), time-domain and frequency-domain signals embedded with variable modulation types [QPSK, 16QAM, 64QAM], delay profiles (i.e., LTE's [EPA EVA ETU]), and signal-to-noise ratios (SNRs) [-10, 20] dB. The methodology utilizes a hybrid model of a Convolutional Neural Network (CNN) and a Long Short Term Memory (LSTM) and two techniques to enhance prediction accuracy, namely, input diversity and binary prediction.