Alatawi, A. (ECE) – Learning-Based Channel Estimation for Next-Generation Wireless Communications
Accurate Channel State Information (CSI) is critical for coherent detection, equalization, and adaptive resource allocation in modern wireless systems. Traditional estimators rely on stationary statistical models, and many learning-based methods assume training and deployment conditions are matched. In practice, these assumptions break down under user mobility and environmental dynamics, leading to degraded performance. This proposal explores machine-learning approaches for channel estimation that address two complementary challenges.
First, we develop an adaptive deep neural network (ADNN) for single-input single-output links over slowly time-varying channels. The method converts readily available physical-layer feedback—cyclic redundancy check (CRC) and automatic repeat request (ARQ)—into reliable self-supervision. Specifically, packets decoded without errors are re-estimated using least squares (LS) across all symbols to generate high-quality labels, and the DNN weights are periodically updated online. This design eliminates the need for ground-truth labels at deployment and enables continual learning. Simulations show that the ADNN tracks distributional shifts and recovers near–linear minimum mean-square error (LMMSE) performance in both mean-square error (MSE) and symbol error rate (SER), whereas a fixed offline-trained DNN degrades as channel statistics change.
Second, we propose a sequence-to-sequence LSTM estimator for orthogonal frequency-division multiplexing (OFDM). The model exploits both temporal and frequency correlation by taking LS pilot estimates from several previous OFDM blocks as input and reconstructing the full channel frequency response of the current block. Trained on realistic time-selective channels such as WINNER II, the LSTM outperforms LS interpolation and recent super-resolution–based methods across a wide range of SNRs, pilot densities, and temporal window sizes.
Finally, the proposal outlines future research on semantic-aware channel estimation using CSI timeliness, and enhanced sequence models with DNN-refined pilots, whole-block inputs, and efficient GRU architectures.
Event Host: Abdulaziz Alatawi, Ph.D. Student, Electrical & Computer Engineering
Advisor: Hamid Sadjadpour & Zouheir Rezki
Zoom- https://ucsc.zoom.us/j/94895993579?pwd=Bs1ppmjqFvNknefRAHoVGXPSXxdZ6i.1
Passcode- 884927