On Using Artificial Neural Networks for Multipath Radio Channel Estimation

Abstract

Line spectral estimation is an important technique widely used in signal processing, e.g., radio channel param- eter estimation. However, the current machine learning-based methods for line spectral estimation are incomplete, and many problems still need to be addressed. We propose an Artificial Neural Network (ANN) architecture that can directly estimate the radio channel delay parameters, including the number of delays present in the radio channel measurements. We propose a robust noise regularization technique, which successfully mitigates the problem of ANN overfitting. Finally, we propose a novel loss function significantly improving the achievable precision of the radio channel parameter estimation. We compare our results with the theoretical limit Cramer-Rao Lower Bound (CRLB) and classical approaches such as the Space-Alternating Generalized Expectation-maximization (SAGE) and Superfast Line Spectral Estimation (SLSE). Our results show that this novel loss function enables the ANN-based delay estimator to approach the CRLB for a single delay case. The proposed method still achieves a super-resolution performance for larger model orders. The ANN- based approach can be approximately 24 times faster than the SAGE algorithm and 180 times faster than the SLSE.