A Neural Network-Based Feature Learning and Processing to Estimate Signal-to-Noise Ratio in Coherent Optical Fiber Systems

Abstract

This paper proposes a novel two-stage neural network (TSNN)-based signal-to-noise ratio (SNR) estimator tailored specifically for coherent optical fiber systems. The proposed TSNN architecture consists of two distinct NN stages. In the first NN stage, a novel technique called feature estimation using NN (FE-NN) is proposed, aiming to decrease the computational complexity by learning feature similarities and estimating some features based on others generated mathematically from the received signal. Subsequent to the FE-NN stage, a second NN stage is meticulously crafted to jointly estimate the linear and nonlinear SNR components with precision. This stage utilizes a novel set of input features generated exclusively from the received signal, without prior knowledge of the transmitted signals. The proposed input features leverage statistical measures such as median absolute deviation, arithmetic mean, and entropy to provide a comprehensive insight into SNR dynamics, thereby enhancing estimation accuracy. A comprehensive analysis of the computational complexity of the proposed TSNN SNR estimator is provided, quantifying the required number of real-valued multiplications and real-valued additions. Performance evaluation of the proposed TSNN estimator is conducted through extensive simulations encompassing 4950 realizations of a standard single-mode fiber wavelength division multiplexing system, employing dual-polarization 16-ary quadrature amplitude modulation. The results underscore the pronounced reduction in computational complexity achieved by the proposed TSNN estimator compared to the most efficient estimators in the literature. Moreover, the proposed TSNN estimator yields superior accuracy in both linear and nonlinear SNR components estimation, thereby highlighting its efficacy in optical communication systems.