Enhanced Signal-to-Noise Ratio Estimation in Optical Fiber Communications: A Pilot-Based Approach

This paper presents two innovative, pilot-assisted, neural network (NN)-based signal-to-noise ratio (SNR) estimators for application in optical fiber communications. These estimators, termed pilot-assisted feature complexity reduction (PAF-CR) and pilot-assisted feature accuracy enhancement (PAF-AE), are designed to jointly estimate both linear and non-linear SNR components. The architectures of these proposed estimators employ feedforward NNs (FFNNs) for the SNR estimation, with PAF-CR utilizing a two-hidden layer FFNN and PAF-AE employing a single-hidden layer FFNN. Novel features are extracted from pilot signals to utilize the pilot overhead in transmitted signals, such as mean absolute error and mean signed deviation, which statistically measure the error between transmitted and received pilot signals. Additionally, features are directly extracted from the received data signal, such as average absolute deviation, entropy, and arithmetic mean, to capture its statistical dispersion characteristics. The proposed features are carefully selected to effectively capture the characteristics of both linear and non-linear SNR components. The estimation accuracy of the SNR components achieved by the proposed estimators is evaluated using the normalized root mean square error and the standard deviation of the estimation errors. A comprehensive computational complexity analysis of the proposed PAF-CR and PAF-AE estimators is conducted, expressed in terms of real-valued multiplications and additions. Numerical results illustrate that the proposed PAF-CR and PAF-AE estimators achieve a favorable trade-off between the SNR estimation accuracy and computational complexity compared with existing literature estimators. The proposed PAF-CR offers significant computational complexity reduction with a slight enhancement in estimation accuracy, while the proposed PAF-AE provides substantial estimation accuracy improvement while slightly decreasing computational complexity.