A Fast Converging Learning-Based Method for Linear and Nonlinear SNR Estimation in Optical Fiber Communication Systems

This paper presents an innovative rapid-convergence learning framework based on the extreme learning machine (ELM) for the estimation of the linear and nonlinear signal-to-noise ratio (SNR) components in coherent optical fiber communication systems. The ELM estimator jointly obtains both SNR components using input features based on the mean squared error (MSE) between the transmitted and the received pilot symbols, as well as the entropy extracted from the amplitude values of the received data symbols. The MSE provides a statistical measure of the errors between the transmitted and received pilot symbols, whereas the entropy quantifies the dispersion within the received data symbols. These input features provide valuable insights into the characteristics of both linear and nonlinear SNR components. A comprehensive computational complexity analysis of the proposed ELM estimator is carried out in terms of real multiplications and additions, to quantify its computational demands. Numerical results reveal that the proposed ELM estimator offers substantial advantages in training time, estimation accuracy, and computational complexity compared to existing estimators in the literature.