Machine Learning Based Linear and Nonlinear Noise Estimation

Operators are pressured to maximize the achieved capacity over deployed links. This can be obtained by operating in the weakly nonlinear regime, requiring a precise understanding of the transmission conditions. Ideally, optical transponders should be capable of estimating the regime of operation from the received signal and feeding that information to the upper management layers to optimize the transmission characteristics; however, this estimation is challenging. This paper addresses this problem by estimating the linear and nonlinear signal-to-noise ratio (SNR) from the received signal. This estimation is performed by obtaining features of two distinct effects: nonlinear phase noise and second-order statistical moments. A small neural network is trained to estimate the SNRs from the extracted features. Over extensive simulations covering 19,800 sets of realistic fiber transmissions, we verified the accuracy of the proposed techniques. Employing both approaches simultaneously gave measured performances of 0.04 and 0.20 dB of standard error for the linear and nonlinear SNRs, respectively.