Abstract

This paper proposes two novel neural network (NN) structures to estimate long-term steady linear and nonlinear signal-to-noise ratio (SNR) components in optical fiber communication systems. The first proposed structure is a parallel NN-based (ParNN) estimator, which estimates each SNR component using a different NN structure and input feature set. A combination of gated recurrent unit and dense layers is used to estimate the linear SNR component. On the other hand, the nonlinear SNR component is estimated using a combination of convolutional layer with dense layer. The proposed input features of the ParNN estimator are generated solely from the received signal without knowledge of the transmitted signal. These features are formed of the lower quartile, upper quartile, and entropy, which can accurately characterize the behavior of the SNR components by measuring the received signal spread and uncertainty. For further improvement of the ParNN estimator, an additional stage is added to form the proposed enhanced ParNN (EParNN) estimator. This additional stage consists of two feedforward NNs (FFNNs), each with a single dense layer, where the first FFNN is used to estimate the linear SNR component and the second one estimates the nonlinear SNR component. The input of this additional stage is a combination of the input features and output of the ParNN estimator. The computational complexity is derived for the proposed estimators. The training and testing dataset is built from 16-ary quadrature amplitude modulation of a dual-polarization on a wide range of standard single-mode fiber system configurations, e.g., number of wavelength division multiplexing channels, optical launch power, and number of spans. Numerical results demonstrate that the proposed ParNN estimator achieves better SNR estimation accuracy with comparable computational complexity compared to the most efficient work in the literature. The proposed ParNN estimator can independently estimate each SNR component, in which the complexity per SNR component is reduced. Moreover, the proposed EParNN estimator improves SNR estimation accuracy compared to the ParNN estimator at the expense of a slight increase in the computational complexity.