Neural Network Enhanced Scaling Factor Estimation in Rate-Adaptive Coherent Systems

We propose a feed-forward fitting artificial neural network (ANN) to estimate scaling factors in the adaptive communication system supporting probabilistically shaped QAM signals with variable entropy. Utilizing simulation data sets generated under different additive white Gaussian noise channels, we enhanced the robustness of the ANN. The impact of hidden layer neuron numbers and probe length on the convergence speed and normalized general mutual information performance is investigated. Experimental verification has been performed in a rate-adaptive 48 to 96 Gbit/s coherent system under different optical signal-to-noise ratios. The result shows that our sequence-based monitoring scheme estimates effectively the scaling factors over a continuous range in the rate-adaptive coherent system.