Mixture-of-domain-experts model for accurate linear and nonlinear SNRs monitoring in unseen optical network conditions

Accurate monitoring of linear and nonlinear signal-to-noise ratios (SNRs) is critical for assessing the signal quality in fiber-optic networks. The existing data-driven methods for joint linear and nonlinear SNRs monitoring experience accuracy degradations under network conditions not represented in the original training data. To address this challenge, we propose a mixture-of-domain-experts (MoDE) model that uses higher-order mixed and central moments extracted from 20 GBaud dual-polarization quadrature phase-shift keying (DP-QPSK) received signals after standard digital signal processing (DSP). Our MoDE model trains domain-specific experts for diverse training conditions and during the inference phase, it adaptively combines their outputs for SNRs estimation under unseen network conditions. The model is experimentally evaluated in several unseen network conditions and compared with a baseline multilayer perceptron (MLP) model. The results validate that MoDE model successfully reduces the mean absolute error (MAE) of linear and nonlinear SNRs, averaged across various testing domains, by up to 0.87 dB and 4.18 dB, respectively, and lowers the worst-case MAE from 8.06 dB to 2.73 dB. These results demonstrate superior generalization performance of our model for joint linear and nonlinear SNRs monitoring and without requiring any labeled data from unseen network conditions for model retraining purpose.