QoT Estimation of a Commercial DWDM Transmission System Using Artificial Neural Network and Its Feature Importance Analysis

In this paper, we proposed an artificial neural network (ANN) architecture to estimate the optical signal-to-noise ratio (OSNR) of an in-use commercial dense wavelength-division multiplexing (DWDM) system. The model yields a very close estimation of the values the optical spectrum analyzer measured during periodical maintenance measurements (mean absolute error is less than $\mathbf{0. 0 4}$ for OSNR values in dB). Since our model is based on periodic measurements, it represents an updated network state. Furthermore, to investigate the compliance of the ANN model with previous theoretical studies, we have examined the importance of the features. The results imply that complex and accurate models like the ANN model for in-use commercial systems do not necessarily behave as the theoretical analyses suggest. This is due to the complex nature of quality of transmission (QoT) analysis in optical systems and other factors such as equipment aging and errors not included in the theoretical studies. Feature importance analysis could help in simplifying the models based on network specifications by removing features with low impacts while maintaining the required estimation accuracy.