Estimation of QoT in an Operational WDM Transmission System Using Artificial Neural Networks With Deep SHAP and Kernel SHAP Algorithms

Abstract

In this work, an artificial neural network (ANN) architecture is proposed to estimate the optical signal-to-noise ratio (OSNR) used in a dense wavelength-division multiplexing (DWDM) transmission system in operational stations. The proposed neural network model, using the TensorFlow platform and Deep SHAP and kernel shape interpretation, provides a very close approximation of the values measured by the optical spectrum analyzer (OSA). The mean absolute error (MAE) estimate in the presented work is less than 0.04 dB for OSNR values. The proposed technique provides an updated network state under test based on real periodic measurements of operational systems. Also, the consistency of the proposed model with previous theoretical studies is evaluated and found acceptable. Using complex and accurate models such as ANN for operational commercial systems does not necessarily perform according to theoretical analyses. This is due to the complexities of analyzing transmission quality in optical systems and other factors such as equipment depreciation that are not usually considered in theoretical studies. Feature importance analysis can be effective in simplifying models based on network characteristics by eliminating less influential features while maintaining the required estimation accuracy.