Transfer learning approach toward joint monitoring of bit rate and modulation format

Abstract

Convolutional neural network based transfer learning (TL) is proposed to achieve joint optical performance monitoring with bit rate and modulation format identification in optical communication systems. TL is used to improve the execution of various tasks by extracting features without knowing other optical link parameters. Eye diagrams of four different modulation formats are generated at optical signal-to-noise ratios (OSNRs) varying from 15 to 30 dB for two distinct bit rates, which are then identified simultaneously with a trained deep neural network. In addition, comparisons of different TL approaches are presented. The database is divided into distinct categories with varying parameter ranges in offline mode, and prediction models are assigned to each class. The results suggest that the proposed system may greatly increase identification performance over existing strategies by utilizing TL techniques. The impacts of training, testing, and validation data size, as well as model structure based on TL, are also thoroughly investigated. The results reveal that the VGG16 achieves the highest accuracies compared to other deep learning algorithms even at low OSNR values of 20 dB. The proposed structure can intelligently evaluate the signals of future heterogeneous optical communications, and the results can be used to enhance optical network management.