A modulation format recognition and optical signal-to-noise ratio monitoring scheme based on residual network and Taylor score pruning

Abstract

Investigating practical methods for real-time monitoring of modulation formats (MF) and optical signal-to-noise ratio (OSNR) in coherent optical communication systems is critical for advancing future dynamic and heterogeneous optical networks. In this work, we propose a residual network with an attention mechanism(SA-ResNet) to perform joint monitoring of MF and OSNR for mainstream quadrature phase shift keying (QPSK) and M-ary quadrature amplitude modulation (MQAM) signals, including 8QAM, 16QAM, 32QAM, 64QAM, and 128QAM. After applying Taylor pruning to the model, its floating-point operations (FLOPs) were reduced from 40.5 M to 9.5 M, and its parameter memory was decreased from 2.6 M to 0.5 M. Notably, following fine-tuning, the model still achieved 100% MF recognition accuracy and an average absolute error of 0.34 dB for OSNR estimation under a sample length of 16,000 and fiber length of 160 km. When the model is evaluated using 5-fold cross-validation, the average MF recognition accuracy is 99.988%, and the mean of average absolute errors for OSNR estimation is 0.32 dB. These results indicate that the proposed model has acceptable monitoring performance and requires relatively low computational resources, which makes it attractive for lightweight application scenarios of optical fiber monitoring systems.