Transfer learning enhanced convolutional neural network for efficient monitoring of a rate-adaptive coherent transmission system

With the demand for capacity and the complex evolution of network topology, flexible coherent communication transceivers are being considered for deployment in a short reach (< 40 km) scenario. The probability constellation shaping (PCS)-based system provides the capability to continuously change the data rate of communication systems, thus calling for the need to monitor both the link rate and the network performance. In this work, a transfer learning (TL)-enhanced entropy and performance estimation scheme is developed to monitor a rate-continuous coherent optical transmission system with various signal-to-noise ratios (SNR). The scheme is demonstrated over a 40-km coherent transmission system with a line rate varying from 128 to 192 Gbps. The entropy is varied from 4 to 6 bits with an interval of 0.05 bits, corresponding to a fine-grained rate of 1.6 Gbps. The mean absolute error (MAE) of entropy tasks reaches 0.158 bits with TL, while that of SNR tasks is reduced to 0.44 dB with TL (for 6 bits of entropy). Our solution allows fine-grained monitoring of link rate and SNR for future adaptive optical communication systems.