Adaptive multi-task convolutional neural network for optical performance monitoring

Abstract

In order to enhance the study of optical communication, we collaboratively monitor modulation format identification (MFI) and optical signal-to-noise ratio (OSNR) estimation using few-shot learning (FSL) and transfer learning techniques. The constellation diagrams were input into the network, with MFI and OSNR classification treated as joint modulation tasks. Utilizing a 16-way-8-shot dataset configuration, the adaptive few-shot transfer learning network (AFTLN) algorithm achieved outstanding performance, attaining 100% accuracy in MFI and 97.29% accuracy in OSNR classification. Ablation studies were conducted on adaptive few-shot meta-learning network (AMCN), adaptive multi-task learning (AMTL), adaptive direct detection network (ADTN), and AFTLN. Notably, only the AFTLN method achieved 100% accuracy in MFI, while its OSNR classification accuracy was approximately 15% higher than that of AMTL. An analysis of the confusion matrix revealed that OSNR classification errors were primarily concentrated around 25 dB for 64QAM and 29 dB for 128QAM. The adaptive weight variation curve was also analyzed, indicating that after approximately 750 epochs, the model primarily concentrated on the first task of MFI. Subsequently, a dataset from optical fibers collected over a distance of 400 km was assembled to validate the feasibility of the proposed algorithms. The classification accuracy for MFI remained at 100%, while the OSNR estimation accuracy was 95.21%, reflecting only a roughly 2% decrease. This demonstrates that the algorithm exhibits strong robustness and is effective in diagnosing optical performance.