Neural network-assisted joint MFI and OSNR monitoring based on multi-model fitting

A neural network-assisted joint modulation format identification (MFI) and optical signal-to-noise ratio (OSNR) estimation method based on multi-model fitting is proposed for elastic optical networks. The method performs MFI by extracting multi-scale amplitude histogram features based on peak-assisted skewed distribution (PASD) fitting. The features are input to a lightweight neural network for robust MFI over a wide OSNR range. For OSNR estimation, a Rician-Gaussian mixture model is used to fit the received signal distribution, with the resulting features processed by a lightweight neural network to achieve accurate and low-complexity estimation. Experimental results in wavelength division multiplexing long-haul systems demonstrate that the proposed method maintains high effectiveness and robustness across varying OSNRs and channel counts. Compared to the transfer-learning cascade neural network (TL-CNN) method, the proposed method achieves 100% MFI accuracy for 4QAM, 16QAM, 32QAM, and 64QAM under minimum OSNR thresholds reduced by 1 dB, 5 dB, 2 dB, and 3 dB, respectively. Furthermore, the system's OSNR estimation mean absolute error is reduced to 0.091 dB, representing an improvement of more than 58.6% compared to the TL-CNN method. In addition, compared to the TL-CNN method, the proposed method reduces the numbers of multiplications, additions, and network parameters by approximately 85.5%, 91.5%, and 99.3%, respectively. These reductions indicate that the proposed method achieves lower computational complexity and higher efficiency.