Multi-feature fusion for joint modulation format recognition and OSNR monitoring based on a lightweight adaptive attention network

To address the challenge of achieving both high accuracy and low computational complexity in joint modulation format recognition (MFR) and optical signal-to-noise ratio (OSNR) monitoring, a multi-feature fusion method based on a lightweight adaptive attention network is proposed. A lightweight feature extraction scheme is first applied to the constant modulus algorithm (CMA)-equalized signal to obtain multiple amplitude statistical and entropy features for efficient signal representation. These features are then input into the proposed lightweight adaptive multi-task squeeze-and-excitation deep neural network (LA-MT-SEDNN) integrated with a hyperparameter tuning strategy based on Bayesian optimization (BO), enabling accurate and low-complexity joint estimation. Experimental results in wavelength-division-multiplexed (WDM) long-haul systems demonstrate strong accuracy and robustness under different OSNR and channel conditions. Compared with a conventional multi-task neural network, the proposed method achieves 100% MFR accuracy at minimum OSNR values reduced by 4 dB, 5 dB, 5 dB, and 2 dB for 4QAM, 8QAM, 16QAM, and 64QAM, respectively. For OSNR monitoring, it attains 0.065 dB MSE, 0.169 dB MAE, and 0.233 dB RMSE, with MAE and RMSE reduced by 63.9% and 65.7% compared with MTL-CNN and adaptive MTL-DNN. Moreover, the multiplication time complexity is reduced by approximately 98.9% and 95.2%, demonstrating significant performance improvements with substantially lower computational complexity.