A Lightweight Multi-Task Neural Network Based on Knowledge Distillation for Joint Modulation Format Identification and Optical Signal-to-Noise Ratio Estimation

A lightweight multi-task neural network (LMTNN) based on knowledge distillation (KD) is proposed to enhance modulation format identification (MFI) accuracy and reduce optical signal-to-noise ratio (OSNR) estimation error under complex channel interference. First, the receiver side signal after carrier frequency recovery (CFR) is normalized by the maximum amplitude. Then, wavelet coefficients are extracted using wavelet transform, and spectral statistics are computed by applying FFT to signals of different orders, thereby constructing a multi-dimensional joint feature representation. Finally, based on the extracted features, a teacher model based on a deep neural network (DNN) architecture is trained. The knowledge of the teacher model is transferred to the structurally simplified student model by leveraging KD technology, thus constructing a computationally efficient LMTNN. Experimental results demonstrate that 100% identification accuracy is achieved at minimum OSNR thresholds of 3 dB, 8 dB, 10 dB, 11 dB, and 7 dB for PDM 4QAM/-8QAM/-16QAM/-32QAM/-64QAM, with OSNR estimation errors remaining within 0.12 dB. Compared with a deep neural network based on high-order FFT features and a convolutional neural network based on constellation diagrams, the proposed method achieves an overall identification accuracy of 98.7%, exceeding the performance of the comparative methods, which reaches 95.1% and 96.3% accuracy, respectively. For the OSNR estimation task, the proposed method maintains a root mean square error (RMSE) below 0.16 dB, which is substantially lower than the RMSE achieved by the comparative methods, measured at 0.28 dB and 0.24 dB, respectively.