Interpretable and lightweight temporal neural network equalizer for nonlinear impairment in DP-64QAM Nyquist-WDM...

With the advancement of high-speed long-haul optical fiber systems, channel nonlinearities have emerged as a key bottleneck to performance. Traditional nonlinear equalizers are constrained by high computational complexity and limited real-time performance. Although machine learning-based methods demonstrate significant potential, they still face challenges such as heavy model complexity, high inference latency, and limited physical interpretability. To address these challenges, a novel perturbation-based nonlinear equalization neural network is proposed. It extracts the principal component features of the signal using first-order perturbation theory-based feature mapping and utilizes a closed-form continuous-time (CFC) neural network to achieve low-complexity signal equalization. The proposed method is well-suited for dynamic real-time transmission systems and features high efficiency, lightweight design, and strong interpretability. The experimental results show that the proposed method improves the Q-factor by 1.2 dB at the optimal launched optical power (LOP) and achieves a maximum Q-factor improvement of 4.7 dB across the measured LOP range. Compared to traditional digital backpropagation (DBP) and Volterra nonlinear equalization (VLNE), it extends the LOP range by 1 dBm and 4 dBm, respectively. Additionally, compared to the long short-term memory (LSTM) neural network method, the time complexity is reduced by up to 79.4%. This work combines perturbation theory with adaptive machine learning, providing an effective solution for nonlinear impairment in next-generation optical communication systems and opening promising prospects for high-capacity, low-latency optical networks.