Nonlinear Impairment Compensation Based on Optical Fiber Characteristic Monitoring Guided Bi-LSTM Network (Invited)

Objective With the rapid development of technologies such as the Internet of Things, cloud computing, and big data, global bandwidth demand has grown explosively. Optical fiber backbone networks, which carry approximately 90% of global traffic, are increasingly constrained by the accumulation of nonlinear impairments in long-haul transmission, limiting further improvements in optical communication capacity. Traditional nonlinear compensation (NLC) algorithms, including digital backpropagation based on the nonlinear Schrödinger equation, Volterra nonlinear equalization, and perturbation-based NLC methods—rely heavily on prior knowledge of the optical fiber system and suffer from high computational complexity. Machine-learning-based NLC approaches are typically categorized into clustering-based and neural-network-based methods. Clustering algorithms effectively mitigate minor impairments in short-distance links and low-order modulation formats, whereas neural-network-based methods can compensate for nonlinear impairments in long-haul transmission systems. However, their robustness degrades under varying transmission conditions, primarily because the trained networks lack sufficient prior knowledge when encountering signals impaired under different channel conditions. Methods To address the issue of limited robustness in fiber nonlinear impairment compensation caused by insufficient prior knowledge of the channel, this paper proposes a nonlinear equalization method based on a bidirectional long short-term memory (Bi-LSTM) network guided by optical fiber feature monitoring. Nonlinear impairment monitoring (NLIM) is achieved by inserting zero-power gaps into the temporal signal sequence, which are then quantified as the generalized nonlinear signal-to-noise ratio (G-NSNR). The G-NSNR, serving as prior information, is jointly input with the signal into the Bi-LSTM network—sensitive to temporal dependencies—to train a neural-network-based nonlinear equalizer with enhanced robustness against optical fiber NLI impairments. Results and Discussions To verify the feasibility of the proposed scheme, experimental validation was conducted on both a dual-polarization 120 Gbit/s uniform 64QAM system and a 138 Gbit/s probabilistic shaping (PS) 64QAM system. For the uniform signal at a launch optical power (LOP) of 1 dBm, the bit error rate (BER) with the proposed NLIM Bi-LSTM scheme is reduced by 0.0045 compared with the Non-NLIM Bi-LSTM scheme. For the PS signal at the same LOP, the BER is reduced by 0.0021. The PS signal exhibits a smaller BER improvement than the uniform signal because its probabilistic distribution inherently provides better tolerance to optical fiber nonlinear impairments (NLI). Since the optimal LOP varies with shaping level and transmission distance, a fair comparison was performed at a fixed LOP of 0 dBm, adhering to the single-variable control principle. Increasing the number of spans exacerbates the impact of fiber NLI on transmission performance, resulting in greater nonlinear phase distortion. At a transmission distance of 750 km, the proposed scheme achieves BER reductions of 0.0068 and 0.0053 for the uniform and PS signals, respectively, compared with the Bi-LSTM scheme without NLIM. In addition, for both uniform and PS signals, the proposed network exhibits performance comparable to that of the retrained matching Bi-LSTM network under channel-matched conditions at a fixed LOP, while delivering improved BER performance as the transmission distance increases. This occurs because, with greater distance, accumulated dispersion renders the fixed time-step length of the Matching Bi-LSTM scheme insufficient to capture the necessary temporal information for effective compensation. In contrast, the proposed NLIM Bi-LSTM scheme maintains ideal compensation performance, as its acquisition of prior information is guided by optical performance monitoring rather than channel-matched retraining. Specifically, at a transmission distance of 750 km, the proposed scheme achieves BER improvements of 5.19% and 6.54% over the Matching Bi-LSTM network for the uniform and PS signals, respectively, demonstrating excellent robustness under varying transmission conditions. The computational complexity of the proposed scheme was evaluated and compared with that of the Non-NLIM Bi-LSTM scheme using the real multiplications per symbol metric. Under optimal performance conditions, the proposed NLIM-guided Bi-LSTM scheme reduces computational complexity by 32.15% and 31.95% for uniform and PS-64QAM signals, respectively, compared with the Non-NLIM Bi-LSTM scheme. Conclusions To address the limited robustness of traditional nonlinear impairment equalization in optical fiber communication systems, this paper proposes a nonlinear impairment compensation scheme based on a Bi-LSTM network guided by optical fiber feature monitoring. Channel prior knowledge is obtained through a zero-power-gap-based NLIM method, which guides the Bi-LSTM network to compensate for nonlinear impairments in long-haul optical transmission. The proposed scheme was experimentally validated on a multi-span single-wavelength coherent optical communication system supporting 120 Gbit/s DP-64QAM and 138 Gbit/s DP-PS-64QAM signals. Experimental results show that, compared with a conventional Bi-LSTM without nonlinear monitoring, the proposed method achieves improved BER performance. Moreover, it enhances the robustness of neural-network-based nonlinear equalizers against variations in channel conditions while reducing computational complexity. This work integrates optical fiber nonlinear impairment monitoring with adaptive machine learning techniques, providing a practical solution for dynamic real-time optical transmission systems.