A data-driven fiber channel modeling method based on deep learning (DL) is introduced in an optical communication system. In this study, bidirectional long short-term memory (BiLSTM) is selected from a diverse range of DL algorithms to perform fiber channel modeling for on—off keying and pulse amplitude modulation 4 signals. Compared with the conventional model-driven split-step Fourier (SSF)-based method, the proposed method yields similar results based on the comprehensive comparison of multiple characteristics associated with the generated optical signals, including the optical amplitude and phase waveforms in the time domain, optical spectrum components in the frequency domain, and eye diagrams after detection in the electrical domain. Additionally, the effects of multiple factors on the modeled fiber channel have also be investigated, including fiber length, fiber nonlinearity, dispersion, data pattern, pulse shaping, and sample rate. The satisfactory fitting results and acceptable mean square errors indicate that the approximate transfer function of the fiber channel is learned by the BiLSTM. Moreover, compared with repetitive iteration SSF, the computing time is significantly reduced by the BiLSTM owing to its independence on fiber length and insensitivity to data size and launch power. Our aim is to demonstrate the BiLSTM is comparable with the conventional model-driven SSF-based method for direct-detection optical fiber system. We think the proposed method could be a supplementary technique that can be used for the existing simulation system and could also be a potential option for future simulation methods.

For any system that follows rigorous mathematical and physical theories like fiber-optic system, system parameter estimation is crucial for system detection and monitoring. In this paper, a physics-informed neural network (PINN)-based method is proposed for optical fiber parameter estimation by solving the inverse problem of the nonlinear Schrödinger equation (NLSE), i.e., inferring the unknown parameters of NLSE. The proposed method considers both the prior knowledge of physical law and the characteristics of observation data from only the transmitted and received waveforms. In PINN, the NLSE and observation data are set as the loss function, and the typical fiber parameters (attenuation, dispersion, and nonlinear coefficient) are optimized iteratively until they satisfy the inferred NLSE by learning the limited observation data. Three scenarios are validated, including single Gaussian pulse propagation, ultrashort pulse propagation in highly nonlinear fiber, and optical signal transmission in standard single-mode fiber. The launch power of the signal is embedded in PINN as a parametric feature, which makes PINN learn the signal transmissions under different powers simultaneously. The weights of mean square error terms for NLSE and observation data in the loss function are properly adjusted to balance learning difficulties. Moreover, the statistical results of 128 different signals and model performance under different signal-to-noise ratios are analyzed. As an extension, fiber length estimation is also studied, and the error is 0.375% when the launch power is 0 dBm. Our work verifiably shows that the proposed approach performs well and can be further extended to multiple applications in fiber optics.

A cost-effective and data size-adaptive optical performance monitoring (OPM) scheme is proposed, which is based on asynchronous delay-tap plot (ADTP) using convolutional neural network (CNN) from the perspective of image processing. First, we design an OPM framework, based on the electrical domain-processing technique for the future optical networks. These networks include coherent detection-based end-to-end channel monitoring at destination node and direct detection-based transmission link monitoring at intermediate node. Aiming at the link monitoring, CNN is applied to recognize and analyze ADTP images that are converted from two-dimension (2D) digital vectors, so that adaptive to the stable algorithm structure. In simulation system, three high-order modulation formats, 16 quadrature amplitude modulation (QAM), 32QAM, 64QAM, are investigated for optical signal-to-noise ratio (OSNR) estimation and modulation format identification (MFI). The 100% accuracies under different chromatic dispersions (CDs) at different iteration epochs are obtained. Compared with asynchronous amplitude histograms (AAH)-based method, the better accuracy and faster convergence rate are achieved, especially in terms of strong CDs. Additionally, the experimental system is also conducted of 16QAM and 64QAM signals. Based on the partially-trained CNN model from simulation, the OSNR estimation accuracies of 16QAM and 64QAM are 97.81% and 96.56%, respectively. The maximum standard deviation is less than 0.45 dB and the MFI accuracies is 99.84%, presenting the satisfactory results and proving the feasibility of ADTP-based image processor for link monitoring at intermediate nodes.

Instead of conventional knowledge-driven modeling, this study introduced data-driven channel modeling methods based on deep learning (DL) in a free-space optical (FSO) communication system to achieve low complexity, less modeling difficulty, high fidelity, and differentiability. The FSO channel was modeled via three DL algorithms: generative adversarial network (GAN), bidirectional long short-term memory (BiLSTM), and Bayesian neural network (BNN). Consequently, their performance was comprehensively compared in terms of optical amplitude waveforms, Kullback-Leibler (KL) divergence, 95% confidence interval, data distribution, and fitted probability density functions. The results indicated that the GAN-based channel exhibited optimal performance, implying its suitability for learning the stochastic distribution of the FSO channel. In addition, the performance of GAN-based channels was examined under different conditions: turbulence intensities (weak, moderate, and strong), transmission distances (3, 4, and 5 km), launch powers (0, 10, and 20 dBm), modulation formats (on-off keying, pulse position modulation, and pulse amplitude modulation 4), and the adaptability to channel variation. The test results were satisfactory and confirmed the ability of the GAN to learn the approximate distribution function of the FSO channel. Thus, the primary aim was to prove that for FSO channels with statistical distribution characteristics, a data-driven DL model can not only serve as an effective surrogate model for specific scenarios but also provide a potential tool for future modeling methods.

A novel low-complexity and nonlinearity-tolerant modulation format identification (MFI) using random forest (RF) is proposed for the flexible coherent receivers (FCRs). The RF-based MFI takes the advantage of distinct features from the amplitude histograms (AHs) and the swarm intelligence of RF to significantly reduce the computational complexity compared with the deep learning-based method. Both numerical simulation and experiments are conducted. The simulation results of polarization multiplexed (PM) 4/8/16/32/64 quadrature amplitude modulation (QAM) wavelength division multiplexing (WDM) systems demonstrate its feasibility and identification performance. In the presence of nonlinear effects, with lower complexity than three other machine learning algorithms (k-nearest neighbors, support vector machine, and deep neural networks), the RF-based MFI can obtain 100% accuracy at the OSNR values greater than or equal to the respective soft decision forward error correction (SD-FEC) threshold. The identification accuracy versus total launch power is experimentally investigated in PM-4/16/32QAM WDM coherent systems. In the case of obvious nonlinear impairments, 100% accuracy can also be achieved above the OSNR corresponding to the SD-FEC threshold. The superiority of the proposed MFI method makes it highly desirable for applications in modulation format-adaptive FCRs.

As the services in elastic optical networks (EONs) are bandwidth-intensive, unpredictable and dynamic, increasing network factors are emerging to affect the performance of network survivable planning and operation from network capacity to efficiency. Most of the traditional protection and restoration approaches may become before long inefficient due to the improvement of a particular network performance metric always at the expense of others. We argue that it would be more beneficial for comprehensive optimization of network performance to consider main network metrics jointly. Moreover, the highly dynamic features of EONs call for the new generation of machine learning-based solutions that are flexible and adaptable to cope with the dynamic nature of services to perform analytics. In this paper, we investigate the problem of global optimization of network performance under survivable EON environment. Specifically, a criterion, named the whole network cost-effectiveness value with survivability (WCES), is defined to measure the overall network performance by balancing the interaction among main network metrics. Then we propose a deep reinforcement learning (DRL) -based heuristic with the objective of improving overall network performance, in which two agents are utilized to provide working and protection schemes converging toward better survivable routing, modulation level and spectrum assignment (S-RMLSA) policies. Numerical results show that the proposed criterion can efficiently measure the overall network performance, and the double-agent DRL-based heuristic can greatly improve WCES while ensuring the network survivability and paying the acceptable extra consumption of request blocking probability.