Investigating Methods of Artificial Intelligence in Fiber Optic Network Performance Monitoring

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Elastic and cognitive optical networks vary modulation format, symbol rate, launch power and route while a service is carrying traffic, so the physical layer has to be measured continuously rather than engineered once at commissioning. Optical performance monitoring (OPM) and modulation format identification (MFI) supply that measurement, and since 2016 the preferred estimator for both has shifted from analytical models to learned ones. The surveys that mapped this shift were assembled narratively, without a published search protocol, without stated eligibility criteria, and without testing whether the accuracies they tabulated were comparable. Objectives: To determine which learning paradigms and architectures dominate AI-assisted OPM, MFI and fibre nonlinearity compensation and how that balance has moved; how the choice of input representation trades accuracy against computational cost; whether the reported performance figures can be pooled or ranked; which bottlenecks separate laboratory demonstrations from field deployment; and what must change in reporting practice for progress to become measurable. Method: Six sources were consulted, namely IEEE Xplore, Optica Publishing Group, ScienceDirect, SpringerLink, MDPI and arXiv, with Google Scholar used for backward and forward snowballing, over the window January 2014 to July 2026. Search strings combined the monitoring terms ("optical performance monitoring" OR "OSNR estimation" OR "modulation format identification" OR "nonlinearity compensation") with the learning terms ("machine learning" OR "deep learning" OR "neural network" OR CNN OR LSTM OR "support vector machine"). Records were screened against six inclusion and five exclusion criteria and appraised on a five-item, ten-point rubric covering problem definition, data provenance, train and test separation, metric definition and complexity reporting, with 6/10 as the retention threshold. Of 612 records identified, 494 were screened after de-duplication, 143 were read in full and 49 were retained. Results are presented through a PRISMA-style selection diagram, two distribution charts and ten comparative tables. Findings: The retained set comprises 25 primary studies and 24 reviews, tutorials and enabling works. Supervised learning accounts for 22 of the 25 primary studies (88 per cent), and no primary study applies reinforcement learning to OPM or MFI although four of the surveys propose it. By architecture the primary studies divide into feed-forward deep networks (8), convolutional networks (5), recurrent networks (5), classical and kernel methods (3), autoencoders (2), transfer learning (1) and learned multi-step compensation (1); eleven learn from image-like representations and fourteen from symbol sequences or link telemetry. Reported outcomes are strong individually and incommensurable collectively: root-mean-square errors of 0.73 dB for OSNR, 1.34 ps/nm for chromatic dispersion and 0.47 ps for differential group delay from one multi-task convolutional monitor; OSNR standard errors of 0.21 dB to 0.48 dB across four polarisation-multiplexed formats from a Stokes-space deep network; a 0.4 dB Q-factor gain from Parzen-window detection; a mean-square error of 3 x 10^-6 for an autoencoder channel model; and 42 Gb/s over 40 km from an end-to-end learned transceiver. Five different accuracy metrics are in use and no two studies share a dataset or a reference link, so these figures can be reported but not ranked. Novelty/Significance: This is the first review in the area to apply a documented selection protocol with explicit eligibility and quality criteria, making its corpus reproducible. It reads the literature along three axes at once, namely input representation, learning paradigm and deployability, where earlier reviews treated these separately, and it shows that representation rather than architecture governs whether a monitor is deployable. It reports a comparability audit quantifying why the published accuracies cannot be pooled, answers it with an eight-item minimum reporting set proposed here, and identifies the complete absence of reinforcement learning from the primary monitoring literature as a structural and actionable gap. Keywords: Optical Performance Monitoring, Modulation Format Identification, Machine Learning, Deep Neural Networks, Fibre Nonlinearity Compensation, Elastic Optical Networks