[Paper Review] An Overview of Machine Learning-aided Optical Performance Monitoring Techniques

Dativa K. Tizikara, Jonathan Serugunda+1 more|arXiv (Cornell University)|Jun 20, 2021

Optical Network Technologies64 references4 citations

TL;DR

This paper reviews machine learning (ML)-based optical performance monitoring (OPM) techniques for dynamic, high-capacity optical networks, focusing on real-time estimation of impairments like OSNR, CD, and PMD without full signal recovery. It highlights deep learning and neuromorphic approaches—especially photonic reservoir computing—that enable accurate, signal-type-transparent monitoring with reduced training complexity, achieving correlations up to 0.997 and 100% modulation format recognition accuracy.

ABSTRACT

Future communication systems are faced with increased demand for high capacity, dynamic bandwidth, reliability and heterogeneous traffic. To meet these requirements, networks have become more complex and thus require new design methods and monitoring techniques, as they evolve towards becoming autonomous. Machine learning has come to the forefront in recent years as a promising technology to aid in this evolution. Optical fiber communications can already provide the high capacity required for most applications, however, there is a need for increased scalability and adaptability to changing user demands and link conditions. Accurate performance monitoring is an integral part of this transformation. In this paper we review optical performance monitoring techniques where machine learning algorithms have been applied. Moreover, since alot of OPM depends on knowledge of the signal type, we also review work for modulation format recognition and bitrate identification. We additionally briefly introduce a neuromorphic approach to OPM as an emerging technique that has only recently been applied to this domain.

Motivation & Objective

• Address the growing need for real-time, scalable optical performance monitoring in elastic, dynamic optical networks with variable bandwidth and signal parameters.
• Overcome limitations of conventional OPM methods that require full signal demodulation and incur high cost and complexity.
• Enable multi-impairment monitoring (OSNR, CD, PMD, non-linearities) simultaneously and transparently to signal type, reducing dependency on prior signal knowledge.
• Explore emerging techniques like photonic reservoir computing to reduce training complexity and support high-bandwidth, real-time monitoring.
• Survey and update the state of the art in ML-based OPM, including modulation format and bitrate identification, with emphasis on deep learning and multi-task learning frameworks.

Proposed method

• Utilize machine learning models—particularly artificial neural networks (ANNs), deep neural networks (DNNs), and convolutional neural networks (CNNs)—to learn relationships between signal features and impairments without full demodulation.
• Extract input features from signal waveforms, spectra, polarization states, or constellation diagrams in Jones or Stokes space for coherent detection systems.
• Apply dimensionality reduction techniques like Principal Component Analysis (PCA) to reduce feature space while maintaining monitoring accuracy.
• Implement multi-task learning and end-to-end deep learning models that jointly predict modulation format, bitrate, and multiple impairments using a single trained network.
• Leverage photonic reservoir computing as a neuromorphic alternative to reduce training overhead and enable high-speed optical-domain signal processing.
• Train models off-line using large, diverse datasets covering various signal formats, bitrates, and impairment combinations, then deploy for real-time inference.

Experimental results

Research questions

• RQ1 How can machine learning enable accurate, real-time monitoring of multiple optical impairments (OSNR, CD, PMD) without full signal recovery?
• RQ2 To what extent can ML models be made transparent to signal type (modulation format and bitrate) to reduce dependency on prior signal knowledge?
• RQ3 What are the performance trade-offs between traditional ML models (SVMs, ridge regression) and deep learning models (DNNs, CNNs) in OPM and modulation format recognition?
• RQ4 How does photonic reservoir computing compare to conventional neural networks in terms of training complexity and real-time monitoring performance for OPM and MFR?
• RQ5 What impact do varying transmission parameters (e.g., launch power, distance) have on the generalization and accuracy of ML-based OPM systems?

Key findings

• Neural network-based OPM techniques achieved correlation coefficients of up to 0.997 between predicted and actual values for OSNR, CD, and PMD, demonstrating high accuracy.
• Deep learning models, particularly DNNs and CNNs, outperformed traditional methods in coherent detection systems, especially when learning features directly from I/Q samples or constellation images.
• Multi-task learning and end-to-end training enabled simultaneous prediction of modulation format, bitrate, and impairments with 100% identification accuracy in multiple studies.
• Photonic reservoir computing emerged as a promising alternative with reduced training complexity, enabling high-speed, real-time optical signal processing suitable for future networks.
• Models trained on a single set of transmission parameters achieved high accuracy, but performance dropped significantly when applied to new parameter combinations unless retrained for each scenario.
• OSNR estimation using only input power as a feature was demonstrated successfully in one study, showing potential for low-complexity, signal-type-transparent monitoring.