Constellation-based identification of linear and nonlinear OSNR using machine learning: a study of link-agnostic...

Area of Science:

• Optical communication systems
• Machine learning applications in telecommunications
• Signal processing for fiber optics

Background:

• Accurate optical signal-to-noise ratio (OSNR) estimation is crucial for optimizing wavelength division multiplexed (WDM) fiber communication systems.
• Traditional OSNR estimation methods can be complex and may require specialized hardware or transmission patterns.
• Distinguishing between linear (e.g., amplified spontaneous emission) and nonlinear noise is essential for effective system management.

Purpose of the Study:

• To develop and demonstrate an accurate method for estimating generalized optical signal-to-noise ratio (GOSNR) in WDM fiber systems.
• To simultaneously estimate linear and nonlinear noise contributions using a single model.
• To assess the model's universality and adaptability for practical deployment in existing networks.

Main Methods:

• Utilized a multi-tasking convolutional neural network (CNN) trained on experimental data from dual-polarized 32-GBaud 16QAM DWDM links.
• Extracted features from constellation density matrices to train the CNN.
• Cross-trained the model with data from different fiber types within metro networks to evaluate universality.

Main Results:

• Achieved highly accurate GOSNR estimation with a mean absolute error (MAE) of less than 0.5 dB.
• Demonstrated accurate estimation of OSNR due to amplified spontaneous emission (OSNR_ASE) and nonlinear noise (OSNR_NL).
• Validated the model's performance and universality across different fiber types, showing a path towards practical, universal training.

Conclusions:

• The developed CNN provides a robust and accurate solution for GOSNR and noise component estimation in WDM systems.
• The method eliminates the need for high-speed sampling, additional hardware, or special transmission symbols, facilitating easy implementation in deployed systems.
• The study highlights the potential of AI-driven signal processing for enhancing the performance and manageability of modern fiber optic networks.