DeepRS: deep neural network-based in-service Rayleigh-scattering monitoring in bidirectional mode-division multiplexing systems

Abstract

In bidirectional mode-division multiplexing (Bi-MDM) systems, which have the potential for significantly higher transmission spectral efficiency, the complex interplay between mode-coupling and Rayleigh scattering (RS) exacerbates channel instability, posing significant challenges for dynamic network management. To address this, we propose DeepRS, an innovative deep-learning-based scheme for high-precision, in-service RS noise monitoring. By utilizing deep neural networks (DNNs) to extract features from the filtered frequency amplitude histogram (FFAH) of received signals, it enables the efficient signal-to-RS ratio (SRR) monitoring without disrupting signal traffic. In a 3-mode coherent Bi-MDM experiment, DeepRS achieves an impressive SRR prediction accuracy, with a coefficient of determination (R2) exceeding 0.9927 when incorporating crosstalk (XT) pre-prediction. It demonstrates strong adaptability across various operating wavelengths, modulation formats, and Baud rates. The scheme’s outstanding performance has also been experimentally validated in a 10 km-long 6-mode fiber. Furthermore, DeepRS exhibits high robustness to XT prediction errors, maintaining an SRR prediction R2 above 0.9884 when the absolute XT error is within 2 dB. Finally, simulation results confirm its insensitivity to optical distortions such as chromatic dispersion (CD), differential mode group delay (DMGD), and laser linewidth (LW), further validating its robustness for practical Bi-MDM systems.