A fiber channel modeling method based on complex neural networks.

Area of Science:

• Optical communications
• Machine learning for communications
• Signal processing

Background:

• Channel modeling is crucial for optical communication networks.
• Real-valued neural networks (RVNNs) do not fully capture complex-valued signal properties.
• Existing methods lack comprehensive feature extraction for optical channels.

Purpose of the Study:

• To propose a complex-valued conditional generative adversarial network (C-CGAN) for optical channel modeling.
• To comprehensively learn channel features using complex-valued signals.
• To evaluate the C-CGAN's performance against real-valued conditional generative adversarial networks (R-CGAN).

Main Methods:

• Developed a C-CGAN architecture with complex-valued windowed input data.
• Evaluated model accuracy and generalization using normalized mean square error (NMSE).
• Benchmarked C-CGAN against R-CGAN across diverse scenarios.

Main Results:

• C-CGAN demonstrated superior generalization across dataset sizes, noise levels, and feature complexities.
• Achieved a stable training process with NMSE below [Formula: see text], outperforming R-CGAN.
• C-CGAN exhibited lower computational complexity (FLOPs) and a self-loop cascading mechanism improved performance by 22.48% on constrained datasets.

Conclusions:

• C-CGAN is an effective model for optical channel modeling, surpassing R-CGAN.
• The proposed method offers improved accuracy, generalization, and computational efficiency.
• C-CGAN provides a robust solution for complex optical channel characteristics.