Training data generation and validation for a neural network-based equalizer.

Area of Science:

• Optical Communications
• Machine Learning in Telecommunications
• Signal Processing

Background:

• Neural networks (NNs) are powerful tools for mitigating impairments in fiber optic communication.
• NN-based equalizers offer superior performance over traditional methods for linear and nonlinear distortions.
• A key limitation is the NN's ability to learn patterns in pseudo-random bit sequences (PRBS) used for training, hindering generalization.

Purpose of the Study:

• To address the issue of neural network (NN) learning PRBS training data in optical communication.
• To propose a novel strategy for generating robust random sequences that resist NN learning.
• To improve the equalization performance and reliability of NN-based systems in optical networks.

Main Methods:

• Developed a combination strategy to construct a highly random sequence resistant to machine learning.
• Trained and tested neural network equalizers using the proposed sequence.
• Validated the approach through simulations on an additive white Gaussian noise channel and experiments on a real intensity modulation/direct detection system.

Main Results:

• The proposed random sequence effectively prevents neural networks from learning its generation rules.
• Demonstrated significantly improved equalization performance compared to traditional PRBS training.
• Experimental validation confirmed the scheme's effectiveness in a practical optical communication setup.

Conclusions:

• The proposed random sequence generation strategy overcomes NN overfitting issues in optical communication equalization.
• This method enhances the robustness and performance of NN-based equalizers.
• The findings are crucial for advancing reliable high-performance optical communication systems.