A robust machine learning approach for DC bias prediction in DCO-OFDM based Li-Fi systems

Area of Science:

• Optical Wireless Communications
• Machine Learning Applications
• Signal Processing

Background:

• Direct Current (DC) bias is crucial for spectrally efficient Light Fidelity (Li-Fi) systems using Optical Orthogonal Frequency Division Multiplexing (DCO-OFDM).
• Previous machine learning (ML) models for DC bias prediction showed limited generalization and robustness to data shuffling.

Purpose of the Study:

• To improve the accuracy and robustness of DC bias prediction in DCO-OFDM systems for Li-Fi.
• To systematically evaluate and select advanced ML regression models for reliable DC bias prediction.
• To identify key features influencing optimal DC bias.

Main Methods:

• Utilized LazyPredict Algorithm (LPA) for systematic regression model evaluation.
• Employed Random Forest (RF) as the optimal ensemble learning algorithm.
• Conducted feature importance analysis, hyperparameter tuning, and bootstrap sampling for validation.

Main Results:

• The Random Forest model achieved a high R-squared of 0.953 and a low RMSE of 0.233.
• A robust ML regressor selection process was established, outperforming previous polynomial regression models.
• Friedman statistical test validated model performance over multiple iterations.

Conclusions:

• The proposed ML approach significantly enhances DC bias prediction accuracy and robustness in Li-Fi.
• The Random Forest model demonstrates superior generalization capabilities across diverse data distributions.
• This work provides a reliable method for optimizing DC bias in DCO-OFDM systems.