Transfer learning based adaptive entropy loading for radio-over-fiber systems.

Area of Science:

• Optical Communications
• Signal Processing
• Machine Learning

Background:

• Radio-over-fiber (RoF) systems offer broadband transmission and flexibility.
• Current entropy loading methods in RoF systems face limitations in dynamic channel conditions.
• Adaptive entropy loading with minimal channel state information is crucial for practical RoF systems.

Purpose of the Study:

• To develop a deep neural network (DNN) based transfer learning model for adaptive entropy prediction.
• To address frequency-selective responses in discrete multi-tone (DMT) signals within RoF systems.
• To enable channel-independent entropy loading for dynamic RoF environments.

Main Methods:

• A deep neural network transfer learning model was employed for adaptive entropy prediction.
• The model utilizes demodulated data and received signal-to-noise ratio (SNR) for entropy estimation.
• Discrete multi-tone (DMT) signals were used to evaluate performance in RoF systems.

Main Results:

• The proposed DNN model achieved capacity-approaching Generalized Mutual Information (GMI).
• Smoother Normalized GMI (NGMI) performance was observed, consistently meeting the 0.83 NGMI threshold.
• The method simplifies implementation by not requiring pre-measured SNR, unlike traditional approaches.

Conclusions:

• The DNN-based transfer learning model provides an effective solution for adaptive entropy loading in RoF systems.
• This approach offers a more practical and channel-independent entropy loading option for dynamic RoF environments.
• The findings contribute to enhancing channel capacity and flexibility in future broadband communication systems.