Collaborative Split Learning-Based Dynamic Bandwidth Allocation for 6G-Grade TDM-PON Systems.

Area of Science:

• Telecommunications Engineering
• Machine Learning Applications
• Network Optimization

Background:

• Time Division Multiplexing Passive Optical Networks (TDM-PON) require efficient upstream bandwidth management.
• Conventional Dynamic Bandwidth Allocation (DBA) methods exhibit suboptimal performance under dynamic traffic conditions.
• Existing solutions often involve high communication overhead and centralized processing.

Purpose of the Study:

• To propose a novel Collaborative Split Learning-Based DBA (CSL-DBA) framework for TDM-PON systems.
• To enhance predictive traffic adaptation and minimize communication overhead.
• To improve the responsiveness of bandwidth allocation to fluctuating traffic.

Main Methods:

• Implementation of Split Learning (SL) between the Optical Line Terminal (OLT) and Optical Network Units (ONUs).
• Decentralized traffic analysis performed locally at ONUs, transmitting only model updates.
• Extensive simulations across various traffic load scenarios (low, fluctuating, high).

Main Results:

• Achieved over 99% traffic prediction accuracy.
• Demonstrated minimal inference latency and scalable learning performance.
• Reduced communication overhead by approximately 60% compared to federated learning approaches.

Conclusions:

• The CSL-DBA framework offers a significant improvement over traditional DBA techniques.
• The proposed method is highly effective in dynamic traffic environments.
• CSL-DBA presents a viable solution for next-generation 6G-grade TDM-PON systems.