Machine learning equalisation techniques for future passive optical networks
Loading...
Files
Full Text E-thesis
Date
2024
Authors
Murphy, Stephen
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
Passive Optical Networks (PON) have become crucial in delivering reliable, high-capacity, and low-latency fiber internet connectivity, largely replacing legacy copper cabling. Research into next generation 100 Gbit/s systems is ongoing to support future bandwidth demands, and the adoption of advanced modulation formats, such as 4-level pulse amplitude modulation (PAM4), will be key to achieving this. However, optical amplification will be required to support PAM4 transmission due to the strict optical loss budgets imposed by legacy network infrastructure. Semiconductor optical amplifiers (SOAs) have been widely proposed, but introduce non-linear signal distortions which can greatly restrict PAM4 performance. Moreover, fiber dispersion impairments are also a concern given PAM4’s high linearity requirements. This thesis explores how neural network-based equalisation (NNE) can offer new solutions to these traditional physical limitations, and enable future 100 Gbit/s PAM4 PON systems.
Through rigorous experimental analysis, this work identifies SOA-induced non-linear patterning as a major barrier to achieving the required 19.5 dB dynamic range requirement for upstream PON transmission. Gated recurrent neural networks are proposed to overcome this effect, and are shown to fully recover PAM4 signals for a range of bit rates up to 128 Gbit/s. The proposed NNE achieves > 28 dB dynamic range in a 100 Gbit/s experimental PAM4 PON scenario, where additional non-linear effects such as fiber dispersion and electrical receiver saturation are also present. Meanwhile, conventional linear equalisation techniques are shown to be insufficient to meet the PON dynamic range requirement. Furthermore, complexity analyses of the proposed NNE solutions are included throughout, and multi-symbol equalisation techniques are investigated to alleviate the challenges of feedback-based NNEs in future hardware implementations.
The thesis’ most significant contribution is the invention of SkipNet, an original adaptive neural network algorithm developed by the author. It overcomes the significant obstacle of complex NNE training, and enables real-time, packet-by-packet adaptation to dynamic transmission conditions in burst-mode PON systems. SkipNet is shown to match the performance of conventionally trained NNE solutions, while using as little as 250 training symbols to mitigate non-linear SOA patterning and fiber dispersion. The introduction of SkipNet opens
new pathways for designing future PON systems that are scalable, cost-effective, and capable of meeting the ever-growing demand for high-speed connectivity.
Description
Keywords
Photonics , Machine learning , PAM4 , Semiconductor optical amplifier , SkipNet , RNN
Citation
Murphy, S. L. 2024. Machine learning equalisation techniques for future passive optical networks. PhD Thesis, University College Cork.