Use this URL to cite or link to this record in EThOS:
Title: End-to-end deep learning in optical fibre communication systems
Author: Karanov, Boris
ISNI:       0000 0005 0288 5786
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Conventional communication systems consist of several signal processing blocks, each performing an individual task at the transmitter or receiver, e.g. coding, modulation, or equalisation. However, there is a lack of optimal, computationally feasible algorithms for nonlinear fibre communications as most techniques are based upon classical communication theory, assuming a linear or perturbed by a small nonlinearity channel. Consequently, the optimal end-to-end system performance cannot be achieved using transceivers with sub-optimum modules. Carefully chosen approximations are required to exploit the data transmission potential of optical fibres. In this thesis, novel transceiver designs tailored to the nonlinear dispersive fibre channel using the universal function approximator properties of artificial neural networks (ANNs) are proposed and experimentally verified. The fibre-optic system is implemented as an end-to-end ANN to allow transceiver optimisation over all channel constraints in a single deep learning process. While the work concentrates on highly nonlinear short-reach intensity modulation/direct detection (IM/DD) fibre links, the developed concepts are general and applicable to different models and systems. Found in many data centre, metro and access networks, the IM/DD links are severely impaired by the dispersion-induced inter-symbol interference and square-law photodetection, rendering the communication channel nonlinear with memory. First, a transceiver based on a simple feedforward ANN (FFNN) is investigated and a training method for robustness to link variations is proposed. An improved recurrent ANN-based design is developed next, addressing the FFNN limitations in handling the channel memory. The systems' performance is verified in first-in-field experiments, showing substantial increase in transmission distances and data rates compared to classical signal processing schemes. A novel algorithm for end-to-end optimisation using experimentally-collected data and generative adversarial networks is also developed, tailoring the transceiver to the specific properties of the transmission link. The research is a key milestone towards end-to-end optimised data transmission over nonlinear fibre systems.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available