Mitigating Fiber Nonlinearity with Machine Learning
Abstract
Nowadays, optical communication transmission is based mainly on optical fiber networks.
Increasing demands for higher-capacity systems are hampered by signal distortions
due to nonlinear effects of the commercial optic fibers. Different techniques have been
proposed to reverse and mitigate this noise effect on the transmitted signal such as the digital
backpropagation (DBP), the Volterra nonlinear compensation, the advanced modulation
transmission, and perturbation pre-compensation techniques. While these techniques
achieve good results they are too complicated for practical industrial implementation and
add more complexity overhead on the system.
This thesis is focused on investigating the merits of optical fiber mitigation using Artificial
Intelligence (AI) techniques instead of analytical methods. Different AI techniques
combined with perturbation-based nonlinear compensation method are used to predict the
added nonlinear noise to a 16-Quadrature Amplitude Modulation (QAM) propagating signal.
A MATLAB simulation program has been used to model the propagation of the signal
and generate the transmitted data.
The AI simulations have been employed using Python on dual-polarization single channel
systems using single-stage AI techniques such as Neural Network (NN) at receiver
or transmitter side and Siamese neural network (SNN), or two-stage AI techniques. In the
two-stage method, different supervised classifiers have been used at the receiver side such
as multi-layer perceptrons (MLP), decision tree, AdaBoosting, GBoosting, random forest,
and extra trees while NN is placed at the transmitter. Additionally, different complexity
reduction techniques have been applied to the proposed systems to achieve more practical
performance in industrial environment applications.
For the first time, a nonlinear-compensation robustness study is applied to the proposed
AI techniques by detecting the performance of each technique while changing the
single-mode fiber’s nonlinear coefficient value. Moreover, empirical equations are developed
to represent the system’s Q-factor enhancement achieved using each of the proposed
techniques as a function of the fiber nonlinear coefficient and the data features.
Collections
Cite this version of the work
Marina Melek
(2021).
Mitigating Fiber Nonlinearity with Machine Learning. UWSpace.
http://hdl.handle.net/10012/17780
Other formats