DATA-DRIVEN MODELING OF ERBIUM DOPED FIBER AMPLIFIERS BY NEURAL NETWORKS
Dependence of EDFA gain shape on input power and input spectrum shape is modelled using a simple neural network-based architecture for amplifiers with different gains and output powers. The model can predict the gain within ±0.1 dB. While the model has good success predicting the performance of an EDFA it is trained with, it is not as successful when predicting a different EDFA, or the same EDFA with different pump power. Retraining the model with a small amount of supplementary data from a separate EDFA makes the model able to predict the performance of the second EDFA with little loss in performance. Experiments show that machine learning model of an EDFA is capable of modelling spectralhole burning effects accurately. As a result, it significantly outperforms black-box models that neglect inhomogenous effects. Model achieves an average RMSE error of 0.016 dB between the model and measurements.
1 . A method for data-driven modeling of erbium doped fiber amplifiers (EDFA) by neural networks, the method comprising:
operating a measurement arrangement including an amplified spontaneous emission (ASE) source, an optical spectrum analyzer (OSA), and an EDFA-device under test interposed between the ASE and OSA, such that input, and output spectra are generated;
determining a power represented by spectra peaks, along with a corresponding gain;
generating a set of input node values using the determined power and corresponding gain;
training a neural network using the set of input node values; and
using the trained neural network to predict the gain of a different EDFA than the one used to train the neural network.
2 . The method of claim 1 wherein a reference power and estimated gain for the different EDFA is applied to the trained neural network to predict the gain of the different EDFA.
3 . The method of claim 2 wherein the reference power and estimated gain for the different EDFA is determined by measuring a gain using a flat input spectrum with several input posers.
4 . The method of claim 3 wherein the neural network is retrained with limited supplementary data including a flat input with varying input power levels.
5 . The method of claim 4 wherein the limited supplementary data includes input spectra with varying linear tilts.