Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Transfer Learning for EDFA Gain Modeling: A Semi-Supervised Approach Using Internal Amplifier Features

Agastya Raj, Dan Kilper, Marco Ruffini

TL;DR

This work tackles the challenge of accurately predicting EDFA gain spectra with limited labeled data while enabling transfer learning across different EDFAs. It introduces a Semi-Supervised Self-Normalizing Neural Network (SS-NN) that leverages unlabeled input spectra and internal EDFA features, including VOA-related signals, to achieve high-accuracy gain predictions. The approach delivers MAEs around 0.07–0.08 dB for same-type gains and under 0.19 dB in cross-type transfer across 22 EDFAs on COSMOS and Open Ireland testbeds, outperforming a benchmark that required far more labeled data. By demonstrating effective one-shot TL with a single new measurement per target gain, the method significantly reduces data collection burdens and enables rapid adaptation to heterogeneous EDFA deployments in practical networks.

Abstract

The gain spectrum of an Erbium-Doped Fiber Amplifier (EDFA) has a complex dependence on channel loading, pump power, and operating mode, making accurate modeling difficult to achieve. Machine Learning (ML) based modeling methods can achieve high accuracy, but they require comprehensive data collection. We present a novel ML-based Semi-Supervised, Self-Normalizing Neural Network (SS-NN) framework to model the wavelength dependent gain of EDFAs using minimal data, which achieve a Mean Absolute Error (MAE) of 0.07/0.08 dB for booster/pre-amplifier gain prediction. We further perform Transfer Learning (TL) using a single additional measurement per target-gain setting to transfer this model among 22 EDFAs in Open Ireland and COSMOS testbeds, which achieves a MAE of less than 0.19 dB even when operated across different amplifier types. We show that the SS-NN model achieves high accuracy for gain spectrum prediction with minimal data requirement when compared with current benchmark methods.

Transfer Learning for EDFA Gain Modeling: A Semi-Supervised Approach Using Internal Amplifier Features

TL;DR

This work tackles the challenge of accurately predicting EDFA gain spectra with limited labeled data while enabling transfer learning across different EDFAs. It introduces a Semi-Supervised Self-Normalizing Neural Network (SS-NN) that leverages unlabeled input spectra and internal EDFA features, including VOA-related signals, to achieve high-accuracy gain predictions. The approach delivers MAEs around 0.07–0.08 dB for same-type gains and under 0.19 dB in cross-type transfer across 22 EDFAs on COSMOS and Open Ireland testbeds, outperforming a benchmark that required far more labeled data. By demonstrating effective one-shot TL with a single new measurement per target gain, the method significantly reduces data collection burdens and enables rapid adaptation to heterogeneous EDFA deployments in practical networks.

Abstract

The gain spectrum of an Erbium-Doped Fiber Amplifier (EDFA) has a complex dependence on channel loading, pump power, and operating mode, making accurate modeling difficult to achieve. Machine Learning (ML) based modeling methods can achieve high accuracy, but they require comprehensive data collection. We present a novel ML-based Semi-Supervised, Self-Normalizing Neural Network (SS-NN) framework to model the wavelength dependent gain of EDFAs using minimal data, which achieve a Mean Absolute Error (MAE) of 0.07/0.08 dB for booster/pre-amplifier gain prediction. We further perform Transfer Learning (TL) using a single additional measurement per target-gain setting to transfer this model among 22 EDFAs in Open Ireland and COSMOS testbeds, which achieves a MAE of less than 0.19 dB even when operated across different amplifier types. We show that the SS-NN model achieves high accuracy for gain spectrum prediction with minimal data requirement when compared with current benchmark methods.

Paper Structure

This paper contains 13 sections, 3 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Measurement Setup and Data Collection
  3. Experimental Setup
  4. Measurement Configuration
  5. Model Architecture
  6. SNN Model
  7. Training Process
  8. Training and Test Sets
  9. SNN Model Performance
  10. Transfer Learning
  11. TL Training Process
  12. Results
  13. Conclusions

Figures (6)

  • Figure 1: Measurement setup for the Booster/Pre-amplifier EDFA in COSMOS and Open Ireland testbeds.
  • Figure 2: SS-NN model structure with 5 layers.
  • Figure 3: SS-NN model training framework. Step 1(a) and (b) show the greedy layer-wise pretraining of hidden layers using unsupervised pretraining. This pre-trained model forms the basis for Step 2, where supervised fine-tuning is performed with 256 labeled measurements.
  • Figure 4: Boxplot distribution of absolute errors across all 11 Booster and 11 Pre-amplifier EDFAs for goalpost and random channel loading. The boxes denote the inter-quartile range, and the whiskers denote the min/95th percentile
  • Figure 5: Boxplot distribution of absolute errors across all 22 EDFAs for (a) Booster to Booster TL, (b) PreAmp to Preamp TL, (c) Booster to Preamp TL and (d) Preamp to Booster TL, for random and goalpost channel loading configurations. The boxes denote the inter-quartile range, and the whiskers denote the min/95th percentile
  • ...and 1 more figures