Abstract

This paper presents an Automated-Machine Learning (AutoML) framework for soft-failure classification in optical networks. Unlike traditional manual approaches, the proposed method automatically selects and optimizes machine learning models with minimal human intervention. Using a large dataset based on the Euro28 topology, multiple algorithms are evaluated under the AutoML pipeline and benchmark them against a baseline Deep-Neural Network (DNN). Results show that the Extra-Trees classifier model achieves the best performance (accuracy > 0.999) while reducing training time by more than 3 times compared to the DNN, demonstrating both computational efficiency and predictive superiority. Feature analysis further reveals that the received power and the ECL current are the dominant indicators of failure, providing valuable insights for network monitoring. These findings highlight AutoML as a practical and scalable solution for failure management in future high-capacity optical networks. © 2025 IEEE.