Generalizable fault-tracing framework for optical networks leveraging RDC-PSO-SNN on imbalanced alarm data

Abstract

The stable operation of optical networks is heavily reliant on efficient fault management. Data-driven intelligent methods have shown great promise for this task, yet their performance is severely hampered by the extreme imbalance inherent in real-world alarm data, where common faults produce abundant samples while rare faults yield few or even none. To address this critical challenge, we propose a fault-tracing scheme based on a repulsion-driven chaotic particle swarm optimization–Siamese neural network (RDC-PSO-SNN). At the data level, our approach leverages a Siamese neural network (SNN), which learns a similarity metric between alarm samples instead of direct input–output mappings. This design enables the model to effectively cluster and identify root alarms, even for fault types absent from the training data, thus directly tackling the data imbalance issue. At the model level, we introduce a RDC-PSO algorithm to automate and optimize the hyperparameter tuning of the SNN. By incorporating a particle collision strategy, RDC-PSO effectively prevents premature convergence to local optima, ensuring robust model performance. Evaluations on real operational data demonstrate the superiority of our scheme: the RDC-PSO-SNN model achieves an accuracy of up to 98.87% and maintains a high accuracy of 92.20% even when three fault types are completely unseen during training, significantly outperforming traditional artificial neural network (ANN) and convolutional neural network (CNN) models. Furthermore, the RDC-PSO algorithm reduced the computational complexity by approximately 17-fold and shortened the hyperparameter search time by nearly two orders of magnitude compared to the exhaustive method. The proposed scheme offers a highly accurate, generalizable, and efficient solution for fault tracing in practical optical networks plagued by extreme data imbalance.