Enhancing Fault Detection and Localization in Passive Optical Networks Through Advanced Deep Learning and Explainability Techniques

The exponential increase in internet usage and data traffic has significantly increased network complexity. Although fiber optic networks are widely deployed and recognized as the backbone of communication infrastructure due to their reliability, security, and high data throughput, they remain susceptible to failures. Traditionally, faults have been detected manually using optical time-domain reflectometry (OTDR) and visual fault locators (VFL). However, these techniques have become impractical due to the rapid expansion of fiber optic networks. In contrast, this work proposes an advanced multitasking learning framework for efficient fault detection, localization, and faulty link identification in passive optical networks (PONs) with reduced prediction delay. The proposed model leveraged a hybrid long and short-term time-series network (LSTNet) model that integrates an autoencoder, a convolutional layer, and a gated recurrent unit to process sequential data. Our model improves fault management in fiber optic communication networks by forecasting fault severity. Additionally, we apply LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive ExPlanations) explainable artificial intelligence (XAI) techniques to ensure transparency. Finally, the performance of our model is compared with that of convolutional neural networks (CNN), gated recurrent unit (GRU) models, and other existing approaches. The proposed model achieved the highest fault detection accuracy of 99.8%, a mean square error (MSE) of 0.00016, and the shortest prediction delay of 3.2s. For fault localization and link identifications, it achieved an accuracy of 94.2% and 99.04%, respectively, with corresponding MSE values of 0.00071 and 0.00019.