A Hybrid Artificial Intelligence Framework for OTDR-Based Anomaly Detection and Fault Classification in Optical Fibre Networks

Optical fibre networks constitute the backbone of modern communication infrastructures, making timely and accurate fault diagnosis essential for maintaining network reliability and service availability. Although Optical Time Domain Reflectometer (OTDR) measurements are widely employed for fibre fault localization, manual interpretation of OTDR traces is labor-intensive, time-consuming, and highly dependent on expert knowledge. Existing artificial intelligence (AI)-based approaches often rely on simulated datasets, focus on a single diagnostic task, or employ a single learning model, thereby limiting their applicability in operational network environments. This paper proposes a hybrid artificial intelligence framework for automated OTDR-based fault detection and multi-class classification using real operational fibre measurements. The proposed framework integrates a convolutional autoencoder for unsupervised anomaly detection, expert-guided ground-truth labelling for reliable fault annotation, and two supervised classifiers—a Random Forest (RF) model and a one-dimensional convolutional neural network (1D-CNN)—for comparative fault classification. The framework was evaluated using 45 healthy OTDR traces for unsupervised model training and 366 operational OTDR traces representing six fault categories: healthy fibre, reflective connector fault, macrobending fault, fibre cut, high-loss splice, and end reflection. Experimental results demonstrate that both the RF and the 1D-CNN achieved an overall classification accuracy of 78% on the expert-labelled operational dataset. The 1D-CNN achieved higher weighted precision (0.80) and weighted F1-score (0.76) by automatically learning hierarchical representations directly from OTDR waveforms, while the RF provided a competitive benchmark with a macro-averaged F1-score of 0.75. The proposed framework combines anomaly detection, expert knowledge, and supervised learning within a unified diagnostic system, offering a practical and scalable solution for intelligent optical network monitoring and predictive maintenance.