High-precision fault classification for optical time-domain reflectometry in optical fiber communication systems based on artificial intelligence

Abstract

Monitoring optical fiber networks requires the analysis of optical time-domain reflectometry (OTDR) traces. However, the conventional approach of expert interpretation of these traces can be time-consuming. This work introduces an artificial intelligence-based approach to solve this problem using a two-layer long short-term memory (LSTM) neural network with dropout regularization to classify OTDR faults automatically. The proposed solution, implemented using TensorFlow/Keras and scikit-learn, is trained on 34 features including signal-to-noise ratio (SNR), power measurements, position, reflectance, and loss characteristics, using a dataset of 125,835 OTDR traces. The model achieved 100 % classification accuracy on 25,167 testing samples across eight different fault classes and generalization performance was confirmed using cross-validation (mean accuracy of 99.92 %) and robustness to noise in the input data (accuracy of 97 % with 5 % noise added to inputs). These results show that the proposed model outperforms the state-of-the-art methods including CNN-BiLSTM [97.3 %] and Random Forest [93.5 %]. The SHAP analysis shows that the loss characteristics (19 %) and reflectance measurements (18 %) are the most important features for making classification decisions. The model also has fast inference time (8.3 ms on GPU and 23.7 ms on CPU) that makes the proposed solution suitable for real-time fault classification in optical communication systems. This can help reduce network downtime and maintenance costs, offering a practical tool for automated fiber-network monitoring.