AI-based OTDR multi-fault detection system for optical fiber networks

Concurrent fault detection is challenging in Fiber optics networks due to the inability of current OTDR-based approaches to detect the presence of coexisting faults. This work proposes a system for concurrent fault detection, localization and classification by utilizing the combination of unsupervised anomaly detection and supervised fault classification, unlike the traditional approach where those two steps are separate. Thresholding reconstruction error values of a Conv1D autoencoder with the dataset of 3910 samples (372 real EXFO OTDR scans and 2000 artificial ones) is performed via a combination of median absolute deviation, Otsu binarisation, and percentile methods. The faults are then classified using CNN + LSTM (validation accuracy = 81.0%) and Random Forest (accuracy = 87.6%) classifiers which identify four types of faults—bend, splice, break, and reflection based on the five-factor quality score. The system showed 90.9% precision, 47.6% recall, and 0.110 km mean localization error on the dataset of 21 faults (artificial and real SOR scans). Nevertheless, evaluation on a larger dataset of 180 real SOR scans demonstrated a huge domain generalization gap (precision = 6.5%, recall = 4.2%, localization error = 0.524 km) implying that this system is just a proof-of-concept solution.