EDFA Soft-Failure Detection and Lifetime Prediction based on Spectral Data using 1-D Convolutional Neural Network

Inhalt:
With the emergence of 5G and the immense demand of data in today’s society the reliability of optical networks is even more important. In this context, maintaining the quality-of-service and reliable operation can only be achieved with effective fault management. Precise fault management requires anomaly detection, identification and reaction. While hard failures, i.e. service disruptions, are detected easily, the task of detecting and identifying gradual degradations in the quality of transmission (QoT), so called soft-failures, is challenging. Machine learning (ML) provides powerful tools to accomplish this task. In this work, a novel three-stage soft-failure detection, identification and lifetime prediction framework of EDFAs based on convolutional neural networks (CNN) is proposed. The CNNs are trained on the spectral data extracted from the electrical power spectrum density (PSD) from a coherent receiver. The output of the framework is the remaining useful life (RUL) of the identified failing EDFA in the link. Extensive simulations are performed to validate the proposed framework. Excellent performance is achieved by the framework with the assumption that just one EDFA in the link is failing. The detection stage reaches an accuracy of 99.89% while the identification stage achieves 98.67% on an unseen data set. The RUL prediction stage achieves high regression (R2-) scores of 0.86 for sudden and 0.97 for gradual degradation of the EDFA pump laser.