Experimental Demonstration of Soft Failure Identification Based on Digital Residual Spectrum and Machine Learning

While machine learning (ML)-based soft failure management in optical network is flexible and efficient, several challenges stand in the way of its actual deployment. The performance of a trained model in unseen systems is a key issue. In this paper, a soft failure identification (SFI) algorithm based on digital residual spectrum and machine learning is verified by experiments. This algorithm facilitates the deployment of ML technology across different lightpaths. Auto Encoder is adopted to extract features from digital residual spectrum, and support vector machine is used to classify soft failures. Without extra hardware, digital residual spectrum can be readily acquired in coherent receiver. A large number of normal samples are used, which are easy to obtain in a working network. Four systems with different transmission distances or different numbers of wavelength selective switches (WSSs) are performed. With the model trained by samples from one system, the algorithm outperforms with 96.20% identification accuracy for the four systems. The outstanding generalization of the algorithm is confirmed by experiments.