Performance of Hybrid Clustering-Classification Approach for Dual-Band System in a Mode-Locked Fiber Laser

Abstract

This paper presents the performance results of a hybrid machine-learning model in the task of classifying light pulse spectra in the regimes (modes) of the following optical system: mode-locked in Erbium-doped fiber laser using nonlinear polarization rotation based on monolayer graphene. The four modes studied are continuous waves, pulses at wavelengths 1533 and $1555~nm$ , and Dual-Band (both wavelengths). The model is a mix of an unsupervised process for identifying pulses at $1533~nm$ and a supervised process for characterizing the remaining modes. The algorithms used are K-means, for the unsupervised stage, and Light Gradient Boosting Machine for supervised learning. Performance is mainly reported by using balanced accuracy, where the model reached 88% compared to manual classification techniques. We also tested the classification speed of our model regarding manual process. We found an average computing time of 10.8 ms for the trained model whereas former technique time was around three orders of magnitude above. This represents a huge improving in time consumption in classification.

Keywords

• Computational processing time
• dual-band
• hybrid machine learning model
• mode-locked lasers