Neural Network Based Perturbation-Location Fiber Specklegram Sensing System Towards Applications With Limited Number of Training Samples

Abstract

Deep learning methods such as convolutional neural networks (CNN) have been shown to be highly effective in complex nonlinear modeling or classification in multimode fiber transmission systems with intensity-only detection. However, such powers are often realized along with time-consuming training processes requiring large number of data samples, which may not be achievable in some practical implementations where only limited numbers of samples are available. In this paper, aiming at high-accuracy perturbation location in a multimode fiber specklegram sensing (FSS) system with small size training data set, we compare and analyze the performance of the CNN based FSS system using fibers under different conditions of mode coupling. We demonstrate that, by utilizing a ring core fiber (RCF) supporting a few weakly-coupled mode groups(MGs), high accuracy of around 100% in the classification of perturbation locations can be more easily achieved with faster convergence speed and fewer training samples, compared with FSS systems using conventional OM3 multimode fibers (MMFs) with hundreds of modes. Furthermore, owing to similar low-level characteristics extracted from the speckle-pattern images, the CNN exhibits good performance of transfer learning in a more practical RCF based FSS system with foot-stepping perturbations, confirming their good potential of extension into practical systems.