Bending Recognition And Curvature Sensing Of Multimode Fiber Based On The Analysis Of Fiber Specklegrams Using Deep Learning

The interference of different modes in multimode fiber will form the specklegrams with a complex bright spot distribution on the end facet of the fiber.Since the formation of the specklegrams is closely related to the structure of the multimode fiber and the surrounding environment,the detection and processing of specklegrams can be used to sense the state of the optical fiber.With the development of optical fiber specklegram detection technology and digital image processing technology,fiber specklegram sensor possesses the unique advantages in many fields and is worthy of investigating.In this paper,the output specklegrams of multimode fiber under different bending radius are simulated and experimentally studied.Through the analysis of specklegrams based on the deep learning method,the recognition of fiber bending state is realized,and the regression relationship between fiber spot and fiber curvature is obtained,and the curvature sensing is realized.The main research contents and innovations are as follows:(1)A detailed simulation of the mode excitation and the formation of specklegram when the multimode fiber under bending status was carried out,and the characteristics of the output specklegrams of the multimode fiber with different structural parameters and different bending radii are analyzed and compared.(2)An experimental setup that can automatically collect the output specklegrams under different fiber curvatures is designed,and a large number of specklegrams were efficiently and automatically obtained using this experimental setup,which shows the good accuracy and robustness of the setup.(3)Two kinds of optical fiber bending state recognition schemes based on specklegrams are proposed.The output specklegrams of the two kinds of multimode fibers under 21 different curvatures are collected as a data set.The first scheme divides the data set into a training set and a test set proportionally,which is used for training the convolutional neural network and testing the performance of the model,respectively.It can accurately identify the curvature of the optical fiber with corresponding specklegrams,and the classification accuracy is up to 96.6%.Since the specklegrams contains rich texture information,the second scheme combining image texture feature extraction and artificial neural network to accurately identify the specklegrams is proposed,and the classification accuracy is 95.95%.Both of the above two schemes show good recognition accuracy,which fully demonstrates the feasibility of the specklegrams as an indicator of the optical fiber state.(4)Regarding the optical fiber curvature sensing based on the specklegrams as a regression problem of deep learning,a regression convolutional neural network for optical fiber curvature sensing is built.A large number of specklegrams output from the multimode fiber under different curvature intervals are divided into training set and test set in proportion.The training set are used to train the convolution neural network and the test set are used to test the convolution neural network model.Based on the above analysis,the generalization ability and sensing ability of the network model are tested by using the specklegrams obtained under new curvatures,and the corresponding fiber curvature can be accurately predicted for any specklegrams within the curvature range1.55-6.93m^-1.For 94.7%of the specklegrams in the test set,the predicted curvature errors of the convolutional neural network are within 0.3m^-1.This proves the feasibility of fiber curvature sensing based on convolution neural network and provides a simple and effective scheme for optical fiber curvature sensing,and an effective method to improve the accuracy of the sensor.