Research On Multi-event Recognition Of ?-OTDR Distributed Optical Fiber Sensing System Based On Machine Learning

Phase-sensitive optical time domain reflectometer(?-OTDR)is a typical distributed optical fiber sensing system,which can detect the occurrence of external vibration.Many studies focus on the positioning accuracy and sensitivity of the system,and lack the ability to classify multiple events,especially similar events,which makes it impossible to further expand the application in many fields.It is a common method to classify disturbance signals in time domain or frequency domain by artificial feature extraction,but artificial feature extraction is more complex and application scenarios have limitations.This paper proposes a Multi-event recognition method based on convolutional neural network(CNN)in deep learning and traditional classifier in machine learning.The experiment involved six types of man-made disturbances and two types of natural events,with a total of 11897 images.First of all,after the construction of the ?-OTDR sensing system is completed,the temporal-spatial data matrix is processed by a simple band-pass filter from the collected disturbance signals in the time and space domain,and then directly input to CNN,which automatically extracts the characteristics of the data and directly classifies the events by CNN,with an accuracy of 92.13%.As CNN works as a black box,gradient-weighted class activation mapping(Grad-CAM)method and T-distributed stochastic neighbor embedding(T-SNE)method are applied to visualize CNN's working process,which illustrates the correctness of feature extraction.In order to further improve the event classification ability,these 192-dimensional features extracted by CNN are input into the traditional classifier for further classification.The experimental results show that when classifying the features of these data,traditional classifiers can obtain higher classification accuracy than the softmax layer of CNN.In this study,the most suitable traditional classifier is support vector machine(SVM).finally,the CNN+SVM strategy performs the best and reaches 94.17 % accuracy,which offers 2.04%improvement compared with using CNN alone.