Research On Pattern Recognition Of Disturbance Events In ?-OTDR Distributed Optical Fiber Sensing System

Phase-Sensitive Optical Time Domain Reflectometer(?-OTDR)has the advantages of high sensitivity,high spatial resolution,long monitoring distance and multi-point positioning,and thus it is widely used in many fields such as perimeter security and pipeline monitoring.However,the long distance safety monitoring system based on ?-OTDR has many problems such as high false alarm rate and poor real-time performance.By combining theory with experiment,this thesis aims to improve the recognition accuracy of disturbance signal,reduce the false alarm rate,and improve the real-time performance of signal recognition for the ?-OTDR system.Boosting ensemble learning algorithm represented by e Xtreme Gradient Boosting(XGBoost),and Convolutional Neural Network(CNN)and Long Short-term Memory Network(LSTM)as representative research on deep learning algorithms are selected for studying the recognition of disturbance signals in the ?-OTDR system.The research work has great significance and practical value for improving the performance of the ?-OTDR system.The main tasks completed in this thesis are described as follows:(1)The principle of the ?-OTDR system is analyzed and an experimental platform for the distributed optical fiber sensing system is setup.Five types of actual disturbance signals,such as watering,knocking,climbing,pressing,and background noise,are collected.Normalization processing,Empirical Mode Decomposition(EMD),data enhancement,and feature extraction are performed on the collected data,which laid the foundation for subsequent recognition work.(2)A disturbance event recognition method based on boosting ensemble learning algorithm is proposed.Three classifier models of adaptive boosting algorithm(Adaptive boosting,Adaboost),gradient boosting decision tree(GBDT)and XGBoost,are constructed.Six IMF components and one Res component are obtained by EMD.The energy features of each component are extracted,and these three algorithms are used for identification.The recognition results based on XGBoost are the best among three algorithms.The accuracy rates of the five events of watering,knocking,climbing,pressing,and background noise are 97.96%,95.90%,91.10%,94.84% and 99.69%,respectively.The average accuracy rate reaches 95.90%,the false alarm rate is 4.10%,and the recognition time is 0.28 s.(3)A method for identifying disturbance events based on deep neural networks is proposed.Four deep neural network models,including One-Dimensional Convolutional Neural Networks(1D CNN),Single-branch Long and Short-term Memory networks(SLSTM),Multi-branch Long and Short-term Memory networks(MLSTM)and Multi-branch Long and Short-term memory Convolutional Neural Networks(MLSTM-CNN),are constructed.Experimental results show that the recognition effect based on the MLSTM-CNN model is better than the other three models.The accuracy rates of the five events of watering,knocking,climbing,rolling,and background noise are 94.38%,94.40%,94.76%,95.25% and 99.71%,respectively.The average accuracy rate reaches 95.70%,the false alarm rate is 4.30%,and the recognition time is 0.27 s.