Research Of Optical Fiber Vibration Detection And Classification Based On ?-OTDR

Distributed optical fiber sensing technology based on ?-OTDR has the advantages of simultaneous multi-point positioning,long-distance detection range and high positioning accuracy.It has huge application prospects and advantages in the field of perimeter security,but the high rate of false positives and real-time requirement are problems that need to be solved urgently.This paper takes the perimeter security system based on ?-OTDR as the research background,starts from the actual problems of real-time and false alarm rate,takes deep learning as the core,and deeply researches fiber vibration detection and classification technology.The main research contents are as follows:1.A distributed optical fiber vibration monitoring experimental platform based on the ?-OTDR system is constructed,the characteristics of ?-OTDR data processing are systematically analyzed,and the GPU-based parallel computing method is used to accelerate the data processing process.At the same time,a multi-level vibration detection model based on RNN neural network is proposed,which ensures the accuracy and improves the real-time performance of the system.2.The fiber vibration classification method is deeply studied in this paper,and a time attention mechanism based on biLSTM is proposed.Firstly,a data set of fiber vibration time-frequency graphs is collected and constructed,and then the classification effects of ANN,RNN,and CNNs on fiber vibration are studied.On this basis,5 CRNN networks with different convolutional layers are designed and constructed.After a lot of experiments,based on the characteristics of fiber vibration time-frequency diagram,CNN and RNN,a biLSTM-based time attention mechanism is proposed,which effectively improves the accuracy rate of the fiber vibration classification model.3.An excellent model has the problems of a large number of parameters and a huge amount of computations.Based on the deep separable convolution and DenseNet network,a compact student network model is designed and built.Furthermore,the classification ability of the large network model is refined into the student model through knowledge distillation,which ensures a small loss of accuracy,and realizes the lightweight of the fiber vibration classification model.