Research On Event Identification Method Of Communication Optical Cable Safety Monitoring Based On DAS

The safety of communication optical cable is related to economic development,social livelihood and other fields.With the development of distributed optical fiber acoustic sensor system(DAS)based on Φ-OTDR technology in the field of long-distance monitoring,researchers began to apply DAS to the safety monitoring of communication optical cable.But,the event signals collected based on DAS still face many problems in current type recognition.In this background,this thesis investigates the current situation of communication cable safety monitoring based on DAS at home and abroad,and finds that the current mainstream signal recognition methods only consider the time information of signal,ignoring the characteristics that DAS can collect the spatial distribution information of signal.At the same time,some related literature points out that the signal-to-noise ratio of optical fiber sensing signal can be increased many times without loss of spatial resolution by using spatial information.Inspired by this idea,this thesis proposes a CNN-BiLSTM spatiotemporal signal processing model based on deep learning.The specific work of this thesis is as follows:(1)Based on the investigation and analysis of the current research status of communication cable safety monitoring,it is concluded that DAS has the characteristics of long monitoring distance,low cost and high sensitivity,which is very suitable for communication cable safety monitoring.However,there is no consideration of signal spatial information in the identification method of DAS signal.On this basis,this thesis proposes CNN-BiLSTM spatiotemporal signal processing method.(2)Five kinds of typical event signals are collected by DAS,and one-dimensional time signal data set and two-dimensional space-time signal data set are established.According to the field data set,the time local feature extraction module of CNN-BiLSTM model is designed and implemented in turn.The module uses multiple parallel 1D-CNN to extract features from time series of each spatial point;the second is the spatial distribution feature extraction module.Based on the time features,the module uses BiLSTM network to extract the distribution features of each spatial point;the second is the full connection layer recognition module.Finally,the training and testing of the whole model are completed by selecting the appropriate loss function and optimization method.The test results show that the recognition rate of CNN-BiLSTM space-time model proposed in this thesis can reach 97.8%.(3)According to the relevant literature,one-dimensional time signal model Xgb,two-dimensional space-time signal model 2D-CNN and BiLSTM are designed as reference models,and the performance of each model is trained and evaluated under the field data set established in this thesis.By comparing the performance of CNN-BiLSTM model and each reference model in classification and evaluation index,feature extraction ability,model stability,recognition speed,it is concluded that the performance of two-dimensional spatiotemporal signal model is better than that of one-dimensional time signal model,and the CNN-BiLSTM model proposed in this thesis is better than the other two-dimensional spatiotemporal signal models,and it is also feasible in timeliness meet the actual application requirements.