Research On DAS Signal Identification Method In Communication Optical Cable Safety Monitoring

In recent years,with the development of ?-OTDR technology,the detection of acoustic signal based on distributed optical fiber acoustic sensor system(DAS)is more and more important.How to effectively extract the characteristics of Das signal for correct and stable identification has always been the biggest problem of Das in the application of communication cable security monitoring.In this paper,the current situation of Das based communication cable security monitoring at home and abroad is investigated.It is found that the methods used in Das signal recognition only consider the artificial features or deep learning features of the signal,and ignore the two methods which can extract different levels of features,and do not make full use of all the information of the signal,It can achieve a better effect than using a certain kind of feature alone.Inspired by this idea,thesis proposes a method to recognize Das signal based on artificial features and deep learning features:(1)The mainstream Das signal recognition mainly uses artificial features or deep learning features now.Using artificial features alone seriously depends on expert knowledge and contains insufficient information.However,using depth features alone has serious over fitting problems and the deep learning method is not reliable enough.On this basis,thesis proposes a method based on the joint use of artificial features and depth features.The method makes full use of the effective information of Das signal.(2)Firstly,under the condition of CNN network accuracy of 90.9%,the parameters of the network before the full connection layer are extracted,and then the artificial features are extracted.The two kinds of features are spliced together to form a highdimensional feature vector,which can reflect the features of Das signal from multiple angles.Too many features may lead to information redundancy.Therefore,thesis adopts the method of feature selection Each extracted feature is scored by the weighted fusion of F value and maximum mutual information coefficient,and the important features are screened out.The experiment shows that the first 110 dimensional features can effectively recognize Das signals,and the classification speed is very fast(10ms level)by XGB(extreme gradient promotion),and the timeliness can meet the requirements of the system and the project,and the average accuracy can reach 95%.(3)Based on the combination of artificial features and deep learning features,thesis also proposes another scheme,which is based on the wide and proposed by Google Deep model is designed to realize the recognition of Das signal.The artificial feature and deep learning feature are mixed by network.The artificial feature module extracts the artificial features of the signal offline through the database,only updates the weight parameters,greatly reduces the training time,and the deep feature module extracts the deep learning characteristics of the signal,and iterations the network through the network Finally,through the full connection layer,the type of event is identified,the appropriate network structure parameters and loss functions are selected.The recognition time is 10 ms,and the final accuracy rate is 98.3%,which is fast and good.It provides an effective way to solve the biggest problem of Das in the application of communication cable safety monitoring.