Investigation Of Disturbance Identification Of Distributed Optical Fiber Sensing System Based On Phase-sensitive OTDR

The fiber-optic distributed sensor based on phase-sensitive optical time domain reflectometer(?-OTDR)is widely used in oil and gas pipeline security,perimeter security and other security fields because of its simple structure,low energy consumption,anti-electromagnetic interference,high accuracy,long sensing distance and other advantages.But in practical applications,false alarms would occur because of environmental noise and harmless human activities.In order to reduce the false alarms in the sensing system,the support vector machine and neural network are introduced in this thesis to identify whether the alarming signal is true alarm or false alarm.Experimental results show that all kinds of disturbance events in the system can be efficiently distinguished through the two algorithms and nuisance alarm rate can be lowered.The main achievements of this dissertation are as follows:(1)The theory of ?-OTDR system is introduced.According to the simulation signal of ?-OTDR system,the principle of sensing system is illustrated.Oriented to the common disturbance signals in the actual environment,the simulation experiment is set up.There are five kinds of disturbance signals,which are watering,knocking,climbing,pressing and background noise.The five kinds of disturbance signals obtained are processed by normalization,data division,difference processing and wavelet packet decomposition.After the signal processing,40 features are extracted in time domain and frequency domain,and the effect of features is analyzed.(2)A new support vector machine(SVM)multi classification method named by near category support vector machine(NC-SVM)is proposed and compared with the commonly used one-versus-one(1-v-1)SVM multi classification method.The one-versus-one SVM multi classification method is tested and the identification precision rates of five disturbance events(watering,knocking,climbing,pressing and background noise)are 94.36%,94.57%,90.36%,94.26%,97.35%,respectively.The NC-SVM multi classification method is tested and the identification precision rates of five disturbance events(watering,knocking,climbing,pressing and background noise)are 94.63%,95.82%,91.74%,94.13%,99.82%,respectively.The average identification precision rate is 95.23%,and the nuisance alarm rate is 4.77%.Compared with 1-v-1 multi classification method,the average identification precision rate is obviously improved.(3)A disturbance events identification method combining fisher feature selection method and extreme learning machine algorithm is proposed and tested.The experimental results show that the five disturbance events can be effectively distinguished through this method.The identification precision rates of five disturbance events(watering,knocking,climbing,pressing and background noise)are 92.74%,96.72%,93.77%,93.85%,99.61%,respectively.The average identification precision rate is 95.34%,and the nuisance alarm rate is 4.66%.Compared with BP(back propagation)neural network and generalized regression neural network,which are used in the experiment,the proposed method has more advantages in identification accuracy and real-time.