Pattern Recognition Of Phi-OTDR Distributed Optical Fiber Disturbance Sensing System Based On Probabilistic Neural Network Classifier

Fiber-optic distributed disturbance sensors(FDDS),with the advantages such as having capable anti-electromagnetic ability,no need for external power supply,high sensitivity,wide detection range as well as convenience of system integration,have become one of the hotspots in the field of sensors.Nowadays,the FDDS is widely used in pipelines monitoring,flaw detection of large structures,territorial security monitoring and many other fields.The sensing system based on phase-sensitive optical time domain reflectometer(?-OTDR)has become one of the mainstream research directions in the research of long-distance distributed optical fiber sensing system due to the high spatial resolution,simple structure and the ability of detecting multipoint disturbances simultaneously.?-OTDR sensor system exposes the problem of high false alarm rate in practical application.In order to solve the issue,a pattern recognition method based on probabilistic neural network(PNN)is proposed in the thesis,which is proved can identify different types of disturbance events effectively.The main achievements of this dissertation are as follows:(1)On the basis of theoretical research on the principle of ?-OTDR distributed optical fiber sensing system and the characteristics of the output signal,an experimental setup based on ?-OTDR distributed optical fiber sensing system is set.Four kinds of disturbance signals,including watering,climbing,knocking and pressing,as well as the output signals without disturbance are collected.Features of time domain signals and differences of the time domain signals are extracted.The signals are divided into samples for further analysis,which establish the foundation for the following research on identification of disturbance events.(2)A disturbance event recognition method based on probabilistic neural network is proposed and validated by experiment with sample data.The average identification rates for five event types(watering,climbing,knocking,pressing and no disturbance)are 97.57%,95.68%,99.92%,99.08%and 99.97%,respectively.The algorithm can identify different events effectively with the average time of 1.1369s for the model establishment and recognition.The problem of low real-time is exposed.(3)According to the process of probabilistic neural network,two improvements by using mean impact value and by using principal component analysis are proposed to improve the real-time performance of the algorithm on the premise of guaranteeing the accuracy of recognition.The two improvements are validated by experiments and the average identification rates reached 93.36%,92.48%,97.01%,96.99%,99.60%and 96.80%,94.13%,99.36%,98.45%,99.95%,respectively.Compared with the original model,0.2624s and 0.2061s are shortened by the two improvements,respectively.(4)Based on the sample data,a method to construct a sample library based on the"deletion-replacement" principle is proposed.The sample libraries with the proportion of 70%,60%,50%and 40%of the total samples for the two improvements are established and tested by experiments.The experiment results indicate that the identification rates can still achieve more than 90%by using the sample libraries with the proportion of 50%.