| Compared with other distributed optical fiber sensing systems,the distributed sensing system based on phase-sensitive optical time domain reflectometer(Phi-OTDR)has the advantages of high sensitivity,simultaneous multi-point positioning,antielectromagnetic interference,relatively low cost,accurate detection,etc.It has broad application prospects in long-distance security monitoring such as around airports,military bases,national borders,and oil and gas pipelines.However,in practical applications,environmental noise or non-invasive disturbances lead to false alarms in the system.Moreover,the problem of accurate identification of events at multiple locations is also a key issue that the system needs to break through when multiple disturbance events occur simultaneously.In this thesis,the study of pattern recognition methods is carried out for the single-point and multi-point disturbance events of the Phi-OTDR distributed sensing system.Through the use of multi-dimensional features,and two rounds of feature adaptive selection,combined with Light Gradient Boosting Machine(LightGBM)to perform pattern recognition of Phi-OTDR sensing system disturbance events,and for the occurrence of multi-point disturbance that may occur,the solutions for three possible situations of multi-point disturbance are given.The main research work completed in this thesis is as follows:(1)According to the principle of Phi-OTDR distributed optical fiber sensing system,an experimental system platform is setup.Pattern recognition is carried out for five disturbance events in the actual scene,namely watering,climbing,knocking,pressing and false disturbance(environmental noise).The collected sensing signals were preprocessed by data difference,group and normalization.According to the characteristics of signals,a total of 40 features are extracted from the time domain,wavelet domain,time-frequency domain after empirical mode decomposition and speech-like signal characteristics of original normalized signal and differential normalized signal.In order to prevent redundancy of features,two hierarchical feature screenings are adopted,and then 14 obvious features are obtained by optimization,which maximizes signal utilization from the characteristics of signals.After two screenings,three methods were compared and analyzed to verify the feasibility of the screening method.(2)According to the actual demand and the characteristics of the classifier itself,an event recognition algorithm based on light gradient boosting Machine classifier is proposed.The parameters of accuracy rate,recall rate and F1 score of this method are tested through the remaining samples.The accuracy rate of watering,knocking,climbing,pressing and false disturbance events is 98.38%,98.13%,96.02%,94.86% and 99.70%respectively.The F1 scores are 97.78%,98.12%,95.62%,96.35% and 99.17%.Thus the average accuracy rate and average F1 score both reach 97.41%.The recognition time of the algorithm after feature optimization is about 0.3s,which ensures good real-time performance.(3)In view of the multi-point intrusion events that may occur in practical applications,a variety of recognition schemes are studied.Aiming at small-scale recognition,a multi-point pattern recognition scheme based on multi-point classifier is proposed.Support vector machine,gradient boosting tree and LightGBM single comprehensive classifier are introduced for multi-point pattern recognition for large-scale recognition.The algorithm based on LightGBM performs best with an average accuracy rate of 96.76%,a recall rate and an F1 score of 96.75%.A multi-point recognition algorithm based on multi-tree is proposed for fuzzy positioning data,which can accurately identify the location and events.The average recognition rates of five events are 95.90%,97.08%,92.72%,94.32% and 99.43% respectively,and thus the total average recognition rate is 95.89%.The feasibility of the multi-point recognition scheme is verified by experimental results. |
PDF Full Text Request