Research On Third-Party Intrusion Monitoring And Recognition Algorithm Based On Optical Fiber Distributed Acoustic Sensor System

Third-party intrusion monitoring is essential to guarantee the safety of the country's new infrastructure.Distributed Acoustic Sensing Technology(DAS)has the advantages of wide sensing coverage,high sensitivity,anti-harsh environment,and all-weather monitoring without blind spots,which makes DAS become the main technical means for third-party intrusion monitoring of infrastructure such as underground structure,rail transit,pipelines,etc.However,in the actual monitoring of the DAS system,the intensity fading phenomenon will lead to bad track data,the background noise will reduce the system's signal-to-noise ratio,meanwhile,the interference signal and the poorly robust recognition algorithm will decrease the accuracy of event recognition.These three problems lead to serious false positives and missed positives in the monitoring and identification of intrusion events,which limits the application of the DAS system in third-party intrusion monitoring.In view of the security monitoring requirements and practical limiting factors of the above-mentioned infrastructure projects,this thesis focuses on the third-party intrusion event recognition algorithm based on the optical fiber DAS system.Kernel algorithms such as the preprocessing of optical fiber sensor signal,location of intrusion event,multi-dimensional feature extraction,and the pattern recognition of intrusion event are proposed to realize accurate location and recognition of intrusion events.The main research work includes:(1)The preprocessing algorithm for optical fiber sensing signals in a complex environment is proposed.The static signal and suspicious signal are separated by the method of floor noise and maximum short-term window energy,and the recognition rate reaches 100%.The bad trace data are removed by the wave packet energy characteristic method,and the removal rate of interference signals can reach more than 98%.the interference of ripple noise is reduced by the empirical mode decomposition(EMD)denoising algorithm,and the signal-to-noise ratio of the sensing signal is increased nearly 3 times.Then,the intrusion event is located by the method of time-domain energy difference,and the experiment proves that the positioning error is less than 5 m.(2)A third-party intrusion event recognition algorithm based on optical fiber DAS technology are developed,including feature extraction and event recognition.The feature extraction part is composed of two feature vectors,one of which is an 84-dimensional feature vector composed of wavelet packet energy,feature frequency band,maximum short-time window energy,and background noise,and the other one of which is a feature vector composed of original time series signals.The back-propagation feedforward neural network(BP)model and the convolutional neural network(CNN)model are designed based on the 84-dimensional feature vector,while the one-dimensional convolutional neural network(1D-CNN)model is designed based on the original signal feature vector.(3)A third-party intrusion event recognition experimental test system based on the optical fiber DAS system is established,and the above three types of models are implemented and optimized using the Pytorch neural network framework.Furthermore,the recognition accuracy of different intrusion events is tested.In the end,the factors affecting the recognition rate are analyzed.Experimental results show that the event recognition accuracy of the BP neural network model,CNN neural network model and 1D-CNN neural network model reach82.47%,83.95% and 93.33%,respectively.