Research On ?-OTDR Vibration Pattern Recognition Based On Image Processing

Distribute optical fiber vibration sensing system based on phase sensitive optical time domain reflectometer(?-OTDR)has many advantages,such as low transmission loss,high sensitivity,long detectable distance,strong anti-interference capability and multi-point positioning,so it's widely used in areas such as intrusion monitoring,oil pipeline monitoring and railway monitoring.Due to the complexity of environment,the system's high FNR and FPR has always been a difficult problem to be solved.This dissertation uses image processing methods to develop theoretical and experimental research about signal denoising and vibration pattern recognition of?-OTDR system.The main contents completed are as follows:(1)Study the theory of ?-OTDR vibration sensing system and establish mathematical model of interference between Rayleigh backscattered light.Then make a experimental plan according to actual experimental conditions,build experimental system and complete data collection.Finally,complete the work about data pre-processing,and the initial SNR of the positioning curve is 6.43 d B.(2)Use image anisotropic diffusion filtering(PM)algorithm and wavelet packet energy-threshold denoising algorithm for signal denoising.Without affecting positioning accuracy,the SNR improvement of PM algorithm reaches 11.60 d B,but the time-intensity signal of the vibration position is seriously affected.And the SNR improvement of the wavelet packet energy-threshold algorithm can reach 15.17 d B,simultaneously the shape of the original signal is well preserved,but the algorithm takes a longer time.(3)Convert the time-intensity signal of the vibration position into a time-frequency distribution image,and do image pre-processing to obtain binary images.Then extract the morphological features of the binary images as eigenvalues for classification,try to train a naive Bayes classifier.The Five-fold cross-validation method is used to improve the generalization ability of the model.Finally,the classification accuracy on the five types of samples is 92.67%,the relative error of the model is 2.64%.In order to avoid the complicated feature extraction process,the convolutional neural network(CNN)is applied to image-classification.By adjusting structure and optimizing parameters,the classification accuracy on the original grayscale images and the binary images are 86.00% and 94.67%,with a relative error of 4.90% and 3.40% respectively.