Research On Pattern Recognition Algorithm Of Disturbance Signal In Optical Fiber Sensing

The perimeter security system based on distributed optical fiber sensing can sensitively monitor environmental vibrations and judge intrusion events through optical fiber sensing signals.In the actual application environment,the optical signal pattern is complex and there is noise interference.It is necessary to design a pattern recognition algorithm to classify the fiber disturbance signal and issue an intrusion warning.How to design a signal feature extraction method that is more discriminative and easier to learn,and choose a suitable and convenient algorithm to design a classifier model to achieve more accurate recognition accuracy are research hotspots and problems in recent years.This thesis focuses on the research of multi-domain feature extraction method and classifier design of optical signal pattern recognition,mainly doing the following work:First,for the distributed optical fiber sensing system,this article compares the principles,advantages and disadvantages of three different interference sensing structures.In this thesis,a dual Mach-Zehnder interference fiber sensing system is built to achieve sensitive sensing of disturbance signals.Secondly,for the non-stationary and non-stationary characteristics of the disturbed signal,a two-stage optical fiber recognition scheme designed by feature extraction +classifier is proposed.The time domain feature extraction method includes short-term energy of the signal and short-term level crossing rate,three dimensionless indicators of waveform factor,margin factor and kurtosis,as well as wavelet packet decomposition entropy features in the wavelet domain,empirical mode decomposition obtains kurtosis features,forming a joint feature vector.This thesis implements support vector machine classification network to classify and identify feature vectors.Then,in view of the limitations of the artificial design feature extraction method,deep learning algorithms are used to improve the recognition accuracy,and two endto-end classification and recognition network architectures are proposed and implemented.First,a one-dimensional convolutional neural network(1D-CNN)model is proposed,which can achieve automatic feature extraction and higher accuracy.Afterwards,based on 1D-CNN,the idea of autoencoder structure and denoising layer is combined to improve it,a one-dimensional convolution stacked denoising autoencoder(1D-CNN-SDAE)model is proposed.Finally,the three signal recognition algorithms proposed in this thesis are verified in a dual M-Z interference fiber sensing system.The fiber optic sensor signals of the four intrusion events of climbing,tapping,flapping and impacting and two nonintrusive events of still and rain are collected as the data set.After experimental verification,the accuracy of the SVM classification algorithm based on time domain features combined with wavelet packet entropy features reaches 97.1%;the 1D-CNN end-to-end network based on deep learning can adaptively extract signal features and improve accuracy up to 99.9%;1D-CNN-SDAE realizes the automatic extraction of signal feature by combining one-dimensional convolutional layer and codec structure,and improves the anti-noise performance of the network,with good robustness and adaptability.