Font Size: a A A

Research On Brillouin Frequency Shift Extract Technology Based On Convolutional Neural Network

Posted on:2020-08-20Degree:MasterType:Thesis
Country:ChinaCandidate:Y Q ChangFull Text:PDF
GTID:2518306104993929Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Stress and temperature need to be efficiently measured for widespread applications such as industrial production and environmental monitoring.Distributed optical fiber sensing system based on Brillouin scattering technology can respond to stress and temperature simultaneously,and therefore have excellent application prospects and economic benefits.The extraction of Brillouin frequency shift is critical for the Brillouin sensing system,which directly determines whether the measurements,such as temperature and stress,can be accurately obtained.However,the traditional Brillouin extraction algorithm based on Lorentz curve fitting is time-consuming as an iterative algorithm.In addition,results will become inaccurate when the signal-to-noise ratio is low.Based on the research of distributed Brillouin time domain analyzer and machine learning algorithm,a Brillouin frequency shift extraction technique utilizing convolutional neural network has been proposed in this thesis.The algorithm treats Brillouin time-frequency signals as two-dimensional data,and directly obtains one-dimensional Brillouin frequency shift information from it.In this thesis,the application requirements of convolutional neural networks are analyzed in detail for Brillouin sensing system.In order to make the network suitable for 2D signals with different noise levels and complete the transformation to 1D,a network model consisting of multiple residual blocks and 1ŚN convolutional layer is constructed.Then the network is trained through a large number of simulation Brillouin scattering data with different signal-to-noise ratio,Brillouin frequency shift and linewidth,and is continuously adjusted according to the results to improve the performance.In the test,the network was comprehensively analyzed through 1224 sets of different simulation data,showing that the Brillouin frequency shift extraction accuracy of the network can reach or even exceed that of traditional curve fitting algorithm without reducing the spatial resolution.After that,25 km Brillouin time domain analyzer system was built,and ten sets of sensor data were used to verify the performance of the algorithm.The experimental results show that the algorithm can accurately extract the Brillouin frequency shift in 7s.For instance,the uncertainty of extracted Brillouin frequency shifts is as low as0.5 MHz when the pump pulse width is 30 ns and the signal-to-noise ratio is 10.5 d B.Compared with most other machine learning methods,the proposed algorithm can be directly used on any Brillouin system without retraining,thus promising and applicable for versatile situations.
Keywords/Search Tags:Optical fiber sensing, Brillouin frequency shift extraction, Convolutional neural network, Brillouin scattering
PDF Full Text Request
Related items