Research On Data Processing Of Brillouin Optical Fiber Sensing System Based On Neural Network

Since the 21 st century,the field of optical fiber and communication has developed rapidly,along with the gradual popularization of the Internet industry and terminals,distributed optical fiber sensing system based on optical fiber sensing has sprung up rapidly.With the development of the Internet,it is gradually entering into millions of households and putting into daily use.Optical fiber sensing system based on stimulated Brillouin Optical Time Domain Analysis(BOTDA)carries optical signal transmission information.The whole system uses optical fiber as the transmission medium.When measuring,the sensing optical fiber is placed on the surface or inside of the object to be measured.Through data acquisition and processing,the temperature and strain information of the object to be measured on the optical fiber can be obtained.It has the advantages of full-scale continuity and long-distance measurement.It has a wide range of applications,such as measuring civil engineering(bridges,tunnels)as well as highways oil pipelines and other fields.It can achieve a maximum of hundreds of kilometers in the length of the fully distributed measurement.With the maturity of the modern Internet and data processing technology,the accuracy of temperature and strain can be further improved by using machine learning technology on the basis of traditional measurement.Firstly,on the basis of data acquisition and building the BOTDA system in the laboratory,principles of Brillouin scattering and the system are analyzed.The acquisition of data acquisition card(NI PCI-5114 data acquisition card)is controlled by LabView software,and the temperature information of each point on the optical fiber is obtained.Secondly,different Brillouin scattering spectra are collected in different temperature environments.Based on non-Brillouin frequency shift fitting method,different neural networks are used to extract temperature characteristics directly.The optical fibers were heated in a constant temperature water bath pot from 16.5? to 80?and Brillouin scattering spectra were collected.The Brillouin scattering spectra signals and corresponding temperatures at different temperatures were trained as highdimensional training sets,and the temperature characteristic model was obtained.The results show that the RMSE of the RBF neural network method is less than the traditional method.With the increase of step frequency,the RMSE increases slowly,and reduces the calculation steps,and improves the convergence to a certain extent.Thirdly,the 2D image which is the overhead view of BOTDA 3D image is used to detect mutation points by using object detection technology.Through feature extraction at the image level,the location information of mutation points can be obtained directly.By training the range of the mutation points of the image and labeling,the model can directly predict the mutation points of the new BOTDA overhead image.Fourthly,based on the analysis of the experiment of extracting temperature information by a neural network,two sweeping strategies based on Gauss centralization and variance value weighting method are proposed on the basis of uniform interval sweeping strategy.In the same environment,281 points were obtained by the above three methods.Each method was trained 10 times and the MSE was tested by the same neural network structure.The results show that the variance value weighting method is better than the Gauss centralization and uniform interval sweeping method.Based on machine learning and neural network,the temperature extraction process of Brillouin scattering spectra is discussed and analyzed in different stages.Some new methods for the temperature extraction process and the improvement of temperature measurement accuracy are proposed.