| Brillouin optical time domain reflectometer(BOTDR)distributed optical fiber sensing system has the advantages of simple structure and single-ended measurement.BOTDR can realize long-range and simultaneous measurement of temperature and strain,and it has shown the unique advantages in the field of health monitoring and fault diagnosis for large-scale structures and equipments,and has been more and more concerned and researched as well.However,due to the mutual restriction of BOTDR sensing system performance indexes,it still can not meet the requirements of higher measurement accuracy and shorter measurement time in many applications.Aiming at optimizing the performance of BOTDR sensing system,the characteristics of BOTDR are in-depth studied in this dissertation.Based on analyzing the influencing factors of signal-to-noise ratio,a method for performance enhancement of BOTDR sensing system is proposed to achieve signal-to-noise ratio improvement and high-accuracy measurement.According to the characteristics of Brillouin scattering spectra,the methods for feature information extraction using artificial neural network,digital image edge detection and convolution neural network are proposed,which can ensure the measurement accuracy and effectively shorten the measurement time of BOTDR sensing system.The main research works are concluded as follows:(1)A method for optimizing the signal-to-noise ratio of BOTDR sensing system is proposed.Firstly,the principles of local-heterodyne detection and self-heterodyne detection are studied,and the influencing factors of signal-to-noise ratio are analyzed.Secondly,a method combined with multi-frequency probe light and frequency shift averaging for enhancing signal-to-noise ratio of BOTDR sensing system is proposed to improve the signal intensity and reduce the coherent Rayleigh noise as well.Finally,the local-heterodyne detection and self-heterodyne detection BOTDR temperature sensing system are constructed for verification.The experiment results have shown that the amplitude fluctuations of Brillouin signals and spectral parameters are decreased.The signal-to-noise ratio and measurement accuracy of BOTDR sensing system are improved by using the three-frequency probe light with shift-averaging.The signal-to-noise ratio of local heterodyne detection system and self-heterodyne detection system are improved by 8.15 dB and 7.92 dB,and the best temperature measurement accuracies are 0.34 ? and 0.36 ?,respectively.(2)An optimized neural network training method for temperature extraction of BOTDR sensing system is proposed.Firstly,the principle of temperature extraction by using artificial neural network is studied,and the difference between the data of training set and test set is analyzed.Secondly,by estimating the noise level of BOTDR sensing system,the method of training the artificial neural network with noisy data is proposed,and the different types of training set data are simulated to train the networks.Finally,the BOTDR temperature sensing system is constructed and the temperature extraction results of three neural networks obtained by using different training sets are compared.The influences of frequency scanning interval and signal-to-noise ratio on the temperature extraction results are analyzed.The experiment results have verified that adding a certain amount of noise to the ideal Brillouin scattering spectra data can effectively enhance the generalization ability and adaptability of artificial neural network,and then improve the measurement accuracy of BOTDR sensing system.(3)A method for temperature extraction of BOTDR sensing system based on extreme learning machine(ELM)network is proposed.The problem that parameters of traditional neural networks are needed to be set artificially is studied firstly,and the characteristics of ELM are also analyzed.Then,a method for extracting temperature of BOTDR sensing system based on ELM network is proposed,and the training set data is simulated to train the network.After constructing the BOTDR sensing system,the temperature extraction results obtained by using curve fitting method and ELM network are compared.The experiment results have confirmed that ELM network has higher accuracy and better tolerance of measurement error even in the case of large frequency scanning interval.The best temperature measurement accuracy is 0.21?.Moreover,ELM can reduce the measurement time significantly compared to curve fitting method.It takes only 3.9812 s to process 9200 Brillouin scattering data when the frequency scanning interval is 16 MHz.(4)An edge feature extraction method of Brillouin scattering spectral image based on second-order edge detection operator is proposed.At first,the characteristics of two-dimensional Brillouin scattering spectral image different from one-dimensional Brillouin scattering spectra are studied,and the feasibility of extracting features from Brillouin scattering spectral image by edge detection is analyzed.Then,the second-order Laplacian edge detection operator is used to extract the Brillouin frequency shift,which is the roof-like edge in Brillouin scattering spectral image.Finally,the accuracy and time of Brillouin frequency shift feature extraction by second-order Laplacian edge detection operator are analyzed.The experiment results have indicated that the second-order Laplacian edge detection operator is feasible and effective to extract Brillouin frequency shift features from Brillouin scattering spectral images.The extraction accuracy is better than that obtained by using the first-order Sobel edge detection operator,and the extraction time is shorter than that consumed by using the curve fitting method.(5)A convolution neural network(CNN)-based method for simultaneous measurement of temperature and strain is proposed.Firstly,the method of simultaneous measurement of temperature and strain using large effective area fiber with multiple Brillouin scattering peaks is studied,and the existing problems are analyzed.Then,based on the study of CNN,a method for simultaneous measurement of temperature and strain by using CNN from multi-peak Brillouin scattering spectral image is proposed.The CNN training set data is constructed,and the CNN is designed.Finally,the CNN is trained and tested.Compared to the time-consuming solving equation method,CNN can extract the temperature and strain information from the multi-peak Brillouin scattering spectral image,reduce the measurement time and improve the measurement accuracy of BOTDR sensing system effectively. |
PDF Full Text Request