Temperature Extraction From BOTDA Sensor Via Machine Learning

Distributed optical fiber sensing technology that utilizes optical fiber as a sensor and transmission medium has been widely used in structural health monitoring,national defense,and other fields due to its advantages of long sensing distance,high resolution,and low cost.Among them,the Brillouin optical time-domain analysis technology(BOTDA)based on stimulated Brillouin scattering has developed rapidly in recent years.With the increase in sensing distance and spatial resolution,the corresponding demodulation technology of sensing information has gradually become a hot topic in the optical fiber sensing field.At present,the extraction of sensing information mainly adopts the curve fitting method.The contour of the Brillouin Gain Spectrum(BGS)is restored through the curve fitting from discrete points,and the corresponding frequency of the peak gain is extracted,then converted into temperature information.Although these methods have good accuracy,they are often highly timeconsuming and have implicit degradation under large frequency scanning steps.In this thesis,echo state network(ESN),kernel extreme learning machine(KELM),and other machine learning methods are proposed to extract the temperature information.Based on their datadriven characteristics and great nonlinear approximation ability,they are utilized to establish a direct mapping from BGS to temperature to meet the real-time,accuracy,and generalization requirements of temperature extraction.Firstly,an experimental platform of the BOTDA system is built,and the sensing data at different temperatures,signal-to-noise ratios(SNRs),pump pulse widths,and frequency sampling steps are collected.Based on the pseudo-Voigt curve,the dataset is augmented by simulation.Then,the echo state network is adopted to extract the BOTDA temperature to explore the performance of the time sequence network in temperature extraction.The optimal parameters and structure of the network are acquired through experiments.The experimental results show that the echo state network can effectively extract the BOTDA sensing data.It has better generalization for BGSs with different pulse widths,and its extraction speed is faster than curve fitting methods by an order of magnitude.To further improve the performance of the algorithm,Extreme Learning Machine(ELM)and Kernel Extreme Learning Machine(KELM)are proposed in this thesis for BOTDA temperature extraction.By ELM and KELM,the raw BGSs are directly divided into different temperature classes.The performance of the proposed methods is investigated both in simulation and experiment under different cases of SNRs,pulse widths,and frequency scanning steps.The results show that KELM has better stability than ELM,with a standard deviation of 0.5°C,which equals only about 30% of that of ELM.Compared with Lorentz curve fitting and pseudo-Voigt curve fitting methods,KELM always keeps good accuracy,and the speed is improved by two orders of magnitude.Meanwhile,the proposed KLEM-p model with zero padding achieves higher accuracy than curve fitting methods in the case of large frequency scanning steps.