Research On Neural Network Based Information Processing Technology Of BOTDA Sensors

Distributed optical fiber sensors have attracted the spotlight due to their characteristics of distributed real-time monitoring,low cost,corrosion resistance and electromagnetic immunity.Among them,BOTDA sensors have the advantages of long-distance sensing,dual-parameter monitoring,and high measurement accuracy.It has good application prospects in the fields of infrastructure structure health monitoring and fault diagnosis.However,with the increasing requirements for BOTDA dynamic sensing range,measurement accuracy,measurement time and so on,how to obtain sensing information efficiently and accurately becomes critical.Therefore,in order to achieve better performance of BOTDA sensor information processing,performance optimization scheme based on neural network is proposed to improve the system SNR and shorten the system measurement time.The main research contents are as follows:(1)In order to extend the application of deep convolutional neural network in BOTDA image denoising,ADNet and BRDNet based on supervised learning are introduced to denoise the Brillouin gain spectrum,with classic Dn CNN denoising algorithm as baseline,the denoising performance of the three networks is compared in four aspects: signal-tonoise ratio(SNR)improvement,Brillouin frequency shift measurement error,full width at half maximum measurement error,and denoising time.And in order to study the impact of different data sets on the performance of the denoising network,three BGS simulation data sets,Single-A,Random-B,and Segment-C,were constructed for network training and testing.And for the problem whether the input data of the network need to be normalized by maximum and minimum value,and the denoising performance of the training and test data before and after the min-max normalization is compared.Experiments show that after the input data of the network is normalized by the min-max value,the denoising performance of the three networks declines,and for the three data sets used for network training,the data of Single-A is simpler and the denoising effect is better.Then,among the three denoising networks,the denoising performance of Dn CNN and BRDNet is similar,and ADNet is slightly worse.(2)In order to realize the noise modeling of the BOTDA,and to generate a noise distribution that approximates the real data to form {clean,noisy} image data pairs,and then to use the supervised image denoising technology to improve the SNR of the Brillouin gain spectrum,an unsupervised blind denoising technique based on SCGAN to model the noise of Brillouin gain spectral data is put forward.At last,the noise map obtained by SCGAN and the ideal data are used to synthesize training data pairs for supervised denoising based Dn CNN,and then the denoising learning was carried out.In this paper,the denoising performance of the proposed network is verified by simulation experiments,and the experimental results show that denoising using the data synthesized by SCGAN can improve the SNR by about 7d B.(3)In order to shorten the temperature extraction time of BOTDA,extreme learning machine(ELM)was considered as the model.In order to improve the defect that the ELM cannot update the newly acquired training data online,an extreme learning machine(OSELM)technology based on online sequence learning is proposed.And in order to solve the problem of high space and time complexity in the training of kernel ELM(KELM),a reduced kernel ELM based on online sequence learning(OS-RKELM)is proposed.The performance of classical ELM and KELM networks,improved OS-ELM and OSRKELM,and traditional Lorentzian fitting algorithms are compared in terms of temperature measurement error,training and testing time in both simulations and experiments.in simulation and experiment.The results show that the training time of OS-ELM is 1.16%of ELM,the test time is 16.7% of ELM,and the temperature measurement errors of them are similar;and the training time of OS-RKELM is 28.6% of KELM,the test time is 25%of KELM,the OS-RKELM measurement error is slightly larger than KELM,but better than ELM and OS-ELM;and LCF has the highest measurement accuracy,but the fitting time is longer than the other four.