Design And Research Of Fast Brillouin Frequency Shift Extraction Algorithm

The sensing system based on Brillouin scatter is of great significance in long-distance temperature and strain monitoring,and is widely used in safety monitoring,with good economic benefits and application prospects.As the core problem of Brillouin sensing system,Brillouin frequency shift extraction often determines the accuracy of system,stability and response time.At the same time,the universality and robustness of the Brillouin frequency domain extraction algorithm are also crucial.In the past reported research work,most of the research only considered the accuracy and processing time of Brillouin frequency shift extraction algorithm,and its scheme is not universal for different Brillouin sensing systems and different sensing environments.Therefore,this paper will comprehensively consider the characteristics of algorithm processing accuracy,stability,response time,general adaptation and robustness,and conduct in-depth research on the real-time Brillouin frequency shift extraction problem.Existing Brillouin frequency shift extraction algorithms have advantages and disadvantages.Traditional curve fitting algorithms,such as Levenberg-Marquardt(LM)algorithm,have strong universality,but have the disadvantages of long processing time,accuracy sensitive to noise,and difficult to apply in real time.Although the crosscorrelation algorithm proposed by the researchers is less sensitive to noise,its accuracy is limited by the frequency sampling intervals.Although the Brillouin frequency shift extraction algorithm based on fully connected neural network has a short processing time,its accuracy is related to the starting position of the frequency sweep,which is relatively unstable.In view of the shortcomings of the above algorithm,the main research content and innovation results are mainly as follows.1.On the basis of the theory of traditional cross-correlation,a fast cross-correlation algorithm is proposed,which solves the defect that the accuracy of the traditional cross-correlation algorithm is limited by the frequency sweep intervals.The principle implementation of fast cross-correlation algorithm is derived through the theory of error minimization,and the performance of fast crosscorrelation algorithm is analyzed from the simulation level,which proves that when the processing time of the fast cross-correlation algorithm is reduced by three orders of magnitude,the accuracy of the fast cross-correlation algorithm is comparable with the LM algorithm and better than the interpolation cross-correlation algorithm,and it is pointed out that the performance improvement of the fast cross-correlation algorithm brings about the consumption of storage space and the decrease of the stable width ratio of the frequency shift extraction results.2.Based on the correlation between the accuracy of Brillouin shift extraction algorithm based on fully connected neural network and the initial scanning frequency,a one-dimensional wavelet convolutional neural network is proposed for Brillouin frequency shift extraction,which solves the problem that the accuracy of Brillouin frequency shift extraction algorithm based on fully connected neural network depends on the starting frequency and provide a method in low signalto-noise ratio,and realizes a Brillouin frequency shift extraction algorithm with good robustness and stability with real-time.By using the translational invariance of convolution and the local feature extraction ability of wavelet function,a Brillouin shift extraction algorithm based on wavelet convolutional neural network is designed.At the simulation level,it is verified that the robustness and stability of the convolutional neural network are superior to the LM algorithm and the fully connected neural network.The Brillouin frequency shift extraction algorithm based on convolutional neural network only takes 0.04s to process 4000 Brillouin gain spectra on the Brillouin optical time domain reflectometer,which is more than two orders of magnitude lower than the processing time of the LM algorithm,and has better robustness and accuracy.3.Aiming at the problem of single application scenario of neural network model,a Brillouin shift extraction algorithm based on adaptive convolutional neural network is proposed,which solves the reusability of neural network under different frequency sweep intervals and different data sampling lengths within a certain range,and realizes a more universal Brillouin frequency shift extraction algorithm based on neural network,which provides a feasible scheme for the universalization of neural networks in Brillouin frequency shift extraction applications.Both simulation and experimental results show that the Brillouin shift extraction algorithm based on adaptive convolutional neural network is superior to the LM algorithm when the frequency sweep interval is large,and its performance improvement in processing time is tens of times.In addition,compared with other neural network algorithms,it can be applied to multiple different frequency sweep intervals and different data sampling lengths,which is more universal.