Research On Signal-to-noise Ratio Improvement Of Φ-OTDR System Based On Image Denoising And Semantic Segmentation

In recent years,distributed optical fiber vibration sensing technology has attracted more and more attention because of its advantages,such as high sensitivity,anti-electromagnetic interference,and large monitoring range.The technology has been widely used in perimeter safety,gas pipeline safety warning,structural health monitoring,railway safety monitoring and other fields.Among them,phase-sensitive optical time-domain reflectometry(Φ-OTDR)system can detect the vibration signal at any position by measuring the coherent superposition change of Rayleigh backscattering(RBS)light.It has significant advantages in distributed,long-distance,multi-point measurement and locating.RBS light is easily affected by the surrounding environment in practical application,leading to a decrease in the signal-to-noise ratio(SNR)of the system.In addition,random noises such as phase noise and electrical noise of the laser will also deteriorate the signal and affect the SNR,thus increasing the difficulty of vibration detection.Therefore,improving the SNR of optical fiber vibration sensing signals has become an important scientific issue.InΦ-OTDR vibration sensing systems,methods such as increasing input optical power,increasing pulse width,introducing optical amplification devices,suppressing fading,and optimizing digital signal processing algorithms can be used to improve the SNR.However,excessive optical power can cause nonlinear effects and increase system noise.The increase in pulse width will directly affect the spatial resolution of the system,leading to a reduction in detection accuracy.Although introducing optical amplifier devices can improve optical power,it also increases system noise,complexity,and cost.Digital signal processing algorithms have attracted much attention in the field of optical fiber sensing due to their advantages such as no need to change the system structure and strong resistance to external interference.They can improve the SNR and other performance without changing the original parameters of the system.However,traditional digital signal processing algorithms for Φ-OTDR systems are generally based on signal denoising in the one-dimensional domain,and do not fully utilize the redundancy and correlation of signals in the multidimensional domain.Therefore,based on the study of the Φ-OTDR system,this article has carried out research on improving the SNR based on image processing technology for Rayleigh temporal-spatial image signals.The main innovative work is as follows:(1)Traditional optical amplification methods can increase the energy of pulsed light while also introducing additional noise to the Φ-OTDR system,and too strong signals can cause nonlinear effects in the system,resulting in limited improvement of the system’s signal-to-noise ratio.To address this problem,a signal to noise ratio improvement method forΦ-OTDR systems based on sparse representation image denoising is proposed.The sparse representation denoising method uses discrete cosine transform to construct the initial dictionary,updates the dictionary matrix and sparse matrix by the K-Singular Value Decomposition(K-SVD)algorithm,and then the RBS signal was reconstructed.In this reconstruction process,the effective signal can be sparsely represented by the linear combination of dictionary atoms,while the noise signal is discarded as the residual between the original RBS signal and the reconstruction target signal,so that the noise can be separated.The experimental results show that after signal noise reduction,the SNR at 10 km,20 km,30 km,and 40 km of optical fiber increases by an average of 3.44 dB compared to that before noise reduction.In addition,based on the sparse representation noise reduction method,the detection distance of the Φ-OTDR system can be increased to 79.75 km,and the SNR at the end of the optical fiber is 3.27 dB.The feasibility of sparse representation noise reduction method applied to Φ-OTDR system noise suppression and long-distance detection is verified through experiments.(2)The polarization orthogonal pulse pair formed by a polarization beam splitter(PBS)and a polarization maintaining-optical switch(PM-PSW)can suppress polarization fading inΦ-OTDR systems,but it also introduces noise that affects the improvement of SNR.Aiming at this problem,a SNR improvement method for Φ-OTDR systems based on non-local means(NLM)image denoising is proposed.The non-local mean image noise reduction method makes full use of redundant texture and self-similarity of multi-dimensional data.By weighted average of pixels with similar neighborhood structure in the image,the estimated noise reduction result value of the current pixel is obtained.Non-local mean noise reduction method breaks through the limitation of local characteristics of local image noise reduction algorithm and extends the calculation of pixel value of target pixels to the whole image region,which improves the noise reduction performance of the algorithm.Moreover,the parameter selection strategy of the NLM algorithm is proposed,and the time complexity of the NLM algorithm is optimized based on the integral graph.The experimental results show that the SNR of the original vibration positioning curve is 17.72 dB at 20.04 km of the optical fiber,and the SNR reaches 23.39 dB after implementing the NLM image denoising method for Rayleigh temporal-spatial image.In addition,the amount of SNR improvement under various experimental conditions also remains relatively stable,with SNR increments of 7 dB or more.Finally,the effectiveness of the proposed method for improving SNR of backrayleigh scattering signals is proved by experiments,and the stability of the proposed method for improving SNR under different experimental conditions is analyzed.(3)Aiming at the problem of limited SNR of differential positioning algorithms inΦ-OTDR systems,which cannot achieve long-distance detection,a deep learning based semantic image segmentation Φ-OTDR system positioning SNR enhancement method is proposed.In this study,a Rayleigh temporal-spatial image is constructed by using RBS curves to detect vibration signals.Encoder-decoder structure and atrous spatial pyramid pooling model are used to obtain the vibration region boundary and multi-scale context information from the Rayleigh temporal-spatial image.In this experiment,the SNR is calculated based on the pixel accumulation positioning curve of the semantic image segmentation results.The experimental results show that the SNR of the vibration positioning results at 4 km,10 km,20 km,and 40 km of optical fiber can reach 37.84 dB 34.28 dB,34.09 dB,and 32.17 dB,respectively.In addition,experimental verification shows that based on the image semantic segmentation method,the vibration detection range of the Φ-OTDR system can be increased to 80.75 km,and the SNR of the end positioning curve can reach over 30 dB.This semantic segmentation model realizes the pixel classification of RBS images based on the full convolutional neural network,and effectively improves the SNR of Φ-OTDR system.This research solves the problem that traditional differential positioning algorithms cannot achieve long-distance sensing positioning by implementing semantic extraction of vibration positions through binary classification of image pixels.In addition,the identification and location of multipoint vibrations have also been verified,and the adaptability and generalization of the model have been evaluated.