Research On Leakage Detection Of Natural Gas Pipeline Based On Dimensionality Reduction Algorithm

Natural gas is an important energy source in the world.With the improvement of living standards,people’s demand for natural gas is gradually increasing,and the pipeline transportation industry is also booming.Due to pipeline corrosion aging,extrusion damage,human factors stealing gas,improper installation and other reasons cause different degrees of pipeline damage,leakage is found less than will cause a certain degree of disaster accidents and property losses,endangering the safety of residents.Therefore,in order to improve the accuracy of pipeline leakage detection,this paper designs a set of natural gas pipeline leakage detection research methods from signal processing,feature extraction to condition recognition.Specific work is as follows:(1)In view of the complexity and non-stationarity of pipeline acoustic signals,which will affect the subsequent signal feature extraction and pipeline leakage detection accuracy,three characteristic parameter indexes including waveform characteristic parameter peak factor,entropy characteristic parameter spread entropy and root-mean-square value of time-frequency domain characteristic parameter are used.The characteristic Mode components of K Intrinsic Mode Functions(IMFs)obtained after signal Decomposition by Variational Mode Decomposition(VMD)algorithm are selected to achieve the pre-processing of pipeline signal.(2)To address the issue of feature modal components still having high dimensionality and possibly containing a large amount of redundant information,which can still affect the accuracy of pipeline leakage detection,it is necessary to reduce the dimensionality of feature modal components.Firstly,calculate the time-frequency domain features of the feature mode components to construct a high-dimensional feature model;Then,a Local Linear Embedding of Asymmetric Jaccard rank-order distances(JRLLE)algorithm based on asymmetric Jaccard rank-order distances is proposed to reduce the dimensionality of high-dimensional feature models.The Jaccard coefficient is used to strengthen the Euclidean distance order table between samples,and the common nearest neighbor information between data points Xi and Xj is obtained.After obtaining appropriate neighborhoods,effective low-dimensional discriminant features are screened out;Finally,low-dimensional features are input into Support Vector Machines(SVM)for classification and recognition.The average classification accuracy of SVM under different neighborhood parameters can reach 96.11%,reflecting the effectiveness of the VMD-JRLLE method.(3)Aiming at the problem that flat Euclidean space cannot effectively describe the nonlinear structure between data,which leads to the non representativeness of the extracted pipeline signal features,a feature extraction method combining VMD and Riemannian manifold(RM)is proposed.Firstly,VMD is used to decompose the eigenmode components and calculate the time-frequency characteristics,which is constructed into the form of symmetric positive definition(SPD)matrix,and the logarithm Euclid Riemann metric method is applied between the SPD matrices to make the data locate on RM from the flat Euclidean space;Then,perform JRLLE dimensionality reduction on RM to filter out effective low dimensional features;Finally,the low-dimensional features were input into SVM for classification and recognition,and the average classification accuracy of the VMD-RM-JRLLE method under different neighborhood parameters reached 98.62%,which was 2.51% higher than the average accuracy of the VMDJRLLE method.This proves the effectiveness of using data on RM for nonlinear structural analysis.