US 12,442,793 B2
Intelligent inversion method for pipeline defects based on heterogeneous field signals
Huaguang Zhang, Shenyang (CN); Jinhai Liu, Shenyang (CN); Lei Wang, Shenyang (CN); Jiayue Sun, Shenyang (CN); Jian Feng, Shenyang (CN); Gang Wang, Shenyang (CN); Dazhong Ma, Shenyang (CN); and Senxiang Lu, Shenyang (CN)
Assigned to NORTHEASTERN UNIVERSITY, Shenyang (CN)
Appl. No. 18/028,010
Filed by Northeastern University, Shenyang (CN)
PCT Filed Nov. 6, 2020, PCT No. PCT/CN2020/126885
§ 371(c)(1), (2) Date Mar. 23, 2023,
PCT Pub. No. WO2022/088226, PCT Pub. Date May 5, 2022.
Claims priority of application No. 202011186863.4 (CN), filed on Oct. 30, 2020.
Prior Publication US 2023/0341354 A1, Oct. 26, 2023
Int. Cl. G01N 27/83 (2006.01); G01B 7/02 (2006.01)
CPC G01N 27/83 (2013.01) [G01B 7/02 (2013.01)] 8 Claims
OG exemplary drawing
 
1. An intelligent inversion method for pipeline defects based on heterogeneous field signals, comprising the following steps:
Step 1: acquiring the heterogeneous field signals of each sampling point in a certain section of a pipeline in real time, performing an abnormality judgement on the acquired heterogeneous field signals, and then correcting base values of the heterogeneous field signals of the each sampling point by an improved average median method;
Step 2: denoising the heterogeneous field signals of the each sampling point, treated in Step 1, by a wavelet analysis method;
Step 3: padding the denoised heterogeneous field signals corresponding to the pipeline defects, unifying the heterogeneous field signals of different sizes into the heterogeneous field signals of the same size, wherein 0 is used for completion in a direction of the sampling point, and a median value of the heterogeneous field signals is used for completion in a direction of signal amplitudes, and performing a nonlinear transformation on the signal amplitudes;
Step 4: converting a heterogeneous field signal matrix corresponding to the pipeline defects, treated by the nonlinear transformation, into a data matrix with the same input dimensions as a sparse autoencoder;
Step 5: designing the sparse autoencoder with an axisymmetric structure, inputting the heterogeneous field signal matrix corresponding to converted pipeline defects into the sparse autoencoder, obtaining primary characteristics of the heterogeneous field signals corresponding to the pipeline defects, and saving a weight of an encoding part of the sparse autoencoder;
Step 6: classifying lengths, widths and depths of the pipeline defects to obtain category labels of the pipeline defects;
Step 7: designing a Softmax-based multi-classification neural network to classify the heterogeneous field signals corresponding to the pipeline defects in a supervised manner, and further extracting deep characteristics containing defect size information;
Step 8: taking the deep characteristics of the heterogeneous field signals corresponding to the pipeline defects extracted in Step 7 as inputs, and real size information of the heterogeneous field signals corresponding to the pipeline defects as outputs to construct a random forest regression model with ntrees decision trees, to realize intelligent inversion for sizes of the pipeline defects; and
Step 9: performing maintenance on the pipeline based on the constructed random forest regression model in Step 8.