A pressure pipeline corrosion online monitoring method and system

Abstract

The present invention provides a pressure pipeline corrosion online monitoring method and system, which relates to the technical field of pipeline monitoring, including acquiring three-dimensional point cloud data, obtaining downsampled point cloud data through an octree filtering algorithm, constructing a geometric feature scattering matrix, inputting a multi-scale fusion twin network, learning feature differences through a contrast loss function, and generating a corrosion probability distribution map; based on an iterative nearest point algorithm, aligning the corrosion probability distribution map, extracting an excessive corrosion area, superimposing and analyzing the excessive corrosion area and a pipeline stress concentration area, and determining a high-risk corrosion area; arranging a sensor array around the high-risk corrosion area, acquiring a sensor corrosion signal, performing variational mode decomposition, extracting an intrinsic function component related to the corrosion feature, sparsely representing the intrinsic function component, constructing a dispersion compensation matrix, determining a corrected sparse feature matrix, and inputting a multi-core extreme learning machine discriminator to determine whether a leakage occurs in the high-risk corrosion area.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline monitoring, and in particular to an online monitoring method and system for pressure pipeline corrosion. Background Art

[0002] Pressure pipelines are important equipment for transporting media in the petroleum, chemical, electric power and other industries. They operate for a long time under complex working conditions such as high temperature, high pressure, and corrosive media. They are prone to corrosion, leakage and other problems, which seriously threaten the safe operation of the pipelines and the surrounding environment.

[0003] Traditional pipeline inspection methods, such as visual inspection and ultrasonic inspection, have defects such as low inspection efficiency, long inspection cycle, and susceptibility to human factors, making it difficult to achieve real-time monitoring and early warning of pipeline corrosion. In order to solve the above problems, researchers have proposed various methods for online monitoring of pipeline corrosion. Among them, acoustic emission technology collects acoustic emission signals released by the pipeline during the corrosion process by laying acoustic emission sensors on the surface of the pipeline, and ultrasonic guided waves can propagate over long distances in the pipeline. By analyzing the amplitude, phase and other characteristics of ultrasonic guided wave signals, pipeline corrosion can be located and evaluated. However, the online monitoring of pipeline corrosion methods still have some shortcomings, such as lack of precision, susceptibility to noise interference, insufficient consideration of the acoustic characteristics of pipeline materials, difficulty in signal analysis and interpretation, and lack of intelligent discrimination and early warning mechanisms.

[0004] In summary, there is an urgent need for a high-precision, intelligent online monitoring method for pressure pipeline corrosion to achieve early warning and trend prediction of pipeline corrosion, ensure the safe operation of the pipeline, and improve the accuracy and intelligence level of pipeline corrosion monitoring. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The present invention provides a pressure pipeline corrosion online monitoring method and system, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] Provided is a pressure pipeline corrosion online monitoring method, comprising:

[0008] The three-dimensional point cloud data of the inner wall of the pressure pipeline is obtained, and the downsampled point cloud data is obtained through the octree filtering algorithm, and the geometric feature descriptor of the inner wall of the pipeline is extracted, and a geometric feature scattering matrix is ​​constructed. The geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network, and the feature differences of different areas of the inner wall of the pipeline are learned through the contrast loss function, and the corrosion probability distribution map of each area of ​​the inner wall of the pipeline is generated;

[0009] Based on the iterative closest point algorithm, the corrosion probability distribution map is aligned with the pre-acquired pipeline design map, and based on the preset corrosion probability threshold, the corrosion limit area is extracted from the corrosion probability distribution map, and based on the pipeline design map, the pipeline stress concentration area is determined, and the corrosion limit area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the area, and the high-risk corrosion area is determined;

[0010] A self-calibrated acoustic emission sensor array and a self-calibrated ultrasonic guided wave sensor array are arranged around the high-risk corrosion area to obtain sensor corrosion signals, perform variational modal decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, construct a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether leakage occurs in the high-risk corrosion area.

[0011] Preferably,

[0012] Obtain the three-dimensional point cloud data of the inner wall of the pressure pipe, and obtain the downsampled point cloud data through the octree filtering algorithm, including:

[0013] Acquire three-dimensional point cloud data of the inner wall of the pressure pipe, and divide the three-dimensional point cloud data into cubic voxels of equal size according to a preset number of voxels, with the point cloud area corresponding to the three-dimensional point cloud data serving as a root node, and each of the cubic voxels corresponding to a child node;

[0014] Traverse each child node from top to bottom to determine the first operation current node. If the number of point clouds in the first operation current node is greater than a preset node point cloud threshold, divide the first operation current node into eight child nodes. Otherwise, the first operation current node is used as a leaf node, and the recursive division is repeated until the number of point clouds in all leaf nodes is less than or equal to the node point cloud threshold, thereby generating an octree.

[0015] Traverse each non-leaf node of the octree from bottom to top to determine the second operation current node, and if all child nodes of the second operation current node are empty, mark the second operation current node as an empty node, and delete all child nodes corresponding to the second operation current node;

[0016] Traversing all non-empty leaf nodes of the octree, determining a third operation current node, calculating an average density of a point cloud in the third operation current node, determining a node point cloud density, and deleting the third operation current node if the node point cloud density of the third operation current node is less than a preset density threshold;

[0017] After the deletion operation, a valid non-empty leaf node is obtained, and all three-dimensional point cloud data contained in the valid non-empty leaf node is obtained to obtain downsampled point cloud data;

[0018] Preferably,

[0019] Extract the geometric feature descriptor of the inner wall of the pipeline and construct the geometric feature scattering matrix including:

[0020] Based on the downsampled point cloud data, a topological space is determined, and for each sampling point in the topological space, a plurality of spherical neighborhoods of different scales are set according to a preset initial spherical radius and a preset radius increasing rate to construct a multi-scale neighborhood;

[0021] In a multi-scale neighborhood, a spherical neighborhood sub-point cloud of multiple scales is determined, a simplicial complex is constructed based on the Alpha shape algorithm, and a homology group at a corresponding scale is obtained by calculating the number of connected components and the Betti number of the simplicial complex; a binary decision vector of the multi-scale neighborhood is constructed, and an optimal feature extraction scale set is determined by iterative search through a particle swarm algorithm with the goal of minimizing the number of generators corresponding to the homology group at each selected scale;

[0022] For the spherical neighborhood sub-point cloud corresponding to each scale in the optimal feature extraction scale set, the normal vector is calculated by the preset principal component analysis method, the curvature at the sampling point is calculated by eigenvalue decomposition, and the shape index is calculated based on the point cloud density distribution to obtain a multi-scale geometric feature descriptor;

[0023] The multi-scale geometric feature descriptor is mapped to a high-dimensional manifold space. Through a local linear embedding algorithm, the multi-scale geometric feature descriptor is used as input to learn the mapping relationship between the high-dimensional manifold space and the original feature space, and the coordinates of the multi-scale geometric feature descriptor in the high-dimensional manifold space are obtained; in the high-dimensional manifold space, the scattering similarity between the multi-scale geometric feature descriptors is calculated through a Gaussian kernel function, and a geometric feature scattering matrix is ​​constructed.

[0024] Preferably,

[0025] The geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network, and the feature differences of different areas of the inner wall of the pipeline are learned through the contrast loss function to generate the corrosion probability distribution map of each area of ​​the inner wall of the pipeline, including:

[0026] Inputting the geometric feature scattering matrix into a pre-trained multi-scale fusion twin network, wherein the multi-scale fusion twin network is composed of two convolutional sub-networks with the same structure and shared parameters, and performing feature extraction on the rows and columns of the scattering matrix respectively;

[0027] In each of the convolutional subnetworks, a stack of multiple convolutional layers and pooling layers is used to perform multi-scale feature extraction on the geometric feature scattering matrix, and local structural features of different scales are captured through different receptive fields of different convolutional layers. The local structural features of different scales are adaptively fused in combination with the attention mechanism to obtain row feature maps and column feature maps, and the similarity between the row feature maps and the column feature maps is measured based on the contrast loss function;

[0028] By back propagation and gradient descent algorithm, the contrast loss function is minimized, the parameters of the multi-scale fusion twin network are adjusted, and the training is repeated until a preset number of iterations is reached;

[0029] Based on the trained multi-scale fusion twin network, dense sampling is performed on the inner wall of the pipeline, the inner wall of the pipeline is divided into local areas of equal size, and feature vectors of the local areas are generated;

[0030] The average Euclidean distance between each of the characteristic vectors and the characteristic vector of the known corrosion area is calculated to determine the corrosion probability of the local area, and generate a corrosion probability distribution map of each area on the inner wall of the pipeline.

[0031] Preferably,

[0032] Also includes:

[0033] The contrast loss function is formulated as follows:

[0034] ;

[0035] in, L represents the contrast loss function, N represents the total number of sample pairs, i represents the index of the sample pair, y i represents a binary label, tanh() represents the hyperbolic tangent function, α represents the distance scaling factor, x i 1 Indicates i The first eigenvector in the sample pairs, x i 2 Indicates i The second eigenvector in the sample pairs, d () represents the distance metric function, max() represents the maximum value function, m Represents the distance sensitivity threshold, exp() represents the exponential function, β Represents the exponential scaling factor.

[0036] Preferably,

[0037] Based on the iterative closest point algorithm, the corrosion probability distribution map is aligned with the pre-acquired pipeline design map. Based on the preset corrosion probability threshold, the corrosion limit area is extracted from the corrosion probability distribution map. Based on the pipeline design map, the pipeline stress concentration area is determined. The corrosion limit area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the area. The high-risk corrosion area is determined to include:

[0038] Binarizing the corrosion probability distribution map, extracting the corrosion area contour, and obtaining a corrosion area contour point set; vectorizing the pre-acquired pipeline design drawing, extracting the pipeline centerline, and obtaining a pipeline centerline point set;

[0039] Inputting the corrosion area contour point set and the pipeline centerline point set into an iterative closest point algorithm;

[0040] Initialize the rigid body transformation matrix as the current rigid body transformation matrix;

[0041] According to the current rigid body transformation matrix, the corrosion area contour point set is transformed into the coordinate system corresponding to the pipeline centerline point set to obtain the transformed corrosion area contour point set;

[0042] Based on each point in the transformed corrosion area contour point set, find the closest point in the pipeline centerline point set to construct an associated point pair;

[0043] Based on the least squares method, the optimal transformation matrix is ​​calculated for the associated point pairs with the goal of minimizing the distance error between the transformed corrosion area contour point set and the pipeline centerline point set;

[0044] If the distance error is greater than a preset error threshold, the current rigid body transformation matrix is ​​updated with the optimal transformation matrix and the iteration continues; otherwise, the iteration is terminated to determine the final optimal transformation matrix;

[0045] Based on the final optimal transformation matrix, generate the registered corrosion probability distribution map;

[0046] Based on the preset corrosion probability threshold, the corrosion excess area is extracted from the registered corrosion probability distribution map, and the pipeline stress concentration area is determined through finite element analysis based on the pipeline design drawings; the corrosion excess area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the two areas, which is determined as the high-risk corrosion area of ​​the pipeline.

[0047] Preferably,

[0048] The sensor corrosion signal is subjected to variational mode decomposition, the intrinsic function components related to the corrosion characteristics are extracted, the intrinsic function components are sparsely characterized, a dispersion compensation matrix is ​​constructed in combination with the acoustic dispersion characteristics of the pipeline material, and the corrected sparse feature matrix is ​​determined and input into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in a high-risk corrosion area, including:

[0049] Performing variational mode decomposition on the sensor corrosion signal, decomposing the sensor corrosion signal into multiple eigenfunction components, calculating the center frequency and bandwidth of each eigenfunction component, and based on a preset frequency range and a preset bandwidth threshold, screening the eigenfunction components within the frequency range and with a bandwidth less than the bandwidth threshold, and determining the corrosion characteristic eigenfunction components;

[0050] Performing time-frequency analysis on the corrosion characteristic intrinsic function components to obtain a time-frequency representation matrix, and reducing the dimension of the time-frequency representation matrix using singular value decomposition to obtain a reduced-dimensional feature matrix;

[0051] Input the reduced dimension feature matrix into the dictionary learning algorithm, based on the preset initial dictionary matrix and the preset initial sparse coefficient matrix, by fixing the initial sparse coefficient matrix, combined with minimizing the reconstruction error, update the initial dictionary matrix to obtain a new dictionary matrix, fix the new dictionary matrix, with the goal of minimizing the weighted sum of the reconstruction error and the sparsity, with the sum of the L1 norms of the atoms in the new dictionary matrix equal to 1 as a constraint, solve and update the initial sparse coefficient matrix through a preset orthogonal matching pursuit algorithm, and obtain a new sparse coefficient matrix; fix and update in turn until the preset number of iterations is reached, and determine the final sparse coefficient matrix;

[0052] Based on the pre-acquired acoustic dispersion curve of the pipeline material, the least square method is used to fit the dispersion curve parameters, determine the phase delay and amplitude attenuation coefficient at different frequencies, and construct the dispersion compensation matrix;

[0053] The dispersion compensation matrix is ​​used to perform dispersion compensation on the final sparse coefficient matrix, a compensated sparse feature matrix is ​​determined, and the compensated sparse feature matrix is ​​normalized to obtain a corrected sparse feature matrix.

[0054] The second aspect of the present invention,

[0055] Provided is a pressure pipeline corrosion online monitoring system, comprising:

[0056] The first unit is used to obtain the three-dimensional point cloud data of the inner wall of the pressure pipeline, obtain the downsampled point cloud data through the octree filtering algorithm, extract the geometric feature descriptor of the inner wall of the pipeline, construct the geometric feature scattering matrix, input the geometric feature scattering matrix into the pre-trained multi-scale fusion twin network, learn the feature differences of different areas of the inner wall of the pipeline through the contrast loss function, and generate the corrosion probability distribution map of each area of ​​the inner wall of the pipeline;

[0057] The second unit is used to align the corrosion probability distribution map with the pre-acquired pipeline design map based on an iterative closest point algorithm, extract the corrosion limit area from the corrosion probability distribution map based on a preset corrosion probability threshold, determine the pipeline stress concentration area based on the pipeline design map, perform superposition analysis on the corrosion limit area and the pipeline stress concentration area, obtain the overlapping part of the area, and determine the high-risk corrosion area;

[0058] The third unit is used to deploy a self-calibration acoustic emission sensor array and a self-calibration ultrasonic guided wave sensor array around the high-risk corrosion area, obtain sensor corrosion signals, perform variational mode decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, construct a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in the high-risk corrosion area.

[0059] The third aspect of the present invention,

[0060] An electronic device is provided, comprising:

[0061] processor;

[0062] a memory for storing processor-executable instructions;

[0063] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0064] A fourth aspect of the present invention,

[0065] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0066] In the present invention, the point cloud data is downsampled by the octree filtering algorithm to effectively reduce the amount of data and improve the efficiency of subsequent feature extraction and analysis; the geometric feature descriptor of the inner wall of the pipeline is extracted, and a geometric feature scattering matrix is ​​constructed to ensure the accurate capture and expression of the local features of the inner wall of the pipeline, providing a reliable basis for subsequent analysis; the geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network to achieve effective fusion of features of different scales and enhance the model's ability to recognize diverse corrosion features; the feature differences of different regions are learned through the contrast loss function to generate a corrosion probability distribution map, which is helpful to accurately evaluate the corrosion condition of the inner wall of the pipeline and guide maintenance and repair work; using the iterative closest point algorithm, the corrosion probability distribution map is accurately aligned with the pipeline design map to ensure the accuracy of the analysis results; by extracting the corrosion limit area, the potential serious corrosion area in the pipeline is accurately identified to improve the effect of risk identification; by deploying self-calibration acoustic emission and ultrasonic guided wave sensor arrays, real-time monitoring of high-risk corrosion areas is achieved and corrosion signals are captured in time; using variational mode decomposition and sparse characterization, the intrinsic function components related to the corrosion characteristics are accurately extracted to improve the accuracy of signal processing; combining the acoustic dispersion characteristics of the pipeline material, a dispersion compensation matrix is ​​constructed to correct the signal characteristics and eliminate errors in the propagation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of a process of an online monitoring method for pressure pipeline corrosion according to an embodiment of the present invention;

[0068] Figure 2 The present invention is a schematic structural diagram of an online pressure pipeline corrosion monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 1 is a flow chart of an online pressure pipeline corrosion monitoring method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] S101. Acquire three-dimensional point cloud data of the inner wall of the pressure pipeline, obtain downsampled point cloud data through an octree filtering algorithm, extract geometric feature descriptors of the inner wall of the pipeline, construct a geometric feature scattering matrix, input the geometric feature scattering matrix into a pre-trained multi-scale fusion twin network, learn the feature differences of different areas of the inner wall of the pipeline through a contrast loss function, and generate a corrosion probability distribution map of each area of ​​the inner wall of the pipeline;

[0073] In this embodiment, the point cloud data is downsampled through the octree filtering algorithm to effectively reduce the amount of data and improve the efficiency of subsequent feature extraction and analysis; the geometric feature descriptor of the inner wall of the pipeline is extracted, and the geometric feature scattering matrix is ​​constructed to ensure the accurate capture and expression of the local features of the inner wall of the pipeline, providing a reliable basis for subsequent analysis; the geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network to achieve effective fusion of features of different scales and enhance the model's ability to recognize diverse corrosion features; the feature differences of different regions are learned through the contrast loss function to generate a corrosion probability distribution map, which helps to accurately evaluate the corrosion condition of the inner wall of the pipeline and guide maintenance and repair work.

[0074] In an optional embodiment, obtaining three-dimensional point cloud data of the inner wall of the pressure pipe and obtaining downsampled point cloud data through an octree filtering algorithm includes:

[0075] Acquire three-dimensional point cloud data of the inner wall of the pressure pipe, and divide the three-dimensional point cloud data into cubic voxels of equal size according to a preset number of voxels, with the point cloud area corresponding to the three-dimensional point cloud data serving as a root node, and each of the cubic voxels corresponding to a child node;

[0076] Traverse each child node from top to bottom to determine the first operation current node. If the number of point clouds in the first operation current node is greater than a preset node point cloud threshold, divide the first operation current node into eight child nodes. Otherwise, the first operation current node is used as a leaf node, and the recursive division is repeated until the number of point clouds in all leaf nodes is less than or equal to the node point cloud threshold, thereby generating an octree.

[0077] Traverse each non-leaf node of the octree from bottom to top to determine the second operation current node, and if all child nodes of the second operation current node are empty, mark the second operation current node as an empty node, and delete all child nodes corresponding to the second operation current node;

[0078] Traversing all non-empty leaf nodes of the octree, determining a third operation current node, calculating an average density of a point cloud in the third operation current node, determining a node point cloud density, and deleting the third operation current node if the node point cloud density of the third operation current node is less than a preset density threshold;

[0079] After the deletion operation, a valid non-empty leaf node is obtained, and all three-dimensional point cloud data contained in the valid non-empty leaf node is obtained to obtain downsampled point cloud data;

[0080] Prepare 3D scanning equipment, such as laser scanner, structured light scanner or other point cloud acquisition equipment; conduct a comprehensive scan of the inner wall of the pressure pipe to ensure that the scanning range covers the entire inner wall surface of the pipe; the scanning equipment collects 3D coordinate point data of the inner wall surface of the pipe, and each point contains its X, Y, and Z coordinate values ​​in 3D space; store the collected 3D coordinate point data in a computer to form an original point cloud data set of the inner wall of the pressure pipe.

[0081] Determine the preset number of voxels and set it according to the size of the point cloud data and the required resolution; calculate the bounding box of the point cloud data, that is, the minimum and maximum coordinate values ​​of the point cloud data in the X, Y, and Z directions; calculate the size of each voxel according to the size of the bounding box and the preset number of voxels, that is, the length of the voxel in the X, Y, and Z directions; divide the bounding box of the point cloud data into cubic voxels of equal size, each voxel represents a sub-area of ​​the point cloud area; assign each point in the original point cloud data set to the corresponding voxel, and establish a correspondence between the point and the voxel; use the point cloud area as the root node of the octree, and each voxel corresponds to a child node of the octree.

[0082] Starting from the root node, traverse each node of the octree from top to bottom. For the currently traversed node (the first operation current node), count the number of point clouds contained in it.

[0083] Compare the number of point clouds of the current node with the preset node point cloud threshold: if the number of point clouds is greater than the threshold, divide the current node into eight child nodes; if the number of point clouds is less than or equal to the threshold, mark the current node as a leaf node.

[0084] If the current node is divided into eight sub-nodes, then: calculate the bounding boxes of the eight sub-regions of the current node; divide the point cloud data of the current node according to the bounding boxes of the sub-regions and distribute them to the corresponding sub-nodes.

[0085] For each child node, do this recursively until all nodes have been processed.

[0086] Starting from the leaf node, traverse each non-leaf node of the octree from bottom to top. For the currently traversed node (the second operation current node), determine whether it is an empty node: if all the child nodes of the current node are empty nodes, mark the current node as an empty node; if any child node of the current node is not empty, the current node is a non-empty node.

[0087] If the current node is marked as an empty node, delete all its child nodes.

[0088] Repeat this process recursively until all the root nodes have been processed.

[0089] Traverse all non-empty leaf nodes of the octree. For each leaf node (the third operation current node), calculate its node point cloud density: count the number of point clouds in the current node, recorded as the number of point clouds; calculate the volume of the current node, that is, the product of the cube size of the leaf node;

[0090] Node point cloud density = point cloud quantity / node volume.

[0091] The node point cloud density is compared with the preset density threshold: if the node point cloud density is less than the density threshold, the current leaf node is marked as a low-density node; if the node point cloud density is greater than or equal to the density threshold, the current leaf node is retained.

[0092] Delete all leaf nodes marked as low-density nodes and the point cloud data they contain.

[0093] Traverse all non-empty leaf nodes of the octree, and for each valid non-empty leaf node, extract all the point cloud data contained in it. Merge the extracted point cloud data into a new point cloud data set to form downsampled point cloud data. Store the downsampled point cloud data in a computer for subsequent analysis and processing.

[0094] In this embodiment, the point cloud data is divided into cubic voxels and recursively divided using an octree structure, which can effectively manage and organize a large amount of three-dimensional point cloud data, reduce data redundancy, highlight key areas, reduce computational complexity, and improve computational efficiency; empty nodes are deleted from the bottom up, and leaf nodes are density-screened to ensure that only data in high-density areas are retained, effectively filter out irrelevant or redundant data, and concentrate resources to process more critical areas, thereby improving data quality and accuracy; by downsampling the point cloud data, the amount of data is significantly reduced, important geometric features are fully retained, storage and computational overhead are reduced, and sufficient information is retained for subsequent geometric feature analysis and corrosion detection; by gradually refining the recursive segmentation and screening strategy, the waste of computing resources caused by global processing is avoided, which helps to efficiently locate and analyze key areas in complex pipeline structures, improve processing speed, and reduce consumption of computing resources.

[0095] In an optional embodiment, extracting the geometric feature descriptor of the inner wall of the pipeline and constructing the geometric feature scattering matrix includes:

[0096] Based on the downsampled point cloud data, a topological space is determined, and for each sampling point in the topological space, a plurality of spherical neighborhoods of different scales are set according to a preset initial spherical radius and a preset radius increasing rate to construct a multi-scale neighborhood;

[0097] In a multi-scale neighborhood, a spherical neighborhood sub-point cloud of multiple scales is determined, a simplicial complex is constructed based on the Alpha shape algorithm, and a homology group at a corresponding scale is obtained by calculating the number of connected components and the Betti number of the simplicial complex; a binary decision vector of the multi-scale neighborhood is constructed, and an optimal feature extraction scale set is determined by iterative search through a particle swarm algorithm with the goal of minimizing the number of generators corresponding to the homology group at each selected scale;

[0098] For the spherical neighborhood sub-point cloud corresponding to each scale in the optimal feature extraction scale set, the normal vector is calculated by the preset principal component analysis method, the curvature at the sampling point is calculated by eigenvalue decomposition, and the shape index is calculated based on the point cloud density distribution to obtain a multi-scale geometric feature descriptor;

[0099] The multi-scale geometric feature descriptor is mapped to a high-dimensional manifold space. Through a local linear embedding algorithm, the multi-scale geometric feature descriptor is used as input to learn the mapping relationship between the high-dimensional manifold space and the original feature space, and the coordinates of the multi-scale geometric feature descriptor in the high-dimensional manifold space are obtained; in the high-dimensional manifold space, the scattering similarity between the multi-scale geometric feature descriptors is calculated through a Gaussian kernel function, and a geometric feature scattering matrix is ​​constructed.

[0100] The shape algorithm specifically refers to an algorithm for describing a point cloud. By adjusting a parameter α, the points in the point cloud are connected to form a polyhedron structure. The polyhedron structure is used to approximate the true geometric shape of the point cloud. The α value determines the degree of refinement of the shape.

[0101] The number of connected components specifically refers to each connected subset in the topological space. If two points are in the same connected component, there is a path between them that can reach each other. The number of connected components refers to the number of such connected subsets; in a simplicial complex, it represents the number of independent parts in the topological structure. By analyzing the number of connected components, we can understand the complexity and connectivity of the geometric structure.

[0102] The Betti number specifically refers to a set of numbers used in topology to describe the number of holes in a topological space. The first number in the Betti number sequence represents the number of independent connected components, the second number represents the number of independent rings, the third number represents the number of independent cavities, and so on. It is an important indicator to measure the complexity of a topological space, reflecting the multidimensional connectivity characteristics of the space. By calculating the Betti number, the topological structure in different dimensions can be quantitatively described.

[0103] The homology group specifically refers to a concept in topology, which describes the shape characteristics of a topological space through algebraic methods. It is composed of topological features such as connected components, rings and cavities, and is represented in the form of algebraic structures in the homology group. Homology groups can help identify and classify different shapes and structures in topological spaces. By studying homology groups, we can have a deeper understanding of the topological characteristics of geometric structures.

[0104] The binary decision vector specifically refers to a vector consisting of 0 and 1, each element of which represents a decision variable. 1 indicates that a scale or feature is selected, and 0 indicates that it is not selected. In optimization problems, a binary decision vector is used to represent the selection of a set of candidate solutions or features.

[0105] The generator specifically refers to a set of basic elements that constitute the homology group. Through the generator, all elements in the homology group can be represented. It is used to describe the structure of the homology group and is the key to understanding the topological characteristics of the homology group at a certain scale. The goal in the optimization problem is to minimize the number of generators to simplify the description of the topological structure.

[0106] Based on the downsampled point cloud data, a topological space is determined, which contains all the sampling points. For each sampling point in the topological space, multiple spherical neighborhoods of different scales are set. A preset initial spherical radius and a preset radius increasing rate are set. With the sampling point as the center and the initial spherical radius as the radius, the first spherical neighborhood is constructed. The spherical radius is gradually increased according to the preset radius increasing rate to construct multiple spherical neighborhoods of different scales. Each sampling point will have multiple spherical neighborhoods of different scales, forming a multi-scale neighborhood.

[0107] For the multi-scale neighborhood of each sampling point, determine the spherical neighborhood sub-point cloud at each scale. For the spherical neighborhood sub-point cloud at each scale, apply the Alpha shape algorithm to construct a simplicial complex. According to the coordinates of the spherical neighborhood sub-point cloud, generate a simplicial complex by the Alpha shape algorithm. A simplicial complex is a topological structure composed of points, edges, faces, etc. For the simplicial complex constructed at each scale, calculate the number of connected components and Betti numbers. The number of connected components represents the number of independent connected regions in the simplicial complex. Betti numbers are a set of values ​​that describe the topological characteristics of the simplicial complex, including the number of connected components, the number of holes, etc. Based on the number of connected components and Betti numbers, the homology group at each scale is obtained.

[0108] Construct a binary decision vector for the multi-scale neighborhood. Assign a binary bit to each scale to indicate whether the scale is selected. If a scale is selected, the corresponding binary bit is 1, otherwise it is 0. Use the particle swarm algorithm for iterative search. Initialize a group of particles, each particle represents a possible scale selection scheme. For each particle, calculate its fitness function value, that is, the sum of the number of generators of the homology group at each selected scale. Search for the optimal scale selection scheme by iteratively updating the position and velocity of the particle. The optimization goal is to minimize the number of generators of the homology group at each selected scale. Determine the optimal feature extraction scale set through the search results of the particle swarm algorithm.

[0109] For each scale in the optimal feature extraction scale set, the corresponding spherical neighborhood sub-point cloud is extracted. For each spherical neighborhood sub-point cloud, the principal component analysis method is applied to calculate the normal vector. The principal component analysis method is used to solve the main direction and secondary direction of the spherical neighborhood sub-point cloud to obtain the normal vector representing the local geometric features. The curvature at the sampling point is calculated by eigenvalue decomposition. The spherical neighborhood sub-point cloud is subjected to eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. The principal curvature and Gaussian curvature at the sampling point are calculated based on the eigenvalues. The shape index is calculated based on the point cloud density distribution. The density distribution of the spherical neighborhood sub-point cloud is analyzed, and the shape index reflecting the local shape characteristics is calculated according to the characteristics of the density distribution. The normal vector, curvature and shape index are combined to obtain the geometric feature descriptor of the sampling point at each scale. The geometric feature descriptors at different scales are combined to form a multi-scale geometric feature descriptor.

[0110] Map the multi-scale geometric feature descriptor to a high-dimensional manifold space. Select a suitable high-dimensional manifold space, such as Euclidean space or Riemannian manifold, and use the multi-scale geometric feature descriptor as the coordinate in the high-dimensional manifold space. Learn the mapping relationship through the local linear embedding algorithm. Take the multi-scale geometric feature descriptor as input and apply the local linear embedding algorithm. By minimizing the local reconstruction error, learn the mapping relationship between the high-dimensional manifold space and the original feature space, and obtain the coordinate representation of the multi-scale geometric feature descriptor in the high-dimensional manifold space. Calculate the scattering similarity in the high-dimensional manifold space. Use the Gaussian kernel function to calculate the similarity between the multi-scale geometric feature descriptors. The Gaussian kernel function gives the similarity value according to the distance between the descriptors in the high-dimensional manifold space. Construct a geometric feature scattering matrix. Based on the calculated scattering similarity, construct a scattering matrix between descriptors. The scattering matrix reflects the geometric feature similarity between different sampling points.

[0111] In this embodiment, by setting ball neighborhoods of different scales and using topological methods such as the Alpha shape algorithm and the Betti number, the complex geometric features of the point cloud data can be captured and extracted at multiple scales, which helps to more accurately reflect the true shape and topological structure of the inner wall of the pressure pipeline; by combining geometric features such as normal vectors, curvatures and shape indices, rich geometric feature descriptors are constructed at multiple scales, which can more comprehensively capture the geometric and topological features of the point cloud data, and help to more efficiently analyze and identify the structural features of the inner wall of the pipeline; the multi-scale geometric feature descriptors are mapped to a high-dimensional manifold space, and the mapping relationship between the original feature space and the high-dimensional manifold space is learned through a local linear embedding algorithm, which can better preserve the relationship between the geometric features and identify and classify the similarities between different sampling points;

[0112] In an optional embodiment, the geometric feature scattering matrix is ​​input into a pre-trained multi-scale fusion twin network, and the feature differences of different areas of the inner wall of the pipeline are learned through a contrast loss function, and the corrosion probability distribution map of each area of ​​the inner wall of the pipeline is generated, including:

[0113] Inputting the geometric feature scattering matrix into a pre-trained multi-scale fusion twin network, wherein the multi-scale fusion twin network is composed of two convolutional sub-networks with the same structure and shared parameters, and performing feature extraction on the rows and columns of the scattering matrix respectively;

[0114] In each of the convolutional subnetworks, a stack of multiple convolutional layers and pooling layers is used to perform multi-scale feature extraction on the geometric feature scattering matrix, and local structural features of different scales are captured through different receptive fields of different convolutional layers. The local structural features of different scales are adaptively fused in combination with the attention mechanism to obtain row feature maps and column feature maps, and the similarity between the row feature maps and the column feature maps is measured based on the contrast loss function;

[0115] By back propagation and gradient descent algorithm, the contrast loss function is minimized, the parameters of the multi-scale fusion twin network are adjusted, and the training is repeated until a preset number of iterations is reached;

[0116] Based on the trained multi-scale fusion twin network, dense sampling is performed on the inner wall of the pipeline, the inner wall of the pipeline is divided into local areas of equal size, and feature vectors of the local areas are generated;

[0117] The average Euclidean distance between each of the characteristic vectors and the characteristic vector of the known corrosion area is calculated to determine the corrosion probability of the local area, and generate a corrosion probability distribution map of each area on the inner wall of the pipeline.

[0118] The geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network for processing. The multi-scale fusion twin network consists of two convolutional sub-networks with the same structure and shared parameters, which extract features from the rows and columns of the scattering matrix respectively. In each convolutional sub-network, a stack of multiple convolutional layers and pooling layers is used to perform multi-scale feature extraction on the geometric feature scattering matrix. By setting different receptive field sizes for different convolutional layers, local structural features of different scales can be captured. At the same time, combined with the attention mechanism, the local structural features extracted at different scales are adaptively fused to obtain row feature maps and column feature maps. Based on the contrast loss function, the similarity between the row feature map and the column feature map is measured to evaluate the effect of feature extraction.

[0119] During the training process, the contrast loss function is minimized through back propagation and gradient descent algorithms, and the parameters of the multi-scale fusion twin network are continuously adjusted. The training is repeated until the preset number of iterations is reached. Through such a training process, the multi-scale fusion twin network can effectively extract the multi-scale local structural features in the geometric feature scattering matrix and establish a suitable similarity measure between the row feature map and the column feature map.

[0120] After the training is completed, the inner wall of the pipeline is densely sampled based on the trained multi-scale fusion twin network. The inner wall of the pipeline is divided into local areas of equal size, and features are extracted from each local area to generate the corresponding feature vector. By inputting the feature vector of the local area into the trained multi-scale fusion twin network, the row feature map and column feature map of the area can be obtained.

[0121] Next, the average Euclidean distance between the feature vector of each local area and the feature vector of the known corrosion area is calculated. The feature vector of the known corrosion area can be obtained by extracting features from the area that has been determined to be corroded. By comparing the similarity between the feature vector of the local area and the feature vector of the known corrosion area, the probability of corrosion in the local area can be determined. The higher the similarity, the closer the features of the local area are to the known corrosion area, and the greater the possibility of corrosion.

[0122] Finally, based on the calculated corrosion probability of each local area, a corrosion probability distribution map of each area on the inner wall of the pipeline is generated. The corrosion probability distribution map intuitively shows the possibility of corrosion in different areas of the inner wall of the pipeline, and different corrosion probabilities can be represented by color or grayscale values. By analyzing the corrosion probability distribution map, areas with higher corrosion risks on the inner wall of the pipeline can be identified, providing guidance for subsequent maintenance and repair work.

[0123] In this embodiment, the geometric feature scattering matrix is ​​processed by a multi-scale fusion twin network, which can effectively capture the local structural features of the inner wall of the pipeline at different scales, and adaptively fuse these features to improve the accuracy of corrosion area identification; the similarity between the row feature map and the column feature map is used to optimize the contrast loss, so that the network can more accurately extract and measure the features of the corrosion area, thereby improving the reliability of corrosion detection; by comparing with the feature vectors of known corrosion areas, the corrosion probability of the local area is calculated, and a corrosion probability distribution map is generated, which helps to intuitively identify high-risk corrosion areas on the inner wall of the pipeline and facilitate timely maintenance measures.

[0124] In an optional embodiment, it also includes:

[0125] The contrast loss function is formulated as follows:

[0126] ;

[0127] in, L represents the contrast loss function, N represents the total number of sample pairs, i represents the index of the sample pair, y i represents a binary label, tanh() represents the hyperbolic tangent function, α represents the distance scaling factor, x i 1 Indicates i The first eigenvector in the sample pairs, x i 2 Indicates i The second eigenvector in the sample pairs, d () represents the distance metric function, max() represents the maximum value function, m Represents the distance sensitivity threshold, exp() represents the exponential function, β Represents the exponential scaling factor.

[0128] The first step is to average the loss values ​​of all sample pairs, that is, to average the loss values ​​of all sample pairs from 1 to N The sample pair index i Sum the values ​​and divide by 2 times the total number of sample pairs N , which is the total number of samples.

[0129] For each sample pair, the loss value consists of two parts, according to the binary label y i Select the value of .

[0130] When the binary label y iWhen it is 1, the first part is selected, that is, the distance between the two feature vectors in the sample pair is measured by the distance metric function d Calculated and then multiplied by the distance scaling factor α , and then transformed by the hyperbolic tangent function tanh.

[0131] When the binary label y i When it is 0, the second part is selected, that is, the distance between the two feature vectors in the sample pair is measured by the distance metric function d Calculate and then multiply by the exponential scaling factor β , and then transformed by the exponential function exp, and finally with the distance sensitivity threshold m Compare and take the larger value as 0.

[0132] The value of the contrast loss function is the average of the loss values ​​of all sample pairs.

[0133] In this embodiment, by combining the hyperbolic tangent function and the exponential function, different types of sample pairs (positive sample pairs and negative sample pairs) are differentiated, thereby enhancing the loss function's ability to distinguish between similar and dissimilar sample pairs; the distance metric is adjusted using the distance scaling factor and the sensitivity threshold, so that the loss function can be flexibly adjusted according to the data characteristics, adapting to different feature spaces and distance scales, and improving the robustness of the model; the hyperbolic tangent function provides a smooth loss curve, and the maximum value function ensures the penalty intensity for negative sample pairs, effectively avoiding the gradient vanishing problem, and ensuring that the loss function can converge stably during the training process; by setting the distance sensitivity threshold, the loss function controls negative sample pairs with excessively large distances, avoids the negative impact of outliers on model training, and improves the reliability of the overall training effect.

[0134] S102. Based on the iterative closest point algorithm, the corrosion probability distribution map is aligned with the pre-acquired pipeline design drawing, and based on the preset corrosion probability threshold, the corrosion exceeding limit area is extracted from the corrosion probability distribution map, and based on the pipeline design drawing, the pipeline stress concentration area is determined, and the corrosion exceeding limit area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the area, and the high-risk corrosion area is determined;

[0135] In this embodiment, the iterative closest point algorithm is used to accurately align the corrosion probability distribution map with the pipeline design map to ensure the accuracy of the analysis results; by extracting the corrosion limit-exceeding area, the potential severe corrosion area in the pipeline is accurately identified to improve the effect of risk identification; the corrosion limit-exceeding area and the pipeline stress concentration area are superimposed and analyzed to identify the overlapping parts of the area, thereby accurately determining the high-risk corrosion area in the pipeline, which helps to provide early warning and prevent accidents.

[0136] In an optional embodiment, based on an iterative closest point algorithm, the corrosion probability distribution map is aligned with a pre-acquired pipeline design map, and based on a preset corrosion probability threshold, a corrosion excess area is extracted from the corrosion probability distribution map, and based on the pipeline design map, a pipeline stress concentration area is determined, and a superposition analysis is performed on the corrosion excess area and the pipeline stress concentration area to obtain an overlapping part of the area, and the high-risk corrosion area is determined to include:

[0137] Binarizing the corrosion probability distribution map, extracting the corrosion area contour, and obtaining a corrosion area contour point set; vectorizing the pre-acquired pipeline design drawing, extracting the pipeline centerline, and obtaining a pipeline centerline point set;

[0138] Inputting the corrosion area contour point set and the pipeline centerline point set into an iterative closest point algorithm;

[0139] Initialize the rigid body transformation matrix as the current rigid body transformation matrix;

[0140] According to the current rigid body transformation matrix, the corrosion area contour point set is transformed into the coordinate system corresponding to the pipeline centerline point set to obtain the transformed corrosion area contour point set;

[0141] Based on each point in the transformed corrosion area contour point set, find the closest point in the pipeline centerline point set to construct an associated point pair;

[0142] Based on the least squares method, the optimal transformation matrix is ​​calculated for the associated point pairs with the goal of minimizing the distance error between the transformed corrosion area contour point set and the pipeline centerline point set;

[0143] If the distance error is greater than a preset error threshold, the current rigid body transformation matrix is ​​updated with the optimal transformation matrix and the iteration continues; otherwise, the iteration is terminated to determine the final optimal transformation matrix;

[0144] Based on the final optimal transformation matrix, generate the registered corrosion probability distribution map;

[0145] Based on the preset corrosion probability threshold, the corrosion excess area is extracted from the registered corrosion probability distribution map, and the pipeline stress concentration area is determined through finite element analysis based on the pipeline design drawings; the corrosion excess area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the two areas, which is determined as the high-risk corrosion area of ​​the pipeline.

[0146] Binarize the corrosion probability distribution map, set the pixels with probability values ​​greater than a certain threshold to 1, and set the remaining pixels to 0. Through binarization, the corrosion area can be separated from the background. Perform contour extraction on the binarized image to obtain a set of contour points of the corrosion area. Contour extraction can be achieved through edge detection algorithms, such as the Canny algorithm or the Sobel algorithm. At the same time, vectorize the pre-acquired pipeline design drawing, extract the center line of the pipeline, and obtain a set of pipeline centerline points. Vectorization can be achieved through image skeleton extraction algorithms, such as thinning algorithms or distance transformation algorithms.

[0147] The corrosion area contour point set and the pipeline centerline point set are input into the iterative closest point algorithm for point set registration. A rigid body transformation matrix is ​​initialized as the current rigid body transformation matrix. According to the current rigid body transformation matrix, the corrosion area contour point set is transformed to the coordinate system where the pipeline centerline point set is located to obtain the transformed corrosion area contour point set. For each point in the transformed corrosion area contour point set, the nearest point is found in the pipeline centerline point set, and the association relationship between the point pairs is constructed to form an associated point pair.

[0148] Based on the least squares method, the optimal transformation matrix is ​​calculated for the associated point pairs with the goal of minimizing the distance error between the transformed corrosion area contour point set and the pipeline centerline point set. Through the least squares method, an optimal transformation matrix can be solved to minimize the distance error between the transformed corrosion area contour point set and the pipeline centerline point set.

[0149] Determine whether the distance error is greater than the preset error threshold. If it is greater than the threshold, update the current rigid body transformation matrix with the optimal transformation matrix and continue iterative optimization. If the distance error is less than or equal to the threshold, terminate the iteration and determine the final optimal transformation matrix.

[0150] According to the final optimal transformation matrix, the corrosion probability distribution map is registered to generate a registered corrosion probability distribution map, which is aligned with the pipeline position and direction in the pipeline design drawing.

[0151] Based on the preset corrosion probability threshold, the corrosion excess area is extracted from the registered corrosion probability distribution map. The corrosion excess area refers to the area where the corrosion probability value exceeds the preset threshold, indicating an area with a high degree of corrosion.

[0152] According to the pipeline design drawings, the stress concentration area of ​​the pipeline is determined through finite element analysis. Finite element analysis is a numerical simulation method that discretizes the pipeline structure into a finite number of units and applies loads and constraints to each unit to calculate the stress distribution of the pipeline. The stress concentration area refers to the area with higher stress values, indicating that the pipeline is prone to stress damage in this area.

[0153] Superimpose the corrosion limit area and the pipeline stress concentration area to find the overlap of the two areas. The overlap is the high-risk corrosion area of ​​the pipeline, indicating that there is both severe corrosion and stress failure. These high-risk corrosion areas need to be paid special attention and repaired to ensure the safe operation of the pipeline.

[0154] In this embodiment, the corrosion area is accurately segmented from the background through binarization processing and contour extraction, thereby improving the accuracy of corrosion area identification; the corrosion area is accurately aligned with the pipeline design drawing using an iterative nearest point algorithm to ensure accurate alignment of the corrosion detection results with the actual pipeline position and direction; by superimposing the corrosion excess area and the pipeline stress concentration area, the high-risk corrosion area of ​​the pipeline is identified to provide accurate guidance for key maintenance; and high-risk areas that are both severely corroded and prone to stress damage are identified to ensure safe operation of the pipeline.

[0155] S103. Deploy a self-calibrated acoustic emission sensor array and a self-calibrated ultrasonic guided wave sensor array around the high-risk corrosion area, obtain sensor corrosion signals, perform variational modal decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, construct a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether leakage occurs in the high-risk corrosion area.

[0156] The corrected sparsified feature matrix is ​​input into the multi-core extreme learning machine discriminator to determine whether leakage occurs in high-risk corrosion areas. The multi-core extreme learning machine is a fast learning neural network model that uses multiple kernel functions to perform nonlinear transformations on input features, and then directly calculates the output weights through randomly generated hidden layer weights and biases to achieve fast learning and classification.

[0157] In the training phase of the multi-core extreme learning machine discriminator, the discriminator is trained using the high-risk corrosion area data where leakage is known. The corrected sparse feature matrix is ​​used as input and the corresponding leakage state is used as output to train the parameters of the discriminator.

[0158] In the discrimination stage, the corrected sparse feature matrix of the high-risk corrosion area to be tested is input into the trained multi-core extreme learning machine discriminator. The discriminator classifies the input features through the learned model and outputs the judgment result of whether the high-risk corrosion area has leaked. If the discriminator outputs leakage, it indicates that the high-risk corrosion area has a leakage risk and needs to be repaired and treated in time; if the discriminator outputs no leakage, it indicates that the high-risk corrosion area is temporarily safe and can continue to be monitored.

[0159] In this embodiment, by deploying self-calibration acoustic emission and ultrasonic guided wave sensor arrays, real-time monitoring of high-risk corrosion areas is achieved, and corrosion signals are captured in time; variational mode decomposition and sparse characterization are used to accurately extract intrinsic function components related to corrosion characteristics, thereby improving the accuracy of signal processing; the acoustic dispersion characteristics of pipeline materials are combined to construct a dispersion compensation matrix, correct signal characteristics, and eliminate errors in the propagation process; and a multi-core extreme learning machine discriminator is used to automatically determine whether a leak occurs in a high-risk corrosion area, thereby improving the intelligence and reliability of detection.

[0160] In an optional embodiment, performing variational mode decomposition on the corrosion signal of the sensor, extracting the intrinsic function components related to the corrosion characteristics, performing sparse characterization on the intrinsic function components, building a dispersion compensation matrix in combination with the acoustic dispersion characteristics of the pipeline material, determining a corrected sparse feature matrix, and inputting it into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in a high-risk corrosion area includes:

[0161] Performing variational mode decomposition on the sensor corrosion signal, decomposing the sensor corrosion signal into multiple eigenfunction components, calculating the center frequency and bandwidth of each eigenfunction component, and based on a preset frequency range and a preset bandwidth threshold, screening the eigenfunction components within the frequency range and with a bandwidth less than the bandwidth threshold, and determining the corrosion characteristic eigenfunction components;

[0162] Performing time-frequency analysis on the corrosion characteristic intrinsic function components to obtain a time-frequency representation matrix, and reducing the dimension of the time-frequency representation matrix using singular value decomposition to obtain a reduced-dimensional feature matrix;

[0163] Input the reduced dimension feature matrix into the dictionary learning algorithm, based on the preset initial dictionary matrix and the preset initial sparse coefficient matrix, by fixing the initial sparse coefficient matrix, combined with minimizing the reconstruction error, update the initial dictionary matrix to obtain a new dictionary matrix, fix the new dictionary matrix, with the goal of minimizing the weighted sum of the reconstruction error and the sparsity, with the sum of the L1 norms of the atoms in the new dictionary matrix equal to 1 as a constraint, solve and update the initial sparse coefficient matrix through a preset orthogonal matching pursuit algorithm, and obtain a new sparse coefficient matrix; fix and update in turn until the preset number of iterations is reached, and determine the final sparse coefficient matrix;

[0164] Based on the pre-acquired acoustic dispersion curve of the pipeline material, the least square method is used to fit the dispersion curve parameters, determine the phase delay and amplitude attenuation coefficient at different frequencies, and construct the dispersion compensation matrix;

[0165] The dispersion compensation matrix is ​​used to perform dispersion compensation on the final sparse coefficient matrix, a compensated sparse feature matrix is ​​determined, and the compensated sparse feature matrix is ​​normalized to obtain a corrected sparse feature matrix.

[0166] The corrosion signal collected by the sensor is subjected to variational mode decomposition. Variational mode decomposition is an adaptive signal decomposition method that can decompose complex non-stationary signals into multiple eigenfunction components. The sensor corrosion signal is decomposed into several eigenfunction components through variational mode decomposition. For each eigenfunction component, its center frequency and bandwidth are calculated. The center frequency represents the main frequency component of the eigenfunction component, and the bandwidth represents the frequency range of the eigenfunction component. According to the preset frequency range and the preset bandwidth threshold, the eigenfunction components within the frequency range and with a bandwidth less than the bandwidth threshold are screened out, and these components are determined as corrosion characteristic eigenfunction components. These corrosion characteristic eigenfunction components contain the key characteristic information of the corrosion signal.

[0167] Time-frequency analysis is performed on the eigenfunction components of the corrosion characteristics. Time-frequency analysis is a method for analyzing the joint distribution of signals in the time and frequency domains, which can reveal the time-frequency characteristics of the signal. By performing time-frequency analysis on the eigenfunction components of the corrosion characteristics, a time-frequency representation matrix is ​​obtained. The time-frequency representation matrix describes the energy distribution of the signal at different times and frequencies. In order to reduce the dimension of the time-frequency representation matrix, singular value decomposition is used to reduce the dimension of the time-frequency representation matrix. Singular value decomposition is a matrix decomposition method that can decompose a matrix into the product of singular values ​​and singular vectors. Through singular value decomposition, a reduced-dimensional feature matrix is ​​obtained, which retains the main feature information of the time-frequency representation matrix.

[0168] The reduced feature matrix is ​​input into the dictionary learning algorithm. Dictionary learning is a sparse representation method that represents the signal by learning an overcomplete dictionary and sparse coefficients. First, a preset initial dictionary matrix and a preset initial sparse coefficient matrix are set. The initial sparse coefficient matrix is ​​fixed, and the initial dictionary matrix is ​​updated with the goal of minimizing the reconstruction error to obtain a new dictionary matrix. The reconstruction error represents the difference between the original signal and the product of the dictionary and the sparse coefficients. Next, the new dictionary matrix is ​​fixed, with the goal of minimizing the weighted sum of the reconstruction error and the sparsity, and the constraint that the sum of the L1 norms of the atoms in the new dictionary matrix is ​​equal to 1. The initial sparse coefficient matrix is ​​solved and updated through the preset orthogonal matching pursuit algorithm to obtain a new sparse coefficient matrix. Sparsity represents the number of non-zero elements in the sparse coefficients. The dictionary matrix and the sparse coefficient matrix are fixed and updated in turn until the preset number of iterations is reached to determine the final sparse coefficient matrix.

[0169] Based on the pre-acquired acoustic dispersion curve of the pipeline material, the least square method is used to fit the dispersion curve parameters. The dispersion curve describes the phase delay and amplitude attenuation at different frequencies when the sound wave propagates in the pipeline material. Through the least square method fitting, the phase delay and amplitude attenuation coefficients at different frequencies are determined, and the dispersion compensation matrix is ​​constructed. The dispersion compensation matrix is ​​used to compensate for the dispersion effect when the sound wave propagates in the pipeline material.

[0170] The dispersion compensation matrix is ​​used to perform dispersion compensation on the final sparse coefficient matrix. The purpose of dispersion compensation is to eliminate the dispersion effect of sound waves when propagating in the pipeline material and restore the original corrosion signal characteristics. Through dispersion compensation, the compensated sparse feature matrix is ​​determined. The compensated sparse feature matrix is ​​normalized so that the element value range in the feature matrix is ​​between 0 and 1, and the corrected sparse feature matrix is ​​obtained. The corrected sparse feature matrix represents the final feature representation of the sensor corrosion signal.

[0171] In this embodiment, the complex corrosion signal is decomposed into multiple eigenfunction components through variational mode decomposition, and the key features of the corrosion signal are accurately extracted to ensure the in-depth analysis of the corrosion signal; the corrosion feature eigenfunction components are subjected to time-frequency analysis and dimensionality reduction processing to extract and retain the main feature information of the signal and enhance the recognizability of the signal features; the sparse coefficient matrix is ​​optimized and obtained through the dictionary learning algorithm to provide an efficient signal representation method, reduce redundant information, and improve signal processing efficiency; the dispersion compensation matrix is ​​used to eliminate the dispersion effect of sound waves propagating in the pipeline material, restore the original corrosion signal features, and improve the accuracy of the corrosion signal.

[0172] Figure 2 FIG. 1 is a schematic diagram of a pressure pipeline corrosion online monitoring system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0173] The first unit is used to obtain the three-dimensional point cloud data of the inner wall of the pressure pipeline, obtain the downsampled point cloud data through the octree filtering algorithm, extract the geometric feature descriptor of the inner wall of the pipeline, construct the geometric feature scattering matrix, input the geometric feature scattering matrix into the pre-trained multi-scale fusion twin network, learn the feature differences of different areas of the inner wall of the pipeline through the contrast loss function, and generate the corrosion probability distribution map of each area of ​​the inner wall of the pipeline;

[0174] The second unit is used to align the corrosion probability distribution map with the pre-acquired pipeline design map based on an iterative closest point algorithm, extract the corrosion limit area from the corrosion probability distribution map based on a preset corrosion probability threshold, determine the pipeline stress concentration area based on the pipeline design map, perform superposition analysis on the corrosion limit area and the pipeline stress concentration area, obtain the overlapping part of the area, and determine the high-risk corrosion area;

[0175] The third unit is used to deploy a self-calibration acoustic emission sensor array and a self-calibration ultrasonic guided wave sensor array around the high-risk corrosion area, obtain sensor corrosion signals, perform variational mode decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, construct a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in the high-risk corrosion area.

[0176] According to a third aspect of the embodiments of the present invention,

[0177] An electronic device is provided, comprising:

[0178] processor;

[0179] a memory for storing processor-executable instructions;

[0180] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0181] A fourth aspect of the embodiments of the present invention is:

[0182] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0183] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pressure pipeline corrosion online monitoring method, characterized in that: include: The three-dimensional point cloud data of the inner wall of the pressure pipeline is obtained, and the downsampled point cloud data is obtained through the octree filtering algorithm, and the geometric feature descriptor of the inner wall of the pipeline is extracted, and a geometric feature scattering matrix is ​​constructed. The geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network, and the feature differences of different areas of the inner wall of the pipeline are learned through the contrast loss function, and the corrosion probability distribution map of each area of ​​the inner wall of the pipeline is generated; Based on the iterative closest point algorithm, the corrosion probability distribution map is aligned with the pre-acquired pipeline design map, and based on the preset corrosion probability threshold, the corrosion limit area is extracted from the corrosion probability distribution map, and based on the pipeline design map, the pipeline stress concentration area is determined, and the corrosion limit area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the area, and the high-risk corrosion area is determined; Deploy a self-calibrated acoustic emission sensor array and a self-calibrated ultrasonic guided wave sensor array around the high-risk corrosion area, obtain sensor corrosion signals, perform variational modal decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, build a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether leakage occurs in the high-risk corrosion area; Obtain the three-dimensional point cloud data of the inner wall of the pressure pipe, and obtain the downsampled point cloud data through the octree filtering algorithm, including: Acquire three-dimensional point cloud data of the inner wall of the pressure pipe, and divide the three-dimensional point cloud data into cubic voxels of equal size according to a preset number of voxels, with the point cloud area corresponding to the three-dimensional point cloud data serving as a root node, and each of the cubic voxels corresponding to a child node; Traverse each child node from top to bottom to determine the first operation current node. If the number of point clouds in the first operation current node is greater than a preset node point cloud threshold, divide the first operation current node into eight child nodes. Otherwise, the first operation current node is used as a leaf node, and the recursive division is repeated until the number of point clouds in all leaf nodes is less than or equal to the node point cloud threshold, thereby generating an octree. Traverse each non-leaf node of the octree from bottom to top to determine the second operation current node, and if all child nodes of the second operation current node are empty, mark the second operation current node as an empty node, and delete all child nodes corresponding to the second operation current node; Traversing all non-empty leaf nodes of the octree, determining a third operation current node, calculating an average density of a point cloud in the third operation current node, determining a node point cloud density, and deleting the third operation current node if the node point cloud density of the third operation current node is less than a preset density threshold; After the deletion operation, a valid non-empty leaf node is obtained, and all three-dimensional point cloud data contained in the valid non-empty leaf node is obtained to obtain the downsampled point cloud data.

2. The method according to claim 1, characterized in that Extract the geometric feature descriptor of the inner wall of the pipeline and construct the geometric feature scattering matrix including: Based on the downsampled point cloud data, a topological space is determined, and for each sampling point in the topological space, a plurality of spherical neighborhoods of different scales are set according to a preset initial spherical radius and a preset radius increasing rate to construct a multi-scale neighborhood; In a multi-scale neighborhood, a spherical neighborhood sub-point cloud of multiple scales is determined, a simplicial complex is constructed based on the Alpha shape algorithm, and a homology group at a corresponding scale is obtained by calculating the number of connected components and the Betti number of the simplicial complex; a binary decision vector of the multi-scale neighborhood is constructed, and an optimal feature extraction scale set is determined by iterative search through a particle swarm algorithm with the goal of minimizing the number of generators corresponding to the homology group at each selected scale; For the spherical neighborhood sub-point cloud corresponding to each scale in the optimal feature extraction scale set, the normal vector is calculated by the preset principal component analysis method, the curvature at the sampling point is calculated by eigenvalue decomposition, and the shape index is calculated based on the point cloud density distribution to obtain a multi-scale geometric feature descriptor; The multi-scale geometric feature descriptor is mapped to a high-dimensional manifold space. Through a local linear embedding algorithm, the multi-scale geometric feature descriptor is used as input to learn the mapping relationship between the high-dimensional manifold space and the original feature space, and the coordinates of the multi-scale geometric feature descriptor in the high-dimensional manifold space are obtained; in the high-dimensional manifold space, the scattering similarity between the multi-scale geometric feature descriptors is calculated through a Gaussian kernel function, and a geometric feature scattering matrix is ​​constructed.

3. The method according to claim 2, characterized in that The geometric feature scattering matrix is ​​input into the pre-trained multi-scale fusion twin network, and the feature differences of different areas of the inner wall of the pipeline are learned through the contrast loss function to generate the corrosion probability distribution map of each area of ​​the inner wall of the pipeline, including: Inputting the geometric feature scattering matrix into a pre-trained multi-scale fusion twin network, wherein the multi-scale fusion twin network is composed of two convolutional sub-networks with the same structure and shared parameters, and performing feature extraction on the rows and columns of the scattering matrix respectively; In each of the convolutional subnetworks, a stack of multiple convolutional layers and pooling layers is used to perform multi-scale feature extraction on the geometric feature scattering matrix, and local structural features of different scales are captured through different receptive fields of different convolutional layers. The local structural features of different scales are adaptively fused in combination with the attention mechanism to obtain row feature maps and column feature maps, and the similarity between the row feature maps and the column feature maps is measured based on the contrast loss function; By back propagation and gradient descent algorithm, the contrast loss function is minimized, the parameters of the multi-scale fusion twin network are adjusted, and the training is repeated until a preset number of iterations is reached; Based on the trained multi-scale fusion twin network, dense sampling is performed on the inner wall of the pipeline, the inner wall of the pipeline is divided into local areas of equal size, and feature vectors of the local areas are generated; The average Euclidean distance between each of the characteristic vectors and the characteristic vector of the known corrosion area is calculated to determine the corrosion probability of the local area, and generate a corrosion probability distribution map of each area on the inner wall of the pipeline.

4. The method according to claim 3, characterized in that Also includes: The contrast loss function is formulated as follows: ； in, L represents the contrast loss function, N represents the total number of sample pairs, i represents the index of the sample pair, y i represents a binary label, tanh() represents the hyperbolic tangent function, α represents the distance scaling factor, x i 1 Indicates i The first eigenvector in the sample pairs, x i 2 Indicates i The second eigenvector in the sample pairs, d () represents the distance metric function, max() represents the maximum value function, m Represents the distance sensitivity threshold, exp() represents the exponential function, β Represents the exponential scaling factor.

5. The method according to claim 1, characterized in that Based on the iterative closest point algorithm, the corrosion probability distribution map is aligned with the pre-acquired pipeline design map. Based on the preset corrosion probability threshold, the corrosion limit area is extracted from the corrosion probability distribution map. Based on the pipeline design map, the pipeline stress concentration area is determined. The corrosion limit area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the area. The high-risk corrosion area is determined to include: Binarizing the corrosion probability distribution map, extracting the corrosion area contour, and obtaining a corrosion area contour point set; vectorizing the pre-acquired pipeline design drawing, extracting the pipeline centerline, and obtaining a pipeline centerline point set; Inputting the corrosion area contour point set and the pipeline centerline point set into an iterative closest point algorithm; Initialize the rigid body transformation matrix as the current rigid body transformation matrix; According to the current rigid body transformation matrix, the corrosion area contour point set is transformed into the coordinate system corresponding to the pipeline centerline point set to obtain the transformed corrosion area contour point set; Based on each point in the transformed corrosion area contour point set, find the closest point in the pipeline centerline point set to construct an associated point pair; Based on the least squares method, the optimal transformation matrix is ​​calculated for the associated point pairs with the goal of minimizing the distance error between the transformed corrosion area contour point set and the pipeline centerline point set; If the distance error is greater than a preset error threshold, the current rigid body transformation matrix is ​​updated with the optimal transformation matrix and the iteration continues; otherwise, the iteration is terminated to determine the final optimal transformation matrix; Based on the final optimal transformation matrix, generate the registered corrosion probability distribution map; Based on the preset corrosion probability threshold, the corrosion excess area is extracted from the registered corrosion probability distribution map, and the pipeline stress concentration area is determined through finite element analysis based on the pipeline design drawings; the corrosion excess area and the pipeline stress concentration area are superimposed and analyzed to obtain the overlapping part of the two areas, which is determined as the high-risk corrosion area of ​​the pipeline.

6. The method according to claim 1, characterized in that The sensor corrosion signal is subjected to variational mode decomposition, the intrinsic function components related to the corrosion characteristics are extracted, the intrinsic function components are sparsely characterized, a dispersion compensation matrix is ​​constructed in combination with the acoustic dispersion characteristics of the pipeline material, and the corrected sparse feature matrix is ​​determined and input into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in a high-risk corrosion area, including: Performing variational mode decomposition on the sensor corrosion signal, decomposing the sensor corrosion signal into multiple eigenfunction components, calculating the center frequency and bandwidth of each eigenfunction component, and based on a preset frequency range and a preset bandwidth threshold, screening the eigenfunction components within the frequency range and with a bandwidth less than the bandwidth threshold, and determining the corrosion characteristic eigenfunction components; Performing time-frequency analysis on the corrosion characteristic intrinsic function components to obtain a time-frequency representation matrix, and reducing the dimension of the time-frequency representation matrix using singular value decomposition to obtain a reduced-dimensional feature matrix; Input the reduced dimension feature matrix into the dictionary learning algorithm, based on the preset initial dictionary matrix and the preset initial sparse coefficient matrix, by fixing the initial sparse coefficient matrix, combined with minimizing the reconstruction error, update the initial dictionary matrix to obtain a new dictionary matrix, fix the new dictionary matrix, with the goal of minimizing the weighted sum of the reconstruction error and the sparsity, with the sum of the L1 norms of the atoms in the new dictionary matrix equal to 1 as a constraint, solve and update the initial sparse coefficient matrix through a preset orthogonal matching pursuit algorithm, and obtain a new sparse coefficient matrix; fix and update in turn until the preset number of iterations is reached, and determine the final sparse coefficient matrix; Based on the pre-acquired acoustic dispersion curve of the pipeline material, the least square method is used to fit the dispersion curve parameters, determine the phase delay and amplitude attenuation coefficient at different frequencies, and construct the dispersion compensation matrix; The dispersion compensation matrix is ​​used to perform dispersion compensation on the final sparse coefficient matrix, a compensated sparse feature matrix is ​​determined, and the compensated sparse feature matrix is ​​normalized to obtain a corrected sparse feature matrix.

7. A pressure pipeline corrosion online monitoring system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the three-dimensional point cloud data of the inner wall of the pressure pipeline, obtain the downsampled point cloud data through the octree filtering algorithm, extract the geometric feature descriptor of the inner wall of the pipeline, construct the geometric feature scattering matrix, input the geometric feature scattering matrix into the pre-trained multi-scale fusion twin network, learn the feature differences of different areas of the inner wall of the pipeline through the contrast loss function, and generate the corrosion probability distribution map of each area of ​​the inner wall of the pipeline; The second unit is used to align the corrosion probability distribution map with the pre-acquired pipeline design map based on an iterative closest point algorithm, extract the corrosion limit area from the corrosion probability distribution map based on a preset corrosion probability threshold, determine the pipeline stress concentration area based on the pipeline design map, perform superposition analysis on the corrosion limit area and the pipeline stress concentration area, obtain the overlapping part of the area, and determine the high-risk corrosion area; The third unit is used to deploy a self-calibration acoustic emission sensor array and a self-calibration ultrasonic guided wave sensor array around the high-risk corrosion area, obtain sensor corrosion signals, perform variational mode decomposition on the sensor corrosion signals, extract intrinsic function components related to corrosion characteristics, perform sparse characterization on the intrinsic function components, construct a dispersion compensation matrix based on the acoustic dispersion characteristics of the pipeline material, determine the corrected sparse feature matrix, and input it into a multi-core extreme learning machine discriminator to determine whether a leakage occurs in the high-risk corrosion area.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images