A pipeline topology relationship construction method based on a graph neural network

Abstract

This invention discloses a method for constructing pipeline topology relationships based on graph neural networks, comprising the following steps: collecting and preprocessing the original spatial data of the pipeline system; generating node feature matrices and edge feature matrices using feature extraction; constructing a pipeline graph structure; performing index alignment and size verification to generate a pipeline graph input matrix; inputting the matrix into an improved EvolveGCN model, introducing a graph attention-controlled evolutionary mechanism to obtain the temporal embedding feature matrix of each pipeline node object; generating a predicted pipeline topology connection set; performing topology consistency verification to obtain the final pipeline topology relationship structure; and forming a pipeline topology relationship model file to achieve automated modeling and optimization of the pipeline system, thereby improving the recognition accuracy and modeling efficiency of the pipeline topology structure.

Description

Technical Field

[0001] This invention relates to the field of spatial data modeling, and in particular to a method for constructing pipeline topology relationships based on graph neural networks. Background Technology

[0002] With the rapid development of urban infrastructure and industrial production systems, pipeline networks have been widely used in municipal water supply and drainage, gas transmission and distribution, power cooling, petrochemicals, and building information modeling (BIM) systems. The topology of a pipeline system is the core foundation for describing the spatial connections between network nodes and pipe segments, and its accuracy directly affects subsequent network analysis, fluid simulation, maintenance planning, and digital twin visualization. However, existing pipeline topology construction technologies still generally suffer from low topology recognition efficiency, poor accuracy in connection relationship reasoning, and insufficient spatial structure representation capabilities, making it difficult to meet the intelligent modeling needs of complex scenarios.

[0003] Traditional pipeline topology construction primarily relies on geometric rule matching and manual logical judgment. For example, in CAD, BIM, or GIS systems, the connection relationship between nodes and pipe segments is typically identified through node coordinate distance, direction vector similarity, or manually defined connection rules. While these methods can achieve basic topology construction in scenarios with simple structures and complete data, they often encounter problems such as missing connections, incorrect matching, or duplicate connections when noisy data, spatial offsets, complex bend connections, or multi-layered structures are present. Furthermore, manual rule-based methods lack generalization ability when dealing with different materials, different diameters, or the fusion of multi-source heterogeneous data, and cannot accurately model the potential high-dimensional correlation information between nodes, resulting in deviations between the generated topology and the actual physical pipeline network. Summary of the Invention

[0004] One objective of this invention is to propose a pipeline topology construction method based on graph neural networks. This invention fully utilizes the graph attention-controlled evolutionary mechanism of the improved EvolveGCN model to perform deep feature extraction and topology inference on node and pipe segment data in the pipeline system. Simultaneously, by establishing node feature matrices and edge feature matrices, the geometric attributes, physical attributes, and spatial relationship parameters of pipeline nodes are uniformly modeled, and a dynamic graph structure learning method is introduced to achieve automatic topology identification and connectivity prediction of pipeline systems in complex scenarios.

[0005] A pipeline topology construction method based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0006] Collect and preprocess the raw spatial data of the pipeline system;

[0007] Feature extraction is performed on the preprocessed raw spatial data to generate node feature matrices and edge feature matrices;

[0008] Based on the node feature matrix and edge feature matrix, and according to the spatial proximity relationship between pipe node objects and pipe segment objects, a pipe graph structure is constructed.

[0009] Perform index alignment and dimension verification on the piping diagram structure to generate the piping diagram input matrix;

[0010] The pipeline graph input matrix is ​​fed into the improved EvolveGCN model, and a graph attention control evolution mechanism is introduced to obtain the temporal embedding feature matrix of each pipeline node object.

[0011] Calculate the connection confidence between node pairs based on the temporal embedding feature matrix of each pipeline node object, and generate a predicted pipeline topology connection set.

[0012] Perform topology consistency verification on the predicted pipeline topology connection set to obtain the final pipeline topology relationship structure;

[0013] The final generated pipeline topology structure is then integrated and encapsulated into a pipeline topology model file.

[0014] Optionally, the raw spatial data includes structural information from lidar scanning, building information modeling, computer-aided design drawings, and geographic information systems, and the preprocessing includes coordinate unification, format conversion, and noise filtering.

[0015] Optionally, the generation of the node feature matrix and the edge feature matrix specifically includes:

[0016] The geometric and physical attributes of each pipe node object are extracted from the preprocessed raw spatial data to form a pipe node feature vector. The geometric attributes include the coordinates and orientation vectors of the pipe node object in three-dimensional space, and the physical attributes include the pipe diameter, material type, and node adjacency of the pipe node object.

[0017] Arrange the feature vectors of the pipeline nodes in order of node number to form the pipeline node feature matrix;

[0018] The spatial relationship parameters of each pipe segment object are extracted from the preprocessed raw spatial data to form a pipe segment feature vector. The spatial relationship parameters include the length of the pipe segment object, the directional angle, and the spatial distance between the two pipe node objects.

[0019] The feature vectors of the pipeline segments are arranged sequentially according to the segment numbers to form the pipeline segment feature matrix;

[0020] The eigenvectors of the pipe node feature matrix and the eigenvectors of the pipe segment feature matrix are sequentially combined vertically to form the node feature matrix and the edge feature matrix.

[0021] Optionally, the construction of the pipeline diagram structure specifically includes:

[0022] Based on the node feature matrix and the edge feature matrix, a set of pipe nodes and a set of pipe connections are defined. The set of pipe nodes is defined as taking each pipe node object with independent spatial coordinates and attribute information as an element of the set according to the elements in the node feature matrix. The set of pipe connections is defined as taking each pipe segment object containing a start node, an end node and spatial relationship parameters as an element of the set according to the elements in the edge feature matrix.

[0023] The spatial distance between two pipeline node objects is calculated using the three-dimensional spatial coordinate information in the node feature matrix, and the attribute similarity between nodes is calculated based on the physical attribute parameters contained in the node feature matrix.

[0024] By combining the spatial relationship parameters in the edge feature matrix, a spatial distance threshold and an attribute similarity threshold are set. When the spatial distance between the node objects of two pipelines is less than the set distance threshold and the attribute similarity is greater than the set similarity threshold, the pipeline node object pair is determined to meet the potential pipeline connection conditions, and a candidate pipeline connection set is formed.

[0025] The connection weights are calculated based on the spatial distance, attribute similarity, and pipe segment feature information in the edge feature matrix of each node pair in the candidate pipeline connection set. A weighted adjacency matrix is ​​constructed by assigning connection weights to each pair of candidate nodes and arranging them according to the node index.

[0026] A complete pipeline graph structure is generated by combining the node feature matrix, edge feature matrix, pipeline node set, pipeline connection set, and weighted adjacency matrix.

[0027] Optionally, the generation of the pipeline diagram input matrix specifically includes:

[0028] By combining the weighted adjacency matrix of the pipeline graph structure with the node feature matrix, a unified pipeline node numbering order is generated, and a pipeline node index mapping relationship is established. The pipeline node numbering order refers to the unique identification and sequential arrangement of all pipeline node objects in the pipeline graph structure. The pipeline node index mapping relationship is established by establishing a one-to-one correspondence between the weighted adjacency matrix and the node feature matrix in rows, columns, and node numbers according to the pipeline node numbering order.

[0029] Based on the pipeline node index mapping relationship, an index alignment operation is performed on the weighted adjacency matrix and the node feature matrix to obtain the index-aligned weighted adjacency matrix and the index-aligned node feature matrix;

[0030] The size of the index-aligned weighted adjacency matrix and the index-aligned node feature matrix is ​​validated to generate the size-validated weighted adjacency matrix and the size-validated node feature matrix.

[0031] The weighted adjacency matrix after size verification is combined and encapsulated with the node feature matrix after size verification to generate the pipeline graph input matrix.

[0032] Optionally, obtaining the temporal embedding feature matrix of each pipeline node object specifically includes:

[0033] The pipeline graph input matrix is ​​input into the improved EvolveGCN model. The improved EvolveGCN model includes a pipeline feature dynamic input module, a dynamic graph weight evolution module, a graph temporal feature propagation module, and a topology prediction and adaptive optimization module. The pipeline feature dynamic input module performs time-step encoding and normalization on the nodes of the pipeline to form a node feature sequence. The dynamic graph weight evolution module calculates the convolution weight matrix using a graph attention-controlled evolution mechanism. The graph temporal feature propagation module performs graph convolution operations on the node feature sequence, and performs weighted aggregation and nonlinear transformation to generate a node hidden representation matrix. The topology prediction and adaptive optimization module calculates the node pair similarity based on the node hidden representation matrix at each time step and obtains the connection strength matrix, forming a temporal embedding representation and uniformly aggregating it to generate a temporal embedding feature matrix.

[0034] In the pipeline feature dynamic input module, the nodes of the pipeline diagram input matrix are encoded and normalized by time steps to form a node feature sequence organized by time steps.

[0035] In the dynamic graph weight evolution module, a graph attention control evolution mechanism is introduced to perform weighted mapping on the node feature sequences and generate the convolution weight matrix for the current time step.

[0036] In the graph time-series feature propagation module, based on the convolution weight matrix, graph convolution operation is performed on the node feature sequence of the current time step to generate the node hidden representation matrix of the current time step. The node hidden representation matrix is ​​obtained by linearly transforming the node feature sequence of the current time step with the convolution weight matrix, combining the node connection relationship in the pipeline graph structure, aggregating the information of adjacent nodes on each node, and processing it with a nonlinear activation function.

[0037] In the topology prediction and adaptive optimization module, the similarity of node pairs is calculated and the connection strength is adjusted based on the node hidden representations at each time step to form a temporal embedding representation. The temporal embedding representation is obtained by calculating the feature similarity between all node hidden representations at each time step to obtain the connection strength matrix of node pairs, and then performing weighted fusion based on the connection strength of the connection strength matrix.

[0038] The temporal embedding representations of each time step are uniformly aggregated to generate a temporal embedding feature matrix.

[0039] Optionally, the generation of the predicted pipeline topology connection set specifically includes:

[0040] The embedding vectors of the temporal embedding feature matrix are normalized, and a symmetric scoring operation is performed on any node pair to form a feature similarity matrix.

[0041] The feature similarity matrix and the weighted adjacency matrix are weighted and fused according to a preset structural weight ratio to obtain the fused relevance matrix;

[0042] A non-linear mapping is performed on the fusion correlation matrix to generate a connection confidence matrix with values ​​between zero and one.

[0043] Based on the preset connection confidence threshold, node pairs with a confidence level greater than the connection confidence threshold in the connection confidence matrix are selected and marked as potential connections. All node pairs that meet the conditions are combined to form a predicted pipeline topology connection set.

[0044] Optionally, obtaining the final pipeline topology structure specifically includes:

[0045] Based on the spatial coordinates of pipe nodes, the direction vectors of pipe nodes, and the pipe diameter parameters of pipe nodes in the predicted pipe topology connection set, a topology consistency check is performed. The topology consistency check includes pipe direction constraint check, pipe diameter matching constraint check, and engineering logic constraint check.

[0046] In the pipeline direction constraint verification, for each connection relationship, a direction vector is calculated based on the spatial coordinates of the pipeline node. The direction vector is calculated based on the difference in spatial coordinates between the starting pipeline node and the ending pipeline node in the connection relationship. The direction vector is then compared with the direction vector of the corresponding pipeline node. When the two direction vectors are less than a preset threshold, it is determined that the direction constraint is satisfied; otherwise, it is recorded as the direction constraint is not satisfied.

[0047] In the process of performing pipe diameter matching constraint verification, the difference in pipe diameter parameters of the pipe nodes at both ends of each connection is calculated and compared with the preset pipe diameter tolerance threshold. When the pipe diameter difference is less than the pipe diameter tolerance threshold, the pipe diameter matching constraint is determined to be satisfied; otherwise, the connection is excluded.

[0048] In the execution of engineering logic constraint verification, based on the engineering design rules of the pipeline network system, the maximum number of allowed connections for each pipeline node is set, and a prohibited connection set is defined. Node pairs that do not conform to the connection rules are eliminated. The prohibited connection set is defined by obtaining the type of each pipeline node according to the pipeline network engineering design rules, and recording the combination of node types that do not meet the requirements according to the engineering logic provisions.

[0049] Based on pipeline direction constraint verification, pipe diameter matching constraint verification, and engineering logic constraint verification, each connection relationship is filtered and sorted according to the corresponding confidence value in the connection confidence matrix to form the final pipeline connection set.

[0050] The final pipeline adjacency matrix is ​​generated based on the final pipeline connection set. The final pipeline connection set and the final pipeline adjacency matrix are combined to output the final pipeline topology structure. The final pipeline adjacency matrix is ​​obtained by establishing a unified number index for all pipeline nodes according to the node pair information in the final pipeline connection set, assigning a value of one to the corresponding position of node pairs with connection relationship in the adjacency matrix, and assigning a value of zero to node pairs without connection relationship.

[0051] Optionally, the formation of the pipeline topology model file specifically includes:

[0052] The geometric and physical attribute information in the final pipeline topology structure is integrated with the connection relationship information in the final pipeline connection set to form a pipeline topology attribute table. The pipeline topology attribute table includes the starting pipeline node number, the ending pipeline node number, and the connection attribute information for each connection relationship.

[0053] A corresponding topology connection mapping matrix is ​​generated based on the final pipeline adjacency matrix. The topology connection mapping matrix is ​​formed by mapping pipeline node pairs with connection relationships to connection paths according to the position of the value of one in the adjacency matrix, and by numerical labeling based on the connection strength between nodes.

[0054] The final pipeline connection set, topology connection mapping matrix, and pipeline topology attribute table are encapsulated according to a unified data structure standard to generate a pipeline topology relationship model file. The pipeline topology relationship model file includes pipeline node information, pipeline connection relationships, and pipeline topology attribute information.

[0055] The beneficial effects of this invention are:

[0056] First, this invention employs an improved EvolveGCN model, introducing a graph attention-controlled evolutionary mechanism. This enables the model to adaptively adjust convolutional weights and dynamically evolve node features over time, effectively capturing the high-dimensional dependencies and temporal variations between nodes in the pipeline network structure. This mechanism not only enhances the model's generalization ability and stability but also significantly improves the prediction accuracy of potential connectivity relationships in complex pipeline systems. By combining the graph attention mechanism with a parameter evolution strategy, this invention achieves dynamic weighting of feature importance during feature propagation, making the topology learning process more consistent with the engineering logic and physical constraints of the pipeline system.

[0057] Secondly, this invention generates a high-confidence predicted connection set by calculating the connection confidence between node pairs and integrating structural similarity, spatial proximity, and attribute constraints, significantly improving the accuracy and robustness of pipeline connection identification. To address potential unreasonable connection relationships in the predicted results, this invention further introduces a topology consistency verification mechanism. This mechanism comprehensively considers multi-dimensional conditions such as pipeline direction constraints, pipe diameter matching constraints, and engineering logic constraints to automatically correct and optimize the connection results, ensuring that the generated topology satisfies both mathematical connectivity and practical engineering feasibility.

[0058] Finally, this invention integrates and encapsulates the generated topological relationship structure, outputting a pipeline topology model file containing node information, connection relationships, and topological attributes. This model file adopts a standardized structure, allowing seamless integration with Building Information Modeling (BIM) systems, Geographic Information Systems (GIS), and intelligent pipeline management platforms, enabling visualized topological relationships, structural analysis, and operation and maintenance management. This invention significantly improves upon existing technologies in terms of topology modeling accuracy, automation, and system adaptability, offering fully automated modeling, accurate structural identification, and intelligent topology analysis. It provides an efficient, scalable, and engineering-oriented solution for the digital modeling and intelligent management of complex pipeline systems. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is an overall flowchart of a pipeline topology relationship construction method based on graph neural networks proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the module structure of the improved EvolveGCN model, which is a pipeline topology relationship construction method based on graph neural networks proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1-2 A method for constructing pipeline topology relationships based on graph neural networks includes the following steps:

[0064] Collect and preprocess the raw spatial data of the pipeline system;

[0065] Feature extraction is performed on the preprocessed raw spatial data to generate node feature matrices and edge feature matrices;

[0066] Based on the node feature matrix and edge feature matrix, and according to the spatial proximity relationship between pipe node objects and pipe segment objects, a pipe graph structure is constructed.

[0067] Perform index alignment and dimension verification on the piping diagram structure to generate the piping diagram input matrix;

[0068] The pipeline graph input matrix is ​​fed into the improved EvolveGCN model, and a graph attention control evolution mechanism is introduced to obtain the temporal embedding feature matrix of each pipeline node object.

[0069] Calculate the connection confidence between node pairs based on the temporal embedding feature matrix of each pipeline node object, and generate a predicted pipeline topology connection set.

[0070] Perform topology consistency verification on the predicted pipeline topology connection set to obtain the final pipeline topology relationship structure;

[0071] The final generated pipeline topology structure is then integrated and encapsulated into a pipeline topology model file.

[0072] In this embodiment, the raw spatial data includes structural information from lidar scanning, building information modeling, computer-aided design drawings, and geographic information systems. The preprocessing includes coordinate unification, format conversion, and noise filtering.

[0073] In this embodiment, the generation of the node feature matrix and the edge feature matrix specifically includes:

[0074] The geometric and physical attributes of each pipe node object are extracted from the preprocessed raw spatial data to form a pipe node feature vector. The geometric attributes include the coordinates and orientation vectors of the pipe node object in three-dimensional space, and the physical attributes include the pipe diameter, material type, and node adjacency of the pipe node object.

[0075] Arrange the feature vectors of the pipeline nodes in order of node number to form the pipeline node feature matrix;

[0076] The spatial relationship parameters of each pipe segment object are extracted from the preprocessed raw spatial data to form a pipe segment feature vector. The spatial relationship parameters include the length of the pipe segment object, the directional angle, and the spatial distance between the two pipe node objects.

[0077] The feature vectors of the pipeline segments are arranged sequentially according to the segment numbers to form the pipeline segment feature matrix;

[0078] The eigenvectors of the pipe node feature matrix and the eigenvectors of the pipe segment feature matrix are sequentially combined vertically to form the node feature matrix and the edge feature matrix.

[0079] In this embodiment, the construction of the pipeline diagram structure specifically includes:

[0080] Based on the node feature matrix and the edge feature matrix, a set of pipe nodes and a set of pipe connections are defined. The set of pipe nodes is defined as taking each pipe node object with independent spatial coordinates and attribute information as an element of the set according to the elements in the node feature matrix. The set of pipe connections is defined as taking each pipe segment object containing a start node, an end node and spatial relationship parameters as an element of the set according to the elements in the edge feature matrix.

[0081] The spatial distance between two pipeline node objects is calculated using the three-dimensional spatial coordinate information in the node feature matrix, and the attribute similarity between nodes is calculated based on the physical attribute parameters contained in the node feature matrix.

[0082] By combining the spatial relationship parameters in the edge feature matrix, a spatial distance threshold and an attribute similarity threshold are set. When the spatial distance between the node objects of two pipelines is less than the set distance threshold and the attribute similarity is greater than the set similarity threshold, the pipeline node object pair is determined to meet the potential pipeline connection conditions, and a candidate pipeline connection set is formed.

[0083] The connection weights are calculated based on the spatial distance, attribute similarity, and pipe segment feature information in the edge feature matrix of each node pair in the candidate pipeline connection set. A weighted adjacency matrix is ​​constructed by assigning connection weights to each pair of candidate nodes and arranging them according to the node index.

[0084] A complete pipeline graph structure is generated by combining the node feature matrix, edge feature matrix, pipeline node set, pipeline connection set, and weighted adjacency matrix.

[0085] In this embodiment, the generation of the pipeline diagram input matrix specifically includes:

[0086] By combining the weighted adjacency matrix of the pipeline graph structure with the node feature matrix, a unified pipeline node numbering order is generated, and a pipeline node index mapping relationship is established. The pipeline node numbering order refers to the unique identification and sequential arrangement of all pipeline node objects in the pipeline graph structure. The pipeline node index mapping relationship is established by establishing a one-to-one correspondence between the weighted adjacency matrix and the node feature matrix in rows, columns, and node numbers according to the pipeline node numbering order.

[0087] Based on the pipeline node index mapping relationship, an index alignment operation is performed on the weighted adjacency matrix and the node feature matrix to obtain the index-aligned weighted adjacency matrix and the index-aligned node feature matrix;

[0088] The size of the index-aligned weighted adjacency matrix and the index-aligned node feature matrix is ​​validated to generate the size-validated weighted adjacency matrix and the size-validated node feature matrix.

[0089] The weighted adjacency matrix after size verification is combined and encapsulated with the node feature matrix after size verification to generate the pipeline graph input matrix.

[0090] In this embodiment, obtaining the temporal embedding feature matrix of each pipeline node object specifically includes:

[0091] The pipeline graph input matrix is ​​input into the improved EvolveGCN model. The improved EvolveGCN model includes a pipeline feature dynamic input module, a dynamic graph weight evolution module, a graph temporal feature propagation module, and a topology prediction and adaptive optimization module. The pipeline feature dynamic input module performs time-step encoding and normalization on the nodes of the pipeline to form a node feature sequence. The dynamic graph weight evolution module calculates the convolution weight matrix using a graph attention-controlled evolution mechanism. The graph temporal feature propagation module performs graph convolution operations on the node feature sequence, and performs weighted aggregation and nonlinear transformation to generate a node hidden representation matrix. The topology prediction and adaptive optimization module calculates the node pair similarity based on the node hidden representation matrix at each time step and obtains the connection strength matrix, forming a temporal embedding representation and uniformly aggregating it to generate a temporal embedding feature matrix.

[0092] In the pipeline feature dynamic input module, the nodes of the pipeline diagram input matrix are encoded and normalized by time steps to form a node feature sequence organized by time steps.

[0093] In the dynamic graph weight evolution module, a graph attention control evolution mechanism is introduced to perform weighted mapping on the node feature sequences and generate the convolution weight matrix for the current time step.

[0094] In the graph time-series feature propagation module, based on the convolution weight matrix, graph convolution operation is performed on the node feature sequence of the current time step to generate the node hidden representation matrix of the current time step. The node hidden representation matrix is ​​obtained by linearly transforming the node feature sequence of the current time step with the convolution weight matrix, combining the node connection relationship in the pipeline graph structure, aggregating the information of adjacent nodes on each node, and processing it with a nonlinear activation function.

[0095] In the topology prediction and adaptive optimization module, the similarity of node pairs is calculated and the connection strength is adjusted based on the node hidden representations at each time step to form a temporal embedding representation. The temporal embedding representation is obtained by calculating the feature similarity between all node hidden representations at each time step to obtain the connection strength matrix of node pairs, and then performing weighted fusion based on the connection strength of the connection strength matrix.

[0096] The temporal embedding representations at each time step are uniformly aggregated to generate a temporal embedding feature matrix:

[0097] ;

[0098] in, For temporal embedding feature matrix, For a weighted adjacency matrix, For time steps The feature matrix of the pipeline nodes, The convolution weight matrix is ​​from the previous time step. For the gated vector, It is a non-linear activation function. It is the hyperbolic tangent function. For element-wise multiplication, and For a trainable parameter matrix, For aggregation operators at the node dimension, For matrix vectorization operators, This represents the number of time steps.

[0099] In this embodiment, the generation of the predicted pipeline topology connection set specifically includes:

[0100] The embedding vectors of the temporal embedding feature matrix are normalized, and a symmetric scoring operation is performed on any node pair to form a feature similarity matrix.

[0101] The feature similarity matrix and the weighted adjacency matrix are weighted and fused according to a preset structural weight ratio to obtain the fused relevance matrix;

[0102] A non-linear mapping is performed on the fusion correlation matrix to generate a connection confidence matrix with values ​​between zero and one.

[0103] Based on the preset connection confidence threshold, node pairs with a confidence level greater than the connection confidence threshold in the connection confidence matrix are selected and marked as potential connections. All node pairs that meet the conditions are combined to form a predicted pipeline topology connection set.

[0104] In this embodiment, obtaining the final pipeline topology structure specifically includes:

[0105] Based on the spatial coordinates of pipe nodes, the direction vectors of pipe nodes, and the pipe diameter parameters of pipe nodes in the predicted pipe topology connection set, a topology consistency check is performed. The topology consistency check includes pipe direction constraint check, pipe diameter matching constraint check, and engineering logic constraint check.

[0106] In the pipeline direction constraint verification, for each connection relationship, a direction vector is calculated based on the spatial coordinates of the pipeline node. The direction vector is calculated based on the difference in spatial coordinates between the starting pipeline node and the ending pipeline node in the connection relationship. The direction vector is then compared with the direction vector of the corresponding pipeline node. When the two direction vectors are less than a preset threshold, it is determined that the direction constraint is satisfied; otherwise, it is recorded as the direction constraint is not satisfied.

[0107] In the process of performing pipe diameter matching constraint verification, the difference in pipe diameter parameters of the pipe nodes at both ends of each connection is calculated and compared with the preset pipe diameter tolerance threshold. When the pipe diameter difference is less than the pipe diameter tolerance threshold, the pipe diameter matching constraint is determined to be satisfied; otherwise, the connection is excluded.

[0108] In the execution of engineering logic constraint verification, based on the engineering design rules of the pipeline network system, the maximum number of allowed connections for each pipeline node is set, and a prohibited connection set is defined. Node pairs that do not conform to the connection rules are eliminated. The prohibited connection set is defined by obtaining the type of each pipeline node according to the pipeline network engineering design rules, and recording the combination of node types that do not meet the requirements according to the engineering logic provisions.

[0109] Based on pipeline direction constraint verification, pipe diameter matching constraint verification, and engineering logic constraint verification, each connection relationship is filtered and sorted according to the corresponding confidence value in the connection confidence matrix to form the final pipeline connection set.

[0110] The final pipeline adjacency matrix is ​​generated based on the final pipeline connection set. The final pipeline connection set and the final pipeline adjacency matrix are combined to output the final pipeline topology structure. The final pipeline adjacency matrix is ​​obtained by establishing a unified number index for all pipeline nodes according to the node pair information in the final pipeline connection set, assigning a value of one to the corresponding position of node pairs with connection relationship in the adjacency matrix, and assigning a value of zero to node pairs without connection relationship.

[0111] In this embodiment, the formation of the pipeline topology model file specifically includes:

[0112] The geometric and physical attribute information in the final pipeline topology structure is integrated with the connection relationship information in the final pipeline connection set to form a pipeline topology attribute table. The pipeline topology attribute table includes the starting pipeline node number, the ending pipeline node number, and the connection attribute information for each connection relationship.

[0113] A corresponding topology connection mapping matrix is ​​generated based on the final pipeline adjacency matrix. The topology connection mapping matrix is ​​formed by mapping pipeline node pairs with connection relationships to connection paths according to the position of the value of one in the adjacency matrix, and by numerical labeling based on the connection strength between nodes.

[0114] The final pipeline connection set, topology connection mapping matrix, and pipeline topology attribute table are encapsulated according to a unified data structure standard to generate a pipeline topology relationship model file. The pipeline topology relationship model file includes pipeline node information, pipeline connection relationships, and pipeline topology attribute information.

[0115] Example 1:

[0116] In a smart water management information system upgrade project in a prefecture-level city, the municipal water affairs bureau planned to digitally reconstruct the existing water supply network system. The original network data was compiled from multiple historical CAD drawings, on-site surveys, and GIS data, and suffered from issues such as node location misalignment, missing pipe diameter information, and inconsistent connection relationships. This resulted in insufficient accuracy of traditional geometric rule-based topology identification methods, a huge workload for manual correction, and a long topology reconstruction cycle. To address this, the project team adopted the pipeline topology relationship construction method based on graph neural networks proposed in this invention to achieve automated and intelligent topology modeling of complex network systems.

[0117] In practical application, spatial data was first collected from the city's main pipeline network area, obtaining raw spatial data containing 3687 pipeline nodes and 4721 pipe segments. The data was preprocessed to remove noise points and perform coordinate normalization and spatial alignment to ensure spatial consistency. Subsequently, in step S2, the algorithm automatically extracted the geometric and physical attributes of each pipeline node and the spatial relationship parameters of each pipe segment, generating node feature matrices and edge feature matrices, laying the data foundation for subsequent graph structure construction.

[0118] Based on the spatial proximity relationships between nodes and edge feature matrix information, a pipeline graph structure is automatically constructed. Index alignment and size verification are performed to generate a pipeline graph input matrix that meets the input requirements, ensuring a one-to-one correspondence between all nodes and edges in the matrix dimension. This mechanism effectively avoids the topological breaks and duplicate connections caused by misaligned node numbers or inconsistent feature dimensions in traditional geometric matching methods.

[0119] The system inputs the generated pipeline graph input matrix into the improved EvolveGCN model. The model internally includes a pipeline feature dynamic input module, a dynamic graph weight evolution module, a graph temporal feature propagation module, and a topology prediction adaptive optimization module. The dynamic graph weight evolution module employs a graph attention-controlled evolution mechanism, enabling the model to dynamically adjust convolutional weights based on the importance of node features, thereby achieving adaptive propagation of temporal features. After 500 training iterations, the model automatically learns the topological evolution patterns between nodes, generating a temporal embedding feature matrix that accurately reflects the potential connections between nodes.

[0120] The connection confidence between node pairs is calculated by embedding a temporal feature matrix, and the prediction results are filtered by setting a threshold of 0.65 to obtain the predicted pipeline topology connection set. Experimental results show that the method of the present invention achieves a node identification accuracy of 97.1% and a topology connection accuracy of 96.8% in municipal water supply network scenarios, with an error connection rate of only 3.2%, which is about 22% higher than the 78.9% of the traditional geometric matching method.

[0121] The predicted topology connection set undergoes topology consistency verification, comprehensively considering pipeline direction constraints, pipe diameter matching constraints, and engineering logic constraints. Reverse connections or connections with mismatched diameters are automatically eliminated to ensure the generated topology structure conforms to actual engineering specifications. Finally, in step S8, the system integrates and encapsulates the verified topology structure to generate a pipeline topology relationship model file. This file contains node information, connection relationships, and topology attributes, and can be directly imported into BIM and GIS systems for pipeline topology visualization, pressure analysis, and operation and maintenance management.

[0122] To verify the universality of the method, the experimental team also selected a comprehensive utility tunnel scenario for testing, with approximately 1593 nodes. The experimental results are shown in Table 1 below.

[0123] Table 1. Performance comparison of different methods in pipeline topology reconstruction task. Experimental results.

[0124]

[0125] As can be seen from the experimental results in Table 1, the method of this invention exhibits significant advantages in pipeline systems of different scales and types. In the scenario of municipal water supply networks, the node identification accuracy of this method is 97.1%, which is 15.9% higher than the geometric rule matching method and 9.5% higher than the CNN-based spatial feature learning method; the topology connection accuracy is improved to 96.8%, the error connection rate is only 3.2%, and the modeling time is only 42 minutes, which is 1 / 10 of the traditional method.

[0126] In integrated utility tunnel scenarios, this method maintains high accuracy and stability, achieving a topology connection accuracy of 95.4% and an error rate of only 4.6%, with a model integrity score as high as 96.1. Experimental results demonstrate that this invention not only significantly improves topology recognition accuracy but also maintains strong robustness and high reliability in complex spatial scenarios.

[0127] In summary, this invention, by introducing an improved EvolveGCN model and a graph attention-controlled evolutionary mechanism, achieves automatic identification and dynamic modeling of pipeline system topology relationships, significantly improving modeling accuracy and efficiency, reducing manual intervention, and possessing high reliability, scalability, and engineering application value.

Claims

1. A method for constructing pipeline topology relationships based on graph neural networks, characterized in that, Includes the following steps: Collect and preprocess the raw spatial data of the pipeline system; Feature extraction is performed on the preprocessed raw spatial data to generate node feature matrices and edge feature matrices; Based on the node feature matrix and edge feature matrix, and according to the spatial proximity relationship between pipe node objects and pipe segment objects, a pipe graph structure is constructed. Perform index alignment and dimension verification on the piping diagram structure to generate the piping diagram input matrix; The pipeline graph input matrix is ​​fed into the improved EvolveGCN model, and a graph attention control evolution mechanism is introduced to obtain the temporal embedding feature matrix of each pipeline node object. Calculate the connection confidence between node pairs based on the temporal embedding feature matrix of each pipeline node object, and generate a predicted pipeline topology connection set. Perform topology consistency verification on the predicted pipeline topology connection set to obtain the final pipeline topology relationship structure; The final generated pipeline topology structure is then integrated and encapsulated into a pipeline topology model file.

2. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The raw spatial data includes structural information from lidar scanning, building information modeling, computer-aided design drawings, and geographic information systems. The preprocessing includes coordinate unification, format conversion, and noise filtering.

3. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The generation of the node feature matrix and edge feature matrix specifically includes: The geometric and physical attributes of each pipe node object are extracted from the preprocessed raw spatial data to form a pipe node feature vector. The geometric attributes include the coordinates and orientation vectors of the pipe node object in three-dimensional space, and the physical attributes include the pipe diameter, material type, and node adjacency of the pipe node object. Arrange the feature vectors of the pipeline nodes in order of node number to form the pipeline node feature matrix; The spatial relationship parameters of each pipe segment object are extracted from the preprocessed raw spatial data to form a pipe segment feature vector. The spatial relationship parameters include the length of the pipe segment object, the directional angle, and the spatial distance between the two pipe node objects. The feature vectors of the pipeline segments are arranged sequentially according to the segment numbers to form the pipeline segment feature matrix. The eigenvectors of the pipeline node feature matrix and the eigenvectors of the pipeline segment feature matrix are sequentially combined vertically to form the node feature matrix and the edge feature matrix.

4. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The construction of the pipeline diagram structure specifically includes: Based on the node feature matrix and the edge feature matrix, a set of pipe nodes and a set of pipe connections are defined. The set of pipe nodes is defined as taking each pipe node object with independent spatial coordinates and attribute information as an element of the set according to the elements in the node feature matrix. The set of pipe connections is defined as taking each pipe segment object containing a start node, an end node and spatial relationship parameters as an element of the set according to the elements in the edge feature matrix. The spatial distance between two pipeline node objects is calculated using the three-dimensional spatial coordinate information in the node feature matrix, and the attribute similarity between nodes is calculated based on the physical attribute parameters contained in the node feature matrix. By combining the spatial relationship parameters in the edge feature matrix, a spatial distance threshold and an attribute similarity threshold are set. When the spatial distance between the node objects of two pipelines is less than the set distance threshold and the attribute similarity is greater than the set similarity threshold, the pipeline node object pair is determined to meet the potential pipeline connection conditions, and a candidate pipeline connection set is formed. The connection weights are calculated based on the spatial distance, attribute similarity, and pipe segment feature information in the edge feature matrix of each node pair in the candidate pipeline connection set. A weighted adjacency matrix is ​​constructed by assigning connection weights to each pair of candidate nodes and arranging them according to the node index. A complete pipeline graph structure is generated by combining the node feature matrix, edge feature matrix, pipeline node set, pipeline connection set, and weighted adjacency matrix.

5. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The generation of the pipeline diagram input matrix specifically includes: By combining the weighted adjacency matrix of the pipeline graph structure with the node feature matrix, a unified pipeline node numbering order is generated, and a pipeline node index mapping relationship is established. The pipeline node numbering order refers to the unique identification and sequential arrangement of all pipeline node objects in the pipeline graph structure. The pipeline node index mapping relationship is established by establishing a one-to-one correspondence between the weighted adjacency matrix and the node feature matrix in rows, columns, and node numbers according to the pipeline node numbering order. Based on the pipeline node index mapping relationship, an index alignment operation is performed on the weighted adjacency matrix and the node feature matrix to obtain the index-aligned weighted adjacency matrix and the index-aligned node feature matrix; The size of the index-aligned weighted adjacency matrix and the index-aligned node feature matrix are validated to generate the size-validated weighted adjacency matrix and the size-validated node feature matrix. The weighted adjacency matrix after size verification is combined and encapsulated with the node feature matrix after size verification to generate the pipeline graph input matrix.

6. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The specific steps for obtaining the temporal embedding feature matrix of each pipeline node object include: The pipeline graph input matrix is ​​input into the improved EvolveGCN model. The improved EvolveGCN model includes a pipeline feature dynamic input module, a dynamic graph weight evolution module, a graph temporal feature propagation module, and a topology prediction and adaptive optimization module. The pipeline feature dynamic input module performs time-step encoding and normalization on the nodes of the pipeline to form a node feature sequence. The dynamic graph weight evolution module calculates the convolution weight matrix using a graph attention-controlled evolution mechanism. The graph temporal feature propagation module performs graph convolution operations on the node feature sequence, and performs weighted aggregation and nonlinear transformation to generate a node hidden representation matrix. The topology prediction and adaptive optimization module calculates the node pair similarity based on the node hidden representation matrix at each time step and obtains the connection strength matrix, forming a temporal embedding representation and uniformly aggregating it to generate a temporal embedding feature matrix. In the pipeline feature dynamic input module, the nodes of the pipeline diagram input matrix are encoded and normalized by time steps to form a node feature sequence organized by time steps. In the dynamic graph weight evolution module, a graph attention control evolution mechanism is introduced to perform weighted mapping on the node feature sequences and generate the convolution weight matrix for the current time step. In the graph time-series feature propagation module, based on the convolution weight matrix, graph convolution operation is performed on the node feature sequence of the current time step to generate the node hidden representation matrix of the current time step. The node hidden representation matrix is ​​obtained by linearly transforming the node feature sequence of the current time step with the convolution weight matrix, combining the node connection relationship in the pipeline graph structure, aggregating the information of adjacent nodes on each node, and processing it with a nonlinear activation function. In the topology prediction and adaptive optimization module, the similarity of node pairs is calculated and the connection strength is adjusted based on the node hidden representations at each time step to form a temporal embedding representation. The temporal embedding representation is obtained by calculating the feature similarity between all node hidden representations at each time step to obtain the connection strength matrix of node pairs, and then performing weighted fusion based on the connection strength of the connection strength matrix. The temporal embedding representations of each time step are uniformly aggregated to generate a temporal embedding feature matrix.

7. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The generation of the predicted pipeline topology connection set specifically includes: The embedding vectors of the temporal embedding feature matrix are normalized, and a symmetric scoring operation is performed on any node pair to form a feature similarity matrix. The feature similarity matrix and the weighted adjacency matrix are weighted and fused according to a preset structural weight ratio to obtain the fused relevance matrix; A non-linear mapping is performed on the fusion correlation matrix to generate a connection confidence matrix with values ​​between zero and one. Based on the preset connection confidence threshold, node pairs with a confidence level greater than the connection confidence threshold in the connection confidence matrix are selected and marked as potential connections. All node pairs that meet the conditions are combined to form a predicted pipeline topology connection set.

8. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The final pipeline topology structure is obtained specifically through: Based on the spatial coordinates of pipe nodes, the direction vectors of pipe nodes, and the pipe diameter parameters of pipe nodes in the predicted pipe topology connection set, a topology consistency check is performed. The topology consistency check includes pipe direction constraint check, pipe diameter matching constraint check, and engineering logic constraint check. In the pipeline direction constraint verification, for each connection relationship, a direction vector is calculated based on the spatial coordinates of the pipeline node. The direction vector is calculated based on the difference in spatial coordinates between the starting pipeline node and the ending pipeline node in the connection relationship. The direction vector is then compared with the direction vector of the corresponding pipeline node. When the two direction vectors are less than a preset threshold, it is determined that the direction constraint is satisfied; otherwise, it is recorded as the direction constraint is not satisfied. In the process of performing pipe diameter matching constraint verification, the difference in pipe diameter parameters of the pipe nodes at both ends of each connection is calculated and compared with the preset pipe diameter tolerance threshold. When the pipe diameter difference is less than the pipe diameter tolerance threshold, the pipe diameter matching constraint is determined to be satisfied; otherwise, the connection is excluded. In the execution of engineering logic constraint verification, based on the engineering design rules of the pipeline network system, the maximum number of allowed connections for each pipeline node is set, and a prohibited connection set is defined. Node pairs that do not conform to the connection rules are eliminated. The prohibited connection set is defined by obtaining the type of each pipeline node according to the pipeline network engineering design rules, and recording the combination of node types that do not meet the requirements according to the engineering logic provisions. Based on pipeline direction constraint verification, pipe diameter matching constraint verification, and engineering logic constraint verification, each connection relationship is filtered and sorted according to the corresponding confidence value in the connection confidence matrix to form the final pipeline connection set. The final pipeline adjacency matrix is ​​generated based on the final pipeline connection set. The final pipeline connection set and the final pipeline adjacency matrix are combined to output the final pipeline topology structure. The final pipeline adjacency matrix is ​​obtained by establishing a unified number index for all pipeline nodes according to the node pair information in the final pipeline connection set, assigning a value of one to the corresponding position of node pairs with connection relationship in the adjacency matrix, and assigning a value of zero to node pairs without connection relationship.

9. The pipeline topology construction method based on graph neural networks according to claim 1, characterized in that, The formation of the pipeline topology model file specifically includes: The geometric and physical attribute information in the final pipeline topology structure is integrated with the connection relationship information in the final pipeline connection set to form a pipeline topology attribute table. The pipeline topology attribute table includes the starting pipeline node number, the ending pipeline node number, and the connection attribute information for each connection relationship. The corresponding topology connection mapping matrix is ​​generated based on the final pipeline adjacency matrix. The topology connection mapping matrix is ​​formed by mapping the pipeline node pairs with connection relationships to the connection paths according to the position of the value of one in the adjacency matrix, and by numerical labeling according to the connection strength between the nodes. The final pipeline connection set, topology connection mapping matrix, and pipeline topology attribute table are encapsulated according to a unified data structure standard to generate a pipeline topology relationship model file. The pipeline topology relationship model file includes pipeline node information, pipeline connection relationships, and pipeline topology attribute information.

Images