Intelligent Analysis Method for Weld Layout of Complex Pipeline Structures Based on Graph Neural Networks

Abstract

The present invention discloses an intelligent analysis method for weld layouts of complex pipeline structures based on graph neural networks, comprising the following steps: S1, obtaining a three-dimensional model to construct a graph structure; S2, predicting weld scores by a graph neural network; S3, constructing a candidate weld graph to execute a reward and punishment Steiner tree; S4, generating a construction dependency graph and executing a minimum feedback arc set; S5, constructing a residual conflict graph to calculate a conflict feedback coefficient; S6, adjusting reward values and weights; S7, re-executing the Steiner tree algorithm; S8, determining whether constraints are satisfied and outputting a final solution. The present invention realizes the joint optimization of intelligent weld layout and construction scheduling, and significantly improves the constructability and intelligent level under complex structures.

Description

Technical Field

[0001] The present invention relates to the technical field of graph neural networks, and particularly to an intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks. Background Art

[0002] Complex pipeline structures are widely used in major engineering fields such as petrochemical, power, nuclear energy, shipbuilding, and ocean engineering. Their spatial configurations usually have characteristics such as high three-dimensional coupling, dense pipe segment intersections, and numerous support nodes. During the pipeline manufacturing and installation process, the layout and construction sequence of welds not only directly affect the connection strength, structural stability, and service safety, but also have a significant impact on the construction period, welding process complexity, and maintenance cost. Therefore, how to complete a reasonable, implementable, and schedulable weld layout and construction sequence planning under the premise of meeting the structural performance requirements has become one of the key issues in the intelligent analysis of engineering drawings.

[0003] In the prior art, the weld layout and path analysis methods are mainly divided into three categories: structural expert rule method, geometric optimization method, and graph model optimization method. The expert rule method relies on manual experience for initial weld point selection and path arrangement, with low efficiency, poor consistency, and difficulty in dealing with large-scale structures. The geometric optimization method usually performs spatial analysis and path traversal based on a CAD model. Although it can automatically generate some reachable paths, it cannot consider engineering semantic factors such as connection importance and structural redundancy, and the layout is often disconnected from the overall process optimization. The graph modeling method developed in recent years has to some extent made up for the above deficiencies. It can abstract the complex pipeline structure into a topological graph composed of nodes and edges, and realize preliminary structural visualization analysis and layout optimization by introducing graph neural networks and path optimization algorithms.

[0004] Existing graph modeling methods generally use static graph neural networks to encode pipeline structures, and use node features (such as spatial position, pipe diameter, connection type) to score the importance of weld points, and then select the weld positions. However, such methods usually do not consider the logical dependencies between welding sequences. For example, some welds must be completed after other welds, or there are occlusion and reachability problems when constructing certain paths in a specific direction. In addition, during the layout graph generation process, the node selection strategy often only depends on score sorting, ignoring the cost of the connection path and the overall connectivity of the structure, which easily leads to local important nodes being isolated or high-cost weld connections being preferentially selected, affecting the overall constructability and stability.

[0005] In terms of path planning, existing research mostly uses graph theory methods such as the shortest path and minimum spanning tree for optimization. However, it lacks the comprehensive modeling of the dual factors of weld "reward value" and connection "cost", and cannot effectively balance the structural importance and welding difficulty. On this basis, introducing constraint optimization algorithms such as Steiner tree can alleviate the path redundancy problem to a certain extent. However, the node selection logic lacking the support of a reward and punishment mechanism is still prone to problems such as high reward and low connectivity or high connection and low benefit. Especially in complex scenarios with multi-redundant node distributions, the generated results are unstable and difficult to generalize.

[0006] On the other hand, there is a clear logical sequence dependence in the weld construction process, that is, some weld nodes must be welded before other nodes. If there are directed loops in the generated path graph, it will directly lead to construction logic conflicts. However, most existing methods adopt static sequence planning or rely on manual adjustment by engineers, lacking a systematic strategy for generating directed acyclic graphs. Although some research has introduced the minimum feedback arc set algorithm for loop elimination, its path selection process is mostly based on a single structural cost or graph topology weight, without considering the weld importance score or structural centrality factor, resulting in the eliminated paths not being optimal, easily destroying structural redundancy or generating unreasonable sequences.

[0007] In addition, existing weld planning methods are generally one-way processes, that is, deriving the weld layout and construction sequence from structural analysis, lacking feedback control capabilities, and unable to use the conflict information in the sequence graph to reverse-correct the weld node reward value or path cost. The absence of a feedback adjustment module makes the method insensitive to engineering uncontrollable factors such as welding conflicts, structural occlusion, and sequence adjustment, and the generated solutions lack robustness. In the multi-round iterative optimization process, existing technologies also lack a scientific termination strategy judgment criterion, making it difficult to effectively control the convergence and stability of the layout results, or there is a risk of resource waste and overfitting.

[0008] In summary, existing weld layout and construction sequence planning technologies still have significant deficiencies in aspects such as graph modeling accuracy, node path evaluation mechanism, sequence graph generation logic, feedback adjustment ability, and iterative convergence control. Especially when facing complex pipeline scenarios with high structural complexity, high redundant connections, and high construction dependence, there is an urgent need for a new intelligent weld analysis method that integrates graph neural network prediction, reward and punishment-based graph selection, conflict feedback adjustment, and multi-strategy termination control to improve the layout rationality, path reachability, and scheduling controllability of the system.

[0009] Therefore, how to provide an intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0010] One purpose of the present invention is to propose an intelligent analysis method for weld layout of complex pipeline structures based on graph neural network. The present invention comprehensively uses graph neural network modeling, reward-penalty Steiner tree optimization algorithm, minimum feedback arc set elimination mechanism and conflict feedback adjustment strategy to construct an iterative optimization system of structure-sequence dual graph linkage, and describes in detail the whole process control logic from weld prediction, path selection to construction sequence optimization, which has the advantages of high structural strength, strong construction accessibility, reasonable process sequence and strong feedback adjustment ability.

[0011] According to an embodiment of the present invention, a method for intelligent analysis of weld layout of a complex pipeline structure based on a graph neural network comprises the following steps:

[0012] S1. Obtain the three-dimensional model data of the complex pipeline structure and construct a graph structure;

[0013] S2. Input the graph structure into the graph neural network model, extract node features, generate the weld layout probability score of each node, and obtain the weld prediction graph;

[0014] S3. Based on the weld prediction graph, a weighted candidate weld graph is constructed, node scores are mapped to bonus values, edge attributes are mapped to welding costs, and a reward-penalty Steiner tree algorithm is introduced. A greedy strategy is adopted to generate a weld layout subgraph.

[0015] S4, constructing a directed dependency graph of weld construction based on the weld layout subgraph, using the minimum feedback arc set algorithm to eliminate loops in the graph, and generating a directed acyclic graph of construction sequence;

[0016] S5. construct a residual conflict graph based on the removed edges of the construction sequence directed acyclic graph, count the participation frequency of the conflict nodes, and calculate the conflict feedback information;

[0017] S6. adjusting the reward value of the corresponding node in the weighted candidate weld diagram according to the conflict feedback coefficient, and adjusting the weight of the conflicting node in the weld selection;

[0018] S7, re-inputting the updated reward value into the reward-penalty Steiner tree algorithm, and iteratively generating a new weld layout subgraph;

[0019] S8. Repeat S4 to S7, classify the iterative result strategies, until the structural strength and construction sequence non-cyclic constraints are met, and output the final weld layout diagram.

[0020] Optionally, it is characterized in that the S2 specifically includes: inputting the graph structure into the graph neural network model, allocating a node attribute vector including spatial coordinates, structure type, and bearing level for each node, allocating an edge attribute vector including connection length, connection angle, welding accessibility, and material heterogeneity for each edge, performing graph convolution processing on the above attribute vectors through at least three graph convolution layers of the graph neural network model, sequentially extracting features from the node attribute vectors, scoring each node through a classifier, generating a weld layout probability score corresponding to each node, and forming a weld prediction graph.

[0021] Optionally, it is characterized in that the S3 specifically includes:

[0022] S31. Construct a weighted candidate weld graph according to the weld prediction graph;

[0023] S32. For each node in the weighted candidate weld graph , set its reward value as , represents the weld layout probability score output by the node in the graph neural network. The weld layout probability score is calculated by the graph neural network based on the spatial position of the node, the pipeline structure type, and its bearing level characteristics. The higher the node score, the more this node needs to be preferentially selected;

[0024] S33. For each edge , set the cost function of the edge , and the cost function is calculated considering the following factors:

[0025] ;

[0026] Among them, represents the connection length of the welding path, represents the connection angle, represents the welding accessibility coefficient, represents the material heterogeneity coefficient, , , , is a preset weighting coefficient;

[0027] S34. Input the reward value of the node and the cost of the edge into the reward and punishment Steiner tree algorithm, and use the greedy strategy for iterative screening. The goal of the reward and punishment Steiner tree algorithm is to select a subgraph that contains the optimal weld nodes and the shortest welding path to minimize the objective function is:

[0028] ;

[0029] Among them, represents the set of selected weld nodes, represents the set of selected weld paths, and the first item represents the total cost of the selected path, and the second item is the penalty term for unselected nodes. By minimizing the objective function, the reward and punishment Steiner tree algorithm can select important weld nodes while optimizing the welding path, reduce the penalty for unselected nodes, and generate an optimized sub-graph of the weld layout.

[0030] Optionally, it is characterized in that the reward and punishment Steiner tree algorithm of S34 adopts a greedy strategy, which specifically includes:

[0031] S341. Define the combined evaluation function of the reward and punishment Steiner tree algorithm , which is used to select the optimal combination of nodes and edges, where:

[0032] ;

[0033] Among them, is the minimum connection distance between node and the currently selected node set, is the adjustment coefficient, which controls the balance between the path cost and the structural connectivity;

[0034] S342. In each round of selection, calculate the evaluation function value of each node and its connecting edge from the set of unselected nodes , select the node with the maximum evaluation function value and its corresponding connecting edge, add the node to the selected node set, and add the edge to the selected edge set;

[0035] S343. Whenever a new node is selected, update the reward values of all unselected nodes. The specific update formula is:

[0036] ;

[0037] Among them, is the reward value of the unselected node, is the updated reward value of the unselected node, is the attenuation factor, is the degree of node , is the number of currently selected nodes;

[0038] S344. Through iteration, execute the above node selection and reward value update steps, select the optimal node and update the path cost in each round until the objective function reaches the minimum value;

[0039] S345. When the objective function reaches the minimum or the preset number of iterations, the algorithm terminates, and the optimized weld layout sub-graph is output, which contains the optimal weld nodes and the shortest welding path.

[0040] Optionally, it is characterized in that the S4 specifically includes:

[0041] S41. Construct a weld construction dependency digraph based on the weld layout sub-graph generated by the reward and punishment Steiner tree algorithm;

[0042] S42. Perform loop detection on the dependency digraph to obtain all sets of feedback edges that form a closed loop;

[0043] S43. For each feedback edge, calculate the static feedback weight , which is jointly defined based on the structural path cost and the difference in weld importance as:

[0044] ;

[0045] Where , are the reward values of the two updated nodes respectively, , is the weighting coefficient;

[0046] S44. Use the minimum feedback arc set algorithm to select the edge set with the minimum total cost for deletion based on the static feedback weight to obtain an acyclic weld sequence graph.

[0047] Optionally, it is characterized in that the S44 specifically includes:

[0048] S441. Based on the set of feedback edges existing in the dependency digraph, combined with the connection density of the weld nodes connected by each edge in the original structure diagram, the position level in the diagram, and the key degree in the process path, calculate the path importance score to measure the structural and technological value of the dependency edge in the overall welding logic;

[0049] S442. Define the comprehensive cost function of the feedback edge according to the basic structure weight of the feedback edge, the difference in node reward values, the path importance, and the sequential tension index between nodes, specifically as follows:

[0050] ;

[0051] Where is the path importance score, comprehensively considering the connection degree and centrality of the node in the structure diagram, is the technological dependency tension between node pairs, reflecting the impact of sequential changes on the stability of the construction process, is the geometric distance of the node in the structure diagram, , is the adjustment parameter;

[0052] S443. Select a set of edges with the minimum sum of the comprehensive cost functions from the feedback edge set as the optimal feedback arc set, delete it, and output an acyclic weld sequence diagram. The sum of the minimum set of edges is used as the optimal feedback arc set, and after deletion, an acyclic weld sequence diagram is output.

[0053] Optionally, it is characterized in that the acyclic weld sequence diagram in S5 is further fused with the construction reachability diagram after the optimization of the feedback arc set is completed to construct a construction sequence directed acyclic graph, wherein the construction reachability diagram is generated based on the operation space layout, the equipment reach range, the weld node space hierarchy, and the adjacent component occlusion relationship. All physically executable paths are retained during the graph fusion process, and the directed edges conflicting with the actual construction path are deleted.

[0054] Optionally, it is characterized in that S5 specifically includes:

[0055] S51. Construct a residual conflict graph according to the difference set between the construction sequence directed acyclic graph and the weld construction directed dependency graph;

[0056] S52. Count the number of feedback edges involved in each node as the starting point or the ending point in the residual conflict graph, and calculate its node conflict frequency coefficient , defined as:

[0057] ;

[0058] Among them, represents the number of feedback edges associated with the node , is the degree value of this node in the weld layout subgraph, and is normalized;

[0059] S53. For each edge in the feedback edge set, calculate its path dynamic conflict intensity , which is defined as:

[0060] ;

[0061] Among them, is the static feedback weight, is the conflict weight adjustment coefficient, representing the amplification factor of the node conflict degree on the path conflict cost;

[0062] S54. Use the node conflict frequency coefficient and the path dynamic conflict intensity as conflict feedback information to guide the reward value function and the cost function Adaptive adjustment

[0063] Optionally, it is characterized in that S6 specifically includes: adaptively adjusting the reward function according to the node conflict feedback coefficient to generate the reward function after conflict response , defined as:

[0064] ;

[0065] Wherein, is the reward weakening factor, is the non-linear control index;

[0066] Replace the original reward value with the updated reward value , update the node attributes in the weighted candidate weld map, and adjust the evaluation function of the greedy strategy in the reward and punishment Steiner tree algorithm

[0067] Optionally, it is characterized in that the S8 strategy classification has four types, and the iteration is terminated as long as any one of the four strategies is satisfied. Specifically, it includes:

[0068] Strategy I, acyclicity sufficient satisfaction strategy: the construction sequence diagram generated in the current round is a directed acyclic graph, and all node conflict frequency coefficients derived from the residual conflict graph all satisfy , where

[0069] is the node conflict threshold;

[0070] ;

[0071] Wherein, represents the shortest equivalent redundant path number of the node in the structure diagram;

[0072] Strategy II, structural strength constraint strategy: all connection paths in the current weld layout subgraph satisfy the minimum connected redundancy defined in the structural mechanics model, that is:

[0073] ;

[0074] Wherein, is the structural convergence threshold, represents the symmetric difference, is the node set of the th round, is the node set of the th round, is the Wheel side set For the Wheel side set Indicates the number of elements;

[0075] Strategy IV, maximum round limit strategy: If the iteration round Reaches the preset maximum round upper limit , the iteration is forced to terminate.

[0076] The beneficial effects of the present invention are:

[0077] (1) The present invention proposes a graph neural network modeling and weld prediction method for complex pipeline structures, which can automatically extract component connection relationships and node features, generate a high-confidence weld layout probability map, avoid the limitations of traditional methods that rely on manual rules or geometric heuristic strategies for weld selection, and improve the accuracy and intelligence level of weld point selection.

[0078] (2) By introducing the reward and punishment Steiner tree algorithm, the present invention jointly models the structural importance (reward value) of nodes and the engineering cost (punishment) of connection paths, realizes the optimal selection of weld subgraphs under the global structural connectivity constraint, effectively solves the problem of the disconnection between node selection and path planning in the prior art, and dynamically adjusts the reward value and cost through a feedback adjustment mechanism to achieve iterative coupling optimization between the layout map and the construction sequence map.

[0079] (3) The present invention further combines the minimum feedback arc set algorithm to construct a construction sequence directed acyclic graph, and proposes a feedback adjustment strategy driven by a residual conflict graph to construct a closed-loop of collaborative optimization of structure-sequence dual graphs. At the same time, a structural difference convergence criterion, process connectivity control and maximum round tolerance mechanism are introduced to ensure that the whole method has adaptive, controllable and convergent optimization performance on the basis of meeting structural strength and constructability, and significantly improves the engineering feasibility and robustness of the scheme. Description of the Drawings

[0080] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0081] Figure 1 Is the overall flow chart of the intelligent analysis method for weld layout of complex pipeline structures based on graph neural network proposed by the present invention;

[0082] Figure 2 Is the flow chart of node-path selection of the reward and punishment Steiner tree based on the greedy strategy proposed by the present invention;

[0083] Figure 3This is the multi-round iterative optimization process and structure difference convergence judgment diagram of the intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks proposed by the present invention. Detailed implementation manners

[0084] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0085] Refer to Figures 1-3 , the intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks includes the following steps:

[0086] S1. Obtain the three-dimensional model data of the complex pipeline structure and construct a graph structure;

[0087] S2. Input the graph structure into the graph neural network model, extract node features, generate the weld layout probability scores of each node, and obtain a weld prediction graph;

[0088] S3. Construct a weighted candidate weld graph based on the weld prediction graph, map the node scores to reward values, map the edge attributes to welding costs, introduce a reward and punishment Steiner tree algorithm, and adopt a greedy strategy to generate a weld layout sub-graph;

[0089] S4. Construct a directed dependence graph of weld construction based on the weld layout sub-graph, and use the minimum feedback arc set algorithm to eliminate the loops in the graph to generate a directed acyclic graph of construction sequence;

[0090] S5. Construct a residual conflict graph based on the removed edges of the directed acyclic graph of construction sequence, count the participation frequencies of conflict nodes, and calculate conflict feedback information;

[0091] S6. Adjust the reward values of the corresponding nodes in the weighted candidate weld graph according to the conflict feedback coefficient, and adjust the weights of conflict nodes in weld selection;

[0092] S7. Re-input the updated reward values into the reward and punishment Steiner tree algorithm, and iteratively generate a new weld layout sub-graph;

[0093] S8. Repeat steps S4 to S7, classify the iterative results, until the structural strength and acyclic constraint of the construction sequence are satisfied, and output the final weld layout graph.

[0094] The present invention innovatively realizes the deep coupling optimization of weld layout and construction scheduling by constructing a complete closed-loop process from 3D modeling, graph neural network prediction, weld selection to construction sequence scheduling. Compared with the problems of mutual separation of weld analysis, path planning and process sequencing in the prior art, the present invention unifies the scheduling structure logic and spatial process information on the basis of the same graph modeling, and has the advantages of accurate prediction, efficient layout, feasible scheduling and strong convergence, significantly improving the intelligent weld analysis ability under complex pipeline structures.

[0095] In this embodiment, it is characterized in that the S2 specifically includes: inputting the graph structure into the graph neural network model, assigning a node attribute vector including spatial coordinates, structure type and bearing grade to each node, and assigning an edge attribute vector including connection length, connection angle, welding accessibility and material heterogeneity to each edge. The graph neural network model performs graph convolution processing on the above attribute vectors through at least three graph convolution layers, sequentially extracts features from the node attribute vectors, scores each node through a classifier, generates the weld layout probability scores corresponding to each node, and forms a weld prediction graph.

[0096] The present invention realizes the deep feature extraction of complex pipeline nodes at the spatial, structural and technological levels by introducing multi-dimensional attribute vectors for each node and edge in the graph structure and performing at least three graph convolution processes in the graph neural network. Compared with the traditional weld scoring method using only single geometric or connection information, the graph neural model of the present invention captures the semantic, topological and structural context of the welding points more comprehensively, provides a high-quality prediction basis for subsequent reward value setting and subgraph selection, and improves the intelligence and credibility of weld recognition.

[0097] In this embodiment, it is characterized in that the S3 specifically includes:

[0098] S31. Construct a weighted candidate weld graph according to the weld prediction graph;

[0099] S32. For each node in the weighted candidate weld graph , set its reward value as , represents the weld layout probability score output by the node in the graph neural network. The weld layout probability score is calculated by the graph neural network based on the spatial position of the node, the pipeline structure type and its bearing grade characteristics. The higher the node score, the more the node needs to be preferentially selected;

[0100] S33. For each edge , set the cost function of the edge , and the cost function is calculated considering the following factors:

[0101] ;

[0102] Among them, represents the connection length of the welding path, represents the connection angle, represents the welding accessibility coefficient, represents the material heterogeneity coefficient, , , , are preset weighting coefficients;

[0103] S34. Input the reward value of the node and the cost of the edge into the reward and punishment Steiner tree algorithm, and use the greedy strategy for iterative screening. The goal of the reward and punishment Steiner tree algorithm is to select a subgraph that contains the optimal weld nodes and the shortest welding path to minimize the objective function which is:

[0104] ;

[0105] Among them, represents the set of selected weld nodes, represents the set of selected weld paths. The first item represents the total cost of the selected path, and the second item is the penalty term for unselected nodes. By minimizing the objective function, the reward and punishment Steiner tree algorithm can select important weld nodes while optimizing the welding path, reduce the penalty for unselected nodes, and generate an optimized subgraph of the weld layout.

[0106] Based on the weld prediction graph, the present invention constructs a weighted candidate weld graph, maps the node scores to reward values, and maps the edge attributes to welding costs. By using the reward and punishment Steiner tree algorithm to generate the layout subgraph, the problem of the disconnection between weld point selection and path optimization in the prior art is solved. By simultaneously modeling the nodes and paths into the objective function and performing joint optimization, the present invention can minimize the welding path cost on the premise of ensuring the coverage of weld importance, and generate a weld layout scheme with a more stable structure and better weldability.

[0107] In this embodiment, it is characterized in that the greedy strategy adopted by the reward and punishment Steiner tree algorithm in S34 specifically includes:

[0108] S341. Define the joint evaluation function of the reward and punishment Steiner tree algorithm, which is used to select the optimal combination of nodes and edges, where:

[0109] ;

[0110] Among them, is the node The minimum connection distance between the current selected node set is the adjustment coefficient, which controls the balance between the path cost and the structural connectivity;

[0111] S342. In each round of selection, from the set of unselected nodes, calculate the evaluation function value of each node and its connecting edge , select the node with the maximum evaluation function value and its corresponding connecting edge, add the node to the selected node set, and add the edge to the selected edge set;

[0112] S343. Whenever a new node is selected, update the reward values of all unselected nodes. The specific update formula is:

[0113] ;

[0114] Among them, is the reward value of the unselected node, is the updated reward value of the unselected node, is the attenuation factor, is the node degree, is the number of currently selected nodes;

[0115] S344. Through iteration, execute the above node selection and reward value update steps, select the optimal node in each round and update the path cost until the objective function reaches the minimum value;

[0116] S345. When the objective function reaches the minimum or reaches the preset number of iterations, the algorithm terminates, and output the optimized weld layout sub-graph, which contains the optimal weld nodes and the shortest welding path.

[0117] The present invention introduces a greedy strategy based on node-edge combination into the reward and punishment Steiner tree algorithm, weighs the node reward value, edge cost and connectivity distance through a joint scoring function, and realizes the expansion of the optimal node round by round. At the same time, a reward value attenuation mechanism is constructed to avoid repeated selection of high-score nodes and ensure the overall rationality of the structure. Compared with the non-feedback and single-point drive strategies in the existing algorithms, the present invention provides a more intelligent node screening and path construction method, making the layout process highly dynamic, self-adaptive and locally optimal controllable.

[0118] In this embodiment, it is characterized in that the S4 specifically includes:

[0119] S41. Construct a weld construction dependency directed graph according to the weld layout sub-graph generated by the reward and punishment Steiner tree algorithm;

[0120] S42. Perform loop detection on the dependency directed graph to obtain all feedback edge sets that form closed loops;

[0121] S43. For each feedback edge, calculate the static feedback weight , which is jointly defined based on the structural path cost and the difference in weld importance as follows:

[0122] ;

[0123] where , are the reward values of the two updated nodes respectively, , is the weighting coefficient;

[0124] S44. Adopt the minimum feedback arc set algorithm, and select the edge set with the minimum total cost for deletion according to the static feedback weight to obtain an acyclic weld sequence diagram.

[0125] Based on the weld layout subgraph, the present invention constructs a construction dependency graph, and through the feedback edge weight function, synthesizes the structural cost and the difference in weld reward values to guide the minimum feedback arc set algorithm for loop elimination. This method breaks through the problem in traditional DAG construction that relies on static graph topology information and lacks node semantic modeling, can more accurately identify process logic conflict paths, improve the optimality of the directed dependency graph and the practicality of the welding process diagram, and provide a theoretical guarantee for subsequent scheduling generation.

[0126] In this embodiment, it is characterized in that the S44 specifically includes:

[0127] S441. Based on the feedback edge set existing in the dependency directed graph, combine the connection density of the weld nodes connected by each edge in the original structure diagram, the position level in the graph, and the key degree in the process path to calculate the path importance score, which is used to measure the structural and technological value of the dependency edge in the overall welding logic;

[0128] S442. Define the comprehensive cost function of the feedback edge according to the basic structure weight of the feedback edge, the difference in node reward values, the path importance, and the sequential tension index between nodes, specifically as follows:

[0129] ;

[0130] where is the path importance score, which synthesizes the connection degree and centrality of the node in the structure diagram, is the process dependency tension between node pairs, which reflects the impact of sequential changes on the stability of the construction process, is the geometric distance of the node in the structure diagram, , is the adjustment parameter;

[0131] S443. Select a set of edges with the minimum sum of comprehensive cost functions from the feedback edge set as the optimal feedback arc set, and delete it to output an acyclic weld sequence diagram. After deletion, an acyclic weld sequence diagram is output.

[0132] In the present invention, by introducing factors such as path importance, sequential tension, and structural hierarchy to the feedback edges to define a comprehensive cost function, the influence degree of each edge on the overall construction logic is accurately measured, and a weighted strategy is used to select the optimal feedback arc set. Compared with the existing method that uses a single structural weight or static path length as the basis for loop resolution, the present invention incorporates structural semantics, weld strategy, and process tension into the optimization scope, realizes a more scientific acyclic reconstruction of the dependency graph, and effectively prevents the destruction of process logic caused by incorrect edge deletion.

[0133] In this embodiment, it is characterized in that after the optimization of the feedback arc set, the acyclic weld sequence diagram in S5 is further fused with the construction reachability graph to construct a directed acyclic graph of construction sequence, wherein the construction reachability graph is generated based on the operation space layout, equipment reach range, spatial hierarchy of weld nodes, and occlusion relationship between adjacent components. During the graph fusion process, all physically executable paths are retained, and the directed edges conflicting with the actual construction path are deleted.

[0134] The present invention fuses the weld sequence diagram with the construction reachability graph, comprehensively considers the spatial layout, equipment operation range, and weld occlusion relationship, screens and retains physically feasible paths, and generates a directed acyclic graph of construction sequence. This mechanism enables the system to transition from "theoretical sequence feasibility" to "engineering feasibility", avoids the problem that the graph structure is logically correct but the on-site construction is unreachable, significantly enhances the executability, reliability, and engineering adaptability of the weld scheduling graph, and improves the implementability of the system in actual projects.

[0135] In this embodiment, it is characterized in that S5 specifically includes:

[0136] S51. Construct a residual conflict graph according to the difference set between the directed acyclic graph of construction sequence and the directed dependency graph of weld construction;

[0137] S52. Count the number of feedback edges involved in each node as the starting point or the ending point in the residual conflict graph, and calculate its node conflict frequency coefficient , defined as:

[0138] ;

[0139] Among them, represents the number of feedback edges associated with the node , is the degree value of this node in the weld layout subgraph, and is normalized;

[0140] S53. For each edge in the feedback edge set , calculate its path dynamic conflict intensity , which is defined as:

[0141] ;

[0142] wherein is the static feedback weight, is the conflict weight adjustment coefficient, representing the amplification factor of the node conflict degree on the path conflict cost;

[0143] S54. Use the node conflict frequency coefficient and the path dynamic conflict intensity as conflict feedback information to guide the adaptive adjustment of the reward function and the cost function .

[0144] The present invention introduces a residual conflict graph to explicitly model the conflict between sequential dependence and structural prediction, and constructs a node conflict frequency coefficient and a path conflict intensity function, providing an accurate quantification index for subsequent reward and cost adjustment. Compared with the "unaware" conflict feedback in traditional weld optimization, the present invention realizes a closed-loop mapping mechanism for sequential-structural conflict feedback, which helps to accurately identify key conflict regions and nodes, thereby improving the convergence efficiency, conflict suppression ability and overall stability of the optimization system.

[0145] In this embodiment, it is characterized in that the S6 specifically includes:

[0146] Adaptive adjustment of the reward function is performed according to the node conflict feedback coefficient to generate a reward function after conflict response, which is defined as:

[0147] ;

[0148] wherein is the reward weakening factor, is the non-linear control index;

[0149] Replace the original reward with the updated reward , update the node attributes in the weighted candidate weld graph, and adjust the evaluation function of the greedy strategy in the reward-punishment Steiner tree algorithm.

[0150] The present invention adjusts the node reward value according to the conflict feedback coefficient, introduces a non-linear attenuation function to control the amplitude of reward value update, and realizes the intelligent dynamic adjustment of node selection weights. By updating the evaluation function in the greedy strategy, the present invention realizes a conflict-oriented controllable feedback mechanism. Compared with the traditional node evaluation method driven by single-shot scoring, this mechanism significantly improves the system's response ability and learning ability to the conflict area, enhances the algorithm's precision control and feedback adaptation ability, and provides a solid foundation for iterative optimization.

[0151] In this embodiment, it is characterized in that there are four types of S8 strategies, and the iteration is terminated as long as any one of the four strategies is satisfied. Specifically, it includes:

[0152] Strategy I, acyclicity sufficient satisfaction strategy: the construction sequence diagram generated in the current round is a directed acyclic graph, and all node conflict frequency coefficients derived from the residual conflict graph all satisfy , where is the node conflict threshold;

[0153] Strategy II, structural strength constraint strategy: all connection paths in the current weld layout subgraph satisfy the minimum connected redundancy defined in the structural mechanics model , that is:

[0154] ;

[0155] Among them, represents the number of the shortest equivalent redundant paths of the node in the structure diagram;

[0156] Strategy III, structural change convergence strategy: the structural difference degree between the current layout diagram and the previous round layout diagram satisfies:

[0157] ;

[0158] Among them, is the structure convergence threshold, represents the symmetric difference, is the node set of the th round, is the node set of the th round, is the edge set of the th round, is the edge set of the th round, represents the number of elements;

[0159] Strategy IV, maximum round limit strategy: if the iteration round reaches the preset maximum round upper limit , the iteration is forced to terminate.

[0160] The present invention proposes four types of termination strategies, including acyclicity sufficient satisfaction, structural strength constraint, structural change convergence, and maximum round number control, for determining the reasonable termination timing of feedback iteration. This set of strategies takes into account process legality, structural stability, and algorithm efficiency. Compared with the existing termination mechanisms that use fixed rounds or single error metrics, it has stronger precision control and flexible adjustment capabilities, significantly improving the controllability, robustness, and engineering practicality of the weld optimization process.

[0161] Example 1:

[0162] To verify the feasibility and performance advantages of the present invention in actual engineering scenarios, the method of the present invention is applied to the welding construction of high-pressure ring network pipelines in the second-phase project of a coastal natural gas liquefaction and treatment base. The core pipe network layout of this project is complex, with the total length of the overall pipeline exceeding 4.3 kilometers, involving 17 large flange connection sections and nearly a hundred multi-branch intersection nodes. There are a series of practical problems such as a high density of weld tasks, serious pipe segment occlusion, and complex construction paths. In this project, traditional weld layout planning methods have repeatedly encountered problems such as layout fragmentation, serious path redundancy, and construction sequence conflicts, seriously affecting the continuity of the welding process and the construction rhythm.

[0163] In this project, the technical team first obtained a complex pipeline structure model exported by a three-dimensional laser scanning and modeling system and used the method of the present invention to perform graph structure conversion on this model. In the converted graph structure, each node corresponds to a potential weld point, and the node attributes cover spatial coordinates, pipe segment types, and load-bearing grades, while the edge attributes reflect welding connection length, angle, accessibility, and material heterogeneity. Subsequently, this graph structure is input into the constructed graph neural network model for feature extraction and weld score prediction, successfully identifying 82 weld nodes with engineering weldability value and giving their layout probability scores, forming a preliminary weld prediction graph.

[0164] Based on the prediction graph, the method of the present invention further constructs a weighted candidate weld graph and applies the reward-punishment Steiner tree algorithm combined with the reward value-cost joint scoring strategy to dynamically screen weld nodes and paths. Through this algorithm, the system preferentially retains nodes with key structural positions, high reward values, and lower path costs, and fully considers construction connection risks and welding tensions in the cost function. Different from the traditional shortest path method, the present invention introduces a feedback adjustment mechanism in the node selection process, adaptively adjusting the reward value based on path-dependent tension and conflict feedback, thus avoiding occlusion conflicts and welding sequence conflicts that occur in early selections.

[0165] During the optimization process, the system went through 2 rounds of conflict feedback adjustment iterations, automatically eliminating 3 redundant connection paths that caused construction occlusion and sequence chaos. Finally, a weld layout plan with a structural stability score of 91.2 was output, and an acyclic construction sequence diagram was constructed. During the process of the construction unit verifying the plan, the proposed layout and scheduling diagram was determined by the on-site welding team as "no need to adjust the sequence again and has direct constructability", and was successfully applied to the welding subsystems of the 3rd and 4th modules of the project, and achieved efficient implementation.

[0166] In the comparative experiment, the technical team also used the traditional graph optimization module in a well-known industrial software to analyze the weld layout in the same scenario. Although the initial weld diagram can be quickly output under this method, its prediction accuracy is less than 85%. There are multiple loops in the final sequence diagram, and the on-site manual adjustment of the sequence logic is about 8 times. The average scheduling planning time is 17 seconds more than that of the method of the present invention, and the structural connection stability score is less than 70 points, which affects the continuous scheduling of the subsequent process path and the construction rhythm.

[0167] The specific experimental data is sorted out in the following table, clearly reflecting the advantages of the present invention in key indicators such as the accuracy of weld node recognition, the optimization convergence speed, and the feasibility of construction scheduling:

[0168] Table 1: Performance comparison data between the method of the present invention and the traditional method in the actual engineering scenario

[0169] ;

[0170] According to the indicators listed in Table 1, it can be clearly seen that the method of the present invention has significant advantages over the traditional graph optimization method in multiple key dimensions.

[0171] First of all, from the perspective of the accuracy of weld node recognition, the method of the present invention correctly predicted 78 out of a total of 82 weld nodes, and the prediction accuracy reached 95.12%. The traditional method only identified 65 accurate nodes, and the accuracy rate was 83.33%. This shows that the graph neural network model constructed by the present invention has stronger learning and expression capabilities in the node attribute extraction and scoring links, can effectively identify high-value weld points, and avoid missing or misjudging key welds.

[0172] Secondly, in terms of the optimization of the construction path, the total number of paths generated by the method of the present invention is 136, which is less than 143 of the traditional method, and the number of feedback conflicts is significantly reduced, from 12 times to only 3 times. It shows that by combining the reward and punishment Steiner tree algorithm with the feedback adjustment mechanism, redundant paths and logical conflicts between paths are effectively avoided, and the overall simplicity and logical consistency of the layout structure are improved.

[0173] In terms of optimizing efficiency, the average layout optimization time of the method of the present invention is 26.4 seconds, which is shortened by more than 39.7% compared with 43.8 seconds of the traditional method. This not only improves the operation efficiency of the optimization process, but also reduces the resource cost of deployment and operation in large-scale industrial projects. At the same time, the constructability rate of the layout plan is increased from 87.4% to 98.6%, fully reflecting the comprehensive ability of the present invention in considering physical accessibility, structural occlusion and process sequence consistency.

[0174] In addition, from the perspective of the final scheduling results, the sequence diagram generated by the traditional method has loops and needs to be adjusted manually, and its structural stability score is only 68.7, while the sequence diagram generated by the present invention is a directed acyclic structure with a structural stability score as high as 91.2, indicating that it is more in line with the engineering structure connection logic and on-site welding process requirements. In terms of feedback convergence, the method of the present invention only needs 2 rounds of iteration to complete the optimization, while the traditional method requires at least 5 rounds, further verifying the accelerating effect of the feedback adjustment module on conflict correction and iterative convergence.

[0175] In summary, the present invention is significantly superior to the prior art in many aspects such as weld node prediction, path optimization, conflict adjustment, convergence efficiency and structural stability, and has extremely high engineering feasibility, system intelligence and on-site adaptability. It is an innovative method with practical value and technical prospects in the field of intelligent analysis of welds and construction scheduling for complex pipeline structures.

[0176] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An intelligent analysis method for the weld layout of complex pipeline structures based on graph neural networks, characterized in that The method includes the following steps: S1. Obtain the three-dimensional model data of the complex pipeline structure and construct a graph structure; S2. Input the graph structure into a graph neural network model, extract node features, generate the weld layout probability scores of each node, and obtain a weld prediction graph; S3. Construct a weighted candidate weld graph based on the weld prediction graph, map the node scores to reward values, map the edge attributes to welding costs, introduce a reward and punishment Steiner tree algorithm, and adopt a greedy strategy to generate a weld layout subgraph; S4. Construct a directed dependence graph of weld construction based on the weld layout subgraph, and use the minimum feedback arc set algorithm to eliminate the loops in the graph to generate a directed acyclic graph of construction sequence; S5. Construct a residual conflict graph based on the removed edges of the directed acyclic graph of construction sequence, count the participation frequencies of conflict nodes, and calculate conflict feedback information; S6. Adjust the reward values of the corresponding nodes in the weighted candidate weld graph according to the conflict feedback coefficient, and adjust the weights of the conflict nodes in weld selection; S7. Re-input the updated reward values into the reward and punishment Steiner tree algorithm, and iteratively generate a new weld layout subgraph; S8. Repeat S4 to S7, classify the iterative results, until the structural strength and the acyclic constraint of the construction sequence are satisfied, and output the final weld layout graph.

2. The intelligent analysis method for the weld layout of a complex pipeline structure based on a graph neural network according to claim 1, characterized in that The specific process of S2 includes: input the graph structure into the graph neural network model, assign a node attribute vector including spatial coordinates, structure type and bearing grade to each node, assign an edge attribute vector including connection length, connection angle, welding accessibility and material heterogeneity to each edge, the graph neural network model performs graph convolution processing on the above attribute vectors through at least three graph convolution layers, extracts the feature of the node attribute vector in turn, scores each node through a classifier, generates the weld layout probability scores corresponding to each node, and forms a weld prediction graph.

3. The intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks according to claim 1, wherein The specific process of S3 includes: S31. Construct a weighted candidate weld graph according to the weld prediction graph; S32. For each node in the weighted candidate weld seam diagram , set its reward value to , denotes the weld seam layout probability score output by the node in the graph neural network. The weld seam layout probability score is calculated by the graph neural network based on the spatial position of the node, the type of pipeline structure, and its bearing grade characteristics. The higher the node score, the more this node needs to be preferentially selected; S33. For each side , set the cost function of the side , and the cost function is calculated by weighted fusion of the connection length, connection angle, welding accessibility coefficient, and material heterogeneity coefficient of the welding path; S34. Input the reward value of the node and the cost of the edge into the reward and penalty Steiner tree algorithm, and use the greedy strategy for iterative screening. The goal of the reward and penalty Steiner tree algorithm is to select a subgraph that contains the optimal weld nodes and the shortest welding path to minimize the objective function as follows: ； Among them, represents the set of selected weld nodes, represents the set of selected weld paths.

4. The intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks according to claim 3, characterized in that, The greedy strategy adopted by the reward and punishment Steiner tree algorithm in S34 specifically includes: S341. Define the combined evaluation function of the reward and punishment Steiner tree algorithm for selecting the optimal combination of nodes and edges; S342. In each round of selection, from the set of unselected nodes, calculate the evaluation function values of each node and its connected edges , select the node with the maximum evaluation function value and its corresponding connected edge, add the node to the set of selected nodes, and add the edge to the set of selected edges; S343. Whenever a new node is selected, update the reward values of all unselected nodes. The specific update formula is: ； Among them, is the reward value of unselected nodes, is the updated reward value of unselected nodes, is the attenuation factor, is the node degree, is the number of currently selected nodes; S344. Through iteration, perform the above node selection and reward value update steps, select the optimal node in each round and update the path cost until the objective function reaches the minimum value; S345. When the objective function reaches the minimum or reaches the preset number of iterations, the algorithm terminates, and outputs the optimized weld layout subgraph, which includes the optimal weld nodes and the shortest welding path.

5. The intelligent analysis method for the weld layout of a complex pipeline structure based on a graph neural network according to claim 1, characterized in that The specific process of S4 includes: S41. Construct a directed dependence graph of weld construction based on the weld layout subgraph generated by the reward and punishment Steiner tree algorithm; S42. Perform loop detection on the dependence directed graph to obtain the set of all feedback edges forming a closed loop; S43. Calculate the static feedback weight for each feedback edge , where the static feedback weight is jointly defined based on the structural path cost and the difference in weld importance as follows: ； Among them, , are the bonus values of the two updated nodes respectively, , is the weighting coefficient; S44. Using the minimum feedback arc set algorithm, based on the static feedback weights select the edge set with the minimum total cost for deletion to obtain an acyclic weld sequence diagram.

6. The intelligent analysis method for the weld layout of a complex pipeline structure based on a graph neural network according to claim 5, characterized in that The specific process of S44 includes: S441. Based on the set of feedback edges existing in the dependence directed graph, combine the connection density of the weld nodes connected by each edge in the original structure diagram, the position level in the graph and the key degree in the process path, calculate the path importance score, which is used to measure the structural and technological value of the dependence edge in the overall welding logic; S442. Define a comprehensive cost function for the feedback edge according to the basic structure weight of the feedback edge, the node reward value difference, the path importance, and the sequential tension index between nodes , as follows: ； Among them, is the path importance score, which comprehensively considers the connectivity and centrality of nodes in the structure diagram, is the process dependence tension between node pairs, reflecting the impact of sequential changes on the stability of the construction process, is an exponential function, is the geometric distance of nodes in the structure diagram, , is an adjustment parameter; S443. Select a set of edges with the minimum sum of comprehensive cost functions from the feedback edge set as the optimal feedback arc set, delete it, and output an acyclic weld seam sequence diagram.

7. The intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks according to claim 1, wherein The acyclic weld sequence diagram described in S5 is further fused with the construction accessibility diagram after the feedback arc set optimization is completed to construct a construction sequence directed acyclic graph. Among them, the construction accessibility diagram is generated based on the operation space layout, equipment reach range, weld node space hierarchy, and adjacent component occlusion relationship. All physically executable paths are retained during the graph fusion process, and the directed edges conflicting with the actual construction path are deleted.

8. The intelligent analysis method for the weld layout of a complex pipeline structure based on a graph neural network according to claim 1, characterized in that The specific content of S5 includes: S51. Construct a residual conflict graph according to the difference set between the construction sequence directed acyclic graph and the weld construction directed dependency graph; S52. Count each node The number of feedback edges involved as a starting point or an ending point in the residual conflict graph, and calculate the node conflict frequency coefficient ; S53. For each edge in the feedback edge set , calculate its path dynamic conflict strength , which is defined as: ； Among them, is the static feedback weight, is the conflict weight adjustment coefficient, representing the amplification factor of the node conflict degree on the path conflict cost; S54. Use the node conflict frequency coefficient and the path dynamic conflict intensity as conflict feedback information to guide the adaptive adjustment of the reward value function and the cost function .

9. The intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks according to claim 1, wherein Specifically, S6 includes: adaptively adjusting the reward function according to the node conflict feedback coefficient to generate the reward function after conflict response , which is defined as: ； Among them, is the reward value attenuation factor, is the non-linear control index; Replace the original award value with the updated award value Replace the original award value , update the node attributes in the weighted candidate weld seam diagram, and adjust the evaluation function of the greedy strategy in the reward and punishment Steiner tree algorithm.

10. The intelligent analysis method for weld layout of complex pipeline structures based on graph neural networks according to claim 1, characterized in that There are four types of S8 strategy classifications, and the iteration is terminated as long as any one of the four strategies is satisfied. Specifically, it includes: Strategy I, a strategy that fully satisfies acyclicity: The construction sequence diagram generated in the current round is a directed acyclic graph, and all node conflict frequency coefficients derived from the residual conflict graph all satisfy , where is the node conflict threshold; Strategy II, Structural Strength Constraint Strategy: All connection paths in the current weld layout sub-graph satisfy the minimum connected redundancy defined in the structural mechanics model , that is: ； Among them, represents the number of the shortest equivalent redundant paths of the node in the structure diagram; Strategy III, Structural Change Convergence Strategy: The structural difference degree between the current layout diagram and the previous round of layout diagrams Meet the following conditions: ； wherein is the structural convergence threshold, represents the symmetric difference, is the node set of the th round, is the node set of the th round, is the edge set of the th round, is the edge set of the th round, represents the number of elements; Strategy IV, the maximum round limit strategy: If the iteration round reaches the preset maximum round upper limit , the iteration is forced to terminate.

Images