A pipe network connectivity assessment method and device based on a DS-PIGNN and related equipment

Abstract

This application relates to the field of urban pipeline network analysis technology, and particularly to a pipeline network connectivity assessment method, device, and related equipment based on DS-PIGNN. The method includes: acquiring pipeline network data, constructing a macroscopic topology map, and collecting the axial microscopic physical field sequence of each pipeline; extracting the features of the microscopic sequence of each pipeline using a one-dimensional residual convolutional neural network, and obtaining a microscopic health state vector through global max pooling; injecting this vector into the edge features of the macroscopic topology map to drive a graph neural network trained with a physical constraint loss function (including Kirchhoff flow conservation), which dynamically calculates the attention weights between nodes based on the hydraulic features of the nodes and the microscopic health state vector; and performing information aggregation based on this dynamic weight, simultaneously outputting the pipeline failure probability and node connectivity reliability. This application achieves dynamic and coupled analysis of microscopic physical damage details and macroscopic network cascade failures, and ensures the physical reliability of the assessment results while maintaining computational efficiency.

Description

Technical Field

[0001] This application relates to the field of urban pipeline network analysis technology, and in particular to a pipeline network connectivity assessment method, device and related equipment based on a dual-scale physical perception graph neural network. Background Technology

[0002] The statements in this section are merely background information related to this application and do not necessarily constitute prior art.

[0003] Urban pipeline systems (such as water supply, gas, and heating networks) are the lifeline of urban operations, and their reliability directly affects public safety and social stability. my country has a vast underground pipeline network, with drainage pipelines alone exceeding 700,000 kilometers in length. However, widespread neglect and disrepair have led to unresolved issues such as leaks, corrosion, cracks, and subsidence, resulting in significant water loss, reduced sewage treatment efficiency, and disruptions to flood control and drainage systems. This has also triggered numerous major accidents, causing loss of life and property and disrupting urban order.

[0004] However, existing technologies face three main technical bottlenecks when conducting reliability assessments of pipeline systems: 1. The scale separation problem between macroscopic simulation and microscopic physics Existing assessment methods (such as Monte Carlo simulations based on hydraulic simulation software like EPANET and WaterGEMS) typically treat pipe networks as abstract topological graphs composed of nodes and edges, simplifying pipe properties to scalar parameters (such as average wall thickness and average corrosion rate). This "averageization" masks the true microscopic physical state inside the pipes (such as localized corrosion defects and stress concentration points). While the finite element method can perform detailed physical analysis on individual pipes, its explosively increased computational cost prevents its scaling to city-level pipe networks. This makes it difficult for existing technologies to capture the "weakest link" risk caused by "localized extreme defects," often resulting in overly optimistic assessments.

[0005] 2. Static evaluation cannot reflect the complexity of dynamic cascading failures. Traditional methods are mostly based on fixed topology and steady-state conditions for assessment, which is a "static snapshot" analysis. However, real pipeline failure is a dynamic process: a single point failure can lead to pressure redistribution, which may then trigger a domino-like cascading failure. Existing technologies lack the ability to model the dynamic coupling mechanism of "microscopic physical damage driving macroscopic topology evolution," and cannot effectively simulate the dynamic coupling effect between branches and main lines and the failure propagation path, resulting in blind spots in risk assessment under complex operating conditions.

[0006] 3. The trade-off between computational efficiency and evaluation reliability Methods that rely on high-fidelity physical simulations (such as fluid-structure coupled simulations) in pursuit of accuracy are computationally extremely time-consuming and cannot meet the timeliness requirements of real-time or near-real-time early warning for urban-level pipe networks (typically requiring millisecond to second-level response). While some purely data-driven AI models (such as simple graph neural networks) are computationally fast, they often lack physical constraints and may output results that violate fundamental physical laws such as the conservation of mass and energy, making them unreliable in high-risk engineering decisions. The industry has long faced the dilemma of "accurate but inefficient physical simulations" versus "efficient but unreliable AI predictions." Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this application provides a method, apparatus, and related equipment for evaluating pipeline connectivity using a dual-scale physical perception graph neural network. For the first time, it realizes dynamic and coupled analysis of microscopic physical damage details and macroscopic network cascade failures in pipeline reliability assessment, and ensures the physical reliability of the assessment results while guaranteeing computational efficiency.

[0008] To achieve the above objectives, this application provides the following technical solution: Firstly, a pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network is provided, the method comprising: Acquire target pipeline network data, construct macroscopic topology map, and obtain microscopic physical field sequence data for each pipeline; For each pipeline, features of its microscopic physical field sequence are extracted using a one-dimensional residual convolutional neural network, and the microscopic health state vector representing the weakest link of the pipeline is obtained by global max pooling aggregation. The micro-health state vector is injected into the corresponding edge features of the macro-topology graph, and the dynamic attention weights between nodes are calculated using a physical perception graph neural network. The physical perception graph neural network is trained using a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. Based on the dynamic attention weights, neighborhood information is aggregated, and the failure probability of each pipeline and the connectivity reliability of each node are output.

[0009] Furthermore, the microscopic physical field sequence data includes at least one of the following: a pipe axially discretized wall thickness sequence, a von Mises stress sequence, and a soil settlement displacement sequence.

[0010] Furthermore, the one-dimensional residual convolutional neural network extracts features using the following formula:

[0011] in, The extracted local feature sequence, For pipe eij The microscopic physical field sequence matrix, This represents a one-dimensional convolution operation. and These are the learnable convolutional weights and biases, respectively.

[0012] Furthermore, the dynamic attention weights The calculation formula is:

[0013]

[0014] in, The hydraulic feature vector of the node; This represents a vector of microscopic health states. This represents a vector concatenation operation; The weight matrix is ​​a learnable matrix; Represents a node and nodes The unnormalized attention scores between them; Represents a node Its neighboring nodes The unnormalized attention scores between them; This is the transpose of the learnable attention weight vector; It is a linear rectified activation function with leakage.

[0015] Furthermore, the loss function that includes Kirchhoff flow conservation constraints for: = +

[0016] in, For the mass conservation loss based on failure probability gating, This is the energy compatibility loss based on Bernoulli's equation.

[0017] Furthermore, the mass conservation loss The expression is:

[0018] Where V is the set of all nodes in the pipeline network; |V| is the number of nodes in set V, that is, the total number of nodes in the pipeline network; v is the node currently being calculated. Let be the set of upstream nodes pointing to node v, that is, the set of starting nodes of the pipe segment flowing into node v; u is the upstream node flowing into node v. Let w be the set of downstream nodes pointed to by node v, that is, the set of terminal nodes of the pipe segment flowing out from node v; w is the downstream node flowing out from node v. Let be the failure probability of pipe segment (u, v); Let be the failure probability of pipe segment (v, w); The design flow rate or estimated flow rate for pipe segment (u, v); The design flow rate or estimated flow rate for the pipe segment (v, w); Let V be the water demand of node v.

[0019] Furthermore, the energy compatibility loss The expression is:

[0020] in, Let i and j be the water heads, i.e., the water pressure potential energy, respectively. This refers to the pipeline resistance coefficient. The flow index; Indicates the connection node and nodes The failure probability of this pipe section; This is the set of all pipes in the pipeline network topology diagram; Indicates a pipe; This represents the predicted flow rate of the fluid within the pipe.

[0021] Furthermore, the method also includes: Based on the Monte Carlo simulation framework, multiple randomized pipeline operating conditions were generated, and the overall reliability of nodes was statistically analyzed. The calculation formula is as follows:

[0022] in, The path connectivity depends on the survival probability of the weakest link in the path. The total number of Monte Carlo random simulations; This refers to the flow regime index or flow rate index in the fluid dynamics resistance equation.

[0023] Furthermore, the pipeline network is a water supply network, a gas network, or a heating network.

[0024] Secondly, a pipeline connectivity assessment device based on a dual-scale physical perception graph neural network is provided to implement the pipeline connectivity assessment method based on a dual-scale physical perception graph neural network as described above. The device includes: The data mapping module is used to acquire target pipeline network data, construct a macroscopic topology map, and acquire pipeline microscopic physical field sequence data. The micro-encoding module is used to extract features of the pipeline's micro-physical field sequence using a one-dimensional residual convolutional neural network, and to obtain a micro-health state vector representing the weakest link of the pipeline through global max pooling aggregation. The topology reconstruction module is used to inject the micro-health state vector into the corresponding edge features of the macro-topology graph and calculate the dynamic attention weights between nodes using a physical perception graph neural network. The physical perception graph neural network is trained by a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. The evaluation output module is used to aggregate neighborhood information based on the dynamic attention weights and output the failure probability of each pipeline and the connectivity reliability of each node.

[0025] Based on the same inventive concept, this application also provides a computer-readable storage medium storing one or more programs, which, when executed, implement the pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network as described above.

[0026] Based on the same inventive concept, this application also provides an electronic device, including a processor, a communication interface, a computer-readable storage medium as described above, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus; the processor is used to execute a program stored in the computer-readable storage medium.

[0027] Compared with the prior art, the beneficial effects of this application are as follows: A macroscopic topology map and microscopic physical field sequence were constructed. Microscopic features were extracted using 1D-ResNet and global max pooling. Dynamic attention weights were calculated through feature injection and a physical perception GNN. The physical perception GNN was trained using a loss function that included Kirchhoff flow conservation constraints. Thus, for the first time, dynamic and coupled analysis of microscopic physical damage details and macroscopic network cascade failures was achieved in pipeline reliability assessment. While ensuring computational efficiency, the physical reliability of the assessment results was also guaranteed.

[0028] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings.

[0029] The present application will be further described below with reference to the accompanying drawings. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating a pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to an embodiment of this application. Figure 2 This is a schematic diagram of a pipeline connectivity evaluation device based on a dual-scale physical perception graph neural network according to an embodiment of this application. Figure 3 This is a schematic diagram of the overall process of a pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] like Figure 1 As shown, one embodiment of this application provides a pipeline network evaluation method based on a dual-scale physical-informed graph neural network (DS-PIGNN), the method comprising the following steps: Acquire target pipeline network data, construct macroscopic topology map, and obtain microscopic physical field sequence data for each pipeline; For each pipeline, features of its microscopic physical field sequence are extracted using a one-dimensional residual convolutional neural network, and the microscopic health state vector representing the weakest link of the pipeline is obtained by global max pooling aggregation. The micro-health state vector is injected into the corresponding edge features of the macro-topology graph, and the dynamic attention weights between nodes are calculated using a physical perception graph neural network. The physical perception graph neural network is trained using a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. Based on the dynamic attention weights, neighborhood information is aggregated, and the failure probability of each pipeline and the connectivity reliability of each node are output.

[0034] The above-mentioned technical solution is the first to realize dynamic and coupled analysis of microscopic physical damage details and macroscopic network cascade failure in pipeline reliability assessment, and ensures the physical reliability of the assessment results while ensuring computational efficiency.

[0035] This solution possesses at least the following characteristics: 1. Constructing macroscopic topology maps and obtaining microscopic physical field sequences: This establishes the data foundation for "dual-scale" assessment. The macroscopic map describes the network topology, while the microscopic sequences (such as high-resolution wall thickness and stress distribution along the pipe axis) capture local damage details that traditional scalar properties cannot reflect, solving the pain point of "averaging processing masking local risks".

[0036] 2. Extracting microscopic features using 1D-ResNet (one-dimensional residual convolutional neural network) and global max pooling: 1D-ResNet can automatically identify local damage patterns such as pitting and stress concentration, like a "sliding probe." Global max pooling is key, as it extracts the most severe damage signal ("bottleneck effect") in the entire pipeline. The generated "microscopic health state vector" directly represents the pipeline's true load-bearing capacity bottleneck, providing accurate physical driving signals for subsequent cascading failure simulation.

[0037] 3. Feature Injection and Dynamic Attention Weight Calculation using a Physically Aware Graph Neural Network: This is the core of the "coupling." Injecting microscopic features into the macroscopic graph and using a physically aware graph neural network means that the network's connection weights are no longer static, but dynamically determined by the physical health of the pipeline. If a microscopic vector shows severe damage, the network automatically learns to reduce the attention weight of that pipeline to an extremely low level, mathematically simulating a physical "break" or "blockage," thus realizing the logic of "microscopic physical damage driving the interruption of macroscopic topological information flow." This method can simulate the pressure redistribution and cascading effects caused by single-point failures, and is closer to reality than static topology analysis.

[0038] 4. Training Based on Physical Constraints: The physical perception graph neural network must be trained using a loss function that incorporates Kirchhoff's flow conservation constraint, ensuring "physical perception." This forces the model to adhere to fundamental engineering physics laws (such as mass conservation) during learning and inference, resolving the issue of purely data-driven models ("black boxes") potentially outputting results that violate common sense physics. This significantly enhances the credibility of the evaluation results in engineering decision-making. Furthermore, the trained model, acting as a high-speed "surrogate model," exhibits extremely fast inference speeds, enabling real-time dynamic evaluation.

[0039] As can be seen, this application provides an intelligent evaluation scheme that can balance computational accuracy and computational efficiency, and unify microscopic physical details with macroscopic network dynamics.

[0040] Among them, the dual-scale physical perception graph neural network is a complete architecture consisting of a "microscopic damage feature encoder" and a "macroscopic cascaded failure reasoning network".

[0041] As a preferred technical solution, the microscopic physical field sequence data includes at least one of the following: the pipe axially discretized wall thickness sequence, the von Mises stress sequence, and the soil settlement displacement sequence.

[0042] This embodiment clarifies the specific physical dimensions for achieving "microscopic physical damage perception," making the technical solution more operable and clearly targeted for engineering applications. Specifically, the wall thickness sequence is directly related to corrosion thinning and residual strength; the von Mises stress sequence reflects the comprehensive stress level of the pipeline under internal and external loads, and is crucial for predicting plastic deformation or fatigue failure; the soil settlement displacement sequence is related to the pipeline's external loads and deformation. These sequences are key data that can be directly or indirectly obtained in engineering practice through technologies such as internal detection and distributed fiber optic sensing. Defining them concretizes the "microscopic physical field" from an abstract concept into measurable engineering parameters, solidifying the industrial application foundation of the technical solution.

[0043] As a preferred technical solution, the one-dimensional residual convolutional neural network extracts features using the following formula:

[0044] in, The extracted local feature sequence, For pipe e ij The microscopic physical field sequence matrix, This represents a one-dimensional convolution operation. and These are the learnable convolutional weights and biases, respectively.

[0045] This embodiment introduces a residual connection mechanism to effectively preserve the original, crucial physical damage signals during deep feature extraction, preventing the loss of important microscopic features due to "gradient vanishing" and ensuring the accuracy and stability of microscopic encoding. The "+" in the above formula... This refers to residual connections. In deep networks, directly learning complex physical damage mappings can lead to gradient problems, making the network difficult to train or causing the loss of input details. Residual connections allow the original input signal to be processed. By directly "bypassing" to the output, it ensures that no matter how deep the network is, the key original physical information (such as a sudden change in low wall thickness) will not be diluted or forgotten in the transformation of layers, thus ensuring the accuracy of the final extracted "microscopic health state vector".

[0046] As a preferred technical solution, the dynamic attention weight The calculation formula is:

[0047]

[0048] in, For the hydraulic feature vector of the node, For microscopic health state vectors, This represents a vector concatenation operation. The weight matrix is ​​a learnable matrix. Represents a node and nodes The unnormalized attention scores between them; Represents a node Its neighboring nodes The unnormalized attention scores between them; This is the transpose of the learnable attention weight vector (used to map the concatenated high-dimensional features to a scalar score). It is a Leaky Rectified Linear Unit, used to introduce non-linear features and prevent gradient vanishing during network training.

[0049] This embodiment discloses the specific mathematical implementation of the creative coupling mechanism of "microscopic damage features directly controlling macroscopic topological connectivity," transforming "feature injection" and "physical soft gating" from concepts into executable algorithms. This formula embodies the core algorithm of this application. It calculates the attention energy value e of node i to j. ij At that time, the micro-health state vector was clearly defined. With node features They are spliced ​​together and participate in the calculations. This means that the physical health condition of the pipeline ( ) directly determines its importance in information transmission. A key input to ")". Through end-to-end training, the network automatically learns: when When it contains severe damage characteristics, it should produce extremely small This leads to the normalized attention weights Approaching 0. This mathematically realizes "dynamically adjusting the topological connection strength according to the pipeline health status", simulating the "logical disconnection" of a failed pipeline, and is the cornerstone of the cascade failure simulation algorithm.

[0050] As a preferred technical solution, the loss function that includes Kirchhoff flow conservation constraints for: = +

[0051] in, For the mass conservation loss based on failure probability gating, This is the energy compatibility loss based on Bernoulli's equation.

[0052] This embodiment clarifies a dual guarantee mechanism for physical constraints, requiring not only flow conservation but also that pressure drop and flow conform to hydraulic relationships. This strengthens the physical consistency of the model output at the multi-physics level, making it closer to the hydraulic behavior of real pipe networks. The formula shows that during training, the model is simultaneously constrained to satisfy mass conservation (Kirchhoff's laws) and energy conservation (a modified form of Bernoulli's equation). Most data-driven models only consider flow balance, while this application further introduces energy compatibility constraints to ensure that the nodal head (pressure) predicted by the model conforms to the intrinsic hydraulic relationship between the pipe segment and the flow rate. This allows the generated "virtual" pipe network state (pressure distribution, flow allocation) to strictly adhere to physical laws even when simulating extreme scenarios with numerous random disruptions and dynamic topological changes, resulting in evaluation results with high engineering simulation fidelity.

[0053] As a preferred technical solution, in an ideal pipe network, for any node, inflow - outflow = water demand, and the mass conservation loss... The expression is:

[0054] Where V represents the set of all nodes in the pipeline network.

[0055] |V| represents the number of nodes in set V, which is the total number of nodes in the pipeline network.

[0056] v is the node currently being calculated (belonging to set V).

[0057] It refers to the set of upstream nodes pointing to node v, that is, the set of starting nodes of the pipe segment flowing into node v.

[0058] u is the upstream node of the inflow node v.

[0059] It is the set of downstream nodes pointed out from node v, that is, the set of terminal nodes of the pipe segment flowing out from node v.

[0060] w represents the downstream node from node v.

[0061] This represents the failure probability (pipeline fault / rupture probability) of pipe segment (u, v).

[0062] Let be the failure probability of pipe segment (v, w).

[0063] The design flow rate / estimated flow rate (the expected flow rate of water under normal, trouble-free conditions) for pipe segment (u, v).

[0064] The design flow rate / estimated flow rate for pipe segment (v, w).

[0065] This represents the water demand of node v (the water demand of users at this node).

[0066] This embodiment innovatively introduces a "failure probability gating" mechanism, enabling physical law constraints to be adaptively applied to dynamically changing pipeline topologies due to random pipeline failures, thus solving the problem that traditional physical constraints cannot handle topological abrupt changes. (1 - P) in the formula... fail The key innovation lies in the "failure gating term." In traditional hydraulic models, the mass conservation equation is based on a fixed topology where all pipes are intact. This application addresses a dynamic scenario where pipes may fail randomly. By using the model to predict the pipe survival probability (1 - P... fail ) as flow item Q est The coefficients of this loss function ensure that the design flow rate is only included in the conservation equation when the pipeline is predicted to have a high probability of survival; if it is predicted to have a high probability of failure, its flow rate contribution is "gated out." This forces the neural network to "understand" the dynamic relationship between physical laws and topological states during training, enabling it to perform physically guided self-supervised learning without real failure labels, significantly improving the model's generalization ability and prediction rationality in unknown failure scenarios.

[0067] As a preferred technical solution, the energy compatibility loss This indicator is used to assess the degree of inconsistency between head loss and actual flow-head relationship in the pipeline network (squared residual), and is part of the optimization objective; the smaller the value, the better. The expression is:

[0068] in, The water head (water pressure potential energy) at nodes i and j are respectively. This is the pipe resistance coefficient, which is related to pipe length, diameter, roughness, etc. The flow index; The failure probability predicted by the model; It is the set of all pipes (i.e., edges of the graph) in the pipeline network topology graph; The predicted flow rate of fluid in the pipe, usually expressed in units of 1000 liters. Or L / s, with the positive or negative sign indicating the direction of flow.

[0069] and The difference: This refers generally to the "probability of failure" physical / mathematical variable type output by the model. (Highly labeled) It is a specialized expression, specifically referring to "connection nodes". and nodes The failure probability of this specific pipe section. (Superscript) This is used to precisely locate specific pipe objects in the overall summation formula. The summation formula must be in the form of superscripts in order to calculate the contribution of each pipe separately.

[0070] This embodiment also applies the "failure probability gating" mechanism to the energy equation, ensuring that the hydraulic relationships (head difference and flow rate) in the remaining intact pipes strictly conform to engineering physics when simulating pipe failure scenarios, thus preventing the model from generating physically impossible pressure-flow combinations. Similar to the previous embodiment, the pressure drop calculation formula in Bernoulli's equation form is also multiplied by (1 - P). fail The gating factor means that for pipelines predicted to fail by the model, their pressure drop calculation is effectively ignored in the loss function, with the constraint focusing on pipelines that remain connected. This represents a lower penalty weight for energy compatibility loss for pipelines with a higher failure probability (because they are likely already failed and no longer involved in normal water transport). This further refines the physical constraints, ensuring that the model learns not only the rationality of the flow distribution but also the rationality of the pressure distribution, making the output of the entire evaluation system under complex dynamic conditions completely self-consistent and reliable at the hydraulic level.

[0071] As a preferred technical solution, the method further includes: Based on the Monte Carlo simulation framework, multiple randomized pipeline operating conditions were generated, and the overall reliability of nodes was statistically analyzed. The calculation formula is as follows:

[0072] in, The path connectivity depends on the survival probability of the weakest link in the path. The total number of Monte Carlo simulations (e.g.) Second-rate); is the flow regime index / discharge index in the fluid dynamics resistance equation, which is usually a constant (such as 1.852 or 2).

[0073] This embodiment combines a high-speed evaluation model with mature probabilistic and statistical methods to achieve efficient and quantitative assessment of system-level connectivity reliability of pipeline networks under the influence of random uncertainties (load, corrosion fluctuations), and outputs intuitive decision indicators. This embodiment represents an upgrade from "single-state assessment" to "system reliability quantification." Using a pre-trained model as a "surrogate model," thousands of Monte Carlo simulations are performed, completing in seconds what traditional simulations would take hours. Its quantification formula R... T Using the "barrel principle," the reliability of a node depends on the connectivity of the "most reliable path" among all possible paths from the source node to that node, and the path connectivity depends on the survival probability of the "weakest link" on that path. This method calculates R... T It is a probabilistic index that integrates network topology redundancy and component vulnerability. It can generate a pipeline network reliability heat map and accurately identify key vulnerable links in the system, providing direct and quantitative decision support for preventive maintenance and optimized sensor deployment.

[0074] As a preferred technical solution, the pipeline network is a water supply network, a gas network, or a heating network.

[0075] This embodiment limits the application type of the pipeline network, clarifies the direct application scenario of this application, and emphasizes its technical value in solving the major practical problem of safety assessment of urban lifeline projects (water supply and gas). Among them, water supply pipeline networks and gas pipeline networks are typical pressure flow pipeline network systems with extremely high requirements for safety and reliability, and their failure modes (leakage, rupture) are closely related to microscopic physical fields (corrosion, stress).

[0076] like Figure 2 As shown, one embodiment of this application provides a pipeline connectivity assessment device based on a dual-scale physical perception graph neural network, used to implement the pipeline connectivity assessment method based on a dual-scale physical perception graph neural network as described above. The device includes: The data mapping module is used to acquire target pipeline network data, construct a macroscopic topology map, and acquire pipeline microscopic physical field sequence data. The micro-encoding module is used to extract features of the pipeline's micro-physical field sequence using a one-dimensional residual convolutional neural network, and to obtain a micro-health state vector representing the weakest link of the pipeline through global max pooling aggregation. The topology reconstruction module is used to inject the micro-health state vector into the corresponding edge features of the macro-topology graph and calculate the dynamic attention weights between nodes using a physical perception graph neural network. The physical perception graph neural network is trained by a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. The evaluation output module is used to aggregate neighborhood information based on the dynamic attention weights and output the failure probability of each pipeline and the connectivity reliability of each node.

[0077] Regarding the apparatus in the above embodiments, the specific manner in which each unit module performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0078] Therefore, this application has at least the following technical effects: 1. A leap forward in assessment efficiency: Single assessments can be completed in milliseconds, enabling the reliability assessment of city-level pipeline networks to shift from offline, batch processing mode to one that can be integrated into real-time online early warning and decision support systems.

[0079] 2. Refined assessment: It can identify "local extreme defects" that traditional methods cannot capture, and realize risk positioning based on the "barrel effect".

[0080] 3. Dynamic evaluation: It can simulate cascading effects and connectivity changes under complex dynamic scenarios such as random failures and load fluctuations.

[0081] To make the above-mentioned pipeline connectivity assessment method and device based on dual-scale physical perception graph neural network clearer, the following further explanation is provided in conjunction with the above-mentioned preferred technical solutions.

[0082] Currently, it is widely believed in the field that the physical field sequences of stress, corrosion, and leakage inside pipelines (such as micro-stress distribution and dynamic evolution of corrosion pits) are only applicable to finite element analysis (FEM) of small-scale local pipelines. However, due to the massive scale of urban pipeline networks (such as over a million pipeline nodes), they must be oversimplified—the pipelines are abstracted as "edges" without internal structure, and their state is characterized only by scalar parameters (such as average pipe age and corrosion rate). Technicians generally believe that directly incorporating the physical field sequences of pipelines into the cascading failure analysis of the entire network is computationally infeasible, and therefore habitually ignore micro-physical details. This bias leads to existing methods (such as simplified hydraulic models like EPANET) assessing connectivity only from the perspective of network topology or single pipeline units, failing to capture the coupling effect between branches and main lines, the dynamic fusion of real-time sensor data, and the cascading propagation of multi-point failures. The assessment results are static and one-sided, failing to reflect the true reliability under complex environments.

[0083] In the field of engineering safety, technical personnel generally have doubts about deep learning models, believing that neural networks are essentially "probabilistic statistical fitting" and cannot strictly adhere to engineering laws such as the conservation of mass and momentum like physical numerical solvers. Therefore, the industry has formed a binary opposition: "either use AI for rough predictions (low accuracy) or use physical simulations for precise calculations (low efficiency)," with the widespread belief that "implanting complex physical equations into neural networks will inevitably lead to model divergence or training failure." This means that existing data-driven methods (such as traditional graph neural networks) can only handle topological structures and cannot incorporate physical laws, resulting in evaluation results lacking engineering credibility and making them unsuitable for decision support in high-risk scenarios.

[0084] The root of the aforementioned technological bias lies in the systemic cognitive limitations of this field regarding the integration of "micro" and "macro" scales: technical personnel have long misjudged the irreconcilable contradiction between "microscopic physical precision" and "macroscopic computational efficiency," while simultaneously viewing "physical constraints" and "data-driven" approaches as opposites. This has resulted in existing technologies remaining stuck on an inefficient path of "simplification and compromise," failing to achieve dynamic, refined, and reliable pipeline reliability assessments, and severely hindering the practical application of intelligent management of urban pipeline networks.

[0085] Based on this, this application proposes a pipeline connectivity assessment method based on a dual-scale physical perception graph neural network (i.e., a pipeline dynamic connectivity reliability assessment method based on a dual-scale physical perception graph neural network). This method combines microscopic physical simulation (internal physical field sequence analysis) with macroscopic graph networks, seamlessly coupling the microscopic features of the pipeline's internal physical field sequence (compressed into high-dimensional feature vectors by a 1D-ResNet encoder) with the pipeline cascade effect (processed by Macro-GAT). Joint training is achieved using a physical loss function (including Kirchhoff flow balance terms). This breaks the inherent perception of the "mutualistic nature of microscopic accuracy and macroscopic efficiency" and the "incompatibility of physical constraints and data-driven approaches," enabling cascade failure simulation and connectivity reliability quantification for complex pipelines such as water supply, gas, and heat systems, establishing a new technological paradigm for pipeline reliability assessment.

[0086] To address the shortcomings of existing technologies in microscopic damage identification, macroscopic cascade inference, and physical consistency, this application proposes a dual-scale evaluation method (i.e., a pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network) that can simultaneously achieve both computational efficiency and physical accuracy. This method comprises five core steps: heterogeneous graph construction, microscopic damage feature extraction, macroscopic cascade failure inference, physical constraint training, and fuzzy reliability assessment; the specific process steps are as follows: Figure 3 As shown.

[0087] Step S1: Construct a heterogeneous attribute graph of the pipeline network containing "serialized physical fields". Unlike traditional graphical models that only store scalar attributes of pipes (such as pipe length and pipe diameter), this step constructs a heterogeneous graph that can reflect the internal details of the pipe.

[0088] The core feature of the "serialized physical field" is: "high-resolution spatial discretization + dynamic distribution of physical quantities along the pipe axis". The microscopic physical field sequence data is the core feature of the "serialized physical field", emphasizing the dynamic distribution of physical quantities along the pipe axis through high-resolution spatial discretization.

[0089] Traditional techniques (such as patent document CN115654381A) treat pipe attributes (diameter, length, wall thickness) as scalars from a "homogeneous unit perspective," treating the entire pipe as a single entity with attributes being global statistical values ​​(e.g., "average wall thickness 10mm," "pipe diameter 200mm"), ignoring physical differences at different locations within the pipe. In contrast, the "serialized physical field" approach adopts a "spatial continuum perspective," viewing the pipe as a carrier with axial (or radial / circumferential, but with a focus on axial longitudinal analysis) spatial dimensions. Physical quantities (wall thickness, stress, corrosion depth, fluid pressure, etc.) are discretely collected along the axial direction at a specific resolution Δx (e.g., per meter, per 0.5 meters, or even higher precision), forming a sequential mapping between location and physical quantities.

[0090] Taking the physical field of wall thickness serialization as an example: For pipes ( Indicates the connection node and nodes The physical field of the wall thickness sequence of this specific pipe segment can be formally represented as:

[0091] In the formula: These are the position coordinates along the axial direction of the pipe (relative to the starting point). (distance) yes The wall thickness at that location.

[0092] For example, a 100-meter-long pipe Meter sampling yielded 100 wall thickness data points, forming a sequence.

[0093] Among them, the local mutation value in the sequence (such as a wall thickness of 5.1 mm at 45 meters) directly corresponds to the local short root weight of "inaccurate response".

[0094] In this implementation plan, the pipeline network detection data (pipe diameter, material, burial depth, soil information) is first imported into a graph database (such as Neo4j).

[0095] Mapping pipeline network data into a multi-dimensional attribute graph (corresponding to the aforementioned macroscopic topology graph): .

[0096] Node set : Basic attributes of storage nodes, including elevation Basic water demand Node type (source / sink / valve).

[0097] side Storage pipe diameter, material, burial depth, and corrosion rate.

[0098] Microscopic physical feature extraction: For each pipe It not only records scalar attributes (pipe diameter, wall thickness), but also regards the stress distribution and corrosion depth distribution along the pipe axis as a "one-dimensional spatial sequence". It is also necessary to extract its axial sequenced physical field: based on internal detection technology (such as leakage magnetic field / ultrasonic detectors moving along the pipeline and collecting data at intervals of Δx=0.5 meters), distributed fiber optic sensors (deployed along the pipeline to monitor axial stress / temperature distribution in real time) or high-resolution detection data, the location-value sequence of physical quantities (wall thickness, corrosion depth, stress) inside the pipeline is generated.

[0099] For example, pipes The physical field for the wall thickness sequence is:

[0100] Among them, local minimum value It is a related indicator for identifying the shortcomings of "inaccurate response", and scalar attributes (such as average wall thickness) cannot capture this kind of information.

[0101] Serialized edge features : Each pipe Discretize along the axial direction Each physical element (e.g., one cross-section per meter). For pipes Construct the micro-feature matrix:

[0102] In the formula, For the feature dimension, it includes the following: Real-time wall thickness (considering corrosion thinning); Von Mises stress; Soil settlement and displacement; This data structure can preserve the "health fingerprint" of the pipeline, enabling subsequent models to identify hidden risks such as "the overall pipeline is in good condition, but there is a fatal crack in the middle".

[0103] The "serialized physical field" in step S1 has the value of identifying the fatal flaw of the "barrel effect". The local weak point of the pipeline determines the overall risk because the core logic of pipeline failure (such as leakage or rupture) is "local extreme value triggers overall failure". That is, the load-bearing capacity and life of the entire pipeline are often determined by the weakest "local location" (such as the minimum wall thickness or the maximum stress point), rather than the global average value.

[0104] Taking corrosion-induced wall thickness reduction as an example, traditional scalar attributes (such as "average wall thickness 10mm") can only reflect the overall trend and cannot identify a "fatal weakness" where the wall thickness drops to 3mm due to corrosion. However, "serialized physical fields" can accurately locate the "weak point" through high-resolution spatial sequences (such as abrupt change points in the wall thickness sequence or peak points in the stress sequence), providing a data foundation for "barrel effect" risk assessment (such as calculating residual strength based on the minimum wall thickness, rather than the average wall thickness). This ability to identify local risks is unattainable by existing technologies (which treat the pipeline as a homogeneous unit), directly solving the technical pain point that "the entire pipeline appears safe, but in reality, it fails due to local defects."

[0105] Step S2: Construct a microscopic damage feature encoder To address the problem of "averaging down local risks" in existing technologies (i.e., the inability to identify local pitting or stress concentration in long-distance pipelines), this step designs a "feature-injection-based dynamic attention coupling mechanism." This mechanism is not a simple network concatenation, but rather directly uses microscopic physical damage features as "edge attributes" and "attention gating factors" of the macroscopic graph network, realizing the logical drive from microscopic damage to macroscopic topological fracture.

[0106] The specific implementation process of step S2 includes the following two sub-steps: Dimensionality reduction encoding of microsequence features The physical field sequence matrix discretized along the axial direction of the pipeline is obtained using a one-dimensional residual convolutional neural network (1D-ResNet). Perform feature extraction and dimensionality reduction.

[0107] S2-1 Local Feature Convolution (Sliding Probe) pipe Physical field sequence matrix generated by discretization along the axis (Including dimensions such as wall thickness, stress, and settlement) are input into a 1D-ResNet model. The one-dimensional convolutional kernel is treated as a "sliding probe," sliding along the pipe length on the sequence feature matrix to automatically identify local physical damage patterns (such as pitting and stress concentration points). To prevent gradient vanishing in deep networks and preserve the original physical signals, a residual connection mechanism is introduced:

[0108] In the formula: For pipelines The original sequence input, Indicates the convolution operation; and For learnable convolutional weights and biases; This is the extracted local feature sequence.

[0109] S2-2 Global Feature Aggregation Based on the "Short Plank Effect" To capture the weakest link that determines the overall reliability of the pipeline (i.e., the shortest plank in the "barrel effect"), a global max pooling operation is used to analyze the local feature sequences. Dimensionality reduction and aggregation were performed to extract the most severe damage signal from the entire pipe:

[0110] In the formula: For the fixed dimension of the output The "pipeline health state vector (corresponding to the aforementioned micro-health state vector)". This operation will... Compression of complex spatiotemporal sequences of dimensionality into A high-level feature vector is generated. This vector not only contains the overall state of the pipe, but also focuses on preserving the feature information of the "most dangerous section" (such as the wall thickness or maximum stress value at the thinnest point in the entire pipe), which is prepared to be injected into the graph neural network in step S3.

[0111] Step S3: Construct a macroscopic cascading failure inference network This step receives the output of step S2. This study utilizes a physical perception graph attention network (i.e., a physical perception graph neural network) to simulate the propagation and cascading of local failures in the network topology. The core logic of this step is "feature injection and attention coupling," which means using microscopic damage features to directly control the connectivity weights of the macroscopic topology, thereby achieving cross-scale cascading failure simulation.

[0112] The specific implementation process of step S3 includes the following two sub-steps: S3-1 Macro-level attention coupling based on feature injection This is a crucial aspect of dual-scale coupling. This application does not use a static graph structure, but instead dynamically computes the dependencies between nodes through a "physical soft gating" mechanism. (Computation nodes...) For nodes When the attention coefficient (connection weight) is adjusted, it shows that microscopic damage features are fused. : 1. Energy score calculation (feature fusion)

[0113] 2. Attention system normalization (weight allocation)

[0114] In the formula: The hydraulic feature vector of the node; The physical health characteristics of the pipeline; The weight matrix is ​​a learnable matrix; This is a vector concatenation operation; These are dynamic connectivity weights (corresponding to the aforementioned dynamic attention weights); dynamic attention weights are the inter-node connection weights calculated by the physical perception graph neural network. Its value changes dynamically.

[0115] In this formula, It was directly used as a key value in the attention score. The calculation. This is a gating mechanism driven by physical states; if the values ​​extracted in step S2... For features containing severe damage (e.g., extremely thin wall thickness), the neural network automatically learns the weights corresponding to those features through end-to-end training. It should produce a very small attention score; the neural network automatically learns to adjust the weights through end-to-end training. Adjustment, resulting in the calculated Minimal, thus making the normalized .

[0116] This realizes the logic of "microscopic physical damage driving macroscopic topological fracture". This mechanism can simulate the cascading failure (such as pressure redistribution) after "pipe break" at the mathematical level without actually removing the mesh at the physical level. It achieves the organic unity of microscopic physical field and macroscopic network topology, thereby blocking the transmission of information (or pressure) between nodes.

[0117] S3-2 Status Update and Failure Prediction Based on the aforementioned dynamic weights that incorporate physical damage information It aggregates neighborhood information and outputs the final state of each pipeline through a nonlinear activation layer:

[0118] In the formula: The sigmoid activation function is used to map the output value of a neural network to... Within the interval, to represent the probability magnitude. This is the learnable weight matrix for the output layer. This is a neighborhood feature aggregation function in a graph neural network (e.g., a weighted summation operation based on weights).

[0119] This represents the output failure probability. Because... The micro-damage has already been redistributed (e.g., the weight of the fractured pipe is 0), and the output here is... The impact of the failure on surrounding nodes is automatically included in the cascading failure effect (i.e., a single failure leads to the reconstruction of the surrounding flow field and pressure redistribution).

[0120] In this step, if the pipeline The microscopic features show severe damage, and the network will automatically learn to adjust the weights. The value is reduced to near 0, simulating a logical "disconnection" in the path, thus achieving dynamic topology reconstruction for risk perception. By aggregating neighborhood information, the failure probability of each pipe is output through a sigmoid activation function. The connectivity state vector of the nodes.

[0121] Step S4: Model training under physical consistency constraints To prevent purely data-driven models from producing results that violate physical principles (e.g., predicting a pipe break but still allowing a large flow rate), this step constructs a composite physics loss function that includes a failure-gated term. The composite loss function is as follows:

[0122] In the formula, The total loss function is the objective function that is ultimately minimized during model training.

[0123] This is the data-driven portion of the mean squared error loss. It is typically used in supervised learning to measure the error between the model's predicted values ​​(flow rate, head, failure probability, etc.) and the true labels.

[0124] This is a hyperparameter for the physical loss weighting coefficient, used to balance the importance of the data fit term (MSE) and the physical consistency term. It is typically set to 0.1~10 and can be fine-tuned based on experiments.

[0125] This is the core component of the physical loss term, used to force the model to obey physical laws. It prevents the model from outputting results that violate physical principles. Based on Kirchhoff's law of conservation of flow, it mandates that the predicted flow distribution satisfy the conservation of node mass. If the model predicts a pipe failure (i.e., the probability of connectivity failure)... If the effective flow rate through the pipe is 0, then the effective flow rate should approach 0.

[0126] in, Includes mass conservation Mass Balance and Energy Conservation (Energy Conservation) two items.

[0127] The loss is determined by the conservation of mass based on Kirchhoff's law of conservation of flow. Requirement: The total flow entering the node equals the total flow leaving the node plus the node's water demand. .

[0128] This is the energy conservation loss, also known as the energy compatibility loss. It measures whether the head difference in each pipe satisfies the pipe hydraulic equations. Mass conservation loss based on failure gating The loss term is based on Kirchhoff's law of conservation of flow (KCL), but unlike existing technologies that only use it for static flow estimation, this application introduces... As a gating term, it is used to adapt to topology mutation scenarios caused by random pipeline failures. For the pipeline failure probability to remain consistent with previous values, this typically means... That is, the failure probability of pipe (i,j).

[0129]

[0130] In the formula: V is the set of all nodes in the pipeline network.

[0131] |V| represents the number of nodes in set V, which is the total number of nodes in the pipeline network.

[0132] v is the node currently being calculated (belonging to set V).

[0133] It refers to the set of upstream nodes pointing to node v, that is, the set of starting nodes of the pipe segment flowing into node v.

[0134] u is the upstream node of the inflow node v.

[0135] It is the set of downstream nodes pointed out from node v, that is, the set of terminal nodes of the pipe segment flowing out from node v.

[0136] w represents the downstream node from node v.

[0137] This represents the failure probability (pipeline fault / rupture probability) of pipe segment (u, v).

[0138] Let be the failure probability of pipe segment (v, w).

[0139] The design flow rate / estimated flow rate (the expected water volume to pass through under normal, trouble-free conditions) for pipe segment (u, v).

[0140] The design flow rate / estimated flow rate for pipe segment (v, w).

[0141] This represents the water demand of node v (the water demand of users at this node).

[0142] This formula introduces As a gating term, this forces the neural network to "understand" that only structurally intact pipes can transmit flow. If a rupture is predicted but a large flow rate is still expected, This will cause a surge, thereby correcting the network parameters, enabling the model to perform physically self-supervised learning on unlabeled data. Conventional techniques lack this gating mechanism and cannot handle topological abrupt changes such as "disappearing pipes."

[0143] Energy compatibility loss based on Bernoulli's equation: To further enhance physical consistency, a pressure drop constraint is introduced:

[0144] In the formula: These are the head values ​​for nodes i and j, respectively. This is the pipe resistance coefficient. It is the flow index (usually taken as 1.852 or 2).

[0145] This step ensures that the flow rate in the undamaged pipeline maintains a strict hydraulic correspondence with the pressure difference at the nodes. This is to guarantee that the virtual pipeline state generated by deep learning does not violate physical laws when simulating numerous random failure scenarios, and is a reasonable constraint for generative simulation.

[0146] The proposed solution deals with dynamic time-varying topologies caused by random pipeline failures, which can only be achieved through the aforementioned "failure gating" mechanism.

[0147] Step S5: Reliability Quantitative Assessment Based on Fuzzy Membership Degree The DS-PIGNN (Dual-Scale Physical Perception Graph Neural Network) trained in step S4 is used as a high-speed surrogate model to replace traditional finite element / hydraulic simulation software (such as EPANET) for Monte Carlo simulations, thus solving the computational efficiency bottleneck.

[0148] 1. Fast stochastic simulation: Utilizing a pre-trained surrogate model, input... Next (such as) (Number of times) Randomly generated pipeline operating conditions. Each simulation includes random pressure fluctuation input and micro-corrosion distribution input, and the model outputs the failure probability distribution of the entire pipeline network within milliseconds.

[0149] 2. Fuzzy path connectivity calculation: In the first... In this simulation, the source point is defined. to the remittance point any path The overall connectivity of a path is not simply equal to the product of the connectivity of all its segments, but rather follows the "barrel principle"—that is, the overall performance of a path is determined by its weakest segment.

[0150] In the formula, For the first In this simulation, a specific path Comprehensive connectivity (value) ).

[0151] For path The set of all pipe segments contained above.

[0152] Connecting nodes output by the model and The probability of pipe section failure.

[0153] In the first The probability that the pipe section remains intact under the next operating condition (i.e., the complement of the failure probability).

[0154] min( The function takes the minimum value, which reflects the "barrel effect", that is, the reliability of the path is limited by the worst-performing (highest failure probability) pipe segment in the path.

[0155] 3. Overall node reliability: Sub-random simulation (considering load fluctuations), calculation nodes The average maximum connectivity (i.e., the connectivity reliability of nodes):

[0156] In the formula, The overall connectivity reliability of the target node.

[0157] The total number of Monte Carlo random simulations (e.g.) Second-rate).

[0158] It is the set of all possible connected paths from the source node to this node.

[0159] The joint probability is that all paths fail (i.e., the system is disconnected).

[0160] This represents the probability that there exists at least one path in the system that connects to the source.

[0161] Through the The average of the simulation results is used to obtain the expected reliability of the node.

[0162] 4. Output Results: Generate a heat map of pipeline connectivity reliability, visually displaying the water / gas supply safety levels in different areas of the city. It also identifies critical vulnerability nodes in the system—those whose failure would lead to widespread downstream reliability issues. The sudden drop in pipe length provides decision support for prioritizing maintenance and sensor deployment.

[0163] To verify the significant improvement in computational efficiency of the method described in this application, a comparative test was conducted on a standard pipeline network model.

[0164] 1. Experimental setup Test subject: The C-Town Network benchmark model for water supply network (containing 388 nodes, 429 pipes, and 7 water tanks) was selected.

[0165] Hardware environment: Intel i9-12900K CPU (for traditional emulation) vs. NVIDIA RTX 3090 GPU (used for model inference in this application) .

[0166] Task objective: Execute 100,000 Monte Carlo random simulations (each simulation requires randomly changing the physical parameters of the pipeline, such as corrosion depth, and determining the overall network connectivity and node water pressure reliability).

[0167] 2. Comparison Method Existing technology (benchmark group): Utilizing the commonly used hydraulic simulation software EPANET 2.2 in conjunction with Python scripts. The process is as follows: generate random parameters → modify the INP file → call the EPANET engine to solve → analyze the results.

[0168] The method used in this application (experimental group) employs a pre-trained DS-PIGNN surrogate model. The process is as follows: generate random tensors → GPU batch inference → output reliability metrics.

[0169] 3. Experimental Results Data

[0170] The accuracy is based on the EPANET calculation results as true values, and is used to determine the consistency of whether a node has failed.

[0171] 4. Comparative Test Conclusions Experimental data show that, while ensuring assessment accuracy (error <1.5%), the method proposed in this application reduces the total time for network-wide reliability assessment from "hours" to "seconds," improving efficiency by more than three orders of magnitude (approximately 1200 times). This significant efficiency leap transforms the previously offline monthly / quarterly assessment of urban pipe networks into an integrated real-time online disaster early warning system, demonstrating outstanding substantive features and significant progress.

[0172] In summary, this application possesses at least the following characteristics: 1. Dual-scale coupled architecture (demonstrating innovation compared to single-scale graph neural networks): Existing technologies often simplify pipelines to scalar edges. This application, through the combination of "1D-ResNet + GAT", achieves for the first time cross-scale feature extraction from "microscopic cross-sectional damage of pipelines" to "macroscopic cascade failure of pipeline networks". This enables the model to identify hidden risks of localized corrosion occurring as little as 1 meter in a long pipeline.

[0173] 2. Physical consistency constraints (demonstrating creativity compared to purely data-driven AI): This application introduces a physical loss function. This is a specific application of "AI for Science" in the pipeline network field. It solves the problem that the prediction results of traditional "black box" models are unreliable when there are no historical fault samples (zero-shot), and ensures that the evaluation results strictly comply with Kirchhoff's laws.

[0174] 3. Balance between computational efficiency and accuracy: Compared to the traditional Monte Carlo method which requires several hours of computation, the model trained in this application, as a surrogate model, only requires milliseconds to infer the state of the entire network, improving efficiency by 3-4 orders of magnitude and making real-time dynamic evaluation possible.

[0175] Based on the same inventive concept, this application also provides a computer-readable storage medium storing one or more programs, which, when executed, implement the pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network as described above.

[0176] Based on the same inventive concept, such as Figure 4 As shown, this application also provides an electronic device, including a processor, a communication interface, a computer-readable storage medium as described above, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus; the processor is used to execute a program stored in the computer-readable storage medium.

[0177] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0178] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0179] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0180] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0181] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0182] The parts not mentioned in the above embodiments are the same as or can be implemented using existing technologies, and will not be further described here.

[0183] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network, characterized in that, The method includes: Acquire target pipeline network data, construct macroscopic topology map, and obtain microscopic physical field sequence data for each pipeline; For each pipeline, features of its microscopic physical field sequence are extracted using a one-dimensional residual convolutional neural network, and the microscopic health state vector representing the weakest link of the pipeline is obtained by global max pooling aggregation. The micro-health state vector is injected into the corresponding edge features of the macro-topology graph, and the dynamic attention weights between nodes are calculated using a physical perception graph neural network. The physical perception graph neural network is trained using a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. Based on the dynamic attention weights, neighborhood information is aggregated, and the failure probability of each pipeline and the connectivity reliability of each node are output.

2. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 1, characterized in that, The microscopic physical field sequence data includes at least one of the following: the pipe axial discretization wall thickness sequence, the von Mises stress sequence, and the soil settlement displacement sequence.

3. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 1, characterized in that, The one-dimensional residual convolutional neural network extracts features using the following formula: in, The extracted local feature sequence, For pipe e ij The microscopic physical field sequence matrix, This represents a one-dimensional convolution operation. and These are the learnable convolutional weights and biases, respectively.

4. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 1, characterized in that, The dynamic attention weight The calculation formula is: in, The hydraulic feature vector of the node; This represents a vector of microscopic health states. This represents a vector concatenation operation; The weight matrix is ​​a learnable matrix; Represents a node and nodes The unnormalized attention scores between them; Represents a node Its neighboring nodes The unnormalized attention scores between them; This is the transpose of the learnable attention weight vector; It is a linear rectified activation function with leakage.

5. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 1, characterized in that, The loss function that includes Kirchhoff flow conservation constraints for: = + in, For the mass conservation loss based on failure probability gating, This is the energy compatibility loss based on Bernoulli's equation.

6. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 5, characterized in that, The mass conservation loss The expression is: Where V is the set of all nodes in the pipeline network; |V| is the number of nodes in set V, that is, the total number of nodes in the pipeline network; v is the node currently being calculated. Let be the set of upstream nodes pointing to node v, that is, the set of starting nodes of the pipe segment flowing into node v; u is the upstream node flowing into node v. Let w be the set of downstream nodes pointed to by node v, that is, the set of terminal nodes of the pipe segment flowing out from node v; w is the downstream node flowing out from node v. Let be the failure probability of pipe segment (u, v); Let be the failure probability of pipe segment (v, w); The design flow rate or estimated flow rate for pipe segment (u, v); The design flow rate or estimated flow rate of the pipe segment (v, w); Let V be the water demand of node v.

7. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 5, characterized in that, The energy compatibility loss The expression is: in, Let i and j be the water heads, i.e., the water pressure potential energy, respectively. This refers to the pipeline resistance coefficient. The flow index; Indicates the connection node and nodes The failure probability of this pipe section; This is the set of all pipes in the pipeline network topology diagram; Indicates a pipe; This represents the predicted flow rate of the fluid within the pipe.

8. The pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to claim 1, characterized in that, The method further includes: Based on the Monte Carlo simulation framework, multiple randomized pipeline operating conditions were generated, and the overall reliability of nodes was statistically analyzed. The calculation formula is as follows: in, The path connectivity depends on the survival probability of the weakest link in the path. The total number of Monte Carlo random simulations; This refers to the flow regime index or flow rate index in the fluid dynamics resistance equation.

9. A pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network according to any one of claims 1-8, characterized in that, The pipeline network is a water supply network, a gas network, or a heating network.

10. A pipeline connectivity assessment device based on a dual-scale physical perception graph neural network, used to implement the pipeline connectivity assessment method based on a dual-scale physical perception graph neural network as described in any one of claims 1-9, characterized in that, The device includes: The data mapping module is used to acquire target pipeline network data, construct a macroscopic topology map, and acquire pipeline microscopic physical field sequence data. The micro-encoding module is used to extract features of the pipeline's micro-physical field sequence using a one-dimensional residual convolutional neural network, and to obtain a micro-health state vector representing the weakest link of the pipeline through global max pooling aggregation. The topology reconstruction module is used to inject the micro-health state vector into the corresponding edge features of the macro-topology graph and calculate the dynamic attention weights between nodes using a physical perception graph neural network. The physical perception graph neural network is trained by a loss function that includes Kirchhoff flow conservation constraints, and the dynamic attention weights are jointly determined by the node hydraulic features and the micro-health state vector. The evaluation output module is used to aggregate neighborhood information based on the dynamic attention weights and output the failure probability of each pipeline and the connectivity reliability of each node.

11. A computer-readable storage medium storing one or more programs, characterized in that, When one or more of the programs are executed, the pipeline connectivity evaluation method based on a dual-scale physical perception graph neural network as described in any one of claims 1-9 is implemented.

12. An electronic device comprising a processor, a communication interface, a computer-readable storage medium as described in claim 11, and a communication bus; wherein, The processor, communication interface, and computer-readable storage medium communicate with each other via a communication bus; Its features are, The processor is used to execute programs stored in a computer-readable storage medium.

Images