Power line tower group structure mechanics response and multi-source remote sensing parameter fusion analysis method based on graph neural network

Abstract

The present application belongs to the technical field of power transmission line structure health monitoring, and discloses a power transmission line tower group structure mechanics response and multi-source remote sensing parameter fusion analysis method based on graph neural network. The method aims at the problems of existing finite element simulation modeling complexity, incomplete sensor monitoring, lack of physical mapping of multi-source remote sensing data and tower mechanics response, acquires and pre-processes tower structure parameters and multi-source remote sensing data, and constructs a tower group graph data model; taking span as a unit, combining conductor mechanics equation to calculate load and unbalanced tension, building a graph neural network model integrated with tower axial force and bending moment bearing capacity physical regularization constraint; after training by historical working condition data, inputting new remote sensing and meteorological data to reconstruct tower group mechanics response field, generating risk early warning according to safety threshold. The present application realizes deep fusion of mechanics mechanism and remote sensing data, quickly and accurately reconstructs tower group mechanics response field, and outputs the results conforming to the structure mechanics balance law, thereby providing technical support for power transmission line health monitoring.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line structural health monitoring technology, specifically involving a method for analyzing the structural mechanical response of power transmission line tower groups and the fusion of multi-source remote sensing parameters based on graph neural networks. Background Technology

[0002] Structural safety assessment of transmission line tower groups is a core task of power grid operation and maintenance. Current main analytical methods include:

[0003] Finite element simulation: It can perform high-precision calculations based on accurate geometric and material models, but the modeling is complex, the calculation cost is high, and it needs to be calculated for each specific working condition, making it difficult to achieve rapid evaluation under multiple working conditions.

[0004] On-site sensor monitoring: The data is reliable, but due to the limitations of installation and maintenance costs, it can only cover key points and cannot achieve comprehensive status perception of the tower group.

[0005] In recent years, remote sensing technologies such as spaceborne synthetic aperture radar interferometry and optical satellite imagery have provided large-scale, periodic observational data. However, these remote sensing data reflect changes in surface deformation around the tower base or structural features in the imagery, lacking a direct physical mapping relationship with the structural mechanical response of the tower itself. How to effectively integrate multi-source remote sensing data with conductor state equations based on mechanical mechanisms to construct an analytical method capable of rapidly assessing the mechanical state of tower groups is a pressing technical problem to be solved in this field. Summary of the Invention

[0006] To address the challenges of complex and costly finite element simulation modeling, which hinders rapid mechanical assessment of power transmission line tower groups under various operating conditions, and the limitations of point-based sensors in covering entire tower groups, as well as the lack of direct physical mapping between multi-source remote sensing data and tower structural mechanical responses, this invention aims to provide a graph neural network-based method for fusing structural mechanical responses and multi-source remote sensing parameters in power transmission line tower groups. This method integrates structural mechanical responses with multi-source remote sensing parameters. The invention uses span spacing as a unit to couple the conductor's mechanical state equation to construct load calculation logic. Combined with a tower group graph network topology that differentiates tower types, multi-source remote sensing data is integrated into the graph neural network as auxiliary identification features. A physical regularization mechanism is constructed through tower axial force, bending moment, and bearing capacity constraints. This solves the long-standing industry problem of deep integration between remote sensing data and tower mechanical mechanisms, enabling rapid, reasonable, and accurate reconstruction of the structural mechanical response field of tower groups while ensuring that the model output conforms to structural mechanical equilibrium laws. This provides technical support for large-scale, high-efficiency, and high-reliability structural health monitoring of power transmission line tower groups.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters, the method comprising the following steps:

[0009] Step S1: Obtain the geographic coordinates, structural design parameters and multi-source remote sensing observation data of the towers in the target area; preprocess the multi-source remote sensing observation data to obtain a remote sensing observation sequence that matches the tower location.

[0010] Step S2: Based on the geographical coordinates and structural design parameters of the towers, each tower is used as a node in the graph network. The graph network edges of adjacent towers are constructed based on the electrical and mechanical topology of the transmission line. The edge attributes are defined by combining the differences in tower type and the edge weights are determined by physical parameters to obtain the tower group graph data model.

[0011] Step S3: Using the span of the structural design parameters as the calculation unit, based on regional meteorological data and the mechanical state equation of the conductor, calculate the conductor load, tension parameters and unbalanced tension on both sides of the tower within the span to obtain the mechanical transmission parameters of the tower.

[0012] Step S4: Based on the tower group graph data model and the tower mechanical transfer parameters, initialize the graph network node features and edge features, and construct a physical information fusion graph neural network model that incorporates the physical regularization constraints of the tower axial force bending moment bearing capacity.

[0013] Step S5: Train the physical information fusion graph neural network model using historical working condition data to obtain the trained tower mechanical response prediction model.

[0014] Step S6: Input the newly acquired remote sensing observation data and meteorological data into the tower mechanical response prediction model, output the mechanical response index of a single tower, and reconstruct the mechanical response field of the tower group structure.

[0015] Step S7: Based on the safety operation standards for transmission lines, a threshold determination is made for the mechanical response field, and risk warning information is generated for towers that exceed the safety threshold.

[0016] Preferably, the multi-source remote sensing observation data in step S1 includes synthetic aperture radar interferometry deformation data and optical image change data; the preprocessing involves performing small baseline set processing and geocoding matching on the synthetic aperture radar interferometry deformation data, and performing scale-invariant feature transformation registration and template matching extraction on the optical image change data.

[0017] Preferably, the tower type mentioned in step S2 is distinguished by numerical coding as a straight tower, a tension tower, and an angle tower; the edge weight is used to characterize the mechanical transmission strength between towers and is obtained by physical parameter calculation or model adaptive learning.

[0018] Preferably, the mechanical state equation of the conductor in step S3 is the catenary equation or the parabolic equation, and the horizontal tension, vertical load and unbalanced tension of the conductor are calculated by combining wind speed, ice thickness and temperature meteorological parameters.

[0019] Preferably, the physical information fusion graph neural network model in step S4 includes a feature embedding layer, a multi-head graph attention layer, a node update layer, and an output layer in sequence; the multi-head graph attention layer calculates attention weights based on node embedding features and edge embedding features.

[0020] Preferably, the total loss function of the physical information fusion graph neural network model in step S4 is:

[0021] ;

[0022] in, , where is the mean square error between the predicted mechanical response and the true value;

[0023] , which is the physical regularization term;

[0024] This is the balance coefficient; This refers to the number of tower nodes; For sample index; For the first The predicted mechanical response vector of each sample; For the first The true mechanical response vector of each sample; For the first Predicted axial force on the base tower; This refers to the axial compressive bearing capacity of the tower. For the first Predicted bending moment of base tower; This refers to the bending bearing capacity of the tower.

[0025] Preferably, the node features in step S4 include tower coordinates, tower type code, nominal height, and remote sensing observation sequence; the edge features include span, height difference, conductor type code, conductor horizontal tension, and vertical load.

[0026] Preferably, the historical operating condition data mentioned in step S5 is generated by calculating and finite element simulation based on historical meteorological parameters combined with the mechanical state equation of the conductor, and includes training samples, verification samples and test samples.

[0027] Preferably, the mechanical response indicators of the single-base tower in step S6 include the horizontal displacement at the top of the tower, the vertical displacement at the top of the tower, and the maximum stress of the main material.

[0028] Preferably, the safety threshold in step S7 includes the tower tilt limit and the main material stress limit.

[0029] The present invention provides a method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters, which achieves several technical advantages:

[0030] 1. Load transfer modeling based on mechanical mechanism: This invention uses span as the basic unit, uses the mechanical state equation of conductor to calculate the load, and transfers the unbalanced tension on both sides of the tower through a graph network, so that the load calculation and transfer process conforms to the basic physical facts of overhead lines.

[0031] 2. Graph Neural Network with Physical Constraints: This invention introduces the relevant equations of axial force and bending moment of the tower as regularization terms into the loss function, so that the model output satisfies the mechanical equilibrium constraints of the tower structure and reduces the probability of non-physical predictions.

[0032] 3. Auxiliary application of remote sensing data: This invention uses InSAR foundation deformation and optical image structure changes as nodal feature inputs to identify abnormal foundation settlement conditions or assist in boundary condition correction, providing supplementary information for tower condition assessment. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to the present invention.

[0034] Figure 2 This is a schematic diagram illustrating the construction of the tower group diagram data model of the present invention.

[0035] Figure 3 This is a structural diagram of the physical information fusion GNN model of the present invention. Detailed Implementation

[0036] The following detailed implementation of the present invention, based on specific embodiments, provides a method for analyzing the structural mechanical response of power line tower groups and the fusion of multi-source remote sensing parameters. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0037] Example 1: Implementation of a method for analyzing the structural mechanical response of power line tower groups and the fusion of multi-source remote sensing parameters based on graph neural networks.

[0038] Combined with appendix Figures 1-3 As shown, this invention provides a method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters.

[0039] like Figure 1 As shown, Figure 1 This is a flowchart of the method of the present invention.

[0040] This invention provides a method for fusing and analyzing the structural mechanical response of power line tower groups with multi-source remote sensing parameters based on graph neural networks. The aim is to construct a graph neural network model based on the conductor's mechanical state equation to fuse remote sensing observation data with the structural stress model, thereby reconstructing the mechanical response field of the tower group. The method includes the following steps:

[0041] Step S1: Data acquisition and preprocessing.

[0042] Step S1 acquires the geographic coordinates, structural design parameters, and multi-source remote sensing observation data of the towers within the target area. The multi-source remote sensing observation data is preprocessed to obtain a remote sensing observation sequence that matches the tower location.

[0043] In step S1, the multi-source remote sensing data includes synthetic aperture radar interferometry deformation data and optical image change data. The preprocessing involves small baseline set processing and geocoding matching of the synthetic aperture radar interferometry deformation data, and scale-invariant feature transformation registration and template matching extraction of the optical image change data. The specific process is as follows:

[0044] Acquire the geographic coordinates and structural design parameters of each tower within the target area, including tower type, nominal height, conductor type, span, and elevation difference, as well as multi-source remote sensing observation data. Perform noise reduction, georegistration, and time alignment on the remote sensing data to form an observation sequence matched to the tower locations.

[0045] In step S1, the InSAR (Synthetic Aperture Radar Interferometry) data preprocessing uses small baseline set technology to extract the time series of surface deformation in the tower foundation area, and then matches it with the tower coordinates after geocoding; the optical image preprocessing uses the scale-invariant feature transformation algorithm for registration, and extracts the relative position change of the tower top structure in the image through template matching.

[0046] Step S1, data acquisition and preprocessing, is the data foundation and core preliminary step for the mechanical response analysis of tower groups. Its core function is to comprehensively collect, standardize, and accurately match basic structural data and multi-source remote sensing data of the towers. This step involves systematically collecting structural design parameters within the target area, such as geographical coordinates, tower type, nominal height, conductor type, span, and elevation difference of the towers, while simultaneously acquiring satellite-borne InSAR and optical satellite imagery remote sensing data. The remote sensing data undergoes denoising, georegistration, and temporal alignment. InSAR data uses small baseline set technology to extract the time series of line-of-sight deformation around the tower base. Optical images are registered using a scale-invariant feature transform algorithm, and template matching is used to extract the relative positional changes of the tower top structure. Finally, a standardized observation sequence that perfectly matches the spatial location and temporal sequence of the towers is formed, transforming remote sensing parameters into auxiliary features recognizable by the model, providing a unified and compliant input data source for subsequent modeling and analysis.

[0047] This step ensures data quality and analytical reliability from the outset, effectively eliminating issues such as noise, spatial misalignment, and temporal mismatch in remote sensing data, and preventing data errors from propagating to subsequent load calculations and model training. By accurately extracting key auxiliary features such as tower base settlement and tower top offset, it provides a reliable basis for the model to identify tower foundation anomalies and correct boundary conditions. Simultaneously, the constructed standardized dataset lays a solid data foundation for the construction of tower group map data models, conductor mechanical load calculations, and physical constraint graph neural network training, ensuring the accuracy and efficiency of subsequent mechanical response field reconstruction. This supports large-scale, high-efficiency, and high-reliability structural health monitoring of tower groups, providing stable data support for mechanical state assessment and risk warning.

[0048] Step S2: Construct a data model of the tower group diagram.

[0049] Step S2 uses the geographical coordinates and structural design parameters of the towers as the basis, takes the individual towers as nodes in the graph network, constructs the graph network edges of adjacent towers based on the electrical and mechanical topology of the transmission line, defines edge attributes based on tower type differences, determines edge weights based on physical parameters, and obtains the tower group graph data model.

[0050] In step S2, the tower type is distinguished as a straight tower, a tension tower, and an angle tower through numerical encoding; the edge weight is used to characterize the mechanical transfer strength between towers and is obtained by calculation of physical parameters or adaptive learning of the model. The specific process is as follows:

[0051] like Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the construction of the tower group graph data model of the present invention. Each tower is defined as a node in the graph. Electrical and mechanical topology construction of transmission lines. Edge connections are established between towers with adjacent spans, and different edge attributes are defined according to the tower type (straight tower, tension tower, angle tower). The edge weight is calculated from physical parameters such as span, height difference, and conductor type, or obtained through model learning, reflecting the true characteristics of the mechanical transmission path.

[0052] Step S2, constructing the tower group graph data model, transforms the physical tower group into a structured topological carrier recognizable by a graph neural network. Its core function is to complete the digital abstraction and modeling of the mechanical relationships within the transmission line. This step uses each tower as a node in the graph network. Based on the electrical and mechanical topological relationships of the transmission line, edge connections are established between towers with adjacent spans. Simultaneously, edge attributes are defined differently according to different tower types, such as straight-line towers, tension towers, and angle towers. Edge weights are calculated through physical parameters such as span, height difference, and conductor type, or generated autonomously by the model. This accurately recreates the true path and correlation strength of mechanical transmission between towers, achieving a standardized mapping of the tower group from physical entities to graph data. This provides a standardized topological framework for subsequent mechanical information aggregation and model calculation.

[0053] This step effectively solves the problem that traditional methods struggle to characterize the overall mechanical relationships of tower groups. By designing differentiated tower type edge attributes and physicalized edge weights, the graph network accurately reflects the actual mechanical transmission laws of transmission lines, providing fundamental support for the graph attention mechanism to distinguish the differences in mechanical transmission among different tower types and spans. The structured graph data model is highly compatible with the computational logic of graph neural networks, reducing the complexity of subsequent model training and inference, ensuring the rationality and accuracy of mechanical information transmission, laying a solid topological foundation for physical constraint fusion and load transmission calculation, and supporting rapid mechanical analysis of tower groups under a wide range of conditions. This improves the adaptability and computational efficiency of the entire method, providing key topological guarantees for accurately reconstructing the mechanical response field of tower groups.

[0054] Step S3: Calculate the conductor load within the span.

[0055] Step S3 uses the span of the structural design parameters as the calculation unit, and calculates the conductor load, tension parameters and unbalanced tension on both sides of the tower within the span based on regional meteorological data and conductor mechanical state equations, to obtain the tower mechanical transmission parameters.

[0056] In step S3, the mechanical state equation of the conductor is either the catenary equation or the parabolic equation. Combined with wind speed, icing thickness, and air temperature meteorological parameters, the horizontal tension, vertical load, and unbalanced tension of the conductor are calculated. The specific process is as follows:

[0057] Based on regional meteorological data, including wind speed, wind direction, icing thickness, and temperature, the mechanical state equation of the conductor is used, with the span as the basic calculation unit, to calculate the equivalent wind load on the conductor within each span. Equivalent icing load and the resulting horizontal tension in the conductor. and vertical load The load within the span is transmitted to the adjacent towers through conductor tension, creating unbalanced tension on both sides of the towers.

[0058] In step S3, the mechanical state equation of the conductor adopts the catenary equation or the parabolic equation in the design code of overhead transmission lines. The horizontal tension and sag of the conductor are calculated according to the meteorological conditions, and then the load acting on the adjacent towers is obtained.

[0059] The catenary equation is a precise theoretical equation for the mechanical calculation of overhead transmission line conductors. It strictly adheres to GB50545-2010 "Design Code for 110kV~750kV Overhead Transmission Lines". Its core purpose is to accurately solve the mechanical state of conductors under complex and heavy-load conditions such as long spans, large elevation differences, icing, and strong winds. Based on the theory of flexible suspension catenary, it comprehensively considers the combined effects of conductor self-weight, icing load, and wind load, realistically reproducing the spatial geometry and mechanical transmission laws of the conductor. Its standard equation is based on the horizontal tension of the conductor. As the core solution quantity, combined with the range Suspension point height difference Comprehensive load per unit length of conductor The equivalent composite values ​​for self-weight, icing, and wind load are established, and the basic form is as follows: ,in, The horizontal coordinate is the independent variable. The vertical coordinate is the dependent variable. For the horizontal tension of the conductor, The total load per unit length of the conductor. It is a hyperbolic cosine function. The integral constant is determined by the boundary conditions of the tower suspension point coordinates. This invention allows for iterative solutions to conductor sag, stress at any point, conductor axis length, and force at the suspension point. Using the span as the basic calculation unit, the catenary equation can accurately calculate the horizontal tension of the conductor. With vertical load This allows for the acquisition of unbalanced tensions on both sides of the tower, providing high-precision mechanical parameters for the initialization of graph network edge features. This ensures that load calculation and transmission fully align with the actual physical mechanism of transmission lines, making it the core method for conductor mechanical modeling under complex working conditions.

[0060] The equation for the oblique parabola is an engineering simplified approximation of the catenary equation. Also compiled according to the GB50545-2010 standard, it is suitable for rapid mechanical calculations under conditions of small to medium spans, small elevation differences, and conventional weather. By discarding higher-order small quantities of the hyperbolic function of the catenary, the spatial shape of the traverse is simplified to an oblique parabola, significantly reducing computational complexity while ensuring engineering accuracy, and adapting to the needs of rapid evaluation under multiple working conditions. Its standard equation form is as follows: ,in This is the elevation difference correction factor. For boundary constants, The horizontal tension component of the conductor along the span, with the span Elevation difference Comprehensive load per unit length Using these as input parameters, the horizontal tension of the conductor, sag, and load at the tower suspension point can be quickly calculated. In this invention, this equation serves as an alternative to the catenary equation, outputting the horizontal tension of the conductor in units of span. With vertical load This provides lightweight mechanical data for the edge features of graph networks, balancing computational efficiency and engineering practicality. Both equations are the core basis for calculating the mechanical state of conductors and can be flexibly selected according to terrain and working conditions. This ensures that the load calculation conforms to the laws of structural mechanics and matches the fast inference requirements of graph neural networks, providing stable and accurate load input support for the reconstruction of the mechanical response field of tower groups.

[0061] Step S3, the calculation of conductor load within the span, is the core physical calculation step in the analysis of the mechanical response of the tower group. It connects data preprocessing and graph network modeling. Its core function is to use the span as the basic calculation unit, relying on the catenary and parabolic equations from the "Design Code for 110kV~750kV Overhead Transmission Lines," to accurately solve for the mechanical state of the conductor and physically initialize the edge characteristics of the graph network. This step uses the tower structural parameters output from step S1, namely span, height difference, conductor type, and comprehensive load per unit length. Using the tower group diagram topology constructed in step S2 as input, the calculation model is flexibly selected for different working conditions: the catenary equation is used for complex working conditions such as large spans, large elevation differences, icing and strong winds, while the oblique parabola equation is used for medium and small spans and conventional working conditions, and the horizontal tension of the conductor is solved iteratively. The sag within the span, the vertical / horizontal load at the suspension point, and the unbalanced tension on both sides of the tower are mapped to the edge attributes of the graph network, injecting real physical constraints into the graph neural network and achieving a deep integration of mechanical laws and data models.

[0062] This step ensures the reliability and accuracy of subsequent analyses from a physical perspective. Rigorous solution of the standard equations guarantees that conductor load calculations fully conform to the mechanical mechanisms of transmission lines, with calculation accuracy meeting engineering design and operation and maintenance requirements. Simultaneously, dual-model adaptation balances high-precision solutions for complex operating conditions with efficient calculations for routine conditions, adapting to the multi-scenario analysis needs of large-scale tower groups. Using mechanical loads as edge features in the graph network effectively constrains the information transmission direction and weights of the graph attention mechanism, allowing the model learning process to follow the laws of real mechanical transmission. This addresses the issues of unreliable physics and poor generalization in purely data-driven models, providing accurate and compliant load inputs for subsequent reconstruction of the mechanical response field of tower groups, identification of abnormal operating conditions, and risk warning. This solidifies the physical foundation of the entire method and enhances its engineering applicability and the reliability of the results.

[0063] Step S4: Construct a physical information fusion graph neural network model.

[0064] Step S4, based on the tower group graph data model and the tower mechanical transfer parameters, initializes the graph network node features and edge features, and constructs a physical information fusion graph neural network model that incorporates the physical regularization constraints of tower axial force bending moment bearing capacity.

[0065] In step S4, the physical information fusion graph neural network model sequentially includes a feature embedding layer, a multi-head graph attention layer, a node update layer, and an output layer; the multi-head graph attention layer calculates attention weights based on node embedding features and edge embedding features.

[0066] In step S4, the node features include tower coordinates, tower type code, nominal height, and remote sensing observation sequence; the edge features include span, elevation difference, conductor type code, conductor horizontal tension, and vertical load. The specific process is as follows:

[0067] Design a multi-layer graph neural network model, whose core mechanisms include:

[0068] Node feature initialization: Each node The initial feature vector includes tower coordinates, tower type code, nominal height, and preprocessed remote sensing observation parameters (InSAR basic deformation and optical image structural changes). These remote sensing observation parameters are used to assist in identifying abnormal settlement and structural offset trends in tower foundations, but are not used as direct mechanical response inputs.

[0069] Edge feature initialization: for each edge The initial feature vector includes span, elevation difference, conductor type code, and conductor horizontal tension. and vertical load .

[0070] Neighborhood information aggregation layer: The graph attention mechanism is used to aggregate the features of adjacent nodes. The attention weight is calculated based on both edge features and node features, enabling the model to distinguish between straight towers and tension towers, as well as the differences in mechanical transmission under different spans.

[0071] Physical constraint node update layer: This layer fuses aggregated neighborhood information with node features and updates node states using a multilayer perceptron. The tower axial force-bending moment correlation equation is introduced as a regularization term in the loss function, ensuring the constraint model output conforms to the basic mechanical equilibrium conditions of the tower structure.

[0072] like Figure 3 As shown, Figure 3 This is a structural diagram of the physical information fusion GNN model of the present invention.

[0073] Loss function definition: Total loss function ,in To predict the mean square error between the mechanical response and the true value, For physical regularization, The balance coefficient and regularization term are used. axial force and bending moment It is obtained by linear transformation of the output of the intermediate layer of the model. , Calculations based on tower design data For the first Predicted axial force on the base tower; This refers to the axial compressive bearing capacity of the tower. For the first Predicted bending moment of base tower; This refers to the bending bearing capacity of the tower. This refers to the number of samples or the number of tower nodes involved in the loss calculation. For sample index, For the first The predicted mechanical response vector of a sample (or tower). For the first The true mechanical response vector of a sample (or tower).

[0074] In step S4, the regularization term uses the following equation form related to the tower axial force and bending moment:

[0075] ;

[0076] in To predict axial force, For the axial compression bearing capacity of the tower, To predict bending moment, This represents the bending bearing capacity of the tower. This equation serves as an additional term in the loss function, guiding the model output to conform to the mechanical state of the tower structure's bearing capacity. If the predicted result does not exceed the bearing capacity limit, the regularization term is 0, and no additional penalty is incurred; constraints are only applied when the predicted value exceeds a reasonable range to ensure the stability of model training.

[0077] Step S4, constructing a physical information fusion graph neural network model, is the core modeling step that deeply integrates the topology of the tower group, the mechanical load of the conductor, and remote sensing monitoring features. Its main function is to build a hybrid neural network architecture that combines data-driven and physical constraints. This model uses the tower group map data constructed in step S2 as its topological foundation, and the conductor horizontal tension, unbalanced load, and span mechanical parameters calculated in step S3 as prior physical information. It also incorporates remote sensing features such as surface deformation and tower top offset extracted in step S1. Through a graph attention layer, it performs weighted learning on the mechanical relationships between different tower nodes and adjacent span edges, distinguishing the differences in mechanical transmission between straight towers, tension towers, and angle towers. Furthermore, it embeds the physical laws of the catenary / parabolic equations into the network loss function, forming a physical information fusion graph neural network model that achieves integrated modeling of mechanical feature extraction, load transfer simulation, and structural response prediction.

[0078] This step, through the deep integration of physical information and graph networks, effectively addresses the problems of traditional data-driven models lacking engineering mechanisms and predictive results deviating from actual mechanical laws, significantly improving the accuracy and reliability of tower group mechanical response analysis. The introduction of physical constraints guides the model to learn within a reasonable mechanical space, avoiding abnormal predictions and non-physical convergence. Simultaneously, the graph attention mechanism adaptively captures the mechanical coupling relationships between towers, adapting to complex terrain and mixed multi-tower scenarios. The node and edge features output by the model can be directly used for subsequent mechanical response field reconstruction, ensuring both computational efficiency for large-scale tower group analysis and meeting the engineering accuracy requirements of transmission line structural health assessment. This provides crucial model support for achieving accurate, efficient, and physically reliable monitoring of the mechanical state of tower groups.

[0079] Step S5: Model training.

[0080] Step S5 uses historical operating data to train the physical information fusion graph neural network model to obtain the trained tower mechanical response prediction model.

[0081] The historical operating condition data mentioned in step S5 is calculated from historical meteorological parameters and the conductor's mechanical state equation, and generated through finite element simulation. It includes training samples, validation samples, and test samples. The specific process is as follows:

[0082] The training foundation for this step is a multi-source historical operating condition dataset, which is generated in two ways: First, based on historical meteorological parameters (ambient temperature, wind speed, icing thickness, etc.), combined with the mechanical state equation of the catenary / parabolic duct in step S3, batch calculations are performed on the mechanical characteristics of the duct horizontal tension, tower suspension point load, span sag, and unbalanced tension under different meteorological combinations. Second, through coupled finite element simulation of towers and ducts, mechanical response data under extreme operating conditions, such as icing overload, strong wind, and uneven settlement of the tower foundation, are constructed to supplement abnormal scenarios that are difficult to cover by measured data. The full dataset is divided into training samples, validation samples, and test samples in a preset ratio of 7:2:1. The training samples are used for iterative updates of model parameters, the validation samples are used for real-time monitoring of the training process and adjustment of hyperparameters, and the test samples are used for final generalization performance evaluation, ensuring that the data distribution covers multiple operating conditions, including normal, extreme, and abnormal conditions, and avoiding the risk of model overfitting from the source.

[0083] The training process employs an end-to-end training strategy, using the physical information fusion graph neural network constructed in step S4 as the carrier. The tower group topology, conductor mechanical loads, and remote sensing monitoring features are used as model inputs. The prediction targets are the mechanical responses of tower nodes, i.e., structural stress, displacement, and unbalanced tension. Simultaneously, the catenary / parabolic mechanics equations and tower structural mechanical constraints are embedded into the joint loss function, forming a dual-driven optimization objective of data fitting error and physical compliance constraints. This guides the model to strictly adhere to the actual mechanical mechanisms of transmission lines while learning data patterns. The model parameters are optimized through multiple iterations until the validation set loss converges and meets engineering accuracy requirements.

[0084] Step S5, model training, is the core step in the parameter optimization and performance finalization of the physical information fusion graph neural network. Its core function is to drive the model to learn the mechanical response laws of tower groups based on multi-source historical operating condition datasets, anchor the model to physical constraints to complete model finalization, and finally output a predictive model of tower mechanical response that can be applied in engineering. This step generates full historical operating condition data in two ways: one is based on historical meteorological parameters, such as ambient temperature, wind speed, and icing thickness, combined with the mechanical state equations of catenary / parabolic conductors, to batch calculate mechanical characteristics such as conductor horizontal tension, tower suspension point load, span sag, and unbalanced tension under multiple meteorological combinations; the other is through coupled finite element simulation of tower conductors to construct mechanical response data for extreme operating conditions such as icing overload, strong wind, and uneven settlement of the tower foundation, supplementing abnormal scenario samples that are difficult to cover by actual measurements. The full dataset is divided into training, validation, and test samples in a 7:2:1 ratio, which are used for parameter iteration and update, hyperparameter tuning, and generalization performance evaluation, respectively, to ensure coverage of various working conditions, including normal, extreme, and abnormal conditions, and to avoid the risk of overfitting from the source. At the same time, an end-to-end training strategy is adopted, embedding mechanical equations and structural constraints to form a dual-drive loss function, guiding the model to learn according to the real mechanical mechanism.

[0085] This training phase ensures the model's engineering reliability and generalization performance from both data and algorithmic dimensions, effectively addressing the pain points of pure data-driven models lacking physical mechanisms and exhibiting poor generalization under extreme conditions. The multi-source fusion dataset covers the mechanical characteristics of conventional operation and maintenance scenarios while supplementing data from extreme and abnormal conditions that are difficult to obtain through actual testing, significantly improving the model's adaptability to complex terrain, multiple tower types, and various operating conditions. Scientific sample partitioning and training strategies effectively suppress overfitting, ensuring stable model performance on unknown data. The dual-drive joint loss function anchors the model's learning direction through physical constraints, avoiding non-physical prediction results. This allows the model to perform efficient inference while strictly adhering to the mechanical laws of transmission lines. The trained tower mechanical response prediction model combines the rapid computational capabilities of data-driven approaches with the engineering accuracy of physical constraints, providing core support for subsequent reconstruction of the tower group's mechanical response field, structural health assessment, and risk warning, significantly improving the method's engineering practicality and result reliability.

[0086] Step S6: Model inference and result output.

[0087] In step S6, the newly acquired remote sensing observation data and meteorological data are input into the tower mechanical response prediction model, the mechanical response index of a single tower is output, and the mechanical response field of the tower group structure is reconstructed.

[0088] In step S6, the mechanical response indicators of the single-base tower include the horizontal displacement at the top of the tower, the vertical displacement at the top of the tower, and the maximum stress of the main material. The specific process is as follows:

[0089] This step involves inputting newly acquired multi-source remote sensing observation data and meteorological data, after preprocessing in step S1 and calculating conductor load within the span in step S3, into the tower mechanical response prediction model trained in step S5. The model performs rapid inference based on a tower group graph network topology and physical information fusion mechanism, outputting the mechanical response indicators for each tower, specifically including the horizontal displacement at the tower top, the vertical displacement at the tower top, and the maximum stress of the main material. Based on the predicted mechanical response values ​​of individual towers and combined with the mechanical transmission correlation characteristics of the tower group, the mechanical state of all towers along the entire line is integrated and mapped to reconstruct the complete structural mechanical response field of the tower group.

[0090] Step S6, model inference and result output, is a crucial execution step for putting the trained prediction model into practical application and realizing the quantitative assessment of the mechanical state of the tower group. Its core function is to complete the model adaptation, rapid inference, and global mechanical state reconstruction of real-time data. This step first transforms the newly acquired multi-source remote sensing observation data and meteorological data into input data that the model can recognize after standardization preprocessing in step S1 and conductor load calculation within the span in step S3. Then, based on the tower group graph network topology and physical information fusion mechanism, inference is carried out to accurately output the three core mechanical response indicators of each tower: tower top horizontal displacement, tower top vertical displacement, and maximum stress of the main material. Finally, combined with the mechanical transmission correlation characteristics of the tower group, the mechanical state data of the entire line towers are integrated and mapped to reconstruct the complete structural mechanical response field of the tower group, realizing the complete presentation from single tower indicators to the global mechanical state.

[0091] This step, through model-based reasoning and global reconstruction, significantly improves the efficiency and engineering practicality of tower group mechanical condition assessment, resulting in remarkable application effects. Relying on a physical information fusion mechanism, the reasoning results strictly adhere to structural mechanics constraints, ensuring accurate and reliable indicators and avoiding non-physical prediction biases. Its rapid reasoning capability is adaptable to large-scale transmission line monitoring scenarios, enabling efficient assessment of the mechanical condition of the entire line's tower group. The reconstructed mechanical response field visually presents the stress and displacement distribution of the entire tower area, clearly locating abnormal sections and weak towers. This provides precise quantitative evidence for subsequent risk warning and maintenance investigation, effectively addressing the pain points of traditional monitoring's incomplete coverage, delayed assessment, and insufficient accuracy. It promotes the upgrade of transmission line structural health monitoring towards automation, precision, and global coverage, providing solid technical support for the safe and stable operation of the power grid.

[0092] Step S7, Risk Warning.

[0093] In step S7, based on the safety operation standards for transmission lines, a threshold determination is made for the mechanical response field, and risk warning information is generated for towers that exceed the safety threshold.

[0094] In step S7, the safety threshold includes the tower tilt limit and the main material stress limit. The specific process is as follows:

[0095] Based on the "Operation Regulations for Overhead Transmission Lines" (DL / T 741-2019), structural safety thresholds for transmission lines are set. These thresholds include a tower tilt limit of 3‰ and a main material stress limit of 0.9 times the yield strength. The system extracts the horizontal and vertical displacements at the top of each tower and the maximum stress index of the main material from the structural mechanical response field of the tower group reconstructed in step S6. The actual tilt of each tower is calculated and compared with the safety thresholds. For abnormal towers with tilt exceeding 3‰ or maximum main material stress exceeding 0.9 times the yield strength, the system automatically generates risk warning information, clearly marking the abnormal tower number, geographical coordinates, type of exceeding limit, and magnitude of exceeding limit. The warning information is pushed to the operation and maintenance management terminal in real time, prompting operation and maintenance personnel to conduct key investigations and emergency response for the abnormal towers. This achieves real-time identification and rapid warning of structural safety hazards in the tower group, ensuring the safe and stable operation of the power supply lines.

[0096] Step S7, risk warning, is the closed-loop implementation link of the tower group structural health monitoring. Its core function is to complete the safety assessment of the mechanical state and issue hazard warnings according to industry standards. This step strictly follows the "Operation Regulations for Overhead Transmission Lines" (DL / T741-2019) to set safety thresholds, specifying a tower tilt limit of 3‰ and a main material stress limit of 0.9 times the yield strength. The system extracts the horizontal displacement, vertical displacement, and maximum stress index of the main material from the mechanical response field reconstructed in Step S6, calculates the actual tower tilt, and compares it with the thresholds one by one. For towers with excessive tilt or excessive main material stress, the system automatically generates warning information including the tower number, geographical coordinates, type of exceedance, and magnitude, and pushes it to the operation and maintenance terminal in real time, guiding operation and maintenance personnel to conduct key investigations and emergency responses, completing the entire closed-loop process from mechanical assessment to safety warning.

[0097] This step, through standardized judgment and automated early warning, enables rapid identification and accurate alerts of safety hazards in tower clusters, significantly improving the efficiency and reliability of transmission line operation and maintenance. Using industry regulations as the basis for judgment ensures authoritative and compliant early warning standards, avoiding the subjectivity and errors of manual judgment. An automated comparison mechanism replaces traditional manual inspections, significantly shortening the hazard discovery cycle and reducing the probability of missed or false detections. Accurate early warning information can directly guide on-site operation and maintenance, helping personnel quickly locate abnormal towers, clarify hazard levels, effectively prevent safety accidents such as tower tilting and main material failure, and ensure stable line operation. This step transforms the mechanical analysis results into implementable operation and maintenance instructions, allowing the technical solution to truly serve power grid operation and maintenance, achieving proactive safety control of large-scale tower clusters, and enhancing the engineering practical value of the entire monitoring method.

[0098] Example 2 is a simulation example based on the method for analyzing the structural mechanical response of power line tower groups and the fusion of multi-source remote sensing parameters based on graph neural networks in Example 1.

[0099] This embodiment uses 50 towers along a power supply line as an example to illustrate how to use the method of the present invention to perform structural mechanical response analysis.

[0100] Step S1: Data acquisition and preprocessing.

[0101] Obtain the latitude and longitude coordinates, tower type (35 straight-line towers and 15 tension towers), nominal height, conductor type (LGJ-300 / 40), span length (40 spans, ranging from 200m to 500m), and elevation difference of the 50 towers. Also obtain the following remote sensing data:

[0102] InSAR data (i.e., synthetic aperture radar interferometry data): Sentinel-1 satellite imagery is used, and small baseline set technology is applied to extract the foundation area of ​​each tower. The time series of line-of-sight deformation within a 20-meter radius centered on the tower coordinates is then geocoded and matched with the tower coordinates.

[0103] Optical satellite imagery: Using Gaofen-2 multi-temporal imagery, registration was performed using a scale-invariant feature transform algorithm. The top structure of the tower was used as a template for matching, and the relative positional change of each tower in the image was extracted.

[0104] Step S2: Construct a data model of the tower group diagram.

[0105] Define each tower as a node. Based on the line topology, edge connections are established between towers with adjacent spans. A total of 49 edges were constructed. Node attributes include tower type encoding (straight-line tower). Tension tower ), height, coordinates; edge attributes include spacing. Elevation difference Wire type code.

[0106] Step S3: Calculate the conductor load within the span.

[0107] Obtain regional meteorological station data for this time period, including wind speed at 10 meters height, wind direction, icing thickness, and temperature. Based on the 110kV-750kV overhead transmission line design code (GB50545-2010) and the conductor mechanical state equation, calculate for each span as a unit:

[0108] Basic wind pressure: Horizontal tension of conductors: solved according to the catenary equation, considering the combined conditions of icing and wind loads; Load acting on the tower: unbalanced tension generated by the difference in horizontal tension between adjacent spans. Forming edge feature vectors ,in For the horizontal tension of the conductor, This is a vertical load.

[0109] in, This is the basic wind pressure; air density; for Wind speed at a height of meters; For the first The unbalanced tension experienced by the base tower; For the first Base and the first Horizontal tension of conductors between base towers; For the first Base and the first Horizontal tension of conductors between base towers; For the first Base and the first Feature vectors of the edges between the base towers; For the first Base and the first Horizontal tension of conductors between base towers; For the first Base and the first Vertical load on conductors in the span between base towers.

[0110] Step S4: Construct a physical information fusion graph neural network model.

[0111] Implement a 3-layer graph neural network model using the PyTorch Geometric framework:

[0112] Feature embedding layer: node features The coordinates (3D), tower shape (1D), call height (1D), and remote sensing parameters (2D) are mapped to 64D through a linear layer; edge features Mapped to 32 dimensions through independent linear layers.

[0113] Graph Attention Layer: A multi-head graph attention mechanism is adopted, specifically 4 heads, each outputting 16 dimensions. The attention weight is calculated based on the splicing of node features and edge features, reflecting the degree of mechanical influence between adjacent towers under different tower types and spans.

[0114] Node update layer: The aggregated neighborhood information is concatenated with the node embedded features, and then processed through two MLP layers, i.e., 128-dimensional to... Via The system outputs the predicted mechanical response values ​​for each tower: horizontal displacement at the top of the tower (mm), vertical displacement at the top of the tower (mm), and maximum stress of the main material (MPa).

[0115] Physical regularization term: Add a regularization term to the loss function. axial force and bending moment It is obtained by linear transformation of the output of the intermediate layer of the model. , Calculations are based on tower design data. If the predicted result does not exceed the bearing capacity limit, the regularization term is 0, and no additional penalty is incurred; constraints are only applied when the predicted value exceeds a reasonable range to ensure the stability of model training.

[0116] Training settings: Total loss function ,in ,Pick The optimizer used was Adam, with an initial learning rate of 0.001, a batch size of 16, an early stopping epoch of 20, and a maximum training epoch of 200.

[0117] Step S5: Model training.

[0118] Training data source: Based on historical meteorological records, 50 typical meteorological days and conductor mechanical state equations from the past 5 years were used to calculate the load for each span. Then, using finite element simulation ANSYS, each tower was modeled according to its actual tower type with a mesh size of 0.1m. Unbalanced tensions were applied between adjacent spans to obtain the mechanical response field of the 50 towers, forming 50 training samples. Additionally, 30 validation samples and 20 test samples were generated through meteorological parameter interpolation. After training, the root mean square error of the model on the test set was: horizontal displacement at the tower top 2.3mm, vertical displacement at the tower top 1.8mm, and principal material stress 4.2MPa.

[0119] Step S6: Model inference and result output.

[0120] For newly acquired remote sensing and meteorological data, after preprocessing and load calculation are completed according to steps S1-S3, the data is input into the model. A single inference iteration takes approximately 0.3 seconds. The operating environment is an Intel Xeon 4210 CPU and an NVIDIA T4 GPU. The model outputs predicted mechanical responses for 50 towers, forming the mechanical response field of the tower group.

[0121] Step S7: Risk Warning.

[0122] According to the "Operation Regulations for Overhead Transmission Lines" (DL / T 741-2019), the warning thresholds are set as follows: tower tilt limit is 3‰, and main material stress limit is 0.9 times the yield strength. If the predicted displacement or stress of a certain tower exceeds the threshold, the system generates a warning message to prompt maintenance personnel to conduct a thorough investigation.

[0123] In this system deployment scenario, this method can be deployed on the server side of a digital twin monitoring platform for transmission lines, receiving satellite remote sensing data, meteorological data, and a tower structure database via an API interface. The system performs periodic assessments based on the satellite revisit cycle (Sentinel-1 is 12 days). Upon obtaining new remote sensing observation data, it automatically triggers model inference, outputs the mechanical response field of the tower group, and compares it with safety thresholds to achieve risk warnings. For abnormal operating conditions such as icing or strong winds, temporary assessments can be triggered through meteorological warnings, without being limited by the remote sensing revisit cycle.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0125] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters, characterized in that, The method includes the following steps: Step S1: Obtain the geographic coordinates, structural design parameters and multi-source remote sensing observation data of the towers in the target area; preprocess the multi-source remote sensing observation data to obtain a remote sensing observation sequence that matches the tower location. Step S2: Based on the geographical coordinates and structural design parameters of the towers, each tower is used as a node in the graph network. The graph network edges of adjacent towers are constructed based on the electrical and mechanical topology of the transmission line. The edge attributes are defined by combining the differences in tower type and the edge weights are determined by physical parameters to obtain the tower group graph data model. Step S3: Using the span of the structural design parameters as the calculation unit, based on regional meteorological data and the mechanical state equation of the conductor, calculate the conductor load, tension parameters and unbalanced tension on both sides of the tower within the span to obtain the mechanical transmission parameters of the tower. Step S4, based on the tower group graph data model and the tower mechanical transfer parameters, initialize the graph network node features and edge features, and construct a physical information fusion graph neural network model incorporating the physical regularization constraints of tower axial force bending moment bearing capacity; Step S4 also includes: The physical information fusion graph neural network model includes, in sequence, a feature embedding layer, a multi-head graph attention layer, a node update layer, and an output layer; the multi-head graph attention layer calculates attention weights based on node embedding features and edge embedding features. The node features include tower coordinates, tower type code, nominal height, and remote sensing observation sequence; the edge features include span, elevation difference, conductor type code, conductor horizontal tension, and vertical load. Design a multi-layer graph neural network model, the mechanism of which includes: Node feature initialization: The initial feature vector of each node includes tower coordinates, tower type code, nominal height, and preprocessed remote sensing observation parameters. The remote sensing observation parameters are used to help identify abnormal settlement of tower foundations and structural offset trends, and are not used as direct mechanical response input. Edge feature initialization: The initial feature vector of each edge includes span, elevation difference, conductor type code, conductor horizontal tension and vertical load; Neighborhood information aggregation layer: The graph attention mechanism is used to aggregate the features of adjacent nodes. The attention weight is calculated based on both edge features and node features, enabling the physical information fusion graph neural network model to distinguish between straight towers and tension towers, and the differences in mechanical transmission under different spans. Physical constraint node update layer: The aggregated neighborhood information is fused with the node’s own features, and the node state is updated through a multilayer perceptron. The tower axial force-bending moment correlation equation is introduced into the loss function as a regularization term. The output of the constraint physical information fusion graph neural network model conforms to the basic mechanical equilibrium conditions of the tower structure. Step S5: Train the physical information fusion graph neural network model using historical working condition data to obtain the trained tower mechanical response prediction model. Step S6: The newly acquired multi-source remote sensing observation data and meteorological data, after preprocessing in Step S1 and calculating the conductor load within the span in Step S3, are input into the tower mechanical response prediction model trained in Step S5, outputting the mechanical response index of a single tower and reconstructing the structural mechanical response field of the tower group. Step S7: Based on the safety operation standards for transmission lines, a threshold determination is made for the mechanical response field, and risk warning information is generated for towers that exceed the safety threshold.

2. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The multi-source remote sensing observation data mentioned in step S1 includes synthetic aperture radar interferometry deformation data and optical image change data; the preprocessing involves performing small baseline set processing and geocoding matching on the synthetic aperture radar interferometry deformation data, and performing scale-invariant feature transformation registration and template matching extraction on the optical image change data.

3. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, In step S2, the tower type is distinguished by numerical coding into straight towers, tension towers, and angle towers; the edge weight is used to characterize the mechanical transmission strength between towers and is obtained by physical parameter calculation or model adaptive learning.

4. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The mechanical state equation of the conductor mentioned in step S3 is either the catenary equation or the parabolic equation. Combined with wind speed, ice thickness, and air temperature meteorological parameters, the horizontal tension, vertical load, and unbalanced tension of the conductor are calculated.

5. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The total loss function of the physical information fusion graph neural network model described in step S4 is: ； in, , where is the mean square error between the predicted mechanical response and the true value; , which is the physical regularization term; in, This is the balance coefficient; This refers to the number of tower nodes; For sample index; For the first The predicted mechanical response vector of each sample; For the first The true mechanical response vector of each sample; For the first Predicted axial force on the base tower; This refers to the axial compressive bearing capacity of the tower. For the first Predicted bending moment of the base tower; This refers to the bending bearing capacity of the tower.

6. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The historical working condition data mentioned in step S5 is calculated by combining historical meteorological parameters with the mechanical state equation of the conductor to generate the load, and finite element simulation is carried out based on the load, including training samples, verification samples and test samples.

7. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The mechanical response indicators of the single-base tower mentioned in step S6 include the horizontal displacement at the top of the tower, the vertical displacement at the top of the tower, and the maximum stress of the main material.

8. The method for analyzing the structural mechanical response of power line tower groups based on graph neural networks and the fusion of multi-source remote sensing parameters according to claim 1, characterized in that, The safety threshold mentioned in step S7 includes the tower tilt limit and the main material stress limit.

Images