Power transmission line spatial form high-precision reconstruction method and system based on physical constraint neural network and unmanned aerial vehicle incomplete point cloud completion

Abstract

The invention discloses a power transmission line spatial form high-precision reconstruction method and system based on a physical constraint neural network and unmanned aerial vehicle incomplete point cloud completion, relates to the technical field of electric power engineering management and digital auditing, and improves the precision and reliability of power transmission project amount auditing. And the urgent requirements of high precision and high reliability in the field of electric power intelligent auditing are met. The method comprises the following steps: extracting quantifiable indexes from a tower detailed statement, and obtaining three-dimensional laser point cloud data; preprocessing the three-dimensional laser point cloud data to separate out a lead linear point cloud set and mark the missing area; constructing a physical constraint neural network, and performing spatial form deduction and completion processing on the missing region according to a physical mechanics rule to obtain a reconstructed conductor model; obtaining engineering quantity data based on the reconstructed conductor model and the tower point cloud of the power transmission line field; and inputting the engineering quantity data and the quantifiable indexes into the multi-modal large model for cross comparison and deviation analysis, and outputting a standardized audit result.

Description

Technical Field

[0001] This application relates to the fields of power engineering management and digital auditing technology, and in particular to a high-precision reconstruction method and system for the spatial morphology of transmission lines based on physical constraint neural networks and incomplete point cloud completion by UAVs. Background Technology

[0002] With the accelerated construction of a new power system based on new energy sources, the scale of power transmission and distribution network projects continues to expand, leading to an explosive growth in the volume of power grid project settlement audits. Under the traditional auditing model, the work relies primarily on manual, point-by-point verification, requiring the review and comparison of massive amounts of as-built documentation and the manual preparation of audit working papers. This is not only labor-intensive and inefficient but also highly susceptible to human error. Particularly in verifying the core engineering quantities of transmission lines, auditors often face practical challenges such as complex site terrain, inconvenient transportation, and numerous blind spots, making it difficult for traditional methods to accurately verify the quantities of work.

[0003] In related technologies, with the popularization of drone inspection technology, collecting point cloud data of overhead power transmission lines through drone micro-application platforms has become an industry trend, providing a data foundation for automated engineering quantity auditing. For point cloud data processing, data-driven methods or polynomial curve fitting algorithms are mainly used to extract conductors and calculate their lengths. Specifically, B-spline fitting, least squares method, or RANSAC (Random Sample Consensus) algorithm are used to cluster and segment the preprocessed conductor point cloud. Then, by calculating the spatial integral of the fitted curve, key geometric parameters such as the actual length, span, and sag of the conductor are derived, aiming to achieve automatic calculation and comparison of engineering quantity data.

[0004] In the process of developing the relevant technology, the applicant recognized that the relevant technology has at least the following technical problems: In complex field environments, due to dense vegetation, severe weather conditions, and limitations of UAV flight strategies, the collected point cloud data often suffers from large-scale missing areas or uneven distribution. When using data-driven fitting algorithms to process the incomplete point cloud data, overfitting is highly likely to occur, leading to severe distortion of the fitted conductor geometry and even violating the physical stress characteristics of the conductor. This results in significant deviations in key indicators such as conductor length, span, and sag, causing numerous invalid false alarms in the audit system. Consequently, it is impossible to generate judgment results that can be used as the basis for project settlement and deduction, making it difficult to meet the urgent needs of the power intelligent audit field for high precision and high reliability. Summary of the Invention

[0005] In view of this, this application provides a method and system for high-precision reconstruction of the spatial morphology of transmission lines based on physical constraint neural networks and incomplete point cloud completion by UAVs. The main purpose is to solve the problem that the current audit system has a large number of invalid false alarms and cannot generate judgment results that can be used as the basis for engineering settlement reduction, which makes it difficult to meet the urgent needs of the power intelligent audit field for high precision and high reliability.

[0006] According to the first aspect of this application, a high-precision reconstruction method for the spatial morphology of power transmission lines based on physical constraint neural networks and incomplete point cloud completion by unmanned aerial vehicles is provided. The method includes: Quantifiable indicators were extracted from the tower list of the audited data, and three-dimensional laser point cloud data were obtained by UAVs using lidar to collect data on the power transmission line site. The three-dimensional laser point cloud data is preprocessed to separate the conductor-shaped point cloud set, and based on the conductor-shaped point cloud set, missing regions are marked in the three-dimensional laser point cloud data; An initial neural network is constructed, and the physical force differential equation of the catenary of the conductor is embedded into the loss function of the initial neural network to obtain a physical constraint neural network. The physical constraint neural network is then used to perform spatial morphology deduction and completion processing on the missing regions marked in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain a reconstructed conductor model. Based on the reconstructed conductor model and the point cloud of the towers at the transmission line site, the engineering quantity is calculated to obtain the engineering quantity data. The engineering quantity data and the quantifiable indicators are input into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and standardized audit results are output based on the suspicious cases of inflated engineering quantities.

[0007] According to a second aspect of this application, a high-precision reconstruction system for the spatial morphology of power transmission lines based on physically constrained neural networks and incomplete point cloud completion by unmanned aerial vehicles is provided. The system includes: The multimodal data access management module is used to extract quantifiable indicators from the tower details table of the data to be audited, and to acquire three-dimensional laser point cloud data collected by UAVs through lidar at the power transmission line site; The neural network point cloud inference module is used to preprocess the three-dimensional laser point cloud data to separate the conductor-shaped point cloud set, and mark the missing regions in the three-dimensional laser point cloud data based on the conductor-shaped point cloud set. The neural network point cloud inference module is also used to construct an initial neural network, embed the catenary physical force differential equation of the conductor into the loss function of the initial neural network to obtain a physical constraint neural network, and use the physical constraint neural network to perform spatial morphology inference and completion processing on the missing regions marked in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain a reconstructed conductor model; The document parsing module is used to perform engineering quantity calculation based on the reconstructed conductor model and the tower point cloud of the transmission line site to obtain engineering quantity data; The audit workflow and working paper generation module is used to input the engineering quantity data and the quantifiable indicators into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and output standardized audit results based on the suspicious cases of inflated engineering quantities.

[0008] According to a third aspect of this application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] Using the above technical solution, this application provides a high-precision reconstruction method and system for the spatial morphology of transmission lines based on physically constrained neural networks and incomplete point cloud completion by UAVs. This application constructs a physically constrained neural network model by embedding the differential equation of the catenary's physical forces into the loss function of the neural network, thus introducing physical laws. Even when point cloud data has large-area gaps due to vegetation obstruction or environmental interference, it can still follow the actual force characteristics of the conductor to perform spatial morphology deduction and completion, thereby accurately restoring the actual length, sag, and other key geometric features of the conductor. The parameters fundamentally avoid deviations in engineering quantity calculation caused by incomplete data. Moreover, the engineering quantity calculation is based on the reconstructed accurate conductor model and tower point cloud. The quantifiable indicators extracted from the tower details table are input into a multimodal large model for cross-comparison and deviation analysis. This can effectively identify suspected cases of inflated engineering quantities and output standardized audit results. It solves the problems of high false alarm rate and inability to form effective deduction basis due to inaccurate data in traditional audit methods. It greatly improves the accuracy and reliability of power transmission engineering quantity audit and meets the urgent needs of the power intelligent audit field for high accuracy and high reliability.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This illustration shows a flowchart of a high-precision reconstruction method for the spatial morphology of power transmission lines based on physical constraint neural networks and incomplete point cloud completion by UAVs, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the architecture of a high-precision reconstruction system for the spatial morphology of power transmission lines based on a physically constrained neural network and incomplete point cloud completion by a UAV, according to an embodiment of this application. Figure 3 A schematic diagram of the device structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a method for high-precision reconstruction of the spatial morphology of power transmission lines based on physical constraint neural networks and incomplete point cloud completion by UAVs, such as... Figure 1 As shown, the method includes: S10: Extract quantifiable indicators from the tower details table of the data to be audited, and obtain three-dimensional laser point cloud data collected by UAVs using lidar at the power transmission line site.

[0015] The technical solution of this application embodiment can be applied to an auditing system, which has a high-precision reconstruction function for the spatial morphology of transmission lines. The auditing system receives engineering submission documents as auditable data and automatically extracts quantifiable engineering quantity indicators from the tower list using OCR (Optical Character Recognition) technology. These indicators include structured data such as the number of towers, tower height, nominal height, and conductor design length. These indicators will serve as benchmark values ​​for subsequent audit comparisons. Simultaneously, the auditing system retrieves high-precision 3D laser point cloud data collected on-site by a drone micro-application platform through a data platform. The 3D laser point cloud data can be in .las format and is collected by a drone using lidar at the transmission line site. It accurately records the physical information of the transmission line site, such as topography, tower locations, and conductor spatial distribution, laying a data foundation for subsequent accurate verification.

[0016] Through the above process, previously scattered unstructured documents and spatial data are unified into a single management system, avoiding the inefficiency and oversight of manual document retrieval and ensuring the comprehensiveness and traceability of audit evidence. For example, in the audit of a 220kV transmission line project, the audit system used OCR recognition technology to extract the design height of 32 towers and the total conductor length of 23.6 kilometers from the tower details table. Simultaneously, it integrated approximately 5 square kilometers of point cloud data collected by drones, providing a complete data source for subsequent comparisons.

[0017] S20: Preprocess the 3D laser point cloud data to separate the conductor-shaped point cloud set, and mark the missing areas in the 3D laser point cloud data based on the conductor-shaped point cloud set.

[0018] In this embodiment, the auditing system preprocesses the acquired 3D laser point cloud data. Specifically, it first uses a filtering algorithm to remove noise points and non-target points on the ground. Then, it uses automatic point cloud classification technology to divide the point cloud into different categories such as poles, conductors, and crossings, thereby separating the linear point cloud set belonging to conductors. Based on this, the auditing system analyzes the continuity of the conductor point cloud, automatically identifies missing areas in the point cloud caused by vegetation obstruction, severe weather, or blind spots of drone flight, and marks the coordinate intervals of these missing areas in 3D space. This achieves the transformation of the original point cloud into structured conductor segments, clarifying the target range that needs to be filled in for subsequent morphological reconstruction.

[0019] Through the above process, missing regions can be accurately identified, avoiding geometric distortion caused by blind fitting, and providing a basis for sampling configuration points in the physical constraint neural network. For example, in the point cloud of an actual mountain line, the auditing system detected that about 30% of the conductor point cloud was missing within a certain span due to tree shading, and automatically marked the x-axis range [1250, 1320] meters corresponding to the missing segment, preparing conditions for subsequent physical deduction.

[0020] S30: Construct an initial neural network, embed the physical force differential equation of the catenary of the conductor into the loss function of the initial neural network to obtain a physically constrained neural network, and use the physically constrained neural network to perform spatial morphology deduction and completion processing on the marked missing regions in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain the reconstructed conductor model.

[0021] In this embodiment, the auditing system constructs a deep fully connected neural network as the initial model. Its input is a one-dimensional spatial coordinate, such as the projected distance along the line direction, and the output is the corresponding elevation value. To overcome the overfitting problem of the pure data-driven method in the missing region, this embodiment embeds the second-order differential equation of the catenary describing the static stress characteristics of the conductor into the network's loss function, forming a physically constrained neural network. The composite loss function of the physically constrained neural network consists of three parts: a data loss term, a physical mechanism loss term, and a boundary condition loss term. The data loss term ensures that the network output matches the existing valid point cloud; the physical mechanism loss term uses automatic differentiation technology to force the network output to satisfy the catenary equation, that is, to ensure that the physical law of gravity and tension balance is followed even in the missing point cloud region; the boundary condition loss term constrains the two ends of the curve to accurately connect to the insulator hanging point of the tower.

[0022] By optimizing network parameters, the physical constraint neural network can ultimately output a continuous, smooth, and strictly conforming high-precision three-dimensional spatial curve of the conductor, thereby enabling scientific deduction and completion of missing areas. Even if there are large areas of missing point cloud data, the introduction of physical laws can ensure that the reconstructed conductor shape conforms to the actual stress characteristics, fundamentally eliminating the geometric distortion caused by traditional polynomial fitting, and providing a reliable spatial model for subsequent engineering quantity calculation.

[0023] For example, in a span that crosses a river valley, the point cloud of the conductor is almost completely missing due to water surface reflection. However, through the physical constraints of PINN (Physics-Informed Neural Network), the physical constraint neural network successfully deduced the sag curve that conforms to gravity. The reconstructed conductor shape has a very small error with the designed sag.

[0024] S40: Based on the reconstructed conductor model and the point cloud of towers at the transmission line site, the engineering quantity is calculated to obtain the engineering quantity data.

[0025] In this embodiment, after obtaining a high-precision reconstructed traverse model, the auditing system further combines the tower point cloud to automatically calculate the engineering quantities. On one hand, the auditing system uses calculus formulas to perform spatial length integration on the reconstructed three-dimensional traverse curve, accurately calculating the actual length of each span of the traverse, and summing up the total length of the entire line. On the other hand, the auditing system automatically measures tower coordinates, tower height, and nominal height from the tower point cloud, and performs a preliminary comparison with the design values ​​to achieve automated extraction of key engineering quantities from the reconstructed model, avoiding errors and inefficiencies caused by manual measurement.

[0026] Thus, through the above process, the engineering quantities calculated by the traverse model reconstructed based on physical constraints have a high degree of accuracy and consistency, truly reflecting the actual situation of on-site construction and providing reliable measured basis for subsequent audit comparison. For example, the audit system calculated the actual length of a certain tension section of the traverse to be 1526.8 meters based on the integral of the reconstructed traverse model, which is close to the design length of 1525 meters. The deviation between the nominal height automatically extracted from the tower point cloud and the design value is within 0.5 meters. These precise data provide solid support for the audit judgment.

[0027] S50: Input the engineering quantity data and quantifiable indicators into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and output standardized audit results based on the suspicious cases of inflated engineering quantities.

[0028] In this embodiment, the audit system inputs quantifiable indicators and calculated engineering quantity data into a multimodal large model. The multimodal large model performs cross-comparison and deviation analysis according to preset audit rules. For example, by comparing the measured value and the declared value of the total conductor length, when the deviation exceeds a preset threshold, it is automatically marked as a suspected case of inflated conductor length. At the same time, the multimodal large model can also identify other anomalies such as inconsistent tower parameters and excessive tower base land occupation. Subsequently, the multimodal large model automatically generates standardized audit working papers as standardized audit results output according to the logical chain of problem description, verification basis, and qualitative conclusion. The standardized audit working papers not only include a deviation data comparison table, but can also embed a visual graph of the PINN reconstruction curve and physical formulas as supporting evidence, and support online editing and optimization by auditors to achieve a fully intelligent closed loop from data comparison to suspicious point identification and report generation.

[0029] In this way, through the above process, and through the in-depth analysis and standardized output of the multimodal large model, the workload of manual verification and subjective misjudgment are significantly reduced, ensuring the objectivity and traceability of audit results. This effectively solves the problems of high false alarm rates and difficulty in forming effective basis for reduction in traditional audits. For example, in the audit of a certain line, the multimodal large model found that the total length of the submitted conductor was 5.2% longer than the measured value. After verification with the catenary physical model, the audit system automatically generated working papers, clearly pointing out that there was an inflated amount of work in that span, and attached the reconstructed curve and deviation calculation as the basis for reduction, so as to achieve accurate auditing.

[0030] Optionally, in this embodiment, quantifiable indicators are extracted from the pole and tower details table of the auditable data, and three-dimensional laser point cloud data collected by the UAV at the transmission line site using LiDAR is obtained. This includes: receiving the project audit data as the auditable data, and obtaining the pole and tower details table from the auditable data; performing OCR recognition on the pole and tower details table, scanning and extracting text from the pole and tower details table, and converting the unstructured settlement quantities in the pole and tower details table into structured quantifiable indicators, wherein the quantifiable indicators include the number of poles, tower height, nominal height, and conductor length; simultaneously, the UAV micro-application platform is called through the data platform to obtain the transmission line point cloud data collected by the UAV at the transmission line site in advance using LiDAR, and the transmission line point cloud data is used as three-dimensional laser point cloud data.

[0031] In this embodiment, the auditing system first receives the submitted project documents as auditable materials. These documents include various formats such as pole and tower detail tables, settlement statements, and bills of quantities. The pole and tower detail tables record key engineering quantity indicators such as pole and tower numbers, tower heights, nominal heights, and conductor design lengths in tabular form. The settlement statements and bills of quantities contain more detailed information on project cost composition and material usage. For the pole and tower detail tables, the auditing system uses OCR technology for scanning and text extraction, automatically converting the unstructured tabular content into structured, quantifiable indicators, such as the number of poles, tower heights, nominal heights, and conductor lengths. This structured data will serve as the benchmark for subsequent audit comparisons. In practical applications, the auditing system can also rely on a multimodal large-scale model to intelligently recognize documents such as settlement statements and bills of quantities. The multimodal large-scale model can simultaneously process information from multiple modalities, including text, tables, and images, extracting key data such as declared engineering quantities, project prices, and material specifications. This data is then cross-validated with the pole and tower detail table data extracted by OCR to ensure the completeness and consistency of the submitted data.

[0032] Furthermore, the auditing system will use the data platform to call the drone micro-application platform to access the point cloud data of the overhead power transmission line pre-collected by the drone using LiDAR at the power transmission line site as 3D LiDAR point cloud data. The 3D LiDAR point cloud data can be in .las format and is acquired by the drone carrying LiDAR during flight operations. It accurately records the physical information of the line site, such as the topography, tower spatial location, and three-dimensional distribution of conductors, providing high-precision basic data for on-site measurement and laying a solid data foundation for subsequent accurate calculation of engineering quantities and deviation analysis.

[0033] Through the above process, previously scattered unstructured documents and spatial data are unified into a single management system. This not only significantly reduces the workload and error probability of manually reviewing documents and entering data, but also enables cross-validation of multi-source data through the intelligent recognition capabilities of the multimodal large-scale model, ensuring the comprehensiveness and traceability of audit evidence. For example, in the audit of a 220kV transmission line project, the audit system used OCR to extract the design nominal height of 32 towers and the total conductor length of 23.6 kilometers from the tower list. At the same time, the multimodal large-scale model identified the corresponding project cost of XXX million yuan from the settlement statement and verified the consistency of the conductor length in the tower list and the bill of quantities. The audit system also accessed approximately 5 square kilometers of point cloud data collected by drones through the data platform, which clearly showed the actual coordinates of each tower and the spatial orientation of the conductors, providing a complete data source for subsequent physical reconstruction and comparison.

[0034] Optionally, in this embodiment, the three-dimensional laser point cloud data is preprocessed to separate a set of linear point clouds resembling conductors, and missing regions are marked in the three-dimensional laser point cloud data based on the set of linear point clouds resembling conductors. This includes: filtering and denoising the three-dimensional laser point cloud data to remove noise points, resulting in preprocessed three-dimensional laser point cloud data; clustering the point clouds based on the normal vector and curvature features of the point clouds in the preprocessed three-dimensional laser point cloud data using an improved region growing algorithm, obtaining clustering results; and using covariance matrix feature analysis to identify linear structures in the clustering results, separating linear point clouds belonging to conductors and constructing the set of linear point clouds resembling conductors as shown in the following expression.

[0035] in, Indicates the first One conductor point, This represents the total number of traverse points; a point cloud continuity localization analysis is performed on the traverse linear point cloud set, and based on the analysis results, missing regions with point cloud discontinuities are identified in the 3D laser point cloud data. The missing regions are marked in the 3D laser point cloud data with reference to the regional coordinate intervals of the missing regions.

[0036] In this embodiment, the auditing system first filters and denoises the acquired 3D laser point cloud data. By removing outliers and noise points caused by factors such as equipment noise, birds, or environmental clutter, clean and reliable pre-processed 3D laser point cloud data is obtained, laying the foundation for subsequent accurate classification and feature extraction. Based on this, the auditing system uses an improved region growing algorithm to cluster and segment the 3D laser point cloud data. The improved region growing algorithm uses the normal vector and curvature of each point in the point cloud as core feature indicators. The normal vector describes the orientation of the local surface of the point cloud, while the curvature reflects the degree of bending of the local surface. The improved region growing algorithm starts with the point with the smallest curvature as the seed point and groups neighboring points with a normal vector angle less than a set threshold and a gentle curvature change into the same region. This achieves preliminary clustering of point clouds of different ground features such as poles, conductors, and crossings, resulting in several geometrically consistent point cloud clusters as the clustering results.

[0037] Subsequently, the auditing system introduced covariance matrix feature analysis based on the clustering results, calculating the covariance matrix and its eigenvalues ​​for each point cloud cluster. Specifically, it is assumed that three eigenvalues ​​are calculated, and they are denoted as follows in descending order: , , The eigenvalues ​​reflect the degree of dispersion of the point cloud in three orthogonal directions. Among them, the largest eigenvalue... The intermediate eigenvalue corresponds to the direction where the point cloud distribution is most dispersed. The minimum eigenvalue corresponding to the secondary dispersion direction The direction in which the point cloud is most concentrated corresponds to the direction of the point cloud. For linear structures like conductors, the point cloud mainly extends along the direction of the conductor, therefore... It will be much greater than and ,and and If the values ​​are similar and relatively small, it indicates that the point cloud is highly clustered in the two directions perpendicular to the orientation. Therefore, by analyzing the proportional relationship of these three characteristic values, the auditing system can accurately determine the geometric shape of the point cloud cluster, that is, if... Significantly greater than and and By observing the proximity of the point cloud clusters, it can be determined that they belong to a linear structure, thus accurately separating the guide wire point cloud from the mixed point cloud and providing a reliable basis for constructing the subsequent guide wire point cloud set. Furthermore, using the separated point clouds belonging to the guide wires, the auditing system will construct the guide wire linear point cloud set as shown in Formula 1: Formula 1:

[0038] in, Indicates the first The three-dimensional spatial coordinates of each guide point This represents the total number of point clouds belonging to the traverse. The traverse linear point cloud set completely preserves the original sampling information of the traverse segments, providing a data foundation for subsequent morphological analysis.

[0039] Finally, the auditing system will perform point cloud continuity location analysis on the linear point cloud set of the conductor. By calculating indicators such as the spacing between adjacent point clouds and the projection density along the conductor direction, it will automatically assess the integrity of the conductor point cloud and accurately locate the "point cloud gaps" caused by tree obstruction, terrain undulations, or drone flight blind spots, i.e., the missing point cloud areas. Then, referring to the coordinate range of these missing areas in three-dimensional space, i.e. the start and end range along the direction of the line, it will mark the corresponding positions in the original point cloud data to achieve the marking of the exact areas.

[0040] Thus, by integrating the clustering capabilities of region growing with the structural recognition capabilities of covariance analysis in the above process, high-precision automatic extraction of conductor point clouds can be achieved, avoiding the inefficiency and subjective errors of manual segmentation. Simultaneously, by accurately locating and marking missing areas in the point cloud, the target range that needs to be filled in is clearly defined for subsequent spatial morphology deduction of the physical constraint neural network, effectively avoiding geometric distortion caused by blind fitting. For example, in a span of a 500kV transmission line passing through a dense mountainous area, the original point cloud showed multiple discontinuities in the conductor point cloud due to tree obstruction. The audit system successfully separated the conductor point cloud from the vegetation point cloud using an improved region growing algorithm. Covariance feature analysis further confirmed that the extracted point cloud was indeed a linear structure, and the constructed conductor point cloud set contained 1278 valid points. Subsequent continuity analysis revealed that the point cloud spacing exceeded the threshold within the x-axis coordinate range of [2340, 2395] meters, which was identified as a "point cloud break" area and marked, providing a clear reconstruction range for subsequent deduction and ensuring that the completed conductor morphology could accurately connect the hanging points on both sides.

[0041] Optionally, in this embodiment of the application, an initial neural network is constructed, and the physical force differential equation of the catenary of the conductor is embedded into the loss function of the initial neural network to obtain a physically constrained neural network. This includes: establishing a second-order ordinary differential equation as shown in the following formula as the physical force differential equation of the catenary of the conductor under gravity, based on classical mechanics.

[0042] in, Represents the catenary constant. Indicates the horizontal position of the conductor. The elevation at the location; the construction parameters are: Deep fully connected neural networks As an initial neural network, the input of the initial neural network is a one-dimensional spatial coordinate. The output is the predicted elevation. or By combining the physical force differential equation of the catenary, a composite loss function is constructed; by minimizing the composite loss function, the initial neural network is trained to obtain a physical constraint neural network that integrates the physical force characteristics of the conductor.

[0043] In this embodiment, the auditing system first establishes a static force physical model of the transmission conductor under gravity, serving as the theoretical basis for the subsequent construction of the physical constraint neural network. Specifically, according to classical mechanics, a uniformly flexible conductor between two suspension points naturally sags under its own weight, and its spatial shape is distributed as a catenary. Therefore, this embodiment of the application refers to this physical law and constructs the second-order ordinary differential equation shown in Equation 2 as the physical force differential equation of the conductor under gravity: Formula 2:

[0044] in, It represents the horizontal projection coordinates along the direction of the line, usually with the projection of the conductor suspension point as a reference; Indicates correspondence Elevation value of the traverse at the location; The catenary constant is the ratio of the horizontal tension of a conductor to its linear density. The catenary constant determines the degree of curvature of the catenary. The larger the value, the smaller the sag and the smoother the curve. The differential equation shown in Formula 2 describes the mechanical equilibrium relationship of any infinitesimal segment on the conductor under the combined action of gravity and tension, and is the essential description of the spatial shape of the conductor.

[0045] Based on this, the auditing system constructs a deep fully connected neural network as the initial neural network, denoted as . ,in This represents all trainable parameters of the network, including weights and biases. The initial input to the neural network is a one-dimensional coordinate system. The output is the predicted elevation value of the traverse at the corresponding location. Furthermore, for wire reconstruction in three-dimensional space, the network can be expanded to simultaneously output... and Elevation coordinates in two directions to accommodate situations where the route curves in the horizontal plane.

[0046] Next, the auditing system embeds the aforementioned catenary physical force differential equations into the neural network training process. Specifically, it needs to construct a composite loss function, which includes not only data fitting terms but also physical constraint terms. The network is trained by minimizing this composite loss function, and the network parameters... Through continuous optimization and updates, the resulting physical constraint neural network can not only fit existing incomplete point clouds, but more importantly, it can strictly follow the physical force law of the conductor in the missing area of ​​the point cloud to generate a continuous curve that conforms to the catenary shape.

[0047] Thus, through the above process, the embodiments of this application integrate classical mechanical equations into a data-driven neural network model using soft constraints. This fundamentally overcomes the overfitting and geometric distortion problems of pure data-driven methods under incomplete point clouds, ensuring that even when large areas of the point cloud are missing, the derived conductor morphology still conforms to the actual stress characteristics. For example, in a span crossing a canyon, the point cloud is completely missing in the middle of the conductor due to wind interference from the canyon. The physical constraint neural network constructed by the audit system uses the coordinates of the anchor points at both ends of the towers as boundaries and trains the catenary differential equation as the core constraint term of the loss function. The resulting conductor sag curve has a very low error compared to the design value, and the entire curve is smooth, continuous, and has reasonable stress.

[0048] In this embodiment of the application, optionally, a composite loss function is constructed by combining the physical force differential equation of the catenary, including: constructing a data loss term, as shown in the following formula, to constrain the error between the network output and the measured traverse point cloud data. ,

[0049] in, This indicates the initial neural network's relationship to spatial coordinates. Predicted output of traverse elevation. This indicates the first measured traverse point cloud participating in the training. The horizontal coordinates of the points This indicates the first measured traverse point cloud participating in the training. The measured elevation of each point This represents the total number of measured point clouds participating in the training; combining the physical force differential equation of the catenary, the physical mechanism loss term shown in the following formula is constructed. ,

[0050] in, Indicates the total number of configuration points. Indicates the index number of the configuration point. Represents the catenary constant. This indicates the initial neural network's relationship to spatial coordinates. Predicted output of traverse elevation. Indicates the first The horizontal coordinates of each configuration point are used to construct a boundary condition loss term to constrain the network output to meet boundary constraints such as wire suspension points. ; These are the data loss items. Physical mechanism loss item and boundary condition loss term Set the corresponding weight coefficients for the loss terms to construct the composite loss function shown in the following formula. ,

[0051] in, Represents data loss items The corresponding loss term weighting coefficients, Represents the physical mechanism loss term The corresponding loss term weighting coefficients, Represents the boundary condition loss term The corresponding weight coefficients for the loss term.

[0052] In this embodiment, the auditing system constructs a composite loss function to train the physical constraint neural network. The composite loss function consists of three parts: a data loss term, a physical mechanism loss term, and a boundary condition loss term. By balancing these three constraints, the network output is made to both match the measured point cloud data and strictly adhere to the physical force characteristics of the conductor. First, the data loss term needs to be constructed. The calculation formula is shown in Formula 3 below: Formula 3:

[0053] in, This indicates the initial neural network for the horizontal coordinate. Predicted output of traverse elevation. and These represent the points in the measured traverse point cloud that participated in the training. The horizontal coordinates and measured elevations of each point. This represents the total number of measured point clouds used in the training process. The data loss term calculates the mean square error between the network's predicted values ​​and the measured point clouds, ensuring that the network output accurately fits the existing valid point cloud data.

[0054] Secondly, by combining the physical force differential equation of the catenary, a physical mechanism loss term is constructed. The calculation formula is shown in Formula 4 below: Formula 4:

[0055] in, Indicates the total number of configuration points. For the index number of the configuration point, Indicates the first The horizontal coordinates of each configuration point Let be the catenary constant. Characterized by the ratio of horizontal tension to linear density of a conductor. For the initial neural network at the placement point The predicted elevation output at the location, as shown in Formula 4, includes the first and second derivatives of the network output with respect to the input, which are accurately calculated using automatic differentiation techniques. Physical mechanism loss term. Its core function is to force the neural network output to satisfy the catenary differential equation by randomly distributing configuration points throughout the entire range, especially in the "blind spots" where the point cloud is missing. This means that every point on the predicted curve must conform to the physical law of the balance between gravity and tension, thereby eliminating the geometric distortion caused by pure data-driven methods in the missing areas.

[0056] Next, we construct the boundary condition loss term. The two ends of the conductor curve output by the boundary condition loss term constraint network must be strictly connected to the coordinates of the insulator hanging points of the towers on both sides to ensure that the reconstruction result is accurately aligned with the actual tower positions on site.

[0057] Finally, corresponding weighting coefficients are set for the data loss term, physical mechanism loss term, and boundary condition loss term, respectively, to construct the complete composite loss function shown in Equation 5 below: Formula 5:

[0058] in, , , These represent the weight coefficients of the data loss term, the physical mechanism loss term, and the boundary condition loss term, respectively. The weight coefficients can be adaptively adjusted or manually set according to the actual data quality and the confidence level of the physical constraints, in order to balance the contribution of the three to the network training.

[0059] The initial neural network is trained by minimizing this composite loss function, and the parameters... Through continuous optimization and updates, the final converged physical constraint neural network can output a high-precision continuous curve that not only matches the existing defect cloud, but also strictly follows the physical laws of the catenary in the missing region, and accurately connects the tower hanging points at both ends.

[0060] In this way, by constructing a composite loss model, data-driven and physics-driven approaches are organically integrated. Data loss ensures the fidelity of the reconstruction results to the measured information, physical mechanism loss gives the model the ability to scientifically extrapolate in areas where data gaps exist, and boundary condition loss ensures the accuracy of the overall spatial location. The synergistic effect of these three factors enables the network to achieve high-fidelity reconstruction of the spatial morphology of transmission lines in complex field environments. For example, in point cloud data for a certain span, approximately 40% of the central area is completely missing due to vegetation shading. (The audit system settings...) , , The corresponding values, while retaining the effective point clouds at both ends, apply physical constraints by randomly placing 500 configuration points in the missing interval. After 5000 iterations of training, the reconstruction curve output by the network not only perfectly fits the measured point clouds at both ends, but also presents a smooth sag shape in the missing area in the middle that conforms to the catenary equation.

[0061] Optionally, in this embodiment of the application, engineering quantity calculation is performed based on the reconstructed conductor model and the tower point cloud at the transmission line site to obtain engineering quantity data, including: calculating the actual spatial stretching length of the conductor within the span using the calculus formula shown in the following formula based on the reconstructed conductor model. ,

[0062] in, The horizontal position of the starting point of the gear distance coordinate, Indicates the horizontal position at the end of the gear interval. coordinate, This indicates the initial neural network's relationship to spatial coordinates. Predicted output of traverse elevation. This indicates that the physical constraint neural network is in The lateral offset value of the conductor output in the direction; based on the actual spatial stretch length. Combine the tower point cloud with the tower coordinates, height, and nominal height to perform parameter measurement operations to obtain engineering quantity data including the actual length of the conductor and the actual parameters of the tower.

[0063] In this embodiment, the auditing system uses a high-precision conductor model reconstructed based on a physical constraint neural network, combined with tower point clouds, to automatically calculate the engineering quantities. First, for each conductor section, the auditing system uses calculus formulas to accurately calculate its actual spatial stretching length. The specific calculus formulas are detailed in Formula 6 below: Formula 6:

[0064] in, This represents the actual spatial stretching length of the conductor to be determined; and These represent the horizontal projected coordinates of the starting and ending points of the span, typically corresponding to the insulator hanging points on both sides of the tower. coordinate; For physically constrained neural networks in The predicted elevation of the traverse output, whose partial derivative in Formula 6 describes the slope change of the traverse in the vertical plane; Physically constrained neural networks in the lateral direction The predicted value of the conductor offset output in the direction, and its partial derivative in Equation 6, describes the change in the conductor's direction in the horizontal plane. The integral formula shown in Equation 6 is essentially a precise calculation of the arc length of the conductor's spatial curve. By substituting the continuous function output by the neural network into the integral, a smoother and more accurate length value can be obtained than the sum of discrete points, which is especially suitable for continuous curves after physical deduction to complete missing areas of the point cloud.

[0065] While calculating the conductor length, the auditing system simultaneously measures parameters using the tower point cloud. Specifically, it uses the clustered tower point cloud to extract the coordinates of the top and bottom of the tower by analyzing the elevation distribution of the point cloud, and automatically calculates the total height of the tower. Combining the local features of the point cloud near the insulator string hanging point, it identifies and measures the nominal height, that is, the vertical distance from the conductor hanging point to the ground. In addition, the auditing system also extracts the center coordinates of each tower to form complete tower spatial position parameters.

[0066] Thus, through the above process, the embodiments of this application transform the reconstructed high-precision conductor model into engineering quantity data that can be directly used for auditing, realizing automated mapping from three-dimensional spatial morphology to quantitative indicators. Moreover, the conductor length calculated based on continuous curve integrals avoids the cumulative errors caused by discrete point fitting, while the automatic measurement of tower parameters significantly improves the efficiency and objectivity of on-site data collection. For example, in the engineering quantity calculation of a 500kV transmission line, the auditing system, based on the conductor model reconstructed in the previous stage, performs integral calculations on the span crossing the valley, obtaining an actual conductor spatial length of 436.5 meters, which matches the design value of 436.2 meters. At the same time, the coordinates and heights of 32 towers are automatically extracted from the tower point cloud, among which the claimed height of tower No. 21 is 42.3 meters, with a deviation of only 0.2 meters from the design claimed height of 42.5 meters. These measured data provide a reliable basis for subsequent audit comparisons.

[0067] Optionally, in this embodiment, the quantity data and quantifiable indicators are input into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated quantities. The standardized audit results are then output based on these suspicious cases. This includes: inputting the quantity data and quantifiable indicators into the pre-trained multimodal large model; calculating the deviation between the actual total conductor length in the quantity data and the conductor length in the quantifiable indicators based on the multimodal large model; comparing the deviation with a preset threshold to mark abnormal deviations; simultaneously identifying problems in the deviation value, the number of towers, tower height, and claimed height in the quantifiable indicators based on the multimodal large model to determine suspicious cases of inflated quantities and deviation analysis results. Suspicious cases of inflated quantities include at least discrepancies in tower parameters and excessive tower base footprint; generating a description of the suspicious cases and audit conclusions based on the results of the deviation analysis; and integrating the suspicious cases of inflated quantities, deviation analysis results, descriptions of the suspicious cases, and audit conclusions to generate standardized audit working papers as standardized audit results and output them.

[0068] In this embodiment, the auditing system inputs quantifiable indicators extracted from the submitted documents and measured engineering quantity data reconstructed based on a physical constraint neural network into a pre-trained multimodal large model. The multimodal large model then performs intelligent cross-comparison and deviation analysis. This multimodal large model is capable of simultaneously processing multiple data formats, including text, tables, and images, and can deeply understand the correspondence between the declared data in the submitted documents and the actual measured data. First, the multimodal large model calculates the deviation between the measured total conductor length and the submitted conductor length, comparing this deviation with a preset threshold, such as 3%. If the deviation exceeds the threshold, it is automatically marked as an abnormal point in conductor length deviation. Simultaneously, the multimodal large model also verifies parameters such as the number of towers, tower height, and nominal height to identify any discrepancies in tower parameters or excessive tower foundation footprint, indicating potential overstatement of engineering quantities.

[0069] Building upon this foundation, the multimodal large-scale model, based on its built-in audit rules and knowledge base, generates problem descriptions for various identified suspicious points. It also automatically extracts clues related to the "three falsehoods and one violation" problem, referring to falsely reporting completion, inflating project quantities, misappropriating construction funds, and illegally subcontracting, in conjunction with the process of reconstructing conductors using physical constraint neural networks. Subsequently, the multimodal large-scale model automatically generates a draft audit working paper according to a standard logical framework of "problem description—verification basis—qualitative conclusion." This draft not only includes comparison tables of deviation data but also automatically embeds a visual graph of the PINN reconstruction curve and the physical calculation formula for the catenary, providing these immutable physical deduction results as a solid basis for adjustments. Finally, the audit system integrates the suspected inflated project quantities, deviation analysis results, problem descriptions, and audit conclusions into a standardized audit working paper as a standardized audit result and outputs it. Simultaneously, it supports online editing and optimization of the working paper by auditors, achieving efficient output of the working paper.

[0070] In this way, the tedious work of manually reviewing documents and comparing data is transformed into automated intelligent analysis driven by a large model, significantly reducing the workload of auditors and the risk of subjective errors. Furthermore, by embedding physical reconstruction curves and formulas as verification criteria, audit conclusions become more objective and persuasive. In addition, standardized output of working papers ensures the standardization and traceability of audit results, thereby accurately identifying issues of inflated project quantities in power grid engineering settlements. For example, in the audit of a 220kV transmission line project, the multimodal large model compared the submitted conductor length of 23.6 km with the measured integral of 22.4 km, finding a deviation of 5.2%, exceeding the 3% threshold, and immediately marking it as a suspected case of inflated conductor length. Simultaneously, it identified that the measured claimed height of three towers deviated from the declared value by more than 1 meter, determining that the tower parameters were inconsistent. The multimodal large model automatically generated a preliminary draft of the audit work report. The problem description section stated that "the declared length of the conductor in this span exceeds the measured value by 1.2 kilometers." The verification basis section embedded a comparison chart of the sag curves reconstructed by PINN and residual analysis of the catenary equation. The qualitative conclusion was "the inflated project quantity is established, and it is recommended to reduce the project cost by XX million yuan." After online confirmation by the auditors, this draft was directly used as the basis for the project settlement reduction, achieving an efficient closed loop in the audit process.

[0071] The method provided in this application embeds the physical force differential equation of the catenary of the conductor into the loss function of the neural network to construct a neural network model with physical constraints, thereby introducing physical laws. Even when there are large areas of missing point cloud data due to vegetation obstruction or environmental interference, it can still follow the actual force characteristics of the conductor to perform spatial morphology deduction and completion, thus accurately restoring the actual length, sag and other key geometric parameters of the conductor. This fundamentally avoids the deviation in engineering quantity calculation caused by data incompleteness. Moreover, the engineering quantity calculation is based on the reconstructed accurate conductor model and tower point cloud, and is input into a multimodal large model for cross-comparison and deviation analysis along with quantifiable indicators extracted from the tower details table. This can effectively identify suspicious points of inflated engineering quantities and output standardized audit results, solving the problem of high false alarm rate and inability to form effective reduction basis due to inaccurate data in traditional audit methods. This significantly improves the accuracy and reliability of power transmission engineering quantity audit, meeting the urgent need for high accuracy and high reliability in the field of power intelligent audit.

[0072] Furthermore, as Figure 1 In a specific implementation of the method, this application provides a high-precision reconstruction system for the spatial morphology of power transmission lines based on a physically constrained neural network and incomplete point cloud completion by unmanned aerial vehicles, such as... Figure 2 As shown, the system includes: a multimodal data access management module 201, a neural network point cloud inference module 202, a document parsing module 203, and an audit workflow and working paper generation module 204.

[0073] The multimodal data access management module 201 is used to extract quantifiable indicators from the tower details table of the data to be audited, and to acquire three-dimensional laser point cloud data collected by UAVs at the power transmission line site via lidar. The neural network point cloud inference module 202 is used to preprocess the three-dimensional laser point cloud data to separate the conductor-shaped point cloud set, and mark the missing regions in the three-dimensional laser point cloud data based on the conductor-shaped point cloud set. The neural network point cloud inference module 202 is also used to construct an initial neural network, embed the catenary physical force differential equation of the conductor into the loss function of the initial neural network to obtain a physical constraint neural network, and use the physical constraint neural network to perform spatial morphology inference and completion processing on the missing regions marked in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain a reconstructed conductor model; Document parsing module 203 is used to perform engineering quantity calculation based on the reconstructed conductor model and the tower point cloud of the transmission line site to obtain engineering quantity data; The audit workflow and working paper generation module 204 is used to input the engineering quantity data and the quantifiable indicators into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and output standardized audit results based on the suspicious cases of inflated engineering quantities.

[0074] In specific application scenarios, the multimodal data access management module 201 is used to receive the engineering submission materials as the auditable materials, and to obtain the pole and tower details table from the auditable materials; by performing OCR recognition on the pole and tower details table, scanning and extracting text from the pole and tower details table, the unstructured settlement engineering quantities in the pole and tower details table are converted into structured quantifiable indicators, wherein the quantifiable indicators include the number of poles and towers, tower height, nominal height, and conductor length; at the same time, the data platform calls the UAV micro-application platform to obtain the overhead transmission line point cloud data pre-collected by the UAV at the transmission line site by the UAV using LiDAR, and uses the overhead transmission line point cloud data as the three-dimensional LiDAR point cloud data.

[0075] In specific application scenarios, the neural network point cloud inference module 202 is used to filter and denoise the 3D laser point cloud data to remove noise points and obtain preprocessed 3D laser point cloud data. Using an improved region growing algorithm, the point clouds are clustered based on the normal vector and curvature features of the preprocessed 3D laser point cloud data to obtain clustering results. Covariance matrix feature analysis is used to identify linear structures in the clustering results to separate linear point clouds belonging to conductors and construct the conductor linear point cloud set shown in the following expression.

[0076] in, Indicates the first The three-dimensional spatial coordinates of each guide point The total number of point clouds belonging to the conductor is represented; point cloud continuity localization analysis is performed on the conductor linear point cloud set, and based on the analysis results, the missing regions with point cloud discontinuities are determined in the three-dimensional laser point cloud data, and the missing regions are marked in the three-dimensional laser point cloud data with reference to the regional coordinate interval of the missing regions.

[0077] In specific application scenarios, the neural network point cloud inference module 202 is used to establish, based on classical mechanics, the second-order ordinary differential equation shown in the following formula as the physical force differential equation of the catenary under gravity.

[0078] in, Represents the catenary constant. Indicates the horizontal position of the conductor. The elevation at the location; the construction parameters are: Deep fully connected neural networks As the initial neural network, the input of the initial neural network is a one-dimensional spatial coordinate. The output is the predicted elevation. or Combined with the physical force differential equation of the catenary, a composite loss function is constructed; by minimizing the composite loss function, the initial neural network is trained to obtain the physical constraint neural network that integrates the physical force characteristics of the conductor.

[0079] In specific application scenarios, the neural network point cloud inference module 202 is used to construct the data loss term shown in the following formula to constrain the error between the network output and the measured traverse point cloud data. ,

[0080] in, This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. This indicates the first measured traverse point cloud participating in the training. The horizontal coordinates of the points This indicates the first measured traverse point cloud participating in the training. The measured elevation of each point This represents the total number of measured point clouds participating in the training; based on the aforementioned catenary physical force differential equation, the physical mechanism loss term shown in the following formula is constructed. ,

[0081] in, Indicates the total number of configuration points. Indicates the index number of the configuration point. Represents the catenary constant. This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. Indicates the first The horizontal coordinates of each configuration point are used to construct a boundary condition loss term to constrain the network output to meet boundary constraints such as wire suspension points. ; are the data loss items respectively. The physical mechanism loss term and the boundary condition loss term Set the corresponding weight coefficients for the loss terms to construct the composite loss function shown in the following formula. ,

[0082] in, Represents the data loss item The corresponding loss term weighting coefficients, Represents the physical mechanism loss term The corresponding loss term weighting coefficients, Represents the boundary condition loss term The corresponding weight coefficients for the loss term.

[0083] In specific application scenarios, the document parsing module 203 is used to calculate the actual spatial stretch length of the conductor within the span based on the reconstructed conductor model and using the calculus formula shown below. ,

[0084] in, The horizontal position of the starting point of the gear distance coordinate, Indicates the horizontal position at the end of the gear interval. coordinate, This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. This indicates that the physical constraint neural network is in The lateral offset value of the conductor output in the direction; based on the actual spatial stretching length. Together with the tower point cloud, perform parameter measurement operations on tower coordinates, height, and nominal height to obtain the engineering quantity data including the actual length of the conductor and the actual parameters of the tower.

[0085] In specific application scenarios, the audit workflow and working paper generation module 204 is used to input the engineering quantity data and the quantifiable indicators into the pre-trained multimodal large model. Based on the multimodal large model, it calculates the deviation value between the actual total length of the conductor in the engineering quantity data and the conductor length in the quantifiable indicators, and compares the deviation value with a preset threshold to mark deviation anomalies. At the same time, based on the multimodal large model, it identifies problems in the deviation value, the number of towers, tower height, and nominal height in the quantifiable indicators to determine the suspicious points of the inflated engineering quantity and the deviation analysis results. The suspicious points of the inflated engineering quantity include at least the problem of inconsistent tower parameters and the problem of tower base land occupation exceeding the standard. Based on the suspicious points of the inflated engineering quantity and the deviation analysis results, it generates a description of the suspicious points and an audit conclusion. It integrates the suspicious points of the inflated engineering quantity, the deviation analysis results, the description of the suspicious points, and the audit conclusion to generate a standardized audit working paper as the standardized audit result and output it.

[0086] The device provided in this application constructs a physically constrained neural network model by embedding the differential equation of the catenary physical forces of the conductor into the loss function of the neural network, thereby introducing physical laws. Even when there are large areas of missing point cloud data due to vegetation obstruction or environmental interference, it can still follow the actual force characteristics of the conductor to perform spatial morphology deduction and completion, thus accurately restoring the actual length, sag and other key geometric parameters of the conductor. This fundamentally avoids the deviation in engineering quantity calculation caused by incomplete data. Moreover, the engineering quantity calculation is performed based on the reconstructed accurate conductor model and tower point cloud, and is input into a multimodal large model for cross-comparison and deviation analysis along with quantifiable indicators extracted from the tower details table. This can effectively identify suspicious points of inflated engineering quantities and output standardized audit results, solving the problem of high false alarm rate and inability to form effective reduction basis due to inaccurate data in traditional audit methods. This significantly improves the accuracy and reliability of power transmission engineering quantity audit, meeting the urgent need for high accuracy and high reliability in the field of power intelligent audit.

[0087] It should be noted that other corresponding descriptions of the functional units involved in the high-precision reconstruction system for the spatial morphology of transmission lines based on physical constraint neural networks and incomplete point cloud completion by UAVs provided in this application embodiment can be found in the following references. Figure 1 The corresponding descriptions in [the document] will not be repeated here.

[0088] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0089] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0091] In an exemplary embodiment, see Figure 3Furthermore, an electronic device is provided, comprising a bus, a processor, a memory, and a communication interface. It may also include input / output interfaces and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the high-precision reconstruction method for the spatial morphology of power transmission lines based on physically constrained neural networks and incomplete point cloud completion by UAVs as described in the above embodiments.

[0092] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for high-precision reconstruction of the spatial morphology of power transmission lines based on physical constraint neural networks and incomplete point cloud completion by unmanned aerial vehicles.

[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause an electronic device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0094] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0095] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0096] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0097] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A high-precision reconstruction method for the spatial morphology of power transmission lines based on physical constraint neural networks and incomplete point cloud completion by unmanned aerial vehicles, characterized in that, include: Quantifiable indicators were extracted from the tower list of the data to be audited, and three-dimensional laser point cloud data were obtained by UAVs using lidar at the power transmission line site. The three-dimensional laser point cloud data is preprocessed to separate the conductor-shaped point cloud set, and based on the conductor-shaped point cloud set, the missing regions are marked in the three-dimensional laser point cloud data; An initial neural network is constructed, and the physical force differential equation of the catenary of the conductor is embedded into the loss function of the initial neural network to obtain a physical constraint neural network. The physical constraint neural network is then used to perform spatial morphology deduction and completion processing on the missing regions marked in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain a reconstructed conductor model. Based on the reconstructed conductor model and the point cloud of the towers at the transmission line site, the engineering quantity is calculated to obtain the engineering quantity data. The engineering quantity data and the quantifiable indicators are input into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and standardized audit results are output based on the suspicious cases of inflated engineering quantities.

2. The method according to claim 1, characterized in that, The process of extracting quantifiable indicators from the tower details table of the audited data and acquiring 3D laser point cloud data collected by UAVs using lidar at the power transmission line site includes: The submitted engineering documents are used as the auditable documents, and the detailed list of poles and towers is obtained from the auditable documents. By performing OCR recognition on the pole and tower details table, scanning and extracting text from the table, the unstructured settlement quantities in the table are converted into structured quantifiable indicators, which include the number of poles and towers, tower height, nominal height, and conductor length. Simultaneously, the data platform calls the UAV micro-application platform to obtain the point cloud data of the overhead power transmission line that the UAV micro-application platform calls the UAV to collect in advance at the power transmission line site using lidar, and uses the point cloud data of the overhead power transmission line as the three-dimensional lidar point cloud data.

3. The method according to claim 1, characterized in that, The preprocessing of the three-dimensional laser point cloud data to separate the conductor-shaped point cloud set, and the marking of missing regions in the three-dimensional laser point cloud data based on the conductor-shaped point cloud set, includes: The three-dimensional laser point cloud data is filtered and denoised to remove noise points, resulting in preprocessed three-dimensional laser point cloud data. By using an improved region growing algorithm, the point cloud is clustered based on the normal vector and curvature features of the point cloud in the preprocessed 3D laser point cloud data, and the clustering results are obtained. Covariance matrix feature analysis is used to identify linear structures in the clustering results, so as to separate the linear point clouds belonging to the conductors and construct the conductor linear point cloud set shown in the following expression. in, Indicates the first The three-dimensional spatial coordinates of each guide point This represents the total number of point clouds belonging to the duct; A point cloud continuity localization analysis is performed on the linear point cloud set of the conductor, and based on the analysis results, the missing regions with point cloud discontinuities are identified in the three-dimensional laser point cloud data. The missing regions are marked in the three-dimensional laser point cloud data with reference to the regional coordinate range of the missing regions.

4. The method according to claim 1, characterized in that, The construction of the initial neural network involves embedding the physical force differential equation of the catenary of the conductor into the loss function of the initial neural network to obtain a physically constrained neural network, including: Based on classical mechanics, the second-order ordinary differential equation shown in the following formula is established as the physical force differential equation of the catenary under the action of gravity. in, Represents the catenary constant. Indicates the horizontal position of the conductor. Elevation of the location; Build parameters are Deep fully connected neural networks As the initial neural network, the input of the initial neural network is a one-dimensional spatial coordinate. The output is the predicted elevation. or ; Based on the aforementioned physical force differential equation of the catenary, a composite loss function is constructed; The initial neural network is trained by minimizing the composite loss function to obtain the physical constraint neural network that integrates the physical force characteristics of the conductor.

5. The method according to claim 4, characterized in that, The composite loss function is constructed by combining the physical force differential equation of the catenary, including: The following formula is used to construct a data loss term to constrain the error between the network output and the measured traverse point cloud data. , in, This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. This indicates the first measured traverse point cloud participating in the training. The horizontal coordinates of the points This indicates the first measured traverse point cloud participating in the training. The measured elevation of each point This represents the total number of measured point clouds participating in the training. Based on the aforementioned physical force differential equation of the catenary, the physical mechanism loss term shown in the following formula is constructed. , in, Indicates the total number of configuration points. Indicates the index number of the configuration point. Represents the catenary constant. This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. Indicates the first The horizontal coordinates of each configuration point; Construct a boundary condition loss term to constrain the network output to meet boundary constraints such as wire suspension points. ; The data loss items are respectively The physical mechanism loss term and the boundary condition loss term Set the corresponding weight coefficients for the loss terms to construct the composite loss function shown in the following formula. , in, Represents the data loss item The corresponding weighting coefficients for the loss term, Represents the physical mechanism loss term The corresponding weighting coefficients for the loss term, Represents the boundary condition loss term The corresponding weight coefficients for the loss term.

6. The method according to claim 1, characterized in that, The engineering quantity calculation is performed based on the reconstructed conductor model and the tower point cloud of the transmission line site to obtain engineering quantity data, including: Based on the reconstructed conductor model, the actual spatial stretch length of the conductor within the span is calculated using the calculus formula shown below. , in, The horizontal position of the starting point of the gear distance coordinate, Indicates the horizontal position at the end of the gear interval. coordinate, This indicates that the initial neural network represents spatial coordinates. Predicted output of traverse elevation. This indicates that the physical constraint neural network is in The lateral offset value of the conductor output in the direction; Based on the actual spatial stretching length Together with the tower point cloud, perform parameter measurement operations on tower coordinates, height, and nominal height to obtain the engineering quantity data including the actual length of the conductor and the actual parameters of the tower.

7. The method according to claim 1, characterized in that, The process of inputting the engineering quantity data and the quantifiable indicators into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and outputting standardized audit results based on the suspicious cases of inflated engineering quantities, includes: The engineering quantity data and the quantifiable index are input into the pre-trained multimodal large model. Based on the multimodal large model, the deviation value between the actual total length of the conductor in the engineering quantity data and the conductor length in the quantifiable index is calculated, and the deviation value is compared with a preset threshold to mark deviation anomalies. Simultaneously, based on the multimodal large model, problems are identified in the deviation value, the number of towers, tower height, and claimed height in the quantifiable indicators, in order to determine the suspicious points of the inflated engineering quantity and the deviation analysis results. Among them, the suspicious points of the inflated engineering quantity include at least the problem of inconsistent tower parameters and the problem of tower foundation land exceeding the standard. Based on the suspected issues of inflated project quantities and the results of the deviation analysis, a description of the issues and an audit conclusion are generated. The suspected cases of inflated project quantities, the results of the deviation analysis, the descriptions of the suspected cases, and the audit conclusions are integrated to generate standardized audit working papers as the standardized audit results and output them.

8. A high-precision reconstruction system for the spatial morphology of power transmission lines based on physically constrained neural networks and incomplete point cloud completion by unmanned aerial vehicles, characterized in that, include: The multimodal data access management module is used to extract quantifiable indicators from the tower details table of the data to be audited, and to acquire three-dimensional laser point cloud data collected by UAVs through lidar at the power transmission line site; The neural network point cloud inference module is used to preprocess the three-dimensional laser point cloud data to separate the conductor-shaped point cloud set, and mark the missing regions in the three-dimensional laser point cloud data based on the conductor-shaped point cloud set. The neural network point cloud inference module is also used to construct an initial neural network, embed the catenary physical force differential equation of the conductor into the loss function of the initial neural network to obtain a physical constraint neural network, and use the physical constraint neural network to perform spatial morphology inference and completion processing on the missing regions marked in the three-dimensional laser point cloud data according to the laws of physical mechanics to obtain a reconstructed conductor model; The document parsing module is used to perform engineering quantity calculation based on the reconstructed conductor model and the tower point cloud of the transmission line site to obtain engineering quantity data; The audit workflow and working paper generation module is used to input the engineering quantity data and the quantifiable indicators into a pre-trained multimodal large model for cross-comparison and deviation analysis to identify suspicious cases of inflated engineering quantities, and output standardized audit results based on the suspicious cases of inflated engineering quantities.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images