A three-dimensional identification modeling method and system for a power grid channel

Abstract

The present application relates to the technical field of point cloud data processing, and more particularly to a three-dimensional identification modeling method and system for power grid channels. The method comprises the following steps: acquiring power grid channel infrared data and power grid channel point cloud data, and extracting thermal anomaly patterns according to the power grid channel infrared data to obtain component infrared feature data; performing component domain division according to the power grid channel point cloud data to obtain component point cloud data; performing component-level three-dimensional thermal structure analysis according to the component point cloud data and the component infrared feature data to obtain thermal structure data; performing component erosion processing according to the thermal structure data to obtain component erosion data; and constructing a power grid channel wear model according to the component erosion data based on the power grid channel infrared data and the power grid channel point cloud data to obtain a power grid channel model. The present application realizes multi-source fusion identification and wear state modeling of power grid channel components, and improves the structure identification precision and the power grid channel model construction capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point cloud data processing, and particularly relates to a three-dimensional identification modeling method and system for power grid channels. BACKGROUND

[0002] With the continuous expansion of the power system transmission network, the operation environment of the transmission channel is increasingly complex, and higher requirements are put forward for the perception and structural health monitoring of its operation state. In recent years, three-dimensional point cloud data and infrared thermal imaging technology have gradually increased in the application of power grid inspection, which are respectively used for geometric modeling of component form and state perception of thermal anomalies. In conventional practice, point cloud data is usually used to restore the spatial structure of power grid components, while infrared data is used to assist in judging the existing thermal defects or energy anomalies. These two types of data are mostly processed in a fragmented way, and there is no unified modeling mechanism between structural information and state information. SUMMARY

[0003] The present application relates to the technical field of point cloud data processing, and particularly relates to a three-dimensional identification modeling method and system for power grid channels.

[0004] The present application provides a three-dimensional identification modeling method for a power grid channel, which comprises the following steps:

[0005] S1, acquiring infrared data of the power grid channel and point cloud data of the power grid channel, and extracting thermal anomaly forms according to the infrared data of the power grid channel to obtain component infrared feature data;

[0006] S2, performing component domain division according to the point cloud data of the power grid channel to obtain component point cloud data; performing component-level three-dimensional thermal structure analysis according to the component point cloud data and the component infrared feature data to obtain thermal structure data;

[0007] S3, performing component erosion processing according to the thermal structure data to obtain component erosion data;

[0008] S4, constructing a power grid channel wear model according to the component erosion data, the infrared data of the power grid channel and the point cloud data of the power grid channel to obtain a power grid channel model.

[0009] In the present application, by introducing the cooperative processing of infrared data and point cloud data, the thermal abnormality form information of the power component can be effectively extracted, and the spatial thermal response modeling of the component operation state can be realized by combining the component-level three-dimensional structure characteristics. Using thermal structure data for erosion evolution analysis can not only identify potential fatigue areas and structural weaknesses, but also dynamically simulate the wear trend, forming erosion data with risk identification capability. The constructed power grid channel wear model integrates geometric shape, thermal characteristics and state evolution information, compared with the traditional single geometric modeling method, the modeling precision and authenticity are improved, and more valuable three-dimensional decision basis is provided for channel operation and maintenance.

[0010] Optionally, the thermal abnormality form extraction comprises:

[0011] Thermal region structure enhancement is performed according to the power grid channel infrared data to obtain thermal region enhancement data;

[0012] Isothermal form boundary spectrum extraction is performed according to the thermal region enhancement data to obtain isothermal data;

[0013] Structure alignment thermal offset extraction is performed on the isothermal data to obtain component infrared feature data.

[0014] In the present application, the thermal region structure enhancement improves the response capability to weak thermal signals and local abnormal temperature rise regions, effectively suppressing background noise interference; secondly, the isothermal form boundary spectrum extraction process can depict the morphological complexity and structural periodicity of the infrared thermal region boundary, thereby realizing the preliminary classification and hierarchical expression of different types of components; finally, combined with the point cloud structure for thermal offset extraction, the spatial misalignment between the infrared hot spots and the actual structure position can be identified, which is helpful to find the structural abnormalities caused by loosening, falling off or shielding.

[0015] Optionally, the component sub-domain comprises:

[0016] Power grid component deconstruction reconstruction is performed according to the power grid channel point cloud data to obtain first component point cloud data;

[0017] Modal resonance fingerprint extraction is performed according to the power grid channel point cloud data and the component infrared feature data to obtain modal resonance fingerprint data;

[0018] The power grid channel point cloud data is divided into regions according to the modal resonance fingerprint data to obtain second component point cloud data;

[0019] Thermal structure confidence voting is performed according to the first component point cloud data and the second component point cloud data to obtain component point cloud data.

[0020] The power grid component deconstruction reconstruction based on the point cloud data in the application can effectively extract main structural units such as conductors and tower bodies, retain the structural topological relationship, and guarantee the overall coherence of the model; modal resonance fingerprints are extracted in combination with component infrared features, periodic response modes of the components under thermal-structural coupling can be mined, and region division is performed on the point cloud data accordingly, so that the recognition accuracy of components such as connectors and fittings is improved. The confidence voting fusion between the first path and the second path result enhances the judgment robustness of the boundary fuzzy area and the structural abnormal area.

[0021] Optionally, the power grid component deconstruction reconstruction comprises:

[0022] The structural saliency response graph is constructed according to the power grid channel point cloud data, and structural response graph data is obtained.

[0023] The point cloud structure graph is processed according to the structural response graph data, and point cloud structure graph data is obtained.

[0024] The structure semantics is deconstructed according to the point cloud structure graph data, and point cloud semantic data is obtained.

[0025] The sub-block structure is extracted according to the point cloud semantic data, and first component point cloud data is obtained.

[0026] The structural saliency response graph construction in the application can highlight the areas with significant geometric features such as tower poles and conductors, effectively suppress background redundant data, and improve the focusing ability of structure recognition; the point cloud structure graph is generated based on the response graph, the topological connection relationship between points is expressed through the graph structure, and the spatial dependence characteristics between components are retained; the semantic deconstruction is performed based on the structure graph, the point cloud is divided into semantic units with function or component attributes, and the semantic readability of the model and the adaptability of the downstream task are enhanced; through the sub-block structure extraction, the first layer component unit in the power grid channel is effectively separated, and good structural integrity and topological connectivity are obtained.

[0027] Optionally, the modal resonance fingerprint extraction comprises:

[0028] The modal analysis region block is divided according to the power grid channel point cloud data, and modal block data is obtained.

[0029] The modal block data is subjected to morphological frequency extraction, and morphological frequency data is obtained.

[0030] The thermal-structural coupling feature fusion is performed according to the component infrared feature data and the morphological frequency data, and thermal-structural coupling data is obtained.

[0031] The modal resonance fingerprint encoding is performed on the thermal-structural coupling data, and modal resonance fingerprint data is obtained.

[0032] In the present application, the complex structure in the power grid channel is effectively disassembled into a local structure unit through modal analysis area block division, providing a clear boundary and reasonable organization processing basis for feature extraction; morphological frequency extraction is performed on the local structure, which can capture periodic configurations such as tower segment repeated structure and conductor bending state, forming a high-resolution spatial modal response spectrum; the infrared features and morphological spectrum are combined for thermal-morphological coupling fusion analysis, which not only reflects the geometric stability of the component, but also introduces the dynamic influence of the operating state, realizing the structure-thermal joint perception of the potential degradation area; the above multi-dimensional features are uniformly expressed through modal resonance fingerprint coding.

[0033] Optionally, the thermal structure confidence voting includes:

[0034] Perform main architecture extraction according to the first component point cloud data to obtain main architecture data;

[0035] Perform thermal risk area division according to the second component point cloud data to obtain thermal risk area data;

[0036] Perform candidate label matching on the first component point cloud data and the second component point cloud data according to the main architecture data and the thermal risk area data to obtain candidate label matching data;

[0037] Perform structure-level semantic alignment according to the candidate label matching data to obtain label reconciliation data;

[0038] Perform confidence voting fusion on the label reconciliation data to obtain preliminary fusion data;

[0039] Perform boundary fuzzy area re-estimation on the preliminary fusion data to obtain component point cloud data.

[0040] The main architecture data extracted from the first component point cloud data in the present application can effectively retain the core structural framework of the tower body and the conductor in the power grid channel, ensuring that the overall topology is not distorted; the thermal risk area divided based on the second component point cloud data has high semantic sensitivity and can accurately mark thermal abnormal components and operating degradation risk points. By jointly guiding candidate label matching with the main architecture and the thermal risk area, the label consistency is improved and the mis-matching rate is reduced; the structure-level semantic alignment performed on this basis makes the multi-path recognition results consistent on the functional boundary, forming a label reconciliation result with multi-source basis. Through confidence weighted fusion, balanced output of semantic enhancement and spatial consistency is realized. The boundary fuzzy area in the preliminary fusion result is locally re-estimated, effectively correcting the label drift problem caused by occlusion, thermal interference, etc.

[0041] Optionally, the component-level three-dimensional thermal structure analysis includes:

[0042] According to the component point cloud data and the component infrared feature data, a three-dimensional thermal response projection modeling is performed to obtain a three-dimensional thermal response model;

[0043] A local heat flow vector field is constructed according to the three-dimensional thermal response model to obtain heat flow direction data;

[0044] A thermal-geometric joint disturbance analysis is performed on the heat flow direction data to obtain thermal geometric distortion data;

[0045] According to the thermal geometric distortion data, a thermal structure region is extracted to obtain thermal structure data.

[0046] In the present application, by mapping the component infrared feature data to the component point cloud, a three-dimensional thermal response model is established, so that the temperature information has spatial positioning capability, providing a real geometric reference for analysis. The local heat flow vector field is constructed in the three-dimensional model, which can simulate the heat diffusion path and its abnormal behavior, thereby revealing the potential thermal conduction abnormality or cooling imbalance problem. Through joint analysis of heat flow and geometric disturbance, the structure distortion region caused by thermal expansion, fatigue deformation or local loosening can be identified, and its deformation trend can be quantitatively expressed. The extracted thermal structure region data can reflect the thermal-structure response mechanism of the component in the real running environment, providing a high-resolution, cross-modal comprehensive basis for judging the physical health state and locating the high-risk area.

[0047] Optionally, S3 comprises:

[0048] According to the thermal structure data, a thermal-fatigue model is mapped to obtain a thermal fatigue model;

[0049] According to the thermal fatigue model, an erosion path is extracted to obtain erosion path data;

[0050] According to the erosion path data, an erosion influence domain simulation is performed to obtain component erosion data.

[0051] In the present application, by thermal-fatigue model mapping, the component thermal structure data is converted into thermal-induced fatigue response indicators, effectively revealing the micro-damage accumulation process that may be caused under long-term high temperature or thermal fluctuation conditions. By extracting the erosion path of the thermal fatigue model, combined with factors such as thermal gradient direction, deformation trend and material response, the potential erosion propagation direction and its influence path can be accurately predicted. Based on the path data, the erosion influence domain simulation can dynamically deduce the spatio-temporal evolution process of the degradation range, and output the fatigue strength distribution results with spatial continuity. The present application not only realizes multi-dimensional reasoning from "surface anomaly" to "structure evolution", but also provides a future evolution trend-oriented core judgment basis for channel safety situation prediction, significantly improving the depth and precision of component-level operation state modeling.

[0052] Optionally, S4 comprises:

[0053] Fuse the channel component wear label according to the component erosion data, the power grid channel infrared data and the power grid channel point cloud data, and obtain the wear fusion data;

[0054] Construct a three-dimensional structure-state coupling grid model according to the wear fusion data, and obtain the three-dimensional structure model.

[0055] Fit the wear evolution trajectory of the three-dimensional structure model, and obtain the power grid channel model.

[0056] In the application, the component erosion data, the power grid channel infrared data and the point cloud data are fused at the component level, the semantic unification and spatial mapping of various risk characteristics such as heat, shape and fatigue can be realized in the channel space, a three-dimensional structure-state coupling grid model is constructed based on the fused label, the component geometric information and state indicators (such as erosion intensity, thermal anomaly value and shape disturbance quantity) are bound to the grid nodes, the structure continuity and operation state are cooperatively expressed, the wear evolution trajectory of the coupling model is fitted, the degradation trend and evolution direction of the key components in the channel can be deduced based on the historical or simulation data, and a dynamic evolution three-dimensional channel state atlas is formed.

[0057] Optionally, the application also provides a three-dimensional identification modeling system for a power grid channel, which is used for executing the three-dimensional identification modeling method for the power grid channel, and the three-dimensional identification modeling system for the power grid channel comprises:

[0058] An infrared thermal anomaly feature extraction module is used for acquiring the power grid channel infrared data and the power grid channel point cloud data, and extracting the thermal anomaly shape according to the power grid channel infrared data, and obtaining the component infrared feature data.

[0059] A three-dimensional thermal structure analysis module is used for performing component domain division according to the power grid channel point cloud data, obtaining the component point cloud data, and performing component-level three-dimensional thermal structure analysis according to the component point cloud data and the component infrared feature data, and obtaining the thermal structure data.

[0060] A structure erosion evolution simulation module is used for performing component erosion processing according to the thermal structure data, and obtaining the component erosion data.

[0061] A channel-level wear model construction module is used for constructing the power grid channel wear model according to the component erosion data, the power grid channel infrared data and the power grid channel point cloud data, and obtaining the power grid channel model.

[0062] The purpose of the present application is to accurately capture the thermal non-uniform response generated by the component in operation through thermal anomaly pattern extraction in infrared data, to provide a thermal feature basis with physical meaning for subsequent analysis; combined with point cloud data, the component is divided into domains, and a structure recognition mechanism driven by geometric topology and modal semantics is constructed, which effectively improves the structural integrity and thermal state perception ability of component recognition; a thermal-fatigue coupling model is established using thermal structure data, and through erosion path deduction and fatigue influence domain simulation, the dynamic prediction of component-level wear trend is realized; by mapping the erosion data to the infrared and structure data of the whole channel, a three-dimensional structure-state coupling grid model is constructed, and combined with the wear evolution trajectory fitting, a channel-level state semantic model containing structure, thermal and fatigue information is output. BRIEF DESCRIPTION OF DRAWINGS

[0063] Other features, objects, and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the drawings:

[0064] Figure 1 A step flow chart of a three-dimensional identification modeling method for a power grid channel is shown in an embodiment;

[0065] Figure 2 A step flow chart of a thermal anomaly pattern extraction method is shown in an embodiment;

[0066] Figure 3 A step flow chart of a component domain division method is shown in an embodiment;

[0067] Figure 4 A step flow chart of a component erosion processing method is shown in an embodiment;

[0068] Figure 5 A step flow chart of a power grid channel wear model construction method is shown in an embodiment;

[0069] The implementation of the present application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0070] The technical method of the present application will be described clearly and completely below in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0071] Furthermore, the drawings are merely schematic and may not be drawn to scale. A software form can be employed to implement the functional entities or the functional entities can be implemented in one or more hardware modules or integrated circuits, or the functional entities can be implemented in different network and / or processor methods and / or microcontroller methods.

[0072] It should be understood that, although the terms "first", "second" or the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of example embodiments. The term "and / or" as used herein encompasses any and all combinations of one or more of the associated associated items.

[0073] Referring to Figures 1 to 5 The application provides a three-dimensional identification modeling method for a power grid channel, which comprises the following steps:

[0074] S1, acquiring infrared data of the power grid channel and point cloud data of the power grid channel, and extracting a thermal anomaly form according to the infrared data of the power grid channel to obtain component infrared feature data;

[0075] In particular, the acquired power channel infrared image is pre-processed, including radiation correction and temperature value normalization processing, to construct a standardized thermal image matrix. On this matrix, a multi-scale edge enhancement filtering method (such as a high-pass filter based on a Laplacian-Gaussian operator or a Gabor filter) is applied to highlight the thermal gradient change area and enhance the thermal anomaly profile. The system performs a thermal anomaly region extraction operation. By calculating the temperature gradient of the temperature field of the infrared image along the horizontal direction and the vertical direction respectively, and taking the square sum and square root, the temperature gradient amplitude value image at each pixel position is obtained. This process can be expressed as a gradient amplitude map, where the value of each pixel is the square root of the square sum of its temperature gradient in two directions. A threshold value is set, and when the gradient amplitude value of a certain pixel point is greater than the threshold value, it is marked as a thermal anomaly candidate region. The hot spot region profile is extracted at the local extreme point. On the basis of the thermal spot profile region, boundary spectral analysis of isothermal morphology is performed. The boundary curve profile of each thermal anomaly region is extracted, and the Fourier descriptor algorithm is used to calculate its morphological frequency spectrum, which can characterize the complexity of the thermal spot morphology, such as distinguishing between strip-shaped concentrated and diffused dispersed thermal anomalies. After the thermal image processing is completed, the two-dimensional thermal image is mapped into the corresponding three-dimensional point cloud coordinate system through the known pose transformation parameters, so as to realize the spatial alignment of the thermal spot and the structure model. The three-dimensional coordinates of the center point (centroid) of each thermal spot region in the infrared image and the point cloud model are calculated, and the distance between the thermal spot centroid and the corresponding point cloud component main axis is defined as the offset. The offset is used to represent the degree of spatial offset of the thermal anomaly region relative to the structure center axis. The component infrared feature data set containing the following information is output: morphological spectral features of the thermal anomaly region, temperature gradient response information, and infrared-structure spatial offset indicators.

[0076] S2, according to the power channel point cloud data, the component domain is obtained, and the component point cloud data is obtained; according to the component point cloud data and the component infrared feature data, the component level three-dimensional thermal structure analysis is carried out, and the thermal structure data is obtained;

[0077] In particular, the three-dimensional point cloud data of the power channel is analyzed for structural saliency. The unit normal vector, local curvature and point cloud density of each point in the point cloud are calculated, and a response function is constructed wherein is the structural saliency score of the point , is the local curvature value weight value, is the local curvature value of the point , is the local point cloud density weight value, is the local point cloud density of the point , is the normal vector gradient module weight value, is the normal vector gradient module of the point a normal vector gradient module of the point , representing a normal change rate thereof, calculated as a difference between a unit normal vector of the point and a neighborhood average normal vector. A structure graph model is constructed based on the point cloud data. Each point in the point cloud is regarded as a node in the graph, and edges in the graph are established based on a nearest neighbor relationship in the Euclidean space (e.g., based on a k-nearest neighbor search). Edge weights are defined based on geometric similarity between the nodes (e.g., a normal vector angle difference, an Euclidean distance, etc.). A graph community division algorithm (e.g., a Louvain modularity algorithm or a spectral clustering algorithm) is used to divide the graph into structural units, to identify component sub-blocks with obvious structural boundaries, and to obtain a preliminary first component point cloud division result. Based on the above division result, combined with the component infrared features extracted from the infrared image, a modal resonance fingerprint is generated. Specifically, a Fourier morphological spectrum of each structural unit region, a spatial position and temperature intensity feature of a thermal anomaly region, and a curvature disturbance index of a local point cloud structure are extracted. The above multi-dimensional features are fused and encoded into a modal resonance fingerprint vector, forming a point cloud region division result under a second segmentation perspective. The two types of division results are fused based on a confidence voting mechanism. Each point calculates a fusion confidence score based on its structural connectivity strength, corresponding thermal anomaly response strength, and neighborhood label consistency, and determines the component category to which the point belongs according to the fusion confidence score. Finally, a component point cloud data with clear boundaries and unified labels is output. wherein is a thermal-geometric disturbance coupling score of the point , is a local temperature gradient vector module weight term, is a local temperature gradient vector module of the point , is a normal vector gradient module weight term, is a normal vector gradient module of the point , is a geometric curvature value weight term, is a geometric curvature value of the point , is the i-th point cloud data. The output three-dimensional thermal structure data includes a thermal response path, a local thermal disturbance intensity distribution, and a high-risk structure label in each component region.

[0078] S3, performing component erosion processing according to the thermal structure data to obtain component erosion data;

[0079] Specifically, a thermal-fatigue mapping model is first established according to the component thermal structure data, which is used to estimate the fatigue response strength of the component under different thermal disturbances. The model can adopt a semi-empirical thermal fatigue formula or a regression prediction model trained based on historical failure data, and the input variables include the steady-state temperature value of each structure point in the point cloud, the temperature change amplitude (i.e. the maximum temperature difference), the local thermal-geometric joint disturbance index, and the external exposure time and other operating parameters. Combining the above factors, a point-level thermal fatigue score (e.g. calculated by a regression model pre-trained based on the aforementioned data through historical samples) is generated, which constitutes a fatigue level field. After completing the construction of the fatigue level field, the extraction process of the potential erosion path is performed. Taking the high fatigue score area as the starting seed point, the local fatigue gradient, i.e. the spatial variation rate of the thermal fatigue score of each point, is calculated to guide the erosion expansion direction. The fatigue gradient field is constructed and converted into a path guiding vector field to indicate the main propagation direction of the erosion trend. Based on the vector field, the graph shortest path algorithm or the structure expansion algorithm (such as the region inflation method based on Voronoi diagram) is used to extract the possible erosion conduction path, which preliminarily outlines the degradation channel caused by thermal fatigue in the component. On the basis of the above path, the spatial diffusion simulation of the erosion evolution process is performed. Combining the path information and the local structure morphology, multi-step erosion propagation simulation is performed on the path neighborhood, i.e. performing multi-round local convolution operation on the path around with the fatigue gradient as the kernel function, to simulate the dynamic evolution of the outward propagation of thermal fatigue in the structure. The simulation results will generate a series of fatigue accumulation regions, and by setting the fatigue threshold, extracting the isosurface or isovalue region, the predicted erosion boundary is formed. The output component erosion data has the following characteristics: its structure boundary is clear, has spatial continuity, and in each erosion unit, it labels information such as fatigue level, propagation path and potential occurrence time window.

[0080] S4, according to the component erosion data, the power grid channel infrared data and the power grid channel point cloud data are used to construct a power grid channel wear model to obtain a power grid channel model.

[0081] Specifically, based on the component erosion data, the mapping of the component-level wear information to the global coordinate system of the whole passage is completed. For the erosion score, thermal disturbance level and deformation offset value of each component, they are projected to the global three-dimensional coordinate system of the power grid passage, and a power grid passage wear label layer covering the entire spatial area is constructed. In this layer, each spatial position point is bound with multi-dimensional attribute information including temperature value, fatigue score, structure disturbance index and component type, thereby forming a global wear data set with multi-source semantics. The point cloud data of the power grid passage is structured into a three-dimensional grid model. The discrete point data is converted into a spatial unit structure with topological connectivity by using the voxel grid division method or the triangular grid reconstruction algorithm. For each voxel unit or grid unit, its thermal state parameters (such as maximum temperature value, thermal gradient), fatigue parameters (such as erosion level) and structure disturbance parameters (such as geometric offset, curvature fluctuation rate) are bound, and a structure-state coupled grid model with thermal-fatigue-geometry attribute fusion is constructed. The system simulates the dynamic evolution trend of wear, uses multi-temporal infrared image data or erosion simulation sequence to model the fatigue score of the key area, and extracts its evolution trajectory over time. The regression process can use a logarithmic growth fitting function wherein is the fatigue score at time , is the initial fatigue score, is the growth amplitude adjustment parameter, is the time scale adjustment coefficient, is the time step, and the fatigue score presents a trend of fast then slow change over time. The fitted fatigue score evolution curve is embedded into the grid model and labeled in the corresponding spatial area, forming a dynamic risk layer with time evolution trend information. The output power grid passage model has the following capabilities: not only accurately reflects the wear distribution of each component in the spatial dimension, but also dynamically presents the propagation trend of potential wear in the time dimension.

[0082] Optionally, the thermal anomaly pattern extraction includes:

[0083] S11, performing thermal region structure enhancement according to the infrared data of the power grid passage to obtain thermal region enhancement data;

[0084] Specifically, the input is infrared image data of the power grid passage, and the image unit can be temperature value or infrared gray value. The system uses bilateral filtering algorithm to smooth the image. The temperature gradient image of the image is calculated, and the temperature gradient value wherein and respectively represent the partial derivatives of the temperature value in the horizontal and vertical directions, represents the temperature gradient amplitude of the corresponding pixel position. The gradient value is fused with the original image as a saliency weight to obtain an enhanced image. The enhancement method is to amplify the original temperature value by weighting, and the specific calculation formula is wherein is the pixel value of the enhanced image, is the pixel value of the original image, is the gradient enhancement coefficient, and the value range is 0.5 to 2.0, is the temperature gradient value. Based on the enhanced heat map, a multi-scale region growing algorithm (such as the maximum stable extreme region MSER algorithm) is used to detect local temperature mutation regions. The system performs clustering analysis on local thermal clusters at multiple scales to generate a mask map of thermal anomaly candidate regions. The mask map is used to mark and cover all regions identified as having thermal anomaly characteristics.

[0085] S12, extracting isothermal morphological boundary spectrum according to the thermal region enhancement data to obtain isothermal data;

[0086] Specifically, based on the temperature enhancement image, a plurality of temperature thresholds are set to extract the corresponding isothermal boundary line. The isothermal line refers to the boundary line of the pixel set with the same temperature value in the image, which represents the isotherm contour of the temperature distribution. The setting method of the temperature threshold can adopt a fixed interval method (such as setting the interval between adjacent thresholds to 1.5°C), or can adaptively select the temperature peak point as the isothermal line extraction reference through gray level histogram analysis of the heat map. Each isothermal line is represented as a sequence of ordered boundary points, and the boundary point coordinates are represented as , is the point number, is the total number of points of the isothermal line. Fourier descriptor coding processing is performed on the sequence to construct the frequency spectrum representation of the boundary, and the specific calculation formula is: wherein is the Fourier descriptor of the order , which is used for the frequency spectrum representation of the isothermal boundary, is the sequence number index of the boundary point, is the total number of points of the current isothermal line boundary point, is the two-dimensional image horizontal coordinate of the th isothermal line boundary point, is the imaginary unit, is the two-dimensional image vertical coordinate of the th isothermal line boundary point, is the base of the natural logarithm, is the constant of pi, is the Fourier frequency component index. Based on the Fourier spectrum feature vector, a clustering algorithm (such as density-based DBSCAN or spectral clustering method) is used to classify the morphologies of multiple isothermal boundaries. Through clustering of spectral features, the system identifies and labels typical boundary morphology categories based on pre-set parameters, including but not limited to long strip, ring, spot, and other thermal anomaly morphologies.

[0087] S13, structure alignment and thermal offset extraction of isothermal data to obtain component infrared feature data.

[0088] Specifically, a cross-modal coordinate system conversion is performed using the calibration extrinsic matrix between the infrared thermal imaging equipment and the three-dimensional sensor (such as a laser radar or a visible light camera). The extrinsic matrix is a pose transformation matrix containing rotation and translation parameters. The system projects the isothermal boundary points in the two-dimensional thermal map into the three-dimensional point cloud coordinate system one by one to obtain the mapping point set of the isothermal boundary in the three-dimensional space. In the point cloud space, the system performs a nearest neighbor search for each projection point and each component point cloud to establish a correspondence between the infrared points and the component points. For each pair of matched points, the three-dimensional spatial position difference is calculated, i.e., the offset vector is constructed, representing the geometric deviation of the thermal anomaly boundary point relative to the actual component surface. The system calculates the mean and variance of the offset vector set. The offset intensity features are calculated for each component type (such as cables, insulators, tower bodies, etc.). If the mean of the offset vector corresponding to a certain type of component exceeds a pre-set threshold, the system identifies the region as a potential thermal structure anomaly region. In addition, the system retains the complete offset direction vector field data to form a three-dimensional thermal offset distribution map.

[0089] Optionally, the component sub-domain includes:

[0090] S21, deconstructing and reconstructing the power grid component according to the power grid channel point cloud data to obtain first component point cloud data;

[0091] Specifically, a structure response graph construction operation is performed. For each point in the power grid channel point cloud, the unit normal vector, local geometric curvature, and point neighborhood density are calculated. The three indicators are combined to construct a response score function for measuring the local structure saliency of each point. The response score can be obtained by weighted summation of multiple indicators, for example, a weighted linear combination of local curvature value, normal vector gradient amplitude, and point density value can be used, with weights of 0.4, 0.3, and 0.3. All response scores are normalized to unify the numerical scale. A structure graph is constructed based on the spatial adjacency relationship of the point cloud. Each point is taken as a node of the graph, and a connection edge is constructed according to the fixed radius neighborhood or K-nearest neighbor rule. The weight of each edge is determined by the similarity function between the Gaussian function of the distance between points and the response score wherein is the weight value of the edge , indicating the strength of the connection between points, is a natural exponential term, is the point cloud, is the point cloud, is a distance Gaussian decay factor, controlling the influence of the distance between points on the edge weight. This graph structure not only embodies the geometric adjacency relationship, but also integrates the semantic consistency of the structural response. The graph partitioning operation is performed on the constructed structure graph. Graph clustering algorithms such as normalized cut or spectral embedding can be used to divide the structure graph into several subgraphs. During the division process, the system preferentially retains the connected subgraph area with relatively high structural response scores and stable edge weights, which is regarded as the component area.

[0092] S22, according to the power grid channel point cloud data and the component infrared feature data, modal resonance fingerprint extraction is carried out, and modal resonance fingerprint data is obtained;

[0093] Specifically, the system extracts modal resonance fingerprint data reflecting the local dynamic characteristics of the component based on the power grid channel point cloud data and the extracted component infrared feature data, combining the component morphology and thermal response characteristics. Local block division of the modal area is performed. The system divides all point clouds according to the principle of spatial uniformity, generates local point cloud blocks using a fixed radius (e.g. 0.5 meters) spherical neighborhood, and each point cloud block contains a center point and its peripheral neighboring point set. This division method ensures the regional stability and geometric consistency of subsequent spectral analysis. Morphological frequency spectrum extraction is performed on each point cloud block. The system establishes a local coordinate system in each block by principal axis fitting, extracts the surface contour features of the point cloud in the local range. Then, the local geometric contour is transformed using a surface shape spectrum analysis method (such as based on Zernike polynomial or spherical harmonic function), and low-order morphological frequency coefficients are extracted as the frequency representation of structural deformation. The geometric spectral features are fused with the component infrared feature data. Specifically, the system extracts the thermal field attributes corresponding to each point cloud block from the infrared image, including the temperature extreme points in the region, the temperature gradient distribution map, and the isothermal shape boundary features. By taking the thermal field gradient value as a weighted correction factor of the spectral coefficient, the thermal coupling modulation of the geometric frequency spectrum is realized, so as to reflect the local thermal-geometric coupling behavior in the frequency domain. The system constructs a modal feature vector based on the above-mentioned geometric frequency characteristics, thermal response indicators and gradient distribution parameters. Each vector contains structural frequency components, thermal gradient response values, isothermal region morphology indicators and other attributes, which are stored as modal resonance fingerprints to form a component-level modal feature data set.

[0094] S23, according to the modal resonance fingerprint data, the power grid channel point cloud data is divided into regions, and the second component point cloud data is obtained;

[0095] Specifically, the system assigns each point in the power grid channel with a modal fingerprint vector corresponding to the modal block it belongs to. The modal fingerprint vector is a previously extracted joint feature representation containing local modal frequency characteristics, thermal response indicators, and temperature gradient information. The system measures the similarity of the modal fingerprints between different points. The cosine similarity calculation method is used to evaluate the angle similarity between any two modal fingerprint vectors. The similarity index value ranges from -1 to 1, and the closer the value is to 1, the more consistent the modal response of the two points. The system constructs a modal similarity graph based on the similarity between the modal fingerprints, and performs regional clustering operations based on the graph. Preferably, a density clustering algorithm (such as DBSCAN) or a feature kernel density-based mean shift clustering algorithm (MeanShift) can be used to automatically determine the number of clusters and boundaries in the modal feature space, thereby realizing the aggregation and identification of structural regions with consistent modalities. The system annotates the original point cloud data based on the clustering results, and each cluster unit is regarded as an independent component region, thereby generating the second component point cloud data.

[0096] S24, performing thermal structure confidence voting according to the first component point cloud data and the second component point cloud data to obtain component point cloud data.

[0097] Specifically, the system pairs the same spatial position points in the two point cloud data sets, and assigns each point with two initial labels representing its structural division result and modal response recognition result. Each point has both a structure label and a modal label. The system constructs a confidence voting function to quantify the label consistency and the degree of influence of local thermal disturbance. The voting function combines the following factors: first, label consistency, i.e., whether the structure label and the modal label are the same, if consistent, a higher score is given; second, local thermal disturbance level, which can be quantified according to the thermal gradient amplitude and thermal shape distortion degree in the neighborhood of the point; third, boundary smoothness, i.e., the label continuity of the point in its neighborhood, if the boundary mutation is significant, the confidence score is reduced. The confidence function is of the form: wherein is the confidence score of point , is the label consistency score weight coefficient, is the label consistency score, which is 1 if consistent, otherwise 0, is the structure division label of point , is the modal response label of point , is the thermal disturbance score weight coefficient, is the thermal disturbance level of the point, is the boundary smoothness score weight coefficient, The boundary smoothness score is obtained. The system normalizes the confidence function result and sets a confidence threshold. When the comprehensive confidence of a certain point is lower than the threshold, or the structure is inconsistent with the modal label, the point is regarded as a label uncertain area. For the uncertain area, the system adopts a soft label assignment strategy, such as a soft-KNN (soft nearest neighbor label weighted estimation) algorithm, to reassign the attribution label of the point to realize the label optimization and correction of the boundary fuzzy area. The system outputs a high-consistency component point cloud data set that fuses the structural morphology information and the modal thermal response characteristics.

[0098] Optionally, the power grid component deconstruction reconstruction comprises:

[0099] According to the power grid channel point cloud data, a structural saliency response map is constructed to obtain structural response map data;

[0100] Specifically, the input is the original three-dimensional point cloud data collected in the power grid channel area, denoted as point set P, containing N spatial points. For each point, the system calculates the following geometric feature indicators based on its fixed radius neighborhood (for example, the neighborhood radius is set to 0.2 meters). Through principal component analysis of the three-dimensional coordinates of the neighborhood points, three eigenvalues of the covariance matrix are extracted. Define the curvature , wherein is the curvature index, reflecting the local flatness or convexity of the region where the point is located, is the maximum eigenvalue of the covariance matrix of the point neighborhood, is the second largest eigenvalue of the covariance matrix of the point neighborhood, is the minimum eigenvalue of the covariance matrix of the point neighborhood. Calculate the distance between the unit normal vector of the current point and the average of the normal vectors of its neighborhood points , wherein is the normal change rate, that is, the Euclidean distance between the normal of the point and the average normal of the neighborhood, is the unit normal vector of the point , is the average normal vector of the neighborhood points of the point . The number of points in the neighborhood of the point is calculated as , then the density inverse factor . The system calculates the structural saliency response score of each point by weighted linear combination of the above three indicators . The weighted formula is as follows , wherein, , , are the weighted coefficients of each geometric indicator, for example, 0.5, 0.3, 0.2, or can be adjusted according to the importance of the actual structural characteristics. Structural response map data is obtained.

[0101] Point cloud structure map data is obtained by processing the structure response map data;

[0102] Specifically, structural response information is incorporated into the point cloud structure to form a weighted graph that can be used for structural clustering and semantic derivation. A kN adjacency graph G=(V,E) is constructed, where each point... An edge connects to its k nearest neighbors (e.g., k=16). Each edge... Empowerment: ,in For the edge The edge weights represent and The structural correlation between them It is an exponential function. For the first A point cloud, For the first A point cloud, Distance attenuation factor ; For response similarity.

[0103] Structural semantic deconstruction is performed based on the point cloud structure diagram data to obtain point cloud semantic data;

[0104] Specifically, the system utilizes the topological information and edge weight features of the point cloud structure graph to construct a graph embedding space for semantic partitioning. Graph neural network methods (such as graph convolutional networks (GCN) or graph attention networks (GAT)) or graph learning methods (such as Laplacian feature mapping) can be used to encode each point into a low-dimensional embedding vector, representing the point's structural semantic feature representation. The goal of generating this embedding vector is to maintain the similarity relationship between the original edge weights in the graph structure; that is, the system optimization objective is to minimize the distance between the embedding vectors of points connected by high weights in the graph, thus preserving the structural consistency of adjacent edges. All embedding vectors are input into the clustering module to perform structural semantic partitioning. Methods such as k-means clustering or spectral clustering can be used to cluster the point cloud, and the number of categories can be set according to the actual component complexity (e.g., 6 to 10 categories). The clustering result assigns a semantic label to each point, representing its structural component type, such as beam, vertical support, insulator attachment point, etc. For clusters initially identified as "lever" or "beam" types, rule filtering can be performed by combining the point cloud height value (Z-axis coordinate) and normal vector direction (e.g., the normal direction is close to horizontal or vertical); point-level reassignment is performed on anomalous cluster boundary regions. The system outputs a structured point cloud dataset with semantic category labels for each point.

[0105] The sub-block structure is extracted based on the point cloud semantic data to obtain the point cloud data of the first component.

[0106] Specifically, for each semantic structure class (e.g. labeled as class k), the system extracts all points with this label from the semantic point cloud data to form a subset k. Then, the system performs a distance-based density clustering algorithm (e.g. DBSCAN) on this subset to identify connected regions with spatial coherence. During the clustering process, the system sets a minimum point number threshold (e.g. 100 points) to remove small noise clusters below this threshold, and only keeps connected regions with sufficient density and complete structure as valid structure sub-component candidates. For each preliminary sub-block, the system calculates its geometric statistical feature indicators, including voxel density: representing the point cloud density within unit volume, for assessing structure reality; bounding box size: the length, width, and height of the three-dimensional minimum axis-aligned bounding box; geometric proportion (aspect ratio): for identifying abnormal slender or flattened redundant structures. The system sets filtering conditions based on the above indicators to remove point cloud clusters that do not meet the geometric characteristics of components, such as isolated noise fragments; linearly distributed and highly sparse non-structure point sets; redundant or merged regions that are significantly beyond the component size range. For all filtered sub-blocks, the system assigns them a unique component number and records their structure class label (derived from the semantic division results in the previous step), forming the first component point cloud data set.

[0107] Optionally, the modal resonance fingerprint extraction includes:

[0108] Modal analysis region block division is performed according to the power grid channel point cloud data to obtain modal block data;

[0109] Specifically, the input is a three-dimensional point cloud data set of a power grid channel Each point is associated with its normal vector and local geometric curvature value . The adjacency graph structure of the point cloud is constructed, and a k-NN graph or a radius-based neighborhood connection method can be optionally used. On the basis of this graph structure, a structure response indicator is introduced as a clustering criterion for local region division of the point cloud. The system defines the structure difference between each pair of adjacent points as the sum of the difference of their normal vectors and the difference of their curvatures: wherein is the structure difference between point and point , defined as the weighted sum of the difference of normal vectors and the difference of curvatures, is the unit normal vector of point , is the unit normal vector of point , is the local geometric curvature value of point , is the local geometric curvature value of point local geometric curvature value. The system takes the difference as the edge weight, performs structural similarity clustering on the adjacency graph, aiming to minimize the total difference within each region, so as to realize the classification of regions with similar geometric features. The number of points within each region block should not be less than 200 points to ensure the statistical stability of subsequent modal modeling; principal component analysis (PCA) is performed on each region block to extract its principal axis variance ratio (i.e. the variance ratio of the first principal direction to the remaining directions); if the ratio is lower than a set threshold (e.g. 0.2), it means that the region may be a pure plane or a boundary region and should be removed. The output is a set of modal analysis region blocks, each of which has significant geometric continuity and potential resonance characteristics.

[0110] modal frequency data is obtained by performing morphological frequency extraction on the modal block data;

[0111] Specifically, for each modal region block, voxel reconstruction is performed, that is, irregular point cloud is converted into a three-dimensional grid structure with equal spacing. The voxel edge length is set to 0.02 meters. A simplified finite element structure is constructed to generate a node connection graph for simulating the geometric response behavior of the region block under static conditions. On the basis of the three-dimensional grid, the system constructs a shape Laplacian matrix (i.e. morphological Laplacian matrix) for each region block to describe the response characteristics of the local structure to external disturbances. The Laplace eigenvalue problem is solved, that is, the characteristic equation of the form: Laplace matrix multiplied by eigenvector equals eigenvalue multiplied by eigenvector. Wherein, the eigenvector represents the modal shape of the structure (i.e. deformation mode), and the eigenvalue represents the square of the frequency of the modal shape. For each region block, the system extracts the first m-order eigenvalue to form a frequency feature vector. In addition to the direct frequency value, the system can also calculate statistical indicators of the frequency spectrum, such as frequency distribution density, low frequency energy ratio, etc. For example, when the total energy ratio of low-order frequencies (such as the first 3 orders) is more than 70%, it is usually determined that the region is a low-frequency dominant region, which has a higher risk of fatigue accumulation and dynamic instability. The output is a set of frequency feature vectors corresponding to each modal analysis region block, which is used for downstream thermal-mechanical coupling coding and modal fingerprint generation.

[0112] thermal-mechanical coupling feature fusion is performed according to the component infrared feature data and the morphological frequency data to obtain thermal-mechanical coupling data;

[0113] Specifically, the thermal coordinates extracted from the infrared image are projected and mapped onto the 3D point cloud coordinate system corresponding to the modal analysis region. Using camera extrinsic parameters and attitude information, the temperature value of each pixel in the infrared image is associated with a 3D point in the point cloud model, constructing the local temperature distribution field for each modal region. For each modal analysis region, the spatial statistical correlation coefficient between its various structural modal deformation modes (i.e., modal shape functions) and the local temperature field is calculated. This correlation is used to measure the consistency between a certain mode and the thermal field distribution, denoted as the i-th... The thermal correlation coefficients corresponding to the first-order modes. Modes with high correlation characterize the principal deformation directions induced by heat. Within each modal region block, a weighted integral is performed on the shape function of each order-order mode, with the weight term being the temperature value of the point. Specifically, for the first... For a given mode, the thermal excitation weight is defined as the sum of the products of the squares of the mode function at each point and the temperature value. ,in For the first The thermal excitation weight of a mode indicates the degree of energy response of that mode under thermal excitation. The coordinates of a three-dimensional point within the modal region block. For the first The point set corresponding to each modal analysis region block For the first First mode at point The squared value of the modal shape function, For point The temperature value is derived from the temperature field information at the corresponding location of the infrared image's three-dimensional projection point. This weight reflects the energy concentration of the local modal response under thermal excitation. The multi-order modal frequencies, spatial correlation coefficients, and thermal offset energy values ​​of each modal region block are concatenated to form a unified fusion feature vector.

[0114] Modal resonance fingerprinting is performed on the thermal coupling data to obtain modal resonance fingerprint data.

[0115] Specifically, the thermal-modal fusion feature vector of each modal region block obtained in the previous step is input into a feature compression network to extract its low-dimensional fingerprint expression. Specifically, a self-encoder structure is used for dimension reduction embedding. The structure is composed of an encoder module and a reconstruction module, and only the encoder part is used to extract the latent expression vector. The input of the encoder is the fusion feature vector of each block region, and the output is a feature vector in a low-dimensional latent space, denoted as a compressed feature code, with a dimension of 16 to 32, which can not only retain sufficient discriminability but also facilitate storage and fast matching. The encoder needs to maintain the order of the modal frequency structure and the mapping consistency of the thermal anomaly contrast during the training process to ensure that the embedded vectors have order-preserving and discriminability in space. The compressed feature vector is normalized to a unit vector, and the L2 norm normalization method is used to make all fingerprint vector lengths uniform to 1, so as to enhance the comparability of distance measurement. At the same time, the corresponding structure number (ID) and thermal anomaly type label (Tag) are bound for each region block to form a complete modal resonance fingerprint triple.

[0116] Optionally, the thermal structure confidence vote comprises:

[0117] performing main architecture extraction according to the first component point cloud data to obtain main architecture data;

[0118] Specifically, a main ridge line extraction algorithm based on geometric morphological skeleton or principal curvature guidance is used to identify a continuous high-stability structure path in the component point cloud. The system calculates the structure consistency score of each point, which is based on the variance of the change of the neighborhood normal vector direction, which is used to reflect whether the surface orientation around the point is consistent; and the principal curvature change rate, which is used to reflect the geometric variation degree of the structure at the point. The structure consistency score is defined as a weighted combination of the above two, wherein the adjustment factor is used to balance the relative importance between the curvature change and the normal consistency. The system preferentially selects the path points with lower structure consistency scores (i.e., points with normal stability and morphological continuity), and constructs a continuous path segment thereof in the point cloud as a candidate skeleton line. For multiple locally extracted skeleton segments, the system uses a path optimization method to generate a main path. The method can be a shortest path network algorithm or a minimum spanning tree algorithm. A path set reflecting the dominant morphology of the overall component structure is formed.

[0119] performing thermal risk region division according to the second component point cloud data to obtain thermal risk region data;

[0120] Specifically, the input is the second component point cloud data, each point contains not only its three-dimensional spatial coordinates, but also a thermal response value, which can be a surface temperature value obtained by thermal imaging or a heat flow intensity index calculated by thermal modeling. The system statistically analyzes the thermal response values of all points, calculates the overall average thermal response value and the standard deviation. Then set the thermal anomaly judgment threshold, which is the "average value plus 1.5 times the standard deviation" of the thermal rise value. The points whose thermal response exceeds this threshold are defined as thermal anomaly points, which constitute a set of thermal anomaly points. The spatial clustering operation is performed on the above-mentioned set of thermal anomaly points. The system can select a density-based clustering algorithm (such as DBSCAN) to adaptively divide the clustering area according to the density of the points; a clustering algorithm based on neighborhood growth, which expands the area layer by layer by setting the spatial distance and thermal attribute difference threshold. After clustering, a number of local thermal risk sub-blocks are obtained, each of which represents a thermal anomaly aggregation area in the component. For each thermal risk sub-block, the system analyzes its boundary shape and thermal stress concentration characteristics. Specifically, based on the temperature gradient change trend in the sub-block, the thermal gradient field is calculated, and the boundary curve or boundary surface is projected along the gradient direction to fit the boundary shape of the thermal stress concentration. The boundary is used to determine the physical boundary range of the thermal risk block, and assists in judging the diffusion direction and potential damage boundary of the thermal influence.

[0121] According to the main architecture data and the thermal risk area data, the first component point cloud data and the second component point cloud data are matched with candidate labels to obtain candidate label matching data;

[0122] Specifically, the system constructs a spatial adjacency graph. The node set of the graph contains all points of the first component point cloud and the second component point cloud, and the edges are connected between point pairs with a spatial distance less than a set threshold. The graph structure is represented as a point cloud mixed graph G=(V,E), where V is the point set and E is the adjacent edge. The system takes the key points on the main architecture path as reference points (i.e. anchor points), and searches for thermal risk area points that are close in space within their spatial neighborhood. The system scores each candidate point pair. The matching score of the candidate point pair is calculated based on the following three factors: normal vector consistency score: the cosine value of the angle between the normal vectors of the two points is used to measure the geometric direction similarity; spatial distance similarity: based on the Euclidean distance between the two points, the spatial proximity is reflected by a Gaussian function decay method; thermal response intensity score: the temperature value or thermal response value of the thermal risk point is used as the risk significance weight. The three indicators are combined by weighting according to the weight coefficients to form a comprehensive scoring model. The weight settings are 0.3, 0.4, and 0.3, respectively, which control the influence of direction consistency, spatial distance, and thermal response on the total score. The system sorts all the scored candidate point pairs and keeps the top k pairs with the highest scores as the candidate semantic label matching pairs between the structural components and the thermal risk areas.

[0123] According to the candidate label matching data, structure-level semantic alignment is performed to obtain label reconciliation data;

[0124] Specifically, the system first extracts all label information in the candidate matching point pair set, and constructs a label semantic relationship graph. The nodes of the label graph represent different labels, and the edges represent the existence of candidate matching or semantic proximity relationship between labels. Based on prior semantic rules, label naming patterns (such as “lead head” and “high-temperature connection end”), and spatial distribution consistency, the system determines whether two labels refer to the same physical structure entity. If the judgment result is semantic equivalence or functional consistency, the system merges them into a unified label. For example, when the structure segments to which the candidate point pairs belong have a common direction or are located at the key hot node position of the main architecture, “lead head” and “high-temperature connection end” can be merged into the “key overheating node” label. To prevent label instability caused by point noise or local mismatch, the system constructs a local consistency propagation graph based on spatial adjacency relationship. In this graph, the nodes are points in the point cloud, and the edges represent the spatial adjacency relationship between point pairs. Through the label consistency propagation mechanism, the system transmits the information of the high-confidence label region to the boundary region, thereby strengthening the coherence of the label boundary and eliminating label repetition or conflict labeling phenomena. Label propagation can be realized by label smoothing or weighted voting, for example, the majority label of neighboring points determines the current point label, or the confidence weighted average is used to determine the label attribution.

[0125] The label reconciliation data is subjected to confidence voting fusion to obtain preliminary fusion data;

[0126] Specifically, the confidence factor is calculated as structure consistency degree temperature anomaly intensity label semantic consistency, wherein is the structure consistency weight, and the value is 0.5, is the temperature anomaly intensity weight data, and the value is 0.1, is the label semantic consistency weight, and the value is 0.4. The system divides the entire point cloud space into regular voxel units (Voxel Grid), and selects the point with the highest confidence in each voxel as the representative point of the region. The label, thermal response value and structure feature carried by the representative point will be the fusion output of the spatial region. When a voxel contains multiple label candidates from different data sources (such as the first component point cloud and the second component point cloud), the system will perform label conflict processing. Label conflict refers to multiple points having different labels but spatially overlapping or adjacent. Conflict resolution strategies include, for example, preferentially retaining labels with higher structure-thermal consistency scores; if the consistency scores are similar, a neighborhood label majority voting or confidence mean strategy can be used for secondary determination.

[0127] The preliminary fusion data is subjected to boundary fuzzy region re-estimation to obtain component point cloud data.

[0128] Specifically, the system performs label consistency analysis on each point, determines the spatial neighborhood (e.g. within a fixed radius or a fixed number of neighbors) of the point; calculates the proportion of points in the neighborhood that have different labels from the point; if the proportion exceeds a certain threshold (e.g. 30%), the point is marked as a "boundary ambiguous point", i.e. an area where the label distribution is unstable and confusing. For the identified boundary ambiguous points, the system can use the following two types of strategies for label re-estimation, such as the system calculates the frequency of each type of label in the neighborhood of the ambiguous point, and combines the confidence values of the neighborhood points as weights for weighted voting. Assign the label with the highest voting score to the point to achieve label smoothing and consistency enhancement. Alternatively, the system constructs a graph structure for all points, and uses spatial adjacency, normal vector consistency, thermal response similarity, etc. as features of the graph edges to construct a conditional random field model. By minimizing the global energy function, the global optimization of the label distribution is realized, making the label distribution more continuous in semantics and smoother in space. The system outputs the point cloud data after the boundary ambiguous area is re-estimated, called the fused component point cloud data.

[0129] Optionally, the component-level three-dimensional thermal structure analysis comprises:

[0130] According to the component point cloud data and the component infrared feature data, a three-dimensional thermal response projection modeling is performed to obtain a three-dimensional thermal response model;

[0131] Specifically, for each three-dimensional spatial point in the component point cloud, it is projected into the two-dimensional thermal image coordinate system through the perspective projection model of the camera. This process is based on the standard pinhole camera model, which uses the camera intrinsic matrix and extrinsic matrix to map the three-dimensional point to the image plane coordinates. For each projected point, the temperature value of its corresponding pixel in the thermal image is obtained. If the projected point is located in the pixel interval, nearest neighbor interpolation or bilinear interpolation method can be used to extract the temperature value to ensure the continuity and stability of the thermal information. For occluded areas or failed projection points, visual angle screening and distance threshold rejection can be performed to improve the projection accuracy. The point cloud data after projection mapping is extended with attributes, and each point is assigned a temperature attribute to form a complete thermal response triple data structure: each point is composed of three attributes: three-dimensional coordinates, unit normal vector and temperature value.

[0132] A local thermal flow vector field is constructed for the three-dimensional thermal response model to obtain thermal flow data;

[0133] Specifically, for any point in the three-dimensional thermal response model, a set of spatial neighborhood points is selected, and the local temperature gradient vector is estimated based on the temperature change and geometric relationship. The temperature difference and distance direction ratio between each pair of points and their neighborhood points are calculated; the average of all neighborhood points is calculated to obtain the thermal gradient vector of the point: the The temperature gradient vector at each point can be expressed as the ratio of the temperature difference between all points in the neighborhood to the distance between those points, obtained by weighted averaging of unit vectors along each direction. ,in For point The temperature gradient vector represents the local upward direction of the thermal field at that point. For point The number of neighboring points, that is, the number of points in its neighborhood. The number of points in The first point in the neighborhood The three-dimensional coordinate vector of a point For point The neighborhood point set is used as A set of points with a fixed radius or a fixed number centered on a point. For neighborhood points Temperature value, For target point Temperature value, For the first A three-dimensional coordinate vector representing the target point. This vector indicates the upward direction of the thermal field near that point. According to the physical laws of heat conduction, the direction of heat flow always points from the high-temperature region to the low-temperature region. Therefore, by taking the negative direction of each temperature gradient vector, a local heat flow vector is obtained. For each point, its heat flow vector is defined as the negative of the temperature gradient, that is, the heat flow vector represents the direction of local heat energy conduction, pointing towards the direction of the steepest temperature decrease. The output is a set of heat flow vector data appended to each point, forming a complete three-dimensional heat flow vector field.

[0134] Thermal-geometric joint perturbation analysis was performed on the heat flow data to obtain thermal geometric distortion data;

[0135] Specifically, for each target point in the point cloud, the degree of normal vector deviation between it and its surrounding neighbor points is calculated to depict whether the local surface structure has mutation, folding or bending phenomenon. For each point, its normal disturbance value is the average value of the cosine difference between the normal vector of the point and the normal vectors of its neighbor points. The higher the index value is, the stronger the surface geometric deformation at the point is, and the risk of structural discontinuity or morphological mutation exists. The temperature variation amplitude in the neighborhood of each point is evaluated to extract the local non-uniformity feature in the heat conduction process. The variance of all temperature values in the neighborhood of the point is calculated. The larger the variance is, the more uneven the heat distribution in the region is, which may be a heat source, a heat barrier or an abnormal heat coupling area. Combined with the geometric disturbance, the intensity of the heat flow direction and the heat fluctuation factor, a multi-factor joint disturbance scoring function is constructed to represent the comprehensive distortion risk value of the point. The joint disturbance score is composed of three linear weighted items: the first item is the geometric normal disturbance degree; the second item is the length of the heat flow direction vector at the point (i.e. the intensity of the direction of the fastest temperature change); the third item is the variance of the temperature variation in the neighborhood; the weights of each item can be set according to experience, for example, 0.4, 0.4 and 0.2 respectively. The joint disturbance scores of all points are statistically analyzed. By comparing the score of each point with the global average value and its standard deviation, when the score value of a certain point exceeds the global average value plus one times the standard deviation, it is marked as a thermal-geometric distortion point, representing the area where thermal deformation or damage occurs.

[0136] According to the thermal-geometric distortion data, a thermal structure region is extracted to obtain thermal structure data.

[0137] Specifically, the region in the thermal-geometric distortion data where the joint disturbance score is significantly higher than the background value is analyzed. A gradient-based density estimation clustering method (such as the Mean Shift algorithm) is used to cluster the continuous distribution region of disturbance intensity in three-dimensional space to identify potential thermal anomaly aggregation areas. This method does not need to pre-set the number of clusters and can adaptively form the boundary of the thermal distortion region. For each clustered region, it is determined whether it can be an independent thermal structure unit. It needs to meet multiple judgment criteria at the same time: spatial connectivity: the point cloud in the clustered region has structural continuity in three-dimensional space, and the boundary does not have significant fracture; thermal gradient consistency: the heat flow direction vector in the region remains highly consistent in direction, indicating that the heat conduction path tends to be stable; geometric disturbance coverage requirement: more than a certain proportion of points (such as 60%) in the region have obvious geometric disturbance response, that is, the normal change rate or warping factor exceeds the set threshold; the clustered region that meets the above three criteria is defined as a thermal structure unit.

[0138] Optionally, S3 comprises:

[0139] S31, mapping a thermal-fatigue model according to the thermal structure data to obtain a thermal fatigue model;

[0140] Specifically, the input is a set of multiple thermal structure units, each containing its spatial boundary information, local temperature gradient value, thermal disturbance amplitude, and thermal flow direction information. If necessary, historical thermal imaging sequence data or simulated thermal field data can be combined as temperature change sources. For each thermal structure unit, extract its highest temperature and lowest temperature in a thermal cycle period, and calculate the temperature change amplitude, denoted as , i.e. Under the condition of thermal elastic approximation, estimate the thermal stress amplitude caused by the temperature fluctuation, and the calculation formula is: , where is the Young's modulus, is the thermal expansion coefficient. Take the thermal stress amplitude calculated above as the input parameter to estimate the fatigue life (i.e. the number of cycles that can be tolerated before failure) of the thermal structure unit under the action of thermal cycle. The calculation formula is: , where is the predicted fatigue life of the th thermal structure unit under the action of thermal cycle, i.e. the number of cycles that can be tolerated, is the material fatigue constant, obtained by experiment, reflecting the benchmark value of the material's fatigue resistance performance, is the equivalent thermal stress amplitude of the th thermal structure unit caused by thermal expansion and temperature change, is the material fatigue index, which controls the sensitivity of fatigue life to stress change. According to the set fatigue life threshold, fatigue level classification is carried out for each thermal structure unit: if the fatigue life is less than the set threshold (e.g. 10-4 times), it is marked as a high fatigue risk area; if the fatigue life is within a safe range, it is marked as a normal area; if the regional temperature disturbance is small or the thermal gradient is flat, it can be considered as a redundant low risk area. Output the thermal fatigue model data set.

[0141] S32, extract the erosion path data according to the thermal fatigue model;

[0142] Specifically, each thermal structure unit is regarded as a node in the graph structure, and the connection edges between nodes are established by spatial adjacency relationship. For each pair of adjacent thermal structure units, the weight of the connection edge between them is defined as: 1 minus the ratio of the mean value of the fatigue life of the two units to the global maximum fatigue life. In other words, the lower the fatigue life of the region, the higher the edge weight, representing the stronger possibility of erosion propagation. Set all structure units with high-risk fatigue level as the source node set, and perform shortest path search on the fatigue level graph using a heuristic search algorithm (such as A* algorithm or Dijkstra algorithm). The path selection gives priority to the connected region with higher edge weight to obtain the potential erosion propagation path. The path termination condition can be set, including but not limited to the path crossing length exceeding the preset threshold (for example, a certain proportion of the length of a certain structure skeleton), or the path connecting multiple structure units of different component types. Any two consecutive nodes in the path must have spatial proximity, i.e. the distance between the center points does not exceed a certain threshold; the overall trend of the path needs to be basically consistent with the direction of the thermal gradient, with a deviation angle of not more than 45 degrees; the point cloud density of the region passed by the path should not be lower than a certain threshold, for example, not less than 120 points per cubic centimeter. The output is an erosion path data set, which includes multiple erosion propagation paths, each path consisting of a series of ordered thermal structure units, and the path is accompanied by attribute information such as spatial trajectory, path weight sum, start and end node identification, etc.

[0143] S33, erosion impact domain simulation is performed according to the erosion path data to obtain component erosion data.

[0144] Specifically, for each erosion path, a local heat flow direction information or equivalent stress distribution is combined to construct an erosion impact kernel function. The impact function is modeled in a Gaussian decay form: for any point in space, its erosion impact value is an exponential function that rapidly decays with the square of its distance to the path , where is the local erosion impact value at point caused by the th erosion path, is a natural exponential function used to model the exponential decay characteristics of erosion impact, is the coordinate vector of any position point in three-dimensional space, is the centerline coordinate set of the th erosion path (which can be approximately represented as the point closest to on the path), For the diffusion scale parameter, which is used to control the spatial diffusion range of the erosion effect, it is related to factors such as thermal conduction length, material damage propagation radius, etc., that is, it is obtained by mapping the preset parameter library through the above-mentioned factors. The scale parameter of the kernel function is used to regulate the diffusion width of the influence domain, which can be set according to the material thermal diffusivity or the actual fatigue propagation range. The influence kernel functions of all erosion paths are superimposed to obtain the cumulative erosion influence field of the entire power grid channel component structure. This field reflects the strength distribution of the potential fatigue accumulation area radiated by the path, and constitutes a continuous three-dimensional space field. Set the influence intensity threshold, and all positions in the global influence field whose influence values are greater than the threshold form the erosion influence domain volume. This volume area is the potential structure weakening or local performance degradation risk area. According to the size of the influence value, the erosion influence domain is further divided into three areas: the core erosion area (influence value greater than 0.8): representing high fatigue accumulation degree, significant structure strength attenuation, suggesting immediate maintenance or replacement; the boundary risk area (influence value between 0.5 and 0.8): representing medium fatigue level, in a subcritical state, suggesting intensive monitoring; the diffusion buffer area (influence value between 0.3 and 0.5): representing the fatigue influence spread area, which should be included in the periodic inspection plan. For each three-dimensional space point within the erosion influence domain, bind the following attribute information of the point: the spatial coordinate position of the point; the erosion path number to which it belongs; the corresponding erosion influence value; the graded risk level label.

[0145] Optionally, S4 comprises:

[0146] S41, according to the component erosion data, the channel component wear label fusion of the power grid channel infrared data and the power grid channel point cloud data is obtained Wear fusion data;

[0147] Specifically, the system adopts octree structure to divide the three-dimensional point cloud data of the power grid channel in space, and establishes an efficient spatial index system. The information containing spatial coordinates and corresponding risk levels in the component erosion data is mapped into the octree nodes to construct a spatial risk heat map at the voxel level, so that each spatial unit has a corresponding risk level label (such as no risk, moderate risk, and high risk). The system performs spatial back projection processing on the infrared image of the power grid channel, and converts the two-dimensional infrared image pixel coordinates into three-dimensional spatial coordinates. This process relies on camera calibration parameters, including intrinsic matrix (K), rotation matrix (R), and translation vector (T), using the following back projection steps: after combining the pixel coordinates and depth estimation, the intrinsic matrix is applied to obtain the three-dimensional ray in the image space; the extrinsic matrix is used to complete the transformation of the image coordinates to the point cloud coordinate system; the nearest neighbor search is performed in the three-dimensional point cloud to complete the spatial mapping of the infrared point and the point cloud point. After alignment, the infrared thermal anomaly label (such as high temperature point label) can be projected into the point cloud space and fused with the structural points. For each spatial unit (such as voxel or point-level data) in the point cloud data, the system calculates the fusion score based on the following three information sources: erosion risk level: derived from the component erosion label, indicating the potential structural wear risk of the current area (assigned levels such as 0 / 1 / 2); infrared thermal intensity value: representing the thermal response degree of the current point, which is normalized and used for calculation; point cloud density standard deviation: representing the volatility of local point density, used to speculate the sparse or unstable area of the structure, also normalized. The above three dimensions correspond to three sub-items in the fusion score model, and the system calculates the fusion risk score of each point by using weighted linear combination. The weights are erosion risk level weight: 0.5; thermal intensity normalized weight: 0.3; point density standard deviation normalized weight: 0.2. The system outputs the wear fusion data.

[0148] S42, constructing a three-dimensional structure-state coupling grid model according to the wear fusion data to obtain a three-dimensional structure model;

[0149] Specifically, the system performs meshing processing on the point cloud space region covered by the wear fusion data, and adopts a three-dimensional Delaunay tetrahedral partitioning algorithm to divide the space into a plurality of topologically coherent tetrahedral units. The Delaunay partitioning mode can effectively avoid the generation of small-angle grid units and is suitable for the modeling needs of irregular point cloud data. For each meshed grid unit, the system counts the fusion point set contained in it and calculates the following three types of state indicators, such as the average wear risk value, i.e. the average risk score of the fusion point set; the local temperature gradient strength, based on the change of infrared thermal intensity, estimates the modulus of the thermal distribution gradient in the grid; the point cloud sparsity change rate, which counts the local change degree of the point cloud density in the grid. The above indicators constitute the state feature vector of the grid unit, denoted as the structure-state description vector. The system takes each grid unit as a node in the graph model, embeds the above state vector into the node attribute, and forms a set of grid graph nodes with state information. The system constructs the connection edges between nodes according to the geometric connection relationship between grids (such as sharing vertices, sharing edges or faces). For each pair of adjacent grid units, calculate the similarity between their state vectors, such as using Euclidean distance, cosine similarity or Mahalanobis distance as the basis of edge weight. The output three-dimensional structure model includes the grid unit set and its spatial topological connection relationship; the structure-state feature vector of each grid (risk average, thermal gradient strength, density difference); the graph model edge weight constructed based on the structure connectivity and state similarity;

[0150] S43, fitting the wear evolution trajectory of the three-dimensional structure model to obtain a power grid channel model.

[0151] Specifically, the system establishes a time series state vector for each grid element in the three-dimensional structure model based on historical observation data or simulated evolution data. Specifically, if a grid element has state indicator records (such as risk value, thermal gradient, point cloud sparsity, etc.) in multiple time segments, a state sequence is constructed in chronological order as a record of the unit's evolution over time. According to the type difference of the grid element evolution behavior, the system adopts two types of modeling strategies: for the state evolution trend of individual points or local sub-blocks, the system uses mathematical models such as exponential curve, logarithmic growth function or polynomial curve to fit the trend of risk level increasing with time. Alternatively, if some wear-out behavior is a spatial diffusion trend, the system will construct a "state transition path graph" with time as the axis and space as the graph nodes. In this graph structure, nodes represent regional state units at different time points, and edges represent state transmission between adjacent nodes in time or space. Then an integrated prediction model is introduced, combining long short-term memory network (LSTM) and graph neural network (GNN) structure, to model and predict the joint change trend of state in space and time dimensions. According to the fitting results, the system identifies the high-risk path that continuously accelerates from the three-dimensional structure model, and explicitly labels the grid area that is expected to reach or exceed the wear threshold at a future preset time (e.g. T time). The labeling content includes the future predicted risk level; the time when the threshold is expected to be reached; whether it is in the main structure area; the type of component, etc. The system outputs a power grid channel three-dimensional state model that integrates structure, state and time dimension information, including the following contents: time series state information of each grid element; fitted local risk evolution curve parameters; regional state transition path and prediction trend; risk warning labeled area and visualization atlas.

[0152] Optionally, the present application also provides a three-dimensional identification modeling system for power grid channels, which is used to execute the three-dimensional identification modeling method for power grid channels as described above, and the three-dimensional identification modeling system for power grid channels comprises:

[0153] An infrared thermal anomaly feature extraction module is configured to obtain power grid channel infrared data and power grid channel point cloud data, and perform thermal anomaly shape extraction based on the power grid channel infrared data to obtain component infrared feature data.

[0154] A three-dimensional thermal structure analysis module is configured to perform component domain division based on the power grid channel point cloud data to obtain component point cloud data, and perform component-level three-dimensional thermal structure analysis based on the component point cloud data and the component infrared feature data to obtain thermal structure data.

[0155] A structure erosion evolution simulation module is configured to perform component erosion processing based on the thermal structure data to obtain component erosion data.

[0156] The channel level wear model construction module is configured to perform power grid channel wear model construction on the power grid channel infrared data and the power grid channel point cloud data according to the component erosion data, and obtain a power grid channel model.

[0157] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application being defined by the appended application file and not by the above description, therefore all variations falling within the meaning and scope of the equivalent requirements of the application file are intended to be included within the present application.

[0158] The above description is merely that of the specific embodiments of the application and allows those skilled in the art to understand or implement the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional recognition and modeling method for power grid channels, characterized in that, The method includes: S1. Acquire infrared data and point cloud data of the power grid channel, and extract thermal anomaly morphology based on the infrared data of the power grid channel to obtain infrared feature data of the component. S2. Based on the power grid channel point cloud data, the power grid components are deconstructed and reconstructed to obtain the first component point cloud data; based on the power grid channel point cloud data and the component infrared feature data, modal resonance fingerprints are extracted to obtain modal resonance fingerprint data; based on the modal resonance fingerprint data, the power grid channel point cloud data is divided into regions to obtain the second component point cloud data; based on the first component point cloud data and the second component point cloud data, thermal structure confidence voting is performed to obtain component point cloud data; based on the component point cloud data and the component infrared feature data, component-level three-dimensional thermal structure analysis is performed to obtain thermal structure data. S3. Perform component erosion processing based on thermal structure data to obtain component erosion data; S4. Based on the component erosion data, construct the power grid channel wear model using the power grid channel infrared data and power grid channel point cloud data to obtain the power grid channel model.

2. The method according to claim 1, characterized in that, The extraction of thermal anomaly morphology includes: Thermal region structure enhancement is performed based on infrared data of the power grid channel to obtain thermal region enhancement data; Isothermal morphological boundary spectrum is extracted based on thermal region enhancement data to obtain isothermal data; The isothermal data were subjected to structural alignment thermal offset extraction to obtain the infrared feature data of the component.

3. The method according to claim 1, characterized in that, The deconstruction and reconstruction of the power grid components includes: Structural saliency response maps are constructed based on power grid channel point cloud data to obtain structural response map data; Point cloud structure map data is obtained by processing the structure response map data; Structural semantic deconstruction is performed based on the point cloud structure diagram data to obtain point cloud semantic data; The sub-block structure is extracted based on the point cloud semantic data to obtain the point cloud data of the first component.

4. The method according to claim 1, characterized in that, The modal resonance fingerprint extraction includes: Modal analysis region blocks are divided based on power grid channel point cloud data to obtain modal block data; Morphological frequency extraction is performed on the modal block data to obtain morphological frequency data; Thermal-shape coupling feature fusion is performed based on component infrared feature data and morphological frequency data to obtain thermal-shape coupling data; Modal resonance fingerprinting is performed on the thermal coupling data to obtain modal resonance fingerprint data.

5. The method according to claim 1, characterized in that, The thermal structure confidence vote includes: The main architecture is extracted based on the point cloud data of the first component to obtain the main architecture data. The thermal risk area is divided based on the point cloud data of the second component to obtain thermal risk area data; Based on the main architecture data and the hot risk area data, candidate label matching is performed on the point cloud data of the first component and the point cloud data of the second component to obtain candidate label matching data. Structural-level semantic alignment is performed based on candidate label matching data to obtain label harmonization data; The label harmonic data is fused using confidence voting to obtain preliminary fused data; The preliminary fused data is re-estimated for blurred boundary regions to obtain component point cloud data.

6. The method according to claim 1, characterized in that, The component-level three-dimensional thermal structure analysis includes: Three-dimensional thermal response projection modeling is performed based on component point cloud data and component infrared feature data to obtain a three-dimensional thermal response model. A local heat flow vector field is constructed from the three-dimensional thermal response model to obtain heat flow direction data; Thermal-geometric joint perturbation analysis was performed on the heat flow data to obtain thermal geometric distortion data; Thermal structure data is obtained by extracting thermal structure regions from thermal geometric distortion data.

7. The method according to claim 1, characterized in that, S3 includes: A thermal fatigue model is obtained by mapping thermal structure data to a thermal fatigue model. Erosion path data is obtained by extracting erosion paths based on a thermal fatigue model. Based on the erosion path data, the erosion influence domain is simulated to obtain the component erosion data.

8. The method according to claim 1, characterized in that, S4 includes: Based on component erosion data, wear labels of channel components are fused with infrared data and point cloud data of power grid channels to obtain wear fused data. A three-dimensional structure-state coupled mesh model is constructed based on wear fusion data to obtain a three-dimensional structure model. The wear evolution trajectory of the three-dimensional structural model is fitted to obtain the power grid channel model.

9. A three-dimensional recognition and modeling system for power grid channels, characterized in that, For executing the three-dimensional identification and modeling method for power grid channels as described in claim 1, the three-dimensional identification and modeling system for power grid channels includes: The infrared thermal anomaly feature extraction module is used to acquire infrared data and point cloud data of the power grid channel, and extract the thermal anomaly morphology based on the infrared data of the power grid channel to obtain infrared feature data of the components. The 3D thermal structure analysis module is used to perform component domain segmentation based on power grid channel point cloud data to obtain component point cloud data; and to perform component-level 3D thermal structure analysis based on component point cloud data and component infrared feature data to obtain thermal structure data. The structural erosion evolution simulation module is used to process component erosion based on thermal structural data to obtain component erosion data; The channel-level wear model construction module is used to construct a power grid channel wear model based on component erosion data, infrared data of the power grid channel, and point cloud data of the power grid channel, thereby obtaining the power grid channel model.

Images