Power transmission line three-dimensional full-attribute reconstruction method, system and device and storage medium

Abstract

The invention discloses a power transmission line three-dimensional full-attribute reconstruction method, system and device and a storage medium, and relates to the field of power transmission line monitoring, and the method comprises the steps: obtaining the visible light data of a power transmission line, and generating a time-aligned original data set; performing low-illumination enhancement on the original data set to obtain an enhanced image and an illumination distribution diagram, and performing multi-modal data registration based on the enhanced image and the illumination distribution diagram to form a multi-source data cube under a unified coordinate system; carrying out multi-scale feature expression and cross-domain mapping on the cube, deeply fusing features to construct a three-dimensional model, evaluating and detecting the model, and starting local super-resolution reconstruction on a defective region to realize dynamic updating of the model; according to the method, the accuracy, integrity and real-time performance of the three-dimensional model of the power transmission line can be effectively improved, potential defects in the power transmission line can be found in time, a powerful guarantee is provided for safe and stable operation of the power transmission line, and the method has important application value and practical significance in operation, maintenance and management of a power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power transmission line monitoring, and in particular to a power transmission line three-dimensional full attribute reconstruction method, system, device and storage medium. BACKGROUND

[0002] As the core artery of the power system, the safe and stable operation of the power transmission line is directly related to the normal operation of the national economy and social life. With the rapid development of the ultra-high voltage power grid and the continuous expansion of the transmission network coverage area, the line inspection work is facing unprecedented challenges. The traditional inspection method exposes many technical bottlenecks in complex environments and adverse weather conditions, especially in low light scenes, and urgent technical innovation is needed.

[0003] Current power transmission line inspection mainly relies on manual inspection and unmanned aerial vehicle visible light camera, both of which have obvious limitations in low light conditions. Manual inspection requires workers to climb the tower or use binoculars and other equipment for ground observation, which is not only inefficient, but also almost impossible to work in low visibility conditions such as night, fog, etc. More seriously, in complex terrain such as mountains and forests, manual inspection poses a great safety risk. According to statistics, more than 60% of power transmission line failures occur at night or in adverse weather conditions, which is exactly the vacuum period for manual inspection. Although unmanned aerial vehicle visible light camera inspection improves efficiency to some extent, its core technical bottleneck lies in its heavy dependence on light conditions. In low light environments, the signal-to-noise ratio of images captured by visible light cameras drops sharply, with severe noise and loss of details in the images. This directly leads to the failure of computer vision-based defect detection algorithms, and makes three-dimensional reconstruction techniques that rely on image feature point matching ineffective. Experimental data shows that in an environment with an illumination of less than 10 lux, the feature point matching success rate of the traditional SFM (Structure from Motion) algorithm is less than 20%, and the reconstructed model will have a large number of holes and distortions. In low light conditions, there are many problems in power transmission line inspection technology that need to be addressed. Various technical routes tried by the industry all have obvious defects: LiDAR technology can provide accurate three-dimensional point cloud data, but the cost is high, and it cannot obtain target material and surface state information. Multi-modal fusion faces technical bottlenecks, with large differences in working mechanism and data characteristics of different sensors, and problems such as temporal and spatial registration difficulties and information redundancy in data fusion, such as registration error of visible light and LiDAR data under dynamic flight exceeding 15 cm. Three-dimensional reconstruction technology has poor adaptability, and the performance of SFM algorithm based on vision decreases sharply in low light, and the point cloud data of LiDAR-based reconstruction method has uneven density, and real-time requirements are difficult to meet. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] Therefore, the technical problem solved by the present application is that the existing power line inspection technology has the problems of poor image quality, low three-dimensional reconstruction accuracy and difficult multi-source data fusion in low light environment.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the embodiments of the present application provide a power line three-dimensional full attribute reconstruction method, comprising:

[0008] Obtaining visible light data of the power line to generate a time-aligned original data set;

[0009] Performing low-light enhancement based on the original data set to generate an enhanced image and a light distribution map;

[0010] Based on the enhanced image and the light distribution map, multi-modal data registration is performed to form a multi-source data cube in a unified coordinate system;

[0011] Multi-scale feature expression and cross-domain mapping are performed on the multi-source data cube, the processed and mapped features are deeply fused, and a three-dimensional model of the power line is constructed;

[0012] The three-dimensional model of the power line is evaluated and detected, local super-resolution reconstruction is started for the area with defects, and dynamic updating of the three-dimensional model is realized.

[0013] As a preferred scheme of the power line three-dimensional full attribute reconstruction method, wherein:

[0014] The low-light enhancement based on the original data set to generate an enhanced image and a light distribution map comprises:

[0015] A deep neural network is designed by fusing a physical imaging model, and illumination invariance constraints and gradient-guided loss are added in the network design: the consistency of the illumination components of the images under different exposures is modeled, the constraint condition is that the values of the estimated illumination components at the spatial corresponding points under different exposures remain consistent, the gradient-guided loss is introduced in the loss function of the deep neural network, the gradient direction of the reflection component is kept consistent with that of the original image, and the light distribution map is output; the illumination and reflection components of the images in the original data set are decomposed by the designed deep neural network, the gradient-guided loss is used to keep the texture details and perform adaptive correction, and the enhanced image with dynamic range expansion is output.

[0016] The beneficial effects of the preferred technical scheme are that by fusing the physical imaging model and adding the illumination invariance constraints and the gradient-guided loss, the illumination distribution of the image can be more accurately estimated, the consistency of the illumination components under different exposures is ensured, and the reflection component better retains the texture information of the original image, thereby providing more reliable illumination information for subsequent multi-modal data processing.

[0017] As a preferred solution for the 3D full attribute reconstruction method of transmission lines, the following are the methods:

[0018] The multimodal data registration based on the enhanced image and the illumination distribution map to form a multi-source data cube in a unified coordinate system includes:

[0019] Initial alignment of point clouds and images is performed based on the inertial measurement unit pose. Visible light feature radar geometric key points are matched through feature pyramids. Differentiable rendering is used to achieve sub-pixel precision registration, and a multi-source data cube is output in a unified coordinate system.

[0020] The beneficial effects of this preferred technical solution are: using the inertial measurement unit pose for initial alignment, combined with feature pyramid matching and sub-pixel precision registration of differentiable rendering, it is possible to more accurately unify data from different modalities into the same coordinate system, forming a multi-source data cube, and providing an accurate and consistent data foundation for subsequent three-dimensional modeling.

[0021] As a preferred solution for the 3D full attribute reconstruction method of transmission lines, the following are the methods:

[0022] The matching of visible light features and radar geometric key points through feature pyramid includes:

[0023] For visible light images, information is constructed by combining key point extraction and deep convolution features;

[0024] For radar data, project the point cloud into a depth image and extract relevant features;

[0025] A shared embedding space is constructed, and the features of each modality are mapped to the common space through the feature encoding network. The contrast loss is used to optimize the embedding space to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs.

[0026] The beneficial effects of this preferred technical solution are: using different feature extraction methods for data from different modalities can fully explore the characteristics and information of each modality. Constructing a shared embedding space and optimizing it with contrastive loss can effectively fuse the features of different modalities, making the features of each modality more distinguishable and relevant in the common space, which is beneficial for subsequent multi-scale feature expression and cross-domain mapping.

[0027] As a preferred solution for the 3D full attribute reconstruction method of transmission lines, the following are the methods:

[0028] The multi-scale feature expression and cross-domain mapping of multi-source data cubes, deep fusion of processed and mapped features, and construction of a three-dimensional model of the transmission line include:

[0029] A three-dimensional reconstruction network with a fusion attention mechanism is designed, and a multi-level attention module is used to realize the fusion of multi-source data and three-dimensional modeling; through the cooperation of spatial attention, channel attention and cross-modal attention mechanisms, a three-dimensional digital model with complete physical properties is constructed;

[0030] Different attention mechanisms are used to extract features for different modal data: for visible light images, a CNN network with a spatial attention mechanism is used to extract texture features; for radar point clouds, a curvature-based geometric attention mechanism is used to extract geometric features.

[0031] As an optimal solution of the three-dimensional full-attribute reconstruction method of the power transmission line, wherein:

[0032] The three-dimensional model of the power transmission line is evaluated and detected, and local super-resolution reconstruction is started for the area with defects to realize dynamic updating of the three-dimensional model, which includes:

[0033] The three-dimensional model of the power transmission line is voxelized, and a three-dimensional attention mechanism is used to fuse multi-scale features to realize real-time identification of defects including broken wires and damaged insulators, calculate the defect probability, and start local super-resolution reconstruction for the area with a defect probability reaching a preset threshold to realize dynamic updating of the three-dimensional model.

[0034] The beneficial effects of the preferred technical solution are: through voxelization and three-dimensional attention mechanism to fuse multi-scale features, defects in the three-dimensional model of the power transmission line can be more accurately identified. Calculate the defect probability and start local super-resolution reconstruction for the area reaching the preset threshold to realize dynamic updating of the three-dimensional model, which can timely discover and repair potential problems in the power transmission line, improve the safety and reliability of the power transmission line, and at the same time avoid unnecessary reconstruction of the entire model, saving computing resources.

[0035] In a second aspect, an embodiment of the present application provides a three-dimensional full-attribute reconstruction system of a power transmission line, comprising:

[0036] A data acquisition module is configured to acquire visible light data of the power transmission line and generate a time-aligned original data set;

[0037] A low-light image enhancement module is configured to perform low-light enhancement based on the original data set to generate an enhanced image and a light distribution map;

[0038] A multi-modal data fusion module is configured to perform multi-modal data registration based on the enhanced image and the light distribution map to form a multi-source data cube in a unified coordinate system;

[0039] A three-dimensional reconstruction module is configured to perform multi-scale feature expression and cross-domain mapping for the multi-source data cube, deeply fuse the processed and mapped features, and construct a three-dimensional model of the power transmission line.

[0040] The defect detection module is used to evaluate and detect the three-dimensional model of the transmission line, initiate local super-resolution reconstruction of the defective areas, and realize dynamic updating of the three-dimensional model.

[0041] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0042] memory and processor;

[0043] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the three-dimensional full attribute reconstruction method for transmission lines as described in any embodiment of the present invention.

[0044] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method for reconstructing the three-dimensional full attributes of a transmission line.

[0045] The invention achieves a deep complementarity between visible light and LiDAR data through an innovative sensor fusion mechanism: LiDAR provides a sub-centimeter (±0.5cm) geometric framework, combined with visible light texture enhanced by deep learning (resolution up to 4096×2160). The resulting 3D model effectively reduces the time required for multi-data fusion. It can be integrated into drones for mobile inspections or deployed at fixed monitoring points to form an intelligent perception network. Backhaul enables cloud-based collaborative analysis to meet monitoring needs in diverse scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0047] Figure 1 This is an overall flow chart of the method for reconstructing all three-dimensional attributes of a transmission line provided by the present invention. DETAILED DESCRIPTION

[0048] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0049] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides a method for reconstructing all three-dimensional attributes of a transmission line, comprising:

[0050] S1: Obtain visible light data of the transmission line and generate a time-aligned raw data set;

[0051] S2: Perform low-light enhancement based on the original dataset to generate enhanced images and light distribution maps;

[0052] S3: Based on the enhanced image and illumination distribution map, multimodal data registration is performed to form a multi-source data cube in a unified coordinate system;

[0053] S4: Perform multi-scale feature expression and cross-domain mapping on multi-source data cubes, deeply fuse the processed and mapped features, and construct a 3D model of the transmission line;

[0054] S5: Evaluate and inspect the 3D model of the transmission line, initiate local super-resolution reconstruction of defective areas, and dynamically update the 3D model.

[0055] It should be noted that through steps S1-S5, visible light data of the transmission line is acquired and a time-aligned raw data set is generated. On this basis, low-light enhancement, multimodal data registration, multi-scale feature expression and cross-domain mapping, and deep fusion are performed to construct a three-dimensional model. Finally, the model is evaluated, tested, and locally super-resolution reconstructed to achieve dynamic updates. This embodiment integrates multiple technical means to effectively improve the accuracy, completeness, and real-time performance of the three-dimensional model of the transmission line, promptly discover potential defects in the transmission line, and provide strong guarantees for the safe and stable operation of the transmission line. It has important application value and practical significance in the operation, maintenance, and management of power systems.

[0056] Example 2, reference Figure 1 , which is an embodiment of the present invention, provides a method for reconstructing all three-dimensional attributes of a transmission line based on the previous embodiment, including:

[0057] In this embodiment, obtaining visible light data of the transmission line in step S1 and generating a time-aligned raw data set includes:

[0058] Specifically, low-light visible light images (Sony IMX585) and LiDAR point clouds (±2cm accuracy) are acquired. A time-aligned raw dataset is generated through black level correction, temperature calibration, and point cloud denoising. A hardware synchronization trigger mechanism ensures microsecond-level temporal alignment between the visible light camera and LiDAR. GPS / IMU provides initial pose estimation, and the ORB-SLAM3 framework implements feature-level coarse registration.

[0059] In this embodiment, performing low-light enhancement based on the original data set in step S2 to generate an enhanced image and a light distribution map includes:

[0060] Based on the original dataset, the physical imaging model is integrated with deep learning, and an improved RetinexNet architecture is adopted. Illumination invariance constraints and gradient-guided loss are added to the network design. This improves the image dynamic range (up to 90dB) while effectively suppressing noise amplification (PSNR is increased by more than 15dB). At the same time, the illumination distribution map is output for radiation consistency correction in subsequent multimodal fusion.

[0061] Specifically, in order to more accurately separate the reflection component and the illumination component of the image, the illumination invariance constraint is introduced. The consistency of the illumination components of images under multiple different exposures is modeled. The formula is as follows:

[0062]

[0063] Among them I i and I j They represent the estimated illumination components of the same scene under different exposures, respectively, and constrain their values ​​at corresponding points (x, y) in space to remain consistent, thereby enhancing the separation accuracy.

[0064] In order to enhance the network's ability to maintain image edges and texture details, a gradient-guided loss is introduced into the loss function. This loss function encourages the gradient of the reflection component to be consistent with the gradient direction of the original image, which is defined as follows:

[0065]

[0066] Where R is the reflection component, S is the input image, Represents the gradient operator of the image. This loss encourages the network to preserve edge details when enhancing the image and prevent image blur.

[0067] The improved RetinexNet network is used to decompose the illumination and reflection components of the image, and a gradient-guided loss (λ1=0.5,λ2=0.2) is used to preserve texture details. Adaptive gamma correction γ=2.2-0.5log10lux is performed in the HSV space to output an enhanced image with a dynamic range extension of 15dB.

[0068] In another possible implementation, the improved Retinex-Net can capture the illumination changes at different spatial scales by introducing a multi-scale feature extraction module, thereby improving the restoration capability of low-light areas. In addition, an attention mechanism is added in the decoding stage to enhance the attention to key areas such as wire surface textures. Finally, according to the illumination-reflectance separation theory in the imaging process, the brightness restoration process of the image is explicitly constrained to avoid noise amplification caused by over-enhancement.

[0069] In the embodiment, the multi-modal data registration based on the enhanced image and the illumination distribution map in the step S3 to form the multi-source data cube in the unified coordinate system comprises:

[0070] Based on the enhanced image and the illumination distribution map, first, the initial alignment of the point cloud and the image is performed based on the IMU (Inertial Measurement Unit) pose, and then the visible light SIFT (Scale-Invariant Feature Transform) features and the LiDAR (Light Detection and Ranging) geometric key points are matched through the feature pyramid matching, and finally the sub-pixel level precise registration (error <0.5 pixels) is realized by using the differentiable rendering, and the multi-source data cube in the unified coordinate system is output.

[0071] Specifically, multi-scale feature representations are constructed for each modality based on the feature pyramid network (FPN): for the visible light image, the hierarchical local texture and structure information are constructed by combining the SIFT key point extraction and the deep convolutional features; the LiDAR data is projected into a depth image, and the gradient edges and geometric structure features thereof are extracted. The pyramid features constructed by each modality are denoted as and wherein i represents the pyramid level.

[0072] To realize the feature alignment between different modalities, a shared embedding space is constructed, and the features of each modality are mapped to the public space through a three-branch feature encoding network, and the feature mapping function is defined as: The distance minimization of the positive sample pair and the distance maximization of the negative sample pair in the embedding space are optimized through the contrast loss (such as TripletLoss):

[0073]

[0074] wherein Z a , Z p , Z nThey represent the anchor point, positive sample and negative sample features respectively, and α is the boundary interval.

[0075] In the geometric registration stage, a differentiable rendering mechanism is introduced to achieve accurate projection between the 3D structure based on the LiDAR point cloud and the image plane. i =[X i ,Y i ,Z i ,1] T Projected to the image coordinate system through the external parameter matrix [R|t] and the camera internal parameter matrix K: i =K[R|t]P i To keep the gradient transferable, Soft Rasterizer or Z-buffer based differentiable technology is used to project the 3D points into a rendered image. The pixel reconstruction loss is defined as:

[0076]

[0077] The final total loss function is:

[0078] L total =λ1L triplet +λ2L render +λ3L reg

[0079] Furthermore, by combining IMU pose (0.1m accuracy) with feature matching (SIFT+ISS key points), accurate registration is achieved by optimizing the objective function, which is expressed as:

[0080]

[0081] LiDAR reliability test: When the point cloud signal-to-noise ratio (SNR) is less than 15dB, pulsed auxiliary lighting (850nm wavelength, meeting eye safety standards) is activated;

[0082] Anti-noise mode: When the lux is lower than 10 lux for 5 seconds, the learning rate of the first layer of CNN is reduced by 50% and time series filtering is enabled to ensure a detection reliability of more than 85%. The weight parameters are calculated by sliding average lux. avg =0.9lux t-1 +0.1lux t Smoothing is performed to avoid model oscillation caused by sudden changes in illumination.

[0083] In this embodiment, the multi-scale feature expression and cross-domain mapping are performed on the multi-source data cube in step S4, and the processed and mapped features are deeply integrated to construct a three-dimensional model of the transmission line, including:

[0084] For multi-source data cubes, a three-dimensional reconstruction network with a fusion attention mechanism is designed to realize intelligent fusion of multi-source data and three-dimensional modeling through multi-level attention modules.

[0085] Specifically, through the cooperative work of spatial attention, channel attention, and cross-modal attention mechanisms, a three-dimensional digital model with complete physical properties is constructed.

[0086] Furthermore, in the feature extraction stage, for visible light images, texture features are extracted through a CNN network with a spatial attention mechanism. This module uses a 7x7 convolution kernel to generate a spatial weight map, highlighting key areas such as structural edges. For LiDAR point clouds, geometric features are extracted through a curvature-based geometric attention mechanism, enhancing the representation ability of areas with significant curvature changes.

[0087] In terms of network architecture design, a cascaded attention fusion strategy is used to achieve deep fusion of multi-source data. First, each modality feature is projected to a unified feature space through a 1x1 convolution. Then, a cross-attention mechanism is used to establish semantic associations between modalities, with point cloud features as Query and visible light features as Key. The fusion weight is calculated through dot product attention.

[0088] In the three-dimensional modeling stage, the network converts the fused features into voxel representation through a 3D convolution encoder-decoder, and combines a physics simulation engine to predict the thermal conductivity, stress distribution, and other physical properties of materials. Finally, a three-dimensional model with geometric accuracy and physical realism is output.

[0089] In this embodiment, the three-dimensional model of the transmission line is evaluated and detected in step S5, and local super-resolution reconstruction is started for the area with defects to realize dynamic updating of the three-dimensional model, including:

[0090] The three-dimensional model of the transmission line is voxelized (0.5 cm precision), multi-scale features are fused through a three-dimensional attention mechanism, real-time recognition of wire breakage (detection rate 96.2%) and insulator damage (F1-score 95.7%) is achieved, and the detection results are fed back to the reconstruction module to trigger local model optimization.

[0091] Specifically, a defect detection module based on a lightweight YOLOv7 was developed. This module employs an innovative 3D attention mechanism for multi-dimensional feature optimization. Its core is to simultaneously enhance the local geometric features and cross-modal channel features of 3D voxel data through a spatial-channel collaborative attention module. This module first uses a 7×7×7 3D convolution kernel to extract spatial attention weights, focusing on the 3D feature regions of defects such as broken conductor strands and damaged insulators. Simultaneously, the channel attention mechanism dynamically adjusts the contribution of multiple sources, such as visible light. A multi-scale feature fusion architecture constructs a four-level pyramid processing pipeline. Through cross-scale feature skip connections and progressive downsampling, deep semantic information is progressively integrated with shallow geometric details. The third stage introduces 3D transposed convolutions for feature map upsampling and concatenation, ensuring effective detection of even tiny defects larger than 2 mm. A feature gating mechanism employs a dual-path architecture for adaptive feature selection. The main path extracts global context through 1×1 convolutions, while the secondary path captures local details using depthwise separable convolutions. The outputs of these two paths are fused using a sigmoid gating function to form discriminative features. This design maintains computational efficiency while making the model more robust to defect features in complex backgrounds. Cross-dimensional information fusion is achieved by establishing voxel-pixel correspondences. First, the point cloud data is converted into an octree-indexed voxel representation. Then, deformable 3D convolution is used to align the multimodal feature space. Finally, a cross-attention mechanism is used to establish an associative mapping between visible light texture and point cloud geometry, enabling the system to simultaneously detect surface damage and internal structural defects.

[0092] Example 3. The above is a schematic diagram of the method for reconstructing the full 3D attributes of a transmission line according to this embodiment. It should be noted that the technical solution of the system for reconstructing the full 3D attributes of a transmission line and the technical solution of the method for reconstructing the full 3D attributes of a transmission line are based on the same concept. For details not described in detail in the technical solution of the system for reconstructing the full 3D attributes of a transmission line in this embodiment, please refer to the description of the technical solution of the method for reconstructing the full 3D attributes of a transmission line.

[0093] This embodiment further provides a transmission line 3D full attribute reconstruction system, including:

[0094] Data acquisition module, used to acquire visible light data of the transmission line and generate time-aligned raw data sets;

[0095] Low-light image enhancement module, used to perform low-light enhancement based on the original dataset and generate enhanced images and light distribution maps;

[0096] Multimodal data fusion module, used to perform multimodal data registration based on enhanced images and illumination distribution maps to form a multi-source data cube in a unified coordinate system;

[0097] The 3D reconstruction module is used to perform multi-scale feature expression and cross-domain mapping on multi-source data cubes, deeply fuse the processed and mapped features, and construct a 3D model of the transmission line;

[0098] The defect detection module is used to evaluate and detect the three-dimensional model of the transmission line, initiate local super-resolution reconstruction of the defective areas, and realize dynamic updating of the three-dimensional model.

[0099] This embodiment further provides an electronic device applicable to the method for reconstructing all three-dimensional attributes of a transmission line, including:

[0100] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the three-dimensional full-attribute reconstruction method of the transmission line proposed in the above embodiment.

[0101] This embodiment further provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for reconstructing the three-dimensional full attributes of a transmission line as proposed in the above embodiment.

[0102] The storage medium proposed in this embodiment and the method for reconstructing the three-dimensional full attributes of the transmission line proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0103] Example 4 is an embodiment of the present invention, which provides a method for reconstructing all three-dimensional properties of a transmission line. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0104] Step 1: Sensor Collaborative Data Collection

[0105] Equipped with a multi-line rotating LiDAR, the system uses a 905nm laser beam for high-speed scanning. The system processes raw point cloud data in real time, first filtering out noise points caused by atmospheric particles and then normalizing the intensity of key areas such as conductors and insulators. To address the slender nature of transmission lines, an adaptive scanning strategy is employed, automatically increasing the scan line density in the conductor area. An integrated inertial measurement unit also compensates for point cloud distortion caused by the drone's flight attitude.

[0106] Step 2: Low-light image enhancement processing

[0107] 2.1 Lighting component extraction

[0108] A deep convolutional encoder network is constructed, progressively extracting illumination features through five downsampling layers. Each layer includes a 3×3 convolution, batch normalization, and a LeakyReLU activation function, ultimately outputting an illumination distribution map at 1 / 32 resolution. A perceptual loss function is used to maintain the spatial continuity of the illumination field.

[0109] 2.2 Reflectivity Restoration and Reconstruction

[0110] A U-Net network architecture with skip connections was designed. The encoder extracts multi-scale features using four downsampling blocks, each containing two residual units. The decoder gradually restores spatial resolution using bilinear upsampling and fuses the features with the encoded features at the corresponding scale. A gradient difference loss was introduced during network training to effectively preserve image edge details.

[0111] 2.3 Adaptive Noise Suppression

[0112] Implement an improved non-local mean filtering algorithm, setting a 21×21 pixel search window and a 7×7 pixel similarity block, and dynamically adjusting the filter strength based on local noise characteristics. For impulse noise, a two-stage median filter structure is used, with the front stage detecting noise points and the back stage performing adaptive replacement.

[0113] Step 3: Accurate data registration

[0114] Multi-frame point clouds are temporally registered using a matching algorithm based on normal distribution transformations, achieving submillimeter alignment by optimizing pose transformation parameters. Point clouds in the conductor area are specifically processed, using the conductor's cylindrical geometry to constrain the registration process and eliminate positional deviations caused by wind sway.

[0115] Step 4: 3D reconstruction of neural radiation field

[0116] The LiDAR point cloud is converted into an occupancy probability grid, which serves as geometric prior knowledge for the neural network. During network training, the accuracy of the 3D structure is continuously optimized by comparing the distance error between the predicted geometric surface and the actual LiDAR measured points. Anisotropic kernel functions are used for special processing of slender structures such as wires.

[0117] Step 5: Intelligent defect detection and analysis

[0118] Based on high-precision 3D models, key parameters such as conductor sag and windage amplitude are measured. Structural anomalies such as insulator tilt and tower deformation are detected through time series comparison. Point cloud reflection intensity information is used to identify corrosion or oxidation of metal components.

[0119] Step 6: System deployment and optimization

[0120] Key enhancements:

[0121] LiDAR data full process:

[0122] 1. Optimize scanning strategies and motion compensation during acquisition; 2. Leverage wire geometry constraints during registration; 3. Provide accurate geometric priors during reconstruction; 4. Integrate reflection intensity information during detection;

[0123] It should be noted that this example fully demonstrates the complete technology chain from LiDAR data acquisition to application, while maintaining a synergistic relationship with visible light processing. All sensor data is ultimately deeply fused within the neural radiation field framework, outputting a holographic 3D model with geometric accuracy, texture detail, and temperature properties.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for reconstructing all three-dimensional attributes of a transmission line, characterized by: include: Obtain visible light data from transmission lines and generate time-aligned raw datasets; Perform low-light enhancement based on the original dataset to generate enhanced images and light distribution maps; Based on the enhanced image and illumination distribution map, multimodal data registration is performed to form a multi-source data cube in a unified coordinate system; Perform multi-scale feature expression and cross-domain mapping on multi-source data cubes, deeply fuse the processed and mapped features, and construct a three-dimensional model of the transmission line; Evaluate and inspect the 3D model of the transmission line, initiate local super-resolution reconstruction of defective areas, and achieve dynamic updating of the 3D model.

2. A method for reconstructing all three-dimensional properties of a transmission line according to claim 1, characterized in that: The low-light enhancement based on the original data set to generate the enhanced image and the light distribution map includes: A deep neural network is designed by integrating the physical imaging model. Illumination invariance constraints and gradient-guided loss are added to the network design. The consistency of the illumination components of images under different exposures is modeled, and the constraint is that the values ​​of the estimated illumination components under different exposures at corresponding points in space remain consistent. Gradient-guided loss is introduced into the loss function of the deep neural network to ensure that the gradient of the reflection component is consistent with the gradient direction of the original image. The network outputs an illumination distribution map.

3. A method for reconstructing all three-dimensional properties of a transmission line according to claim 2, characterized in that: The performing low-light enhancement based on the original data set to generate the enhanced image and the light distribution map further includes: The designed deep neural network decomposes the illumination and reflection components of the image in the original dataset, uses gradient-guided loss to maintain texture details, and performs adaptive correction to output enhanced images with extended dynamic range.

4. A method for reconstructing all three-dimensional properties of a transmission line according to claim 3, characterized in that: The multimodal data registration based on the enhanced image and the illumination distribution map to form a multi-source data cube in a unified coordinate system includes: Initial alignment of point clouds and images is performed based on the inertial measurement unit pose. Visible light feature radar geometric key points are matched through feature pyramids. Differentiable rendering is used to achieve sub-pixel precision registration, and a multi-source data cube is output in a unified coordinate system.

5. A method for reconstructing all three-dimensional properties of a transmission line according to claim 4, characterized in that: The matching of visible light features and radar geometric key points through feature pyramid includes: For visible light images, information is constructed by combining key point extraction and deep convolution features; For radar data, project the point cloud into a depth image and extract relevant features; A shared embedding space is constructed, and the features of each modality are mapped to the common space through the feature encoding network. The contrast loss is used to optimize the embedding space to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs.

6. A method for reconstructing all three-dimensional properties of a transmission line according to claim 5, characterized in that: The multi-scale feature expression and cross-domain mapping of multi-source data cubes, deep fusion of processed and mapped features, and construction of a three-dimensional model of the transmission line include: Design a 3D reconstruction network that integrates attention mechanisms, using multi-level attention modules to achieve multi-source data fusion and 3D modeling. By synergizing spatial attention, channel attention, and cross-modal attention, a 3D digital model with complete physical properties is constructed. Different attention mechanisms are used to extract features for different modal data: for visible light images, a CNN network with a spatial attention mechanism is used to extract texture features; for radar point clouds, a curvature-based geometric attention mechanism is used to extract geometric features.

7. A method for reconstructing all three-dimensional properties of a transmission line according to claim 6, characterized in that: The evaluation and detection of the three-dimensional model of the transmission line, initiating local super-resolution reconstruction of defective areas, and achieving dynamic updating of the three-dimensional model include: The three-dimensional model of the transmission line is voxelized, and the three-dimensional attention mechanism is used to fuse multi-scale features to achieve real-time identification of defects such as broken conductors and damaged insulators. The defect probability is calculated, and for areas where the defect probability reaches a preset threshold, local super-resolution reconstruction is initiated to achieve dynamic updating of the three-dimensional model.

8. A transmission line three-dimensional full attribute reconstruction system, applying the method according to any one of claims 1 to 7, characterized in that: include: Data acquisition module, used to acquire visible light data of the transmission line and generate time-aligned raw data sets; Low-light image enhancement module, used to perform low-light enhancement based on the original dataset and generate enhanced images and light distribution maps; Multimodal data fusion module, used to perform multimodal data registration based on enhanced images and illumination distribution maps to form a multi-source data cube in a unified coordinate system; The 3D reconstruction module is used to perform multi-scale feature expression and cross-domain mapping on multi-source data cubes, deeply fuse the processed and mapped features, and construct a 3D model of the transmission line; The defect detection module is used to evaluate and detect the three-dimensional model of the transmission line, initiate local super-resolution reconstruction of the defective areas, and realize dynamic updating of the three-dimensional model.

9. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that It stores computer-executable instructions, which implement the steps of the method according to any one of claims 1 to 7 when executed by a processor.

Images