Lossless precise detection and three-dimensional reconstruction system and method for hidden engineering of transformer substation

Abstract

The invention discloses a transformer substation hidden project lossless precise detection and three-dimensional reconstruction system and method, and relates to the technical field of transformer substation project detection, and the system comprises the steps that a multi-source full-period acquisition and transmission module deploys multiple types of equipment, and acquires, encrypts and transmits images and non-contact signal data in three stages; the multi-source heterogeneous fusion module performs targeted preprocessing on the data to generate a standardized data set; the high-precision identification metering module locates the classification component, calculates parameters and outputs a report; the three-dimensional modeling calculation amount correction module constructs theoretical and actual models and adjusts the theoretical and actual models; the model comparison early warning chain storage module compares the models, triggers early warning and archives data; according to the method, a standardized data set is generated through multi-device collaborative acquisition and encrypted transmission in combination with targeted preprocessing and weighted fusion; based on lightweight detection and classification architecture, precise positioning and metering of components are realized, and through BIM and multi-view reconstruction technology and in combination with bidirectional closed-loop correction, model precision is ensured.

Description

Technical Field

[0001] This application relates to the field of substation engineering inspection technology, specifically to a non-destructive precision detection and three-dimensional reconstruction system and method for concealed works in substations. Background Technology

[0002] Concealed works in substations are a core component of substation infrastructure construction, involving the arrangement and installation of various key components. Their construction quality directly affects the long-term safe and stable operation of the substation. Due to the characteristic that components are concealed after construction, subsequent inspection, quality traceability, and maintenance of components are extremely difficult. With the continuous improvement of substation construction standards and safety operation requirements in the power industry, how to achieve accurate acquisition, efficient integration, and visual management of data throughout the entire lifecycle of concealed works has become an important issue that urgently needs to be addressed in the industry's development. Currently, the application of multi-source sensing technology, digital image processing technology, artificial intelligence recognition technology, and building information modeling technology in the engineering field is gradually deepening, providing the possibility of breaking through the technical bottlenecks in concealed works management. The industry's demand for non-destructive testing and 3D reconstruction solutions that can integrate the advantages of multiple technologies and cover the entire engineering process is becoming increasingly urgent.

[0003] Traditional technologies for detecting and managing concealed works in substations have significant limitations, failing to meet the high-quality requirements of modern power engineering. In the data acquisition phase, traditional methods rely heavily on contact-based detection, which is not only complex but also potentially causes unnecessary damage to already constructed concealed components. Furthermore, data acquisition often focuses only on a specific stage of the project, failing to achieve continuous data coverage throughout the entire lifecycle—before construction, before concealment, and after concealment. This results in an incomplete data chain throughout the entire project, lacking sufficient evidence for subsequent quality traceability. In terms of data processing and component identification, traditional technologies lack effective multi-source heterogeneous data fusion mechanisms, resulting in limited image data... Traditional 3D modeling relies heavily on manual experience for component identification, which is not only inefficient but also prone to subjective biases. It is difficult to accurately distinguish the type, material, and specifications of components. In the modeling and quality control stages, traditional 3D modeling is mostly based on theoretical models built from design drawings, which differs significantly from actual construction conditions. The accuracy of engineering quantity calculation is insufficient, and there is a lack of systematic model comparison and deviation early warning mechanisms. Quality hazards are difficult to detect in a timely manner, and data archiving is also relatively scattered, failing to form a complete traceable data system. This seriously restricts the efficiency of quality control and subsequent operation and maintenance of hidden works. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings of existing technologies and provide a non-destructive and accurate detection and 3D reconstruction system and method for concealed works in substations. The system acquires images and non-contact signals through a multi-source full-cycle acquisition and transmission module, and generates a standardized dataset through a multi-source heterogeneous fusion module; a high-precision identification and measurement module locates and classifies components and calculates parameters; a 3D modeling and quantity calculation correction module constructs theoretical and actual construction 3D models and adjusts them through a bidirectional closed-loop correction algorithm; and a model comparison and early warning chain storage module compares the theoretical and actual models, triggers early warnings according to a hierarchical mechanism, and archives the data. This method realizes accurate detection and 3D reconstruction of concealed works.

[0005] To address the aforementioned technical problems, this application provides the following technical solution: Firstly, a non-destructive and precise detection and three-dimensional reconstruction system for concealed works in substations. Multi-source full-cycle data acquisition and transmission module: used to deploy zoom gimbal cameras, 5G dual-camera mobile control spheres, electromagnetic induction detectors and ground radars, to acquire image data and non-contact signal data in three stages: before construction, before concealment and after concealment. The data is processed using lossless compression and transmitted via AES-256 encryption protocol. Multi-source heterogeneous fusion module: used to perform Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing on image data, wavelet transform filtering and mean filtering preprocessing on non-contact signals, and to associate the data of the same component before and after concealment through a weighted fusion algorithm to generate a standardized fusion dataset; High-precision identification and measurement module: Used to locate and classify components using a combination of lightweight target detection model and deep fine-grained classification model, and calculate component length, size, burial depth and deformation, generating a structured analysis report; 3D Modeling Quantity Calculation Correction Module: Used to construct a theoretical 3D model using BIM modeling software, integrate the structured analysis report, generate an initial actual construction 3D model by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology, and adjust the model size and engineering quantity calculation value through a two-way closed-loop correction algorithm; Model Comparison and Early Warning Chain Storage Module: Used to compare the theoretical model with the corrected actual model using the ICP Iterative Closest Point Algorithm, trigger early warnings according to the hierarchical mechanism, and archive the entire process data through the Minio distributed storage system.

[0006] Furthermore, in the multi-source full-cycle acquisition and transmission module, four zoom gimbal cameras are deployed at the highest point of the construction site, fixed on anti-shake brackets with an installation height of no less than 15 meters, covering the 220kV and 110kV site areas; six 5G dual-camera mobile control spheres are deployed in the cable reel laying area and grounding welding points, using both magnetic suction and tripod mounting methods; one electromagnetic induction detector is deployed every 5 meters along the component path, with the density increased to 3 meters at turns and installed close to the ground; the ground-penetrating radar and electromagnetic induction detectors are alternately arranged every 10 meters along the same path, with the antenna perpendicular to the ground and the gap from the ground not exceeding 5 centimeters; before construction, the terrain of the construction area and the initial state of the components are collected; before concealment, panoramic video and detailed images of the components from 8 angles are collected, covering 4 lighting scenarios; after concealment, the component path coordinates, deformation, and burial depth signals are collected; all equipment undergoes coordinate calibration with standard reference objects before acquisition.

[0007] Furthermore, in the multi-source heterogeneous fusion module, the preprocessing operations of Gaussian filtering, adaptive histogram equalization, and Canny edge detection on the image data are as follows: Gaussian filtering uses a 5×5 kernel size, and filters out image noise by calculating the gray-level mean of the target pixel and its 24 surrounding neighboring pixels and weighting the average; adaptive histogram equalization divides the image into 8×8 pixel blocks, statistically analyzes the gray-level distribution within each block, and adjusts the dynamic range of gray-level within the block to balance the overall contrast of the image and avoid local over-brightness or under-brightness; Canny edge detection first calculates the gray-level gradient magnitude and direction of the image, uses a high threshold of 150 and a low threshold of 50 to filter gradient pixels, directly determines pixels with gradient magnitude greater than the high threshold as edges, and determines pixels between the high and low thresholds and connected to edge pixels as edges, and finally extracts the complete contour of the component.

[0008] Furthermore, in the multi-source heterogeneous fusion module, the operation of performing wavelet transform filtering and mean filtering preprocessing on non-contact signals is as follows: For electromagnetic induction signals and ground-penetrating radar signals, mean filtering is first performed. A 3×3 sliding window is used to slide point by point along the signal data sequence, and the arithmetic mean of the amplitude of all signals in each window is calculated to replace the original signal value at the center of the window. Then, layered filtering is performed to decompose the signal into multiple components according to frequency characteristics, filter out the low-frequency components and effective mid-frequency components that reflect the actual state of the component, remove high-frequency clutter components caused by electromagnetic interference and soil inhomogeneity, and reconstruct a pure signal through component recombination to remove random fluctuations and environmental interference components in the signal.

[0009] Furthermore, in the multi-source heterogeneous fusion module, the mathematical expression of the weighted fusion algorithm is: ;in: It is the first The first component The fused feature values ​​of each feature dimension It is the first The first component The original image data values ​​of each feature dimension, It is the signal-to-noise ratio of image data. It is the positional correlation coefficient. This is the position offset compensation value before and after concealment. No. The first component Image data weights for each feature dimension It is the first The first component Non-contact signal weights for each feature dimension It is the first The first component The original non-contact signal values ​​of each feature dimension, It is the first The first component Signal confidence in each feature dimension.

[0010] Furthermore, in the high-precision identification and measurement module, the specific operations for calculating the length, size, burial depth, and deformation of the component are as follows: Length is fused with the multi-view image path extraction results and the non-contact signal coordinate sequence, and accumulated in segments according to the component's orientation to compensate for the influence of path curvature and sag; Size is based on the pixel ratio of the candidate area and calibration with a standard reference object, extracting the component's edge feature points and calculating the actual distance between feature points; Burial depth is analyzed by considering the relationship between the ground-penetrating radar signal propagation time and the medium velocity, and the depth of the component's top is determined by combining the signal reflection intensity; Deformation is compared with the initial state before concealment and the signal data after concealment, calculating the coordinate offset and morphological difference value, and quantifying the degree of lateral and longitudinal deformation.

[0011] Furthermore, the specific steps for constructing a theoretical 3D model using BIM modeling software in the 3D modeling quantity calculation correction module are as follows: import design data, analyze component design dimensions, spatial coordinates, material parameters, and installation specifications, model layer by layer according to component type and area number, label the path direction and cross-sectional dimensions of the cable layer, restore the grid distribution and welding nodes of the grounding layer, clarify the length, width, depth, and support structure of the cable trench layer, bind the design dimensions, material density, quantity calculation unit, and installation standards to each component, perform collision detection to check the component spacing after modeling is completed, conduct specification verification to check the installation angle, connection method, and burial depth requirements, output the FBX format model file after verification, and associate the theoretical quantity calculation parameters of the project settlement to form a complete theoretical model system.

[0012] Furthermore, in the 3D modeling quantity calculation correction module, an initial actual construction 3D model is generated by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology. The specific steps are as follows: First, feature points are extracted from the multi-view images. Key feature points on the component surface are captured based on image texture and edge information. Feature points from different perspectives are matched, and preliminary point cloud data of the component is generated through spatial triangulation. The preliminary point cloud is denoised and registered, redundant and outlier points are removed, and the point cloud density and uniformity are optimized. Then, the feature information in the non-contact signal is analyzed, and the signal features are mapped to the corresponding spatial positions to supplement the geometric information of the hidden areas in the point cloud data. The component type, specifications, and size parameters in the structured analysis report are integrated, and the optimized point cloud data is triangulated to construct a component surface mesh model, bind the physical properties of the component material, restore the component connection relationship and spatial posture, and finally generate an initial actual construction 3D model consistent with the actual state of the construction site.

[0013] Furthermore, in the 3D modeling calculation correction module, the mathematical expression of the bidirectional closed-loop correction algorithm is: ; ;in, It is the first The dimensions of the 3D model after the component was corrected; It is the first The revised quantity calculation value for each component. It is the first The design theoretical model dimensions of each component It is the first The theoretical engineering quantity calculation value of each component. It is the first The deviation rate between the initial model and the theoretical dimensions of each component. It is the first Deviation rate between manual and theoretical quantity calculations for each component It is a two-way correction factor. These are the fusion feature weight coefficients. It is the size correction factor for the i-th component. It is the first A set of measured dimensional data for each component. It is the first Quantity correction factor for each component It is the first The set of influence factors of the measured quantity of each component. It is the first The length class of the component is a fusion feature value. It is the first Volume class fusion feature value of each component.

[0014] Furthermore, in the model comparison and early warning chain storage module, the mathematical expression of the ICP iterative nearest point algorithm is: ,in, It is the first actual model point cloud One point, It is the first point cloud in the theoretical model One point, It is a rotation matrix. It is a translation vector. It is the square of the Euclidean distance; the graded early warning mechanism is as follows: the first-level early warning is triggered when the spatial position deviation is greater than 5 cm, the size deviation is greater than 3%, or the deformation exceeds the allowable value of the project, triggering an audible and visual alarm and freezing the acceptance process; the second-level early warning is triggered when the spatial position deviation is 3 to 5 cm, the size deviation is 2% to 3%, or the deformation is 80% to 100% of the allowable value of the project, triggering a platform pop-up and SMS reminder and requiring rectification within 48 hours; the third-level early warning is triggered when the spatial position deviation is 1 to 3 cm or the size deviation is 1% to 2%, only triggering a platform early warning and recording the deviation information.

[0015] On the other hand, the method for non-destructive and precise detection and three-dimensional reconstruction of concealed works in substations involves the following steps: Multi-source full-cycle data acquisition and transmission: Deploy zoom gimbal camera, 5G dual-camera mobile control ball, electromagnetic induction detector and ground radar to collect image data and non-contact signal data in three stages: before construction, before concealment and after concealment. After lossless compression, the data is transmitted through AES-256 encryption protocol. Multi-source heterogeneous data fusion processing: Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing are performed on image data; wavelet transform filtering and mean filtering preprocessing are performed on non-contact signals; and a weighted fusion algorithm is used to associate the data of the same component before and after concealment to generate a standardized fusion dataset. High-precision component identification and measurement: A combination of a lightweight target detection model and a deep fine-grained classification model is used to locate and classify components, calculate component length, size, burial depth and deformation, and generate a structured analysis report; 3D Modeling and Quantity Calculation Correction: A theoretical 3D model is constructed using BIM modeling software, integrated with a structured analysis report, and an initial actual construction 3D model is generated by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology. The model size and engineering quantity calculation values ​​are adjusted through a two-way closed-loop correction algorithm. Model comparison: Early warning and data archiving. The ICP iterative nearest point algorithm is used to compare the theoretical model with the corrected actual model. Early warnings are triggered according to the hierarchical mechanism, and the entire process data is archived through the Minio distributed storage system.

[0016] This application addresses the problems of traditional contact-based detection methods that damage engineering and break data chains by employing a multi-source, full-cycle data acquisition and transmission module with multiple types of non-contact devices for phased data acquisition and encrypted transmission. A multi-source heterogeneous fusion module performs targeted preprocessing and weighted fusion of data to generate a standardized dataset, overcoming the shortcomings of traditional methods where data is isolated and difficult to coordinate. A high-precision identification and measurement module uses a lightweight detection and fine-grained classification architecture to achieve automatic and accurate component positioning, classification, and parameter measurement, changing the situation of low efficiency and large errors caused by reliance on manual methods. A 3D modeling and quantity calculation correction module constructs a theoretical model and generates an actual model by combining multi-view reconstruction and signal mapping. Through bidirectional closed-loop correction, it ensures that the model and quantity calculation are consistent with reality, correcting the previous problems of modeling being disconnected from working conditions and inaccurate quantity calculations. A model comparison, early warning, and chain storage module uses the ICP algorithm to compare models and issue graded early warnings. Simultaneously, it utilizes distributed storage to archive the entire process data, achieving automatic quantitative control and complete data traceability of quality deviations, solving the problems of delayed hazard detection and scattered, untraceable data in traditional methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] in: Figure 1 A flowchart for a non-destructive and precise detection and 3D reconstruction system for concealed works in substations; Figure 2 A flowchart for a method of non-destructive and precise detection and three-dimensional reconstruction of concealed works in substations; Figure 3 This diagram illustrates the data transmission between modules of a non-destructive precision detection and 3D reconstruction system for concealed works in substations. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Example 1: Application of concealed engineering in substations with mixed voltage levels of 220kV and 110kV.

[0021] Four zoom PTZ cameras were installed at the highest point of the construction site, fixed to anti-shake brackets at a height of 16 meters, to ensure complete coverage of the 220kV and 110kV site areas and avoid missing key construction areas due to insufficient field of view. 5G (5th Generation Mobile Communication Technology) dual-camera mobile surveillance cameras were deployed at six key nodes, including the cable laying area and grounding electrode welding points, using a combination of magnetic attachment and tripod mounting to ensure stable operation of the equipment in complex construction environments. Electromagnetic induction detectors were deployed every 5 meters along the paths of grounding electrodes, cable trenches, and other components, with the density increased to 3 meters at bends and the detectors installed close to the ground to improve the density and accuracy of data collection along the component paths. Ground-penetrating radars were alternately deployed every 10 meters along the same paths as the electromagnetic induction detectors, with the antennas perpendicular to the ground and a ground clearance of 3 centimeters to reduce interference during signal acquisition. Data acquisition was conducted in three phases: Before construction, the terrain and initial state of the components in the construction area were collected to provide a benchmark for subsequent comparisons; before concealment, panoramic video and detailed images of the components from eight angles were acquired, covering four lighting scenarios: sunny days with strong sunlight, cloudy days, early morning, and evening, ensuring complete recording of component information under different lighting conditions; after concealment, the component's path coordinates, deformation, and burial depth signals were collected to capture changes in the component's state after concealment. All equipment underwent coordinate calibration using standard reference objects before acquisition to ensure spatial consistency of the collected data; the acquired data was processed using lossless compression to reduce data transmission volume while preserving complete information; the processed data was securely transmitted using the AES-256 encryption protocol to prevent data leakage or tampering during transmission. Figure 1 As shown.

[0022] The AES-256 encryption protocol is a symmetric encryption algorithm that is widely used to protect the confidentiality of data.

[0023] The acquired image data undergoes Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing. Canny edge detection is a technique that extracts useful structural information from different visual objects while significantly reducing the amount of data to be processed. Gaussian filtering uses a 5×5 kernel size, calculating and weighting the gray-level mean of the target pixel and its 24 neighboring pixels to filter out random noise and improve image detail. Adaptive histogram equalization divides the image into 8×8 pixel blocks, statistically analyzes the gray-level distribution within each block, and adjusts the dynamic range to avoid loss of component information due to local overexposure or underexposure. Canny edge detection first calculates the image gray-level gradient magnitude and direction, using a high threshold of 150 and a low threshold of 50 to filter gradient pixels, ultimately extracting the complete outline of the component and providing clear boundaries for subsequent component localization. Electromagnetic induction signals and ground-penetrating radar signals were preprocessed using mean filtering and wavelet transform filtering. First, a 3×3 sliding window was used to slide point-by-point along the signal data sequence, calculating the arithmetic mean of the signal amplitude within each window to replace the original central signal value, thus smoothing signal fluctuations. Then, the signal was decomposed into multiple components based on frequency characteristics, filtering low-frequency and effective mid-frequency components, and removing high-frequency clutter components caused by electromagnetic interference and soil inhomogeneity. Key information such as component path coordinates, burial depth, and deformation was retained to ensure the signal data accurately reflects the actual state of the component. Finally, a weighted fusion algorithm was used to link the data of the same component before and after concealment, allowing data from different sources to complement each other and generate a standardized fused dataset. The mathematical expression of the weighted fusion algorithm is: ;in: It is the first The first component The fused feature values ​​of each feature dimension It is the first The first component The original image data values ​​of each feature dimension, It is the signal-to-noise ratio of image data. It is the positional correlation coefficient. This is the position offset compensation value before and after concealment. No. The first component Image data weights for each feature dimension It is the first The first component Non-contact signal weights for each feature dimension It is the first The first component The original non-contact signal values ​​of each feature dimension, It is the first The first component The signal confidence level of each feature dimension provides high-quality data support for subsequent high-precision identification and measurement.

[0024] The necessity of the dual denoising strategy of "mean filtering + wavelet transform" lies in balancing efficiency and depth. Mean filtering, as a preprocessing step, can quickly smooth random high-frequency noise in the signal, providing a preliminary stable data base for subsequent analysis. Wavelet transform performs deep frequency domain decomposition, effectively separating and eliminating structural high-frequency clutter and low-frequency drift caused by electromagnetic pulses, abrupt changes in soil medium, etc., while retaining effective frequency bands that reflect the true state of the components. This layered processing approach overcomes the limitations of single filtering methods, significantly improving the signal-to-noise ratio and feature fidelity in complex field environments, laying a reliable foundation for subsequent accurate feature extraction and mapping.

[0025] The position offset compensation value is calculated by comparing the difference in three-dimensional spatial coordinates of the same component before (baseline state) and after concealment. Specifically, based on the coordinate data output by the component identification and measurement module, the Euclidean distance offset of the key feature points of the component before and after construction is calculated, and the normalized value is used as a compensation item to correct the spatial position changes caused by construction processes such as backfilling and settlement.

[0026] The value of the position correlation coefficient γ is primarily determined by the spatial registration accuracy during multi-source data acquisition. After the equipment completes coordinate calibration using a standard reference, the system evaluates the spatial alignment residuals of the image data and non-contact signal data in a unified coordinate system. If the alignment accuracy is high and the spatial consistency is good, a higher value for γ (e.g., close to 1) is assigned to fully utilize the corrective effect of position offset compensation. If complex field conditions lead to certain registration errors, the value of γ is appropriately reduced (e.g., 0.3 to 0.7) to weaken the impact of potential registration errors on the fusion results and ensure the robustness of data fusion.

[0027] A lightweight object detection model combined with a deep fine-grained classification model is employed for component processing. The feature extraction network contains 32 convolutional layers. The first 16 layers use 8 groups of convolutions to reduce computation while ensuring efficient feature extraction. The last 16 layers have 8 residual blocks, which preserve gradients through cross-layer connections, avoiding gradient vanishing during deep network training and ensuring effective extraction of deep features. The object detection model pre-sets 18 anchor boxes in 6 scales and 3 ratios to adapt to components of different sizes. Non-maximum suppression is used to filter candidate regions and remove duplicate detection results. Bounding box regression is optimized using smoothL1 loss to improve the accuracy of component location. The deep fine-grained classification model integrates channel attention and spatial attention modules to focus on key component features and reduce irrelevant background interference. Candidate region features are first reduced to 512 dimensions to reduce computational complexity, and then global average pooling is used to extract global features. The component feature library contains feature vectors of more than 1000 categories of power components. Similarity matching uses cosine distance calculation to improve the accuracy of component type, material, and specification parameter classification. After the input image is adjusted to 640×640 pixels, it is processed by the aforementioned model architecture to locate the component's position and classify its attributes. Only recognition results with a confidence level of over 95% are retained to ensure the reliability of the recognition results. Simultaneously, relevant component parameters are calculated: length is calculated by fusing multi-view image path extraction results with non-contact signal coordinate sequences, accumulating segments according to the component's orientation, and compensating for the effects of path curvature and sag, making the length calculation more realistic; dimensions are calculated based on the pixel ratio of candidate regions and standard reference objects, extracting component edge feature points to calculate the actual distance, ensuring accurate dimension data; the depth is determined by analyzing the relationship between the ground-penetrating radar signal propagation time and medium velocity, combined with signal reflection intensity, avoiding errors caused by a single data source; deformation is quantified by comparing the initial state before concealment with signal data after concealment, fully reflecting the morphological changes of the component after concealment, and finally generating a structured analysis report that clearly presents the component identification and measurement results.

[0028] In this system, the optimal selection of core parameters such as the number of network layers and training cycle is primarily based on benchmark testing and convergence analysis conducted on image datasets of concealed engineering components in typical substations. The 32-layer convolutional neural network depth was determined as a balance point after comprehensively considering the complexity of substation component features and the deployment efficiency of the embedded platform. This depth is sufficient to extract multi-level features from edges and textures to component combinations, while the introduction of grouped convolutions and residual connections suppresses gradient vanishing and computational redundancy caused by excessively deep networks. The training cycle is set to 1000 epochs, a reasonable upper limit set to ensure sufficient convergence and avoid underfitting, based on the observation that the model's loss function and accuracy metrics on the validation set typically enter a stable plateau after 600-800 epochs. Validation is performed every 20 epochs to identify overfitting trends early in training and dynamically adjust hyperparameters such as the learning rate. This parameter combination achieved the optimal accuracy-efficiency balance in historical engineering datasets and was therefore established as the system's default configuration.

[0029] The lightweight target detection model refers to a target detection neural network with optimized structure and significantly fewer parameters than conventional detection models. It aims to maintain high positioning accuracy while meeting the real-time requirements of edge computing devices in engineering sites. This lightweight target detection model can utilize architectures such as YOLOv5s, SSD-MobileNetV3, or EfficientDet-D0, and achieves real-time component localization on embedded devices through depthwise separable convolution, lightweight feature pyramid design, and INT8 quantization technology.

[0030] The deep fine-grained classification model refers to a deep learning model capable of accurately distinguishing highly similar subclasses with subtle differences within the same broad category. Its core challenge lies in the small inter-class differences and large intra-class differences, requiring the use of attention mechanisms, local feature alignment, and higher-order feature extraction to enhance the model's ability to identify detailed features. Deep fine-grained classification models can employ ResNet50+CBAM attention mechanisms, Vision Transformer, or Bilinear CNN structures, achieving accurate classification of components with subtle differences, such as grounding electrode type, cable specifications, and weld joints, through local feature alignment and higher-order feature fusion.

[0031] SmoothL1 loss optimization is model optimization based on the SmoothL1 loss function. The SmoothL1 loss function is a commonly used loss function for regression tasks in deep learning.

[0032] By importing substation concealed works design data into BIM modeling software, the design dimensions, spatial coordinates, material parameters, and installation specifications of components are analyzed. Modeling is then performed layer by layer according to component type and area number. The cable layer is labeled with path directions and cross-sectional dimensions; the grounding electrode layer restores the grid distribution and welding nodes; and the cable trench layer clearly defines the length, width, depth, and support structure, ensuring a clear presentation of design information for each type of component. Each component is bound with design dimensions, material density, unit of measurement, and installation standards, providing foundational data for subsequent quantity calculations. After modeling, collision detection is performed to check component spacing, avoiding spatial conflicts present in the design phase. Standard verification is conducted to check installation angles, connection methods, and burial depth requirements, ensuring the theoretical model conforms to engineering specifications. After successful verification, an FBX format (cross-platform 3D data exchange format) model file is output, linked to theoretical quantity calculation parameters for engineering settlement to form a theoretical 3D model, providing a standard basis for subsequent comparison with actual models. Subsequently, an initial actual construction 3D model was generated by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology: First, feature points were extracted from the multi-view images, and key feature points on the component surface were captured based on image texture and edge information. After matching feature points from different viewpoints, preliminary point cloud data was generated through spatial triangulation. After denoising and registration, the point cloud density and uniformity were optimized so that the point cloud could accurately reflect the surface morphology of the component. Then, feature information such as path coordinates, burial depth, and deformation in the non-contact signal was analyzed and mapped to the corresponding spatial location to supplement the geometric information of the hidden area, solving the problem that the image could not penetrate the hidden layer. The component parameters in the structured analysis report were integrated, and the optimized point cloud data was triangulated to construct a surface mesh model and bind the physical properties of the material, restoring the component connection relationship and spatial posture to ensure that the initial actual model closely resembled the actual state of the construction site. Finally, a two-way closed-loop correction algorithm was used to adjust the model size and engineering quantity calculation value. The mathematical expression of the two-way closed-loop correction algorithm is as follows: ; ;in, It is the first The dimensions of the 3D model after the component was corrected; It is the first The revised quantity calculation value for each component. It is the first The design theoretical model dimensions of each component It is the first The theoretical engineering quantity calculation value of each component. It is the first The deviation rate between the initial model and the theoretical dimensions of each component. It is the first Deviation rate between manual and theoretical quantity calculations for each component It is a two-way correction factor. These are the fusion feature weight coefficients. It is the size correction factor for the i-th component. It is the first A set of measured dimensional data for each component. It is the first Quantity correction factor for each component It is the first The set of influence factors of the measured quantity of each component. It is the first The length class of the component is a fusion feature value. It is the first The volume-class fusion feature value of each component makes the dimensional accuracy and quantity calculation data of the corrected actual construction 3D model more consistent with the actual construction situation.

[0033] Optionally, the DR function can be in the following form: DR(D_i) = 1 + α * tanh( (avg(D_i) - M_{0,i}) / M_{0,i} ) Where avg(D_i) is the weighted average of the measured dimensions of the i-th component, D_i, and M_{0,i} is its theoretical design dimension. α is the size sensitivity coefficient (usually taken as 0.1~0.3), and tanh is the hyperbolic tangent function, used to map the deviation to a smooth correction interval. This function can output a coefficient that smoothly varies around 1 based on the positive or negative deviation of the measured dimension from the theoretical dimension, achieving a gradual, non-linear correction of the model dimensions.

[0034] Optionally, the HR function can be in the following form: HR(H_i) = ∏_{k=1}^{n} (1 + θ_k * h_{i,k}) Where h_{i,k} represents the k-th measured quantity influencing factor (e.g., material loss rate, construction margin, etc.) of the i-th component, forming the set H_i. θ_k is the weight coefficient of the corresponding influencing factor (determined through regression of historical data). This function combines the correction effects of multiple influencing factors on the theoretical quantity in a multiplicative form. When there is a positive influencing factor (such as an increase in actual loss), it outputs a coefficient greater than 1; otherwise, it outputs a coefficient less than 1, thereby dynamically correcting the engineering quantity calculation value.

[0035] The bidirectional correction coefficient λ and the fusion feature weight coefficient β are key adjustable parameters in this algorithm used to balance the contributions of theoretical bias correction and measured feature fusion. Their typical values ​​and ranges are as follows: λ (two-way correction coefficient): Its value is between 0 and 1, and it is calibrated by analyzing the correlation data between the model and the quantity calculation deviation of historical qualified projects. The recommended value range is 0.1 to 0.3. This range can ensure a smooth and stable correction process and avoid over-adjustment.

[0036] β (fusion feature weight coefficient): Its value is dynamically adjusted mainly based on the comprehensive confidence of the multi-source fusion data, and the recommended value range is 0.2 to 0.8. When the image and non-contact signal are clear and the signal-to-noise ratio is high, it approaches the upper limit to enhance the effect of measured features; when the on-site interference is large and the data quality is average, it approaches the lower limit to maintain the robustness of the correction process.

[0037] The length-type fusion feature value is generated by a weighted fusion algorithm: the component path length feature extracted from the image before concealment and the length feature parsed from the non-contact signal (such as electromagnetic induction coordinate sequence) after concealment are correlated, weights are assigned according to their respective confidence-to-noise ratio or confidence level and the weighted sum is calculated, and the construction process analysis (such as bending compensation amount) is fused at the same time, and finally a fusion value representing the actual length of the component is output.

[0038] In practice, a set of initial values ​​for λ and β suitable for the current working conditions can be determined through testing and calibration in typical engineering scenarios in the early stage, and β can be fine-tuned based on the data quality assessment results during system operation.

[0039] The ICP iterative closest point algorithm is used to compare the theoretical 3D model with the corrected actual construction 3D model, accurately calculating the spatial differences between the two. The mathematical expression of the ICP iterative closest point algorithm is as follows: Where, min is used to calculate the minimum value. It is the first actual model point cloud One point, It is the first point cloud in the theoretical model One point, It is a rotation matrix. It is a translation vector. It is the square of the Euclidean distance; providing accurate basis for subsequent early warnings; triggering corresponding early warnings according to the graded early warning mechanism: if the detected spatial position deviation is greater than 5 cm, the dimensional deviation is greater than 3%, or the deformation exceeds the allowable value of the project, a first-level early warning is triggered, audible and visual alarms are activated and the acceptance process is frozen to promptly stop the unqualified project from continuing; if the spatial position deviation is 3 to 5 cm, the dimensional deviation is 2% to 3%, or the deformation is 80% to 100% of the allowable value of the project, a second-level early warning is triggered, a pop-up window appears on the platform and an SMS reminder is sent, requiring rectification to be completed within 48 hours to avoid the deviation from expanding; if the spatial position deviation is 1 to 3 cm, or the dimensional deviation is 1% to 2%, only a third-level early warning is triggered and the deviation information is recorded for subsequent project traceability. At the same time, all relevant data, including data collected before construction, before concealment, and after concealment, are archived through the Minio distributed storage system, forming a complete data archive to facilitate subsequent project acceptance, maintenance, and traceability. Minio is a high-performance distributed object storage system based on open-source technology, focusing on solving the needs of rapidly growing unstructured data, and its core design is compatible with the Amazon S3 protocol.

[0040] In summary, in this embodiment, when the non-destructive precision detection and 3D reconstruction system for concealed works in a 220kV and 110kV mixed substation is applied, the multi-source full-cycle data acquisition module ensures data comprehensiveness and security through scientific equipment deployment and phased data acquisition; the multi-source heterogeneous fusion module generates a high-quality dataset through preprocessing and weighted fusion; the high-precision identification and measurement module achieves accurate component identification and measurement based on the combined model and strict screening; the 3D modeling and quantity calculation correction module constructs a realistic 3D model through BIM modeling and bidirectional closed-loop correction; the model comparison, early warning, and chain storage module timely manages deviations through ICP algorithm and hierarchical early warning; and Minio storage enables data traceability. Figure 3 As shown. The entire system covers the entire process of detection, processing, identification, modeling, and early warning, providing strong support for the quality control of concealed works in substations. The ICP algorithm is an algorithm based on data registration and utilizing the nearest point search method to solve problems based on freeform surfaces.

[0041] Example 2: Implementation process of non-destructive testing and 3D reconstruction of concealed works in substation cable trenches.

[0042] Clearly define the construction scope, design parameters, and key detection areas for the concealed cable trench project to avoid directional deviations in subsequent work; complete the debugging and calibration of equipment such as zoom pan-tilt dome cameras, 5G dual-camera mobile surveillance dome cameras, electromagnetic induction detectors, and ground-penetrating radar to ensure that the equipment is in normal working condition and to reduce the impact of equipment errors on data acquisition; deploy equipment according to system requirements: four zoom pan-tilt dome cameras are installed on anti-shake brackets at the highest point of the construction site, at a height of 15.5 meters, covering the cable trench construction area and surrounding related areas to ensure a complete field of view; six sets of 5G dual-camera mobile surveillance dome cameras... At key locations such as cable trench excavation nodes and support structure installation points, a dual installation method using magnetic attachment and tripods is employed to enhance equipment stability during construction and prevent data acquisition deviations caused by equipment displacement. Electromagnetic induction detectors are deployed every 5 meters along the planned cable trench path, with increased spacing of 3 meters at turns and corners, and fixed close to the ground to increase the density of data acquisition along the path and ensure no component information is missed at turns. Ground-penetrating radar and electromagnetic induction detectors are alternately arranged every 10 meters along the same path, ensuring that the antennas are perpendicular to the ground and the gap between them and the ground does not exceed 5 centimeters, reducing external interference during signal acquisition. After all equipment is deployed, coordinate calibration is performed using standard reference points to ensure spatial consistency of data collected by different devices, laying the foundation for subsequent data fusion and modeling. Figure 2 As shown.

[0043] Data collection was conducted in three phases: before construction, before concealment, and after concealment. Before construction, topographical data, construction materials, and initial state information of components in the cable trench construction area were collected using various devices, providing an initial benchmark for data comparison after concealment. Before concealment, panoramic video of the cable trench construction process and detailed images of the trench body, support structure, and embedded components from eight angles were collected, ensuring coverage of four different lighting scenarios and fully recording the appearance and positional relationships of components after installation, avoiding information loss due to lighting differences. After concealment, component path coordinates, structural deformation, and burial depth-related signal data were collected after backfilling the cable trench, capturing the actual state changes of components after the concealment work was completed. During the collection process, lossless compression was used on image and non-contact signal data to reduce data storage space and transmission bandwidth usage while fully preserving key information. Data was transmitted to the system data processing center using the AES-256 encryption protocol to prevent theft or tampering during transmission, ensuring data security and integrity.

[0044] For image data transmitted to the data processing center, Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing are performed sequentially. Gaussian filtering removes noise interference from the image, making image details clearer; adaptive histogram equalization adjusts the dynamic range of image grayscale to avoid local overexposure or underexposure affecting component observation; Canny edge detection extracts the complete outline of the component, providing clear boundaries for subsequent component positioning. For non-contact signal data, mean filtering is first performed through a 3×3 sliding window to smooth random fluctuations in the signal; then wavelet transform layered filtering is performed to filter out effective signal components that reflect the actual state of the component, eliminate high-frequency noise caused by environmental interference, and retain key information such as component path coordinates, burial depth, and deformation. A weighted fusion algorithm is used to correlate and fuse the preprocessed image data and non-contact signal data of the same component before and after concealment, integrating the advantages of data from different sources, compensating for the limitations of single data, and generating a standardized fusion dataset, providing high-quality and reliable data support for subsequent high-precision identification and measurement.

[0045] A lightweight object detection model combined with a deep fine-grained classification model is used to process the standardized fusion dataset. The mathematical expression of the weighted fusion algorithm is as follows: After adjusting the input image to 640×640 pixels, features are extracted using a 32-layer convolutional neural network. The first 16 layers employ grouped convolutions to reduce computation and improve feature extraction efficiency; the last 16 layers retain deep features through residual connections, avoiding gradient vanishing in deep networks and ensuring comprehensive extraction of component features. Multi-scale feature maps are generated to adapt to the feature requirements of components of different sizes. The object detection model traverses the feature map using a sliding window, determining component positions through anchor box matching and bounding box regression, and outputting candidate regions containing component coordinates and confidence scores to improve component localization accuracy. A deep fine-grained classification model enhances the features of candidate regions, focusing on key parts of the component through an attention mechanism to reduce background interference. The extracted feature vectors are matched with a pre-set component feature library using cosine distance similarity to accurately distinguish component type, material, and specifications. During the model training phase, ImageNet pre-trained weights were used for initialization to shorten the training cycle and improve training performance. Transfer learning was employed to initialize weights, and training samples were expanded through random rotation, scaling, flipping, and color perturbation to enhance the model's generalization ability. The Adam optimizer was used with an initial learning rate of 0.001, decaying by a factor of 10 every 200 epochs. Iterative optimization was performed using mini-batch gradient descent, with a training cycle of 1000 epochs. Model performance was validated every 20 epochs, and recognition results with a confidence level below 95% were removed to ensure reliability. Simultaneously, the component's length, dimensions, burial depth, and deformation were calculated. Length was compensated for by multi-source data to mitigate path effects; dimensions were calibrated based on standard references; burial depth analyzed signal propagation relationships; and deformation was compared before and after. All recognition and measurement results were integrated to generate a structured analysis report, clearly presenting the component's relevant information.

[0046] First, a mature convolutional neural network model (e.g., ResNet50 or YOLOv5s) pre-trained on a large general image dataset (such as ImageNet) is selected as the base model, and its weights are used as initialization parameters. For the identification and classification task of hidden engineering components in substations, the parameters of the front and middle convolutional layers of this base model (used to extract general image features) are retained, while the top fully connected layer or detector head is removed and replaced, and the output layer is reconstructed to match the number of target component categories and the detection task. In the initial training phase, the parameters of the pre-trained convolutional layers are frozen, and only the newly added top-level network is trained, allowing the model to quickly adapt to the feature distribution of the new task. Subsequently, all or part of the deep networks are unfrozen, and a small learning rate (e.g., 1 / 10 of the initial learning rate) is used to perform end-to-end fine-tuning training using substation component image data collected and labeled by this system. This allows for efficient adaptation to fine-grained features in the power engineering field while inheriting general visual representation capabilities, achieving rapid convergence of model training and a significant improvement in recognition accuracy.

[0047] Import cable trench design data into BIM modeling software, analyze component design dimensions, spatial coordinates, material parameters, and installation specifications, and model layer by layer according to component type, clearly presenting different components such as cables, grounding electrodes, and cable trenches. After modeling, perform collision detection to check component spacing and eliminate spatial conflicts in the design phase. Conduct specification verification to check installation angles, connection methods, and burial depth requirements to ensure that the theoretical model conforms to engineering specifications. After verification, output the theoretical 3D model in FBX format, link it with engineering settlement theoretical quantity calculation parameters, and form a complete theoretical model system to provide a standard reference for the construction of the actual model. Integrate component parameters from the structured analysis report, and use a combination of multi-view image 3D reconstruction technology and non-contact signal feature mapping technology to generate the initial actual construction 3D model: first, process multi-view images to generate point cloud data and perform noise reduction and registration to optimize point cloud quality; then, map non-contact signal features to spatial positions to supplement geometric information of hidden areas; finally, integrate parameters to construct a surface mesh model and bind material properties to restore the component connection relationships and spatial posture. A two-way closed-loop correction algorithm is used to adjust the size and engineering quantity calculation values ​​of the initial actual model. The mathematical expression of the two-way closed-loop correction algorithm is: ; This will allow the corrected model dimensions and calculation data to better reflect actual construction conditions, thereby improving the model's practicality.

[0048] The ICP iterative closest point algorithm is used to accurately compare the theoretical 3D model with the corrected actual 3D model, calculating the differences between the two in spatial position, size, and shape. The mathematical expression of the ICP iterative closest point algorithm is as follows: This provides accurate data for early warning judgment. Based on the comparison results, early warnings are triggered according to a tiered mechanism: Level 1 warnings activate audible and visual alarms and freeze the acceptance process, promptly preventing substandard projects from proceeding; Level 2 warnings send platform pop-ups and SMS reminders, requiring rectification within 48 hours to prevent deviations from escalating; Level 3 warnings only record deviation information for easy subsequent traceability. The entire process data is archived and stored using the Minio distributed storage system, forming a complete and traceable engineering data archive, facilitating subsequent project acceptance, maintenance inquiries, and problem tracing, providing data support for the full lifecycle management of cable trench concealed works.

[0049] In summary, this embodiment presents a method for concealed cable trench engineering in substations. Starting with construction preparation and equipment deployment, it lays the foundation for data acquisition through equipment debugging, calibration, and scientific layout. Multi-stage acquisition and encrypted transmission ensure data integrity and security. Multi-source data preprocessing and weighted fusion improve data quality. Component identification and measurement rely on combined models and rigorous training to ensure reliable results. 3D modeling and quantity calculation correction utilizes BIM and two-way closed-loop correction to generate accurate actual models. Model comparison, early warning, and archiving enable deviation control and data traceability. The entire method is tightly integrated, covering the entire process of concealed cable trench engineering, providing a feasible implementation path for non-destructive testing and 3D reconstruction, and contributing to engineering quality control and data management.

[0050] Compared with existing technologies, the non-destructive precision detection and three-dimensional reconstruction system and method for concealed works in substations has the following advantages: I. This application constructs a full-cycle data acquisition and transmission system through multi-device collaboration, covering all stages of construction for non-destructive data acquisition. Combined with encrypted transmission protocols, it ensures data security and integrity. Targeted preprocessing of images and non-contact signals removes noise and environmental interference. Then, a weighted fusion algorithm is used to correlate data from different dimensions, generating a standardized fusion dataset. This breaks down heterogeneous data barriers. Relying on a lightweight detection and deep, fine-grained classification architecture, it focuses on key component features to achieve precise positioning and classification. Simultaneously, it integrates multi-dimensional data to measure core parameters such as length and dimensions, outputting a structured analysis report. This effectively solves problems such as incomplete data acquisition, ambiguous identification and classification, and insufficient measurement accuracy in concealed works. It provides high-quality data support for subsequent modeling, improves the scientific rigor and reliability of engineering inspection, and reduces quality risks caused by missing or incorrect data in concealed works.

[0051] II. This application constructs a complete theoretical 3D model using BIM (Building Information Modeling) technology, integrates multi-view image reconstruction and non-contact signal feature mapping technology to generate an initial actual construction model, and dynamically adjusts the model size and engineering quantity calculation values ​​using a two-way closed-loop correction algorithm to achieve a precise match between theory and reality. It uses an iterative nearest-point algorithm to conduct model comparisons and establishes a hierarchical early warning mechanism to promptly trigger corresponding handling procedures for different degrees of deviation, thus avoiding potential engineering quality hazards in advance. Simultaneously, it utilizes a distributed storage system to archive the entire process data, forming a traceable data chain. This not only solves problems such as the disconnect between concealed works modeling and actual working conditions, large quantity calculation deviations, and difficulty in timely detection of hidden dangers, but also provides comprehensive data support for project acceptance and subsequent operation and maintenance, ensuring the construction quality and long-term stable operation of substation concealed works and improving the level of precision in project management.

[0052] Other advantages, objectives and features of this application will be set forth in part in the description which follows, and in part will be obvious to those skilled in the art based on an examination of the following, or may be taught from practice of this application.

[0053] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0054] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0055] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0056] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A non-destructive and precise detection and three-dimensional reconstruction system for concealed works in substations, characterized in that, The system includes: Multi-source full-cycle data acquisition and transmission module: used to deploy zoom gimbal cameras, 5G dual-camera mobile control spheres, electromagnetic induction detectors and ground radars, to acquire image data and non-contact signal data in three stages: before construction, before concealment and after concealment. The data is processed using lossless compression and transmitted via AES-256 encryption protocol. Multi-source heterogeneous fusion module: used to perform Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing on image data, wavelet transform filtering and mean filtering preprocessing on non-contact signals, and to associate the data of the same component before and after concealment through a weighted fusion algorithm to generate a standardized fusion dataset; High-precision identification and measurement module: Used to locate and classify components using a combination of lightweight target detection model and deep fine-grained classification model, and calculate component length, size, burial depth and deformation, generating a structured analysis report; 3D Modeling Quantity Calculation Correction Module: Used to construct a theoretical 3D model using BIM modeling software, integrate the structured analysis report, generate an initial actual construction 3D model by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology, and adjust the model size and engineering quantity calculation value through a two-way closed-loop correction algorithm; Model Comparison and Early Warning Chain Storage Module: Used to compare the theoretical model with the corrected actual model using the ICP Iterative Closest Point Algorithm, trigger early warnings according to the hierarchical mechanism, and archive the entire process data through the Minio distributed storage system.

2. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the multi-source full-cycle acquisition and transmission module, four zoom gimbal cameras are deployed at the highest point of the construction site, fixed on anti-shake brackets with an installation height of no less than 15 meters, covering the 220kV and 110kV site areas; six 5G dual-camera mobile control spheres are deployed in the cable reel laying area and grounding welding points, using both magnetic suction and tripod mounting methods; one electromagnetic induction detector is deployed every 5 meters along the component path, with the density increased to 3 meters at turns and installed close to the ground; the ground-penetrating radar and electromagnetic induction detectors are alternately arranged every 10 meters along the same path, with the antenna perpendicular to the ground and the gap from the ground not exceeding 5 centimeters; before construction, the terrain of the construction area and the initial state of the components are collected; before concealment, panoramic video and detailed images of the components from 8 angles are collected, covering 4 lighting scenarios; after concealment, the component path coordinates, deformation, and burial depth signals are collected; all equipment undergoes coordinate calibration with standard reference objects before acquisition.

3. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the multi-source heterogeneous fusion module, the preprocessing operations of Gaussian filtering, adaptive histogram equalization, and Canny edge detection on the image data are as follows: Gaussian filtering adopts a 5×5 kernel size, and the image noise is filtered out by calculating the gray value mean of the target pixel and its 24 surrounding neighboring pixels and weighting the average. Adaptive histogram equalization divides the image into 8×8 pixel blocks, statistically analyzes the gray-level distribution within each block, and adjusts the dynamic range of gray-levels within the block to achieve overall image contrast balance and avoid local over-brightness or under-brightness. Canny edge detection first calculates the gray-level gradient magnitude and direction of the image, uses a high threshold of 150 and a low threshold of 50 to filter gradient pixels, directly identifies pixels with gradient magnitudes greater than the high threshold as edges, and identifies pixels between the high and low thresholds that are connected to edge pixels as edges, ultimately extracting the complete contour of the component.

4. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the multi-source heterogeneous fusion module, the operation of wavelet transform filtering and mean filtering preprocessing for non-contact signals is as follows: For electromagnetic induction signals and ground-penetrating radar signals, mean filtering is first performed. A 3×3 sliding window is used to slide point by point along the signal data sequence, and the arithmetic mean of the amplitude of all signals in each window is calculated to replace the original signal value at the center of the window. Then, layered filtering is performed to decompose the signal into multiple components according to frequency characteristics, filter out the low-frequency components and effective mid-frequency components that reflect the actual state of the component, remove high-frequency clutter components caused by electromagnetic interference and soil inhomogeneity, and reconstruct a pure signal through component recombination to remove random fluctuations and environmental interference components in the signal.

5. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the high-precision identification and measurement module, the specific steps for locating and classifying components using a combination architecture of a lightweight target detection model and a deep fine-grained classification model are as follows: After the input image is resized to 640×640 pixels, features are extracted through a 32-layer convolutional neural network. The first 16 layers use grouped convolution to reduce the computational load, and the last 16 layers retain deep features through residual connections to generate multi-scale feature maps. The object detection model traverses the image using a feature map sliding window, determines the component location through anchor box matching and bounding box regression, and outputs candidate regions containing component coordinates and confidence scores. The deep fine-grained classification model enhances the features of the candidate regions, focuses on key parts of the component through an attention mechanism, and performs similarity matching between the extracted feature vectors and a pre-set component feature library to distinguish component type, material, and specifications. During the model training phase, transfer learning is used to initialize weights, and training samples are expanded through random rotation, scaling, flipping, and color perturbation. Mini-batch gradient descent is used for iterative optimization, with 1000 training cycles set. Model performance is verified every 20 cycles, and recognition results with confidence scores below 95% are removed.

6. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, The specific steps for constructing a theoretical 3D model using BIM modeling software in the 3D modeling and quantity calculation correction module are as follows: import design data, analyze component design dimensions, spatial coordinates, material parameters, and installation specifications, model layer by layer according to component type and area number, label the path and cross-sectional dimensions of the cable layer, restore the grid distribution and welding nodes of the grounding layer, clarify the length, width, depth, and support structure of the cable trench layer, bind the design dimensions, material density, quantity calculation units, and installation standards to each component, perform collision detection to check the component spacing after modeling is completed, conduct specification verification to check the installation angle, connection method, and burial depth requirements, output the FBX format model file after verification, and associate the theoretical quantity calculation parameters of the project settlement to form a complete theoretical model system.

7. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the 3D modeling and quantity calculation correction module, an initial actual construction 3D model is generated by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology. The specific steps are as follows: First, feature points are extracted from the multi-view images. Key feature points on the surface of the component are captured based on image texture and edge information. Feature points from different perspectives are matched, and preliminary point cloud data of the component is generated through spatial triangulation. The preliminary point cloud is then denoised and registered to remove redundant and outlier points and optimize the point cloud density and uniformity. Next, the feature information in the non-contact signal is analyzed, and the signal features are mapped to the corresponding spatial positions to supplement the geometric information of the hidden areas in the point cloud data. By integrating the component type, specifications, and size parameters from the structured analysis report, the optimized point cloud data is triangulated to construct a surface mesh model of the component, bind the physical properties of the component material, restore the connection relationship and spatial posture of the component, and finally generate an initial actual construction 3D model that is consistent with the actual state of the construction site.

8. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the 3D modeling calculation correction module, the mathematical expression of the bidirectional closed-loop correction algorithm is: ; ;in, It is the first The dimensions of the 3D model after the component was corrected; It is the first The revised quantity calculation value for each component. It is the first The design theoretical model dimensions of each component It is the first The theoretical engineering quantity calculation value of each component. It is the first The deviation rate between the initial model and the theoretical dimensions of each component. It is the first Deviation rate between manual and theoretical quantity calculations for each component It is a two-way correction factor. These are the fusion feature weight coefficients. It is the size correction factor for the i-th component. It is the first A set of measured dimensional data for each component. It is the first Quantity correction factor for each component It is the first The set of influence factors of the measured quantity of each component. It is the first The length class of the component is a fusion feature value. It is the first Volume class fusion feature value of each component.

9. The non-destructive precision detection and three-dimensional reconstruction system for concealed works in substations according to claim 1, characterized in that, In the model comparison and early warning chain storage module, the mathematical expression of the ICP iterative nearest point algorithm is: Where, min is used to calculate the minimum value. It is the first actual model point cloud One point, It is the first point cloud in the theoretical model One point, It is a rotation matrix. It is a translation vector. It is the square of the Euclidean distance; the graded early warning mechanism is as follows: the first-level early warning is triggered when the spatial position deviation is greater than 5 cm, the size deviation is greater than 3%, or the deformation exceeds the allowable value of the project, triggering an audible and visual alarm and freezing the acceptance process; the second-level early warning is triggered when the spatial position deviation is 3 to 5 cm, the size deviation is 2% to 3%, or the deformation is 80% to 100% of the allowable value of the project, triggering a platform pop-up and SMS reminder and requiring rectification within 48 hours; the third-level early warning is triggered when the spatial position deviation is 1 to 3 cm or the size deviation is 1% to 2%, only triggering a platform early warning and recording the deviation information.

10. A method for non-destructive and precise detection and three-dimensional reconstruction of concealed works in substations, the method being applicable to the non-destructive and precise detection and three-dimensional reconstruction system for concealed works in substations as described in any one of claims 1-9, characterized in that, The specific steps of this method are as follows: Multi-source full-cycle data acquisition and transmission: Deploy zoom gimbal camera, 5G dual-camera mobile control ball, electromagnetic induction detector and ground radar to collect image data and non-contact signal data in three stages: before construction, before concealment and after concealment. After lossless compression, the data is transmitted through AES-256 encryption protocol. Multi-source heterogeneous data fusion processing: Gaussian filtering, adaptive histogram equalization, and Canny edge detection preprocessing are performed on image data; wavelet transform filtering and mean filtering preprocessing are performed on non-contact signals; and a weighted fusion algorithm is used to associate the data of the same component before and after concealment to generate a standardized fusion dataset. High-precision component identification and measurement: A combination of a lightweight target detection model and a deep fine-grained classification model is used to locate and classify components, calculate component length, size, burial depth and deformation, and generate a structured analysis report; 3D Modeling and Quantity Calculation Correction: A theoretical 3D model is constructed using BIM modeling software, integrated with a structured analysis report, and an initial actual construction 3D model is generated by combining multi-view image 3D reconstruction technology with non-contact signal feature mapping technology. The model size and engineering quantity calculation values ​​are adjusted through a two-way closed-loop correction algorithm. Model comparison: Early warning and data archiving. The ICP iterative nearest point algorithm is used to compare the theoretical model with the corrected actual model. Early warnings are triggered according to the hierarchical mechanism, and the entire process data is archived through the Minio distributed storage system.

Images