Oil-immersed transformer inspection method and system based on three-dimensional space perception

Abstract

The invention discloses an oil-immersed transformer inspection method and system based on three-dimensional space perception. The method comprises the following steps: S1, obtaining a visible light video stream and synchronous three-dimensional point cloud original data in a transformer; s2, performing frame extraction on the visible light video stream, and performing optimization processing on each frame of image by adopting a multi-scale adaptive image enhancement algorithm to obtain an enhanced video frame; s3, performing registration on the enhanced video frame and three-dimensional point cloud original data, establishing a mapping relation between video frame pixel coordinates and three-dimensional point cloud space coordinates, and generating a dense three-dimensional laser point cloud; s4, inputting the dense three-dimensional laser point cloud into a pre-constructed three-dimensional space understanding model, extracting spatial features, performing feature alignment with a semantic model, and outputting a transformer internal space structure description text and defect position information; and S5, planning an inspection path according to the space structure description text and the defect position information. The method has the advantages of accurate identification of internal defects of the transformer and the like.

Description

Technical Field

[0001] This invention mainly relates to the field of oil-immersed transformer inspection technology, specifically to an oil-immersed transformer inspection method and system based on three-dimensional spatial perception. Background Technology

[0002] As the "core hub of energy transmission" in the power system, oil-immersed transformers undertake the critical functions of voltage conversion and power distribution. Their operational reliability directly determines the safety of the power grid and the stability of energy supply. Currently, oil-immersed transformers are gradually entering a period of high incidence of defects such as "insulation aging and component loosening." Key equipment such as 1000kV converter transformers in ultra-high voltage substations are the "lifeline" for nuclear safety power supply and cross-regional energy transmission. Once a fault is caused by internal defects (such as short circuits between winding turns, carbon trace discharge in insulation paper, and burnt contact of tap changer), it will lead to regional power outages and industrial shutdowns. The economic loss of a single fault can reach tens of millions of yuan.

[0003] Currently, internal inspections of oil-immersed transformers mainly rely on the traditional method of "manual inspection by lifting the cover + indirect monitoring." This method can no longer meet the "efficient, accurate, and safe" operation and maintenance requirements of the new power system, and has four major pain points: 1. Manual inspections have poor safety and high secondary risks. Traditional manual inspections require a complex process including "transformer shutdown → insulating oil venting → cover hoisting → personnel entry → maintenance → oil filling → oil circulation → partial discharge withstand voltage test". Maintenance personnel must wear protective clothing to enter the enclosed oil tank, facing risks to personal safety, secondary contamination of equipment, and power outage losses.

[0004] 2. Manual inspection is costly and inefficient. Manual inspection requires a significant investment of human and material resources.

[0005] 3. Indirect monitoring technologies have limited accuracy and are difficult to locate faults. Existing indirect monitoring technologies (such as oil chromatography, online partial discharge monitoring, and infrared thermometry) can only provide "fault warnings" but cannot accurately locate fault points, which has significant limitations.

[0006] 4. Existing robot inspection technologies face significant bottlenecks. In recent years, the industry has gradually explored robot inspection technologies (such as floating robots and tracked robots), but two major technical bottlenecks have yet to be overcome: First, poor image perception quality: the internal light intensity of transformers is low (≤5 lux), and the insulating oil medium causes visible light images to exhibit problems such as "high scattering noise, low contrast, and blurred details." Second, weak three-dimensional spatial understanding: existing robots mostly rely on two-dimensional images for defect identification and lack the ability to analyze the three-dimensional spatial structure inside transformers.

[0007] In summary, neither traditional manual inspection nor existing technologies can simultaneously meet the requirements of "safety, economy, accuracy, and autonomy" for internal inspection of oil-immersed transformers. As the power grid develops towards "ultra-high voltage and intelligent" directions, there is an urgent need for a targeted inspection technology. By optimizing image enhancement algorithms and overcoming the bottleneck of three-dimensional spatial understanding, a robotic fish can achieve autonomous, accurate, and efficient internal inspection, replacing manual inspection and providing technical support for the full lifecycle operation and maintenance of oil-immersed transformers. Summary of the Invention

[0008] To address the technical problems existing in the prior art, this invention provides an oil-immersed transformer inspection method and system based on three-dimensional spatial perception for accurate identification of internal defects in transformers.

[0009] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for inspecting oil-immersed transformers based on three-dimensional spatial perception includes the following steps: S1. Obtain the visible light video stream and synchronized 3D point cloud raw data inside the transformer; S2. Frame extraction is performed on the visible light video stream to obtain continuous video frames, and a multi-scale adaptive image enhancement algorithm is used to optimize each frame to obtain enhanced video frames; the multi-scale adaptive image enhancement algorithm includes sequential multi-scale Retinex enhancement and improved bilateral filtering denoising; S3. Register the enhanced video frame with the original three-dimensional point cloud data, and establish a mapping relationship between the pixel coordinates of the video frame and the spatial coordinates of the three-dimensional point cloud through feature point matching to generate a dense three-dimensional laser point cloud. S4. Input the dense three-dimensional laser point cloud into the pre-constructed three-dimensional spatial understanding model, extract spatial features and align them with the semantic model, and output the text describing the internal spatial structure of the transformer and the location information of defects. S5. Based on the spatial structure description text and defect location information, plan the inspection path, complete the full coverage inspection, and generate a standardized inspection report.

[0010] Preferably, in step S2, the multi-scale Retinex enhancement includes: S201. Perform multi-scale Gaussian blur on the extracted video frames to generate blurred images at three scales. ; S202, Based on the blurred images at each scale Calculate Retinex-enhanced images :

[0011] in, These are the original video frames. For scale weights, To avoid the logarithmically meaningless minimum; S203, Enhance Retinex image Gray-level normalization is performed, and the gray-level range of 0-255 is mapped through linear transformation to complete the initial enhancement.

[0012] Preferably, the weights of the improved bilateral filter are dynamically adjusted based on the spatial domain standard deviation and the gray-level domain standard deviation. The calculation formula is:

[0013] in, The current pixel coordinates, For the neighboring pixel coordinates, The standard deviation of the spatial domain is dynamically set to 5-15 based on the image texture, with 12-15 for complex texture areas and 5-8 for simple texture areas. The standard deviation for the grayscale range is dynamically set to 10-30 based on the image's grayscale distribution. For areas with large grayscale differences, a value of 25-30 is used, while for areas with uniform grayscale, a value of 10-15 is used. , These are the grayscale values ​​of the current pixel and its neighboring pixels, respectively.

[0014] Preferably, the registration in step S3 employs feature extraction. coarse matching The optimized three-level registration method specifically includes: S301, Feature Extraction: Extract SIFT feature points from enhanced video frames and ISS feature points from raw 3D point cloud data; S302, Coarse matching: The FLANN matching algorithm is used to perform preliminary matching between the SIFT feature points and the ISS feature points to obtain initial matching pairs; S303. Fine-tuning: The RANSAC algorithm is used to remove mismatches from the initial matching pairs, retain the optimal matching pairs, and calculate the rotation matrix and translation vector based on the optimal matching pairs. S304. Dense point cloud generation: Based on the rotation matrix and translation vector, each pixel of the video frame is mapped to a three-dimensional space, and the sparse point cloud is interpolated and completed to generate a dense three-dimensional laser point cloud.

[0015] Preferably, the three-dimensional spatial understanding model in step S4 includes: An improved point cloud encoder, based on the Point Transformer V3 architecture, adds a "transformer target attention mechanism" to extract target-oriented features from dense 3D laser point clouds using a preset transformer internal target feature template; The cross-modal feature alignment module receives the point cloud feature matrix. It adopts a dual mechanism of "dynamic projection + semantic constraint". First, it maps the point cloud feature matrix to the basic feature vector through a two-layer fully connected network. Then, it introduces a transformer-specific semantic dictionary to perform semantic constraints on the basic feature vector. It calculates the cosine similarity between the feature vector and the embedding vector of key terms in the dictionary, optimizes the feature vector, and achieves deep alignment between the point cloud feature vector and the semantic model. The semantic generation layer, based on a lightweight language model, takes aligned feature vectors as input and outputs text describing the internal spatial structure of the transformer and information on the location of defects through an autoregressive generation mechanism.

[0016] Preferably, the improved point cloud encoder includes 6 attention layers, each with 8 attention heads and a hidden layer dimension of 256. During forward propagation, the point cloud features and feature templates are subjected to attention calculation, thereby enhancing the feature extraction of the target part and outputting a point cloud feature matrix with dimension K×D, where K is the number of features and D is the hidden layer dimension.

[0017] Preferably, the defect location information includes defect type, size, spatial coordinates, and grayscale value; the defect type includes loose winding, carbon marks, electrical trees, foreign objects, lead displacement, and poor contact of tap changers.

[0018] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0019] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0020] The present invention also discloses an oil-immersed transformer inspection system based on three-dimensional spatial perception, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0021] Compared with the prior art, the advantages of the present invention are as follows: This invention addresses the problems of poor image quality and weak 3D spatial understanding caused by dim lighting inside oil-immersed transformers and interference from insulating oil. It utilizes a robotic fish equipped with a visible light imaging module and a lidar module working in tandem. First, a multi-scale adaptive image enhancement algorithm is employed to optimize the quality of visible light video frames. Then, pixel-point cloud registration technology is used to transform the enhanced video into a dense 3D laser point cloud. Finally, a 3D spatial understanding model based on "point cloud feature-semantic association dual-drive" is introduced to perform semantic analysis of the laser point cloud. This enables the robotic fish to navigate autonomously and accurately identify internal transformer defects. It can effectively replace manual labor in the internal inspection of oil-immersed transformers of various voltage levels from 110kV to 1000kV, meeting the "efficient, accurate, and safe" requirements of power grid operation and maintenance.

[0022] This invention provides an inspection method for oil-immersed transformers. Addressing the complex conditions inside oil-immersed transformers, such as dim lighting, interference from insulating oil, and confined spaces, this method utilizes image enhancement, video-point cloud fusion, and 3D spatial understanding technologies to achieve accurate identification of internal defects and spatial structure analysis. Its core application is in the internal inspection of 110kV and above high-voltage oil-immersed transformers, and it can be extended to the internal inspection of oil-immersed power equipment such as reactors and instrument transformers. This invention uses an "image enhancement-video point cloud fusion-3D spatial understanding" technology chain to solve the problems of poor image quality and weak spatial understanding in the internal inspection of oil-immersed transformers. It enables a robotic fish to autonomously complete high-precision, high-efficiency internal inspections, replacing manual inspection and reducing maintenance costs and safety risks. Attached Figure Description

[0023] Figure 1 This is a flowchart of an embodiment of the oil-immersed transformer inspection method of the present invention.

[0024] Figure 2 This is an architectural diagram of the transformer-specific three-dimensional spatial understanding model of the present invention in an embodiment. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] like Figure 1 As shown, the oil-immersed transformer inspection method based on three-dimensional spatial perception provided in this embodiment of the invention includes the following steps: S1. Visible light and laser point cloud data acquisition.

[0027] The robotic fish, equipped with a visible light imaging module, a lidar module, and an embedded computing unit, enters the interior of an oil-immersed transformer. The visible light imaging module captures a visible light video stream from inside the transformer at a frame rate of 25fps, while the lidar module simultaneously acquires raw 3D point cloud data of the environment at a sampling frequency of 10Hz (point cloud density 100-200 points / m²). 2 ); The robotic fish is equipped with a visible light imaging module that uses a low-light 1080P camera and two 3W high-brightness LED lighting units; a lidar module that uses a miniature lidar; and an embedded computing unit that uses an NVIDIA Jetson NX chip.

[0028] S2. Visible light image preprocessing.

[0029] The embedded computing unit extracts frames from the visible light video stream to obtain continuous video frames. Due to the extremely weak lighting inside the transformer and severe light scattering caused by the insulating oil, the original image has low contrast and high noise. Therefore, a multi-scale adaptive image enhancement algorithm is used to optimize each frame, resulting in enhanced video frames. The multi-scale adaptive image enhancement algorithm combines multi-scale Retinex enhancement with improved bilateral filtering denoising. First, it enhances image details and contrast through multi-scale Retinex enhancement, separating the image's illumination and reflection components to strengthen defect texture features. The specific implementation process of multi-scale Retinex enhancement is as follows: S201. Perform multi-scale Gaussian blur on the extracted video frames to generate blurred images at three scales (5×5, 15×15, 25×25). The 5×5 scale captures the image detail blur component, the 15×15 scale captures the image medium texture blur component, and the 25×25 scale captures the image global illumination blur component. S202. Calculate the Retinex-enhanced image according to the following formula, separate the image illumination component and reflection component, and enhance the texture features of defects:

[0030] in, These are the original video frames. Scale weights (values ​​of 0.5, 0.3, and 0.2; 0.5 for 5×5 scales to highlight details, and 0.2 for 25×25 scales to suppress global illumination interference). To avoid the logarithmically meaningless minimum; S203, to Perform grayscale normalization and then map the grayscale value to the 0-255 grayscale range through linear transformation to complete the initial enhancement. To further eliminate scattering noise from the oil medium while preserving key defect edges, an improved bilateral filter is applied to the normalized image. By improving bilateral filtering denoising, high weights are used to preserve details in image edge regions, while low weights are used to enhance denoising in smooth regions, thus eliminating image noise caused by the insulating oil medium and obtaining the final enhanced video frame. .

[0031] The weights for the improved bilateral filter are calculated as follows:

[0032] in, The current pixel coordinates, For the neighboring pixel coordinates, The standard deviation of the spatial domain is dynamically set to 5-15 based on the image texture, with 12-15 for complex texture areas and 5-8 for simple texture areas. The standard deviation of the grayscale range is dynamically set to 10-30 based on the grayscale distribution of the image, with 25-30 for areas with large grayscale differences and 10-15 for areas with uniform grayscale. , These are the grayscale values ​​of the current pixel and its neighboring pixels, respectively.

[0033] S3, Pixel-Point Cloud Registration To establish a precise mapping relationship between two-dimensional images and three-dimensional space, this invention adopts a three-level registration method of "feature extraction-coarse matching-fine optimization" to solve the problems of "feature blurring and sparse point cloud caused by oil medium". Specifically, the enhanced video frame from step S2 is pixel-to-point cloud registration is performed with the original 3D point cloud data acquired by the LiDAR module. A mapping relationship between the video frame pixel coordinates and the 3D point cloud spatial coordinates is established through feature point matching, generating a dense 3D LiDAR point cloud. The specific steps are as follows: S301, Feature Extraction: From Enhanced Video Frames SIFT feature points are extracted (using a 16×16 neighborhood to calculate the gradient direction histogram, generating a 128-dimensional descriptor, with the number of feature points controlled at 600-800 per frame); ISS feature points are extracted from the original 3D point cloud data (by calculating the local curvature changes of the point cloud, retaining curvature abrupt change points, with the number of feature points controlled at 400-600 per frame), and the normal vectors (using the neighborhood point fitting plane method, with a normal vector calculation error ≤0.05) and curvature (using the principal component analysis method, with a curvature calculation error ≤0.02) of the ISS feature points are calculated simultaneously. S302. Coarse matching: The fast nearest neighbor search FLANN matching algorithm (constructing a KD tree index with a search radius of 0.8) is used to perform preliminary matching of SIFT feature points and ISS feature points to obtain initial matching pairs. The matching accuracy is controlled at 70%-75%. S303, Fine-tuning: The RANSAC algorithm (with 1000 iterations and an inlier threshold of 1.5 pixels) is used to remove mismatches from the initial matching pairs and retain the best matching pairs (the number is controlled at 250-350 per frame). The rotation matrix R (a 3×3 orthogonal matrix with a determinant of 1 to ensure rotation consistency) and the translation vector t (a 3×1 vector) are calculated using the least squares method based on the optimal matching pair, as shown below:

[0034] in, These are the pixel coordinates of the video frame. The coordinates of the three-dimensional point cloud; S304. Dense Point Cloud Generation: Using the obtained rotation matrix R and translation vector t, each pixel of the video frame is mapped to three-dimensional space according to the mapping relationship. Since the original point cloud from the LiDAR is sparse, gaps still exist in the projected point cloud. This is combined with the LiDAR point cloud density (100-200 points / m²). 2 The sparse point cloud is interpolated and completed (using inverse distance weighted interpolation method, with interpolation error ≤1mm) to generate a dense 3D laser point cloud. .

[0035] S4. Construct a 3D spatial understanding model based on "point cloud feature-semantic association dual-drive" to transform the dense 3D laser point cloud generated in step S3. The model is input, and it first extracts spatial features through an improved point cloud encoder. Then, it achieves deep alignment between features and semantic model through a cross-modal feature alignment module. Finally, it outputs text describing the internal spatial structure of the transformer (including the three-dimensional coordinates and dimensions of building elements such as walls, windings, and tap changers) and defect location information (including defect type, size, and spatial coordinates) through a semantic generation layer. To address the shortcomings of existing 3D spatial understanding models, such as poor adaptability to general scenarios and poor recognition of transformer-specific targets, this invention proposes three innovative improvements to construct a transformer-specific 3D spatial understanding model. The model architecture is as follows: Figure 2 As shown: An improved point cloud encoder. Based on the Point Transformer V3 architecture, it adds a "transformer target attention mechanism," which uses preset feature templates of internal transformer targets (windings, tap changers, tank walls) to process dense 3D laser point clouds. Target-oriented feature extraction is performed; the encoder contains 6 attention layers, each with 8 attention heads and a hidden layer dimension of 256; during forward propagation, the point cloud features are used to perform attention calculations with these feature templates, thereby enhancing the feature extraction of the target part and outputting a point cloud feature matrix with dimension K×D, where K is the number of features and D is the hidden layer dimension.

[0036] Cross-modal feature alignment module. This module receives the point cloud feature matrix; it adopts a dual mechanism of "dynamic projection + semantic constraint". First, it maps the point cloud feature matrix to the basic feature vector through a two-layer fully connected network (input dimension 512, output dimension 768, the first layer activation function is ReLU, and the second layer activation function is GELU). Then, it introduces a transformer professional semantic dictionary (containing 300+ professional terms such as "winding turn distance", "insulation paper thickness", "tap switch contact gap" etc.) to perform semantic constraints on the basic feature vector. It calculates the cosine similarity between the feature vector and the embedding vector of key terms in the dictionary, and sets the similarity threshold to 0.85 as a constraint to optimize the feature vector, thereby achieving deep alignment between the point cloud feature vector and the semantic model.

[0037] Semantic Generation Layer. Based on a lightweight language model, the layer takes aligned feature vectors as input and outputs text describing the internal spatial structure of the transformer (e.g., “Wall W1: coordinates (0,0,0)-(2000,0,3000), height 3000mm, thickness 10mm”) and defect location information (e.g., “Defect 1: carbon mark, location (550,550,1500), size 4mm×3mm, grayscale value 180-200”) through an autoregressive generation mechanism (using a bundle search algorithm with a bundle width of 5). The layer also provides information on the location of defects.

[0038] The lightweight language model is based on the Qwen2.5-0.5B architecture, with parameters compressed to 0.3B, resulting in a 40% improvement in inference speed.

[0039] S5. Autonomous Inspection by the Robotic Fish. Based on the spatial structure description text and defect location information obtained in step S4, the embedded computing unit plans the inspection path of the robotic fish (using the A* algorithm, with a path planning error ≤ 5mm), controls the robotic fish to complete a full-coverage inspection along the planned path, and simultaneously records defect information to generate a standardized inspection report.

[0040] This invention addresses the problems of poor image quality and weak 3D spatial understanding caused by dim lighting inside oil-immersed transformers and interference from insulating oil. It utilizes a robotic fish equipped with a visible light imaging module and a lidar module working in tandem. First, a multi-scale adaptive image enhancement algorithm is employed to optimize the quality of visible light video frames. Then, pixel-point cloud registration technology is used to transform the enhanced video into a dense 3D laser point cloud. Finally, a 3D spatial understanding model based on "point cloud feature-semantic association dual-drive" is introduced to perform semantic analysis of the laser point cloud. This enables the robotic fish to navigate autonomously and accurately identify internal transformer defects. It can effectively replace manual labor in the internal inspection of oil-immersed transformers of various voltage levels from 110kV to 1000kV, meeting the "efficient, accurate, and safe" requirements of power grid operation and maintenance.

[0041] This invention provides an inspection method for oil-immersed transformers. Addressing the complex conditions inside oil-immersed transformers, such as dim lighting, interference from insulating oil, and confined spaces, this method utilizes image enhancement, video-point cloud fusion, and 3D spatial understanding technologies to achieve accurate identification of internal defects and spatial structure analysis. Its core application is in the internal inspection of 110kV and above high-voltage oil-immersed transformers, and it can be extended to the internal inspection of oil-immersed power equipment such as reactors and instrument transformers. This invention uses an "image enhancement-video point cloud fusion-3D spatial understanding" technology chain to solve the problems of poor image quality and weak spatial understanding in the internal inspection of oil-immersed transformers. It enables a robotic fish to autonomously complete high-precision, high-efficiency internal inspections, replacing manual inspection and reducing maintenance costs and safety risks.

[0042] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0043] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0044] The present invention also discloses an oil-immersed transformer inspection system based on three-dimensional spatial perception, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0045] The products, media, and systems of the present invention, corresponding to the methods described above, also possess the advantages described above.

[0046] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0047] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for inspecting oil-immersed transformers based on three-dimensional spatial perception, characterized in that, Including the following steps: S1. Obtain the visible light video stream and synchronized 3D point cloud raw data inside the transformer; S2. Frame extraction is performed on the visible light video stream to obtain continuous video frames, and a multi-scale adaptive image enhancement algorithm is used to optimize each frame to obtain enhanced video frames; the multi-scale adaptive image enhancement algorithm includes sequential multi-scale Retinex enhancement and improved bilateral filtering denoising; S3. Register the enhanced video frame with the original three-dimensional point cloud data, and establish a mapping relationship between the pixel coordinates of the video frame and the spatial coordinates of the three-dimensional point cloud through feature point matching to generate a dense three-dimensional laser point cloud. S4. Input the dense three-dimensional laser point cloud into the pre-constructed three-dimensional spatial understanding model, extract spatial features and align them with the semantic model, and output the text describing the internal spatial structure of the transformer and the location information of defects. S5. Based on the spatial structure description text and defect location information, plan the inspection path, complete the full coverage inspection, and generate a standardized inspection report.

2. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 1, characterized in that, In step S2, the multi-scale Retinex enhancement includes: S201. Perform multi-scale Gaussian blur on the extracted video frames to generate blurred images at three scales. ; S202, Based on the blurred images at each scale Calculate Retinex-enhanced images : in, These are the original video frames. For scale weights, To avoid the logarithmically meaningless minimum; S203, Enhance Retinex image Gray-level normalization is performed, and the gray-level range of 0-255 is mapped through linear transformation to complete the initial enhancement.

3. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 2, characterized in that, The weights of the improved bilateral filter are dynamically adjusted based on the spatial domain standard deviation and the gray-level domain standard deviation. The calculation formula is: in, The current pixel coordinates, For the neighboring pixel coordinates, The standard deviation of the spatial domain is dynamically set to 5-15 based on the image texture, with 12-15 for complex texture areas and 5-8 for simple texture areas. The standard deviation for the grayscale range is dynamically set to 10-30 based on the image's grayscale distribution. For areas with large grayscale differences, a value of 25-30 is used, while for areas with uniform grayscale, a value of 10-15 is used. , These are the grayscale values ​​of the current pixel and its neighboring pixels, respectively.

4. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 1, 2, or 3, characterized in that, The registration in step S3 uses feature extraction. coarse matching The optimized three-level registration method specifically includes: S301, Feature Extraction: Extract SIFT feature points from enhanced video frames and ISS feature points from raw 3D point cloud data; S302, Coarse matching: The FLANN matching algorithm is used to perform preliminary matching between the SIFT feature points and the ISS feature points to obtain initial matching pairs; S303. Fine-tuning: The RANSAC algorithm is used to remove mismatches from the initial matching pairs, retain the optimal matching pairs, and calculate the rotation matrix and translation vector based on the optimal matching pairs. S304. Dense point cloud generation: Based on the rotation matrix and translation vector, each pixel of the video frame is mapped to a three-dimensional space, and the sparse point cloud is interpolated and completed to generate a dense three-dimensional laser point cloud.

5. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 1, 2, or 3, characterized in that, The three-dimensional spatial understanding model in step S4 includes: An improved point cloud encoder, based on the Point Transformer V3 architecture, adds a "transformer target attention mechanism" to extract target-oriented features from dense 3D laser point clouds using a preset transformer internal target feature template; The cross-modal feature alignment module receives the point cloud feature matrix. It adopts a dual mechanism of "dynamic projection + semantic constraint". First, it maps the point cloud feature matrix to the basic feature vector through a two-layer fully connected network. Then, it introduces a transformer-specific semantic dictionary to perform semantic constraints on the basic feature vector. It calculates the cosine similarity between the feature vector and the embedding vector of key terms in the dictionary, optimizes the feature vector, and achieves deep alignment between the point cloud feature vector and the semantic model. The semantic generation layer, based on a lightweight language model, takes aligned feature vectors as input and outputs text describing the internal spatial structure of the transformer and information on the location of defects through an autoregressive generation mechanism.

6. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 5, characterized in that, The improved point cloud encoder contains 6 attention layers, each with 8 attention heads and a hidden layer dimension of 256. During forward propagation, point cloud features and feature templates are subjected to attention calculations to enhance feature extraction of target parts. The output is a point cloud feature matrix with dimension K×D, where K is the number of features and D is the hidden layer dimension.

7. The method for inspecting oil-immersed transformers based on three-dimensional spatial perception according to claim 1, 2, or 3, characterized in that, The defect location information includes defect type, size, spatial coordinates, and grayscale value; the defect type includes loose winding, carbon marks, electrical trees, foreign objects, lead displacement, and poor contact of tap changers.

8. A computer program product, comprising a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

10. An oil-immersed transformer inspection system based on three-dimensional spatial perception, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that... The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-7.

Images