A machine vision-based current transformer oil leakage detection method and system

Abstract

The present application relates to the field of substation oil leakage detection, and provides a current transformer oil leakage detection method and system based on machine vision, comprising: obtaining a substation map, obtaining an identification picture, obtaining a current transformer three-dimensional data graph, and obtaining a smart patrol process of a smart patrol vehicle; constructing an oil leakage detection map based on the substation map, and constructing a 3D model of each current transformer based on the current transformer three-dimensional data graph; obtaining an oil leakage fusion picture based on the identification picture and a double-branch U-Net++ semantic segmentation model; constructing an oil leakage 3D model based on the oil leakage fusion picture and the smart patrol process; and constructing a maintenance route based on the oil leakage 3D model and the oil leakage detection map. When the present application is used, maintenance personnel on the way to maintenance can determine the condition of the oil leakage point according to the oil leakage 3D model, and formulate a maintenance plan based on the oil leakage point condition, thereby effectively reducing the maintenance time of the oil leakage point.

Description

Technical Field

[0001] This invention relates to the field of substation oil leakage detection, and more specifically, to a method and system for detecting oil leakage in current transformers based on machine vision. Background Technology

[0002] The main functions of a substation are voltage transformation, power distribution, and safety control. Therefore, the safety of a substation is related to the stability of power supply, which in turn is related to production and people's livelihood. At the same time, since oil leakage of current transformers is an important factor affecting the safety of substations, current transformers are monitored in real time during substation operation to ensure that the oil leakage on the surface of the current transformers can be reported in a timely manner.

[0003] Because current transformers have irregular shapes and large overall areas, and their surfaces are uneven, flat graphics cannot accurately show the exact location of oil leaks. In addition, substations have large overall areas, and oil leaks cannot be associated with corresponding current transformers. This prolongs the time for maintenance personnel to find the oil leaks, thus increasing the difficulty of current transformer maintenance.

[0004] To address this issue, we propose a machine vision-based method and system for detecting oil leakage in current transformers. Summary of the Invention

[0005] The technical problem to be solved:

[0006] In view of the problems existing in the prior art, the purpose of this invention is to provide a current transformer oil leakage detection method and system based on machine vision, which solves the problem of the high difficulty of current transformer maintenance.

[0007] Technical solution:

[0008] To solve the above problems, the present invention adopts the following technical solution.

[0009] A machine vision-based method and system for detecting oil leakage in current transformers, comprising:

[0010] Obtain substation maps, obtain recognition images, obtain 3D data maps of current transformers, and obtain the intelligent patrol process of the intelligent patrol vehicle;

[0011] An oil leak detection map is constructed based on the substation map, and 3D models of each current transformer are constructed based on the three-dimensional data map of the current transformer.

[0012] Based on the identified image and the dual-branch U-Net++ semantic segmentation model, an oil leak fusion image is obtained;

[0013] A 3D model of the oil leak is constructed based on the fused oil leak images and the intelligent patrol process.

[0014] A repair route is constructed based on the oil leak 3D model and the oil leak detection map.

[0015] The process of acquiring substation maps, acquiring recognition images, acquiring 3D data of current transformers, and acquiring intelligent patrol vehicle data includes:

[0016] The substation map was independently constructed based on on-site investigation;

[0017] The identified image is a bimodal image, which includes an infrared image and a high-resolution image;

[0018] The three-dimensional data of the current transformers were obtained based on on-site inspections of all the current transformers contained in the substation;

[0019] The intelligent patrol process of the intelligent patrol vehicle is obtained, wherein the intelligent patrol process includes the intelligent patrol location.

[0020] The steps of constructing an oil leak detection map based on the substation map and constructing the corresponding 3D model based on the current transformer 3D data include:

[0021] A smart patrol map is constructed based on the aforementioned substation map;

[0022] Based on the substation map, the location of each current transformer is obtained, and the location of the current transformer is marked on the intelligent patrol map to construct an oil leak detection map.

[0023] Based on the three-dimensional data of the current transformers, 3D models of each current transformer are constructed, and the positions of the current transformers are associated with the 3D models.

[0024] The step of obtaining the oil leak fusion image based on the identified image and the dual-branch U-Net++ semantic segmentation model includes:

[0025] The infrared image and the high-resolution image are obtained based on the identified image;

[0026] Copy the aforementioned high-resolution image as a backup high-resolution image;

[0027] Based on the infrared image, an effective region is obtained by adaptive threshold segmentation. The effective region is then denoised by median filtering. Temperature anomaly regions are marked based on the denoised effective region.

[0028] The high-definition image is denoised and enhanced to crop out the device area image;

[0029] The rotation matrix and translation vector of the infrared camera and the high-definition camera are obtained based on the Zhang Zhengyou calibration method, and the effective area is aligned with the device area map based on the rotation matrix and the translation vector.

[0030] The effective region and the device region are input into the internal output oil stain mask of the dual-branch U-Net++ semantic segmentation model;

[0031] The oil stain mask is mapped onto the surface of the backup high-resolution image to construct the oil leak fusion image.

[0032] The steps for constructing a 3D model of the oil leak based on the fused oil leak image and the intelligent patrol process include:

[0033] The intelligent patrol location is confirmed based on the intelligent patrol process;

[0034] Based on the location of the intelligent patrol, the corresponding 3D model is confirmed, and the corresponding 3D model is copied as a backup 3D model;

[0035] The pixel coordinates of the oil leak marker in the oil leak fusion image are extracted as oil trace coordinates, and the oil trace coordinates are converted into 3D global oil trace coordinates;

[0036] The 3D global oil trace coordinates are fused to the surface of the backup 3D model to construct the oil leak 3D model.

[0037] The step of constructing a repair route based on the oil leak 3D model and the oil leak detection map includes:

[0038] Based on the oil leak detection map, the location points of all the oil leak 3D models were confirmed;

[0039] A repair route is constructed based on all the aforementioned location points and the oil leak detection map.

[0040] The step of inputting the effective region and the device region into the internal output oil stain mask of the dual-branch U-Net++ semantic segmentation model includes:

[0041] The effective region and the device region are simultaneously resized and aligned to obtain the standard effective region and the standard device region;

[0042] The standard effective region is input into the infrared branch to obtain temperature difference characteristics;

[0043] The standard device area is input into the high-definition branch to obtain texture morphology features;

[0044] The contribution weights of the bimodal features are calculated based on the channel attention module, and the oil trace fusion features are calculated based on the contribution weights.

[0045] The oil trace fusion feature is based on multi-scale skip connections and dense decoder block output of the initial oil trace mask;

[0046] The initial oil stain mask is optimized through post-processing to obtain the oil stain mask.

[0047] The step of converting the oil trail coordinates into 3D global oil trail coordinates includes:

[0048] The oil trail coordinates are transformed into normalized oil trail coordinates based on the camera coordinate system;

[0049] The depth value of the oil stain attachment location is extracted based on the metadata of the 3D model;

[0050] The normalized oil trail coordinates and the depth value are transformed into 3D oil trail coordinates based on the camera coordinate system;

[0051] The 3D oil trail coordinates are converted into the 3D global oil trail coordinates based on the corresponding rotation matrix and translation vector.

[0052] The training steps of the dual-branch U-Net++ semantic segmentation model are as follows:

[0053] Collect 1000 sets of dual-modal oil leakage images of substations and mark oil trace masks on the surface of the high-resolution images to form a dataset;

[0054] The dataset is divided into training set, validation set, and test set according to the specified proportions.

[0055] A dual-modal synchronous enhancement training set is used to align the high-resolution icon annotations in the training set at the pixel level.

[0056] The training set is input into the dual-branch U-Net++ semantic segmentation model to train the model;

[0057] Input the validation set into the dual-branch U-Net++ semantic segmentation model to avoid over-convergence;

[0058] The test set was input into the dual-branch U-Net++ semantic segmentation model to evaluate the model's capabilities.

[0059] A machine vision-based current transformer oil leakage detection system includes:

[0060] The image acquisition module is used to acquire substation maps, acquire and recognize images, acquire 3D data maps of current transformers, and acquire the intelligent patrol process of the intelligent patrol vehicle.

[0061] The image integration module is used to construct an oil leak detection map based on the substation map and to construct 3D models of each current transformer based on the current transformer 3D data map.

[0062] The feature extraction module is used to obtain the oil leak fusion image based on the identified image and the dual-branch U-Net++ semantic segmentation model;

[0063] The feature fusion module is used to construct a 3D model of the oil leak based on the fused oil leak image and the intelligent patrol process;

[0064] The route construction module is used to construct a repair route based on the oil leak 3D model and the oil leak detection map.

[0065] Beneficial effects: Compared with the prior art, the advantages of this invention are:

[0066] This application detects oil stains using a dual-modal approach combining infrared and high-resolution images. It also adjusts the contribution weights of the dual-modal features based on illumination, ensuring that the high-resolution image dominates under high-intensity illumination and the infrared image dominates under low-light conditions, thereby improving the accuracy of oil stain recognition.

[0067] Simultaneously, after detecting oil stains, the oil stains are mapped onto the surface of the corresponding 3D model. The 3D model is built based on the current transformer of the substation. Therefore, when the oil stains are mapped onto the surface of the 3D model, maintenance personnel can quickly check the oil leak point based on the oil leak 3D model. This allows maintenance personnel to judge the oil leak point on their way to the repair site and formulate a repair plan based on the oil leak point, effectively reducing the repair time of the oil leak point.

[0068] Since the substation has a large area and contains multiple sets of current transformers, this application can obtain the corresponding location based on the 3D model and construct the maintenance route based on the 3D model location. When searching for the 3D model of the oil leak, the corresponding current transformer can be quickly found based on the maintenance route, saving maintenance time. Attached Figure Description

[0069] Figure 1 This is a flowchart of the oil leak detection method of the present invention;

[0070] Figure 2 This is a flowchart of the data acquisition method of the present invention;

[0071] Figure 3 This is a flowchart of the 3D model construction method of the present invention;

[0072] Figure 4 This is a flowchart of the oil stain mask extraction method of the present invention;

[0073] Figure 5 This is a flowchart of the oil leak 3D model construction method of the present invention;

[0074] Figure 6 This is a flowchart of the route construction method of the present invention. Detailed Implementation

[0075] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0076] This application provides a machine vision-based method and system for detecting oil leakage in current transformers, which solves the problem of high maintenance difficulty of current transformers. In use, the method marks the oil leakage point on the surface of a 3D model, making it convenient for maintenance personnel to quickly find the oil leakage point on the surface of the current transformer to be repaired.

[0077] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0078] Please see Figure 1 A method and system for detecting oil leakage in current transformers based on machine vision, comprising the following steps:

[0079] S1. Obtain the substation map, obtain the identification image, and obtain the 3D data map of the current transformer. Since the current transformer will have surface oxidation over time and will also be replaced due to maintenance, the current transformer needs to re-obtain the relevant 3D data map and draw the latest 3D model every once in a while. Obtain the intelligent patrol process of the intelligent patrol vehicle. The intelligent patrol process is mainly to record the corresponding location in real time.

[0080] S2. Construct an oil leak detection map based on the substation map. The oil leak detection map is mainly used to view the distribution of current transformers in the substation. Based on the three-dimensional data map of the current transformers, construct 3D models of each current transformer. The 3D models are used to restore the shape of the current transformers to the greatest extent possible.

[0081] S3. Based on the recognized image and the dual-branch U-Net++ semantic segmentation model, the oil leak fusion image is obtained. The dual-branch U-Net++ semantic segmentation model is set to have two parts to process the infrared image and the high-definition image respectively.

[0082] S4. Based on the oil leak fusion image and intelligent inspection process, a 3D model of the oil leak is constructed. The surface of the 3D model of the oil leak will be mapped with oil traces, which makes it convenient for maintenance personnel to analyze the oil leak situation.

[0083] S5. Based on the 3D model of the oil leak and the oil leak detection map, a maintenance route is constructed. The maintenance route is mainly used to help maintenance personnel quickly find the current transformer that is leaking oil.

[0084] Please see Figure 2A flowchart illustrating the data acquisition method of a machine vision-based current transformer oil leakage detection method and system, including:

[0085] S11. The substation map is independently constructed based on on-site investigation. The current transformers are placed in different locations in each substation, so on-site investigation is necessary to create an accurate substation map.

[0086] S12. The image is identified as a bimodal image, which includes an infrared image and a high-definition image. The infrared image is dominant in detecting oil stains under low light conditions, while the high-definition image is dominant in detecting oil stains under sufficient light conditions.

[0087] S13. The three-dimensional data of the current transformer is obtained based on the on-site inspection of all the current transformers in the substation. The three-dimensional data is mainly to better restore the current transformer. At the same time, the three-dimensional data is obtained through high-definition equipment, which makes it convenient for the 3D model to restore the actual color of the current transformer one-to-one.

[0088] S14. Obtain the intelligent patrol process of the intelligent patrol vehicle, wherein the intelligent patrol process includes the intelligent patrol location, and the location of the measured current transformer can be determined based on the intelligent patrol location.

[0089] Please see Figure 3 A flowchart illustrating a 3D model construction method for a current transformer oil leakage detection method and system based on machine vision, including:

[0090] S21. Construct a smart inspection map based on the substation map, obtain the location of each current transformer based on the substation map, mark the current transformer location on the smart inspection map to construct an oil leak detection map.

[0091] The current transformers are marked on the surface of the substation, which makes it easier for maintenance personnel to see the location of the current transformers more intuitively. Furthermore, associating the current transformers with their corresponding locations can increase the independence of each current transformer.

[0092] S22. Construct 3D models of each current transformer based on the 3D data of the current transformers, and associate the positions of the current transformers with the 3D models.

[0093] 3D models can more intuitively represent the overall appearance of current transformers. When oil leaks occur, the corresponding positions on the 3D model can be rendered, allowing maintenance personnel to better understand the oil leak situation based on the oil leak points on the 3D model surface.

[0094] Please see Figure 4 A flowchart of an oil trace mask extraction method for a current transformer oil leakage detection method and system based on machine vision, including:

[0095] S31. Based on the recognized image, obtain the infrared image and the high-definition image, copy the high-definition image as a backup high-definition image, and fuse the backup high-definition image with the infrared image for easy reference during subsequent fusion.

[0096] S32. Based on the infrared image, the effective area is obtained by adaptive threshold segmentation. The effective area is denoised by median filtering. The temperature abnormal area is marked based on the denoised effective area. Since there is an obvious temperature difference between the surface of the oil leak and other locations, the temperature difference of the infrared image can be used to find the oil trace location better under low light conditions.

[0097] S33. Denoising and enhancing the high-definition image to crop out the device area image. High-definition images contain many irrelevant backgrounds, so in order to reduce workload, irrelevant backgrounds are usually removed.

[0098] S34. Obtain the rotation matrix and translation vector of the infrared camera and the high-definition camera based on Zhang Zhengyou's calibration method. Align the effective area with the device area map based on the rotation matrix and translation vector. Simultaneously resize and calibrate the effective area and the device area to obtain the standard effective area and the standard device area.

[0099] Synchronous resizing refers to strictly adhering to a unified target size, a unified interpolation method, and a unified scaling anchor point when adjusting the image size of the infrared and high-resolution images in dual-modal detection, thereby reducing pixel offset.

[0100] Infrared images are single-channel grayscale images, which can clearly show the temperature differences on the model surface and do not depend on strong light.

[0101] The high-resolution image is denoised based on wavelet thresholding and then enhanced. When cropping the recognition area, the original RGB colors and oil stain texture need to be preserved.

[0102] Zhang Zhengyou's calibration method is based on the known geometric information of the checkerboard calibration object. It involves taking multi-view images of the checkerboard, solving for camera parameters using the homography matrix, and then calculating the corresponding rotation matrix and translation vector based on the parameters.

[0103] S34. Based on the recognized image and the dual-branch U-Net++ semantic segmentation model, an oil leak fusion image is obtained;

[0104] The dual-branch U-Net++ semantic segmentation model mainly consists of two branches that specifically adjust the infrared image and the high-resolution image respectively;

[0105] S341. Input the standard effective area into the infrared branch to obtain temperature difference features, and input the standard device area into the high-definition branch to obtain texture morphology features;

[0106] Acquiring the texture features of oil stains is a key distinguishing factor in dual-modal oil leak detection. Its core function is to complement infrared temperature features, enabling the identification of oil stains from interference and small oil stains that cannot be distinguished by temperature alone, thus ensuring that the model can accurately identify oil stains in all substation scenarios.

[0107] S342. Calculate the contribution weights of the bimodal features based on the channel attention module, and calculate the oil trace fusion features based on the contribution weights. The specific formula is as follows:

[0108] ;

[0109] ;

[0110] in, Infrared characteristics, It is characterized by high definition. The characteristics of oil stain fusion For infrared weighting, For high-definition weight;

[0111] The contribution weight primarily determines the dominant image; in low-light conditions, infrared radiation is increased. This makes the infrared image dominant, while under sufficient lighting conditions, it increases the resolution of the image. This makes high-definition images the dominant feature;

[0112] S343. The oil stain fusion feature is based on multi-scale skip connections and the output of the dense decoder block as an initial oil stain mask. The initial oil stain mask is optimized through post-processing to obtain the final oil stain mask. The specific steps of the optimization are as follows: First, morphological closing operation is performed on the initial oil stain mask. Through the operation of "dilation first, erosion later", the small holes inside the mask are filled and the broken areas are connected. Then, all connected components in the mask after the closing operation are extracted, the pixel area of ​​each connected component is calculated, and connected components with an area less than a preset threshold are judged as noise and removed. Finally, the remaining mask area is back-projected onto the infrared temperature matrix, and the average temperature of the area is calculated. If the difference between the average temperature and the ambient temperature is less than a preset temperature difference threshold, it is judged as a false detection and the mask is cleared, and the final oil stain mask is obtained.

[0113] Oil stain mask is the core intermediate carrier in the entire process of dual-modal oil leak detection. Its function is to accurately mark the location of oil stains, isolate invalid backgrounds, connect various technical links, and run through the entire process of model training, infrared coarse screening, high-definition fine judgment, 3D fusion, and operation and maintenance visualization.

[0114] S344, Training steps for the dual-branch U-Net++ semantic segmentation model:

[0115] Collect 1000 sets of dual-modal oil leakage images of substations and mark oil trace masks on the surface of high-resolution images to create a dataset;

[0116] The dataset is divided into training, validation, and test sets proportionally, with the following ratios: 60% training set, 20% validation set, and 20% test set.

[0117] The training set is augmented using dual-modal synchronous methods, and the high-resolution icons in the training set are aligned pixel-level.

[0118] The training set is input into the dual-branch U-Net++ semantic segmentation model to train the model;

[0119] Inputting the validation set into the dual-branch U-Net++ semantic segmentation model avoids over-convergence;

[0120] The test set was input into the dual-branch U-Net++ semantic segmentation model to evaluate the model's capabilities.

[0121] The training set should include all typical scenarios of a substation, such as strong midday light (20%), nighttime (25%), rainy weather (15%), dense equipment (20%), and different levels of oil leakage (20%).

[0122] At the same time, bimodal enhancement should be implemented based on the training set, such as rotation, scaling, noise superposition, and amplification of small oil stains, to expand the number of training sets, so that the model can learn the variant features of oil stains and improve its anti-interference ability.

[0123] During training, the model is evaluated using a validation set every 5 epochs. The model with the highest validation set mIoU and the lowest false negative rate is saved. For example, if the validation set mIoU reaches 96.5% when trained to 260 epochs and does not improve or even decreases in subsequent epochs, then the model is the optimal one.

[0124] mIoU is the core evaluation metric for semantic segmentation tasks. Its core function is to quantify the degree of overlap between the oil slick region predicted by the model and the real oil slick region annotated by humans.

[0125] The test set is used only once after the model training and validation are completed, and is not used for parameter tuning or model modification (if the test set is found to perform poorly in a certain scenario, the test set data cannot be used directly to supplement the training), so as to avoid data leakage that could lead to distorted evaluation results.

[0126] S35. Map the oil stain mask onto the surface of the backup high-resolution image to construct a fused image of the oil leak.

[0127] When constructing the fused oil leak image, it is referenced with the high-resolution image to avoid deviations in the oil stain markings and 3D anchor points after fusion, ensuring that the fusion result can directly meet the substation operation and maintenance needs.

[0128] Please see Figure 5 A flowchart illustrating a machine vision-based method for detecting oil leakage in current transformers and a method for constructing a 3D model of the leakage, including:

[0129] S41. Confirm the intelligent patrol location based on the intelligent patrol process, confirm the corresponding 3D model based on the intelligent patrol location, and copy the corresponding 3D model as a backup 3D model.

[0130] After obtaining the location of the intelligent patrol, the location of the nearest current transformer is determined based on the location of the intelligent patrol, and then the corresponding 3D model is obtained based on the location;

[0131] S42. Extract the pixel coordinates of the oil leak mark in the oil leak fusion image as oil trace coordinates, and transform the oil trace coordinates into normalized oil trace coordinates based on the camera coordinate system.

[0132] S43. When constructing the 3D model of each current transformer, a multi-view depth map is generated simultaneously based on the rendering of the model surface. The camera pose of each view is used as an index and stored together with the 3D model as metadata. Therefore, the depth value of the oil stain attachment position can be extracted based on the metadata of the 3D model. The normalized oil stain coordinates and depth values ​​are transformed into 3D oil stain coordinates based on the camera coordinate system. The 3D oil stain coordinates are then transformed into 3D global oil stain coordinates based on the corresponding rotation matrix and translation vector.

[0133] The depth value is used to fill the spatial dimension gap of 2D pixel coordinates. It defines the height and distance information of the oil stain in the physical world and is the key to achieving accurate mapping from 2D coordinate position to 3D physical position.

[0134] The core difference between global oil trace coordinates and 3D oil trace coordinates is that the coordinate system reference is different. Global oil trace coordinates are more suitable for the operation and maintenance positioning of the entire substation. The core of its conversion is to achieve mapping through the camera extrinsic rotation matrix R and translation vector T, ensuring that the oil trace position is uniquely locatable in the entire substation scenario.

[0135] S44. Merge the 3D global oil stain coordinates onto the surface of the spare 3D model to construct an oil leak 3D model;

[0136] The mapping of 3D global oil trace coordinates onto the surface of the standby 3D model is essentially the process of accurately binding abstract global 3D coordinates (X,Y,Z) to the device surface of the standby 3D model, generating surface anchor points.

[0137] Please see Figure 6 A flowchart illustrating the route construction method of a machine vision-based current transformer oil leakage detection method and system, including:

[0138] S51. Based on the oil leak detection map, confirm the location points of all oil leak 3D models;

[0139] S52. Construct a repair route based on all location points and the oil leak detection map;

[0140] Since there may be more than one current transformer leaking oil in a substation, maintenance may require a tour of the substation, and it is easy to forget about the current transformers that have already been repaired.

[0141] Therefore, given multiple source points, a genetic algorithm can be used to find the shortest closed path that passes through all points only once and eventually returns to the starting point.

[0142] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A machine vision-based method for detecting oil leakage in current transformers, characterized in that, include: Obtain substation maps, obtain recognition images, obtain 3D data maps of current transformers, and obtain the intelligent patrol process of the intelligent patrol vehicle; An oil leak detection map is constructed based on the substation map, and 3D models of each current transformer are constructed based on the three-dimensional data map of the current transformer. Based on the identified image and the dual-branch U-Net++ semantic segmentation model, an oil leak fusion image is obtained; A 3D model of the oil leak is constructed based on the fused oil leak images and the intelligent patrol process. A repair route is constructed based on the oil leak 3D model and the oil leak detection map.

2. The current transformer oil leakage detection method based on machine vision as described in claim 1, characterized in that, The process of acquiring substation maps, acquiring recognition images, acquiring 3D data of current transformers, and acquiring intelligent patrol vehicle data includes: The substation map was independently constructed based on on-site investigation; The identified image is a bimodal image, which includes an infrared image and a high-resolution image; The three-dimensional data of the current transformers were obtained based on on-site inspections of all the current transformers contained in the substation; The intelligent patrol process of the intelligent patrol vehicle is obtained, wherein the intelligent patrol process includes the intelligent patrol location.

3. The current transformer oil leakage detection method based on machine vision as described in claim 2, characterized in that: The steps of constructing an oil leak detection map based on the substation map and constructing the corresponding 3D model based on the current transformer 3D data include: A smart patrol map is constructed based on the aforementioned substation map; Based on the substation map, the location of each current transformer is obtained, and the location of the current transformer is marked on the intelligent patrol map to construct an oil leak detection map. Based on the three-dimensional data of the current transformers, 3D models of each current transformer are constructed, and the positions of the current transformers are associated with the 3D models.

4. The current transformer oil leakage detection method based on machine vision as described in claim 2, characterized in that: The step of obtaining the oil leak fusion image based on the identified image and the dual-branch U-Net++ semantic segmentation model includes: The infrared image and the high-resolution image are obtained based on the identified image; Copy the aforementioned high-resolution image as a backup high-resolution image; Based on the infrared image, an effective region is obtained by adaptive threshold segmentation. The effective region is then denoised by median filtering. Temperature anomaly regions are marked based on the denoised effective region. The high-definition image is denoised and enhanced to crop out the device area image; The rotation matrix and translation vector of the infrared camera and the high-definition camera are obtained based on the Zhang Zhengyou calibration method, and the effective area is aligned with the device area map based on the rotation matrix and the translation vector. The effective region and the device region are input into the internal output oil stain mask of the dual-branch U-Net++ semantic segmentation model; The oil stain mask is mapped onto the surface of the backup high-resolution image to construct the oil leak fusion image.

5. The current transformer oil leakage detection method based on machine vision as described in claim 2, characterized in that: The steps for constructing a 3D model of the oil leak based on the fused oil leak image and the intelligent patrol process include: The intelligent patrol location is confirmed based on the intelligent patrol process; Based on the location of the intelligent patrol, the corresponding 3D model is confirmed, and the corresponding 3D model is copied as a backup 3D model; The pixel coordinates of the oil leak marker in the oil leak fusion image are extracted as oil trace coordinates, and the oil trace coordinates are converted into 3D global oil trace coordinates; The 3D global oil trace coordinates are fused to the surface of the backup 3D model to construct the oil leak 3D model.

6. The current transformer oil leakage detection method based on machine vision as described in claim 1, characterized in that: The step of constructing a repair route based on the oil leak 3D model and the oil leak detection map includes: Based on the oil leak detection map, the location points of all the oil leak 3D models were confirmed; A repair route is constructed based on all the aforementioned location points and the oil leak detection map.

7. The current transformer oil leakage detection method based on machine vision as described in claim 4, characterized in that: The step of inputting the effective region and the device region into the internal output oil stain mask of the dual-branch U-Net++ semantic segmentation model includes: The effective region and the device region are simultaneously resized and aligned to obtain the standard effective region and the standard device region; The standard effective region is input into the infrared branch to obtain temperature difference characteristics; The standard device area is input into the high-definition branch to obtain texture morphology features; The contribution weights of the bimodal features are calculated based on the channel attention module, and the oil trace fusion features are calculated based on the contribution weights. The oil trace fusion feature is based on multi-scale skip connections and dense decoder block output of the initial oil trace mask; The initial oil stain mask is optimized through post-processing to obtain the oil stain mask.

8. The current transformer oil leakage detection method based on machine vision as described in claim 5, characterized in that: The step of converting the oil trail coordinates into 3D global oil trail coordinates includes: The oil trail coordinates are transformed into normalized oil trail coordinates based on the camera coordinate system; The depth value of the oil stain attachment location is extracted based on the metadata of the 3D model; The normalized oil trail coordinates and the depth value are transformed into 3D oil trail coordinates based on the camera coordinate system; The 3D oil trail coordinates are converted into the 3D global oil trail coordinates based on the corresponding rotation matrix and translation vector.

9. The current transformer oil leakage detection method based on machine vision as described in claim 5, characterized in that: The training steps of the dual-branch U-Net++ semantic segmentation model are as follows: Collect 1000 sets of dual-modal oil leakage images of substations and mark oil trace masks on the surface of the high-resolution images to form a dataset; The dataset is divided into training set, validation set, and test set according to the specified proportions. A dual-modal synchronous enhancement training set is used to align the high-resolution icon annotations in the training set at the pixel level. The training set is input into the dual-branch U-Net++ semantic segmentation model to train the model; Input the validation set into the dual-branch U-Net++ semantic segmentation model to avoid over-convergence; The test set was input into the dual-branch U-Net++ semantic segmentation model to evaluate the model's capabilities.

10. A machine vision-based current transformer oil leakage detection system, said system being used to implement the method as described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire substation maps, acquire and recognize images, acquire 3D data maps of current transformers, and acquire the intelligent patrol process of the intelligent patrol vehicle. The image integration module is used to construct an oil leak detection map based on the substation map and to construct 3D models of each current transformer based on the current transformer 3D data map. The feature extraction module is used to obtain the oil leak fusion image based on the identified image and the dual-branch U-Net++ semantic segmentation model; The feature fusion module is used to construct a 3D model of the oil leak based on the fused oil leak image and the intelligent patrol process; The route construction module is used to construct a repair route based on the oil leak 3D model and the oil leak detection map.

Images