Transformer area fault positioning method and device based on machine vision

Abstract

The invention discloses a transformer area fault positioning method and device based on machine vision, and relates to the technical field of low-voltage transformer area fault detection. The method comprises the following steps: acquiring original image data of a low-voltage distribution area; screening similar pixel blocks from the original image data; stacking a target pixel block and the corresponding similar pixel block into a 3D block group; processing the 3D block group to obtain an initial polarization image; calculating to obtain an enhanced image according to the original polarization image and the target parameter, constructing a target detection model, and inputting the image into the target detection model to obtain a target result; by collecting the original image data of the low-voltage distribution area, the equipment fault is accurately positioned and the fault probability is evaluated by using the target detection model in a low-visibility environment after enhancement processing, so that timely early warning of the fault and cloud uploading of a detection report are realized, and the operation and maintenance efficiency and the fault response speed of the power equipment in the area are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of low-voltage distribution area fault detection, and particularly relates to a low-voltage distribution area fault positioning method and device based on machine vision. BACKGROUND

[0002] A low-voltage distribution area is a power supply area of a low-voltage distribution network in a power system, which is usually composed of a distribution transformer, a low-voltage distribution box, a low-voltage line, and related protection equipment to provide low-voltage power supply to users within a certain range. The low-voltage distribution area is mainly responsible for converting high-voltage power into low-voltage power suitable for household, commercial, and small industrial use, usually 220 volts or 380 volts. It is the last end of power transmission, and its operating state directly affects the power supply quality and reliability of users.

[0003] Common faults of a low-voltage distribution area include transformer overload burnout, line short circuit / disconnection, switch device aging failure, electric leakage, and joint oxidation, line insulation layer damage, etc. These faults can cause user power outage, unstable voltage, and even electric shock or fire risk. In low-visibility environments (such as night, fog, heavy rain), manual inspection efficiency is low, and images of power equipment collected by cameras appear blurred, overall dark, and details are lost, so that abnormal heating of area equipment, arc spark, and insulator damage are not easy to be found, making it difficult to accurately detect equipment failure and locate the fault point, which may result in misjudgment and delay of repair, exacerbating equipment damage and causing economic losses. SUMMARY

[0004] The purpose of the present application is to solve the problem of inaccurate fault detection of a low-voltage distribution area in a low-visibility environment, and to propose a low-voltage distribution area fault positioning method based on machine vision.

[0005] In the first aspect of the present application, a low-voltage distribution area fault positioning method based on machine vision is first proposed, which comprises:

[0006] Obtaining original image data of a low-voltage distribution area, and screening similar pixel blocks from the original image data; the original image data includes original polarization images and infrared images of different polarization angles;

[0007] Stacking a target pixel block and its corresponding similar pixel block into a 3D block group, processing the 3D block group to obtain an initial polarization image, extracting the gray value of the initial polarization image and calculating a target parameter, and calculating an enhanced image according to the original polarization image and the target parameter; the target parameter includes Stokes parameters, transmittance, polarization angle, and atmospheric light value;

[0008] Fusing the enhanced image and the infrared image to obtain a fused image;

[0009] A target detection model is constructed, and the fused image is input into the target detection model to obtain the target result; the target result includes the probability of device failure and the location of the failure;

[0010] Based on the target results, an early warning is issued and a detection report is generated and uploaded to the cloud server.

[0011] Optionally, the original image data is filtered to select similar pixel blocks, and the original polarized images with different polarization angles include those with a polarization angle of 1. The original polarization image The polarization angle is The original polarization image and polarization angle is The original polarization image ,include:

[0012] The original polarization image of the target is segmented into n×n pixel blocks; the original polarization image of the target is the original polarization image. Original polarization image and the original polarization image any one of them;

[0013] The target pixel block in the original polarization image of the target is compared with other pixel blocks to obtain an error value. Similar pixel blocks are then selected by comparing the error value with a threshold. The target pixel block is any one of n×n pixel blocks.

[0014] Each target pixel block and each similar pixel block are combined to form a pixel block set; the target pixel block and the similar pixel block correspond one-to-one; the pixel block set contains multiple pixel block groups, and each pixel block group is composed of target pixel blocks and similar pixel blocks;

[0015] An enhanced image is obtained by performing a preset operation on the target pixel block and its corresponding similar pixel blocks in the pixel block set.

[0016] Optionally, an enhanced image is obtained by performing a preset operation based on the target pixel block of the pixel block set and its corresponding similar pixel blocks, including:

[0017] The target pixel block is stacked with its corresponding similar pixel block to obtain a 3D block group. The discrete cosine transform is performed on the 3D block group to convert the spatial domain 3D block group into a frequency domain 3D block group. The filtered 3D block group is obtained by performing hard threshold filtering and Wiener filtering operations on the frequency domain 3D block group.

[0018] Perform an inverse discrete cosine transform on the filtered 3D frequency domain blocks to convert the frequency domain 3D blocks back into spatial domain 3D blocks. Then, fuse multiple 3D blocks by weighted averaging and finally use the inverse polarization transformation matrix to generate a denoised initial polarization image. , the initial polarization image , the initial polarization image ;

[0019] traversing each pixel of the image, extracting the gray value in the initial polarization image , the initial polarization image , the initial polarization image , the Stokes parameter is calculated according to the gray value;

[0020] The transmittance is calculated according to the Stokes parameter, and the pixel polarization angle is calculated according to the initial polarization image and the Stokes parameter, and the pixel region with a polarization angle not less than 80° and not greater than 100° is screened out, and the initial polarization image The mean value of the light intensity value is taken as the global atmospheric light value;

[0021] The enhanced image is calculated according to the original polarization image , the transmittance and the atmospheric light value.

[0022] Optionally, the target detection model is based on the improvement of the YOLOv8 model, comprising:

[0023] The C2f module of the 2nd layer, the 4th layer, the 6th layer and the 8th layer of the backbone network in the YOLOv8 model is replaced by a C2f_SEBlockV module; the C2f_SEBlockV module is used to enhance the small infrared target feature extraction capability of the input enhanced image of the backbone network;

[0024] The C2f module of the 12th layer, the 15th layer, the 18th layer and the 21st layer of the neck network in the YOLOv8 model is replaced by a C2f_SEBlockV module;

[0025] One WultiSEAM module is added in front of each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is used to reduce the image background noise in the head network;

[0026] The WultiSEAM module is connected with the C2f_SEBlockV module of the 15th layer, the 18th layer and the 21st layer in the improved neck network.

[0027] Optionally, the C2f_SEBlockV module is improved based on the C2f module, wherein:

[0028] The DarknetBottleneck module in the C2f module is replaced by an SEBlockV module;

[0029] The working principle of the SEBlockV module is:

[0030] The input image is taken as a first feature map, the first feature map is sequentially input into a Conv module, a DWConv module and a SELayer module to obtain a first channel feature map, the first feature map and the first channel feature map are fused to obtain a second channel feature map and output;

[0031] The working principle of the WultiSEAM module is:

[0032] The input image is taken as a first original feature map, the first original feature map is input into three CSMM modules respectively to obtain a first convolution feature map, a second convolution feature map and a third convolution feature map;

[0033] The first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map are fused to obtain a fourth convolution feature map, and the fourth convolution feature map is sequentially subjected to average pooling and full connection processing to obtain a fifth convolution feature map and output.

[0034] In the second aspect of the implementation of the present application, a machine vision-based transformer area fault positioning device is provided, which comprises a screening module, an enhanced calculation module, an image fusion module, a model construction module and an information transmission module, wherein:

[0035] The screening module is used for acquiring original image data of a low-voltage distribution transformer area, and screening similar pixel blocks from the original image data; the original image data comprises original polarization images and infrared images of different polarization angles;

[0036] The enhanced calculation module is used for stacking a target pixel block and its corresponding similar pixel block into a 3D block group, processing the 3D block group to obtain an initial polarization image, extracting a gray value of the initial polarization image and calculating a target parameter, and calculating an enhanced image according to the original polarization image and the target parameter; the target parameter comprises Stokes parameters, transmittance, a polarization angle and an atmospheric light value;

[0037] The image fusion module is used for fusing and processing the enhanced image and the infrared image to obtain a fused image;

[0038] The model construction module is used for constructing a target detection model, inputting the fused image into the target detection model to obtain a target result; the target result comprises a position of a device fault and a probability of the device fault;

[0039] The information transmission module is used for prewarning according to the target result and generating a detection report uploaded to a cloud server.

[0040] Optionally, the screening module comprises a division module, an error calculation module and a processing module, wherein:

[0041] The division module is configured to divide the target original polarized image to obtain n×n pixel blocks; the target original polarized image is any one of an original polarized image , an original polarized image and an original polarized image ;

[0042] The error calculation module is configured to calculate a target pixel block in the target original polarized image with other pixel blocks to obtain an error value, and screen out similar pixel blocks by comparing the error value with a threshold value; the target pixel block is any one of the n×n pixel blocks;

[0043] The combination module is configured to combine each target pixel block and each similar pixel block to obtain a pixel block group set; the target pixel block and the similar pixel block are in one-to-one correspondence; the pixel block group set comprises a plurality of pixel block groups, and each pixel block group is composed of a target pixel block and a similar pixel block;

[0044] The processing module is configured to perform a preset operation on the target pixel block of the pixel block group set and the corresponding similar pixel block to obtain an enhanced image.

[0045] Optionally, the enhancement calculation module comprises a filtering module, a denoising module, a first calculation module, a second calculation module and a third calculation module, wherein:

[0046] The filtering module is configured to stack a target pixel block and a corresponding similar pixel block to obtain a 3D block group, perform a discrete cosine transform on the 3D block group, convert the 3D block group in a spatial domain into a 3D block group in a frequency domain, and perform a hard threshold filtering and a Wiener filtering operation on the 3D block group in the frequency domain to obtain a filtered 3D block group;

[0047] The denoising module is configured to perform an inverse discrete cosine transform on the filtered 3D frequency domain block group, convert the 3D block group in the frequency domain back to a 3D block group in the spatial domain, fuse a plurality of 3D block groups by weighted average, and generate an initial polarized image , an initial polarized image and an initial polarized image after denoising by using an inverse polarized transformation matrix;

[0048] The first calculation module is configured to traverse each pixel of the image, extract a gray value in the initial polarized image , an initial polarized image and an initial polarized image , and calculate a Stokes parameter according to the gray value;

[0049] The second calculation module is configured to calculate the transmittance according to the Stokes parameters, and calculate the initial polarization image by using the transmittance and the Stokes parameters to calculate the pixel polarization angle, and screen out a pixel region with a polarization angle close to π / 2, and extract the initial polarization image in the pixel region The mean value of the light intensity value is taken as the global atmospheric light value.

[0050] The third calculation module is configured to calculate an enhanced image according to the original polarization image , the transmittance and the atmospheric light value.

[0051] Optionally, the model construction module comprises:

[0052] The target detection model is based on an improved YOLOv8 model, wherein:

[0053] The C2f modules in the 2nd layer, the 4th layer, the 6th layer and the 8th layer of the backbone network in the YOLOv8 model are replaced by C2f_SEBlockV modules; the C2f_SEBlockV modules are configured to enhance the small infrared target feature extraction capability of the input enhanced image in the backbone network;

[0054] The C2f modules in the 12th layer, the 15th layer, the 18th layer and the 21st layer of the neck network in the YOLOv8 model are replaced by C2f_SEBlockV modules;

[0055] One WultiSEAM module is added in front of each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is configured to reduce the image background noise in the head network;

[0056] The WultiSEAM module is connected with the C2f_SEBlockV modules in the 15th layer, the 18th layer and the 21st layer of the replaced neck network.

[0057] Optionally, the C2f_SEBlockV module is based on an improved C2f module, wherein:

[0058] The DarknetBottleneck module in the C2f module is replaced by an SEBlockV module;

[0059] The working principle of the SEBlockV module is as follows:

[0060] An input image is taken as a first feature map, the first feature map is sequentially input into a Conv module, a DWConv module and an SELayer module to obtain a first channel feature map, and the first feature map and the first channel feature map are fused to obtain a second channel feature map and output.

[0061] The working principle of the WultiSEAM module is as follows:

[0062] An input image is taken as a first original feature map, and the first original feature map is input into three CSMM modules respectively to obtain a first convolution feature map, a second convolution feature map and a third convolution feature map.

[0063] The first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map are fused to obtain a fourth convolution feature map, and the fourth convolution feature map is sequentially subjected to average pooling and full connection processing to obtain a fifth convolution feature map and output.

[0064] The beneficial effects of the present application are as follows:

[0065] The present application provides a transformer area fault positioning method based on machine vision, which collects original image data of low-voltage distribution transformer area, processes and fuses infrared images, uses a target detection model to accurately locate equipment faults in low-visibility environments and evaluate fault probability, realizes timely early warning and cloud uploading of detection reports, solves the problem of low efficiency of artificial inspection and low image quality of cameras, effectively improves the operation and maintenance efficiency and fault response speed of transformer area power equipment. BRIEF DESCRIPTION OF DRAWINGS

[0066] The present application will be further described below in conjunction with the drawings.

[0067] Figure 1 A flowchart of a transformer area fault positioning method based on machine vision provided by the present application embodiment is shown in the figure.

[0068] Figure 2 A framework diagram of a transformer area fault positioning device based on machine vision provided by the present application embodiment is shown in the figure.

[0069] Figure 3 A structural diagram of a transformer area fault positioning method model based on machine vision provided by the present application embodiment is shown in the figure.

[0070] Figure 4 A schematic diagram of a WultiSEAM module provided by the present application embodiment is shown in the figure.

[0071] Figure 5 A schematic diagram of a C2f_SEBlockV module provided by the present application embodiment is shown in the figure.

[0072] Figure 6 A schematic diagram of a SEBlockV module provided by the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. The term "and / or" in this document is only used to describe the association relationship of associated objects, and can represent three relationships, for example, A and B can represent three cases of existence of A alone, existence of A and B at the same time, and existence of B alone. In addition, the description of "first", "second" and the like in the present application is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that a person skilled in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope of the present application.

[0074] Based on the embodiments in the present application, all other embodiments obtained by a person skilled in the art without creative labor are within the protection scope of the present application.

[0075] The embodiment of the present application provides a low-voltage distribution area fault positioning method based on machine vision. Figure 1 , Figure 1 The embodiment of the present application provides a flow chart of a low-voltage distribution area fault positioning method based on machine vision. The method comprises the following steps:

[0076] S101, acquiring original image data of a low-voltage distribution area, and screening similar pixel blocks from the original image data;

[0077] S102, stacking the target pixel block and its corresponding similar pixel block into a 3D block group, processing the 3D block group to obtain an initial polarization image, extracting a gray value of the initial polarization image and calculating a target parameter, and calculating an enhanced image according to the original polarization image and the target parameter;

[0078] S103, fusing the enhanced image and an infrared image to obtain a fusion image;

[0079] S104, constructing a target detection model, inputting the fusion image into the target detection model to obtain a target result;

[0080] S105, performing early warning according to the target result and generating a detection report uploaded to a cloud server.

[0081] The original image data comprises original polarization images of different polarization angles and infrared images.

[0082] The target parameters include Stokes parameters, transmittance, polarization angle, and atmospheric light value.

[0083] The target results include the probability of equipment failure and the location of the failure.

[0084] The present invention provides a machine vision-based method for locating transformer substation faults. By collecting raw image data of low-voltage distribution substations, enhancing and fusing infrared images, and using a target detection model, the method accurately locates equipment faults and assesses the probability of faults in low-visibility environments. This enables timely early warning of faults and cloud-based uploading of detection reports, effectively improving the operation and maintenance efficiency and fault response speed of power equipment in the substation.

[0085] In one implementation, the polarization angle refers to the angle between the vibration direction of linearly polarized light and the reference direction, which is usually horizontal, corresponding to 0°, and ranges from... The raw image data was acquired using a binocular camera.

[0086] In one implementation, a detection report is generated and uploaded to a cloud database based on the probability of power facility failure and its location within the distribution network area in the target result. If the probability of power facility failure in the detection report exceeds the preset failure probability, an alarm is triggered to remind staff to check.

[0087] In one implementation, the enhanced image and the infrared image are fused to obtain a fused image. The specific process includes:

[0088] A multi-scale decomposition method is used to decompose the enhanced image and the infrared image into sub-band sequences of different scales; the multi-scale decomposition method is the Laplacian pyramid method.

[0089] For the low-frequency subband that characterizes the image contour, a weighted averaging strategy is adopted to evenly preserve the background information of the enhanced image and the basic energy of the thermal target in the infrared image.

[0090] For high-frequency subbands containing details and edges, the "largest absolute value" fusion rule is adopted. The fusion rule is: prioritize the position with the largest gray-level gradient difference between corresponding positions in the two images, so as to fuse the texture details of the enhanced image with the prominent thermal target contour in the infrared image and reconstruct the fused image.

[0091] In one embodiment, similar pixel blocks are filtered from the original image data, and the original polarized images with different polarization angles include those with polarization angles of 1. The original polarization image The polarization angle is The original polarization image and polarization angle is The original polarization image ,include:

[0092] segmenting the target raw polarization image to obtain n×n pixel blocks; the target raw polarization image is any one of a raw polarization image , a raw polarization image , and a raw polarization image ;

[0093] calculating a target pixel block in the target raw polarization image and other pixel blocks to obtain an error value, and screening similar pixel blocks by comparing the error value with a threshold value; the target pixel block is any one of the n×n pixel blocks;

[0094] combining each target pixel block and each similar pixel block to obtain a pixel block group set; the target pixel block and the similar pixel block are in one-to-one correspondence; the pixel block group set includes a plurality of pixel block groups, and each pixel block group is composed of a target pixel block and a similar pixel block;

[0095] performing a preset operation on the target pixel block and the corresponding similar pixel block in the pixel block group set to obtain an enhanced image.

[0096] In an implementation manner, n×n can be set by human, and segmentation facilitates processing of a single pixel block; the reason for taking three polarization angle images of 0°, 45°, and 90° is that the Stokes parameters can be calculated, the key directions of linear polarization and orthogonality are covered to maximize the polarization difference between scattered light and target reflected light, the standardized operation of the polarization imaging device is met, and the algorithm denoising and filtering are supported.

[0097] In an implementation manner, the error value is obtained by calculating the target pixel block in the target raw polarization image and other pixel blocks, and the process is as follows:

[0098]

[0099] wherein, E is the error value, N is the number of pixels matched between the target pixel block and other pixel blocks, represents the i th target pixel block in a raw polarization image , represents the i th pixel block other than the target pixel block in a raw polarization image , represents the i th target pixel block in a raw polarization image , represents the i th pixel block other than the target pixel block in a raw polarization image , represents the i th target pixel block in a raw polarization image , represents the i th pixel block other than the target pixel block in a raw polarization image ;

[0100] The threshold is obtained by historical error experimental data, and candidate pixel blocks with E less than the threshold are screened as similar pixel blocks.

[0101] In an implementation, the enhanced image is generated by dividing the pixel blocks, matching and screening the similar blocks, and finally fusing the infrared image, so as to improve the image details and contrast, enhance the fault feature recognition capability, and provide clearer image for subsequent target detection.

[0102] In one embodiment, the enhanced image is obtained by performing a preset operation on the target pixel block and the corresponding similar pixel block of the pixel block group, including:

[0103] Stacking the target pixel block and the corresponding similar pixel block to obtain a 3D block group, performing discrete cosine transform on the 3D block group to convert the 3D block group in the spatial domain into a 3D block group in the frequency domain, and obtaining a filtered 3D block group by performing hard threshold filtering and Wiener filtering operations on the 3D block group in the frequency domain;

[0104] Performing inverse discrete cosine transform on the filtered 3D frequency block group to convert the 3D block group in the frequency domain back to a 3D block group in the spatial domain, fusing multiple 3D block groups by weighted average, and generating the initial polarized image after denoising by using the inverse polarization transformation matrix , the initial polarized image , and the initial polarized image ;

[0105] Traversing each pixel of the image, extracting the gray value in the initial polarized image , the initial polarized image , and the initial polarized image , and calculating the Stokes parameters according to the gray value;

[0106] Calculating the transmittance according to the Stokes parameters, calculating the pixel polarization angle according to the initial polarized image and the Stokes parameters, and screening a pixel region with a polarization angle not less than 80° and not greater than 100°, and extracting the mean value of the light intensity value in the pixel region as the global atmospheric light value;

[0107] Calculating the enhanced image according to the original polarized image , the transmittance, and the atmospheric light value.

[0108] In one implementation, the Stokes parameters are calculated according to the gray value, and the calculation process is as follows:

[0109]

[0110]

[0111] ​

[0112] wherein, represents a first Stokes parameter of pixel coordinates at represents a second Stokes parameter of pixel coordinates at represents a third Stokes parameter of pixel coordinates at is a pixel value of the initial polarization image at is a pixel value of the initial polarization image at is a pixel value of the initial polarization image at

[0113] In an implementation, the transmittance is calculated according to the Stokes parameters, and the pixel polarization angle is calculated according to the initial polarization image and the Stokes parameters, and the specific process is as follows:

[0114]

[0115]

[0116]

[0117]

[0118] wherein, represents a degree of polarization of pixel coordinates at , is a transmittance of pixel coordinates at is a polarization angle of pixel coordinates at A is an atmospheric light value, R is a pixel region in which the polarization angle is not less than 80° and not greater than 100°, and K represents a total number of pixels in the pixel region.

[0119] In an implementation, the enhanced image is calculated according to the original polarization image , the transmittance, and the atmospheric light value, and the process is as follows:

[0120]

[0121] wherein, represents an enhanced pixel value at coordinates, represents an enhanced pixel value at ​​​​​​​​pixel value of the original polarization image at the coordinates, for the atmospheric light value;

[0122] The enhanced image is obtained by combining reconstruction.

[0123] In an implementation manner, the 3D block group is denoised by a discrete cosine transform and a filtering operation to retain image details, the 3D block group is fused by a weighted average to generate a denoised polarization image, the transmittance and the polarization angle are calculated through Stokes parameters, the atmospheric light value is extracted, and finally the enhanced image is generated in combination with the original polarization image to improve the image quality and provide accurate data as support for an input target detection module.

[0124] In one embodiment, the target detection model is improved based on a YOLOv8 model, and the improvement includes:

[0125] The C2f module of the 2nd layer, the 4th layer, the 6th layer and the 8th layer of the backbone network in the YOLOv8 model is replaced by a C2f_SEBlockV module; the C2f_SEBlockV module is used to enhance the small infrared target feature extraction capability of the input enhanced image of the backbone network;

[0126] The C2f module of the 12th layer, the 15th layer, the 18th layer and the 21st layer of the neck network in the YOLOv8 model is replaced by a C2f_SEBlockV module;

[0127] One WultiSEAM module is added in front of each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is used to reduce the image background noise in the head network.

[0128] The WultiSEAM module is connected with the C2f_SEBlockV module of the 15th layer, the 18th layer and the 21st layer in the improved neck network.

[0129] In an implementation manner, referring to Figure 3 , Figure 3 A structure diagram of a method model for locating a fault in a transformer area based on machine vision is provided in an embodiment of the application. The target detection model core is used to solve the key problems of serious background interference and insufficient multi-scale feature extraction in infrared small target detection, maintains strong robustness under serious occlusion and low signal-to-noise ratio conditions, and meets the efficient and real-time requirements of infrared small target detection.

[0130] In an implementation manner, referring to Figure 4 , Figure 4 ​A schematic diagram of the WultiSEAM module is provided for the embodiment of the present application, the WultiSEAM module is located in the head network, and is mainly used to solve the problems of unobvious features, background noise interference and difficult identification of small target features caused by slight occlusion of the target in infrared small target detection.

[0131] In one embodiment, the C2f_SEBlockV module is improved based on the C2f module, wherein:

[0132] The DarknetBottleneck module in the C2f module is replaced by the SEBlockV module;

[0133] The working principle of the SEBlockV module is as follows:

[0134] The input image is taken as the first feature map, the first feature map is sequentially input into the Conv module, the DWConv module and the SELayer module to obtain the first channel feature map, and the first feature map and the first channel feature map are fused to obtain the second channel feature map and output;

[0135] The working principle of the WultiSEAM module is as follows:

[0136] The input image is taken as the first original feature map, and the first original feature map is input into three CSMM modules to obtain the first convolution feature map, the second convolution feature map and the third convolution feature map;

[0137] The first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map are fused to obtain the fourth convolution feature map, and the fourth convolution feature map is sequentially subjected to average pooling and full connection processing to obtain the fifth convolution feature map and output.

[0138] In one implementation manner, referring to Figure 5 , Figure 5 A schematic diagram of the C2f_SEBlockV module is provided for the embodiment of the present application; the main function of the C2f_SEBlockV module is to enhance the feature extraction capability of the network in the spatial and channel dimensions, help the network to more accurately locate and detect small infrared targets in complex backgrounds, reduce the problems of missed detection and false detection caused by insufficient feature extraction, and lay a good foundation for subsequent multi-scale feature fusion; the DarknetBottleneck module, the Conv module, the DWConv module and the SELayer module are all existing modules.

[0139] In one implementation manner, referring to Figure 6 , Figure 6The schematic diagram of the SEBlockV module is provided for the embodiments of the present invention. The function of SEBlockV is to optimize the feature channel weight allocation and retain key spatial details to improve the detection accuracy of small infrared targets, while avoiding the loss of subtle spatial details (such as target edges and temperature difference features) in deep networks, and enhancing the efficiency of information transmission and gradient propagation.

[0140] In one implementation, the object detection model is tested using the UAVPowerDefectDataset dataset. The object detection model is developed in the CUDA 11.8 and PyTorch 2.0 environment, and the detection model is trained on a Windows 11 system equipped with an NVIDIA GeForce RTX 4070 Ti Super graphics card.

[0141] Table 1. Detection results of different fault types based on the UAVPowerDefectDataset dataset.

[0142] Category YOLOv8 Target Model Accuracy Recall mAP@0.5 Accuracy Recall mAP@0.5 Fire 0.963 0.931 0.935 0.983 0.944 0.941 Electric leakage 0.852 0.855 0.861 0.889 0.871 0.879 Device heating 0.957 0.963 0.953 0.973 0.974 0.962 Line damage 0.839 0.853 0.843 0.872 0.852 0.864 Average 0.903 0.901 0.898 0.932 0.910 0.912

[0143] Table 1 shows that the target detection model has improved the accuracy, recall, and mAP@0.5 of accident type identification compared to the previous YOLOv8. Therefore, the target detection model is more accurate and efficient in identifying fault types of low-voltage distribution transformer equipment in low-visibility lighting environments.

[0144] Based on the same inventive concept, this invention also provides a machine vision-based fault location device for transformer substations. See also Figure 2 , Figure 2 This invention provides a framework diagram of a machine vision-based fault location device for transformer substations. The device includes a screening module, an enhancement calculation module, an image fusion module, a model building module, and an information transmission module, wherein:

[0145] The filtering module is used to acquire the raw image data of the low-voltage distribution radio station area and filter similar pixel blocks from the raw image data; the raw image data includes raw polarized images and infrared images with different polarization angles;

[0146] The enhancement calculation module is used to stack the target pixel block and its corresponding similar pixel blocks into a 3D block group, process the 3D block group to obtain the initial polarization image, extract the gray values ​​of the initial polarization image and calculate the target parameters, and calculate the enhanced image based on the original polarization image and the target parameters; the target parameters include Stokes parameters, transmittance, polarization angle and atmospheric light value;

[0147] The image fusion module is used to fuse the enhanced image with the infrared image to obtain a fused image;

[0148] The model construction module is configured to construct a target detection model, input the fused image into the target detection model, and obtain a target result; the target result includes a location of a device failure and a probability of the failure;

[0149] The information transmission module is configured to perform early warning according to the target result and generate a detection report uploaded to a cloud server.

[0150] The low-voltage distribution area fault positioning device based on machine vision provided by the embodiment of the application can accurately locate a device failure and evaluate a failure probability in a low-visibility environment by collecting original image data of a low-voltage distribution area, performing enhancement processing and fusing an infrared image, realizing timely early warning of the failure and cloud uploading of a detection report, and effectively improving the operation and maintenance efficiency and failure response speed of the low-voltage distribution area power device.

[0151] In one embodiment, the screening module includes a division module, an error calculation module, and a processing module, wherein:

[0152] The division module is configured to divide a target original polarization image to obtain n x n pixel blocks; the target original polarization image is any one of an original polarization image , an original polarization image , and an original polarization image .

[0153] The error calculation module is configured to calculate a target pixel block and other pixel blocks in the target original polarization image to obtain an error value, and screen out similar pixel blocks by comparing the error value with a threshold value; the target pixel block is any one of the n x n pixel blocks.

[0154] The combination module is configured to combine each target pixel block and each similar pixel block to obtain a pixel block group set; the target pixel block and the similar pixel block are in one-to-one correspondence; the pixel block group set includes a plurality of pixel block groups, and each pixel block group is composed of a target pixel block and a similar pixel block.

[0155] The processing module is configured to perform a preset operation on the target pixel block of the pixel block group set and the corresponding similar pixel block to obtain an enhanced image.

[0156] In one embodiment, the enhancement calculation module includes a filtering module, a denoising module, a first calculation module, a second calculation module, and a third calculation module, wherein:

[0157] The filtering module is configured to stack the target pixel block and the corresponding similar pixel block to obtain a 3D block group, perform a discrete cosine transform on the 3D block group, convert the 3D block group in a spatial domain into a 3D block group in a frequency domain, and perform a hard threshold filtering operation and a Wiener filtering operation on the 3D block group in the frequency domain to obtain a filtered 3D block group.

[0158] The de-noising module is configured to perform inverse discrete cosine transform on the filtered 3D frequency domain block group, convert the 3D block group in the frequency domain back to the 3D block group in the spatial domain, fuse multiple 3D block groups through weighted average, and generate the initial polarized image after de-noising by using the inverse polarization transformation matrix , the initial polarized image , and the initial polarized image

[0159] The first calculation module is configured to traverse each pixel of the image, extract the gray value in the initial polarized image , the initial polarized image , and the initial polarized image , and calculate the Stokes parameter according to the gray value;

[0160] The second calculation module is configured to calculate the transmittance according to the Stokes parameter, calculate the pixel polarization angle through the initial polarized image and the Stokes parameter, and screen out a pixel region in which the polarization angle is close to , extract the mean value of the light intensity value in the initial polarized image in the pixel region as the global atmospheric light value;

[0161] The third calculation module is configured to calculate the enhanced image according to the original polarized image , the transmittance, and the atmospheric light value.

[0162] In one embodiment, the model construction module comprises:

[0163] The target detection model is based on an improved YOLOv8 model, wherein:

[0164] The C2f modules in the second layer, the fourth layer, the sixth layer and the eighth layer of the backbone network in the YOLOv8 model are replaced by C2f_SEBlockV modules; the C2f_SEBlockV modules are used to enhance the small infrared target feature extraction capability of the input enhanced image in the backbone network;

[0165] The C2f modules in the twelfth layer, the fifteenth layer, the eighteenth layer and the twenty-first layer of the neck network in the YOLOv8 model are replaced by C2f_SEBlockV modules;

[0166] One WultiSEAM module is added in front of each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is used to reduce the image background noise in the head network;

[0167] The WultiSEAM module is connected with the C2f_SEBlockV modules in the fifteenth layer, the eighteenth layer and the twenty-first layer of the neck network after replacement.

[0168] ​In one embodiment, the C2f_SEBlockV module is improved based on the C2f module, wherein:

[0169] The DarknetBottleneck module in the C2f module is replaced by the SEBlockV module;

[0170] The SEBlockV module works as follows:

[0171] An input image is taken as a first feature map, the first feature map is sequentially input into a Conv module, a DWConv module and an SELayer module to obtain a first channel feature map, the first feature map and the first channel feature map are fused to obtain a second channel feature map and output;

[0172] The WultiSEAM module works as follows:

[0173] An input image is taken as a first original feature map, the first original feature map is input into three CSMM modules respectively to obtain a first convolution feature map, a second convolution feature map and a third convolution feature map;

[0174] The first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map are fused to obtain a fourth convolution feature map, the fourth convolution feature map is sequentially subjected to average pooling and full connection processing to obtain a fifth convolution feature map and output.

[0175] The above describes one embodiment of the present application in detail, but the content is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A machine vision-based method for locating a fault in a transformer area, characterized by, The method comprises: Obtaining original image data of a low-voltage power distribution area, and screening similar pixel blocks from the original image data; the original image data comprises original polarization images and infrared images of different polarization angles; Stacking a target pixel block and its corresponding similar pixel block into a 3D block group, processing the 3D block group to obtain an initial polarization image, extracting a gray value of the initial polarization image and calculating a target parameter, and calculating an enhanced image from the original polarization image and the target parameter; the target parameter comprises Stokes parameters, transmittance, polarization angles, and atmospheric light values; Fusing the enhanced image and the infrared image to obtain a fused image; Building a target detection model, inputting the fused image into the target detection model to obtain a target result; the target result comprises a probability of equipment failure and a location of failure; Performing early warning according to the target result and generating a detection report to be uploaded to a cloud server.

2. The machine vision-based fault location method for a transformer area according to claim 1, characterized in that, Screening similar pixel blocks from the original image data, the original polarization images of different polarization angles include the original polarization image of the original polarization image of the original polarization image of the original polarization image of and the original polarization image of the original polarization image of , comprising: segmenting the target original polarimetric image to obtain n×n pixel blocks; the target original polarimetric image is any one of an original polarimetric image , an original polarimetric image , and an original polarimetric image ​ Calculating an error value from the target pixel block and other pixel blocks in the target original polarization image, and screening similar pixel blocks by comparing the error value with a threshold value; the target pixel block is any one of n*n pixel blocks; Combining each target pixel block and each similar pixel block to obtain a pixel block group set; the target pixel block and the similar pixel block correspond to each other; the pixel block group set comprises a plurality of pixel block groups, and each pixel block group comprises a target pixel block and a similar pixel block; Performing a preset operation on the target pixel block and its corresponding similar pixel block in the pixel block group set to obtain an enhanced image. 3.The machine vision-based fault location method for a transformer area according to claim 2, characterized in that, Performing a preset operation on the target pixel block and its corresponding similar pixel block in the pixel block group set to obtain an enhanced image, comprising: Stacking the target pixel block and its corresponding similar pixel block to obtain a 3D block group, performing discrete cosine transform on the 3D block group to convert the 3D block group in the spatial domain into a 3D block group in the frequency domain, and performing hard threshold filtering and Wiener filtering operations on the 3D block group in the frequency domain to obtain a filtered 3D block group; performing inverse discrete cosine transform on the filtered 3D frequency domain block group to convert the 3D block group in frequency domain back to 3D block group in spatial domain, fusing multiple 3D block groups by weighted average, and generating the initial polarized image after denoising by using inverse polarized transform matrix , initial polarized image and initial polarized image ; traversing each pixel of the image to extract the initial polarization image , the initial polarization image and the initial polarization image , the gray value in the initial polarization image, and calculating the Stokes parameters according to the gray value The transmittance is calculated according to the Stokes parameters, and the initial polarization image is extracted from a pixel region with a polarization angle not less than 80° and not greater than 100° The polarization angle of the pixel is calculated according to the Stokes parameters, and a pixel region with a polarization angle not less than 80° and not greater than 100° is screened out, and the initial polarization image is extracted from the pixel region The mean value of the light intensity value is taken as the global atmospheric light value; According to the original polarization image The enhanced image is obtained by calculating the transmittance and the atmospheric light value.

4. The machine vision-based fault location method for a transformer area according to claim 1, characterized in that, The target detection model is based on an improved YOLOv8 model, comprising: Replacing C2f modules in the second layer, the fourth layer, the sixth layer, and the eighth layer of a backbone network in the YOLOv8 model with C2f_SEBlockV modules; the C2f_SEBlockV modules are used to enhance the small infrared target feature extraction capability of the backbone network input enhanced image; Replacing C2f modules in the twelfth layer, the fifteenth layer, the eighteenth layer, and the twenty-first layer of a neck network in the YOLOv8 model with C2f_SEBlockV modules; Adding one WultiSEAM module before each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is used to reduce image background noise in the head network; The WultiSEAM module is connected to the C2f_SEBlockV modules in the fifteenth layer, the eighteenth layer, and the twenty-first layer of the improved neck network.

5. The machine vision-based fault location method for a transformer area according to claim 4, characterized in that, The C2f_SEBlockV module is improved based on the C2f module, wherein: Replace the DarknetBottleneck module in the C2f module with an SEBlockV module; The SEBlockV module works as follows: Take the input image as a first feature map, input the first feature map into a Conv module, a DWConv module and an SELayer module in sequence to obtain a first channel feature map, and fuse the first feature map and the first channel feature map to obtain a second channel feature map and output the second channel feature map; The working principle of the WultiSEAM module is as follows: Take the input image as a first original feature map, input the first original feature map into three CSMM modules respectively to obtain a first convolution feature map, a second convolution feature map and a third convolution feature map; Fuse the first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map to obtain a fourth convolution feature map, and perform average pooling and full connection processing on the fourth convolution feature map in sequence to obtain a fifth convolution feature map and output the fifth convolution feature map.

6. A machine vision-based substation fault locating device, characterized by, The device comprises a screening module, an enhanced calculation module, an image fusion module, a model construction module and an information transmission module, wherein: The screening module is configured to obtain original image data of a low-voltage distribution area, and screen similar pixel blocks from the original image data; the original image data includes original polarization images and infrared images of different polarization angles; The enhanced calculation module is configured to stack target pixel blocks and their corresponding similar pixel blocks into 3D block groups, process the 3D block groups to obtain initial polarization images, extract gray values of the initial polarization images and calculate target parameters, and calculate enhanced images based on the original polarization images and the target parameters; the target parameters include Stokes parameters, transmittance, polarization angles and atmospheric light values; The image fusion module is configured to fuse the enhanced images and the infrared images to obtain a fusion image; The model construction module is configured to construct a target detection model, input the fusion image into the target detection model to obtain a target result; the target result includes the location of a device failure and the probability of the device failure; The information transmission module is configured to perform early warning based on the target result and generate a detection report to be uploaded to a cloud server.

7. The machine vision-based substation fault locating apparatus according to claim 6, wherein, The screening module comprises a division module, an error calculation module and a processing module, wherein: The division module is configured to divide the target original polarization image to obtain n×n pixel blocks; the target original polarization image is any one of an original polarization image , an original polarization image , and an original polarization image ​ The error calculation module is configured to calculate error values of target pixel blocks in the target original polarization image and other pixel blocks, and screen similar pixel blocks by comparing the error values with a threshold value; a target pixel block is any one of n×n pixel blocks; The combination module is configured to combine each target pixel block and each similar pixel block to obtain a pixel block group set; a target pixel block and a similar pixel block correspond to each other; the pixel block group set includes a plurality of pixel block groups, and each pixel block group is composed of a target pixel block and a similar pixel block; The processing module is configured to perform a preset operation on the target pixel blocks and their corresponding similar pixel blocks in the pixel block group set to obtain an enhanced image. 8.The machine vision-based fault location device for a transformer area according to claim 6, wherein, The enhanced computing module comprises a filtering module, a denoising module, a first computing module, a second computing module and a third computing module, wherein: The filtering module is configured to stack a target pixel block and its corresponding similar pixel block to obtain a 3D block group, perform a discrete cosine transform on the 3D block group to convert the 3D block group in a spatial domain into a 3D block group in a frequency domain, and obtain a filtered 3D block group by performing hard threshold filtering and Wiener filtering operations on the 3D block group in the frequency domain; The denoising module is configured to perform inverse discrete cosine transform on the filtered 3D frequency domain block group, convert the 3D block group in the frequency domain back to a 3D block group in the spatial domain, fuse a plurality of 3D block groups through weighted average, and generate an initial polarized image after denoising by using an inverse polarization transformation matrix , the initial polarized image , and the initial polarized image ; The first calculation module is used to traverse each pixel of the image and extract the initial polarization image. Initial polarization image and initial polarization image The grayscale values ​​in the image are used to calculate the Stokes parameters. The second calculation module is configured to calculate the transmittance according to the Stokes parameters, and calculate the pixel polarization angle according to the Stokes parameters , and screen out a pixel region with a polarization angle close to 0° or 180°, and extract the initial polarization image in the pixel region , and take the mean value of the light intensity values as the global atmospheric light value . The third calculation module is used to calculate based on the original polarization image. The enhanced image is obtained by calculating the transmittance and the atmospheric light value.

9. The machine vision-based substation fault locating apparatus of claim 6, wherein, The model construction module comprises: The target detection model is based on an improved YOLOv8 model, wherein: The C2f modules in the second layer, the fourth layer, the sixth layer and the eighth layer of the backbone network in the YOLOv8 model are replaced by C2f_SEBlockV modules; the C2f_SEBlockV modules are configured to enhance the small infrared target feature extraction capability of the input enhanced image of the backbone network; The C2f modules in the twelfth layer, the fifteenth layer, the eighteenth layer and the twenty-first layer of the neck network in the YOLOv8 model are replaced by C2f_SEBlockV modules; One WultiSEAM module is added in front of each of the three detection heads in the head network in the YOLOv8 model; the WultiSEAM module is configured to reduce image background noise in the head network; The WultiSEAM module is connected to the fifteenth layer of the replaced neck network, the C2f_SEBlockV modules in the eighteenth layer and the twenty-first layer.

10. The machine vision-based substation fault locating apparatus of claim 9, wherein, The C2f_SEBlockV module is based on an improved C2f module, wherein: The DarknetBottleneck module in the C2f module is replaced by an SEBlockV module; The working principle of the SEBlockV module is as follows: An input image is taken as a first feature map, the first feature map is sequentially input into a Conv module, a DWConv module and an SELayer module to obtain a first channel feature map, and the first feature map and the first channel feature map are fused to obtain a second channel feature map and output; The working principle of the WultiSEAM module is as follows: An input image is taken as a first original feature map, and the first original feature map is input into three CSMM modules to obtain a first convolution feature map, a second convolution feature map and a third convolution feature map; The first original feature map, the first convolution feature map, the second convolution feature map and the third convolution feature map are fused to obtain a fourth convolution feature map, and the fourth convolution feature map is sequentially subjected to average pooling and full connection processing to obtain a fifth convolution feature map and output.

Images