A method, system, device and medium for identifying oil-paper insulation defects in high-voltage equipment

Abstract

The present invention relates to the technical field of high-voltage equipment electrical variable detection, and specifically to a method, system, device and medium for identifying oil-paper insulation defects in high-voltage equipment. In order to solve the technical problem of low accuracy in identifying oil-paper insulation defects inside a transformer, the present invention first obtains images of several positions inside the high-voltage equipment by performing image registration and pixel fusion processing of repeated areas to obtain a mosaic image inside the high-voltage equipment; then, different depth features of the mosaic image inside the high-voltage equipment are extracted, and feature cross-fusion processing of bottom-up fusion and top-down fusion is performed to obtain several cross-fusion features; finally, the several cross-fusion features are spliced and segmented to obtain an oil-paper insulation defect identification result; the present invention is used in the process of detecting electrical variables inside high-voltage equipment, can improve the accuracy of identifying oil-paper insulation defects, and improve the accuracy of detecting electrical variables of high-voltage equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical variable detection of high-voltage equipment, and in particular to a method, system, equipment and medium for identifying oil-paper insulation defects of high-voltage equipment. Background Art

[0002] Detecting oil-paper insulation defects during electrical variable monitoring of high-voltage equipment is crucial for preventing power system accidents. In actual operation, the metal-sealed design of large transformers makes effective internal inspection difficult without disassembling them. Currently, existing detection technologies are unable to visually detect internal transformer faults, posing a significant challenge to their safe operation.

[0003] To address this issue, transformer internal inspection robots have become a key research area in the field of internal fault detection. Equipped with high-precision sensors and imaging equipment, these robots can conduct in-depth inspections without disassembling the transformer. Their application not only improves inspection efficiency but also reduces the risk of manual intervention. However, internal inspection robots still face the challenge of accurately and quickly identifying and segmenting the discharge traces of internal transformer defects. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system, equipment and medium for identifying defects in oil-paper insulation of high-voltage equipment.

[0005] The technical solutions of the present invention are as follows:

[0006] A method for identifying defects in oil-paper insulation of high-voltage equipment includes the following operations:

[0007] S1. Acquire images of several locations inside the high-voltage equipment, register them, and then perform pixel fusion processing on repeated areas to obtain a mosaic image of the interior of the high-voltage equipment;

[0008] S2. Extract different depth features of the internal mosaic of the high-voltage equipment to obtain several depth features; specifically, the internal mosaic of the high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment; the shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment; the middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment;

[0009] S3, cross-fusion processing is performed on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature;

[0010] S4, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

[0011] The specific operations of feature cross-fusion processing in S3 are as follows: the deep features inside the high-voltage equipment and the medium-deep features inside the high-voltage equipment, the medium-level features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment are respectively processed by local feature extraction to obtain deep local features, medium-deep local features, medium-level local features, and shallow local features; the deep local features are respectively processed with the medium-deep local features, the medium-level local features, and the shallow local features based on upsampling local feature fusion to obtain the first fusion feature, the second fusion feature, and the third fusion feature; the first fusion feature is processed by multi-scale feature aggregation to obtain the first cross-fusion feature; the first cross-fusion feature is respectively processed with the second fusion feature, the third fusion feature and the deep local feature through local attention feature aggregation based on downsampling to obtain the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature.

[0012] The specific operation of local feature extraction processing is as follows: the input feature map is processed by the local attention mechanism in the horizontal direction and the vertical direction respectively, and then multiplied element-by-element with the input feature map to obtain a local attention feature map; the local attention feature map and the input feature map are multiplied element-by-element and then subjected to two-dimensional convolution processing to obtain an output feature map; the input feature map is the deep features inside the high-voltage equipment, or the medium-deep features inside the high-voltage equipment, or the medium-level features inside the high-voltage equipment, or the shallow features inside the high-voltage equipment; the output feature map is the deep local features, or the medium-deep local features, or the medium-level local features, or the shallow local features.

[0013] The specific operation of the local feature fusion processing based on upsampling is as follows: the deep local features are subjected to several times of deconvolution processing to obtain deep deconvolution local features; the deep deconvolution local features are processed by the local attention mechanism in the horizontal and vertical directions respectively, and then multiplied element-by-element with the deep deconvolution local features and the first feature to be fused to obtain the initial first fusion feature; the initial first fusion feature and the deep deconvolution local features are added element-by-element to obtain the fusion feature; the first feature to be fused is a medium-deep local feature, a medium-level local feature, or a shallow local feature.

[0014] The specific operation of the local attention feature aggregation processing based on downsampling is as follows: the first cross-fusion feature is processed by several two-dimensional convolutions to obtain the first cross-fusion convolution feature; the first cross-fusion convolution feature is processed by the local attention mechanism in the horizontal and vertical directions respectively, and then multiplied element-by-element with the first cross-fusion convolution feature and the second feature to be fused to obtain the initial second fusion feature; the initial second fusion feature and the first cross-fusion convolution feature are processed by element-by-element addition and multi-scale feature aggregation to obtain the cross-fusion feature; the second feature to be fused is the second fusion feature, or the third fusion feature, or the deep local feature.

[0015] Before the second feature to be fused is subjected to the element-by-element multiplication operation, the second feature to be fused is also included in the local feature enhancement operation to obtain the second enhanced feature to be fused; when the second feature to be fused is the second fusion feature or the third fusion feature, the local feature enhancement operation can be achieved through multi-scale feature aggregation processing, local attention mechanism processing and two-dimensional convolution processing; when the second feature to be fused is a deep local feature, the local feature enhancement operation can be achieved through local attention mechanism processing and two-dimensional convolution processing.

[0016] After the operation of obtaining the internal mosaic image of the high-voltage equipment in S1, it also includes performing image enhancement processing on the internal mosaic image of the high-voltage equipment to obtain an enhanced internal image of the high-voltage equipment for executing the operation in S2; the image enhancement processing operation is specifically: the internal mosaic image of the high-voltage equipment is processed by Gaussian filtering of different scales to obtain several Gaussian filtering feature maps; after the several Gaussian filtering feature maps and the internal mosaic image of the high-voltage equipment are subjected to feature aggregation processing, they are normalized to obtain the enhanced internal image of the high-voltage equipment.

[0017] A high-voltage equipment oil-paper insulation defect identification system, used to implement the above-mentioned high-voltage equipment oil-paper insulation defect identification method, comprising:

[0018] The high-voltage equipment internal mosaic image generation module is used to obtain images of several locations inside the high-voltage equipment, perform pixel fusion processing on repeated areas after image registration, and obtain the internal mosaic image of the high-voltage equipment;

[0019] Several deep feature generation modules are used to extract different depth features from the internal mosaic of high-voltage equipment to obtain several deep features. Specifically, the internal mosaic of high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment. The shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment. The middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment.

[0020] A cross-fusion feature generation module is used to perform feature cross-fusion processing on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature;

[0021] The oil-paper insulation defect recognition result generation module is used to obtain the feature map to be detected after the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are processed with multi-scale feature enhancement and channel dimension enhancement respectively; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

[0022] A device for identifying defects in oil-paper insulation of high-voltage equipment comprises a processor and a memory, wherein the processor implements the above-mentioned method for identifying defects in oil-paper insulation of high-voltage equipment when executing a computer program stored in the memory.

[0023] A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the above-mentioned method for identifying oil-paper insulation defects of high-voltage equipment is implemented.

[0024] The beneficial effects of the present invention are:

[0025] The present invention provides a method for identifying defects in oil-paper insulation of high-voltage equipment. First, images of several positions inside the high-voltage equipment are obtained, and pixel fusion processing of repeated areas is performed to obtain a mosaic image of the internal part of the high-voltage equipment. Then, different depth features of the mosaic image of the internal part of the high-voltage equipment are extracted, and the different depth features of the mosaic image of the internal part of the high-voltage equipment are subjected to feature cross-fusion processing from bottom to top and from top to bottom, so that the image can not only focus on local details, but also combine overall semantic information, thereby improving the feature expression ability of the carbon mark image, and obtaining the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature. The combined features reflect the fusion feature information at different levels respectively; finally, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to enhance the dynamic and expressiveness of image aggregation, realize the separation of oil-paper insulation carbon mark defects from the complex environmental background, and obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result; this method is used for oil-paper insulation defect recognition in the process of internal electrical variable detection of high-voltage equipment, which can improve the defect recognition accuracy and improve the accuracy of electrical variable detection of high-voltage equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] By reading the detailed description of the preferred embodiment below, the solutions and advantages of the present application will become clear to those skilled in the art. The accompanying drawings are only for illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0027] In the attached figure:

[0028] Figure 1 This is a schematic diagram of the effect of image enhancement processing in the embodiment. Figure 1 (a) is the original image of the carbon mark on the oil-paper insulation. Figure 1 (b) is the carbon mark image of oil-paper insulation after image enhancement;

[0029] Figure 2 This is a schematic diagram of the result of the segmentation process in the embodiment. Figure 2 (a) is the result of identifying dendritic carbon traces. Figure 2 (b) in the figure is the result of identifying carbon clusters. DETAILED DESCRIPTION

[0030] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0031] This embodiment provides a method for identifying oil-paper insulation defects in high-voltage equipment, including the following operations:

[0032] S1. Acquire images of several locations inside the high-voltage equipment, register them, and then perform pixel fusion processing on repeated areas to obtain a mosaic image of the interior of the high-voltage equipment;

[0033] S2. Extract different depth features of the internal mosaic of the high-voltage equipment to obtain several depth features; specifically, the internal mosaic of the high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment; the shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment; the middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment;

[0034] S3, cross-fusion processing is performed on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature;

[0035] S4, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

[0036] S1. Acquire images of several locations inside the high-voltage equipment, register them, and perform pixel fusion processing on repeated areas to obtain a mosaic image of the interior of the high-voltage equipment.

[0037] By acquiring images of several positions inside the high-voltage equipment and performing image registration and pixel fusion processing on repeated areas, close-range and wide-field image acquisition and stitching of internal components of the high-voltage equipment are achieved, and a stitching map of the interior of the high-voltage equipment is obtained, which facilitates a comprehensive analysis of the internal situation of the high-voltage equipment and is conducive to improving the accuracy of overall oil-paper insulation defect identification.

[0038] First, images of several locations within the high-voltage equipment are acquired as inspection data. Specifically, several cameras are installed around the oil submersible. As the oil submersible navigates within the high-voltage equipment, these cameras simultaneously capture images and videos from multiple directions, including but not limited to top, bottom, left, right, front, and back. These images are used to extract images of several locations within the high-voltage equipment, improving the efficiency of oil-paper image acquisition. The captured images and videos are then wirelessly transmitted to an industrial computer outside the high-voltage equipment for display, processing, and storage. The number of these cameras must be no less than six.

[0039] During the capture process, the camera on the vehicle's perimeter is relatively narrow, and the individual cameras on the vehicle have a short shooting distance and a narrow field of view. Relying on only one or two images to identify oil-paper insulation defects can affect the accuracy of overall defect recognition. To address this technical issue, this embodiment captures images from several locations within the high-voltage equipment, performs image registration, and then performs pixel fusion processing on the duplicated areas. This creates a mosaic image of the high-voltage equipment interior, facilitating comprehensive analysis of the equipment's internal conditions and improving overall oil-paper insulation defect recognition accuracy. This image registration can be achieved using the scale-invariant feature transform (SIFT) method.

[0040] When the camera takes images, repeated areas will appear on the two images at adjacent positions. In order to improve the stitching effect and enhance the quality of the stitching image inside the high-voltage equipment, this embodiment performs pixel fusion processing on the repeated areas of the two images at adjacent positions to make the transition at the stitching natural, thereby improving the subsequent recognition accuracy.

[0041] The pixel fusion processing of the repeated area is achieved through the following formula:

[0042] ,

[0043] ,

[0044] The coordinates of the repeated area in the stitched image arex 、 y The pixel value at the corresponding position, is the pixel mixing weight of the previous image (after registration), The coordinates of the repeated area in the previous image (after registration) are x 、 y The pixel value at the corresponding position, is the pixel mixing weight of the current image (after registration), The coordinates of the repeated area in the current image (after registration) are x 、 y The pixel value at the corresponding position, For the i The pixel blending weights of the image, is the standard image width, Standard image height, 、 Respectively i The image width and height of the image.

[0045] In order to solve the problem of exposure and color distortion of the internal image of the high-voltage equipment and improve the subsequent recognition accuracy, after obtaining the internal mosaic image of the high-voltage equipment in S1, it also includes image enhancement processing of the internal mosaic image of the high-voltage equipment, see Figure 1 , obtain the internal enhancement map of the high-voltage equipment, which is used to perform the operation in S2.

[0046] The specific operation of image enhancement processing is as follows: the internal splicing image of the high-voltage equipment is processed by Gaussian filtering of different scales to obtain several Gaussian filtering feature maps; several Gaussian filtering feature maps and the internal splicing image of the high-voltage equipment are subjected to feature aggregation processing and normalization processing to obtain the internal enhanced image of the high-voltage equipment.

[0047] Feature aggregation and normalization are achieved through the following formulas:

[0048] ,

[0049] ,

[0050] The coordinates in the aggregation diagram of the internal features of high-voltage equipment are x 、 y The pixel value at the corresponding position, For the n The Gaussian filter feature map of each scale corresponds to the weight, N is the total number of Gaussian filters, The coordinates in the internal splicing diagram of the high-voltage equipment are x 、 y The pixel value at the corresponding position, The coordinates in the Gaussian filter feature map of the nth scale are x 、 y The pixel value at the corresponding position, The coordinates in the diagram for the internal enhancement of high-voltage equipment are x 、 y The pixel value at the corresponding position, The coordinates of several Gaussian filter feature maps are x 、 y The maximum pixel value at the corresponding position, The coordinates of several Gaussian filter feature maps are x 、 y The minimum pixel value at the corresponding position.

[0051] S2. Extract different depth features of the internal mosaic image of the high-voltage equipment to obtain several depth features.

[0052] Extracting different depth features of the internal mosaic image of high-voltage equipment and obtaining rich feature information is conducive to improving the subsequent ability to obtain slender and curved local structures in the internal mosaic image of high-voltage equipment, facilitating the capture of the dendritic structure features on the oil paper, and realizing the accurate extraction of the fine dendritic structure features of the oil paper carbon mark defects, thereby achieving higher recognition accuracy.

[0053] Specifically, the internal splicing image of the high-voltage equipment is processed by convolution and cross-level tortuosity feature extraction to obtain the texture change characteristics of the oil paper and obtain the shallow internal features of the high-voltage equipment; the shallow internal features of the high-voltage equipment are processed by cross-level tortuosity feature extraction once and twice respectively to obtain the middle internal features of the high-voltage equipment and the middle and deep internal features of the high-voltage equipment, which can respectively reflect details such as tiny cracks in the oil paper and local color differences; the middle and deep internal features of the high-voltage equipment are processed by cross-level tortuosity feature extraction and spatial pyramid pooling (preferably implemented through the SPPF model) to obtain the deep internal features of the high-voltage equipment that can reflect the overall structural state of the oil paper.

[0054] Cross-layer zigzag feature extraction can be achieved by replacing convolutions with C2f modules that implement dynamic snake convolutions. Specifically, the key convolutions in the C2f module in the YOLO model are replaced with dynamic snake convolutions (DySnakeConv) to obtain an improved C2f module. Cross-layer zigzag feature extraction can be achieved through the improved C2f module.

[0055] S3. The deep features inside the high-voltage equipment are cross-fused with the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature.

[0056] The different depth features of the internal mosaic image of the high-voltage equipment are cross-fused from bottom to top and from top to bottom, so that the acquired image can not only focus on local details but also combine the overall semantic information, thereby improving the feature expression ability of the carbon mark image. The first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are obtained, which respectively reflect the fusion feature information at different levels, which is conducive to the subsequent accurate and comprehensive oil-paper insulation defect detection.

[0057] The specific operations of feature cross fusion processing are as follows.

[0058] First, the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment are respectively processed by local feature extraction to obtain deep local features, medium-deep local features, medium local features, and shallow local features.

[0059] The specific operation of local feature extraction processing is as follows: after the input feature map is processed by the local attention mechanism in the horizontal and vertical directions respectively, it is multiplied element-by-element with the input feature map to focus on key details, and obtain a local attention feature map that can focus on small targets on the oil paper, so that the characteristics of small defects can be captured more clearly in the future; the local attention feature map and the input feature map are multiplied element-by-element and then subjected to two-dimensional convolution processing to further enhance the local features and obtain the output feature map.

[0060] The input feature map is the deep features inside the high-voltage equipment, or the medium-deep features inside the high-voltage equipment, or the medium-layer features inside the high-voltage equipment, or the shallow features inside the high-voltage equipment; the output feature map corresponds to the deep local features, or the medium-deep local features, or the medium-layer local features, or the shallow local features.

[0061] Then, the deep local features are fused with the medium-deep local features, the medium-layer local features, and the shallow local features based on upsampling to achieve complementary refinement of features at different levels, and obtain the first fusion features, the second fusion features, and the third fusion features. These features are used for subsequent defect identification to accurately locate and judge the degree and scope of aging defects, facilitate analysis of the impact range of discharge and the specific changes to local materials, and thus judge whether the oil paper is in the aging process or in a damp state more accurately through detailed differences, thereby improving the accuracy and stability of identifying oil-paper insulation defects in complex environments.

[0062] The above-mentioned upsampling-based local feature fusion processing operation is specifically as follows: the deep local features are subjected to several deconvolution processes to obtain deep deconvolution local features; the deep deconvolution local features are processed by the local attention mechanism in the horizontal and vertical directions respectively, and then multiplied element-by-element with the deep deconvolution local features and the first feature to be fused to obtain the initial first fused feature; the initial first fused feature and the deep deconvolution local features are added element-by-element to obtain the fused feature. The initial first feature to be fused is a medium-deep local feature, a medium-level local feature, or a shallow local feature. The fused features correspond to the third fused feature, the second fused feature, or the first fused feature, respectively.

[0063] Among them, in the process of obtaining the first fusion feature, the number of times the deep local feature is processed by deconvolution is 3, and in the process of obtaining the second fusion feature, the number of times the deep local feature is processed by deconvolution is 2.

[0064] Finally, the first fusion feature is processed by multi-scale feature aggregation (which can be achieved through the C2f module) to obtain the first cross-fusion feature; the first cross-fusion feature is respectively aggregated with the second fusion feature, the third fusion feature and the deep local feature through downsampling-based local attention feature aggregation to further fuse features at different levels and enhance the ability to express image details, thereby obtaining the second cross-fusion feature, the third cross-fusion feature and the fourth cross-fusion feature.

[0065] The specific operation of the local attention feature aggregation processing based on downsampling is as follows: the first cross-fusion feature is processed by several two-dimensional convolutions to obtain the first cross-fusion convolution feature; the first cross-fusion convolution feature is processed by the local attention mechanism in the horizontal direction and the vertical direction respectively, and then multiplied element-by-element with the first cross-fusion convolution feature and the second feature to be fused to obtain the initial second fusion feature; the initial second fusion feature and the first cross-fusion convolution feature are processed by element-by-element addition and multi-scale feature aggregation to obtain the cross-fusion feature; the second feature to be fused is the second fusion feature, or the third fusion feature, or the deep local feature; the cross-fusion features correspond to the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature respectively.

[0066] In order to further enhance the ability to express image details, before the second feature to be fused is subjected to element-by-element multiplication, the second feature to be fused is also subjected to local feature enhancement processing to obtain the second enhanced feature to be fused, which is used for element-by-element multiplication with the first cross-fused convolution feature and the first cross-fused convolution feature processed by the local attention mechanism.

[0067] When the second feature to be fused is the second fusion feature or the third fusion feature, the operation of local feature enhancement processing can be achieved through multi-scale feature aggregation processing, local attention mechanism processing and two-dimensional convolution processing.

[0068] When the second feature to be fused is a deep local feature, the operation of local feature enhancement processing can be achieved through local attention mechanism processing and two-dimensional convolution processing.

[0069] S4, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

[0070] The second, third, and fourth cross-fusion features obtained through bottom-up and top-down fusion are processed with multi-scale feature enhancement and channel dimension enhancement respectively, and then spliced with the first cross-fusion feature to enhance the dynamics and expressiveness of image aggregation, and to separate the carbon mark defects of oil-paper insulation from the complex environmental background. The feature map to be detected is obtained, which can clearly reflect the overall state of the image and highlight the key details. The feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result, see Figure 2 The channel dimension enhancement processing includes but is not limited to being implemented by the Reshape function; the segmentation processing operation can be implemented by a training neural network obtained by training a neural network on a standard oil-paper insulation dataset.

[0071] Taking the second cross-fusion feature as an example, the above-mentioned multi-scale feature enhancement processing operation is specifically as follows: the second cross-fusion feature is spliced after several convolutions, batch normalization, Silu activation function processing, and two-dimensional convolution at different scales to obtain the second cross-fusion enhanced feature, which is used to perform the channel dimension enhancement processing operation.

[0072] This embodiment further provides a high-voltage equipment oil-paper insulation defect identification system, which is used to implement the above-mentioned high-voltage equipment oil-paper insulation defect identification method, including:

[0073] The high-voltage equipment internal mosaic image generation module is used to obtain images of several locations inside the high-voltage equipment, perform pixel fusion processing on repeated areas after image registration, and obtain the internal mosaic image of the high-voltage equipment;

[0074] Several deep feature generation modules are used to extract different depth features from the internal mosaic of high-voltage equipment to obtain several deep features. Specifically, the internal mosaic of high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment. The shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment. The middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment.

[0075] A cross-fusion feature generation module is used to perform feature cross-fusion processing on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature;

[0076] The oil-paper insulation defect recognition result generation module is used to obtain the feature map to be detected after the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are processed with multi-scale feature enhancement and channel dimension enhancement respectively; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

[0077] This embodiment also provides a high-voltage equipment oil-paper insulation defect identification device, including a processor and a memory, wherein the processor implements the above-mentioned high-voltage equipment oil-paper insulation defect identification method when executing a computer program stored in the memory.

[0078] This embodiment further provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for identifying oil-paper insulation defects of high-voltage equipment.

[0079] This embodiment provides a method for identifying defects in oil-paper insulation of high-voltage equipment. First, images of several locations inside the high-voltage equipment are obtained, and pixel fusion processing of repeated areas is performed to obtain a mosaic image of the internal part of the high-voltage equipment. Then, different depth features of the mosaic image of the internal part of the high-voltage equipment are extracted, and the different depth features of the mosaic image of the internal part of the high-voltage equipment are subjected to feature cross-fusion processing from bottom to top and from top to bottom, so that the image can not only focus on local details but also combine overall semantic information, thereby improving the feature expression ability of the carbon mark image, and obtaining the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature. The combined features reflect the fusion feature information at different levels respectively; finally, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to enhance the dynamic and expressiveness of image aggregation, realize the separation of oil-paper insulation carbon mark defects from the complex environmental background, and obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result; this method is used for oil-paper insulation defect recognition in the process of internal electrical variable detection of high-voltage equipment, which can improve the defect recognition accuracy and improve the accuracy of electrical variable detection of high-voltage equipment.

Claims

1. A method for identifying defects in oil-paper insulation of high-voltage equipment, characterized in that: The following operations are included: S1. Acquire images of several locations inside the high-voltage equipment, register them, and then perform pixel fusion processing on repeated areas to obtain a mosaic image of the interior of the high-voltage equipment; S2. Extract different depth features of the internal mosaic of the high-voltage equipment to obtain several depth features; specifically, the internal mosaic of the high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment; the shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment; the middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment; S3, cross-fusion processing is performed on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature; S4, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are respectively processed by multi-scale feature enhancement and channel dimension enhancement, and then spliced with the first cross-fusion feature to obtain the feature map to be detected; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

2. The method for identifying oil-paper insulation defects of high-voltage equipment according to claim 1, characterized in that: In S3, the operations of feature cross-fusion processing are specifically as follows: The deep internal features of the high-voltage equipment and the medium-deep internal features, the medium-layer internal features, and the shallow internal features of the high-voltage equipment are respectively processed by local feature extraction to obtain deep local features, medium-deep local features, medium-layer local features, and shallow local features; the deep local features are respectively processed with the medium-deep local features, the medium-layer local features, and the shallow local features based on upsampling local feature fusion to obtain the first fusion feature, the second fusion feature, and the third fusion feature; the first fusion feature is processed by multi-scale feature aggregation to obtain the first cross-fusion feature; the first cross-fusion feature is respectively processed with the second fusion feature, the third fusion feature and the deep local feature through local attention feature aggregation based on downsampling to obtain the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature.

3. The method for identifying defects in oil-paper insulation of high-voltage equipment according to claim 2, characterized in that: The specific operations of local feature extraction processing are: The input feature map is processed by the local attention mechanism in the horizontal and vertical directions respectively, and then multiplied element-wise with the input feature map to obtain the local attention feature map; the local attention feature map and the input feature map are multiplied element-wise and then subjected to two-dimensional convolution to obtain the output feature map; The input feature map is the deep features inside the high-voltage equipment, or the medium-deep features inside the high-voltage equipment, or the medium-layer features inside the high-voltage equipment, or the shallow features inside the high-voltage equipment; the output feature map is the deep local features, or the medium-deep local features, or the medium-layer local features, or the shallow local features.

4. The method for identifying defects in oil-paper insulation of high-voltage equipment according to claim 2, characterized in that: The specific operations of local feature fusion processing based on upsampling are: The deep local features are processed by several deconvolutions to obtain the deep deconvolution local features. The deep deconvolution local features are processed by the local attention mechanism in the horizontal and vertical directions respectively, and then multiplied element-by-element with the deep deconvolution local features and the first feature to be fused to obtain the initial first fusion feature. The initial first fusion feature and the deep deconvolution local features are added element-by-element to obtain the fusion feature. The first feature to be fused is a mid-deep local feature, a mid-layer local feature, or a shallow local feature.

5. The method for identifying defects in oil-paper insulation of high-voltage equipment according to claim 2, characterized in that: The specific operations of local attention feature aggregation processing based on downsampling are: The first cross-fusion feature is processed by several two-dimensional convolutions to obtain the first cross-fusion convolution feature; the first cross-fusion convolution feature is processed by the local attention mechanism in the horizontal direction and the vertical direction respectively, and then multiplied element-by-element with the first cross-fusion convolution feature and the second feature to be fused to obtain the initial second fusion feature; the initial second fusion feature and the first cross-fusion convolution feature are subjected to element-by-element addition and multi-scale feature aggregation to obtain the cross-fusion feature; The second feature to be fused is the second fusion feature, or the third fusion feature, or the deep local feature.

6. The method for identifying oil-paper insulation defects of high-voltage equipment according to claim 5, characterized in that: Before the second feature to be fused is subjected to the element-by-element multiplication operation, the second feature to be fused is further subjected to local feature enhancement processing to obtain a second enhanced feature to be fused; When the second feature to be fused is the second fused feature or the third fused feature, the local feature enhancement processing operation can be achieved through multi-scale feature aggregation processing, local attention mechanism processing and two-dimensional convolution processing; When the second feature to be fused is a deep local feature, the operation of local feature enhancement processing can be achieved through local attention mechanism processing and two-dimensional convolution processing.

7. The method for identifying defects in oil-paper insulation of high-voltage equipment according to claim 1, characterized in that: After the operation of obtaining the internal mosaic image of the high-voltage equipment in S1, the process further includes performing image enhancement processing on the internal mosaic image of the high-voltage equipment to obtain an enhanced internal image of the high-voltage equipment for performing the operation in S2; The specific operation of image enhancement processing is as follows: the internal splicing image of the high-voltage equipment is processed by Gaussian filtering of different scales to obtain several Gaussian filtering feature maps; several Gaussian filtering feature maps and the internal splicing image of the high-voltage equipment are subjected to feature aggregation processing and normalization processing to obtain the internal enhanced image of the high-voltage equipment.

8. A high-voltage equipment oil-paper insulation defect identification system, characterized in that: The method for identifying oil-paper insulation defects of high-voltage equipment according to claim 1 is characterized by comprising: The high-voltage equipment internal mosaic image generation module is used to obtain images of several locations inside the high-voltage equipment, perform pixel fusion processing on repeated areas after image registration, and obtain the internal mosaic image of the high-voltage equipment; Several deep feature generation modules are used to extract different depth features from the internal mosaic of high-voltage equipment to obtain several deep features. Specifically, the internal mosaic of high-voltage equipment is processed by convolution and cross-level zigzag feature extraction to obtain shallow features of the high-voltage equipment. The shallow features of the high-voltage equipment are processed by cross-level zigzag feature extraction once and twice to obtain middle features and middle-deep features of the high-voltage equipment. The middle-deep features of the high-voltage equipment are processed by cross-level zigzag feature extraction and spatial pyramid pooling to obtain deep features of the high-voltage equipment. A cross-fusion feature generation module is used to perform feature cross-fusion processing on the deep features inside the high-voltage equipment, the medium-deep features inside the high-voltage equipment, the medium-layer features inside the high-voltage equipment, and the shallow features inside the high-voltage equipment to obtain the first cross-fusion feature, the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature; The oil-paper insulation defect recognition result generation module is used to obtain the feature map to be detected after the second cross-fusion feature, the third cross-fusion feature, and the fourth cross-fusion feature are processed with multi-scale feature enhancement and channel dimension enhancement respectively; the feature map to be detected is segmented to obtain the oil-paper insulation defect recognition result.

9. A device for identifying defects in oil-paper insulation of high-voltage equipment, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the method for identifying oil-paper insulation defects of high-voltage equipment according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that Used to store a computer program, wherein when the computer program is executed by a processor, the method for identifying oil-paper insulation defects of high-voltage equipment according to any one of claims 1 to 7 is implemented.

Images