Distribution network line skeleton type defect debugging method, system, equipment and medium

Abstract

The invention relates to the technical field of image processing, and discloses a distribution network line skeleton type defect debugging method, system and device and a medium. The method comprises the following steps: acquiring a distribution network line inspection image, and enabling the distribution network line inspection image to sequentially pass through a skeleton region detection model, a large target identification model and a small target identification model to obtain a plurality of first residual defect rectangular frames and corresponding first residual defect types; performing debugging on each first residual defect rectangular frame and the corresponding first residual defect type by using a first debugging rule to obtain a plurality of second residual defect rectangular frames and the corresponding second residual defect types, and performing debugging on each second residual defect rectangular frame and the corresponding second residual defect type by using a second debugging rule to obtain a plurality of second residual defect rectangular frames and the corresponding second residual defect types. Obtaining a distribution network line defect debugging result; and recording a debugging result defect list according to the distribution network line defect debugging result, and storing the debugging result defect list. According to the invention, the debugging efficiency of electric power inspection work can be greatly improved, and the false detection phenomenon is effectively prevented.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, device and medium for troubleshooting skeleton-type defects in power distribution network lines. Background Technology

[0002] Currently, there are numerous drone inspections of power distribution lines, and the resulting defect images urgently need processing. Conventional power distribution line defect detection models generate a large number of false alarms when detecting defects in these images. On the one hand, many detected defects are not actually on the power distribution line, but rather in other background areas; on the other hand, many detected defects, although on the power distribution line, are not located within the actual defect area according to prior knowledge of the actual defects. This results in a large number of false alarms requiring troubleshooting, posing a significant challenge to the staff.

[0003] Existing technologies for identifying defects in distribution network lines often employ detection algorithms such as the YOLO series and DETR series. However, due to the indistinct characteristics of defects in distribution network lines, the numerous types of defects, and the complex and varied scenarios they occur in, directly using detection algorithms such as the YOLO series and DETR series often results in insufficient detection and a high false alarm rate. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to overcome the shortcomings of the prior art and provide a method, system, device and medium for troubleshooting defects in the skeleton type of distribution network lines.

[0005] This invention provides the following technical solution: In a first aspect, the present invention provides a method for troubleshooting defects in the skeleton type of distribution network lines, the method comprising: Obtain distribution network line inspection images, input the distribution network line inspection images into the skeleton region detection model to obtain multiple skeleton region rectangles and multiple skeleton region types, merge and mark each skeleton region rectangle to obtain a first error correction region and a first error correction region image, and mark each skeleton region rectangle individually to obtain multiple second error correction regions and multiple second error correction region types; The distribution network line inspection image is input into the large target recognition model to obtain multiple large target defect rectangles and multiple large target defect types. The first error correction area image is input into the small target recognition model to obtain multiple small target defect rectangles and multiple small target defect types. The large target defect rectangles and the large target defect types, and the small target defect rectangles and the small target defect types are filtered and fused to obtain multiple first residual defect rectangles and their corresponding first residual defect types. The first error-checking rule is used to check each first remaining defect rectangle and its corresponding first remaining defect type to obtain multiple second remaining defect rectangles and their corresponding second remaining defect types. The second error-checking rule is used to check each second remaining defect rectangle and its corresponding second remaining defect type to obtain multiple third remaining defect rectangles and their corresponding third remaining defect types. Each third remaining defect rectangle and its corresponding third remaining defect type is determined as the error-checking result of the distribution network line defect. Determine whether the fault diagnosis result of the distribution network line is empty. If so, record the fault diagnosis result defect list of the distribution network line inspection image as empty. Otherwise, record each third remaining defect rectangle and its corresponding third remaining defect type into the fault diagnosis result defect list and save it.

[0006] In an optional implementation, the step of using a first error-correction rule to correct errors in each of the first remaining defect rectangles and their corresponding first remaining defect types to obtain multiple second remaining defect rectangles and their corresponding second remaining defect types includes: For each of the first remaining defect types, a corresponding first mutation overlap rate threshold is set; Traverse each first remaining defect rectangle, calculate the variation overlap rate between each first remaining defect rectangle and the first error elimination region, and use the first error elimination formula to exclude the first target remaining defect rectangle. The variation overlap rate of the first target remaining defect rectangle is less than the first variation overlap rate threshold of its corresponding first remaining defect type, and obtain multiple second remaining defect rectangles and their corresponding second remaining defect types. The first error-correction formula is:

[0007] In the formula, Indicate whether to exclude the first remaining defect type. The The first remaining defect rectangle ,when When the value is 1, the corresponding first remaining defect rectangle is... As the remaining defect rectangle of the first target, and excluded, when When the value is 0, the corresponding first remaining defect rectangle is... reserve; This indicates that the first remaining defect type is The The first remaining defect rectangle With the first error correction area The overlap rate of the variation, Indicates that the first remaining defect type is The first variation overlap rate threshold.

[0008] In an optional implementation, the step of using a second error-correction rule to correct errors in each of the second remaining defect rectangles and their corresponding second remaining defect types to obtain multiple third remaining defect rectangles and their corresponding third remaining defect types includes: For each second remaining defect type, multiple second error-correction region types are associated, and for each second error-correction region type, a corresponding second variation overlap rate threshold is set; Traverse each second remaining defect rectangle, and search for the second error correction regions of multiple second error correction region types associated with each second remaining defect rectangle according to the second remaining defect type, to obtain the search results of the second error correction region list; Determine whether the search results of the second error-correction area list are empty. If so, retain each second remaining defect rectangle and its corresponding second remaining defect type. Otherwise, obtain the corresponding second error-correction area list and the corresponding second error-correction area type list. The second error-correction area type list includes multiple second error-correction area types. Traverse each second error correction region in the second error correction region list, calculate the variation overlap rate of each second remaining defect rectangle and its corresponding multiple second error correction regions, and use the second error correction formula to exclude the second target remaining defect rectangle. The variation overlap rate of the second target remaining defect rectangle is less than the second variation overlap rate threshold of the multiple second error correction region types associated with its corresponding second remaining defect type, and obtain multiple third remaining defect rectangles and their corresponding third remaining defect types. The second error-correction formula is as follows:

[0009] In the formula, Indicate whether to exclude the second residual defect type. The The second remaining defect rectangle ,when When the value is 1, the corresponding second remaining defect rectangle is... As the remaining defect rectangle of the second target, and excluded, when When the value is 0, the corresponding second remaining defect rectangle is... reserve; This indicates that the second residual defect type is The associated list of second error correction region types, Indicates the second residual defect type The associated first The second type of misaligned area, This indicates that the second residual defect type is The The second remaining defect rectangle The second error correction area type is The The second misaligned area The overlap rate of the variation, This indicates that the second residual defect type is The associated second error correction region type is The second variation overlap rate threshold.

[0010] In an optional implementation, the step of filtering and fusing the large target defect rectangles with the large target defect types and the small target defect rectangles with the small target defect types to obtain multiple first remaining defect rectangles and their corresponding first remaining defect types includes: Obtain the confidence level of each of the large target defect rectangles, and delete rectangles with confidence levels lower than the first preset confidence threshold to obtain multiple first large target defect rectangles; The overlap rate of each first largest target defect rectangle is calculated using a non-maximum suppression algorithm, and rectangles with an overlap rate greater than a preset overlap rate threshold are deleted to obtain multiple second largest target defect rectangles. Obtain the confidence level of each of the small target defect rectangles, and delete rectangles with confidence levels lower than the second preset confidence threshold to obtain multiple first small target defect rectangles; The variation overlap rate of each first small target defect rectangle is calculated using the area non-maximum suppression processing algorithm, and rectangles with variation overlap rates greater than a preset variation overlap rate threshold are deleted to obtain multiple second small target defect rectangles. The second large target defect rectangle and its corresponding large target defect type, and the second small target defect rectangle and its corresponding small target defect type are merged to obtain multiple first remaining defect rectangles and their corresponding first remaining defect types.

[0011] In an optional implementation, the step of inputting the distribution network line inspection image into a large target recognition model to obtain multiple large target defect bounding boxes and multiple large target defect types, and inputting the first error-correction area image into a small target recognition model to obtain multiple small target defect bounding boxes and multiple small target defect types, includes: The network line inspection image is input into the large target recognition model, and the large target recognition model is used to detect target defects to obtain multiple large target defect rectangles and multiple large target defect types. The first error-correction area image is cropped using a sliding window at a preset cropping resolution to obtain a sliding window image set. The sliding window image set is then input into the small target recognition model, which is used to detect target defects. The defect rectangles and defect types of the sliding window image set are obtained. The coordinates of the defect rectangles of the sliding window image set are mapped back to the coordinates of the distribution network line inspection image to obtain multiple small target defect rectangles and multiple small target defect types.

[0012] In an optional implementation, the skeleton region detection model is based on an improved YOLO11 detection algorithm. The improved YOLO11 detection algorithm includes a backbone network, a feature pyramid network, and a decoupling head. The feature pyramid network includes a first self-attention convolutional module, a second self-attention convolutional module, a first convolutional module, a second convolutional module, a first efficient attention convolutional module, a second efficient attention convolutional module, and a third efficient attention convolutional module. The decoupling head includes a first classification regression module, a second classification regression module, and a third classification regression module. The distribution network line inspection image is input into the skeleton region detection model to obtain multiple skeleton region bounding boxes and multiple skeleton region types, including: The inspection images of the distribution network lines are input into the backbone network to obtain the first layer feature map, the second layer feature map and the third layer feature map; The third-layer feature map and the second-layer feature map are input into the first self-attention convolution module to obtain the first self-attention convolution feature map. The first self-attention convolution feature map and the second-layer feature map are concatenated to obtain the first fused feature map. The first fused feature map and the first layer feature map are input into the second self-attention convolution module to obtain the second self-attention convolution feature map, and the second self-attention convolution feature map is concatenated with the first layer feature map to obtain the second fused feature map. The second fused feature map is input into the first convolution module to obtain the first feature map, and the first feature map is concatenated with the first fused feature map to obtain the third fused feature map; The third fused feature map is input into the second convolution module to obtain the second feature map, and the second feature map is concatenated with the third layer feature map to obtain the fourth fused feature map; The second fused feature map is input into the first efficient attention convolution module to obtain the first efficient attention convolution feature map. The third fused feature map is input into the second efficient attention convolution module to obtain the second efficient attention convolution feature map. The fourth fused feature map is input into the third efficient attention convolution module to obtain the third efficient attention convolution feature map. The first efficient attention convolution feature map, the second efficient attention convolution feature map, and the third efficient attention convolution feature map are respectively input into the first classification regression module, the second classification regression module, and the third classification regression module to obtain multiple skeleton region rectangles and multiple skeleton region types.

[0013] In an optional implementation, the step of inputting the third-layer feature map and the second-layer feature map into the first self-attention convolution module to obtain the first self-attention convolution feature map includes: The second layer feature map is convolved twice with 1×1 to obtain the first projection matrix and the second projection matrix respectively. The first projection matrix and the second projection matrix are matched for similarity to obtain a similarity matching matrix. The similarity matching matrix is ​​then processed by an activation function to obtain an attention weighting matrix. The third layer feature map is sequentially convolved with 1×1 and upsampled to obtain an upsampled feature map. The upsampled feature map is then convolved with 1×1 to obtain the third projection matrix. The third projection matrix and the attention weighting matrix are multiplied together to obtain the first self-attention convolution feature map. The step of inputting the second fused feature map into the first efficient attention convolution module to obtain the first efficient attention convolution feature map includes: Based on the number of channels in the second fused feature map, the second fused feature map is divided into two fused feature sub-maps. The two fused feature sub-maps are then subjected to spatial max pooling and spatial average pooling, respectively, and then the channels are concatenated to obtain a channel feature map. The channel feature map is convolved with 1×1 and activated by ReLU to obtain a channel weighted feature map. The second fused feature map is multiplied with the channel weighted feature map to obtain a channel attention feature map. Based on the number of channels in the channel attention feature map, the channel attention feature map is divided into two channel attention feature sub-maps. The two channel attention feature sub-maps are then subjected to channel max pooling and channel average pooling, respectively, and then concatenated to obtain a spatial feature map. The spatial feature map is convolved with 1×1 and activated by the ReLU function to obtain a spatially weighted feature map. The channel attention feature map and the spatially weighted feature map are then multiplied by a matrix to obtain the first efficient attention convolution feature map.

[0014] Secondly, the present invention provides a distribution network line skeleton-type defect troubleshooting system, the system comprising: The marking module is used to acquire distribution network line inspection images, input the distribution network line inspection images into the skeleton region detection model to obtain multiple skeleton region rectangles and multiple skeleton region types, merge and mark each skeleton region rectangle to obtain a first error correction region and a first error correction region image, and mark each skeleton region rectangle individually to obtain multiple second error correction regions and multiple second error correction region types. The identification module is used to input the inspection image of the distribution network line into the large target identification model to obtain multiple large target defect rectangles and multiple large target defect types, and to input the image of the first error correction area into the small target identification model to obtain multiple small target defect rectangles and multiple small target defect types. The fusion module is used to filter and fuse each of the large target defect rectangles with each of the large target defect types and each of the small target defect rectangles with each of the small target defect types to obtain multiple first residual defect rectangles and their corresponding first residual defect types; The error-detection module is used to detect errors in each of the first remaining defect rectangles and their corresponding first remaining defect types using a first error-detection rule, to obtain multiple second remaining defect rectangles and their corresponding second remaining defect types, and to detect errors in each of the second remaining defect rectangles and their corresponding second remaining defect types using a second error-detection rule, to obtain multiple third remaining defect rectangles and their corresponding third remaining defect types, and to determine each of the third remaining defect rectangles and their corresponding third remaining defect types as the error-detection results of the distribution network line defects; The recording module is used to determine whether the fault diagnosis result of the distribution network line is empty. If so, the fault diagnosis result defect list of the distribution network line inspection image is recorded as empty. Otherwise, each third remaining defect rectangle and its corresponding third remaining defect type are recorded in the fault diagnosis result defect list and saved.

[0015] Thirdly, this disclosure provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the distribution network line skeleton defect troubleshooting method described in the first aspect.

[0016] Fourthly, this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the distribution network line skeleton defect troubleshooting method described in the first aspect.

[0017] The beneficial effects of this application are: The distribution network line skeleton-based defect troubleshooting method provided in this application embodiment is based on a deep learning algorithm. First, a skeleton region detection model is used to obtain the skeleton region rectangle and skeleton region type, thereby obtaining the first-level troubleshooting region, the first-level troubleshooting region image, the second-level troubleshooting region, and the second-level troubleshooting region type. Then, a large target defect detection model is used to detect large target defect types, and a small target defect detection model is used to detect small target defect types. Finally, the first troubleshooting rule and the second troubleshooting rule are used sequentially to perform troubleshooting processing for large target defect types and small target defect types, obtaining the final distribution network line defect troubleshooting result. This method can greatly improve the troubleshooting efficiency of power inspection work and effectively prevent false detections.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the various drawings, similar components are numbered similarly.

[0020] Figure 1 A flowchart of a method for troubleshooting defects in a distribution network line skeleton provided in an embodiment of this application is shown; Figure 2 A schematic diagram of the network structure of an improved YOLO11 detection algorithm provided in an embodiment of this application is shown; Figure 3 This illustration shows a schematic diagram of the network structure of a first self-attention convolutional module provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of the network structure of a first high-efficiency attention convolutional module provided in an embodiment of this application; Figure 5 This paper shows a schematic diagram of a distribution network line skeleton-type defect troubleshooting system provided in an embodiment of this application. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] Example 1 like Figure 1 The diagram shown is a flowchart of a method for troubleshooting skeleton-type defects in distribution network lines according to an embodiment of this application. The method for troubleshooting skeleton-type defects in distribution network lines provided in this embodiment includes the following steps: Step S110: Obtain the distribution network line inspection image, input the distribution network line inspection image into the skeleton area detection model to obtain multiple skeleton area rectangles and multiple skeleton area types, merge and mark the skeleton area rectangles to obtain the first error correction area and the first error correction area image, and mark the skeleton area rectangles individually to obtain multiple second error correction areas and multiple second error correction area types.

[0023] In this embodiment, firstly, inspection images of distribution network lines are acquired. Based on these images, the components of the distribution network lines are skeletonized, resulting in various skeleton regions distributed along the distribution network lines (specifically including pole areas, crossarm equipment areas with many devices mounted on crossarms, transformer areas with many devices mounted on transformers, vacuum equipment areas with many devices mounted on vacuum equipment, connecting hardware areas, conductor areas, etc.). In the distribution network line inspection images, each skeleton region is bounded to a rectangle and labeled with its type, resulting in an original skeleton region sample library. Based on this original sample library, an augmented skeleton region sample library is constructed (augmentation methods include left / right flipping, angle rotation, Gaussian noise, image mixing, etc.). The original and augmented skeleton region sample libraries are merged to obtain a skeleton region sample library. Then, based on this sample library, an improved YOLO11 detection algorithm is trained to obtain a skeleton region detection model.

[0024] Understandably, distribution network inspection images have high resolution and varied scenes. The various skeleton regions distributed on the distribution network lines present targets of different sizes and complex shapes and textures. In order to accurately detect the skeleton regions of the distribution network lines, this application adopts an improved YOLO11 detection algorithm to build a skeleton region detection model, which can not only effectively cover the scale range of each skeleton region, but also effectively perceive the features of each skeleton region itself.

[0025] In order to accurately obtain high-level detailed features and quickly extract the skeleton region of the distribution network, this application introduces a self-attention convolution module (SACM) into the feature pyramid network part of the original YOLO11 detection algorithm, so that the network pays more attention to the skeleton region. It also introduces an efficient attention convolution module (EACM) to enable the network to quickly capture the detailed features of the skeleton region.

[0026] like Figure 2As shown, the improved YOLO11 detection algorithm includes a backbone network, a feature pyramid network, and a decoupling head. The feature pyramid network includes a first self-attention convolutional module, a second self-attention convolutional module, a first convolutional module, a second convolutional module, a first efficient attention convolutional module, a second efficient attention convolutional module, and a third efficient attention convolutional module. The decoupling head includes a first classification regression module, a second classification regression module, and a third classification regression module.

[0027] Understandably, inputting the distribution network line inspection image into the skeleton region detection model yields multiple skeleton region rectangles and multiple skeleton region types. The specific process is as follows: First, such as Figure 2 As shown, firstly, as Figure 2As shown, the distribution network line inspection images are processed through a backbone network (Conv3~Conv5) constructed by the C3K2 module to obtain the first layer feature map F3, the second layer feature map F4, and the third layer feature map F5. Then, F3~F5 are input into the feature pyramid network PA-SACM-FPN. The third layer feature map F5 and the second layer feature map F4 are processed by the first self-attention convolution module to generate the first self-attention convolution feature map SACF4. Then, the first self-attention convolution feature map SACF4 is concatenated with the second layer feature map F4, and then... A first fused feature map ~P4 is generated. Then, the first fused feature map ~P4 and the first layer feature map F3 are passed through a second self-attention convolution module to generate a second self-attention convolution feature map SACF3. Next, the second self-attention convolution feature map SACF3 is concatenated with the first layer feature map F3 to generate a second fused feature map P3. Then, the second fused feature map P3 is passed through a first convolution module to generate a first feature map PAF4 with the same resolution as the first fused feature map ~P4. The first feature map PAF4 is further concatenated with the first fused feature map ~P4. The process involves generating a third fused feature map P4. Then, the third fused feature map P4 is passed through a second convolution module to generate a second feature map PAF5 with the same resolution as the third layer feature map F5. The second feature map PAF5 is further concatenated with the third layer feature map F5 to generate a fourth fused feature map P5. Next, the second fused feature map P3, the third fused feature map P4, and the fourth fused feature map P5 are passed through the first, second, and third efficient attention convolution modules, respectively, to generate the first efficient attention convolution feature map T3, the second efficient attention convolution feature map T4, and the third efficient attention convolution feature map T5. This results in cross-level detail feature maps (T3~T5) arranged from top to bottom and bottom to top. Finally, the first efficient attention convolution feature map T3, the second efficient attention convolution feature map T4, and the third efficient attention convolution feature map T5 are classified and regressed by the first, second, and third classification regression modules in the decoupling head, respectively, to obtain multiple skeleton region rectangles and multiple skeleton region types, thus completing the detection task of the distribution network line skeleton region.

[0028] It should be noted that, in Figure 2 In this context, "concat" refers to the operation of stitching multiple images together along a certain dimension to facilitate subsequent calculations.

[0029] Understandably, such as Figure 3As shown, the working principle of the self-attention convolution module in the above process is as follows (taking the third-layer feature map F5 and the second-layer feature map F4 as inputs to the first self-attention convolution module to obtain the first self-attention convolution feature map SACF4 as an example): First, the low-layer high-resolution feature map (second-layer feature map F4) is convolved twice with 1×1 to obtain the first projection matrix Q and the second projection matrix K respectively; the first projection matrix Q and the second projection matrix K are similarly matched to obtain the similarity matching matrix, and the similarity matching matrix is ​​processed through the activation function to obtain the attention weighting matrix W; then, the high-layer low-resolution feature map (third-layer feature map F5) is convolved and upsampled sequentially to obtain an upsampled feature map with the same resolution as the low-layer high-resolution feature map (second-layer feature map F4), and the upsampled feature map is convolved with 1×1 to obtain the third projection matrix V; the third projection matrix V and the attention weighting matrix W are multiplied to obtain the first self-attention convolution feature map SACF4, and its specific formula is as follows:

[0030] In the formula, The first self-attention convolutional feature map is SACF4. Let K be the vector dimension for similarity matching in the second projection matrix K.

[0031] It should be noted that in the process of inputting the first fused feature map ~P4 and the first layer feature map F3 into the second self-attention convolution module to obtain the second self-attention convolution feature map SACF3, the first fused feature map ~P4 is a high-level low-resolution feature map, and the first layer feature map F3 is a low-level high-resolution feature map. The principle is the same as that of the first self-attention convolution module mentioned above, and will not be elaborated here in this embodiment.

[0032] Understandably, such as Figure 4As shown, the working principle of the efficient attention convolution module in the above process is as follows (taking the second fused feature map P3 as input to the first efficient attention convolution module to obtain the first efficient attention convolution feature map T3 as an example): First, according to the number of channels c of the second fused feature map P3, the second fused feature map P3 is divided into two fused feature sub-maps with the number of channels c / 2 and c / 2 respectively; the two fused feature sub-maps are subjected to spatial max pooling and spatial average pooling respectively, and then the channels are concatenated to obtain a channel feature map with the number of channels c; then, the channel feature map is convolved with a 1×1 convolution and activated by the ReLU function to obtain a channel-weighted feature map, and the second fused feature map P3 is then concatenated with the channel-weighted feature map. The weighted feature map is multiplied by a matrix to obtain the channel attention feature map. Then, based on the number of channels c in the channel attention feature map, the channel attention feature map is divided into two equal channel attention feature sub-maps with the number of channels c / 2 and c / 2, respectively. The two channel attention feature sub-maps are then subjected to channel max pooling and channel average pooling, and concatenated to obtain a spatial feature map with 2 channels. The spatial feature map is then convolved with a 1×1 convolution and activated by the ReLU function to obtain a spatial weighted feature map with 1 channel. Finally, the channel attention feature map and the spatial weighted feature map are multiplied by a matrix to obtain the first efficient attention convolution feature map T3.

[0033] It should be noted that the principle of inputting the third fused feature map P4 into the second efficient attention convolution module to obtain the second efficient attention convolution feature map T4, and inputting the fourth fused feature map P5 into the third efficient attention convolution module to obtain the third efficient attention convolution feature map T5, is the same as the principle of the first efficient attention convolution module mentioned above, and will not be elaborated here in this embodiment.

[0034] After obtaining multiple skeleton region rectangles and multiple skeleton region types through the skeleton region detection model, the multiple skeleton region rectangles are merged and marked to obtain the first error correction area and the first error correction area image of the distribution network line. Each skeleton region rectangle is marked separately to obtain multiple second-level error correction areas and multiple second-level error correction area types of the distribution network line.

[0035] The above steps employ a skeleton region detection model for distribution network line skeletonization, effectively separating the background region from the distribution network line region (the first error correction region) and refining the distribution area of ​​each distribution network device (the second error correction region). This not only narrows the detection range for both small and large target defects but also plays a crucial role in subsequent defect target error correction. Furthermore, the skeleton region detection model uses an improved YOLO11 detection network, utilizing a self-attention convolution module to impart detail attention to lower-level feature maps within high-level feature maps, and employing an efficient attention convolution module for rapid detail learning of target features. This not only enhances the high-level semantic feature representation of the target but also effectively controls the computational efficiency of target features, laying the foundation for both the accuracy and speed of skeleton region detection.

[0036] Step S120: Input the distribution network line inspection image into the large target recognition model to obtain multiple large target defect rectangles and multiple large target defect types; input the first error correction area image into the small target recognition model to obtain multiple small target defect rectangles and multiple small target defect types.

[0037] Furthermore, after the error-checking area marking process is completed, large target defect detection and small target defect detection can be performed simultaneously. First, the distribution network line inspection images are input into the large target defect detection model, and the large target recognition model is used to detect target defects, resulting in multiple large target defect rectangles and multiple large target defect types (including tower top damage, crossarm bending deformation, tower bird nests, conductor foreign objects, missing drop-out fuses, transformer bushings lacking insulation covers, vacuum bushings lacking insulation covers, etc.). Then, the first error-checking area image is cropped into a sliding window according to a preset resolution size (1500×1500 in this embodiment) to obtain a sliding window image set (in this embodiment, if the width or height of the first-level error-checking area is less than 1500 pixels, it will be expanded outward; if the width and height of the image are less than 1500 pixels, padding will be performed); the sliding window image set is input into the small target defect detection model, and the small target recognition model is used to detect target defects to obtain the defect rectangle and defect type of the sliding window image set; the coordinates of the defect rectangle of the sliding window image set are mapped back to the coordinates of the distribution network line inspection image to obtain multiple small target defect rectangles and multiple small target defect types (including safety pin detachment, connecting hardware ball head corrosion, spring pin detachment, etc.).

[0038] It should be noted that, in this embodiment, the training process of the large target defect detection model is as follows: obtain the original sample library of large target defects, and construct the augmented sample library of large target defects (the augmentation methods include left and right flipping, angle rotation, Gaussian noise, image mixing (mixup), etc.); merge the original sample library of large target defects and the augmented sample library of large target defects to obtain the large target defect training sample library; train the YOLO11 detection algorithm based on the large target defect training sample library to obtain the large target defect detection model.

[0039] It should be noted that, in this embodiment, the training process of the small target defect detection model is as follows: Based on the original sample library of small target defects, image cropping is performed using a sliding window (in this embodiment, the cropping resolution is 1500 pixels × 1500 pixels; if the cropped image width or height is less than 1500 pixels, it will be expanded outwards; if the width and height of the original image are less than 1500 pixels, padding will be performed), thereby obtaining the original sub-image library of small target defects; based on the original sample library of small target defects, an augmented sample library of small target defects is constructed (the augmentation methods include left and right flipping, angle rotation, Gaussian noise, etc.). Image mixing (mixup, etc.) and image cropping through a sliding window (in this embodiment, the cropping resolution is 1500 pixels × 1500 pixels; if the cropped image width or height is less than 1500 pixels, it will be expanded outward; if the width and height of the original image are less than 1500 pixels, padding will be performed) are used to obtain an augmented sub-image library of small target defects; the original sub-image library of small target defects and the augmented sub-image library of small target defects are merged to obtain a training sample library of small target defects; based on the training sample library of small target defects, the YOLO11 detection algorithm is trained to obtain a small target defect detection model.

[0040] The above steps employ a strategy of large target defect classification and detection combined with small target defect classification and detection. This approach specifically addresses the core issues of large differences in target scale and weak, easily missed features in small targets within distribution network line inspection images. It achieves full coverage detection of defects at different scales in distribution network lines, providing comprehensive and accurate initial detection results for subsequent defect filtering and troubleshooting.

[0041] Step S130: Filter and fuse the large target defect rectangles and large target defect types, and the small target defect rectangles and small target defect types to obtain multiple first residual defect rectangles and their corresponding first residual defect types.

[0042] Specifically, for large target defect rectangles and large target defect types: (1) First, for different large target defect types, set different first preset confidence thresholds, obtain the confidence of each large target defect rectangle, and delete rectangles with confidence lower than the first preset confidence threshold (e.g., 0.30) of their corresponding large target defect type to obtain multiple first large target defect rectangles; (2) Then, for different large target defect types, set different preset overlap rate thresholds, use the Non-Maximum Suppression (NMS) algorithm and preset overlap rate formula to calculate the overlap rate of each first large target defect rectangle, and delete rectangles with overlap rates greater than the preset overlap rate threshold (e.g., 0.70) to obtain multiple second large target defect rectangles.

[0043] Understandably, the default overlap rate formula is:

[0044] In the formula, This indicates the overlap rate between rectangle ABCD and rectangles EFGH. Let be the area of ​​the intersection of rectangles ABCD and EFGH. Let be the area of ​​rectangle ABCD. Let EFGH be the area of ​​the rectangle. Let A be the x-coordinate of the top-left point A of rectangle ABCD. Let A be the ordinate of the top-left point A of rectangle ABCD. Let C be the x-coordinate of the bottom right point C of rectangle ABCD. Let C be the ordinate of the bottom right point C of rectangle ABCD. Let E be the x-coordinate of the top-left point E of the rectangle EFGH. Let E be the ordinate of the top-left point E of the rectangle EFGH. Let G be the x-coordinate of the bottom right point G of the rectangle EFGH. Let G be the ordinate of the bottom right point G of the rectangle EFGH.

[0045] Further, regarding the small target defect rectangles and small target defect types: (1) First, for different small target defect types, different second preset confidence thresholds are set to obtain the confidence of each small target defect rectangle, and rectangles with confidence lower than the second preset confidence threshold (e.g., 0.30) are deleted to obtain multiple first small target defect rectangles; (2) Then, for different small target defect types, different preset mutation overlap rate thresholds are set to calculate the mutation overlap rate of each first small target defect rectangle using the Area Non-Maximum Suppression (ANMS) algorithm, and rectangles with mutation overlap rates greater than the preset mutation overlap rate threshold (e.g., 0.80) are deleted to obtain multiple second small target defect rectangles.

[0046] In the above process, since the small target defect detection model detects sliding window images, some sliding window images contain the complete target contour, while others only contain a part of the target contour. After detection by the small target defect detection model, the bounding boxes of the complete target contour and the bounding boxes of the partial target contour are obtained. Therefore, it is necessary to use the area non-maximum suppression processing algorithm to retain the bounding boxes of the complete target contour detected by the small target defect and delete the bounding boxes of the partial target contour detected by the small target defect.

[0047] Preferably, the process of the area non-maximum suppression algorithm is as follows: First, all rectangles are sorted according to their area from largest to smallest (or from largest to smallest). Then, the rectangle with the largest area is selected as the registration rectangle. The variation overlap rate between each remaining rectangle and the registration rectangle is calculated using a preset variation overlap rate formula. Thus, each remaining rectangle corresponds to a variation overlap rate. If the variation overlap rate is greater than a preset variation overlap rate threshold (e.g., 0.80), the corresponding remaining rectangle is deleted. If the variation overlap rate is less than or equal to the preset variation overlap rate threshold, the corresponding remaining rectangle is retained. Then, in From the remaining rectangles, the rectangle with the largest area is selected as the new registration rectangle. The variation overlap rate is calculated between the other remaining rectangles and the new registration rectangle using a preset variation overlap rate formula. Remaining rectangles with variation overlap rates greater than a preset variation overlap rate threshold are deleted, while those with variation overlap rates less than or equal to the preset variation overlap rate threshold are retained. This process is repeated until no rectangle has a variation overlap rate greater than the preset variation overlap rate threshold among all selected registration rectangles. Finally, all selected registration rectangles are used as the output rectangles of the final algorithm.

[0048] Understandably, the presupposed formula for the variation overlap rate is:

[0049] In the formula, This represents the variation overlap rate between rectangles ABCD and EFGH. This represents the area of ​​the intersection of rectangles ABCD and EFGH. Let be the area of ​​rectangle ABCD. Let EFGH be the area of ​​the rectangle. Let A be the x-coordinate of the top-left point A of rectangle ABCD. Let A be the ordinate of the top-left point A of rectangle ABCD. Let C be the x-coordinate of the bottom right point C of rectangle ABCD. Let C be the ordinate of the bottom right point C of rectangle ABCD. Let E be the x-coordinate of the top-left point E of the rectangle EFGH. Let E be the ordinate of the top-left point E of the rectangle EFGH. Let G be the x-coordinate of the bottom right point G of the rectangle EFGH. Let G be the ordinate of the bottom right point G of the rectangle EFGH.

[0050] It should be noted that the first preset confidence threshold and the second preset confidence threshold can be the same or different, and the preset overlap rate threshold and the preset variation overlap rate threshold can be the same or different. This embodiment does not limit this.

[0051] Finally, the second largest target defect rectangles and their corresponding large target defect types, and the second smallest target defect rectangles and their corresponding small target defect types are merged to obtain multiple first remaining defect rectangles and their corresponding first remaining defect types.

[0052] The above steps remove low-confidence false detection boxes by using differentiated confidence thresholds, and use NMS and ANMS deduplication strategies for large and small targets respectively to effectively delete redundant duplicate boxes and partial outline boxes of small targets. Finally, high-confidence, non-redundant first residual defect data is obtained, which greatly reduces the burden of subsequent error debugging and lays a high-quality data foundation for accurate error debugging.

[0053] Step S140: Use the first error-checking rule to check each first remaining defect rectangle and its corresponding first remaining defect type to obtain multiple second remaining defect rectangles and their corresponding second remaining defect types. Use the second error-checking rule to check each second remaining defect rectangle and its corresponding second remaining defect type to obtain multiple third remaining defect rectangles and their corresponding third remaining defect types. Then, determine each third remaining defect rectangle and its corresponding third remaining defect type as the error-checking result of the distribution network line defect.

[0054] Understandably, when large-target defect detection models and small-target defect detection models detect target defects, many detected defects are not on the distribution network lines but in other background areas. To eliminate such obvious false alarms, this application designs a first error correction rule. Since the first error correction area is a merged frame of the skeleton area, its range is relatively large, and due to the influence of the skeleton area, its range fluctuates significantly. In contrast, the detection frame range for the first residual defect types (including tower top damage, crossarm bending deformation, tower bird nests, conductor foreign objects, missing drop-out fuses, transformer bushings lacking insulation covers, vacuum bushings lacking insulation covers, detached safety pins, corroded connecting hardware ball heads, detached spring pins, etc.) is relatively small. Because it is not affected by other areas, the detection frame range for the first residual defect types does not fluctuate much. In order to exclude the first remaining defect detection boxes outside the first error-checking area and obtain a reasonable quantitative overlap rate, this application uses a first error-checking rule to check each first remaining defect rectangle and its corresponding first remaining defect type, which can effectively exclude defect detection boxes outside the first error-checking area. The specific process of the first error-checking rule is as follows: (1) For each type of first remaining defect, set the corresponding first variation overlap rate threshold; (2) Traverse each first remaining defect rectangle, calculate the variation overlap rate between each first remaining defect rectangle and the first error correction region, and use the first error correction formula to exclude the first target remaining defect rectangle. The variation overlap rate of the first target remaining defect rectangle is less than the first variation overlap rate threshold of the first remaining defect type. (3) After all the first remaining defect rectangles have completed the above operations, the remaining first remaining defect rectangles and their corresponding first remaining defect types are determined as multiple second remaining defect rectangles and their corresponding second remaining defect types, and the troubleshooting process of the first remaining defect rectangles ends.

[0055] The first error correction formula is:

[0056] In the formula, Indicates whether to exclude the first residual defect type. The The first remaining defect rectangle ,when When the value is 1, the corresponding first remaining defect rectangle is... The remaining defect rectangle is used as the first target and is excluded when... When the value is 0, the corresponding first remaining defect rectangle is... reserve; Indicates the first residual defect type as The The first remaining defect rectangle With the first misaligned area The overlap rate of the variation, Indicates the first residual defect type as The first variation overlap rate threshold.

[0057] It should be noted that the specific positional relationship between the first misalignment area and the first remaining defect rectangle is as follows: Generally speaking, defects such as tower top damage, crossarm bending deformation, missing drop-out fuses, transformer bushings lacking insulation covers, vacuum bushings lacking insulation covers, detached safety pins, corroded connecting hardware ball heads, and dislodged spring pins are all faults occurring in the distribution network line components. Their corresponding locations are relatively fixed, almost all within the first misalignment area. Therefore, the first variation overlap rate threshold set for these defect types is generally relatively high. On the other hand, defects such as bird nests on poles and foreign objects on conductors are not defects inherent to the distribution network line itself, but rather external hazards. Their corresponding locations are not fixed and may not be entirely within the first misalignment area, but some areas of these defects are definitely located within the first misalignment area. Therefore, the first variation overlap rate threshold set for these defect types is relatively low.

[0058] Understandably, when large-target defect detection models and small-target defect detection models detect defects, many detected defects, although located on the distribution network line, may not fall within the actual defect area according to prior knowledge of the distribution network line's defects. To eliminate this common false alarm, this application designs a second error correction rule. Since the second error correction area is an independent frame of the skeleton area, it involves different second error correction areas of different types (including pole areas, crossarm equipment areas, transformer areas, vacuum equipment areas, connecting hardware areas, conductor areas, etc.). At the same time, distribution network line defects are also independent frames, thus involving different defect rectangles of different defect types (defect types include tower top damage, crossarm bending deformation, tower bird nests, conductor foreign objects, missing drop-out fuses, transformer bushings lacking insulation covers, vacuum bushings lacking insulation covers, detached safety pins, corroded connecting hardware ball heads, detached spring pins, etc.). In order to exclude second residual defect detection boxes outside the second error-checking region and obtain a reasonable quantified overlap rate, this application uses a second error-checking rule to check each second residual defect rectangle and its corresponding second residual defect type. This can effectively exclude defect detection boxes outside the second error-checking region. The specific process of the second error-checking rule is as follows: (1) Associate multiple second error correction region types with each second residual defect type, and set a corresponding second variation overlap rate threshold for each second error correction region type; (2) Traverse each second remaining defect rectangle, and search for the second error correction area of ​​the multiple second error correction area types associated with each second remaining defect rectangle according to the second remaining defect type, and obtain the search results of the second error correction area list; (3) Determine whether the search results of the second error correction area list are empty. If so, retain each second remaining defect rectangle and its corresponding second remaining defect type, and the error correction process of the second remaining defect rectangle ends. Otherwise, obtain the corresponding second error correction area list and the corresponding second error correction area type list. The second error correction area type list includes multiple second error correction area types. (4) Traverse each second error correction region in the second error correction region list, calculate the variation overlap rate of each second remaining defect rectangle and its corresponding multiple second error correction regions, and use the second error correction formula to exclude the second target remaining defect rectangle. The variation overlap rate of the second target remaining defect rectangle is less than the second variation overlap rate threshold of the multiple second error correction region types associated with its corresponding second remaining defect type. (5) After all the second remaining defect rectangles have completed the above operations, the remaining second remaining defect rectangles and their corresponding second remaining defect types are determined as multiple third remaining defect rectangles and their corresponding third remaining defect types. The troubleshooting process of the second remaining defect rectangles ends. Finally, the multiple third remaining defect rectangles and their corresponding third remaining defect types are determined as the troubleshooting results of the distribution network line defects.

[0059] The second error correction formula is:

[0060] In the formula, Indicate whether to exclude the second residual defect type. The The second remaining defect rectangle ,when When the value is 1, the corresponding second remaining defect rectangle is... As the second objective, the remaining defect rectangle is excluded when... When the value is 0, the corresponding second remaining defect rectangle is... reserve; Indicates the second residual defect type as The associated list of second error correction region types, Indicates the second residual defect type The associated first The second type of misaligned area, Indicates the second residual defect type as The The second remaining defect rectangle The second misalignment area type is The The second misaligned area The overlap rate of the variation, Indicates the second residual defect type as The associated second error correction region type is The second variation overlap rate threshold.

[0061] It should be noted that the specific positional relationship between the second misalignment area and the second remaining defect rectangle is as follows: Generally speaking, tower top damage is located within the pole area; crossarm bending and deformation, and missing drop-out fuses are located within the crossarm equipment area; transformer bushing insulation cover missing is located at the edge of the transformer area; vacuum bushing insulation cover missing is located near the vacuum equipment area; defects such as detached safety pins, corroded connecting hardware ball heads, and dislodged spring pins are all located within the connecting hardware area; and foreign objects in the conductors are located around the conductor area. However, the types of second misalignment areas corresponding to tower nests are numerous and may be located around the pole area, crossarm equipment area, transformer area, vacuum equipment area, etc. Therefore, for these defect types, different second misalignment area types will be associated with each defect type based on the different positional relationships between the defect area and the second misalignment area, and then different second variation overlap rate thresholds will be set for each second misalignment area type.

[0062] The above steps employ the first and second error-checking rules, which can not only effectively filter out a large number of false defect detection boxes in the background area (defect detection boxes outside the first error-checking area), but also effectively remove false defect detection boxes that do not belong to the actual defect occurrence area (defect detection boxes outside the second error-checking area), thereby obtaining effective and correct defect-checking results.

[0063] Step S150: Determine whether the fault diagnosis result of the distribution network line is empty. If so, record the fault diagnosis result defect list of the distribution network line inspection image as empty. Otherwise, record each third remaining defect rectangle and its corresponding third remaining defect type to the fault diagnosis result defect list and save it.

[0064] Understandably, after obtaining the results of troubleshooting distribution network line defects, it is determined whether the results are empty. If the results are empty, it means that no third residual defect rectangle or third residual defect type was detected. In this case, the defect list of the troubleshooting results for the distribution network line inspection image is recorded as empty, and the corresponding record is saved. If the results are not empty, it means that the third residual defect rectangle and third residual defect type were detected. In this case, each third residual defect rectangle and its corresponding third residual defect type are recorded in the defect list of the troubleshooting results, and the corresponding record is saved. The troubleshooting process for this distribution network line inspection image ends.

[0065] The above steps form a closed loop of "detection-troubleshooting-recording" by judging the troubleshooting results and recording them in a standardized manner. This ensures that the results of each inspection map are traceable, provides maintenance personnel with accurate information on the location and type of defects, and the accumulated structured data can support preventive maintenance and model iteration, thus helping to standardize distribution network inspections and continuously optimize performance.

[0066] The distribution network line skeleton-based defect troubleshooting method provided in this application embodiment is based on a deep learning algorithm. First, a skeleton region detection model is used to obtain the skeleton region rectangle and skeleton region type, thereby obtaining the first-level troubleshooting region, the first-level troubleshooting region image, the second-level troubleshooting region, and the second-level troubleshooting region type. Then, a large target defect detection model is used to detect large target defect types, and a small target defect detection model is used to detect small target defect types. Finally, the first troubleshooting rule and the second troubleshooting rule are used sequentially to perform troubleshooting processing for large target defect types and small target defect types, obtaining the final distribution network line defect troubleshooting result. This method can greatly improve the troubleshooting efficiency of power inspection work and effectively prevent false detections.

[0067] Example 2 like Figure 5 The diagram shown is a structural schematic of a distribution network line skeleton-type defect troubleshooting system 500 according to an embodiment of this application. The system includes: The marking module 510 is used to acquire distribution network line inspection images, input distribution network line inspection images into skeleton area detection model to obtain multiple skeleton area rectangles and multiple skeleton area types, merge and mark each skeleton area rectangle to obtain the first error correction area and the first error correction area image, and mark each skeleton area rectangle individually to obtain multiple second error correction areas and multiple second error correction area types. The identification module 520 is used to input the distribution network line inspection image into the large target identification model to obtain multiple large target defect rectangles and multiple large target defect types, and input the first error correction area image into the small target identification model to obtain multiple small target defect rectangles and multiple small target defect types. The fusion module 530 is used to filter and fuse the large target defect rectangles with the large target defect types and the small target defect rectangles with the small target defect types to obtain multiple first residual defect rectangles and their corresponding first residual defect types. The error troubleshooting module 540 is used to troubleshoot each first remaining defect rectangle and its corresponding first remaining defect type using the first error troubleshooting rule to obtain multiple second remaining defect rectangles and their corresponding second remaining defect types. It then uses the second error troubleshooting rule to troubleshoot each second remaining defect rectangle and its corresponding second remaining defect type to obtain multiple third remaining defect rectangles and their corresponding third remaining defect types. Finally, it determines each third remaining defect rectangle and its corresponding third remaining defect type as the error troubleshooting result of the distribution network line defect. The recording module 550 is used to determine whether the fault diagnosis result of the distribution network line is empty. If so, the fault diagnosis result defect list of the distribution network line inspection image is recorded as empty. Otherwise, each third remaining defect rectangle and its corresponding third remaining defect type are recorded in the fault diagnosis result defect list and saved.

[0068] The distribution network line skeleton defect troubleshooting system provided in this application embodiment can realize each process of the distribution network line skeleton defect troubleshooting method corresponding to Embodiment 1, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0069] The distribution network line skeleton-type defect troubleshooting system provided in this application embodiment is based on a deep learning algorithm. First, it uses a skeleton region detection model to obtain the skeleton region rectangle and skeleton region type, thereby obtaining the first-level troubleshooting region, the first-level troubleshooting region image, the second-level troubleshooting region, and the second-level troubleshooting region type. Then, it uses a large target defect detection model to detect large target defect types and a small target defect detection model to detect small target defect types. Finally, it sequentially uses the first troubleshooting rule and the second troubleshooting rule to perform troubleshooting processing for large target defect types and small target defect types, thereby obtaining the final distribution network line defect troubleshooting result. This can greatly improve the troubleshooting efficiency of power inspection work and effectively prevent false detections.

[0070] This disclosure also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the distribution network line skeleton defect troubleshooting method described in Embodiment 1.

[0071] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the distribution network line skeleton defect troubleshooting method described in Embodiment 1.

[0072] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A network configuration line skeleton type defect troubleshooting method, characterized in that, The method comprises: obtaining a distribution network line inspection image, inputting the distribution network line inspection image into a skeleton region detection model to obtain a plurality of skeleton region rectangular frames and a plurality of skeleton region types, marking each skeleton region rectangular frame with a combined frame to obtain a first fault area and a first fault area image, marking each skeleton region rectangular frame with an individual frame to obtain a plurality of second fault areas and a plurality of second fault area types; inputting the distribution network line inspection image into a large target recognition model to obtain a plurality of large target defect rectangular frames and a plurality of large target defect types, inputting the first fault area image into a small target recognition model to obtain a plurality of small target defect rectangular frames and a plurality of small target defect types; filtering and fusing each large target defect rectangular frame, each large target defect type, each small target defect rectangular frame and each small target defect type to obtain a plurality of first remaining defect rectangular frames and corresponding first remaining defect types; using a first fault elimination rule to eliminate faults of each first remaining defect rectangular frame and corresponding first remaining defect type to obtain a plurality of second remaining defect rectangular frames and corresponding second remaining defect types, using a second fault elimination rule to eliminate faults of each second remaining defect rectangular frame and corresponding second remaining defect type to obtain a plurality of third remaining defect rectangular frames and corresponding third remaining defect types, and determining each third remaining defect rectangular frame and corresponding third remaining defect type as a distribution network line defect elimination result; judging whether the distribution network line defect elimination result is empty, if yes, recording an elimination result defect list of the distribution network line inspection image as empty, if not, recording each third remaining defect rectangular frame and corresponding third remaining defect type in the elimination result defect list, and saving.

2. The network provisioning line skeleton type defect troubleshooting method according to claim 1, characterized by, The method comprises: for each first remaining defect type, setting a corresponding first variation overlap rate threshold value; traversing each first remaining defect rectangular frame, calculating the variation overlap rate of each first remaining defect rectangular frame and the first fault area, and using a first fault elimination formula to exclude a first target remaining defect rectangular frame, the variation overlap rate of the first target remaining defect rectangular frame being less than the first variation overlap rate threshold value of the first remaining defect type corresponding thereto, to obtain a plurality of second remaining defect rectangular frames and corresponding second remaining defect types; wherein the first fault elimination formula is: In the formula, indicates whether the first remaining defect type is excluded from the first remaining defect type of the first remaining defect rectangular frame When is 1, the corresponding first remaining defect rectangular frame is taken as the first target remaining defect rectangular frame, and is excluded when is 0, the corresponding first remaining defect rectangular frame is retained; indicates the first variation overlap rate of the first remaining defect type of the first remaining defect rectangular frame with the first error correction area , indicates the first variation overlap rate threshold of the first remaining defect type of the first remaining defect type .​​​​ 3. The network provisioning line skeleton type defect troubleshooting method according to claim 1, characterized by, The method comprises: for each second remaining defect type, associating a plurality of second fault area types, and for each second fault area type, setting a corresponding second variation overlap rate threshold value; traversing each of the second remaining defect rectangular frames, according to the second remaining defect type corresponding to each of the second remaining defect rectangular frames, searching for the second error correction area of the plurality of second error correction area types associated therewith to obtain a second error correction area list search result; determining whether the second error correction area list search result is empty, if yes, retaining each of the second remaining defect rectangular frames and the second remaining defect type corresponding thereto, if not, obtaining a corresponding second error correction area list and a corresponding second error correction area type list, the second error correction area type list including a plurality of the second error correction area types; traversing each of the second error correction areas in the second error correction area list, calculating the variation overlap rate of each of the second remaining defect rectangular frames and the plurality of second error correction areas corresponding thereto, and using a second error correction formula to exclude a second target remaining defect rectangular frame, the variation overlap rate of the second target remaining defect rectangular frame being less than the second variation overlap rate threshold of the plurality of second error correction area types associated with the second remaining defect type corresponding thereto, to obtain a plurality of third remaining defect rectangular frames and third remaining defect types corresponding thereto; wherein the second error correction formula is: In the formula, indicates whether to exclude the second remaining defect type of the first second remaining defect rectangular frame When is 1, the corresponding second remaining defect rectangular frame is taken as the second target remaining defect rectangular frame, and is excluded, and when is 0, the corresponding second remaining defect rectangular frame is retained; indicates the second remaining defect type associated second error area type list, indicates the second remaining defect type associated first second error area type, indicates the second remaining defect type of the first second remaining defect rectangular frame and the second error area type of the first second error area second variation overlap rate, indicates the second remaining defect type associated second error area type second variation overlap rate threshold.

4. The network provisioning line skeleton type defect troubleshooting method according to claim 1, characterized by, the filtering and fusion of each of the large target defect rectangular frames and each of the large target defect types, each of the small target defect rectangular frames and each of the small target defect types, to obtain a plurality of first remaining defect rectangular frames and first remaining defect types corresponding thereto, including: obtaining the confidence of each of the large target defect rectangular frames, and deleting the rectangular frame with a confidence lower than a first preset confidence threshold to obtain a plurality of first large target defect rectangular frames; calculating the overlap rate of each of the first large target defect rectangular frames using a non-maximum suppression processing algorithm, and deleting the rectangular frame with an overlap rate greater than a preset overlap rate threshold to obtain a plurality of second large target defect rectangular frames; obtaining the confidence of each of the small target defect rectangular frames, and deleting the rectangular frame with a confidence lower than a second preset confidence threshold to obtain a plurality of first small target defect rectangular frames; calculating the variation overlap rate of each of the first small target defect rectangular frames using an area non-maximum suppression processing algorithm, and deleting the rectangular frame with a variation overlap rate greater than a preset variation overlap rate threshold to obtain a plurality of second small target defect rectangular frames; merging each of the second large target defect rectangular frames and the large target defect type corresponding thereto, each of the second small target defect rectangular frames and the small target defect type corresponding thereto, to obtain a plurality of first remaining defect rectangular frames and first remaining defect types corresponding thereto.

5. The network provisioning line skeleton type defect troubleshooting method according to claim 1, characterized by, the input of the distribution network line inspection image into a large target recognition model to obtain a plurality of large target defect rectangular frames and a plurality of large target defect types, the input of the first error correction area image into a small target recognition model to obtain a plurality of small target defect rectangular frames and a plurality of small target defect types, including: inputting the distribution network line inspection image into the large target recognition model, and performing target defect detection using the large target recognition model to obtain a plurality of the large target defect rectangular frames and a plurality of the large target defect types; The first error region image is cropped by a sliding window according to a preset cropping resolution to obtain a sliding window image set, the sliding window image set is input into the small target recognition model, target defect detection is performed by using the small target recognition model, and a defect rectangular frame and a defect type of the sliding window image set are obtained. The coordinates of the defect rectangular frame of the sliding window image set are mapped back to the coordinates of the network distribution line inspection image to obtain a plurality of small target defect rectangular frames and a plurality of small target defect types.

6. The network provisioning line skeleton type defect troubleshooting method according to claim 1, wherein The skeleton region detection model is based on an improved YOLO11 detection algorithm, the improved YOLO11 detection algorithm includes a backbone network, a feature pyramid network and a decoupling head, the feature pyramid network includes a first self-attention convolution module, a second self-attention convolution module, a first convolution module, a second convolution module, a first efficient attention convolution module, a second efficient attention convolution module and a third efficient attention convolution module, the decoupling head includes a first classification regression module, a second classification regression module and a third classification regression module, and the network distribution line inspection image is input into the skeleton region detection model to obtain a plurality of skeleton region rectangular frames and a plurality of skeleton region types, including: The network distribution line inspection image is input into the backbone network to obtain a first layer feature map, a second layer feature map and a third layer feature map; The third layer feature map and the second layer feature map are input into the first self-attention convolution module to obtain a first self-attention convolution feature map, and the first self-attention convolution feature map and the second layer feature map are concatenated to obtain a first fusion feature map; The first fusion feature map and the first layer feature map are input into the second self-attention convolution module to obtain a second self-attention convolution feature map, and the second self-attention convolution feature map and the first layer feature map are concatenated to obtain a second fusion feature map; The second fusion feature map is input into the first convolution module to obtain a first feature map, and the first feature map and the first fusion feature map are concatenated to obtain a third fusion feature map; The third fusion feature map is input into the second convolution module to obtain a second feature map, and the second feature map and the third layer feature map are concatenated to obtain a fourth fusion feature map; The second fusion feature map is input into the first efficient attention convolution module to obtain a first efficient attention convolution feature map, the third fusion feature map is input into the second efficient attention convolution module to obtain a second efficient attention convolution feature map, and the fourth fusion feature map is input into the third efficient attention convolution module to obtain a third efficient attention convolution feature map; The first efficient attention convolution feature map, the second efficient attention convolution feature map and the third efficient attention convolution feature map are respectively input into the first classification regression module, the second classification regression module and the third classification regression module to obtain a plurality of skeleton region rectangular frames and a plurality of skeleton region types.

7. The network provisioning line skeleton type defect troubleshooting method according to claim 6, wherein, The third layer feature map and the second layer feature map are input into the first self-attention convolution module to obtain a first self-attention convolution feature map, including: The second layer feature map is subjected to 1*1 convolution twice respectively to obtain a first projection matrix and a second projection matrix respectively; The first projection matrix and the second projection matrix are subjected to similarity matching to obtain a similarity matching matrix, and the similarity matching matrix is processed by an activation function to obtain an attention weighting matrix; The third layer feature map is subjected to 1*1 convolution and up-sampling in sequence to obtain an up-sampled feature map, and the up-sampled feature map is subjected to 1*1 convolution to obtain a third projection matrix; The third projection matrix and the attention weighting matrix are subjected to matrix multiplication operation to obtain the first self-attention convolution feature map; The second fusion feature map is input into the first efficient attention convolution module to obtain a first efficient attention convolution feature map, including: According to the number of channels of the second fusion feature map, the second fusion feature map is evenly divided into two fusion feature subgraphs, the two fusion feature subgraphs are subjected to spatial maximum pooling and spatial average pooling respectively, and channel splicing is performed to obtain a channel feature graph; The channel feature graph is subjected to 1*1 convolution and is processed by a Relu activation function to obtain a channel weighting feature graph, and the second fusion feature graph and the channel weighting feature graph are subjected to matrix multiplication operation to obtain a channel attention feature graph; According to the number of channels of the channel attention feature graph, the channel attention feature graph is evenly divided into two channel attention feature subgraphs, the two channel attention feature subgraphs are subjected to channel maximum pooling and channel average pooling respectively, and channel splicing is performed to obtain a spatial feature graph; The spatial feature graph is subjected to 1*1 convolution and is processed by the Relu activation function to obtain a spatial weighting feature graph, and the channel attention feature graph and the spatial weighting feature graph are subjected to matrix multiplication operation to obtain the first efficient attention convolution feature graph.

8. A network configuration line skeleton type defect troubleshooting system, characterized by, The system comprises: A marking module is configured to acquire a distribution network line inspection image, input the distribution network line inspection image into a skeleton region detection model, obtain a plurality of skeleton region rectangular frames and a plurality of skeleton region types, mark each skeleton region rectangular frame with a combined frame to obtain a first fault area and a first fault area image, and mark each skeleton region rectangular frame with an individual frame to obtain a plurality of second fault areas and a plurality of second fault area types. An identification module is configured to input the distribution network line inspection image into a large target identification model to obtain a plurality of large target defect rectangular frames and a plurality of large target defect types, and input the first fault area image into a small target identification model to obtain a plurality of small target defect rectangular frames and a plurality of small target defect types. A fusion module is configured to filter and fuse each large target defect rectangular frame and each large target defect type, each small target defect rectangular frame and each small target defect type to obtain a plurality of first residual defect rectangular frames and corresponding first residual defect types. The error elimination module is configured to eliminate errors of each of the first remaining defect rectangular frames and the corresponding first remaining defect type by using a first error elimination rule to obtain a plurality of second remaining defect rectangular frames and the corresponding second remaining defect type, eliminate errors of each of the second remaining defect rectangular frames and the corresponding second remaining defect type by using a second error elimination rule to obtain a plurality of third remaining defect rectangular frames and the corresponding third remaining defect type, and determine each of the third remaining defect rectangular frames and the corresponding third remaining defect type as the distribution network line defect elimination result. The recording module is configured to determine whether the distribution network line defect elimination result is empty, record the error elimination result defect list of the distribution network line inspection image as empty if the distribution network line defect elimination result is empty, or record each of the third remaining defect rectangular frames and the corresponding third remaining defect type in the error elimination result defect list and save if the distribution network line defect elimination result is not empty.

9. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the distribution network line skeleton type defect elimination method in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the distribution network line skeleton type defect elimination method in any one of claims 1-7.

Images