A method for detecting defects of a 220kV power cable intermediate joint compression tube

Abstract

The application provides a kind of 220kV power cable intermediate joint crimping tube defect detection method, method includes: collecting power equipment joint image, using lightweight UNet network for pretreatment;The processed image is input into improved YOLO11 network for defect detection, to obtain defect class, pixel position and confidence.The improved YOLO11 network includes: all shallow C3K2 modules of Backbone layer are replaced by C3K2-SEDM sparse edge detail enhancement module;All original up-sampling layers of Neck layer are replaced by CARAFE-FD defect perception up-sampling module;The Detect detection head of Head layer is replaced by ST-DFL spatial perception small defect detection head.The application realizes the high-precision automatic detection of 220kV power cable joint surface burr and scratch, improves the quality control efficiency and cable line safety.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to a method for detecting defects in crimped pipes of intermediate joints in 220kV power cables. Background Technology

[0002] 220kV power cables, as a key component of high-voltage transmission systems, are widely used in urban core load areas, underground utility tunnels, and other scenarios. Their operational reliability directly affects the safe and stable power supply of the power grid. Cable joints are the weakest link in cable lines, undertaking multiple core functions such as electrical connection, insulation protection, and mechanical fixation. Their manufacturing quality is a key factor determining the long-term safe operation of cable lines.

[0003] As the core conductive component of a 220kV power cable intermediate joint, the crimping tube achieves a reliable electrical connection of the cable conductors through hydraulic crimping. Its surface quality directly determines the contact resistance and electric field distribution characteristics of the joint. During the processing and on-site crimping of the crimping tube, two typical micro-defects are easily generated on the surface due to the extrusion process: one is burrs, which are mostly manifested as metal flash or small protrusions on the crimping edge or processed surface, which are small in size and irregular in shape; the other is scratches, which are usually caused by wear of the crimping die, conductor extrusion, or surface scratches, and appear as linear fine texture defects with shallow depth and narrow width, which are difficult to identify with the naked eye.

[0004] Although these defects are small in size, they can pose serious safety hazards in a 220kV high-voltage electric field environment: First, burrs and scratches alter the electric field distribution on the surface of the crimped connector, leading to localized electric field concentration. When the local field strength exceeds the withstand field strength of the insulating medium, insulation breakdown is highly likely, resulting in flashover or burnout of the connector. Second, burrs may pierce the semiconductive or insulating layer inside the connector, damaging the integrity of the insulation structure, forming creepage paths, and accelerating insulation aging. Finally, scratches increase the contact resistance of the crimped connector, which can easily cause localized overheating under long-term load current, further exacerbating insulation degradation, shortening the connector's service life, and even leading to power outages.

[0005] Currently, the detection of surface defects in crimped connectors mainly relies on manual visual inspection. However, due to factors such as uneven lighting, background interference, differences in inspector experience, and fatigue, the rate of missed detection for minute burrs and fine scratches is high, resulting in low inspection efficiency and making it difficult to meet the needs of mass production and quality control. Furthermore, images of crimped connectors acquired on-site suffer from interference factors such as complex backgrounds, uneven lighting, and the target being obscured by other components of the connector. Traditional machine vision inspection methods, including the general-purpose YOLO series of target detection algorithms, suffer from insufficient ability to extract weak edge features, poor background noise suppression, and a high rate of missed detection for minute defects when faced with such minute defects, thus failing to achieve stable and reliable automated inspection. Summary of the Invention

[0006] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method for detecting defects in the crimped pipe of intermediate joints of 220kV power cables.

[0007] To achieve the above-mentioned objectives of this invention, this invention provides a method for detecting defects in the crimped pipe of intermediate joints in 220kV power cables, the method comprising: Images of power equipment joints are acquired and preprocessed using a lightweight UNet segmentation network to obtain preprocessed joint images. The processed connector image is input into an improved YOLO11 network for defect detection to obtain inspection results including defect type, pixel location, and detection confidence. The improved YOLO11 network includes: Replace all shallow C3K2 modules in the original YOLO11 network Backbone layer with C3K2-SEDM sparse edge detail enhancement modules; Replace all native upsampling layers in the original YOLO11 network Neck layer with the CARAFE-FD defect-aware upsampling module; Replace the Detect head in the original YOLO11 network Head layer with the ST-DFL spatial awareness small defect detection head.

[0008] The beneficial effects of this invention are as follows: This invention preprocesses the original joint image using a lightweight UNet semantic segmentation network, optimizes the initial mask for denoising using a breadth-first search algorithm, and automatically crops the defect foreground region based on the minimum bounding rectangle. This effectively eliminates interference factors such as complex backgrounds, uneven lighting, and target occlusion by other parts of the joint in the on-site image, providing a high-quality region of interest for subsequent detection. In the Backbone layer, a C3K2-SEDM sparse edge detail enhancement module is introduced. Through the synergistic effect of the sparse edge perception branch and the detail feature weight gating mechanism, the edge detail feature response of minor defects such as burrs and scratches is adaptively amplified, significantly improving the network's feature extraction capability for weak edge defects. In the Neck layer… A CARAFE-FD defect-aware upsampling module replaces the native upsampling layer. It utilizes orientation-aware asymmetric convolution and depthwise separable convolution to learn an adaptive spatial upsampling kernel adapted to the shape of slender scratches, and integrates shallow detail features to supplement edge information lost during upsampling. This effectively suppresses background noise while retaining the advantages of content-aware upsampling. In the Head layer, an ST-DFL spatial-aware small defect detection head is designed. Through a collaborative mechanism of classification, regression, and small defect confidence weighting, spatial orientation-aware distribution modeling and center-focused narrow distribution modeling are used for linear scratches and point-like burrs, respectively. An adaptive weight adjustment mechanism amplifies the feature response of small defects and suppresses background interference, significantly reducing the false negative rate of small defects. In summary, this invention, through multi-level collaborative optimization of preprocessing denoising, edge detail enhancement, defect-aware upsampling, and spatial-aware detection head, achieves high-precision, high-reliability automated detection of burrs and scratches on the surface of crimped pipes of 220kV power cable intermediate joints. This effectively improves the efficiency and accuracy of cable joint manufacturing and quality control, ensuring the safe and stable operation of high-voltage cable lines.

[0009] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0010] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a method for detecting defects in the crimped pipe of a 220kV power cable intermediate joint according to the present invention; Figure 2 This is a schematic diagram of the structure of the improved YOLO11 network of this invention; Figure 3 This is an image of the electrical equipment connector of the present invention; Figure 4 This is the image of the connector after processing according to the present invention; Figure 5 This is a schematic diagram of the C3K2-SEDM sparse edge detail enhancement module of the present invention; Figure 6 This is a schematic diagram of the CARAFE-FD defect sensing upsampling module of the present invention; Figure 7 This is a schematic diagram of the ST-DFL spatial sensing small defect detection head of the present invention; Figure 8 This is a diagram showing the inspection results of the improved YOLO11 network of this invention. Detailed Implementation

[0011] 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.

[0012] like Figures 1 to 8 As shown, a method for detecting defects in the crimped pipe of a 220kV power cable intermediate joint is provided, the method comprising: Images of power equipment joints are acquired and preprocessed using a lightweight UNet segmentation network to obtain preprocessed joint images. like Figure 3 and 4 As shown, to remove background interference, standardize image size, and highlight defect features in power equipment joint images, this embodiment employs an automated image preprocessing scheme based on a lightweight UNet segmentation network. This scheme sequentially performs target segmentation, size standardization, and image format normalization. The specific implementation process is as follows: First, the trained UNetSmall semantic segmentation model is loaded, deployed to the corresponding computing device, and switched to inference mode. The model uses the 1152×1536 input size set during training to ensure image segmentation accuracy. The original RGB format image is read, and its size is verified to ensure that both the width and height are integer multiples of 8, meeting the network inference size requirements. The image is then converted to the tensor format required by the model and normalized before being input into the segmentation model for forward inference. A defect target probability map is output using the Sigmoid activation function, and the probability map is binarized with a threshold of 0.5 to obtain the initial target mask. To eliminate interference from scattered noise in the initial mask, this embodiment uses an existing breadth-first search algorithm, retaining only the largest connected region within the mask to optimize mask denoising and ensure the defect target outline is complete and the region is clean.

[0013] Based on the denoised and optimized target mask, the defective foreground image in RGBA format is extracted from the original image. Then, the minimum bounding rectangle is calculated according to the mask region, and an automatic cropping operation is performed on the foreground image. A 1024×1024 square canvas is set as the uniform output size, and a 3% edge fill area is set based on the canvas side length. The cropped image is scaled proportionally, centered, and pasted into the canvas, and the blank areas of the canvas are filled with a fixed RGB color. Finally, the RGBA format image is converted to RGB format and saved.

[0014] This embodiment supports automated batch image processing. The program iterates through all legally formatted image files in the specified input directory, executes the preprocessing procedure described above for each file, and automatically counts the number of files that were successfully processed and those that failed. After this process, standardized images with uniform specifications, no background interference, and centered defect targets can be obtained in batches, providing high-quality input data for subsequent defect detection models. The comparison of the effects before and after image preprocessing is shown in the figure below.

[0015] The processed connector image is input into an improved YOLO11 network for defect detection to obtain inspection results including defect type, pixel location, and detection confidence. This embodiment presents an improved YOLO11 network for detecting defects in power equipment joints. Addressing issues in detecting weak-edge and fine-texture defects such as scratches and burrs on power joints, including insufficient extraction of weak defect features, blurred upsampling edges, inaccurate localization of small defects, high false negative rates, and the difficulty of adapting existing networks to edge deployments in power field environments, this embodiment makes targeted lightweight improvements to the three core layers of the YOLO11 network: Backbone, Neck, and Head. The overall architecture is as follows: Figure 2 As shown.

[0016] The core improvements in this embodiment include three levels, with each module working together to achieve a dual optimization of defect detection performance and lightweight design: 1. Backbone level: Replace all shallow C3K2 modules with C3K2-SEDM sparse edge detail enhancement modules to enhance the extraction of weak edges and fine texture features of scratches and burrs, while suppressing background noise such as metallic reflections and oil stains, without increasing the amount of additional computation. 2. Neck layer: Replace all native upsampling layers with CARAFE-FD defect-aware upsampling module, optimize upsampling convolution kernels for slender defects, fuse shallow detail features to avoid edge blurring, and achieve lightweighting of the entire link; 3. Head level: Replace the native Detect head with the ST-DFL spatial awareness small defect detection head, optimize the regression distribution strategy for slender / point defects, add a small defect confidence weighted branch to reduce the false negative rate, and remove redundant large target branches to achieve lightweighting.

[0017] Through the above three-level collaborative improvements, this embodiment effectively improves the detection accuracy of weak defects in power connectors, reduces the missed detection rate, and at the same time, the overall network is lightweight and adaptable to the deployment requirements of power field edge devices.

[0018] The improved YOLO11 network includes: Replace all shallow C3K2 modules in the original YOLO11 network Backbone layer with C3K2-SEDM sparse edge detail enhancement modules; like Figure 5 As shown, this embodiment proposes a sparse edge detail enhancement module C3K2-SEDM for the backbone network of the power equipment joint defect detection network. All shallow C3K2 modules in the backbone are uniformly replaced with this module to solve the technical problems of the traditional C3K2 module's insufficient ability to extract weak edge and fine texture defect features such as scratches and burrs on power joints, and its susceptibility to background noise interference such as metal reflection and oil stains. At the same time, without increasing the amount of additional computation, it takes into account the semantic extraction capability and defect detail enhancement effect of the original module.

[0019] Combined with appendix Figure 5 The diagram shows the overall structure of the C3K2-SEDM module. The workflow and specific implementation of this module are as follows: 1. Initial Feature Processing The input features first pass through the initial convolution (Conv) module to complete basic channel adjustment and preliminary feature extraction, which is consistent with the input processing flow of the original C3K2 module, providing a unified feature foundation for subsequent branch processing.

[0020] 2. Parallel Feature Extraction with Two Branches This module adopts a dual-branch parallel structure to simultaneously achieve semantic feature extraction and defect edge detail capture: (1) Main semantic branch: The features output by the initial convolution enter the main branch composed of n C3K (3×3) modules stacked together. This branch completely retains the structure of the original C3K2 module, which can stably extract the deep semantic features of the image and maintain the semantic extraction capability of the original module. (2) Sparse edge-aware branch (EdgeConv): At the same time, the features of the initial convolution output are introduced into the newly added EdgeConv (sparse edge-aware convolution) branch. This branch is specifically designed for the linear fine texture of electrical connector scratches and the dotted edges of burrs. It can accurately capture weak edge and fine texture defect features that are difficult to extract by traditional convolution, and realize the directional extraction of defect edge information.

[0021] 3. Detailed feature weight gating (DetailGate) The semantic features output from the main branch and the edge detail features extracted by the EdgeConv branch are jointly input into the DetailGate (detail feature weight gating) module. DetailGate automatically amplifies the response weights of defect detail features such as scratches and burrs through an adaptive weight adjustment mechanism, while suppressing the interference of background noise such as metallic reflections and oil stains. This achieves the enhancement of defect details and the filtering of background noise, solving the problems of insufficient response to weak defect features and susceptibility to noise interference in traditional modules.

[0022] 4. Feature fusion and channel adjustment The features enhanced by DetailGate are concatenated with the output features of the initial convolutional module, and then fused and adjusted in dimension by a 1×1 convolution to output the final enhanced feature map. This fusion method can efficiently fuse the initial features, enhanced semantic features, and defect detail features, achieving complementarity of multi-scale features without introducing additional computational overhead.

[0023] 5. Lightweight and compatible design This module retains the original C3K2 backbone structure and adds only the lightweight EdgeConv and DetailGate branches, without increasing the number of network parameters or computational load. At the same time, the overall structure is fully compatible with the input and output interfaces of the original C3K2 module, and can directly replace the original shallow modules in the Backbone without modifying other network structures, thus achieving the lightweight goal of "enhancing defect details without increasing computational load".

[0024] Replace all native upsampling layers in the original YOLO11 network Neck layer with the CARAFE-FD defect-aware upsampling module; To address the upsampling stage of the Neck layer in a power equipment defect detection network, an improved content-aware upsampling module, CARAFE-FD (Content-Aware ReAssembly of Features for Feature Defect), is proposed. This module replaces all native upsampling layers in the Neck layer, overcoming the shortcomings of traditional upsampling methods, such as blurred defect edges, insufficient reconstruction capabilities of fine defect features like scratches / burrs, and the difficulty of adapting existing upsampling modules to the deployment of edge devices in power fields.

[0025] Combined with appendix Figure 6 The diagram shown illustrates the overall structure of the CARAFE-FD module. This embodiment of the CARAFE-FD module employs a dual-input architecture, and its workflow and specific implementation are as follows: 1. Dual-input feature acquisition The module contains two inputs: the main input is the upsampled feature map X output by the Neck network, with dimensions C×H×W; the auxiliary input is the SEDM (Shallow Detail Feature), which contains defect edges and texture information, with dimensions equal to or greater than the main feature. Figure 1 To.

[0026] 2. Early fusion and lightweight channel compression The module, Fusion & Lightweight Compressor, simultaneously inputs the backbone feature map X and the shallow detail features of SEDM into a fusion and lightweight channel compression module. This module uses lightweight 1×1 convolutions to achieve multi-source feature fusion, pre-introducing defect detail information into the subsequent kernel generation process, guiding the subsequent convolution kernel generation process to focus on defect features; at the same time, it reduces the feature channel dimension through low-dimensional channel compression operations, reducing the subsequent computational load, and achieving a lightweight design across the entire link to adapt to the deployment requirements of power field edge devices, corresponding to the lightweight improvement goal of this embodiment.

[0027] 3. Defect-Adaptive Kernel Generation The fused feature input after channel compression is a defect-adaptive kernel encoder. For the morphological features of elongated defects such as scratches and burrs on the surface of power equipment, this encoder uses orientation-aware asymmetric convolution (a combination of 3×1 longitudinal convolution and 1×3 transverse convolution) and depth-separable convolution to replace the ordinary convolution structure of the native CARAFE, generating a spatially adaptive upsampling convolution kernel adapted to the morphology of elongated defects. This significantly improves the upsampling reconstruction capability of defect features, achieving the adaptive kernel improvement for defect features in this embodiment.

[0028] 4. Upsampling kernel normalization processing The encoder outputs the original upsampled kernel input kernel normalizer module, which normalizes the convolutional kernel corresponding to each spatial location through the Softmax function, so that the sum of all weights in a single convolutional kernel is 1, avoiding feature distortion during the upsampling process and ensuring the effectiveness of the upsampled features.

[0029] 5. Feature Recombination and Dual-Path Detail Enhancement The normalized upsampling kernel, the original backbone feature map X, and the SEDM shallow detail features are jointly input into the feature reassembly & detail enhancement module. The specific operation is as follows: First, the normalized defect adaptive convolution kernel and the original backbone feature map are used to perform content-aware feature reassembly to complete the upsampling operation, retaining the content-aware upsampling capability of the original CARAFE module; at the same time, the SEDM shallow detail features are directly input into this module after spatial resolution alignment through bilinear upsampling, and fused with the reassembled upsampling features to supplement the defect edge details lost during the upsampling process, solve the edge blurring problem caused by traditional upsampling, and completely retain the contour information of scratches and burrs, thus achieving the improvement goal of fusing shallow detail features to suppress edge blurring in this embodiment.

[0030] 6. High-resolution feature output The module ultimately outputs a high-resolution upsampled feature map X′ with dimensions C×σH×σW (where σ is a preset upsampling factor). The output features are directly fed into the subsequent network detection branch to complete the power equipment defect detection task.

[0031] Replace the Detect head in the original YOLO11 network Head layer with the ST-DFL spatial awareness small defect detection head.

[0032] This embodiment proposes a spatially aware small defect detection head, ST-DFL (Spatial DFL), for the detection head link at the end of the power equipment defect detection network. This module replaces the native Detect detection head in the network to solve the technical problems of traditional detection heads, such as large sub-pixel positioning error for slender / point-like small defects like scratches and burrs on power joints, low confidence of small defects, high false negative rate, redundant calculation of large target detection branches, and incompatibility with power defect size distribution and edge deployment.

[0033] Combined with appendix Figure 7 The diagram shown illustrates the overall structure of the ST-DFL inspection head. This inspection head abandons the original general DFL regression framework and optimizes the regression strategy and branch structure specifically for the morphological characteristics of electrical connector defects, which are mainly linear scratches and dotted burrs. The overall workflow and specific implementation method are as follows: 1. Multi-scale feature input and preprocessing The detection head receives multi-scale features output from the Neck module and first feeds them into the Feature Preprocessing module. Basic convolutions unify the feature channels and dimensions, providing standardized feature input for subsequent multi-branch parallel computation. This module is compatible with the preprocessing flow of the native Detect head, ensuring network interface consistency without requiring modifications to the upstream module structure.

[0034] 2. Multi-branch parallel defect processing The preprocessed features are processed simultaneously in three parallel branches, and the functions of each branch are as follows: (1) Classification Branch: Completes the determination of defect category, distinguishes different defect types such as scratches and burrs, and retains the basic classification capability of the detection head; (2) ST-DFL Regression Branch: As the core improvement module of this embodiment, it abandons the regression distribution strategy of the original general DFL and optimizes the regression distribution modeling for the morphological characteristics of electrical connector scratches (linear slender shape) and burrs (small point targets): For linear scratch defects, spatial direction perception distribution modeling is adopted to strengthen the position weight along the defect extension direction and improve the boundary positioning accuracy of linear defects; For point burr defects, center-focused narrow distribution modeling is adopted to improve the sub-pixel-level center positioning accuracy of point defects by shrinking the distribution interval and strengthening the center position weight, effectively solving the sub-pixel positioning error problem of traditional detection head for slender / small point defects; (3) Small Defect Confidence Weight Branch: This is a new functional branch added in this embodiment. In response to the problem that small defects in the power scenario have weak feature response and are easily affected by background noise such as metal reflection and oil stains, the branch automatically amplifies the feature response weight of small defects through an adaptive weight adjustment mechanism, while suppressing the interference of background noise, thereby improving the detection confidence of small defects and effectively reducing the false negative rate.

[0035] 3. Lightweight branch structure design The overall structure of this detection head retains only three core branches adapted to the size distribution of power defects, completely eliminating redundant branches in the original Detect head adapted to large target detection, and removing computational overhead unrelated to power defect detection. While ensuring the performance of small defect detection, the network structure is simplified, the number of parameters and computational complexity are reduced, and the detection head is transformed into a lightweight version, adapting to the deployment requirements of power field edge terminals.

[0036] 4. Summary and Output of Test Results The results of the three branches (defect category, location coordinates, and weighted confidence score) are summarized and then output by the Output module to obtain the defect category, pixel location, and detection confidence score, thus completing the task of detecting and locating defects in power equipment. The ST-DFL detection head is fully compatible with the original Detect detection head in terms of input and output interfaces and can be directly replaced without modifying other modules of the network. It has low engineering implementation difficulty and strong versatility.

[0037] Compared to the native YOLO11 network, this embodiment has the following advantages: Significantly improved weak defect detection capability: The C3K2-SEDM module enhances the extraction of weak edge features of scratches and burrs, the CARAFE-FD module avoids the blurring of upsampling edges, and the ST-DFL detection head optimizes the localization and confidence of small defects. These three improvements effectively enhance the detection accuracy of weak defects in power connectors and reduce the false negative rate. Lightweight end-to-end design adapted to edge deployment: C3K2-SEDM, CARAFE-FD, and ST-DFL all adopt a lightweight design with no additional computational overhead, significantly reducing the overall number of network parameters and computational load, and adapting to the deployment requirements of power field edge devices; Strong structural compatibility: All improved modules are compatible with the native module interface and can be directly replaced without modifying the rest of the network structure, resulting in low engineering implementation cost and strong versatility.

[0038] As an optional embodiment of the present invention, the preprocessing of the connector image using a lightweight UNet segmentation network may be performed, including: Load the trained UNetSmall semantic segmentation model and switch the model to inference mode; It should be noted that the training method of the UNetSmall semantic segmentation model in this embodiment is as follows: collect detection images of 220kV power cable intermediate joints under different processes during the joint manufacturing process, construct a pixel-level labeled dataset, and label the crimping pipe area and background area (including non-detection target areas such as cable insulation layer, metal sheath, and environmental clutter) in each image separately; randomly divide the dataset into training set and validation set in an 8:2 ratio, add random horizontal flipping, random rotation, and random brightness perturbation operations to the training set images to complete data augmentation, expand the diversity of training samples, and improve the model's segmentation generalization ability.

[0039] The normalized training images are input into the UNetSmall semantic segmentation model, which outputs corresponding pixel-level segmentation prediction results. The segmentation loss is calculated using a weighted cross-entropy loss function, where the weight of the pipe region loss is set to 5 times that of the background region to address the class imbalance problem in the dataset where the background accounts for a large proportion and the foreground accounts for a small proportion. The AdamW optimizer is used to update the model parameters, with an initial learning rate of 1×10⁻⁴ and a weight decay coefficient of 1×10⁻⁵. The learning rate is dynamically adjusted using a cosine annealing strategy. Training stops when the intersection-over-union (IoU) of the validation set segmentation no longer improves after 10 consecutive rounds, and the optimal model parameters are saved, thus obtaining the trained UNetSmall semantic segmentation model.

[0040] In the inference stage, the original image to be detected is input into the trained UNetSmall model, which outputs a binary segmentation mask of the crimping pipe region. The segmentation mask is multiplied pixel by pixel with the original input image to filter out background interference regions outside the crimping pipe, and only retains the effective image output of the crimping pipe region. This image is then fed into the subsequent defect detection network to complete the detection. This can further reduce the interference of irrelevant background on defect feature extraction and further improve the detection accuracy of small defects.

[0041] The image size of the power equipment connector image is verified and normalized to obtain the processed image data; It should be noted that this embodiment requires the input image to the defect detection network to be a square of uniform size. After the image to be detected is segmented and preprocessed, if the original image size does not meet the network input requirements, the image is first adjusted to the target input size by proportional borderless filling to avoid distortion of the defect shape caused by direct stretching and scaling, thus ensuring the integrity of the original geometric features of the defect. Subsequently, the image pixel value range is linearly normalized from the original 0~255 to the 0~1 range, and then further standardized according to the pixel mean and standard deviation obtained from the overall statistics of the training set. This can effectively eliminate the differences in pixel distribution caused by different acquisition devices and different on-site exposure conditions, improve the stability of the network detection results, and adapt to the defect detection needs under different on-site acquisition conditions. All of the above preprocessing processes can be completed on the edge device without additionally uploading the original image to the cloud, meeting the requirements of power field data security and real-time detection.

[0042] The processed image data is input into the UNetSmall semantic segmentation model for forward inference, and a defect target probability map is output through the Sigmoid activation function. It should be noted that in this embodiment, the region with a pixel value greater than a preset threshold of 0.5 in the probability map is determined as the foreground target region of the compression pipe, and the region with a pixel value less than or equal to 0.5 is determined as the background region. Based on this, a binary segmentation mask is generated: the pixel value of the foreground region in the mask is set to 1, and the pixel value of the background region is set to 0, to ensure that the segmentation mask can accurately distinguish the compression pipe region to be detected from various irrelevant interference backgrounds.

[0043] The resulting binary segmentation mask is then multiplied pixel-by-pixel with the original input image to filter out all background interference outside the crimping nozzle area, retaining only the valid image output of the crimping nozzle region. This image is then fed into the subsequent defect detection network for detection. This preprocessing step allows the subsequent defect detection network to concentrate all its computational resources on feature extraction of the crimping nozzle region, avoiding interference from background metallic reflections and noise on the features of minor defects, thus further improving the detection accuracy of small defects.

[0044] Meanwhile, the UNetSmall semantic segmentation model used in this embodiment, through its lightweight full-network structure design, ensures that the segmentation inference process for a single 1024×1024 resolution image takes no more than 100ms on the edge terminal, without increasing the overall detection process by too much latency. This fully meets the speed requirements for real-time detection at power sites and also complies with the computing power constraints of edge devices.

[0045] The defect target probability map is binarized using a judgment threshold to obtain an initial target mask; It should be noted that the preset threshold of 0.5 can be flexibly adjusted according to the actual on-site detection needs: when the detection scenario has a high requirement for the false negative rate, the threshold can be appropriately lowered to expand the range of the foreground region to avoid the false filtering of defects at the edge of the crimping pipe; when the background interference of the scene is strong and the detection accuracy requirement is higher, the threshold can be appropriately increased to more strictly filter out background interference and further improve detection stability. The initial target mask obtained by binarization is processed by morphological closing operation to fill the small holes inside the mask, smooth the mask edges, and eliminate the fragmented noise generated during the segmentation process, so as to obtain the final continuous and complete binary segmentation mask of the crimping pipe region, ensuring the quality of mask segmentation and avoiding the introduction of additional noise in the subsequent feature extraction process.

[0046] The initial target mask is optimized by mask denoising using an existing breadth-first search algorithm to obtain the target mask. It should be noted that breadth-first search traverses all connected foreground regions in the mask, calculates the pixel area of ​​each connected region, retains only the largest connected region of the main pipe, and directly removes the remaining small isolated noisy connected regions, further eliminating the noise caused by missegmentation, and ensuring that the final mask contains only a complete single pipe region and does not introduce irrelevant interference regions.

[0047] Based on the target mask, the defect foreground image is extracted from the power equipment joint image, and the minimum bounding rectangle is solved based on the target mask. The defect foreground image is then automatically cropped based on the minimum bounding rectangle to obtain the preprocessed joint image.

[0048] It should be noted that after extracting the crimped pipe area, it is directly cropped to the smallest image size containing only the target area. Subsequent defect detection is performed only on this area, which further reduces redundant calculations, lowers the computing power consumption of edge devices, and avoids interference from the background area outside the crimped pipe, thereby further improving the detection accuracy of small defects. After obtaining the preprocessed joint image, it can be input into the improved defect detection network of this invention to complete the detection and output of crimped pipe defects according to the aforementioned module process.

[0049] As an optional embodiment of the present invention, optionally, replacing all shallow C3K2 modules in the original YOLO11 network Backbone layer with C3K2-SEDM sparse edge detail enhancement modules includes: The preprocessed connector image is used to perform preliminary feature extraction using the initial convolution module to obtain initial features; It should be noted that after the original input image is cropped to the target size image containing only the piping, it is first fed into the initial convolutional layer of the Backbone network to complete preliminary downsampling and basic feature extraction, resulting in a shallow initial feature map. This process is completely consistent with the native YOLO11 and requires no additional adjustment to the convolutional parameters or output dimensions. Subsequently, the original shallow C3k2 basic module in the Backbone network is replaced with the C3K2-SEDM sparse edge detail enhancement module. This module is based on a lightweight structure built with depthwise separable convolution. Without increasing the number of parameters significantly, it introduces a sparse edge attention mechanism to automatically enhance the response of weak edge features such as scratches and burrs on the piping, suppress redundant feature calculations in flat areas, and achieve targeted enhancement of defect edge features in the shallow feature extraction stage.

[0050] The initial features are input into a trunk branch composed of n stacked C3K modules to extract deep semantic features of the image; It should be noted that the main branch is responsible for extracting general deep semantic features of the image. Its structure is consistent with the main branch of the original YOLO11. Only the original shallow modules are replaced, and there is no need to adjust the deep network structure, so as to ensure the feature extraction capability while reducing the transformation cost.

[0051] The features output from the main branch are fed into a parallel sparse edge attention sub-branch. After compressing the channel dimension through 1×1 convolution, deformable convolution is used to extract the sparse response of the defect edge. Then, an edge attention weight map is generated by Sigmoid activation. The weight map is multiplied with the main branch features channel by channel to complete the targeted enhancement of the defect edge features. Finally, the enhanced shallow features are output and fed into the subsequent feature pyramid Neck module to ensure that weak defect edge features can be completely transmitted to the subsequent detection stage and avoid the loss of defect edge information during the shallow feature extraction process.

[0052] The initial features are input into a sparse edge-aware convolutional branch to extract defect detail features; It should be noted that the sparse edge-aware convolutional branch first performs channel compression on the input features using 1×1 convolutions, reducing the computational cost of the branch while preserving key defect edge information. Then, two layers of 3×3 deformable convolutions are stacked, adaptively adjusting the convolution sampling point positions to better fit the irregular shapes of scratches and burrs, accurately capturing the sparse response features of defect edges. Finally, an edge attention weight map in the 0-1 range is generated using the Sigmoid activation function, with high weights for defect edge areas and low weights for flat background areas. Multiplying the generated edge attention weight map channel-by-channel with the features output from the main branch completes the targeted enhancement of defect edge features in the main features, suppressing feature responses in irrelevant flat areas. The final enhanced shallow defect features are fed into the subsequent Neck feature pyramid module, ensuring that weak defect edge features are fully transmitted to subsequent detection stages, avoiding premature loss of defect edge information during shallow feature extraction, and improving the basic detection capability of weak defects from the source of feature extraction.

[0053] The deep semantic features and defect features are input into the detail feature weight gating module. The detail feature weight gating module uses an adaptive weight adjustment mechanism to automatically amplify the response weight of the defect detail features to obtain the enhanced features. It should be noted that the detail feature weight gating module generates channel weights for the deep semantic features and defect edge features input from different branches. The weights are dynamically allocated according to the contribution of the feature channels to defect recognition: the channel corresponding to defect edge features has a higher overall weight, while the channel corresponding to flat background semantic features has a lower weight. This further enhances the response of detail defect features without destroying the original feature information, suppresses redundant feature interference, and finally outputs the enhanced fused features to be passed to the subsequent feature pyramid network. This ensures that weak defect features always maintain a sufficient response during the forward propagation of the entire network and are not covered or submerged by background features.

[0054] The enhanced features and the initial features are sequentially concatenated and convolved to obtain the final enhanced feature map.

[0055] It should be noted that the splicing operation can fuse the original features and the weighted defective features along the channel dimension, and then unify the channel dimension of the fused features through a 1×1 convolution. This perfectly matches the input channel requirements of the subsequent Neck module, without the need for additional adjustments to the subsequent network structure. It is compatible with the processing flow of the native network, ensuring the plug-and-play replacement feature of the module, without having any additional impact on the overall network structure, and further reducing the difficulty of engineering modification and deployment.

[0056] As an optional embodiment of the present invention, optionally, replacing all native upsampling layers in the original YOLO11 network Neck layer with the CARAFE-FD defect-aware upsampling module includes: The final enhanced feature map and SEDM shallow detail features are input into the fusion and lightweight channel compression module for channel compression to obtain the fused features after channel compression. The fused features are input into the defect adaptive kernel encoder for kernel parameter learning to obtain a spatial adaptive upsampling convolution kernel that adapts to the shape of slender defects. Based on the upsampling convolution kernel, the Softmax function is used to normalize the convolution kernel corresponding to each spatial location to obtain the normalized upsampling kernel; The normalized upsampling kernel, the final enhanced feature map, and the SEDM shallow detail features are input into the feature recombination and detail enhancement module to obtain the upsampling feature map.

[0057] It should be noted that the defect-aware upsampling module first performs channel-dimensional concatenation and fusion of multi-scale features from adjacent layers. Then, it performs channel compression on the fused features using 1×1 convolutions. This reduces the computational load of subsequent processing while preserving multi-scale defect feature information. The entire channel compression process uses depthwise separable convolutions instead of ordinary convolutions, further reducing the number of parameters and computational overhead, adapting to the computing power constraints of edge devices. After compression, an adaptive upsampling kernel is generated based on the compressed features. The parameters of the upsampling kernel are directly predicted from the input features and can adaptively adjust the weights of the upsampling region according to the defect location and shape. Compared to traditional upsampling kernels with fixed structures, this better preserves the detailed features of defect edges and avoids the smooth loss of defect details during upsampling. Next, the generated upsampling kernel is normalized to ensure that the sum of the upsampling kernel weights at each output position is 1. This avoids overall shift in output pixel values ​​caused by normalization, maintains stable feature distribution, and prevents abnormal feature activation from affecting subsequent detection accuracy. Finally, the generated adaptive upsampling kernel is used to resample the input features, completing the feature map size upsampling and outputting high-resolution defect features. This better preserves the detailed information of small defects and improves the accuracy of subsequent detection heads in locating and classifying tiny defects.

[0058] After completing the multi-scale feature fusion of the feature pyramid, the final multi-scale high-resolution feature map is fed into the YOLO11 native detection head. The detection head outputs prediction results containing defect category, defect location confidence, and bounding box parameters. Redundant and duplicate prediction boxes are then eliminated through non-maximum suppression. Finally, the category, location, and confidence information of all detected crimping pipe defects in the image are output, completing the entire defect detection process for 220kV power cable intermediate joint crimping pipes. This method significantly reduces the number of network parameters and computational load without sacrificing detection accuracy through lightweight structure improvements and detailed feature-oriented enhancement. It can be directly deployed on power field edge detection equipment, balancing detection accuracy, speed, and deployment compatibility. It effectively solves the problems of high false negative rates for minor defects in crimping pipes and large model sizes that are difficult to deploy in the field in existing detection methods.

[0059] As an optional embodiment of the present invention, the fusion and lightweight channel compression module may optionally employ lightweight 1×1 pointwise convolution.

[0060] It should be noted that 1×1 pointwise convolution can efficiently adjust and compress the channel dimension without changing the feature map spatial size. Compared with multi-layer large convolution kernel compression schemes, it has lower parameter and computational cost, can better adapt to the computing power limitations of edge devices, and can retain the complete spatial information of multi-scale defect features without loss of defect features due to excessive compression.

[0061] As an optional embodiment of the present invention, the defect adaptive kernel encoder may use orientation-aware asymmetric convolution and depth-separable convolution to replace the ordinary convolutional structure CARAFE in the original YOLO11 network Neck layer.

[0062] It should be noted that orientation-aware asymmetric convolution can extract convolution features along different directions to better fit the irregular elongated shape of the defect, which is often characterized by thin scratches and longitudinal burrs in pipe fittings. Compared with ordinary square symmetric convolution, it can more accurately capture the directional information and contour details of the defect edge and generate upsampling kernel parameters that are more suitable for the defect shape. At the same time, it is combined with depthwise separable convolution to split ordinary convolution into depthwise convolution and pointwise convolution, which further reduces the number of convolution parameters and computation while ensuring feature extraction capabilities, meeting the requirements of lightweight design and adapting to the deployment needs of edge devices.

[0063] As an optional embodiment of the present invention, the feature recombination and detail enhancement module may use the normalized upsampling kernel and the final enhanced feature map to perform content-aware feature recombination, complete the upsampling operation, and retain the content-aware upsampling capability of the CARAFE module in the original YOLO11 network Neck layer. The shallow detail features of SEDM are spatially aligned by bilinear upsampling, and then input into the feature recombination and detail enhancement module. They are then fused with the recombined upsampled features to supplement the defect edge details lost during the upsampling process.

[0064] It is important to explain in detail that in the original CARAFE upsampling process, the deep features, after multiple downsampling operations, have already lost a significant amount of shallow detail information. Relying solely on the upsampling kernel generated by feature adaptation is insufficient to fully restore the clear edge contours of small and elongated defects, easily leading to blurred defect edges in the upsampled output. This, in turn, causes subsequent detection heads to mislocate and misclassify weak defects. By introducing the native shallow defect detail features extracted by the SEDM module, complete original response information of defect edges can be directly supplemented during the upsampling feature reconstruction stage. This effectively avoids the problem of defect edges being smoothed out after upsampling, further enhancing the richness of defect details in the upsampled output features. This provides the subsequent detection head with clearer defect location and contour information, reducing the probability of missing small defects from the feature level. Simultaneously, this improvement only increases the computational load of bilinear alignment and feature fusion by a small amount, without significantly increasing the overall computational cost or parameter count of the network. It achieves targeted enhancement of defect details while adhering to the overall lightweight design goal of this method, better adapting to the computational constraints of power field edge detection terminals without adding additional detection latency.

[0065] As an optional embodiment of the present invention, optionally, replacing the Detect head in the original YOLO11 network Head layer with the ST-DFL spatially aware small defect detection head includes: The upsampled feature map is preprocessed to obtain the preprocessed defect features. It should be noted that feature preprocessing includes performing channel-dimensionality reduction convolutions on the upsampled feature maps of multi-scale inputs, adjusting the output channels of each feature map to a detection output dimension that matches the preset number of anchor boxes. Depth-separable convolutions are used in the process to achieve dimension adjustment, further compressing the number of parameters and computation of the detection head.

[0066] Based on the preprocessed defect features, the defect category is determined using classification branches to obtain the defect category determination result; It should be noted that the determination of the defect category is accomplished through an independent classification convolution branch. To address the class imbalance problem between defective and normal samples, a weighted cross-entropy loss function can be used to train the classification branch, thereby increasing the classification loss weights for small defects and difficult samples, guiding the network to focus more on learning defect features, and reducing classification errors for small defects.

[0067] Based on the preprocessed defect features, the ST-DFL regression branch is used to identify the defect location and obtain the defect location identification result. It should be noted that the defect location recognition extracts the orientation-aware features of the bounding box position through spatially perceptual convolution. Targeting the narrow boundary characteristics of small defects and slender defects in the compression pipe, it adaptively enhances the feature response of the boundary region, improving the perception accuracy of the regression branch for the boundary position. Compared with the native DFL regression branch, it can more accurately fit the bounding box parameters of irregular defects and reduce the positioning error of small defects.

[0068] Based on the preprocessed defect features, the adaptive weight adjustment mechanism of the small defect confidence weighted branch is used to automatically amplify the feature response weight of small defects, while suppressing the interference of background noise, and obtain the weighted and corrected defect detection confidence. Regarding the adaptive weight adjustment mechanism of the small defect confidence weighted branch, it should be noted that this branch generates weights using only two 1×1 convolution layers, without requiring additional complex computational structures. The increase in parameter quantity is negligible, meeting the overall lightweight design requirements. This branch generates confidence correction weights based on the feature statistics corresponding to each candidate detection box: for low-response candidate boxes corresponding to small defects, higher confidence gain weights are automatically assigned to increase their retention probability in the non-maximum suppression stage; for high-response false alarm candidate boxes corresponding to background noise, confidence suppression weights are automatically assigned to reduce their false detection probability. Existing native detection heads often determine the confidence of small defects as below the detection threshold due to their weak overall feature response, ultimately leading to their elimination in the non-maximum suppression stage and resulting in missed detections. The adaptive correction mechanism of this branch can effectively improve this problem, further reducing the missed detection rate of small defects in the crimped pipe from the detection output. Finally, the ST-DFL spatial awareness small defect detection head summarizes the defect category determination results, location recognition results, and corrected detection confidence, outputting all crimped pipe defect detection results that meet the threshold requirements, completing the entire detection process.

[0069] Based on the defect category determination result, defect location identification result, and weighted and corrected defect detection confidence, the Output module outputs the inspection result including the defect category, pixel location, and detection confidence.

[0070] Regarding the Output module, it's important to note that this module primarily handles the format conversion and standardization of the raw inspection results output by the inspection head. First, it maps the detected defect pixel coordinates back to the coordinate system of the original input inspection image, converting them into location annotations that match the on-site image acquisition system. This facilitates quick location of defects on the actual press-fit pipe by grinding personnel. Suspected defects with confidence levels between the warning threshold and the pass threshold are automatically marked individually and a specific warning message is generated, reminding on-site quality inspectors to manually verify the location. This prevents potentially defective joints from entering subsequent construction stages, further supplementing the inspection process to prevent the omission of weak and small defects. For confirmed defects with confidence levels above the pass threshold, the defect category and location are clearly labeled directly. Areas with confidence levels below the warning threshold are directly judged as passable without additional labeling. Finally, all inspection results are output.

[0071] As an optional embodiment of the present invention, when the preprocessed defect features are linear and elongated, the ST-DFL regression branch adopts spatial orientation-aware distribution modeling and increases the position weight along the defect extension direction. When the preprocessed defect features are point-like small target defects, the ST-DFL regression branch adopts a center-focused narrow distribution model, which increases the weight of the center position by shrinking the distribution interval.

[0072] Regarding the ST-DFL regression branch, it's important to note that different types of crimped pipe defects correspond to different distribution modeling methods. This approach can specifically strengthen the probability weights of the true boundary locations of defects and weaken the interference from irrelevant locations. Compared to the fixed uniform distribution modeling used in the original DFL, it better adapts to the diverse morphological characteristics of crimped pipe defects, effectively improving the regression accuracy of irregular defect bounding boxes. For linear, slender defects, increasing the positional weights along the extension direction can more accurately capture the overall length range of the defect, avoiding the segmentation of continuous slender defects into multiple scattered detection boxes or missing longitudinal scratch defects with extremely small short diameters. For point-like micro-defects, shrinking the distribution interval to focus on the central location can reduce the interference of background features on boundary regression, making the localization results more concentrated on the actual location of the defect and avoiding defect omissions caused by regression box offsets. This modeling scheme only adaptively adjusts the distribution calculation logic of the original DFL branch, without adding a large number of convolution parameters, thus avoiding excessive computational overhead. It improves the localization accuracy of small and irregularly shaped defects while meeting the overall lightweight design requirements and adapting to the computing power constraints of edge deployment.

[0073] Evaluation indicators and experimental results This embodiment uses five commonly used evaluation metrics in the field of target detection to compare and analyze the performance of the native YOLO11 model and the improved YOLO11 model in detecting defects in power equipment joints. The definitions of each metric are as follows: 1. Precision: Characterizes the model's ability to resist false detections. It refers to the proportion of samples predicted as defects that are actually defects. The higher the value, the lower the probability that the model will misclassify background noise as defects.

[0074] 2. Recall: Characterizes the model's ability to detect false negatives. It refers to the proportion of real defective samples that are successfully detected. The higher the value, the lower the false negative rate of the model.

[0075] 3. mAP50: The average accuracy when the Intersection over Union (IoU) threshold is 0.5. It is a core basic indicator for industrial defect detection and reflects the model's basic ability to identify defects.

[0076] 4.mAP75: The average accuracy when the Intersection over Union (IoU) threshold is 0.75. It has stringent requirements for the accuracy of defect boundary positioning and reflects the sub-pixel level positioning capability of the model.

[0077] 5. mAP50-95: The average accuracy of IoU in the range of 0.5 to 0.95, which comprehensively evaluates the overall detection performance of the model, taking into account both basic recognition and fine localization capabilities.

[0078] The experimental results are as follows: This embodiment uses point-like burrs and linear scratches on power equipment joints as the detection targets. It compares and analyzes the defect detection performance of the native YOLO11 model and the improved YOLO11 model. The results show that the improved model achieves improvements in all evaluation indicators, and its overall performance is superior to the native model. The overall precision of the model increases from 0.9505 to 0.9699, the recall from 0.9742 to 0.9784, and the mAP50 from 0.9806 to 0.9868, enhancing basic recognition stability and simultaneously optimizing resistance to false positives and false negatives. Simultaneously, mAP75 increases from 0.6315 to 0.7180, and mAP50-95 increases from 0.6330 to 0.6987, significantly improving sub-pixel defect localization accuracy and multi-scale detection robustness.

[0079] For point-like burr defects that are easily affected by background interference, the improved model accuracy increased from 0.9026 to 0.9413, the false negative rate continued to decrease, and mAP75 and mAP50-95 improved by nearly 48% and 21.7% respectively, significantly enhancing the ability to detect and locate weak defects. For linear scratch defects, the model maintained excellent performance with zero false negatives, while mAP75 and mAP50-95 were slightly improved, the accuracy of defect edge positioning was further optimized, and the edge blurring problem caused by upsampling was effectively solved.

[0080] The aforementioned performance improvements correspond closely to the three core improvements in this embodiment: the C3K2-SEDM module suppresses background noise and enhances defect details, the CARAFE-FD module optimizes the preservation of slender defect edges, and the ST-DFL detection head improves the confidence and positioning accuracy of small defects. Moreover, the end-to-end lightweight transformation did not negatively impact performance; on the contrary, it achieved a dual improvement in detection accuracy and edge deployment adaptability, verifying the effectiveness and engineering practicality of the improvement scheme.

[0081] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for detecting defects in the crimped pipe of a 220kV power cable intermediate joint, characterized in that, The method includes: Images of power equipment joints are acquired and preprocessed using a lightweight UNet segmentation network to obtain preprocessed joint images. The processed connector image is input into an improved YOLO11 network for defect detection to obtain inspection results including defect type, pixel location, and detection confidence. The improved YOLO11 network includes: Replace all shallow C3K2 modules in the original YOLO11 network Backbone layer with C3K2-SEDM sparse edge detail enhancement modules; Replace all native upsampling layers in the original YOLO11 network Neck layer with the CARAFE-FD defect-aware upsampling module; Replace the Detect head in the original YOLO11 network Head layer with the ST-DFL spatial awareness small defect detection head.

2. The method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 1, characterized in that, Preprocessing using a lightweight UNet segmentation network yields the following preprocessed connector images: Load the trained UNetSmall semantic segmentation model and switch the model to inference mode; The image size of the power equipment connector image is verified and normalized to obtain the processed image data; The processed image data is input into the UNetSmall semantic segmentation model for forward inference, and a defect target probability map is output through the Sigmoid activation function. The defect target probability map is binarized using a judgment threshold to obtain an initial target mask; The initial target mask is optimized for mask denoising using a breadth-first search algorithm to obtain the target mask. Based on the target mask, the defect foreground image is extracted from the power equipment joint image, and the minimum bounding rectangle is solved based on the target mask. The defect foreground image is then automatically cropped based on the minimum bounding rectangle to obtain the preprocessed joint image.

3. The method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 1, characterized in that, The original YOLO11 network Backbone layer replaces all shallow C3K2 modules with the C3K2-SEDM sparse edge detail enhancement module, including: The preprocessed connector image is used to perform preliminary feature extraction using the initial convolution module to obtain initial features; The initial features are input into a trunk branch composed of n stacked C3K modules to extract deep semantic features of the image; The initial features are input into a sparse edge-aware convolutional branch to extract defect detail features; The deep semantic features and defect features are input into the detail feature weight gating module. The detail feature weight gating module uses an adaptive weight adjustment mechanism to automatically amplify the response weight of the defect detail features to obtain the enhanced features. The enhanced features and the initial features are sequentially concatenated and convolved to obtain the final enhanced feature map.

4. The method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 1, characterized in that, The original upsampling layers in the YOLO11 network's Neck layer are replaced with the CARAFE-FD defect-aware upsampling module, including: The final enhanced feature map and SEDM shallow detail features are input into the fusion and lightweight channel compression module for channel compression to obtain the fused features after channel compression. The fused features are input into the defect adaptive kernel encoder for kernel parameter learning to obtain a spatial adaptive upsampling convolution kernel that adapts to the shape of slender defects. Based on the upsampling convolution kernel, the Softmax function is used to normalize the convolution kernel corresponding to each spatial location to obtain the normalized upsampling kernel; The normalized upsampling kernel, the final enhanced feature map, and the SEDM shallow detail features are input into the feature recombination and detail enhancement module to obtain the upsampling feature map.

5. A method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 4, characterized in that, The fusion and lightweight channel compression module uses lightweight 1×1 pointwise convolution.

6. A method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 4, characterized in that, The defect adaptive kernel encoder uses orientation-aware asymmetric convolution and depth-separable convolution to replace the ordinary convolution structure CARAFE in the original YOLO11 network Neck layer.

7. A method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 4, characterized in that, The feature recombination and detail enhancement module uses the normalized upsampling kernel and the final enhanced feature map to perform content-aware feature recombination, complete the upsampling operation, and retain the content-aware upsampling capability of the CARAFE module in the original YOLO11 network Neck layer. The shallow detail features of SEDM are spatially aligned by bilinear upsampling, and then input into the feature recombination and detail enhancement module. They are then fused with the recombined upsampled features to supplement the defect edge details lost during the upsampling process.

8. A method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 1, characterized in that, The replacement of the original YOLO11 network Head layer Detect head with the ST-DFL spatial awareness small defect detection head includes: The upsampled feature map is preprocessed to obtain the preprocessed defect features. Based on the preprocessed defect features, the defect category is determined using classification branches to obtain the defect category determination result; Based on the preprocessed defect features, the ST-DFL regression branch is used to identify the defect location and obtain the defect location identification result. Based on the preprocessed defect features, the adaptive weight adjustment mechanism of the small defect confidence weighted branch is used to automatically amplify the feature response weight of small defects, while suppressing the interference of background noise, and obtain the weighted and corrected defect detection confidence. Based on the defect category determination result, defect location identification result, and weighted and corrected defect detection confidence, the Output module outputs the inspection result including the defect category, pixel location, and detection confidence.

9. A method for detecting defects in crimped pipes of intermediate joints in 220kV power cables as described in claim 8, characterized in that, When the preprocessed defect features are linear and elongated, the ST-DFL regression branch adopts spatial orientation-aware distribution modeling and increases the position weight along the defect extension direction. When the preprocessed defect features are point-like small target defects, the ST-DFL regression branch adopts a center-focused narrow distribution model, which increases the weight of the center position by shrinking the distribution interval.

Images