Weak boundary image-oriented ultrahigh voltage cable insulation structure measurement method

Abstract

The application discloses a kind of weak boundary image-oriented ultra-high voltage cable insulation structure measurement methods, comprising: collecting cable cross-sectional image, and pre-processing, obtain original input image;The original input image is input into the first neural network model trained, and segmentation mask is obtained;The mask obtained is used to extract the region of interest image block containing the inside and outside boundary information of insulating layer in original input image, the second neural network model constructed is used to carry out image repair and reconstruction to the region of interest image block, and high-quality reconstruction image is obtained;Based on the high-quality reconstruction image obtained, the measurement of cable insulation structure is carried out;The second neural network model uses conditional diffusion model.The present application solves the measurement problem caused by the fuzzy boundary of cable image and physical interference in complex industrial environment by the two-stage strategy of "rough segmentation positioning-fine repair reconstruction", and realizes high-precision, high-robustness non-contact automatic online detection.

Description

Technical Field

[0001] This invention relates to the field of cable insulation structure detection technology, and more specifically, to a method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images. Background Technology

[0002] Ultra-high voltage cables play a core role in power transmission systems. The thickness and uniformity of their insulation layer directly affect their electrical performance, mechanical strength, and service life, and are important indicators for ensuring the safe and stable operation of the power grid. Uneven insulation structure or insulation below standard requirements can easily cause abnormal internal electric field distribution, leading to serious faults such as partial discharge and thermal breakdown.

[0003] Existing measurement methods are mainly divided into two categories:

[0004] Contact measurement, such as section microscopy, is highly accurate but also highly destructive, making it unsuitable for online full inspection and inefficient.

[0005] Non-contact measurement, such as ultrasonic or X-ray detection, is non-destructive, but the equipment is expensive and maintenance is complex. Furthermore, for multi-layered structures with similar acoustic or optical properties, the imaging of interlayer boundaries is easily blurred, limiting accuracy.

[0006] In recent years, machine vision technology has been introduced into the field of cable structural dimension measurement. However, traditional image processing algorithms, such as threshold segmentation, edge detection, and Hough transform, heavily rely on high-quality imaging conditions. In real industrial production environments, due to factors such as uneven lighting, the reflective properties of cable materials, and burrs and debris generated during cutting, the acquired cable cross-section images often suffer from low interlayer contrast, blurred boundaries, or discontinuities. Especially at the junction of stranded conductors and insulation layers, traditional algorithms struggle to accurately locate the boundary, leading to a significant decrease in measurement accuracy and robustness, making it impossible to achieve stable and reliable fully automated inspection.

[0007] While deep learning technology has been applied in fields such as cable surface defect detection, simple segmentation networks often struggle to achieve sub-pixel-level measurement accuracy when dealing with weak boundary issues in industrial images. Currently, there is a lack of a comprehensive, high-precision automated measurement method capable of actively enhancing and reconstructing low-quality industrial images, and establishing a complete system from image acquisition to automatic output of geometric parameters. Therefore, there is an urgent need to develop a high-precision, high-efficiency, non-contact automated measurement technology for ultra-high voltage cable insulation structures that can adapt to complex industrial environments. Summary of the Invention

[0008] The main objective of this invention is to provide an automatic measurement method for the insulation structure of ultra-high voltage cables, aiming to solve the problems of low measurement accuracy, poor robustness, insufficient automation, and weak model generalization ability in the existing technology when using traditional image processing algorithms or single deep learning models to process complex industrial images with uneven lighting, noise interference, blurred boundaries, and low contrast.

[0009] To achieve the above objectives, this invention provides a method for measuring the insulation structure of ultra-high voltage cables based on images with weak boundaries. The method includes: acquiring a cross-sectional image of the cable; preprocessing the image to obtain an original input image; inputting the preprocessed original input image into a trained first neural network model to obtain a segmentation mask; using the obtained mask to extract regions of interest (ROIs) image blocks containing information about the inner and outer boundaries of the insulation layer from the original input image; using a constructed second neural network model to perform image restoration and reconstruction on the ROI image blocks to obtain a high-quality reconstructed image; and measuring the cable insulation structure based on the obtained high-quality reconstructed image. The second neural network model employs a conditional diffusion model.

[0010] More specifically, the method includes the following steps:

[0011] (1) Image acquisition and standardization: A digital image of the cross-section of the ultra-high voltage cable is acquired through an image acquisition device, and the digital image is preprocessed to generate an input image with standardized size and shape. The preprocessing may include region of interest extraction, size normalization, and image correction for correcting perspective or radial distortion.

[0012] (2) Coarse localization of the target region: The input image is input into the first neural network model; the first neural network model is a trained region recognition network used to perform high-level semantic analysis on the input image to identify and output preliminary segmentation results that characterize the approximate location and shape of the insulating layer, thereby obtaining a coarse segmentation mask.

[0013] (3) Boundary Condition Repair and Reconstruction: Using the coarse segmentation mask generated in the previous step, the region of interest containing the insulating layer and its adjacent boundaries is extracted from the original image. This region image is used as a conditional guide and input into the second neural network model. The second neural network model is a trained conditional diffusion model. This model generates a high-fidelity, high-resolution image step by step from a random noise distribution through a multi-step iterative denoising process. By learning the conditional probability distribution from the low-quality image domain to the high-quality image domain, the model can not only sharpen the edges and reconstruct the texture of the input low-contrast, blurred image, but also intelligently repair the tiny physical scratches or debris generated during the cutting process. Finally, it outputs a high-quality, clear image with clear boundaries, realistic details, and complete structure. After size restoration, the final high-quality reconstructed image is obtained.

[0014] (4) Accurate calculation of geometric parameters: The high-quality reconstructed image is post-processed to accurately extract the inner and outer boundary contours of the insulating layer; then, based on the contour data, the geometric center of the insulating layer is determined by a geometric fitting algorithm, and at least one key geometric parameter of the insulating layer is calculated and output using the multi-angle ray method or an equivalent geometric measurement algorithm. The geometric parameters include, but are not limited to: average thickness, maximum thickness, minimum thickness and concentricity.

[0015] Furthermore, the first neural network model is a semantic segmentation model.

[0016] Furthermore, the first deep learning model (i.e., the first neural network model) adopts an encoder-decoder architecture, such as U-Net or DeepLab series networks. Even further, the first neural network model is selected from U-Net or DeepLab semantic segmentation models. To improve the ability to identify small or irregular insulating regions, the model can integrate an attention mechanism module.

[0017] Furthermore, the U-Net or DeepLab semantic segmentation model integrates an attention mechanism module into the encoder. This attention mechanism module is a spatial attention mechanism module. This invention improves upon existing U-Net or DeepLab semantic segmentation models by optimizing the encoder and adding a spatial attention mechanism module.

[0018] Furthermore, taking the U-Net architecture semantic segmentation model with integrated attention mechanism module as an example, the encoder of the improved U-Net architecture semantic segmentation model includes:

[0019] A multi-branch processing unit contains at least two parallel processing paths. Each processing path receives the original input image and obtains the output feature map corresponding to the parallel processing path.

[0020] The first fusion unit fuses the output feature maps from all parallel processing paths to obtain a comprehensive feature map.

[0021] The spatial attention module receives the comprehensive feature map from the first fusion unit and performs spatial attention weighting to obtain the weighted feature map.

[0022] The second fusion unit fuses the weighted feature map with the original input image through residual connection to obtain the feature map after residual connection.

[0023] Output processing unit: Performs channel integration and dimensionality reduction on the feature map after residual connection, and outputs the feature map.

[0024] The multi-branch processing unit includes at least one processing path containing a standard convolutional layer, while other processing paths contain dilated convolutional layers with different dilation rates.

[0025] Furthermore, the spatial attention module includes:

[0026] Three parallel convolutional layers are used to generate intermediate feature maps that functionally correspond to the query, key, and value from the input feature map, respectively.

[0027] The first dimension reshaping and transpose unit performs dimension reshaping and transpose operations on the intermediate feature map corresponding to the query.

[0028] The first matrix multiplication unit performs matrix multiplication on the feature map corresponding to the query after dimension reshaping and transpose operations and the intermediate feature map corresponding to the key to obtain the correlation score matrix.

[0029] The normalization unit normalizes the correlation score matrix output by the matrix multiplication unit to obtain the attention weight map.

[0030] The second dimension reshaping and transpose unit performs dimension reshaping and transpose operations on the intermediate feature map corresponding to the value;

[0031] The second matrix multiplication unit performs matrix multiplication on the feature map corresponding to the value obtained after dimension reshaping and transpose operations and the attention weight map to obtain a weighted feature map that aggregates global information.

[0032] The convolutional layer adjustment unit performs deep feature fusion and channel dimension adjustment on the obtained weighted feature map to obtain the initial output feature map;

[0033] The residual connection unit adds the initial output feature map element-wise to the original input features (i.e., the comprehensive feature map) that have not undergone attention processing, to obtain the final output feature map of the attention module.

[0034] Preferably, the normalization unit uses the Softmax function.

[0035] Furthermore, the second deep learning model is preferably a conditional diffusion model. This model is trained to learn a denoising function, and during the inference phase, conditioned on a low-quality image and a coarse segmentation mask, it gradually reconstructs a clear target image from pure Gaussian noise through a back-diffusion process. This method avoids common problems in generative adversarial networks such as pattern collapse, training instability, and artifact generation, ensuring high realism and structural fidelity of the reconstructed image.

[0036] As a preferred option, the second neural network model is the Palette model.

[0037] Furthermore, a high-quality, clear image is obtained using the second neural network model, and the image is reconstructed to restore its proportions to obtain the high-quality reconstructed image.

[0038] Furthermore, based on the obtained high-quality reconstructed images, the thickness of the insulation layer is calculated using the ray intersection method.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) Extremely high measurement accuracy and robust repair: The two-stage progressive reconstruction strategy of this invention effectively avoids the problem of a single model directly processing complex global images. Especially in the fine reconstruction stage, this invention innovatively adopts a conditional diffusion model instead of a traditional generative adversarial network. Compared with GAN models that rely on adversarial training, the diffusion model generates images of higher quality and richer details, and the training process is more stable, fundamentally avoiding mode collapse and artifact problems. More importantly, the iterative denoising mechanism of the diffusion model gives it a unique ability to intelligently repair minor physical damage in images (such as scratches and debris caused by cutting), rather than just enhancing it. This makes this invention exhibit unparalleled robustness when processing low-quality industrial images with physical interference, thereby achieving sub-pixel-level boundary localization and measurement accuracy far exceeding that of existing technologies.

[0041] (2) Full automation and higher efficiency: The entire measurement process, from image acquisition to parameter output, achieves a fully automated closed loop from end to end, without any manual intervention. The processing speed is fast, meeting the stringent requirements of high-speed and online modern industrial production lines, and significantly improving production quality inspection efficiency.

[0042] (3) Non-destructive and universal testing: Non-contact measurement using machine vision does not cause any physical damage to the cable sample, ensuring the integrity of the product and is suitable for comprehensive quality monitoring of all offline products.

[0043] (4) Excellent scalability and adaptability: The method framework of this invention has high modularity and scalability, and the deep learning model can be easily replaced or upgraded according to technological development. More importantly, the method is based on data-driven, and when it is necessary to adapt to new cable models or other products with similar layered structures, only a small number of new samples are needed for model fine-tuning or incremental training to quickly achieve functional transfer, which has strong versatility and industrial adaptability. Attached Figure Description

[0044] Figure 1This is an overall flowchart of the insulation structure measurement method in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the thickness calculation principle based on the ray intersection method in an embodiment of the present invention.

[0046] Figure 3 This is a schematic block diagram of the measuring device in an embodiment of the present invention.

[0047] Figure 4 This is the original image of one of the inputs in the embodiment.

[0048] Figure 5 This is the image after preprocessing such as cropping.

[0049] Figure 6 This is a structural framework diagram of the editor section of the improved U-Net.

[0050] Figure 7 for Figure 6 Structural framework diagram of the spatial attention module.

[0051] Figure 8 The first neural network outputs the image.

[0052] Figure 9 The second neural network outputs an image.

[0053] Figure 10 This is the image after its dimensions have been restored.

[0054] Figure 11 The absolute error distribution histogram obtained for the example is shown. Detailed Implementation

[0055] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. However, the scope of protection of the present invention is not limited thereto.

[0056] Reference Figure 1 This invention provides a method and system for measuring the insulation structure of ultra-high voltage cables based on deep learning. The method includes: acquiring and preprocessing a cross-sectional image of an ultra-high voltage cable; performing semantic segmentation on the image using a first neural network model to obtain a coarse localization mask for the insulation layer; extracting the region of interest based on the mask, and using a second neural network model, preferably a conditional diffusion model, to perform image inpainting and reconstruction on the region of interest to sharpen interlayer boundaries and repair minor physical damage caused by cutting; finally, based on the reconstructed high-quality image, calculating parameters such as the thickness and eccentricity of the insulation layer using a geometric algorithm; specifically including the following steps:

[0057] (1) Image acquisition and preprocessing

[0058] An image acquisition system consisting of an industrial camera, a telecentric lens, and a backlight or ring light source is used to acquire high-resolution digital images of the cross-section of ultra-high voltage cables. To establish an accurate mapping relationship between pixel size and physical size, the camera system needs to be calibrated in advance using a calibration board to obtain the physical size conversion factor K (e.g., unit: mm / pixel).

[0059] The original images acquired (such as) Figure 4 Preprocessing is required, and the process includes: First, performing Huffman detection and ROI cropping on the image, recording the cropping location information, removing irrelevant background, and retaining only the core area of ​​the cable cross-section; then, scaling the cropped image to a fixed resolution and normalizing it to meet the input requirements of subsequent deep learning models, thus obtaining the original input image, such as... Figure 5 As shown.

[0060] (2) Coarse localization of insulating regions based on semantic segmentation

[0061] The preprocessed cable cross-section image (original input image) is input into the first neural network model for segmentation. This model can employ a semantic segmentation model or network based on the U-Net or DeepLabV3+ architecture, and can incorporate an attention mechanism module to enhance the extraction of insulation layer edge features. This embodiment uses the U-Net neural network model as an example, improving the encoder portion of the existing U-Net neural network structure by introducing an attention mechanism module, while keeping the rest of the structure unchanged. Figure 6 and Figure 7 As shown. Figure 6 The encoder module of the U-Net neural network model, which introduces a spatial attention mechanism module, mainly includes a multi-branch processing unit, a first fusion unit, a spatial attention module, a second fusion unit, and an output processing unit.

[0062] The multi-branch processing unit receives an input feature map (corresponding to the original input image) and contains at least two parallel processing paths. In this embodiment, the unit includes four parallel processing paths for synchronous processing of the input feature map. At least one processing path contains a standard convolutional layer (3 x 3 Conv), while the other processing paths contain dilated convolutional layers with different dilation rates (represented by rate in the figure). By setting different dilation rates, each path can obtain different receptive fields, thereby capturing local details and global contextual information of the input feature map respectively, achieving multi-scale feature extraction.

[0063] First fusion unit: This unit is connected to the output of the multi-branch processing unit. Its function is to fuse the output feature maps from all parallel processing paths. In this embodiment, the fusion operation is implemented by element-wise summation to generate a comprehensive feature map that aggregates multi-scale information.

[0064] Spatial attention module: such as Figure 7 As shown, this module receives the synthesized feature map from the first fusion unit. Its internal structure and working principle are as follows:

[0065] The module's computation begins with an input feature map of dimension C×H×W. To compute self-attention, this input feature map is first fed in parallel into three independent 1x1 convolutional layers (1x1 Conv), which generate intermediate feature maps functionally corresponding to the query, key, and value, respectively. The feature map corresponding to the query, generated by the first (upper) 1x1 convolutional layer, is transposed after dimensionality reshaping; while the feature map corresponding to the key, generated by the second (middle) 1x1 convolutional layer, is not transposed and is directly used for subsequent computation. The transposed feature map from the upper layer is then multiplied by the untransposed feature map from the middle layer (as shown in the diagram). (Symbol representation). The resulting correlation score matrix is ​​normalized using a Softmax function to generate an attention weight map. Each value in this weight map represents the degree of attention a location on the feature map gives to all other locations.

[0066] Meanwhile, the feature map representing the value generated by the third 1x1 convolutional layer (the lower layer) also undergoes corresponding transpose and reshaping operations. Multiplying the attention weight map obtained in the previous step with this value feature map dynamically weights the features according to contextual relationships, thus generating a weighted feature map that aggregates global information. This weighted feature map is then further processed through a 1x1 convolutional layer for deep feature fusion and channel dimension adjustment, resulting in a preliminary output feature map with dimensions C1×H×W.

[0067] Finally, to ensure effective information transfer and prevent gradient vanishing, this module also introduces residual connections. Specifically, the module's initial output feature map is added element-wise to the original input feature map that has not undergone attention processing (as shown in the figure). (Symmetric representation). The final sum is the output feature map of the attention module. Through this series of operations, the module effectively enhances the representation of the original features, enabling it to focus on a broader receptive field and more critical information regions.

[0068] The second fusion unit (residual connection): The input of this unit is connected to both the output of the spatial attention module and the input of the entire multi-scale feature fusion module (i.e., the encoder module). It fuses the attention-weighted feature map with the original input feature map through residual connections (e.g., element-wise addition). This structure aims to preserve the integrity of the original features, prevent information degradation in multi-layer processing, and help stabilize the entire processing flow.

[0069] Output processing unit: This unit is connected to the output of the second fusion unit and typically uses a 1x1 convolutional layer (also known as a unit convolution). Its function is to perform channel integration and dimensionality reduction on the feature map after residual connection, and finally output a feature map with stronger representational power after being enhanced by multi-scale information and spatial attention.

[0070] The model of this invention learns the mapping from image pixels to categories such as "insulation layer," "conductor," and "background" by training on a large number of labeled samples. During the inference phase, the model outputs a segmentation mask of the same size as the original image, based on the input complete cable region of interest image. This achieves coarse localization of the target region and effectively filters out interference from conductor stranding textures and background noise.

[0071] Step (2) yields a mask that retains only information about the insulating layer region; see [link to relevant documentation]. Figure 8 .

[0072] (3) Boundary repair and reconstruction based on guided diffusion model

[0073] Using the mask generated in step (2), the region of interest (ROI) image block containing the inner and outer boundaries of the insulating layer is extracted from the original image (i.e., the corresponding image obtained after processing in step (1)) to obtain the insulating region image, i.e., the ROI image block. In view of the problems of blurred boundaries, low contrast and cutting burrs in industrial field imaging, this step uses the second neural network model to perform image repair and reconstruction on the ROI image block.

[0074] The second neural network model preferably employs a conditional diffusion model (the Palette model is used in this embodiment), the core idea of ​​which is to generate data by learning a reverse noise addition process. This invention performs image reconstruction based on the aforementioned generative network.

[0075] The above conditional diffusion model includes two processes: forward diffusion and backward diffusion.

[0076] Forward diffusion process: During the training phase, a small amount of Gaussian noise is gradually and repeatedly added to a perfect, high-quality, clear image until it becomes pure random noise.

[0077] Backdiffusion process: The goal of model learning is to predict added noise given a noisy image and a time step. During the inference phase, the model starts with a pure Gaussian noise map, using the low-quality ROI image to be tested as a conditional guide. Through multiple iterations of calling the noise prediction network (i.e., the conditional diffusion model mentioned above), the predicted noise is gradually subtracted from the noise, ultimately "sculpting" a clear, complete, high-quality image.

[0078] Given the difficulty of directly capturing perfectly clear images for labeling in industrial settings, this embodiment also employs a "data synthesis and degradation" strategy to construct the training set. The specific steps are as follows:

[0079] High-quality, well-defined cable cross-section images are acquired under ideal laboratory lighting conditions or using high-precision scanning equipment, serving as the target domain image (i.e., a high-quality, clear image).

[0080] The high-quality, clear images were artificially degraded using existing image processing algorithms to simulate the challenges of a real industrial environment. This included applying Gaussian blur of random degree, superimposing salt and pepper or Gaussian noise, adjusting local gamma values, and artificially adding simulated scratches or debris patterns to obtain the artificially degraded image (degraded image).

[0081] Artificially degraded images are used as conditions, forming training pairs with corresponding high-quality, sharp images. The conditional diffusion model is trained to learn how to recover the original high-quality, sharp image from noise given a degraded image as guidance. The loss function is typically the mean squared error between the predicted noise and the actual added noise.

[0082] During the online detection phase, the low-quality ROI image to be tested is input into a trained conditional diffusion model. Through a back-diffusion process, a reconstructed image with clear binarization features, sharp boundaries, and effective repair of minor imperfections is generated. This results in a high-quality, clear image obtained from the aforementioned conditional diffusion model. (See [link to relevant documentation]). Figure 9 .

[0083] After obtaining a high-quality, clear image, based on the cropping position information corresponding to step (1), the image ratio of the obtained high-quality, clear image is restored to obtain a high-quality, clear image (i.e., a high-contrast image or a high-quality reconstructed image) with the same ratio as the original image. See the image after ratio restoration. Figure 10 .

[0084] (4) Calculation of measurement parameters including thickness based on the ray intersection method

[0085] Reference Figure 2 Automated geometric measurements are performed on the reconstructed high-contrast image (i.e., the high-quality clear image obtained in step (3)).

[0086] Contour extraction: The inner and outer contours of the insulating layer are extracted using the Canny operator or morphological gradient algorithm.

[0087] Center Fitting: For the extracted outer contour point set, a circle is fitted using the least squares method to determine the center coordinates of the outer contour of the insulating layer. and inner contour center coordinates .

[0088] Ray scanning: Using the center coordinates of this outer contour as the origin, multiple virtual measurement rays are emitted within a range of 0° to 360° with preset angle increments. For each ray, the coordinates of its intersection with the inner and outer boundary contours are calculated (x, y, y). in , y in ) and (x out , y out ).

[0089] Thickness and eccentricity calculation: based on the center coordinates of the outer contour of the insulation layer obtained by fitting. and inner contour center coordinates First, calculate the Euclidean distance between the centers of the two circles. The formula is as follows:

[0090]

[0091] Subsequently, the average thickness of the insulation layer obtained by X-ray diffraction was combined with... Calculate the eccentricity of the insulation layer The formula for calculating eccentricity is defined as the distance between the centers of the circles. With average insulation layer thickness The ratio of is expressed mathematically as follows:

[0092]

[0093] Finally, the local thickness values ​​measured in all ray directions are statistically analyzed, and their average, maximum, and minimum values ​​are calculated. These pixel-level values, along with the calculated eccentricity parameter, are then multiplied by the physical dimension conversion factor obtained beforehand through camera calibration. This will give you the final physical geometric parameters in millimeters (mm).

[0094] To verify the feasibility and advantages of the above method, 100 images were randomly selected from 3300 images for thickness measurement. Python programming was used to calculate the insulation layer thickness, obtaining the corresponding thickness measurements. The measurement error of the system was statistically analyzed by comparing the automated measurement results with the reference values.

[0095] Experimental results show that the mean absolute error (MAE) of this measurement system is 0.009 mm, and the maximum absolute error does not exceed 0.015 mm. Figure 11 The absolute error distribution histogram shows that over 90% of the sample measurement errors are concentrated within 0.012 mm, indicating that the system not only has high accuracy but also exhibits good consistency and stability when processing different samples.

[0096] Reference Figure 3 The invention employs a high-voltage cable insulation thickness measuring device 300, which includes:

[0097] Image acquisition module (301): used to acquire digital images of the cross-section of high-voltage cables, such as industrial cameras and light sources.

[0098] Preprocessing module (302): Connected to the image acquisition module, used for preprocessing the image such as cropping and size normalization.

[0099] Region segmentation module (303): Used to run the first neural network model, obtain the mask corresponding to the preprocessed image, and then segment the preprocessed image to output the region of interest, i.e., ROI image block.

[0100] Boundary reconstruction module (304): Used to run the second neural network model, reconstruct the image of the region of interest, output an image with clear boundaries, restore the size of the image, and obtain a high-quality clear image (or a high-quality reconstructed image).

[0101] Thickness calculation module (305): used to perform geometric measurements on the reconstructed image, calculate and output insulation thickness parameters, etc.

[0102] Storage module (306): Used to store neural network models, intermediate data and final measurement results.

[0103] Control and processing module (307): Used to schedule and run the functions of the above modules.

[0104] The above modules can be implemented through software, hardware, or a combination of both. For example, the trained first and second neural network models can be integrated and deployed on a high-performance edge computing unit equipped with a GPU to achieve online, real-time detection of images on high-voltage cable production lines, meeting the needs of industrial automation quality control.

[0105] In summary, this invention constructs a complete automated measurement process by cascading two complementary deep learning models and combining them with precise automated geometric calculations. This method effectively solves the common image quality degradation problem in industrial settings, achieving high-precision, high-efficiency, non-destructive thickness measurement, and possesses significant practical value and broad application prospects.

Claims

1. A method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images, characterized in that, include: Acquire images of the cable cross-section and perform preprocessing to obtain the original input image; The original input image is input into the trained first neural network model to obtain a segmentation mask; the obtained mask is used to extract the region of interest (ROI) image patch containing the inner and outer boundary information of the insulation layer in the original input image; the constructed second neural network model is used to perform image inpainting and reconstruction on the ROI image patch to obtain a high-quality reconstructed image; the cable insulation structure is measured based on the obtained high-quality reconstructed image; the second neural network model adopts the conditional diffusion model.

2. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 1, characterized in that, The first neural network model is a semantic segmentation model.

3. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 1, characterized in that, The first neural network model is selected from either the U-Net semantic segmentation model or the DeepLab semantic segmentation model.

4. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 3, characterized in that, The U-Net architecture semantic segmentation model or the DeepLab architecture semantic segmentation model is an improved model structure that integrates an attention mechanism module into the encoder part.

5. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 4, characterized in that, In the improved U-Net semantic segmentation model structure, the encoder includes: A multi-branch processing unit contains at least two parallel processing paths. Each processing path receives the original input image and obtains the output feature map corresponding to the parallel processing path. The first fusion unit fuses the output feature maps from all parallel processing paths to obtain a comprehensive feature map. The spatial attention module receives the comprehensive feature map from the first fusion unit and performs spatial attention weighting to obtain the weighted feature map. The second fusion unit fuses the weighted feature map output by the spatial attention module with the original input image through residual connection to obtain the feature map after residual connection. Output processing unit: Performs channel integration and dimensionality reduction on the feature map after residual connection, and outputs the feature map.

6. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 5, characterized in that, The spatial attention module includes: Three parallel convolutional layers are used to generate intermediate feature maps that functionally correspond to the query, key, and value from the input feature map, respectively. The first dimension reshaping and transpose unit performs dimension reshaping and transpose operations on the intermediate feature map corresponding to the query. The first matrix multiplication unit performs matrix multiplication on the feature map corresponding to the query after dimension reshaping and transpose operations and the intermediate feature map corresponding to the key to obtain the correlation score matrix. The normalization unit normalizes the correlation score matrix output by the matrix multiplication unit to obtain the attention weight map. The second dimension reshaping and transpose unit performs dimension reshaping and transpose operations on the intermediate feature map corresponding to the value; The second matrix multiplication unit performs matrix multiplication on the feature map corresponding to the value obtained after dimension reshaping and transpose operations and the attention weight map to obtain a weighted feature map that aggregates global information. The convolutional layer adjustment unit performs deep feature fusion and channel dimension adjustment on the obtained weighted feature map to obtain the initial output feature map; The residual connection unit adds the preliminary output feature map to the original input comprehensive feature map without attention processing element by element to obtain the final output feature map of the attention module.

7. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 1, characterized in that, The second neural network model is the Palette model.

8. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 1, characterized in that, A high-quality, clear image is obtained using the second neural network model, and the image is then reconstructed to its original scale to obtain the high-quality reconstructed image.

9. The method for measuring the insulation structure of ultra-high voltage cables based on weak boundary images according to claim 1, characterized in that, Based on the obtained high-quality reconstructed images, the thickness of the insulation layer was calculated using the ray intersection method.

Images