Intelligent qualitative analysis method and system for incentives of surface protrusions of cable shielding material

Abstract

The invention provides an intelligent qualitative analysis method and system for incentives of protrusions on the surface of a cable shielding material, and belongs to the technical field of analytical testing. Further, an automatic processing method based on deep learning development is adopted to carry out parameter extraction and automatic classification on the defect area, intelligent qualitative analysis is carried out on incentives of shielding layer surface protrusions, the method is high in automation degree and flexibility, a good recognition effect can be obtained through light-weight training, the operation process is simple, and the efficiency is high. Corresponding empirical parameters can be obtained based on different application scenes, and the obtained conclusion is more real and reliable.

Description

Technical Field

[0001] This invention belongs to the field of analytical testing technology, specifically relating to an intelligent qualitative analysis method and system for the causes of surface protrusions in cable shielding materials. Background Technology

[0002] High-voltage cables, as core electrical equipment for power transmission, play a crucial role in urban power grid construction and offshore wind power deployment. Against the backdrop of a new round of national power grid transformation, my country's power grid construction is accelerating, and urban power grid upgrades are being vigorously promoted. This is driving the growth in demand for high-voltage and ultra-high-voltage cables, which in turn will significantly increase the demand for domestically produced ultra-smooth high-voltage semi-conductive shielding materials.

[0003] The semi-conductive shielding layer is a key component of the insulation system of cross-linked polyethylene insulated high-voltage power cables. It is prepared by melt-blending and extrusion of matrix resin, conductive carbon black, dispersants, and other processing aids. In the cable structure, it plays a crucial role in eliminating air gaps at the interface between the metal core and the insulation layer, and in ensuring a uniform electric field distribution at the interface. As a multi-layered insulation system, the service life and reliability of high-voltage cables are closely related to the performance of the semi-conductive shielding layer. Its key properties, such as mechanical properties, electrical properties, and surface finish, determine the upper limit of the voltage rating of the high-voltage cable. Because the semi-conductive shielding layer contains very small amounts of carbon black, surface protrusions can easily form during processing, affecting the interface matching between the insulation layer and the semi-conductive shielding layer. According to the Mason equation, surface protrusions in the semi-conductive shielding layer lead to significant distortion of the electric field at the insulation-shielding layer interface, thereby reducing the service life of the high-voltage cable. According to the Chinese standard GB / T 18890.2—2015, the semi-conductive shielding layer should be uniformly extruded onto the insulation and firmly bonded to it. Therefore, the higher the voltage level of the cable, the higher the requirements for the surface smoothness of the semiconductive shielding layer, and the more stringent the control of the size of the surface protrusions, thus ensuring a stronger ability to ensure a uniform electric field in the semiconductive shielding layer.

[0004] According to existing literature, traditional methods for evaluating surface finish mainly include industrial optical microscopy, white light three-dimensional interferometry, laser profilometry, and atomic force microscopy (AFM). None of these four methods can accurately reflect the true morphology of protrusions on the surface of a semiconductive shielding layer. Industrial optical microscopy can only reflect the 2D morphology of the protrusions; laser profilometry can obtain data with some theoretical guidance; white light three-dimensional interferometry can characterize the outline size of the protrusions, not their actual size; and AFM can effectively reflect the true three-dimensional morphology of the material. However, since the height of most surface protrusions is greater than 2µm, distortion is easily introduced during AFM probe scanning, making it impossible to effectively characterize the true morphology of the protrusions on the shielding layer surface. Patent authorization number CN 113183434 B proposes a detection system and method for surface protrusions on a high-voltage cable semiconductive shielding tape. It employs continuous real-time optical detection, resulting in more accurate detection results and timely reconstruction of the protrusion size for a more comprehensive evaluation of the surface finish of the extruded tape. However, it fails to effectively analyze the causes of surface protrusions. Furthermore, traditional methods for analyzing the formation of protrusions on the surface of cable shielding layers include proton-induced X-ray fluorescence (PIXE) and neutron activation analysis (NAA). These methods quantitatively measure the concentration and distribution of mineral impurities in semiconductive shielding materials produced from furnace black and acetylene black, indirectly indicating that the main source of surface protrusions is impurities in the carbon black. While effective and feasible, these methods only provide indirect conclusions and cannot directly and qualitatively analyze the causes of protrusions on the cable shielding layer surface. The conclusions are somewhat limited and time-consuming, making them impractical for direct application in real-world situations.

[0005] In summary, this paper proposes an intelligent qualitative analysis method and system for the causes of surface protrusions in cable shielding materials. An automated processing method based on deep learning is used to extract parameters and automatically classify defect areas of surface protrusions in the cable shielding layer. Furthermore, the causes of these surface protrusions are directly analyzed. This method has significant theoretical guiding significance for optimizing the surface smoothness of cable shielding layers, a key performance characteristic. It can effectively provide feedback to the upstream of shielding material production, thereby optimizing the raw materials for the production of ultra-smooth high-voltage cable semiconductive shielding materials to achieve practical production goals. Summary of the Invention

[0006] In view of this, the present invention provides an intelligent qualitative analysis method and system for the causes of surface protrusions in cable shielding materials, in order to solve the problem that the existing technology cannot directly and qualitatively analyze the causes of surface protrusions in cable shielding layers, and the conclusions are somewhat one-sided, time-consuming and labor-intensive, and not conducive to direct use in practice.

[0007] The technical solution adopted in this invention is as follows: A method for intelligent qualitative analysis of the causes of surface protrusions in cable shielding materials, comprising: Step 1: Obtain a shielding layer sample. Slice the protrusions of the shielding layer sample to obtain the polished surface after the protrusions are cut open. Mark any abnormal areas on the polished surface to obtain a sample image; specifically including: Self-made semi-conductive shielding material was extruded into tape. Protrusions with a 2D outer diameter greater than 50µm were screened and marked using an industrial optical microscope. The tape was cut to ensure the marking points were close to the sample edge. The sample surface was repeatedly cleaned with ethanol, and then the marking points were covered with PTFE tape. The sample thickness was 0.5-1mm, width 1-3mm, and length 5-10mm. The obtained samples were transferred to a laser confocal microscope for more intuitive and three-dimensional morphological characterization of the circular protrusions on the sample surface. The height distribution of the protrusions and their surrounding areas was obtained, and the distance-Z height curve between the protrusions and the sample edge was analyzed. Characterization was performed in live mode at 20x magnification. The Z-track scanning depth and laser wavelength were determined according to actual conditions. The pinhole was adjusted to 30-50µm, and the gain master value was set to 300-500. With an offset of 0 and a digital gain of 1, the image is designed to avoid overexposure and minimize background noise. A cryogenic ultramicrotome is used to polish the raised areas on the sample surface to further analyze the causes of the circular raised areas and expose their underlying nature. The slicing temperature is -40℃ to -60℃. The slicing depth is determined based on the Distance-Z height curve obtained in step 3, which specifies the X-axis dimension of the surface protrusion and its distance from the sample edge. The slicing depth is strictly less than the distance *a* between the far edge (F-Far) of the circular raised area and the sample edge, and greater than the distance *b* between the near edge (C-Close) of the circular raised area and the sample edge. A polarizing microscope is used to monitor the slicing depth in real time to ensure appropriate depth. Slicing is stopped when the slice position is observed between the F and C ends. The polished surface is then qualitatively analyzed to observe and mark any abnormal areas, yielding images for subsequent analysis. The basic components of the self-made semi-conductive shielding material extruded tape include an EBA (ethylene butyl acrylate copolymer) resin matrix, high-structure CB (carbon black), a DBP absorption value of 120-150 ml / 100g, an ash content of <0.1%, and a pH range of 8.5-9.5. The extruded tape has a thickness of 1 mm, a width of 3 cm, and an observation area of ​​1 cm² under an industrial optical microscope. 2 The screening sites were repeatedly cleaned with ethanol and then sealed with PTFE tape to ensure no interference from impurities.

[0008] Step 2: Convert the sample image to grayscale and refine it into blocks. While removing the background, highlight the regional features and avoid overexposure to obtain the preprocessed image. Step 2 specifically includes: The sample image is grayscaled using OpenCV; the grayscaled image is further subdivided into blocks according to the principle that the number of pixels in a single block is greater than 16; the image histogram information of each subdivided block is statistically analyzed, and an equalization histogram H is generated; a threshold T is set according to the grayscale features of the grayscaled sample image, and the pixel grayscale range from H[0] to H

[255] in the equalization histogram H is traversed sequentially, and pixels with grayscale values ​​greater than the threshold T are filtered out, and the number of filtered pixels is counted, and the number of filtered pixels is defined as Sub-pixel; the mean number of pixels is calculated, and the mean number of pixels is defined as Mean-pixel; all pixel grayscale values ​​in the range from H[0] to H

[255] in the equalization histogram H are added to Mp, and the equalization histogram H is redistributed to obtain the preprocessed image.

[0009] Step 3: Extract the region boundaries from the preprocessed image obtained in Step 2, and overlay the boundaries onto the preprocessed image in Step 2 to form a multi-information fusion image; Step 4: Import the fused image into the convolutional neural network for hierarchical processing, and based on the Sigmoid function, superimpose and fuse all the hierarchically processed feature maps to obtain the corresponding binary mask image; In step 4, the convolutional neural network uses 3×3 pixel blocks as convolutional kernels and includes an encoding layer, a decoding layer, and a skip link component. The encoding layer employs multiple convolutions, followed by linear rectified function activation and pooling, halving the fused image each time to double the number of feature channels, extracting multi-level features of the image, progressively compressing spatial information, and expanding the receptive field. The layered processing refers to the results obtained from multiple encoding layers in the encoder, with each encoder layer producing one layer. The decoding layer employs multiple deconvolutions, concatenating the outputs of the corresponding encoding layers, and then importing them into the convolutional layer for feature extraction, outputting a segmentation mask of the same size as the input. The skip link component specifically involves concatenating the output of each encoding layer with the corresponding layer of the decoding layer to ensure enhanced boundary clarity. Subsequently, the feature maps of each layer are fused based on the sigmoid function, and the final binary mask image is output. This invention preferentially uses the U-Net structure or related variants. The U-Net structure is an existing structure and will not be described in detail here.

[0010] Step 5: Analyze and calculate parameters of the feature regions extracted from the binary mask image. The parameters include: total defect area, defect proportion and distribution density, number of defects, defect size and centroid coordinates. Then, convert the length of a single pixel into data measured in micrometers according to the scale. Calculate the dispersion index U of the defect sites in the feature regions through cluster analysis. Classify the feature regions according to the U value. After classification, generate a defect heatmap and a defect area distribution histogram. In cluster analysis, the dispersion index U of defect sites is calculated based on the spatial distribution index (SDI), nearest neighbor index (R), and coefficient of variation (CV) of defect sites within the feature region. After normalization, the information entropy of each coefficient is calculated using the entropy weight method. The weight of each coefficient is then determined based on the information entropy to obtain the dispersion index. The images are initially screened and classified based on the U value. The calculation formulas and steps for each coefficient are as follows: First, divide the image into n×n grids of equal area; ,in The number of red sites within the i-th grid. It is the average number of sites within each grid cell. If it is a complete color patch, SDI << 1. Where r is the average nearest distance of the site, For the red site, R is the areal density. If it is a complete color patch, R > 1 and is close to... ; , where S is the standard deviation of the number of red spots in each grid; if it is a complete color block, CV≈0; Information entropy is calculated based on m samples. SDI, R, and CV are calculated for each of the m similar images, with each coefficient corresponding to a set of data. These three coefficients are then organized into an m×3 matrix X. For each element in matrix X... (where j = 2 represents SDI, j = 3 represents R, and j = 3 represents CV) After normalization, its information entropy and information utility value are calculated, and the weights are determined. The relevant calculation formulas are as follows: Normalized ,in It is the minimum value in the j-th column of data. It is the maximum value in the j-th column of data.

[0011] Information entropy ,in , representing the proportion of the j-th coefficient in the i-th sample, is defined as follows: hour, ; Weight The corresponding weights of each coefficient are calculated. , and .

[0012] In this scheme, the spatial distribution index, nearest neighbor index, and coefficient of variation are parameters used to evaluate the location, area, and uniformity of distribution of certain feature regions or sites in an image. The objective entropy weighting method is used to calculate the information entropy corresponding to SDI, R, and CV, obtaining the corresponding weights to determine the final evaluation index of color patch dispersion or distribution. The greatest advantage of this method is that it can be trained on a small number of image inputs to calculate the weights of each coefficient, better adapting to a specific domain and ensuring the accuracy and universality of the method. For example, in the domain covered by this scheme, if the execution standards of the samples are consistent, although the samples are different, the information they contain, such as the dispersion of fillers and the embedding of impurities, is the same, and the weights of the corresponding coefficients will be within a similar range. This method, when performing image recognition in this domain, can ensure the accuracy of the classification results by assigning values ​​based on the objective entropy weighting method, and at the same time, it also increases the error tolerance and universality of the recognition.

[0013] Step 6: Calculate the fractal dimension D of the defect site area of ​​the feature region based on the parameters in Step 5. According to the pattern in the generated defect heat map and defect site area distribution histogram, and combined with the fractal dimension D of the feature region and the empirical parameter X in the established database, automatically classify the feature region after classification, so that the automatically classified feature region corresponds to the verified cause of the protrusion on the shielding layer surface, and automatically output the inference result.

[0014] The extracted parameters are further used to generate a heatmap of defect site distribution and an area distribution histogram, and a database is established. Based on the concentration or dispersion of defect sites and the concentration or dispersion of defect site area distribution in the area distribution histogram, manual classification is performed, with single peaks or multiple peaks. Based on the empirical parameter X obtained from micro-training on a small number of samples, the defect area can be automatically classified into two types: "carbon black agglomerates" or "hard carbon embeddings". Defect areas that do not correspond to those in the database are transferred to the manual identification library. After manual assisted analysis, the X empirical parameter is transferred back to the database for optimization.

[0015] A smart qualitative analysis system for the causes of surface protrusions in cable shielding material includes: Sample acquisition module: used to acquire shielding layer samples, slice the protrusions of the shielding layer samples to obtain the polished surface after the protrusions are cut open, mark abnormal areas of the polished surface, and obtain sample images; Preprocessing module: Used to perform grayscale processing on sample images and refine the blocks, while removing the background, highlighting regional features and avoiding overexposure, to obtain a preprocessed image; Fusion module: Extracts region boundaries from the preprocessed image obtained in step 2 and overlays the boundaries onto the preprocessed image in step 2 to form a multi-information fusion map; Overlay module: The fused image is imported into the convolutional neural network for hierarchical processing, and based on the Sigmoid function, all the hierarchically processed feature maps are overlaid and fused to obtain the corresponding binary mask image; Calculation module: Analyzes and calculates parameters of the feature regions extracted from the binary mask image. The parameters include: total defect area, defect proportion and distribution density, number of defects, defect size and centroid position coordinates. Cluster analysis is used to calculate the dispersion index U of defect sites in the feature region. The feature region is classified according to the value of U. After classification, a defect heat map and a defect area distribution histogram are generated. Output module: Based on the parameters in the calculation module, calculate the fractal dimension D of the defect site area of ​​the feature region, and according to the pattern in the generated defect heat map and defect site area distribution histogram, combine the fractal dimension D of the feature region with the empirical parameter X in the established database for comparative analysis, automatically classify the classified feature region, so that the automatically classified feature region corresponds to the verified cause of the protrusion on the shielding layer surface, and automatically output the inference result.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention directly analyzes the protrusion sites on the surface of the shielding layer, exposing their corresponding defect areas. Furthermore, it employs an automated processing method based on deep learning to extract parameters and automatically classify the defect areas, and intelligently qualitatively analyzes the causes of the protrusions on the surface of the shielding layer. This method has a high degree of automation and flexibility, and can achieve good recognition results with lightweight training. The operation process is simple, and corresponding empirical parameters can be obtained based on different application scenarios, resulting in more realistic and reliable conclusions.

[0017] 2. This invention is based on the physical differences in internal defects, impurities, and processing marks of shielding materials. It can segment images by combining color / brightness differences and adjusting thresholds. The constructed convolutional neural network Python+OpenCV script can classify each pixel individually, achieving high-precision segmentation while preserving minute structures. This ensures good recognition of small dark areas and clarity of irregular defect boundaries, thereby improving the accuracy of extracting relevant parameters of defect areas, such as defect quantity, size, and centroid coordinates. This effectively determines the attribution of defect areas and provides a theoretical basis for the automated classification of protrusion causes on the shielding layer surface. The multi-level screening and self-optimizing characteristics ensure accurate qualitative results for abnormal areas and high efficiency of the method in multiple scenarios. Attached Figure Description

[0018] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 is a schematic diagram of the process of the present invention; Figure 2 The results of screening the surface protrusions of the shielding layer of the present invention using an industrial optical microscope; Figure 3 The morphology characterization results and Distance-Z height curves of the protrusions on the surface of the shielding layer of the present invention were obtained by laser confocal microscopy. Figure 4 is a schematic diagram of the slicing process of the cryo-ultrathin section of the present invention; Figure 5 is a polarizing microscope photograph of the present invention; Figure 6 is a multi-information fusion diagram of the feature region of the present invention; Figure 7 is a binary mask diagram of the feature region of the present invention; Figure 8 shows the centroid coordinates of the defect sites in the feature region of the present invention, as well as the extraction of the area, cluster analysis, and dispersion index output. Figure 9 shows the defect thermal map and density distribution map of the characteristic region of the present invention; Figure 10 shows the histogram of the area distribution of defects and the fractal dimension diagram of the experimental object of the present invention; Figure 11 shows the histogram of the area distribution of defects and the fractal dimension diagram of the comparative object of the present invention; Figure 12 shows the results of abnormal regions in the AFM modulus scanning experimental object (left) and the control object (right) of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. Example

[0025] This invention discloses an intelligent qualitative analysis method for the causes of surface protrusions in cable shielding materials, wherein: The experimental subject was a sample with known surface protrusions in the shielding layer caused by carbon black agglomerates; The comparison object is a sample with known surface protrusions in the shielding layer caused by hard carbon embedding; The above samples ensure the universality and accuracy of the results.

[0026] The experimental procedure is as follows: The flowchart of the method for intelligent qualitative analysis of the causes of protrusions on the surface of a shielding layer based on convolutional neural networks is as follows: Figure 1 As shown, this method is lightweight, highly accurate, highly automated, and flexible. It can obtain the critical dispersion index U and empirical parameters X based on lightweight training and generate a database. Through continuous optimization and iteration based on image recognition results, it can significantly improve the degree of automation and even form a fully automated process. The following mainly explains the processing flow of the experimental subjects, and verifies the reliability of the method based on the results of intelligent qualitative analysis of the experimental and control subjects (which is also the training process of the method).

[0027] First, a traditional industrial optical microscope was used to screen defects based on their 2D size (>50µm), and the results were as follows: Figure 2 As shown, the 2D size of the screening site in the image is approximately 89µm, consistent with the characteristics of protrusions on the surface of the shielding layer. Subsequently, a higher-precision laser confocal microscope with image depth synthesis technology was used to characterize the morphology of the sample, as shown below. Figure 3 As shown, the results indicate that the surface protrusions of the shielding layer are relatively regular "hill-shaped" ridges with some degree of depression in the surrounding local areas. The dimension along a specific X-axis is 150µm, and the overall average height is approximately 20µm. The top of the surface protrusions is relatively flat but has undulations of 2-3µm. Based on the Distance-Z height curve, the distance 'a' between the far edge (F-Far) of the circular protrusion and the sample edge (right side) is approximately 220µm, which is greater than the distance 'b' between the near edge (C-Close) of the circular protrusion and the sample edge, approximately 75µm. Therefore, the subsequent ultrathin sectioning depth can be determined to be greater than 75µm and less than 220µm. The sectioning process is as follows... Figure 4 As shown. After successful sample sectioning, real-time observation was performed using a polarizing microscope. Due to the high resolution of the polarizing microscope and its suitability for observing birefringent materials, it is easier to confirm the presence or distribution of protruding sites and "heterogeneous regions" on the polished surface of the sample, such as... Figure 5 As shown. This result indicates that determining the advance depth of the ultrathin slicer tip based on the Distance-Z height curve is appropriate and feasible, and the protrusions on the shielding layer surface were successfully cut open (e.g., Figure 4 Left), its polished surface (such as Figure 4 Observation on the right revealed a heterogeneous region below the convex point. The interface between the two regions was obvious, with a strong hard texture and a width of about 200µm, which also facilitated the acquisition of images for subsequent analysis.

[0028] First, OpenCV is used to convert the image to grayscale and then refine it into blocks. Thresholds for each block are redistributed to remove the background while highlighting regional features and avoiding overexposure. The Sobel operator is then used to extract region boundaries, which are then superimposed onto the original image to form a multi-information fusion map. Figure 6 Subsequently, it is imported into the constructed convolutional neural network and encoded, decoded, and skipped through connections using 3×3 pixel blocks as convolutional kernels to achieve high-precision extraction of multi-level image features and refinement of boundaries and textures. While expanding the receptive field, layered binarization is performed, and the layers are superimposed and fused to obtain the corresponding binary mask image, such as... Figure 7 As shown.

[0029] Further analysis was conducted using a pre-developed Python-OpenCV script to analyze the extracted portion of the binary mask image, calculating the total area, proportion, and distribution density of defects (dark areas). Simultaneously, the number and size of defects were statistically analyzed, and the centroid coordinates were extracted. Based on the original samples, the dispersion index calculation formula was derived. The dispersion index U of the defect sites in the characteristic region was calculated and compared with the critical dispersion index C (0.09) obtained from training. In Example 1, the dispersion index U of the defect sites in the characteristic region was 0.39. After preliminary classification, a defect heatmap and a defect area distribution histogram were further generated, and the results are as follows. Figure 8 and Figure 9 As shown. Based on the information obtained above (defect centroid coordinates, density, and area), the fractal dimension D of the feature region is calculated. According to the patterns in the obtained defect heatmap and distribution histogram, the fractal dimension D of the feature region is compared and analyzed with the empirical parameter X in the established database. Further, the feature region is automatically classified or transferred to a manual identification database. For example... Figure 1 and Figure 10 As shown, the red area in the thermal distribution map of the defect in the intelligent analysis experimental object is concentrated, and the histogram distribution of the defect area is unimodal. The calculated fractal dimension D is 1.049 < empirical parameter X (1.192), and it is automatically classified as carbon black agglomerates. In contrast, the analysis of the comparative object shows that the histogram distribution of the defect area is multimodal, indicating the existence of multiple heterogeneous regions. The calculated fractal dimension D is 1.385 > empirical parameter X (1.192), and it is automatically classified as a hard carbon embedded region. The classification of this result is consistent with the definition of fractal dimension in this algorithm. The fractal dimension value is between 1 and 2. The closer it is to 1, the more uniform the size and distribution of the defect area. The closer it is to 2, the more heterogeneous the overall structure is, and the more discrete the overall distribution is. When large hard carbon impurities are embedded, they are often broken into relatively intact but dispersed small pieces under shear stress during screw extrusion. The atomic force microscopy (AFM) characterization results further verify the results obtained by this intelligent analysis method, proving the reliability of the algorithm. Figure 12As shown. In exceptional cases, the system will automatically switch to a manual identification database to determine the specific situation and feed the results back to the database for incremental optimization of the empirical parameter X. Lightweight training can achieve relatively accurate identification results. Furthermore, with continuous improvement through extensive training, it can eventually achieve highly automated or even fully automated intelligent qualitative analysis of the causes of protrusions on the shielding layer surface, and further feed these results back into production practice.

[0030] Based on the above conclusions, it is inferred that the surface protrusions of the shielding layer of the experimental object are due to the solid sieve residue impurities in the carbon black product breaking down under the shear field during processing. These impurities then combine with the matrix during the mixing process. Because the impurities are large and difficult to disperse, carbon black agglomerates form in situ during the extrusion of the shielding material strip, causing bulges on the strip surface. The control object, on the other hand, has hard carbon impurities introduced during processing embedded in the material. Under the shear field, these impurities break down into multiple small, intact pieces, exhibiting heterogeneous "domains" at a certain depth as described above. Due to the density difference, the surface protrusions are "hill-like," hence the surface protrusions of the shielding layer. Further, continuous optical microscopy (OMC) was used to randomly inspect the continuously extruded cable semiconductive shielding layer. This method was used for qualitative analysis of the causes, and the results are shown in Table 1. The inspection results indicate that the surface defects of the produced semiconductive shielding layer are mainly due to the effective dispersion of carbon black during processing, resulting in the formation of multiple carbon black agglomerates. A small number of surface defects are also caused by the introduction of hard carbon or other impurities. Based on the comprehensive analysis of these results, it is necessary to further optimize the processing technology to ensure sufficient dispersion of carbon black. At the same time, there are still a small amount of impurities in the carbon black product used, and further optimization is needed in terms of processing technology and impurity control of carbon black product.

[0031] In summary, the intelligent qualitative analysis method for the causes of surface protrusions in cable shielding materials allows for more accurate and intuitive observation of the true three-dimensional morphology of protrusions on the surface of cable shielding layers and intelligent qualitative analysis of their causes. This method has significant theoretical guiding significance for optimizing the key performance of surface smoothness in cable shielding layers. It can effectively provide feedback to the upstream of shielding material production, thereby optimizing the raw materials for the production of ultra-smooth high-voltage cable semiconductive shielding materials to achieve practical production goals.

[0032] Table 1. Results of intelligent qualitative analysis of surface defects: ; Example 2 This invention discloses an intelligent qualitative analysis system for the causes of surface protrusions in cable shielding materials, comprising: Sample acquisition module: used to acquire shielding layer samples, slice the protrusions of the shielding layer samples to obtain the polished surface after the protrusions are cut open, mark abnormal areas of the polished surface, and obtain sample images; Preprocessing module: Used to perform grayscale processing on sample images and refine the blocks, while removing the background, highlighting regional features and avoiding overexposure, to obtain a preprocessed image; Fusion module: Extracts region boundaries from the preprocessed image obtained in step 2 and overlays the boundaries onto the preprocessed image in step 2 to form a multi-information fusion map; Overlay module: The fused image is imported into the convolutional neural network for hierarchical processing, and based on the Sigmoid function, all the hierarchically processed feature maps are overlaid and fused to obtain the corresponding binary mask image; Calculation module: Analyzes and calculates parameters of the feature regions extracted from the binary mask image. The parameters include: total defect area, defect proportion and distribution density, number of defects, defect size and centroid position coordinates. Cluster analysis is used to calculate the dispersion index U of defect sites in the feature region. The feature region is classified according to the value of U. After classification, a defect heat map and a defect area distribution histogram are generated. Output module: Based on the parameters in the calculation module, calculate the fractal dimension D of the defect site area of ​​the feature region, and according to the pattern in the generated defect heat map and defect site area distribution histogram, combine the fractal dimension D of the feature region with the empirical parameter X in the established database for comparative analysis, automatically classify the classified feature region, so that the automatically classified feature region corresponds to the verified cause of the protrusion on the shielding layer surface, and automatically output the inference result.

[0033] The circuits, electronic components, and modules involved are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The scope of protection of this invention does not involve any improvement to the software and methods.

[0034] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0035] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for surface protrusion cause intelligent qualitative analysis of cable shielding compound, characterized in that, The application relates to a shielding layer defect detection method based on deep learning. Step 1: a shielding layer sample is obtained, the protrusions of the shielding layer sample are sliced to obtain polished surfaces after the protrusions are cut open, and the abnormal areas of the polished surfaces are marked to obtain a sample picture; Step 2: the sample picture is preprocessed to obtain a pretreatment image; Step 3: a region boundary is extracted from the pretreatment image obtained in step 2, and the boundary is superimposed on the pretreatment image in step 2 to form a multi-information fusion image; Step 4: the fusion image is imported into a convolutional neural network for hierarchical processing, and all the hierarchical processed feature maps are superimposed and fused based on a Sigmoid function to obtain a corresponding binary mask image; Step 5: the feature regions extracted from the binary mask image are analyzed and parameters are calculated, the parameters including a total defect area, a defect proportion and a distribution density, a defect number, a defect size and a centroid position coordinate; The dispersion index U of the defect sites of the feature region is calculated through cluster analysis, the feature region is classified according to the size of the U value, and a defect heat map and a defect area distribution histogram are generated after classification; Step 6: the fractal dimension D of the defect site area of the feature region is calculated based on the parameters in step 5, and the feature region is automatically classified according to the regularity in the generated defect heat map and defect site area distribution histogram, the fractal dimension D of the feature region and the experience parameters X in the established database are compared and analyzed, the automatically classified feature region corresponds to the verified causes of the shielding layer surface protrusions, and an inference result is automatically output.

2. The method of claim 1, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The step 2 specifically comprises: The sample picture is subjected to gray processing based on Open CV; The gray-processed image is subjected to fine block division according to the principle that the number of single block pixels Pixel is greater than 16; Image histogram information of each fine block is counted, and an equalization histogram H is generated; A threshold T is set according to the gray features of the sample picture after gray processing, the pixel gray interval of the equalization histogram H from H[0] to H[255] is sequentially traversed, pixels with a gray value greater than the threshold T are filtered out, and the number of the filtered pixels is counted, and the number of the filtered pixels is defined as Sub-pixel; The mean pixel number is calculated, and the mean pixel number is defined as Mean-pixel; All pixel gray values in the interval from H[0] to H[255] in the equalization histogram H are added to Mp, and the equalization histogram H is redistributed to obtain a pretreatment image.

3. The method of claim 1, wherein the method is characterized by, In step 4, the convolutional neural network takes a 3*3 pixel block as a convolution kernel, and comprises an encoding layer, a decoding layer and a jump link part.

4. The method of claim 3, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The encoding layer adopts multiple convolutions, linear rectifier function activation and pooling, the fusion image is halved each time to double the feature channel number, the multi-level features of the image are extracted, the spatial information is gradually compressed and the receptive field is expanded.

5. The method of claim 3, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The decoding layer adopts multiple deconvolutions, splices the outputs of the corresponding layers of the encoding layer and imports a convolution layer for feature extraction to output a segmentation mask with the same size as the input.

6. The method of claim 3, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The jump link part is specifically a splicing of the output of each layer of the encoding layer and the corresponding layer of the decoding layer.

7. The method of claim 1, wherein the method is a surface protrusion cause intelligent qualitative analysis method of a cable shielding material. In step 5, the calculation of the dispersion index U of the defect sites of the feature region by cluster analysis specifically comprises: The information entropy of each coefficient data is calculated by using the entropy weight method after normalization processing based on the spatial distribution index, the nearest neighbor index and the coefficient of variation of the defect site inside the feature area, and the weight of each coefficient is determined according to the information entropy to obtain the dispersion index wherein, SDI, R and CV respectively represent the spatial distribution index, the nearest neighbor index and the coefficient of variation, , and respectively represent the spatial distribution index weight, the nearest neighbor index weight and the coefficient of variation weight.

8. The method of claim 7, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The information entropy and the weight are calculated by the following method: Based on m samples, SDI, R, CV are calculated respectively for m similar images, each coefficient corresponds to a group of data, and the data of the three coefficients are arranged into an m x 3 matrix X. For the elements in the matrix X j = 1 represents SDI, j = 2 represents R, and j = 3 represents CV) is normalized, the information entropy is calculated, and the weight is determined.

9. The method of claim 8, wherein the surface protrusion cause of the cable shielding material is intelligently qualitatively analyzed. The calculating information entropy after the normalization processing and determining the weight specifically include: the normalized wherein is the minimum value in the jth column data, is the maximum value in the jth column data; Information entropy wherein represents the proportion of the jth coefficient in the ith sample, and is defined as ;​ weights , the corresponding calculation of each coefficient weight , and .

10. A surface protrusion causation intelligent qualitative analysis system for a cable shielding compound, characterized by, Comprise: The sample acquisition module is used to acquire the shielding layer sample, slice the protrusions of the shielding layer sample to obtain the polished surface after the protrusions are cut open, mark the abnormal areas of the polished surface, and obtain a sample picture; The preprocessing module is used to perform grayscale processing and fine block division on the sample picture, peel off the background, highlight the area features, and avoid overexposure to obtain a preprocessed image; The fusion module extracts the region boundary of the preprocessed image obtained in step 2, and superimposes the boundary into the preprocessed image in step 2 to form a multi-information fusion image; The superposition module imports the fusion image into a convolutional neural network for hierarchical processing, and based on the Sigmoid function, superimposes and fuses all the feature maps after hierarchical processing to obtain a corresponding binary mask image; The calculation module analyzes and calculates parameters of the feature regions extracted from the binary mask image, and the parameters include: total defect area, defect proportion and distribution density, defect number, defect size and centroid position coordinates; The dispersion index U of the defect sites of the feature region is calculated by cluster analysis, the feature regions are classified according to the size of the U value, and a defect heat map and a defect area distribution histogram are generated after classification; The output module calculates the fractal dimension D of the defect site area of the feature region based on the parameters in the calculation module, and according to the rules in the generated defect heat map and defect site area distribution histogram, combined with the fractal dimension D of the feature region and the experience parameters X in the established database, comparative analysis is performed on the classified feature regions, so that the automatically classified feature regions correspond to the verified causes of the shielding layer surface protrusions, and the inference result is automatically output.

Images