An intelligent cable fault analysis platform and method

Abstract

The application relates to the technical field of cable fault analysis, in particular to an intelligent cable fault analysis platform, which comprises a table body, a multi-axis moving platform arranged on the table body, a cutting mechanism arranged at the moving end of the multi-axis moving platform, a fixing mechanism arranged on the table body and used for fixing the analyzed cable, and an identification mechanism arranged on the table body and used for identifying and analyzing the cable cut. The application is convenient to carry and use, is convenient to move to a work scene, can absorb and filter dust in the process of dissecting the cable, avoids causing pollution to the environment, can analyze the cable fault type through the identification mechanism, improves the analysis efficiency and the accuracy of the result, and makes the analysis result more standardized.

Description

Technical Field

[0001] This invention relates to the field of cable fault analysis technology, and in particular to an intelligent cable fault analysis platform and method. Background Technology

[0002] As power distribution network operation and maintenance develops towards intelligence and speed, cable fault analysis is gradually shifting from the laboratory environment to on-site operations.

[0003] Traditional cable dissection equipment is mostly fixed in structure, which is large in size, difficult to transport, and cannot adapt to complex field environments.

[0004] Meanwhile, manual dissection remains the dominant method, which is not only inefficient and labor-intensive, but also generates dust that is difficult to control effectively, posing environmental pollution and occupational health risks.

[0005] In addition, existing equipment generally lacks integrated design, with dissection, cleaning and analysis functions being scattered, making it impossible to form a standardized operating procedure and failing to meet the modern power operation and maintenance requirements for efficiency, safety, environmental protection and intelligence.

[0006] Therefore, it is necessary to design a portable cable fault dissection and cleaning platform that is compact, easy to move, and highly integrated in function.

[0007] To address this, a smart cable fault analysis platform and method are proposed. Summary of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to design a portable cable fault dissection and cleaning platform that is compact, easy to move, and highly integrated in function.

[0009] The above-mentioned technical problems are solved by the following technical solution: This invention proposes an intelligent cable fault analysis platform, comprising, Platform; A multi-axis mobile platform, which is mounted on the platform body; A cutting mechanism is located at the moving end of the multi-axis moving platform; A fixing mechanism, which is provided on the platform, is used to fix the cable being analyzed; An identification mechanism, located on the platform, is used to identify and analyze cable cuts.

[0010] In a preferred embodiment of the intelligent cable fault analysis platform of the present invention, it further includes a dust collection mechanism, which comprises: An induced draft fan is mounted on the platform. A duct is provided at the air inlet end of the induced draft fan; A dust filter element is disposed at the exhaust end of the induced draft fan; The platform is equipped with a dust collection station, and the dust collection station is equipped with a dust hopper. The end of the dust hopper is connected to an air duct.

[0011] In a preferred embodiment of the intelligent cable fault analysis platform of the present invention: the platform body is provided with casters; The platform is equipped with a support base.

[0012] In a preferred embodiment of the intelligent cable fault analysis platform of the present invention: the platform body is provided with a protective cover.

[0013] This invention also proposes an intelligent cable fault analysis method for the identification mechanism to identify and analyze cable cuts. The method includes the following steps: Acquire cross-sectional image data of the cable; Preprocess the cable cross-section image to obtain the target image data; The target image data is segmented to obtain the target region; Extract texture features of the target region from the target image data; The cable status is identified using the texture features, and the identification result of the cable status is output.

[0014] In a preferred embodiment of the intelligent cable fault analysis method of the present invention, the step of obtaining the target area includes: The first model is used to extract deep features from the target image data, and the predicted segmentation result is output based on the deep features; An optimization objective function is constructed based on the predicted segmentation results. The first model is then optimized based on the optimization objective function to obtain the second model. The second model is used to extract deep features from the target image data, and a segmentation mask is generated based on the deep features; The target region is defined based on the segmentation mask.

[0015] In a preferred embodiment of the intelligent cable fault analysis method of the present invention, the step of extracting the texture features includes: The grayscale values ​​of pixels within the target area are statistically analyzed to construct a grayscale co-occurrence matrix; Calculate the contrast eigenvalues, energy values, and entropy values ​​of the gray-level co-occurrence matrix, and use these values ​​as texture features. Calculate the edge intensity value of all pixels within the target area, and use the edge intensity value as the edge intensity feature.

[0016] In a preferred embodiment of the intelligent cable fault analysis method of the present invention, the step of identifying the cable condition includes: Texture features are fused to obtain deep feature values; The deep feature values ​​are nonlinearly activated to obtain deep fusion features; The predicted probability of cable status belonging to each category is calculated using deep fusion features. The category corresponding to the highest predicted probability is selected as the preliminary identification result. The texture features are compared with a preset normal threshold range; If the category corresponding to the preliminary identification result is consistent with the comparison conclusion, the preliminary identification result is output as the final identification result.

[0017] In a preferred embodiment of the intelligent cable fault analysis method of the present invention, the step of obtaining the target image data includes: The cross-sectional image of the cable is converted to grayscale to obtain grayscale image data; The grayscale image data is filtered to obtain standard image data; The target image data is obtained by optimizing the standard image data.

[0018] In a preferred embodiment of the intelligent cable fault analysis method of the present invention, the step of obtaining the grayscale image data includes: The weighting coefficients for red, green, and blue pixel values ​​are set based on the color sensitivity of the human eye; The grayscale value of a pixel is calculated using weighted coefficients for red, green, and blue pixel values, and the grayscale value of the pixel is used as grayscale image data.

[0019] The beneficial effects of this invention are as follows: This invention is easy to carry and use, convenient to move to the work scene, and can absorb and filter dust during the cable dissection process to avoid environmental pollution. The identification mechanism can analyze the cable fault type, improve the efficiency of analysis and the accuracy of results, and make the analysis results more standardized. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0021] Figure 1 A schematic diagram of the planar structure of the intelligent cable fault analysis platform is shown.

[0022] Figure 2 A three-dimensional structural diagram of the intelligent cable fault analysis platform is shown.

[0023] Figure 3 A schematic diagram of the overall process of intelligent cable fault analysis method is shown.

[0024] In the diagram: 1. Platform; 11. Vacuuming platform; 111. Dust hopper; 12. Casters; 13. Support base; 14. Protective cover; 2. Multi-axis moving platform; 3. Cutting mechanism; 4. Fixing mechanism; 5. Identification mechanism; 6. Vacuuming mechanism; 61. Exhaust fan; 62. Air duct; 63. Dust filter; 631. Filter nozzle; 7. Remote control. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0026] The terminology used in this invention is that which is currently widely used in the art in consideration of the function of the invention; however, these terms may vary according to the intent of those skilled in the art, precedent, or new technology in the art. Furthermore, specific terms may be chosen by the applicant, and in such cases, their detailed meanings will be described in the detailed description of the invention. Therefore, the terms used in this specification should not be construed as simple names, but rather based on their meanings and the overall description of the invention.

[0027] Reference Figure 1 This embodiment provides an intelligent cable fault analysis platform, including a platform 1 and a multi-axis moving platform 2, which is mounted on the platform 1. The cutting mechanism 3 is located at the moving end of the multi-axis moving platform 2; Among them, the multi-axis moving platform 2 is at least a three-axis moving platform, including the X-axis, Y-axis and Z-axis. The multi-axis moving platform 2 can adjust the position of the cutting mechanism 3 so as to adjust the cutting part of the cable. The cutting mechanism 3 is a circular saw cutting machine.

[0028] The fixing mechanism 4 is located on the platform 1 and is used to fix the cable being analyzed; wherein, the fixing mechanism 4 is a bench vise structure, which can fix the position of the cable.

[0029] Preferably, the fixing mechanism 4 is provided in two sets, both of which are fixed on the platform 1, for fixing the two ends of the cable under test.

[0030] The identification mechanism 5, which is located on the platform 1, is used to identify and analyze cable cuts and determine the type of cable fault based on the analysis results.

[0031] It also includes a dust collection mechanism 6, which includes a blower 61 mounted on the platform 1; an air duct 62 located at the air inlet of the blower 61; and a dust filter 63 located at the exhaust end of the blower 61. The dust filter element 63 has a filter nozzle 631 at its end for secondary filtration.

[0032] The platform 1 is equipped with a dust collection platform 11, and the dust collection platform 11 is equipped with a dust hopper 111. The end of the dust hopper 111 is connected to the air duct 62.

[0033] The dust collection table 11 is a grid-like structure with several openings, and the dust hopper 111 is located at the bottom of the dust collection table 11 to collect particles and waste generated during cutting.

[0034] The platform 1 is equipped with casters 12; the platform 1 is equipped with a support base 13, wherein the support base 13 has a telescopic function.

[0035] The platform 1 is equipped with a protective cover 14, which is made of a transparent material, such as acrylic or glass. The protective cover 14 can reduce dust and block the generated dust.

[0036] It also includes a remote controller 7, which is used to control the multi-axis moving platform 2 to adjust the position of the cutting mechanism 3.

[0037] During use, the platform 1 is moved by the caster 12. After moving to the work area, the support base 13 is used to fix the position of the platform 1 on the ground to improve the stability during operation. Then, the cable is fixed above the platform 1 by the fixing mechanism 4. The position of the cutting mechanism 3 is moved by the multi-axis moving platform 2. After the cutting mechanism 3 is moved to the vicinity of the cutting area, the cutting mechanism 3 starts to work and cuts the cable. The identification mechanism 5 is used to collect cable cross-section image data and analyze the cable fault type based on the cross-section image data.

[0038] During the cutting process of the cutting mechanism 3, the induced draft fan 61 is started and the air duct 62 is used to draw air. The dust particles and waste generated by cutting enter the dust hopper 111 through the dust collection table 11. After being collected in the dust hopper 111, they are transported to the dust filter 63. The dust filter 63 has a honeycomb filter structure to adsorb particles. The filter nozzle 631 at the end of the dust filter 63 performs secondary filtration, so that the dust particles and waste in the air passing through the dust filter 63 are blocked, avoiding the impact on the environment.

[0039] In summary, this invention is easy to carry and use, convenient to move to the work scene, and can absorb and filter dust during cable dissection to avoid environmental pollution. The identification mechanism 5 can analyze the cable fault type, improving the efficiency and accuracy of the analysis and making the analysis results more standardized.

[0040] Reference Figure 3 As an optional embodiment, this embodiment provides a cable fault intelligent analysis method for identification mechanism 5 to identify and analyze cable cuts. The cable fault intelligent analysis method includes the following steps: S1: Acquire cable cross-section image data; In one optional embodiment, step S1 specifically performs the following operations: First, the cable to be tested is fixed on the fixing mechanism 4 and cut using the cutting mechanism 3 to obtain the cable cross-section. Then, the cable is removed and placed horizontally with the cross-section facing upwards, and the plane of the cross-section is perpendicular to the identification mechanism 5. In this embodiment, the identification mechanism 5 is a high-definition industrial camera with a resolution of 1920×1080. The optical axis of the camera lens is perpendicular to the cross-section to ensure that the cable cross-section is completely presented in the center of the captured image.

[0041] Before data acquisition, the camera's shooting parameters were uniformly set: exposure time was fixed at 100ms, focal length at 50mm, and white balance mode was set to manual white balance with the parameter values ​​locked. Simultaneously, based on the actual ambient light conditions, the camera's physical aperture and the brightness of the supplementary light were adjusted appropriately to ensure moderate brightness in the cable cut area, avoiding overexposure or underexposure. All shooting parameters remained unchanged during subsequent continuous data acquisition after being set once.

[0042] After setting the parameters, the camera captures a color image of the cable cross-section, resulting in a 1920×1080 resolution RGB three-channel cable cross-section image. In the initial image, the cable cross-section edges are clear and complete, without obvious shadows, reflections, or external debris obstructing the view. The conductor, insulation layer, and shielding layer structures inside the cross-section are clearly identifiable.

[0043] By fixing the relative position of the camera and the cable cross-section and setting uniform shooting parameters, the consistency of imaging conditions is ensured every time an image is captured.

[0044] S2: Preprocess the cable cross-section image to obtain the target image data; It should be noted that the cable cross-section image acquired in step S1 is an RGB three-channel color image, with each pixel containing pixel values ​​for the red, green, and blue channels, and each channel having a value range of 0 to 255.

[0045] The steps for obtaining the target image data include: S21: Perform grayscale processing on the cable cross-section image to obtain grayscale image data; In this embodiment, converting a color image to a grayscale image can reduce the amount of data processing and eliminate color redundancy.

[0046] The steps to obtain grayscale image data include: S211: Set weighting coefficients for red, green, and blue pixel values ​​based on the color sensitivity of the human eye; It should be noted that cone cells in the human retina are most sensitive to green light, followed by red light, and least sensitive to blue light. Based on these biological characteristics, this embodiment uses a weighted average method for grayscale processing, setting weighting coefficients for the red, green, and blue channels that match the aforementioned sensitivity characteristics of the human eye.

[0047] For example, the weighting coefficient of the red channel is set to 0.299, the weighting coefficient of the green channel is set to 0.587, and the weighting coefficient of the blue channel is set to 0.114. The weighting coefficients of the red channel, the green channel, and the blue channel meet the normalization conditions, which can ensure that the range of grayscale values ​​is consistent with the range of original pixel values.

[0048] It should be noted that the weighting coefficients for the red channel, green channel, and blue channel are fixed values ​​that are set once during system initialization and are directly used in subsequent processing of all cable cross-section images without needing to be set again.

[0049] S212: Calculate the gray value of a pixel using weighted coefficients of red, green, and blue pixel values, and use the gray value of the pixel as grayscale image data; After determining the weighting coefficients, the grayscale value of the cable cross-section image is calculated pixel by pixel. For any pixel with coordinates (x,y) in the image, its red channel pixel value R(x,y), green channel pixel value G(x,y), and blue channel pixel value B(x,y) are obtained, and substituted into the following weighted summation formula: In the formula, This represents the grayscale value of the pixel, which is an integer in the range of 0 to 255, where 0 represents pure black and 255 represents pure white.

[0050] Taking a pixel in a cable cross-section image as an example, its coordinates are (500, 300), and its corresponding three-channel pixel values ​​are R(500, 300) = 120, G(500, 300) = 150, and B(500, 300) = 90. Substituting into the formula, the calculation is as follows: After rounding, the grayscale value of this pixel is 134, which is presented as a mid-gray tone.

[0051] After performing the above calculations on every pixel of the initial image, a single-channel grayscale image with the same size as the original image can be obtained. In this grayscale image, the conductor area of ​​the cable cross-section exhibits a brighter grayscale value due to its dense material and strong reflectivity, the insulation layer area exhibits a medium grayscale value, while the background and fixture areas exhibit different grayscale distributions due to different materials and lighting conditions.

[0052] After executing step S21, the image data size is reduced from the original 1920×1080×3 bytes to 1920×1080×1 byte, a reduction of approximately two-thirds, effectively reducing the computational burden. Simultaneously, because the weighting coefficients are set in accordance with the sensitivity characteristics of the human eye, the converted grayscale image retains the image details perceptible to the human eye in terms of brightness levels, avoiding the brightness distortion problem caused by the traditional direct averaging method, and providing a grayscale image foundation that better reflects actual visual perception for subsequent processing.

[0053] S22: Filter the grayscale image data to obtain standard image data; In this embodiment, the preferred filtering method is median filtering. Although the grayscale image data obtained in step S21 has removed color redundancy and reduced the amount of data, there may still be random noise in the image due to factors such as light fluctuations, camera sensor thermal noise, or electromagnetic interference during the acquisition process. The above noise is manifested in the image as isolated abnormal pixel points with significantly different grayscale values ​​from their neighboring pixels. The above noise can be filtered out by median filtering.

[0054] For example, a 3×3 square filter window is selected, that is, the current pixel to be processed is taken as the center, and the surrounding 8 neighboring pixels are taken, plus the current pixel itself, to form a filter neighborhood of 9 pixels. After determining the filter window, each pixel in the grayscale image data is traversed, and the following operations are performed on each pixel: The first step is to extract the grayscale values ​​of all pixels in the 3×3 neighborhood centered on the pixel. For pixels on the image edge and the four corners, their neighborhood will extend beyond the image boundary. In this embodiment, a mirror filling method is used, that is, with the image boundary as the axis, the pixel values ​​inside the boundary are symmetrically copied to the outside of the boundary, thereby completing all the pixel values ​​required for the 3×3 neighborhood.

[0055] The second step is to sort the nine gray values ​​in the 3×3 neighborhood in ascending order to obtain an ordered sequence.

[0056] The third step is to take the gray value of the fifth position in the middle of the ordered sequence as the filtered gray value of the pixel and replace the original gray value of the pixel with this value.

[0057] For example, taking a pixel with coordinates (500, 300) in a grayscale image as an example, the grayscale values ​​of the 9 pixels in its 3×3 neighborhood are [132, 134, 133, 135, 134, 136, 133, 134, 135]. After sorting these values, we get [132, 133, 133, 134, 134, 134, 135, 135, 136], with a midpoint value of 134. Therefore, after filtering, the grayscale value of this pixel remains 134, indicating that the grayscale distribution in this area is uniform and there are no abnormal noise points.

[0058] If a noise point is introduced into the neighborhood, for example, if the noise point has a grayscale value of 200, while the grayscale values ​​of the remaining 8 normal pixels remain as described above, then the 9 grayscale values ​​are [132, 134, 133, 135, 134, 136, 133, 134, 200]. After sorting, we get [132, 133, 133, 134, 134, 134, 135, 136, 200], with the median value still being 134. The noise point with a grayscale value of 200 is placed at the end of the sequence and is naturally discarded when taking the median value. The filtered pixel grayscale value is still 134, and the noise is effectively removed.

[0059] It should be noted that, compared to mean filtering, median filtering can fundamentally avoid the interference of extreme noise values ​​on the filtering results.

[0060] After filtering, the median replacement operation is performed on each pixel of the grayscale image, resulting in the denoised standard image data. The size of this standard image data remains unchanged at 1920×1080.

[0061] After performing step S22, random noise in the image is effectively suppressed. At the same time, the edge boundaries between the various structures inside the cable cross section, as well as the boundaries between the defect area and the normal area, are completely preserved without blurring or shifting.

[0062] S23: Obtain the target image data by optimizing the standard image data; In this embodiment, the optimization process involves histogram equalization of the standard image data. Although the standard image data obtained in step S22 has filtered out random noise, the grayscale distribution of the image may still have the following problems due to the different materials and reflective properties of different areas of the cable cross-section, as well as the limitations of the ambient lighting conditions: grayscale values ​​in some areas are concentrated in a narrow range, resulting in an overall image that is either too dark or too bright; and the fine structures of the cable cross-section, such as subtle texture changes inside the insulation layer, tiny ablation spots, and early discharge traces, do not show significant grayscale differences from the surrounding normal areas, making it difficult to discern details. However, by using histogram equalization, the grayscale distribution range of the image is stretched, enhancing the overall contrast and highlighting the detailed features of the cable cross-section.

[0063] For example, histogram equalization processing includes the following steps: First, iterate through all pixels of the standard image data and count the number of pixels appearing at each gray level. Let the gray level be... The corresponding number of pixels is ,in At the same time, calculate the total number of pixels in the standard image data. ,in M represents the number of pixels in the horizontal direction of the image, M=1920. N is the number of pixels in the vertical direction of the image, N=1080, and the total number of pixels. .

[0064] Secondly, starting from gray level 0, the number of pixels at each gray level is incremented progressively to calculate the cumulative pixel count for each gray level. That is, the grayscale value ranges from 0 to the current grayscale level. The sum of the number of all pixels. Then, the cumulative probability density of each gray level is calculated based on the cumulative number of pixels, which serves as the basis for calculating the corresponding mapped gray value of that gray level after equalization processing.

[0065] Furthermore, grayscale value remapping is performed on each pixel in the standard image data. For a pixel with coordinates (x, y), its original grayscale value is... equalized grayscale values Calculate using the following formula: In the formula, L represents the total number of gray levels, L=256. This is the maximum grayscale value; This represents the total number of pixels in the image. Gray values ​​range from 0 to the original gray value of the current pixel. The cumulative number of pixels.

[0066] Taking a certain pixel as an example, its original grayscale value is 134, and the corresponding cumulative pixel count is... That is, pixels with gray values ​​between 0 and 134 account for half of the total number of pixels in the image. Substitute into the formula to calculate: After rounding, the equalized grayscale value of this pixel is 128.

[0067] After performing the above remapping calculation on each pixel, the target image data after histogram equalization is obtained. The size of this target image data remains 1920×1080, and the grayscale value range remains 0~255.

[0068] After executing step S23, the grayscale distribution of the target image data, which was originally concentrated in a narrow range, is stretched to a uniform distribution across the entire range of 0 to 255. Adjacent areas in the cable cross-section where grayscale differences were not initially significant, such as dense areas and heat-aged areas within the insulation layer, and normal oxide layers and partial discharge erosion points on the copper core surface, have their grayscale differences widened after equalization. This makes subtle textures and boundaries that were previously difficult to distinguish with the naked eye clearer. Simultaneously, the grayscale jump amplitude at the boundary between defective and normal areas increases.

[0069] S3: Perform image segmentation on the target image data to obtain the target region; Although the target image data output in step S2 already has the characteristics of high contrast and low noise, the image still contains a complete background area and various structural areas inside the cable cross-section. By segmenting the image, the target area of ​​the cable cross-section can be separated from the background, and the normal structural areas and suspected defect areas inside the cable cross-section can be distinguished.

[0070] The steps to obtain the target region include: S31: Use the first model to extract deep features of the target image data, and output the predicted segmentation result based on the deep features; In this embodiment, the first model is the U-Net semantic segmentation model. The target image data obtained in step S23 is used as the input of the first model. The target image data contains the complete cable cross-section structure and some background areas, such as the fixing mechanism 4 and the cover 14. The U-Net semantic segmentation model adopts an encoder and decoder structure. The encoder is responsible for extracting deep features from the input image layer by layer, from edge texture to structural semantics. The decoder is responsible for restoring these deep features to the original image size and outputting pixel-level predicted segmentation results.

[0071] In the encoder, the input target image data undergoes layer-by-layer convolution and pooling operations, gradually compressing the image's spatial dimensions while simultaneously increasing the number of feature channels. The shallow layers extract low-level information such as edge orientation, texture direction, and local grayscale gradients in different regions of the cable cross-section. As the layers deepen, the extracted features gradually evolve into higher-level semantic information, including the conductor's circular cross-sectional outline, the annular interlocking structure of the insulation layer, the closed irregular boundaries of the ablation region, and the radial textures around the discharge vents. At the deepest layer of the encoder, the output is a deep feature map with a spatial dimension compressed to 120×68 but containing 512 channels. This feature map condenses the multi-level feature representation of the entire image, from local details to overall structure.

[0072] In the decoder, deep feature maps are progressively upsampled layer by layer to restore spatial dimensions. At each layer, shallow feature maps of the corresponding scale from the encoder are stitched together via skip connections. This process preserves the semantic judgment capabilities of the deep layers while incorporating the spatial boundary information of the shallow layers. The decoder ultimately outputs a predicted segmentation result with a size of 1920×1080, which is exactly the same as the original input image.

[0073] The specific form of the predicted segmentation result is as follows: for each pixel in the input target image data, the model outputs the judgment conclusion of the category to which the pixel belongs. There are three categories: background, cable cross-section target area, and suspected defect area.

[0074] For example, the predicted segmentation results are presented in the form of a visual segmentation mask: the first color, such as black, represents the pixel area determined to be the background, corresponding to the fixed mechanism 4, cover 14, platform 1 and other parts in the image that are unrelated to the cable cross section itself; the second color, such as white, represents the pixel area determined to be the target area of ​​the cable cross section, corresponding to the normal structural parts of the cable cross section such as the conductor copper core, insulation layer, shielding layer, armor layer and so on; the third color, such as gray, represents the pixel area determined to be the suspected defect area, corresponding to the abnormal structural parts of the cable cross section such as ablation pits, discharge pores, carbon traces and so on that may exist.

[0075] Through the pixel-level division of the three categories of black, white, and gray, the predicted segmentation result output in step S31 achieves the initial separation of the target area and suspected defect area of ​​the cable cross-section from the background area.

[0076] S32: Construct an optimization objective function based on the predicted segmentation results, optimize the first model based on the optimization objective function, and obtain the second model; In this embodiment, step S32 is the process of training and optimizing the first model so that it gradually acquires the ability to segment cable cross-section images from its initial random parameter state.

[0077] It should be noted that in step S31, the first model performs forward propagation on the input target image data and outputs the predicted segmentation result. However, in the initial stage of model training, the convolutional kernel weights and bias parameters inside the first model are randomly initialized. At this time, the model does not have the ability to accurately identify the categories of each region of the cable cross-section, and its output predicted segmentation result has a significant deviation from the actual cable cross-section structure. This can manifest as: misclassifying large background areas as cable cross-section target areas, misclassifying normal structural areas of the cable cross-section as suspected defect areas, or omitting real ablation pits and discharge pores, etc.

[0078] Preferably, in order to measure the deviation between the segmentation results predicted by the first model and the actual segmentation results, this embodiment uses the cross-entropy loss function as the optimization objective function.

[0079] The cross-entropy loss function uses a single pixel as the basic unit of calculation. It calculates the loss value of each pixel by comparing the difference between the predicted result and the true label. Then, it sums the loss values ​​of all pixels in the entire image to obtain the overall loss value of the whole image.

[0080] Specifically, from the predicted segmentation results output in step S31, for each pixel, the predicted probability value of the first model for that pixel belonging to a certain target category, such as belonging to the cable cut target area rather than the background, can be obtained. ,in, The value range is 0 to 1.

[0081] Meanwhile, during the preparation of training data, technicians manually labeled each training target image pixel by pixel using a manual annotation tool. In binary classification scenarios, such as determining whether a pixel belongs to the target region of a cable cut, if the pixel does indeed belong to the target region, then the true label is... If it does not belong to, then .

[0082] Based on predicted probability and real labels The cross-entropy loss value for a single pixel is calculated using the following formula: The total loss value of the entire image is obtained by summing the loss values ​​of all pixels in the image. : When the predictions of the first model match the true labels, for example and When it approaches 1, The loss value is close to 0. A value close to 0 indicates a small penalty for that pixel; this is useful when the first model's prediction significantly deviates from the true label, for example... and When approaching 0, Approaching negative infinity, loss value Increasing the value indicates that a penalty is imposed for incorrect predictions of that pixel.

[0083] After constructing the cross-entropy loss function as the optimization objective function, the first model is iteratively optimized using this objective function, specifically by employing the gradient descent algorithm. The optimization process is as follows.

[0084] First, initialize all learnable parameters of the first model. The kernel weights of each convolutional layer in the model are assigned values ​​using a random initialization method, and the bias parameters of each convolutional layer are uniformly initialized to 0.

[0085] Then, the target image data used for training is input into the first model, and the predicted segmentation result is output according to the forward propagation process in step S31. The predicted segmentation result is compared pixel by pixel with the corresponding manually labeled real segmentation mask, and the cross-entropy loss function formula is substituted to calculate the overall loss value under the current model parameter state. In the initial training phase, because the model parameters are random values, there are numerous missegments and missed segments in the predicted segmentation results, leading to a high loss value. It is at a relatively high level.

[0086] Next, using the backpropagation algorithm, from the loss value Starting from the beginning, the model's computational graph is reversed layer by layer to calculate the gradient of the loss value with respect to each learnable parameter, i.e., the weights and biases of each convolutional kernel. Based on the calculated gradients of each parameter, the parameters are updated according to the gradient descent rule: each parameter is adjusted a small step in the opposite direction of its gradient, with the step size controlled by a preset learning rate. The updated model parameters make the loss value... Moving in the downward direction reduces the deviation between the model's predicted segmentation result and the actual segmentation mask.

[0087] In summary, as the number of training rounds increases, the loss value... The loss rate decreases rapidly in the early stages of training, as the model quickly learns the approximate boundary between the target region and the background of the cable cross-section. In the middle stages of training, the rate of decrease slows, and the model begins to fine-tune the segmentation boundaries. In the later stages of training, the loss value... It drops to a lower level and tends to converge, no longer decreasing significantly.

[0088] When the loss value When convergence reaches a stable state, it indicates that the model's predicted segmentation results are highly consistent with the manually labeled actual segmentation mask. The background region, cable cross-section, and suspected defect area can be accurately marked without significant missegmentation or omissions. At this point, the parameters of the first model have been optimized from initial random values ​​to a set of parameter configurations capable of accurately performing the cable cross-section image segmentation task. This parameter configuration is fixed and saved, becoming the second model in step S32.

[0089] It should be noted that the second model is structurally identical to the first model, both being U-Net semantic segmentation models. The only difference lies in the parameter values ​​within the models. The first model uses randomly initialized parameters and lacks effective segmentation capabilities; the second model uses parameters that have been iteratively optimized using the cross-entropy loss function and gradient descent algorithm, enabling it to accurately output predicted segmentation results from the input target image data, which will be used to generate segmentation masks in subsequent steps.

[0090] S33: Use the second model to extract deep features from the target image data, and generate a segmentation mask based on the deep features; The target image data obtained in step S23 is used as the input of the second model. The target image data is a single-channel grayscale image with a size of 1920×1080 pixels. The image contains complete structural information of the cable cross-section, as well as possible defect areas such as ablation pits, discharge pores, and carbon traces. It also contains some background areas. Every texture, every edge, and every grayscale change area in the image will be analyzed and judged pixel by pixel by the second model.

[0091] After the target image data is input into the second model, its encoder first performs deep feature extraction. The encoder compresses the input image from its original 1920×1080 spatial size through layer-by-layer convolution and pooling operations, while simultaneously increasing the number of feature channels from a single channel to 512 channels. In the shallow layers, the encoder extracts the most basic low-level features of the image, such as the edge orientation at the junction of different materials in a cable cross-section, the texture direction within each region, and the local grayscale gradient. As the layers deepen, the extracted features gradually sublimate into structural features such as the circular cross-sectional outline of the conductor, the ring-shaped interlocking structure of the insulation layer, and the complete cross-sectional layout of multiple materials, as well as defect-level features such as the irregular closed boundaries of ablation areas, the material melting texture around discharge pores, and the dendritic branching morphology of carbon traces. Finally, the encoder outputs a deep feature map with a spatial size compressed to 120×68 but containing 512 channels. This feature map contains complete semantic information about the structure and anomalies in the entire target image data.

[0092] The deep feature map output by the encoder then enters the decoder of the second model. The decoder gradually restores the spatial size from 120×68 to 1920×1080 through layer-by-layer upsampling; the decoder finally outputs a segmentation mask with the exact same size as the original input image. This segmentation mask is a pixel-level classification label map, where the value of each pixel represents the category to which that pixel belongs.

[0093] S34: Define the target region based on the segmentation mask.

[0094] In this embodiment, the segmentation mask contains three categories of markers: the first category marker, for example, a pixel value of 0, is displayed as black, corresponding to the background area; the second category marker, for example, a pixel value of 1, is displayed as white, corresponding to the target area of ​​the cable cross section; and the third category marker, for example, a pixel value of 2, is displayed as gray, corresponding to the suspected defect area.

[0095] Through the above three types of pixel-level markings, the segmentation mask generated in step S33 achieves complete classification of the target image data, and the background area is removed, the normal structure of the cable cross-section target area is preserved, and the suspected defect areas are marked.

[0096] S4: Extract texture features of the target region from the target image data; The steps for extracting texture features include: S41: Statistically analyze the pixel grayscale values ​​within the target area and construct a grayscale co-occurrence matrix; It should be noted that the target area defined by the segmentation mask is used to clarify which pixels in the target image data belong to the cable cross-section target area and the suspected defect area, and which pixels belong to the background.

[0097] By statistically analyzing pixel pairs in an image that satisfy specific distance and orientation conditions, texture characteristics can be quantized into computable matrix data from the frequency of occurrence of combinations of grayscale values ​​in the image.

[0098] Before constructing the gray-level co-occurrence matrix, two key construction parameters are first set: the distance between pixel pairs. and direction In this embodiment, distance Setting it to 1 means counting adjacent pixel pairs, where the interval between two pixels is 1 pixel, and the direction... Setting it to 0°, i.e., the horizontal right direction, counts pixel pairs formed by the current pixel and its horizontal right-side adjacent pixel. The horizontal direction was chosen because in the cable cross-section image, each layer of the structure unfolds horizontally in the radial cross-section, and the texture changes in the horizontal direction best reflect the differences in the characteristics of each layer of material and the texture anomalies caused by defects.

[0099] Based on the set parameters, the grayscale values ​​of pixels within the target area are statistically analyzed. First, the grayscale value of each pixel within the target area is obtained from the target image data, with the grayscale value ranging from 0 to 255. Then, each pixel within the target area is traversed, excluding pixels located at the rightmost boundary of the target area that cannot find a horizontally adjacent pixel to the right. For each pixel, its own grayscale value is then calculated. grayscale value of the pixel to its right horizontally adjacent pixel This forms a grayscale value pair. .

[0100] Count all possible combinations of grayscale values The number of times it appears within the target area. Since the grayscale value ranges from 0 to 255, there are 256 grayscale levels, therefore the grayscale value pairs... There are 256 × 256 = 65536 possible combinations. After recording the occurrence count of each combination, construct a 256 x 256 matrix. The nth row of the matrix is... Line 1 The element values ​​of the column are grayscale pairs. The number of times it appears in the target area.

[0101] Based on this, the occurrence matrix is ​​transformed into a probability matrix: each element value in the matrix is ​​divided by the total number of all counted pixel pairs within the target area to obtain each grayscale value combination. probability of occurrence This probability matrix is ​​the gray-level co-occurrence matrix.

[0102] Once constructed, the gray-level co-occurrence matrix (GLCM) intuitively reflects the texture characteristics of the target region. For a normal cable cross-section, the material texture is uniform, and the gray-level values ​​of adjacent pixels are mostly similar. For example, in the insulation layer region, the gray-level values ​​of adjacent pixels are all between 130 and 140. Therefore, in the GLCM, elements with higher probability values ​​are concentrated near the diagonal of the matrix. and The positions where the values ​​are close, and and Locations with significant differences have very low probability values. In defective regions, due to ablation or discharge causing surface unevenness, adjacent pixels exhibit large differences in grayscale values. For example, a pixel may have a very low grayscale value within a pit, while an adjacent pixel may have a very high grayscale value at the edge of a protrusion. The probability distribution in the grayscale co-occurrence matrix is ​​relatively dispersed, with non-negligible probability values ​​even at locations far from the diagonal. By analyzing the differences in the matrix distribution, it is possible to distinguish between normal and defective regions.

[0103] S42: Calculate the contrast feature value, energy value, and entropy value of the gray-level co-occurrence matrix, calculate the edge intensity value of all pixels in the target area, and use the contrast feature value, energy value, entropy value, and edge intensity value as texture features; Step S41 has constructed a gray-level co-occurrence matrix based on the statistical results of pixel gray-level values ​​within the target area, and obtained the probability of occurrence of each gray-level value combination. Step S42, based on this, further calculates three quantized texture metrics from the gray-level co-occurrence matrix: contrast, energy, and entropy. At the same time, it calculates the edge intensity values ​​of all pixels in the target area, and finally uses these four values ​​together as texture features.

[0104] Contrast ratio measures the difference in grayscale values ​​between adjacent pixels within a target area, reflecting the clarity of the texture and the depth of the grooves. A higher contrast ratio indicates a more pronounced difference in grayscale values ​​between adjacent pixels, resulting in a coarser and clearer texture.

[0105] Contrast ratio is calculated using the following formula: In the formula, and These are the grayscale values ​​of two adjacent pixels, It is the square of the difference between two grayscale values. It is the probability of the gray value combination appearing in the gray co-occurrence matrix obtained in step S41.

[0106] By traversing all possible combinations of gray values ​​in the gray-level co-occurrence matrix For each combination, calculate the square of its grayscale value difference. Multiply by the probability of that combination occurring. Finally, sum the products of all combinations.

[0107] when A larger value indicates a large difference in grayscale between the pixel pair. If such combinations of large differences occur frequently, When it is large, then The product of these will be very large, and the contrast value obtained by summing them up will also increase accordingly.

[0108] Taking a normal cable cross-section area as an example: the material in this area is uniform in texture, and the grayscale values ​​of adjacent pixels are mostly similar, such as (134, 134), (134, 135), etc. In these combinations... Extremely small, generally 0 or 1, even their The contrast is relatively high, but the product is still very small, so the contrast value in the normal area is low.

[0109] Taking the ablation defect area as an example: the surface in this area is uneven, with very low gray values ​​at the pits and very high gray values ​​at the raised edges. Combinations with large differences in gray values ​​between adjacent pixels frequently occur, such as (100, 200). Among these combinations... Maximum, multiplied by a non-negligible probability Afterwards, the product contribution is significant, so the contrast value of the defective area is significantly increased, much higher than that of the normal area.

[0110] In this embodiment, energy, also known as the second moment of the angle, is used to measure the uniformity and orderliness of the texture within the target region. The larger the energy value, the more uniform the gray-level distribution and the more ordered the texture in the image; the smaller the energy value, the more dispersed the gray-level distribution and the more chaotic the texture.

[0111] Energy is calculated using the following formula: In the formula, Also derived from the gray-level co-occurrence matrix in step S41. By iterating through all probability values ​​in the gray-level co-occurrence matrix, for each... Perform the square operation, and then sum all the squared values.

[0112] because Since the base value is between 0 and 1, squaring will further amplify the difference between high-probability and low-probability values. If certain grayscale combinations... Larger values, when squared and summed, will become the main contributors to energy, resulting in a larger final energy value. However, if the grayscale distribution is highly dispersed, with no single combination clearly dominant, all... Since they are all relatively small and close, summing them after squaring results in a smaller final energy value.

[0113] Taking a normal cable cross-section area as an example: the grayscale distribution is highly concentrated, such as P(134,134)=0.3, P(134,135)=0.2, P(135,135)=0.2. Substituting into the formula for calculation: The high energy value indicates that the texture of the normal area is uniform and orderly.

[0114] Taking a defective area as an example: the probability distribution of each grayscale combination is relatively even and dispersed. For example, there are 20 grayscale combinations, each with a probability of approximately 0.05, and no single combination is dominant. Substituting into the formula for calculation: The energy value is significantly lower than that of the normal area, reflecting that the texture of the defective area is disordered and lacks uniformity.

[0115] It should be noted that entropy is used to measure the degree of disorder in texture within a target area. Based on the concept of entropy in information theory, it reflects the uncertainty and randomness of gray-level distribution. The larger the entropy value, the more disordered the gray-level distribution; the smaller the entropy value, the more regular the gray-level distribution.

[0116] Entropy is calculated using the following formula: In the formula, These are the probability values ​​in the gray-level co-occurrence matrix. It is the natural logarithm. When At that time, it was stipulated This is to avoid the problem of the logarithmic function being undefined at 0.

[0117] Calculate the probability for each grayscale combination. and The product of , then sum the products of all combinations. Since Between 0 and 1, Since the product is negative, the summation is negative to ensure that the entropy value is positive.

[0118] Taking a normal cable cross-section area as an example: the number of high-probability grayscale combinations is limited, such as P(134,134)=0.3, P(134,135)=0.2, etc. The contributions of each element are calculated separately: , After summing all the values, the entropy value of the normal region is approximately 1.38, which is at a low level and reflects the ordered texture.

[0119] Taking the defect area as an example: the probability distribution of grayscale combinations is uniformly dispersed, with each of the 20 combinations having a probability of approximately 0.05. The individual contribution is... The entropy value after summing 20 terms is approximately 3.0. The entropy value is significantly higher than that of the normal area, reflecting that the gray-scale distribution in the defective area is extremely random and the texture is highly disordered.

[0120] It should be noted that entropy and energy are inverse indicators when characterizing texture properties: normal areas have high energy and low entropy; defective areas have low energy and high entropy. They describe the uniformity and disorder of the same texture from different perspectives, and when used together, they can more comprehensively depict the texture state of the target area. Furthermore, edge intensity is used to measure the severity of gray-level abrupt changes within a target area, i.e., the strength of a clear boundary or contour line in the image. When a cable has discharge or ablation defects, a clear gray-level jump will appear at the boundary between the defect area and the normal area, and the magnitude of the edge intensity value can effectively reflect the clarity of this physical damage boundary.

[0121] This embodiment uses the Sobel operator to calculate the edge intensity value of each pixel. The specific calculation steps are as follows.

[0122] First, for each pixel within the target area, calculate its horizontal grayscale gradient. and the gray-level gradient in the vertical direction The Sobel operator consists of two 3×3 convolution kernels: For example: Sobel operator in the horizontal direction: [-1,0,1] [-2,0,2] [-1,0,1] Vertical Sobel operator: [-1,-2,-1] [ 0, 0, 0] [ 1, 2, 1] For the target area with coordinates as For a given pixel, take its 9 grayscale values ​​within a 3×3 neighborhood, multiply them element-wise with both the horizontal and vertical operators, and then sum them to obtain the pixel's value. and .

[0123] Then, according to and Calculate the edge intensity value of the pixel using the following formula. : This formula fuses the gray-level gradients in both the horizontal and vertical directions. Regardless of the direction in which the gray-level abrupt change occurs, the final result after summing the squares and taking the square root is... The values ​​can comprehensively reflect the edge strength at that pixel, for example, the horizontal boundary causes... Larger, vertical boundary leads to The boundary in the direction of inclination is relatively large, causing both to be relatively large simultaneously.

[0124] After performing the above calculations on each pixel within the target area, take the statistical results of the edge intensity values ​​of all pixels, such as the mean or the maximum value, as the edge intensity feature value of the target area.

[0125] Taking a normal cable cross-section area as an example: the internal structure of this area is continuous, and the grayscale transition is smooth. and The values ​​are all relatively small. Typical values ​​are as follows: , Substitute into the formula to calculate The edge intensity value of normal areas is usually in the range of 5 to 8.

[0126] Taking an ablation defect area as an example: the grayscale value at the edge of the defect pit changes drastically, dropping abruptly from the grayscale value of normal insulating material to the grayscale value inside the pit. For example, the normal value is around 140, while the grayscale value inside the pit is around 60, a drop of 80 grayscale levels. The Sobel operator responds strongly to this jump, with typical values ​​as follows: , Substitute into the formula to calculate The value is much higher than that in the normal area. This significant difference makes edge intensity one of the key features for identifying physical damage defects such as ablation and discharge.

[0127] After the above four calculation steps, step S42 extracts four quantized feature values ​​from the target region: Contrast, Energy, Entropy, and Edge Intensity. These four feature values ​​together constitute the texture features, describing the characteristics of the target region in four dimensions: texture roughness, uniformity, disorder, and boundary sharpness.

[0128] The texture characteristics of a normal cable cross-section area are: low contrast, high energy, low entropy, and low edge strength. The texture characteristics of a defective area are: high contrast, low energy, high entropy, and high edge strength.

[0129] S5: Use texture features to identify the cable status and output the identification result of the cable status.

[0130] The steps for identifying the cable condition include: S51: Fuse texture features to obtain deep feature values; Step S42 has extracted four texture feature values ​​from the target region: contrast, energy, entropy, and edge intensity. These four feature values ​​quantify the texture characteristics of the target region from four different dimensions: texture roughness, uniformity, disorder, and boundary sharpness. However, a single feature value can only reflect one aspect of the target region's characteristics and cannot independently support a comprehensive judgment of the cable's condition. For example, high contrast alone cannot directly determine the presence of ablation defects, as moderate contrast increases may also occur at some normal material interfaces; low energy alone cannot rule out local image noise interference. Therefore, it is necessary to fuse these four independent texture feature values ​​to generate a comprehensive deep feature value, forming an overall expression of the texture characteristics of the target region.

[0131] Before fusion, the four texture feature values ​​are first normalized. Since the typical ranges for contrast are approximately 0.150, energy is 0.1, entropy is approximately 0.5, and edge intensity is approximately 0.360, the dimensions and numerical ranges of these four feature values ​​differ significantly. If fusion is performed directly, features with large numerical ranges, such as edge intensity, will have an overwhelming weight in the fusion result, while the contribution of features with small numerical ranges, such as energy, will be submerged. This would result in the fused deep feature values ​​actually reflecting only information from a single feature, losing the significance of multi-dimensional fusion.

[0132] Preferably, this embodiment uses convolution operations in a convolutional neural network to fuse the four normalized texture feature values. The convolution operation uses a set of learnable weight parameters to weight and combine the input feature values, while adding a bias term for overall offset adjustment, and finally outputs a fused deep feature value.

[0133] Let the four normalized texture feature values ​​be: contrast ,energy ,entropy Edge strength The weight parameters of the convolution kernel are: The bias term is Convolution operations are performed according to the following formula: In the formula, This refers to the deep feature values ​​obtained after fusion.

[0134] Weight parameters The values ​​of these values ​​reflect the relative importance of the corresponding texture features in the cable condition recognition task. It should be noted that the weight parameters are not manually set, but are automatically learned through the gradient descent algorithm. After training, features that contribute significantly to defect recognition are typically assigned larger absolute weight values, while features that contribute relatively less are assigned smaller absolute weight values. Bias term This is used to adjust the overall output baseline, ensuring that when all input feature values ​​are close to 0, the deep feature values ​​may still have reasonable values.

[0135] Taking a target area of ​​a cable cross-section as an example, contrast ,energy ,entropy Edge strength The weight parameters are: Bias term Substitute into the formula to calculate: The deep feature value of 0.3085 integrates texture information from four dimensions, comprehensively reflecting the overall tendency of the target area's texture to be within the normal range.

[0136] Taking a certain ablation defect area as an example, the contrast... ,energy ,entropy Edge strength Substitute the same parameters into the calculation: The deep feature value of 0.475 is significantly higher than that of the normal area of ​​0.3085, reflecting the comprehensive characteristics of the texture abnormality in the defect area.

[0137] Through the above convolution operation, four independent texture feature values ​​are fused into a single deep feature value. This deep feature value not only integrates information from four dimensions—texture roughness, uniformity, disorder, and boundary clarity—but also automatically retains the most valuable components for cable condition identification in each dimension through learnable weight parameters.

[0138] S52: Nonlinear activation is applied to deep feature values ​​to obtain deep fusion features; Step S51 uses convolution operations to fuse four texture feature values—contrast, energy, entropy, and edge intensity—into a single deep feature value. However, this deep eigenvalue Essentially, it remains a linear weighted combination of the input features, and its calculation process involves only multiplication and addition. No matter how many convolutional layers are stacked, the output of the linear combination is always a linear mapping of the input. However, cable state classification is a typical nonlinear problem: the relationship between texture features and classification results is not a simple linear one, but rather involves threshold effects and complex nonlinear correlations.

[0139] For example, increasing the edge strength value from 5 to 10 might only slightly increase the probability of a cable being defective; however, increasing it from 40 to 50 could cause a dramatic increase in the probability. This nonlinear phenomenon of probability jumps near a certain critical point cannot be characterized by linear models. Therefore, a nonlinear activation step needs to be introduced after convolutional fusion to inject nonlinear expressive power into the model, enabling it to fit arbitrarily complex nonlinear mappings between features and classification results.

[0140] This embodiment uses the ReLU (Modified Linear Unit) activation function for nonlinear activation. When the input value is greater than or equal to 0, the output is equal to the input value itself; when the input value is less than 0, the output is set to 0. This can be expressed by the formula: In the formula, The input value is the deep feature value obtained in step S51. ; The output value after nonlinear activation is the deep fusion feature.

[0141] The nonlinearity of ReLU originates from its... The turning point. For all inputs greater than or equal to 0, ReLU maintains linear propagation without modification, ensuring that effective feature information is not attenuated; for all inputs less than 0, ReLU forces its output to 0, making that feature channel completely silent in the current calculation; taking the target area of ​​the normal cable cross-section in step S51 as an example, its deep feature values , which is a positive value. Substituting this into the ReLU activation function: The output value is 0.3085, consistent with the input value. This deep feature value is fully preserved and continues to be passed on.

[0142] Taking the ablation defect region in step S51 as an example, its deep characteristic value Also a positive value. Substituting into the ReLU activation function: The output value is 0.475, which is also fully preserved.

[0143] If the deep feature values ​​obtained after convolutional fusion of the texture features of a target region are negative, for example... This indicates that the overall representation of this set of features contradicts the expected direction of the current classification task, and is considered an invalid feature that does not contribute to the classification decision or even interferes with it. In this case, ReLU sets its output to 0. The invalid feature is completely suppressed and will not interfere with the subsequent classification probability calculation. After the nonlinear mapping of the ReLU activation function, the deep feature value output in step S51 is... Transformed into deep fusion features This deep fusion feature has two key properties: First, it retains the positive information of all effective features; the higher the feature strength and the more valuable the signal for classification, the more completely it is passed to the next layer. Second, it completely eliminates invalid negative signals, avoiding interference from irrelevant features in the classification decision. The deep fusion feature will be used as input to the Softmax classifier in step S53 to calculate the predicted probability of the cable state belonging to each category.

[0144] S53: Calculate the predicted probability of the cable state belonging to each category using deep fusion features, and select the category corresponding to the highest predicted probability as the preliminary identification result; In this embodiment, step S53 specifically performs the following operations: Step S52 involves nonlinearly activating the deep feature values ​​using the ReLU activation function to obtain the deep fusion feature. This deep fusion feature is a comprehensive expression of the four texture features—contrast, energy, entropy, and edge intensity—after convolutional fusion and nonlinear mapping, containing all the effective information about the texture characteristics of the target region. However, at this point, the deep fusion feature is still a numerical value and needs to be transformed into a classification criterion with clear physical meaning, namely, the probability of the cable condition belonging to each of the three categories: normal, discharge defect, or ablation defect.

[0145] First, the deep fusion features are fed into a fully connected layer, which maps the deep fusion features into multiple score values. The number of score values ​​equals the total number of cable condition categories. In this embodiment, the cable condition is divided into three categories: normal (category 1), discharge defect (category 2), and ablation defect (category 3). Therefore, the fully connected layer outputs three score values, denoted as follows: , and These correspond to the raw scores of the three categories, respectively.

[0146] The score output by the fully connected layer , , While the scores can reflect the degree of bias of deep fusion features towards different categories, the range and relative differences of the scores do not have an intuitive probabilistic meaning. The scores can be any real numbers, and the sum of the scores for each category is not equal to 1. Therefore, it is necessary to convert the scores into predicted probabilities that take values ​​between 0 and 1 and where the sum of the probabilities of all categories is 1.

[0147] This embodiment uses the Softmax classifier to perform this transformation. Softmax is performed in two steps: First, the exponential function value is calculated for each score value. The exponential function maps any real number to a positive number, while simultaneously amplifying the differences between scores. Categories with higher scores have their exponential values ​​disproportionately amplified, further widening the gap between them and categories with lower scores. The second step involves summing the exponential values ​​of all categories as the denominator and using the exponential value of each category as the numerator to calculate the proportion of each category, which is the predicted probability of that category. The complete formula is: In the formula, This indicates the cable condition belongs to a category. The predicted probability, For category The score, denominator This is the sum of the score index values ​​for all categories, used for normalization.

[0148] After Softmax processing, the output is , , Satisfies: Each probability value is between 0 and 1, and This constitutes a complete probability distribution.

[0149] Taking a target area of ​​a cable cross-section as an example, after convolutional fusion in step S51 and ReLU nonlinear activation in step S52, the deep fusion features are mapped by the fully connected layer into three categories of scores: normal score. Discharge defect score ablation defect score .

[0150] First, calculate the index value for each category score: Then calculate the sum of all category index values, and use this as the normalized denominator: Finally, calculate the predicted probabilities for each category: (Approximately 60.7%) (Approximately 30.2%) (Approximately 9.1%) The sum of the three probability values ​​is 100%, and the predicted probability of the normal category is the highest.

[0151] The category with the highest predicted probability is selected as the initial identification result. After obtaining the predicted probabilities for each category, the category with the highest predicted probability is selected as the initial identification result by comparing the probability values. This can be expressed by the formula: In the formula, This indicates the preliminary identification results. The function is used to find the Category when the maximum value is obtained .

[0152] In the example above, the predicted probability for the normal category is 60.7%, for discharge defects it is 30.2%, and for ablation defects it is 9.1%. The normal category has the highest probability value, therefore the initial identification result is normal.

[0153] S54: Compare the texture features with a preset normal threshold range; In this embodiment, step S54 specifically performs the following operations: Step S53 has calculated the predicted probabilities of the cable condition belonging to each category using the Softmax classifier, and selected the category corresponding to the highest predicted probability as the preliminary identification result. However, this preliminary identification result is entirely based on the statistical learning of texture features by the convolutional neural network. In some cases with blurred boundaries, such as when the texture of normal areas appears slightly rough due to mild aging, or when the texture of defective areas is still in its early stages of development and its features are not yet significant, the probability distributions output by Softmax may be quite similar. Judgments made solely based on probability magnitude carry a certain risk of misjudgment.

[0154] To enhance the reliability of the recognition results, this embodiment introduces a cross-validation mechanism based on physical rules: the original texture feature values ​​extracted in step S42 are compared one by one with the pre-set normal threshold range to verify the preliminary recognition results.

[0155] It should be noted that the normal threshold range is preset by statistically analyzing the distribution range of each texture feature value after performing complete feature extraction steps S1 to S4 on a large number of cable cross-section samples known to be in normal condition.

[0156] For example, in this embodiment, the normal threshold range for each texture feature is as follows: The normal threshold range for contrast is 20-30. In a normal cable cross-section, the insulation layer and conductor area have a uniform texture, and the grayscale values ​​of adjacent pixels have small differences, so the contrast value is stable within this range.

[0157] The normal threshold range for energy is 0.2~0.3. Normal cable cross-sections exhibit uniform and orderly textures, with a few grayscale combinations dominating the grayscale co-occurrence matrix, resulting in higher energy values.

[0158] The normal threshold range for entropy is 1.0 to 1.5. The cross-sectional texture of a normal cable is regular and predictable, with an ordered grayscale distribution and a low entropy value.

[0159] The normal threshold range for edge strength is 5 to 8. A normal cable cross-section has a continuous internal structure, clear but not abrupt boundaries between layers, a gentle grayscale gradient, and a relatively low edge strength value.

[0160] These threshold ranges are pre-stored in the system's decision rule base and are directly invoked when executing step S54.

[0161] S55: If the category corresponding to the preliminary identification result is consistent with the comparison conclusion, the preliminary identification result will be output as the final identification result. In this embodiment, the four texture feature values ​​extracted from the current target region in step S42—contrast, energy, entropy, and edge intensity—are compared one by one with the aforementioned normal threshold range. The comparison judgment rule is: if the current feature value falls within the normal threshold range, the feature is determined to be normal; if the current feature value exceeds the normal threshold range, the feature is determined to be abnormal.

[0162] Taking a target area of ​​a cable cross-section as an example, the preliminary identification result of step S53 is normal. Step S54 compares the feature values ​​of this area: The actual contrast value is 25.3, which falls within the normal threshold range of 20-30 and is therefore considered normal; the actual energy value is 0.26, which falls within the normal threshold range of 0.2-0.3 and is therefore considered normal; the actual entropy value is 1.38, which falls within the normal threshold range of 1.0-1.5 and is therefore considered normal; and the actual edge intensity value is 5.83, which falls within the normal threshold range of 5-8 and is therefore considered normal.

[0163] All four features were judged to be normal, and the comparison conclusion was that the features were consistent and normal.

[0164] Taking another cable cross-section target area as an example, the preliminary identification result of step S53 is also normal, but the feature comparison result is: The actual contrast value is 45.6, which exceeds the normal threshold range of 20~30 and is therefore considered abnormal; the actual energy value is 0.08, which is lower than the normal threshold range of 0.2~0.3 and is therefore considered abnormal; the actual entropy value is 3.2, which exceeds the normal threshold range of 1.0~1.5 and is therefore considered abnormal; the actual edge intensity value is 50, which far exceeds the normal threshold range of 5~8 and is therefore considered abnormal.

[0165] All four features were judged to be abnormal, and the comparison conclusion was that the features were severely abnormal.

[0166] S56: If there is a discrepancy, a secondary recognition process will be triggered.

[0167] It should be noted that the comparison conclusion output in step S54 is not used to directly determine the cable status, but rather serves as the basis for cross-validation with the preliminary identification result in step S55. The consistency between the comparison conclusion and the preliminary identification result will determine whether the final identification result is output directly or a secondary identification process is triggered. This can improve the reliability and robustness of cable status identification.

[0168] Finally, it should be noted that the methods and devices described in detail above are merely embodiments, and those skilled in the art can modify these embodiments in different ways as long as they do not depart from the scope of the present invention.

Claims

1. A cable fault intelligent analysis platform, characterized in that: include, Platform (1); A multi-axis moving platform (2) is mounted on the platform body (1); A cutting mechanism (3) is located at the moving end of the multi-axis moving platform (2); A fixing mechanism (4) is provided on the platform (1) for fixing the cable being analyzed; The identification mechanism (5) is located on the platform (1) and is used to identify and analyze cable cuts.

2. The intelligent cable fault analysis platform according to claim 1, characterized in that: It also includes a vacuuming mechanism (6), which includes, An induced draft fan (61) is mounted on the platform (1); Air duct (62), which is located at the air inlet end of the induced draft fan (61); A dust filter element (63) is provided at the exhaust end of the induced draft fan (61); The platform (1) is provided with a dust collection platform (11), and the dust collection platform (11) is provided with a dust hopper (111). The end of the dust hopper (111) is connected to the air duct (62).

3. The intelligent cable fault analysis platform according to claim 1 or 2, characterized in that: The platform (1) is equipped with casters (12); The platform (1) is provided with a support base (13).

4. The intelligent cable fault analysis platform according to claim 1 or 2, characterized in that: The platform (1) is provided with a protective cover (14).

5. A method for intelligent analysis of cable faults, characterized in that: The method for identifying and analyzing cable cuts using the identification mechanism (5) as described in any one of claims 1 to 4 includes the following steps: Acquire cross-sectional image data of the cable; Preprocess the cable cross-section image to obtain the target image data; The target image data is segmented to obtain the target region; Extract texture features of the target region from the target image data; The cable status is identified using the texture features, and the identification result of the cable status is output.

6. The intelligent cable fault analysis method according to claim 5, characterized in that: The steps to obtain the target region include: The first model is used to extract deep features from the target image data, and the predicted segmentation result is output based on the deep features; An optimization objective function is constructed based on the predicted segmentation results. The first model is then optimized based on the optimization objective function to obtain the second model. The second model is used to extract deep features from the target image data, and a segmentation mask is generated based on the deep features; The target region is defined based on the segmentation mask.

7. The intelligent cable fault analysis method according to claim 6, characterized in that: The steps for extracting the texture features include: The grayscale values ​​of pixels within the target area are statistically analyzed to construct a grayscale co-occurrence matrix; Calculate the contrast feature value, energy value, and entropy value of the gray-level co-occurrence matrix, calculate the edge intensity value of all pixels in the target area, and use the contrast feature value, energy value, entropy value, and edge intensity value as texture features.

8. The intelligent cable fault analysis method according to claim 7, characterized in that: The steps for identifying the cable condition include: Texture features are fused to obtain deep feature values; The deep feature values ​​are nonlinearly activated to obtain deep fusion features; The predicted probability of cable status belonging to each category is calculated using deep fusion features. The category corresponding to the highest predicted probability is selected as the preliminary identification result. The texture features are compared with a preset normal threshold range; If the category corresponding to the preliminary identification result is consistent with the comparison conclusion, the preliminary identification result is output as the final identification result.

9. The intelligent cable fault analysis method according to claim 8, characterized in that: The steps for obtaining the target image data include: The cross-sectional image of the cable is converted to grayscale to obtain grayscale image data; The grayscale image data is filtered to obtain standard image data; The target image data is obtained by optimizing the standard image data.

10. The intelligent cable fault analysis method according to claim 9, characterized in that: The steps for obtaining the grayscale image data include: The weighting coefficients for red, green, and blue pixel values ​​are set based on the color sensitivity of the human eye; The grayscale value of a pixel is calculated using weighted coefficients for red, green, and blue pixel values, and the grayscale value of the pixel is used as grayscale image data.

Images