Power cable production link monitoring method

Abstract

The invention discloses a power cable production link monitoring method, and relates to the technical field of power cables. The method comprises the following steps: splitting production nodes according to an execution sequence, and collecting process parameters and image data of each production node in real time; correcting the image data by using a convolutional neural network to obtain a corrected image; based on pre-stored reference features, extracting corrected image features through a semi-supervised semantic segmentation method, respectively comparing the corrected image features and the process parameter features with preset reference values, and calculating to obtain a preliminary qualification index; constructing a real-time feature vector, inputting the real-time feature vector into the graph neural network model, outputting a reconstructed feature vector, and calculating a real-time reconstruction error; and weighting the preliminary qualification index and the normalized reconstruction error through an analytic hierarchy process to obtain an abnormal score, and based on the abnormal score, obtaining an overall qualification index and dividing risk levels.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power cable, in particular to a power cable production link monitoring method. BACKGROUND

[0002] As the core carrier of power transmission system, the production quality of power cable directly determines the safety and stability of power transmission. With the popularization of high-voltage and ultra-high-voltage power engineering, the cable production process needs to cover conductor extrusion, insulation coating, shielding weaving, sheath forming and other processes. Process fluctuations in each link may cause quality problems such as insulation breakdown and shielding failure.

[0003] The traditional monitoring method widely used in the current industry is a power cable production monitoring method based on single-node process parameter threshold. This method first splits the cable production process into independent production nodes, sets process parameter baseline thresholds for each production node, and collects process parameters of each production node in real time through sensors. The real-time parameters are compared with the preset thresholds to determine whether the production status of a single production node is qualified.

[0004] However, this traditional method only focuses on whether the process parameters of a single production node exceed the threshold, and cannot capture the process dependency relationship between production nodes, so it cannot identify the hidden quality problems of subsequent production nodes caused by slow drift of previous production node parameters. Ultimately, potential quality risks in the production process are missed, making it difficult to ensure the quality stability of the entire cable production process. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a power cable production link monitoring method to solve the problems in the background art.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: a power cable production link monitoring method, comprising the following steps: Step S1: split the production link of the cable to be produced according to the execution order to obtain a plurality of production nodes, and collect process parameters and image data of each production node in real time; Step S2: correct the image data by an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image; Step S3: obtain pre-stored baseline image features and baseline process parameters, extract features from the corrected image by a semi-supervised image semantic segmentation method to obtain corrected image features, compare the corrected image features with the baseline image features to calculate an image comparison score, compare the process parameters with the baseline process parameters to calculate a parameter comparison score, and calculate the image comparison score and the parameter comparison score by a weighted summation method to obtain a preliminary qualification index of each production node; Step S4: Constructing a real-time feature vector based on the process parameters, the corrected image and the preliminary qualification index, inputting the real-time feature vector into the graph neural network model, outputting a reconstructed feature vector, and calculating a real-time reconstruction error based on the reconstructed feature vector; Step S5: Based on the preliminary qualification index and the real-time reconstruction error, performing weighted summation by using the analytic hierarchy process to calculate an abnormal score, calculating an overall qualification index based on the abnormal score, and dividing a risk level to realize monitoring of the cable production link.

[0007] Preferably, the production link of the cable to be produced is split according to the execution order to obtain a plurality of production nodes, including the following specific steps: First, the core process links of the industrialized production of power cables are combed, and four continuous production links of "conductor extrusion, insulation layer coating, shielding layer weaving and outer sheath coating" are determined. The process characteristics and quality influence weights of each link are directly related, and are split according to the execution order to finally obtain the extrusion node, the insulation node, the weaving node and the sheath node.

[0008] Preferably, the image data is corrected by the image distortion correction algorithm based on the convolutional neural network to obtain a corrected image, including the following steps: The preprocessed single-channel image is input into the CNN model, and the model outputs five predicted distortion parameters. Based on the predicted distortion parameters, a nonlinear distortion model of the cable image is constructed to describe the mapping relationship between the pixel coordinates of the distorted image and the ideal non-distorted coordinates. The calculation formula is:

[0009] wherein, is the coordinate of the pixel in the ideal non-distorted image, corresponding to the pixel position of the real form of the cable, is the coordinate of the pixel in the distorted image, r is the distance from the coordinate of the pixel in the ideal non-distorted image to the center of the image, is a low-order radial distortion parameter, and is a high-order radial distortion parameter, is a horizontal direction tangential distortion parameter, is a vertical direction tangential distortion parameter; The distortion correction adopts the inverse mapping method. First, each pixel of the ideal non-distorted image is traversed, and its corresponding coordinate in the distorted image is calculated by the above formula. Then, the gray value is extracted from the distorted image and filled into to obtain the corrected image. If is a non-integer coordinate, the gray value is calculated by using the quadratic linear interpolation method:

[0010] wherein, is the gray value of the pixel in the ideal undistorted image, is the gray value of the pixel in the distorted image, is the gray value of the pixel in the distorted image, is the gray value of the pixel in the distorted image, is the interpolation weight coefficient, , {0, 1}, represents the offset of the integer part of the target coordinate in the distorted image relative to the 2x2 pixel neighborhood in which the target coordinate is located; The morphological consistency of the corrected image is verified, a cable cross-section contour is extracted by a Hough circle detection algorithm, the circularity of the contour is calculated, and if the circularity is greater than or equal to 0.95, it is determined that the correction is qualified; if the circularity is less than 0.95, the CNN model is re-inputted for secondary prediction of distortion parameters and correction until it is qualified.

[0011] Preferably, the feature extraction of the corrected image by the semi-supervised image semantic segmentation method comprises the following steps: The feature extraction of the corrected image by the semi-supervised image semantic segmentation method comprises the following steps: A confidence threshold of 0.8 is set, and a region with a pixel value greater than or equal to 0.8 is selected as a candidate prompt region; 5 prompt points are sampled from the candidate prompt region; The corrected image is input into the SAM model, the SAM model is optimized by weight adjustment, and finally three candidate masks are output; the final mask calculation formula is as follows:

[0012] wherein, is the final output key region pseudo-label mask, is the fine scale candidate mask output by the SAM, is the medium scale candidate mask output by the SAM, is the coarse scale candidate mask output by the SAM, and are mask control weights; The generated pseudo-label mask The region integrity index is calculated, if the coverage rate of the key region mask is greater than or equal to 90%, the mask is determined to be qualified; if the coverage rate is less than 90%, the prompt points are re-generated and fine-tuned again until it is qualified, to ensure that there is no region missing in the mask; Based on the qualified pseudo-label mask, features directly related to the quality of the cable are extracted, which are divided into two categories: "visible light image features" and "infrared image features"; the feature definition, calculation method and value are as follows: ​​​For visible light image, two features, insulation layer thickness and surface defect density, need to be calculated. The formula for calculating insulation layer thickness is: wherein, is the number of radial pixels of insulation layer in mask, is the camera calibration coefficient, taking 0.01 mm / pixel, is the image zoom factor, is the insulation layer thickness; the formula for calculating surface defect density is: wherein, is the number of defect pixels in mask, is the area of key region, is the surface defect density, and the defect pixels are identified by edge detection, ≦0.1 per cm2 is qualified; For infrared image, two features, internal temperature standard deviation and hot spot proportion, need to be calculated. The formula for calculating internal temperature standard deviation is: wherein, is the pixel temperature of infrared image, is the average temperature, is the total number of pixels, is the internal temperature standard deviation, ≦3° is qualified to avoid local overheating leading to insulation aging; the formula for calculating hot spot proportion of infrared image is: wherein, is the hot spot proportion of infrared image, is the number of pixels with temperature >220℃.

[0013] Preferably, the image comparison score is: The image comparison score is calculated by comparing the corrected image features with the reference image features. The formula for calculating the image feature similarity score is:

[0014] wherein, is the similarity score of a single image feature, is the feature value of the current image, is the reference value of the feature, is the attenuation coefficient, taking 5.

[0015] Preferably, the parameter comparison score is: The parameter comparison score is calculated by comparing the process parameters with the reference process parameters. The formula for calculating the parameter comparison score is as follows:

[0016] wherein, is the parameter comparison score, representing the overall qualification degree of the process parameters of the node, the number of process parameters of the node, the standardized value of the nth parameter, the reference value of the kth parameter, the deviation tolerance threshold of the nth parameter.

[0017] Preferably, the image comparison score and the parameter comparison score are calculated by using the weighted summation method to obtain the preliminary qualification index of each production node, including the following specific steps: The image comparison score and the parameter comparison score are calculated by using the weighted summation method to obtain the preliminary qualification index of each production node, and the calculation formula of the preliminary qualification index is:

[0018] wherein, the preliminary qualification index of the production node, the parameter comparison weight, which is 0.6, the image comparison weight, the parameter comparison score, the image comparison score.

[0019] Preferably, the real-time reconstruction error is calculated based on the reconstructed feature vector, including the following specific steps: In real-time monitoring, the real-time feature vector is input into the trained graph neural network model, and the sequence data of the previous 4 time steps is intercepted with the current time as the terminal point to construct the real-time feature vector of each node , which contains the process parameter sequence, the image feature sequence, and the preliminary qualification index, and the real-time graph structure is input into the graph neural network model to output the reconstructed feature vector of each node , and then the real-time reconstruction error of each node is calculated, and the calculation formula of the real-time reconstruction error is:

[0020] wherein, the real-time reconstruction error of the ith node, the number of time steps of the real-time sequence, the real-time feature vector of node i at the tth time step, the reconstructed feature vector of node i at the tth time step.

[0021] Preferably, the abnormal score is calculated by using the analytic hierarchy process to perform weighted summation based on the preliminary qualification index and the real-time reconstruction error, including the following specific steps: The index weight is calculated by using the analytic hierarchy process, the sum of each column element is calculated first, each element of the judgment matrix A is divided by the sum of the column where it is located to obtain the normalized matrix The preliminary characteristic vector is calculated by a normalization matrix, and the calculation formula of the preliminary characteristic vector is: The preliminary characteristic vector is normalized by dividing all elements of the preliminary characteristic vector by the sum of the preliminary characteristic vector, and the calculation formula of the weight of the fth index is:

[0022] Wherein, is the weight of the fth index, is the sum of the preliminary characteristic vector, is the preliminary characteristic vector; The calculation formula of the abnormal score is:

[0023] Wherein, is the abnormal score, and the larger the value is, the better the node quality is, is the index weight of the preliminary qualified index, is the index weight of the normalized reconstruction error.

[0024] Preferably, the overall qualified index is calculated based on the abnormal score, and the calculation comprises the following specific steps: The calculation formula of the corrected node qualified index is:

[0025] Wherein, is the corrected node qualified index, is the preliminary qualified index of the ith node, is the monitoring weight of the ith node, is the abnormal score, is the correction coefficient; The calculation formula of the overall qualified index is:

[0026] Wherein, is the overall qualified index, is the monitoring weight of the ith node, is the corrected node qualified index.

[0027] The application provides a power cable production link monitoring method, relates to machine learning and deep learning technology, and has the following beneficial effects: ​(1) Based on the image distortion correction algorithm of convolutional neural network, the image radial and tangential distortion caused by lens process error and camera installation deviation in power cable production scene is solved. Through accurate prediction of distortion parameters and correction, the image distortion of conductor, insulation layer and other key areas is eliminated, ensuring the authenticity and reliability of image information in multi-modal monitoring data, and laying a high-quality data foundation for cable production whole-process monitoring.

[0028] (2) The semi-supervised semantic segmentation method relies on the corrected distortion-free cable image, and can accurately extract the conductor surface state, insulation layer thickness, braided layer density and other key features without a large amount of manual annotation, which not only greatly reduces the data labeling labor cost of cable production monitoring, but also retains complete process details in the corrected image, making the feature extraction more in line with the actual quality state of cable production, and making the calculation of preliminary qualified index of each production node more valuable.

[0029] (3) Through the graph neural network model, the abnormality of cable production sequence is recognized, which breaks through the limitation of traditional single node threshold judgment, accurately models the process dependence relationship between extrusion, insulation, braiding and sheath nodes, can sensitively capture the implicit abnormal trend of production parameter slow drift, and early warning of potential quality risk such as insufficient crosslinking of insulation layer, avoiding the abnormal omission caused by isolated monitoring of single node, and improving the comprehensiveness and forward-looking of cable production abnormality recognition.

[0030] (4) Through the dynamic abnormal trend recognized by the analytic hierarchy process combined with the graph neural network, the static quality index and process trend information of cable production are fused to accurately correct the preliminary qualified index, which not only avoids the one-sidedness of single dimension score, but also makes the cable whole-link quality grade division more in line with the needs of different application scenes such as high-voltage and low-voltage, significantly improves the reliability of monitoring results, provides accurate guidance for cable production process adjustment and unqualified product rework, and ensures the stability of power cable production quality. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating laborious work.

[0032] Fig. 1 A step flow chart of a power cable production link monitoring method is proposed for the present application. Fig. 2 A step level diagram for obtaining a preliminary qualified index in a power cable production link monitoring method is proposed for the present application. Fig. 3A step level diagram for obtaining an abnormal score in a power cable production link monitoring method according to the present application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0034] Please refer to Figs. 1-3 The present application provides a technical solution: a power cable production link monitoring method.

[0035] Step S1: split the production link of the cable to be produced according to the execution order to obtain a plurality of production nodes, and collect the process parameters and image data of each production node in real time.

[0036] The production link of the current cable to be produced is obtained and split according to the execution order to obtain a plurality of production nodes, and the specific steps are as follows: First, the core process links of the industrialized production of power cables are sorted out, and four continuous production links of "conductor extrusion, insulation layer coating, shielding layer weaving, and outer sheath coating" are determined. The process characteristics and quality influence weight of each link are directly related, and the links are split according to the execution order to obtain a plurality of production nodes.

[0037] Secondly, for each production node, a plurality of modal monitoring items are allocated in combination with its process target: The extrusion node: the target is the conductor forming quality; the parameter monitoring items include the extrusion temperature (180~220°), the die head pressure (5~8MPa, to avoid the appearance of depressions on the surface of the conductor), and the traction speed (0.3~0.5m / s, to match the extrusion efficiency); the image monitoring item is the visible light image of the conductor (the resolution is 1920*1080, to capture the oxidation layer and scratches on the surface of the conductor); The insulation node: the target is to meet the insulation performance; the parameter monitoring items include the coating temperature (200~230°, to ensure that the insulation material is fully cured), the insulation layer thickness (0.8-1.2mm), and the curing time (30-40min, to ensure that the insulation crosslinking degree is ≥70%); the image monitoring items include the visible light image of the insulation layer (to monitor the surface bubbles and impurities) and the infrared image (the resolution is 384*288, to monitor the internal temperature uniformity and avoid local overheating to cause insulation aging); The weaving node: the target is to stabilize the shielding performance; the parameter monitoring items include the weaving wire diameter (0.15-0.2mm, to ensure the shielding effect) and the weaving density (≥90%); the image monitoring item is the visible light image of the weaving layer (to monitor the broken wires and missed weaving); Sheath node: target for weather resistance and protection; parameter monitoring items include sheath thickness (1.0-1.5mm), extrusion temperature (190-210°); image monitoring item is visible light image of sheath layer (monitoring color unevenness, surface cracks) Real-time acquisition of process parameters and image data of each production node is implemented as follows: The temperature of the extrusion / sheath link uses a PT100 platinum resistance sensor (measurement range -50-300°, accuracy ±0.5℃), the head pressure uses a diffusion silicon pressure sensor (measurement range 0~15MPa, accuracy ±0.1MPa), and the insulation / sheath thickness uses a laser thickness gauge (measurement range 0~5mm, accuracy ±0.01mm). All sensors are connected to the edge gateway through an RS485 bus, the sampling frequency is set to 1 / s, and according to the cable production speed (0.3~0.5m / s), the cable moves less than 0.5m in 1s, which can ensure that at least 2 parameter data points are collected per meter of cable, avoiding data omission. Image acquisition equipment: 4 visible light cameras with 2 million pixels CMOS camera, 1 at 0.5m downstream of the conductor extrusion die outlet, 1 at 0.5m downstream of the insulation layer coating equipment outlet, 1 at 0.5m downstream of the braiding machine outlet, and 1 at 0.5m downstream of the sheath forming equipment outlet; 2 infrared cameras with non-cooled infrared focal plane camera (resolution 384*288, temperature measurement range -20-500°), 1 at the outlet of the extrusion node (monitoring conductor extrusion temperature distribution) and 1 at the outlet of the insulation node; camera deployment position: installation height 1.5m from the cable surface (according to lens focal length 16mm, ensuring that a single frame image covers 1m long cable area), camera lens axis perpendicular to cable axis (avoiding tangential distortion), acquisition frequency set to 1 frame / 5s, 4 visible light cameras are respectively assigned a unique hardware ID, and the camera automatically embeds the node ID in the collected image file after starting, directly associated with the corresponding production node. The edge gateway uses a 5G industrial router (transmission rate ≥100Mbps) to transmit the collected parameter data and image data to the control computer in real time. The computer performs preliminary cleaning on the data: parameter data uses "sliding window filtering method" (window size 5, i.e. taking the mean value of 5 consecutive sampling points to replace the original point, eliminating impulse noise), and image data uses "size normalization" (unified scaling to 360*360 pixels, reducing the calculation amount of subsequent distortion correction), and the preprocessed data is temporarily stored in the local database of the computer.

[0038] Step S2: correcting the image data by an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image.

[0039] In industrial scenarios, the distortion of power cable images mainly comes from two factors: one is the radial distortion caused by lens processing errors (such as "fish-eye effect", which shows that the edge pixels of the cable are stretched or compressed), and the other is the tangential distortion caused by the non-perpendicularity of the camera axis and the cable axis (which shows that the cross-section of the cable is elliptical instead of circular). To correct these two types of distortion, a CNN model needs to be trained based on MATLAB to cover different distortion levels. The specific steps are as follows: Take a standard cable sample as the shooting object, adjust the lens focal length (4-300mm, covering the commonly used focal length range in industry), and the camera tilt angle (0°-15°, simulating the actual installation deviation) to shoot 500 real distortion images. At the same time, based on the principle of machine vision imaging, 11500 synthetic images with different distortion parameters are generated through MATLAB simulation. The total number of images in the training set is 12000 (training set: validation set: test set = 8:1:1).

[0040] The distortion parameters of the synthetic images follow the distribution of the actual industrial scene. The specific value range is as follows: , [ ], is the low-order radial distortion parameter, is the high-order radial distortion parameter, and [ ], is the horizontal tangential offset, is the vertical tangential offset.

[0041] A CNN model with 4 layers of convolution, 4 layers of pooling, and 2 layers of full connection is designed to adapt to the cable image distortion parameter prediction task. The specific structure and parameter values are as follows: The input layer is 1 layer with an input size of 360*360*1. The convolution layer is 4 layers with a convolution kernel of 3×3, which can reduce parameters while preserving edge features. The step is 1, padding=“SAME” to avoid edge pixel loss, and the output channel number is 32→64→128→256 to adapt to the cable long strip feature.

[0042] The activation function is 4 layers of ReLU function to solve the gradient disappearance problem. The pooling layer is 4 layers of max pooling to preserve the key information of the image edge with a pooling kernel of 2×s2 and a step of 2. The full connection layer is 2 layers with 512 neurons in the first layer and 5 neurons in the second layer. The 512 neuron layer adapts to the high-dimensional feature mapping of the image, and the 5 neuron layer corresponds to the output of the 5 distortion parameters.

[0043] The loss function uses the mean square error (MSE) function, and the distortion parameter prediction is a regression task. MSE can accurately measure the deviation between the predicted value and the true value.

[0044] The model was trained on an industrial control computer with the following training parameters: Batch Size = 20 (to balance memory usage and training efficiency), Initial Learning Rate = 0.001 (using the Adam optimizer, the learning rate decays to 1 / 10 of its original value every 50 epochs), and Epochs = 200 (using early stopping to stop training when the validation set loss does not decrease for 20 consecutive epochs to avoid overfitting). After training, the model's distortion parameter prediction error on the test set was ≤5%, meeting the industrial correction accuracy requirements.

[0045] After the cable image acquired in step 1 is input into the trained CNN model, it needs to go through four steps: "distortion parameter prediction → nonlinear distortion model construction → pixel coordinate correction → grayscale value interpolation" to achieve distortion-free output. The specific implementation is as follows: The preprocessed single-channel image is input into a CNN model, which outputs five predicted distortion parameters. Based on these predicted distortion parameters, a nonlinear distortion model for the cable image is constructed to describe the mapping relationship between pixel coordinates in the distorted image and ideal distortion-free coordinates. The calculation formula is as follows:

[0046] in, These are the pixel coordinates in an ideal, distortion-free image, corresponding to the pixel positions of the actual cable shape. Let be the coordinates of a pixel in the distorted image, and r be the distance from the coordinates of a pixel in the ideal, distortion-free image to the image center. These are low-order radial distortion parameters. and These are higher-order radial distortion parameters. The horizontal tangential distortion parameter. This represents the tangential distortion parameter in the vertical direction.

[0047] It should be noted that the formula for calculating the distance from the pixel coordinates to the image center in an ideal distortion-free image is: It is used to quantify the variation of radial distortion with pixel distance.

[0048] Distortion correction employs a "reverse mapping method." First, it iterates through each pixel of the ideal, distortion-free image, calculating its corresponding coordinates in the distorted image using the formula described above. Then, it extracts grayscale values ​​from the distorted image to fill the distortion. To obtain the corrected image, if For non-integer coordinates, quadratic linear interpolation is used to calculate grayscale values ​​to avoid "pixel holes". The interpolation formula is as follows:

[0049] in, For an ideal distortion-free image The grayscale value of a pixel, In distorted images The grayscale value of a pixel, These are the interpolation weighting coefficients. , {0, 1} represents the target coordinates relative to the distorted image. The offset of the integer part, through Locate target coordinates The four integer pixel coordinates of the 2×2 pixel neighborhood where it is located.

[0050] It should be noted that, The calculation formula is: The significance of this formula is that quadratic linear interpolation can better fit the gray-scale gradation characteristics of cable images, and the edge blurring of the corrected image is reduced.

[0051] The morphological consistency of the corrected image is verified by extracting the cable cross-sectional profile using the Hough circle detection algorithm and calculating the roundness of the profile (roundness = ...). ×Area / Perimeter 2 (Ideal circle roundness = 1). If the roundness is ≥ 0.95, the correction is considered qualified. If the roundness is < 0.95, the distortion parameters of the CNN model are re-inputted for secondary prediction and correction until qualified, ensuring that there is no significant distortion in each frame of the image.

[0052] The process parameters collected in Step 1 have significant dimensional differences and dispersed numerical ranges. Direct comparison will lead to an imbalance in the weighting of temperature and pressure deviations. Therefore, standardization is required to map all parameters to a unified range. Specific implementation details are as follows: Step 1 has already used sliding window filtering to initially eliminate transient noise. This step further employs 3σ outlier removal: the mean and standard deviation of the historical 1000 normal sample values ​​for each parameter are calculated. If the current sample value x satisfies... If the value is not found, it is determined to be impulse noise and replaced with the normal sampled value from the previous moment to ensure the continuity of the parameter sequence.

[0053] The filtered parameter values ​​are mapped to the [0,1] interval, and the standardization formula is as follows: ,in, These are the standardized parameter values. These are the original process parameter values. and These are the lower and upper limits of the parameter reference range, respectively. The standardized parameter data and the corresponding corrected image data are associated with each other through timestamps and node IDs and stored in the time-series database of the industrial control computer.

[0054] Step S3: Obtain the pre-stored reference image features and reference process parameters, extract features from the correction image by semi-supervised image semantic segmentation method, obtain the correction image features, compare the correction image features with the reference image features, calculate the image comparison score, compare the process parameters with the reference process parameters, calculate the parameter comparison score, and calculate the image comparison score and the parameter comparison score by using the weighted summation method to obtain the preliminary qualification index of each production node.

[0055] The process parameters of 1000 batches of qualified cables produced in the past year (3 samples were randomly selected from each batch, a total of 3000 samples) were collected, and the reference process parameters of each parameter were constructed by using the "3 sigma principle". This principle can cover 99.73% of normal production data and exclude extreme abnormal value interference. All reference process parameters are stored in the reference database of the industrial control computer.

[0056] Select 10% of the qualified cable images (a total of 300, covering four production nodes), manually label the conductor, insulating layer and other key areas by experienced quality inspection personnel, generate real label masks as semi-supervised learning seed data; input the remaining 90% of unannotated images into the SAM model, generate class activation maps (confidence threshold =0.8) by Grad-CAM, select high contribution pixel regions, and use the farthest point sampling to extract 5 hint points, input the SAM to generate pseudo-label masks, and select masks with a confidence of ≥0.9 to ensure accuracy: Based on the real label mask and the qualified pseudo-label mask, extract core image features, visible light image focus edge smoothness, defect density, infrared image focus internal temperature standard deviation, and hot spot proportion. All features are stored in the format "node-feature name-mean-standard deviation" to form a reference image feature library.

[0057] Extract features from the correction image by semi-supervised image semantic segmentation method to obtain the correction image features, the specific steps are as follows: Input the correction image into the pre-trained image classification model (ResNet50, fine-tuned on the cable image dataset), and output the conductor / insulating layer / braided layer / sheath layer 4 categories. Calculate the gradient weight of the target category by Grad-CAM to generate class activation map CAM. The higher the pixel value in the CAM map, the greater the contribution of the position to the identification of the target category. For example, in the CAM map of the insulating layer category, the pixel value of the insulating layer region is significantly higher than that of other regions (mean≥180, background region≤50).

[0058] Set the confidence threshold =0.8 (based on experimental verification, When the pixel value is 0.8, the overlap rate between the activation map and the real area is ≥85%, which can effectively filter out high-contribution pixel areas and avoid background pixel interference. Areas with a pixel value ≥0.8 are selected as "candidate hint areas".

[0059] Five prompt points are sampled from the candidate prompt area (the value is verified by experiments; five prompt points can cover the center and edge of the critical area of ​​the cable). The sampling algorithm follows the principle of "maximizing the point spacing" to ensure that the prompt points are evenly distributed. The specific algorithm steps and formulas are as follows: Initialize the sampling point set During the initial sampling, one initial point is randomly selected from the candidate suggestion area. (Prioritize the pixel with the highest CAM value, i.e., the center of the critical area). Calculate all other points in the candidate suggestion area to In the Euclidean distance algorithm, iterate through all Euclidean distance values ​​and select the point with the largest value as the endpoint. ,join in ,at this time ={ , }, sampling point set When there are two or more points, calculate the distance to all other points in the candidate suggestion area. The minimum Euclidean distance to all existing points in the candidate hint area, and the distance to all other points in the candidate hint area. The formula for calculating the minimum Euclidean distance between points in the middle is:

[0060] in, For all other points in the candidate suggestion area The minimum Euclidean distance between all points in the middle. For the coordinates of all other points in the candidate suggestion area, for The coordinates of each point in the middle.

[0061] Filter out If a unique maximum point exists, then directly use that as the maximum point. join in If multiple points have the same and highest OU value, extract the Class Activation Map (CAM) heatmap values ​​of all candidate points, prioritizing points with higher heatmap values. If the heatmap values ​​are still the same, sort the points in ascending order by their x-axis coordinates and select the points with smaller x-coordinates; if the x-coordinates are still the same, sort them in ascending order by their y-axis coordinates and select the points with smaller y-coordinates. join in Repeat the above steps until... It contains 5 prompts.

[0062] It should be noted that the significance of the algorithm is to ensure that the prompt points cover the edges of the key areas (such as the junction of the insulating layer and the conductor, the surface of the insulating layer) by maximizing the point distance, providing "comprehensive and non-redundant" positioning information for the SAM model, and improving the edge coincidence of the pseudo-label mask compared with randomly sampled prompt points.

[0063] The original SAM model has insufficient segmentation accuracy for cable-like "long strip, strong texture" targets (edge detection rate ≥ 15%), so the SAM needs to be fine-tuned for cable image features: The SAM model is divided into "image encoder (ViT-H), prompt encoder, and mask decoder" three parts, among which the pre-training weight of the image encoder (trained on the SA-1B dataset) is frozen, only the "mask control weight" is trained, and the three candidate masks (original SAM outputs three different scale masks) output by the SAM are optimized through weight adjustment. The final mask calculation formula is as follows:

[0064] Among them, is the final output key area pseudo-label mask, is the fine scale candidate mask output by SAM, is the medium scale candidate mask output by SAM, is the coarse scale candidate mask output by SAM, and is the mask control weight.

[0065] It should be noted that the pseudo-label mask is a binary image, 1 represents the key area, and 0 represents the background. The initial value of the mask control weight is 1 / 3, which is obtained through training by minimizing the "intersection over union (IoU) loss" of the mask and the artificial seed label.

[0066] For the generated pseudo-label mask Calculate the area integrity index. If the mask coverage rate of the key area (such as the insulating layer) is ≥ 90% (coverage rate = number of 1 pixels in the mask / number of real key area pixels), the mask is determined to be qualified; if the coverage rate is < 90%, generate a new prompt point (add 1 prompt point) and fine-tune it again until it is qualified, ensuring that the mask has no "area missing" (such as missing the braided layer filament area).

[0067] Based on the qualified pseudo-label mask, extract features directly related to cable quality, which can be divided into "visible light image features" and "infrared image features". The feature definition, calculation method and value are as follows: For visible light images, two types of features, insulating layer thickness and surface defect density, need to be calculated. The calculation formula of the insulating layer thickness is: Among them, The number of radial pixels in the insulating layer of the mask, The camera calibration coefficient is 0.01 mm / pixel, The image scaling factor, The thickness of the insulating layer; the calculation formula of the surface defect density is: Wherein, The number of defect pixels in the mask, The area of the key region, The surface defect density is identified by edge detection (Canny operator), ≦0.1 / μm2 Pass.

[0068] For infrared images, the internal temperature standard deviation and two types of features need to be calculated. The calculation formula of the internal temperature standard deviation is: Wherein, The pixel temperature of the infrared image, The average temperature, The total number of pixels, The internal temperature standard deviation, ≦3° is qualified, avoiding local overheating leading to insulation aging; the calculation formula of the infrared image hotspot ratio is: Wherein, The infrared image hotspot ratio, The number of pixels with temperature >220℃.

[0069] It should be noted that the significance of feature extraction is to convert image "qualitative features" into "quantitative indicators", providing a unified dimension for subsequent fusion calculation of process parameters and qualification index.

[0070] Compare the corrected image features with the reference image features to calculate the image comparison score. The calculation formula of the image feature similarity score is:

[0071] Wherein, The similarity score of a single image feature, The feature value of the current image, The feature reference value, The attenuation coefficient is 5.

[0072] Compare the process parameters with the reference process parameters to calculate the parameter comparison score. The calculation formula of the parameter comparison score is as follows:

[0073] Wherein, The parameter comparison score represents the overall qualification degree of the process parameters at this node, the number of process parameters of the node, the standardized value of the nth parameter, the reference value of the kth parameter, the deviation tolerance threshold of the nth parameter.

[0074] It should be noted that the reference value takes the median value of the reference interval, and the deviation tolerance threshold takes 10% of the width of the reference interval.

[0075] For all image features of each production node, the mean value is taken as the image comparison score .

[0076] The image comparison score and the parameter comparison score are calculated by using the weighted summation method to obtain the preliminary qualified index of each production node, and the calculation formula of the preliminary qualified index is:

[0077] wherein, the preliminary qualified index of the production node, the parameter comparison weight, which is 0.6, the image comparison weight, the parameter comparison score, the image comparison score.

[0078] It should be noted that, ≥0.8 indicates preliminary qualification, <0.8 indicates preliminary disqualification.

[0079] Step S4: Based on the process parameters, the corrected image and the preliminary qualified index, a real-time feature vector is constructed, the real-time feature vector is input into a graph neural network model, a reconstructed feature vector is output, and a real-time reconstruction error is calculated based on the reconstructed feature vector.

[0080] Cable production is a continuous process of "extrusion → insulation → braiding → sheath", and the process state of the previous node (such as extrusion) directly affects the quality of the subsequent node (such as insulation) (such as low extrusion temperature, which will lead to insufficient cross-linking degree of the insulation layer). Traditional single-node threshold judgment cannot capture this correlation. This step selects a graph neural network model, and the core process includes two parts of "node feature definition" and "edge weight definition": The nodes of GNN need to include dynamic sequence information (capture trends) and static feature information (reflect current state). Based on process parameters, image data and preliminary qualified index, a real-time feature vector is constructed, and the calculation formula is:

[0081] wherein, the real-time feature vector of the ith production node, the process parameter sequence segment of the i-th node, the correction image feature sequence segment of the i-th node, the preliminary pass index of the i-th node, the vector splicing operation.

[0082] It should be noted that the process parameter sequence segment and the correction image feature sequence segment take the normalized parameter values and the image feature values of the current time and the previous 4 times, a total of 5 time steps, The time step interval is 1 second, The time step interval is 5 seconds, and the sequence length of 5 time steps can cover a range of 2.5-5 meters of cable production, which is sufficient to capture the trend of slow drift of parameters.

[0083] Based on the power cable production process, a topological graph representing the process association relationship between nodes is constructed, the edges of the GNN are used to represent the process dependency relationship between production nodes, only the directed edges of “predecessor node→successor node” (such as extrusion→insulation, insulation→braid, braid→sheath) are retained, there is no reverse edge (the successor node does not affect the predecessor node), and the calculation formula of the edge weight is:

[0084] Among them, is the edge weight of the predecessor node i to the successor node j, is the number of times that node j is abnormal due to the abnormality of node i in historical production, is the total number of abnormal times of node j.

[0085] In addition, in order to adapt to process fluctuations, the edge weight can be automatically adjusted during the GNN training process through the graph attention layer (the initial weight is the calculated value, and the attention coefficient is updated during training), and the adjustment formula of the attention coefficient is:

[0086] Among them, is the adjusted attention coefficient, reflecting the influence strength of node i on j learned by the model, is the feature transformation matrix, is the attention vector, used to calculate the similarity between nodes, is the neighbor node set of node j, is the activation function, is the neighbor node index representing node j.

[0087] It should be noted that through the attention mechanism, the model can dynamically adjust the edge weight, avoid abnormal misjudgment caused by fixed weight, and improve the robustness of abnormal identification.

[0088] The core goal of the GNN model is to learn the correlation pattern of node features in normal production state, and then judge the abnormality through the deviation of the current sequence from the normal pattern, so the training data needs to be the sequence data of historical fault-free production, and the specific training process is as follows: Collect the sequence data of 500 batches of normal production cables in the past year (each batch contains 4 nodes, and each node has 500 time step feature vectors), divide them into training set, validation set and test set according to the ratio of 7:2:1, and convert each batch of data into the input graph structure of GNN , wherein, represents 4 production nodes, represents 3 directed edges, and the feature vector of each node i is .

[0089] The GNN structure of 2-layer graph attention layer + 1-layer fully connected layer is selected, and the specific configuration and hyperparameter values are as follows: The input dimension of the input layer is 31, which is used to match the dimension of , to ensure that the features are not truncated, the graph attention layer is 2 layers, the hidden layer dimension is 64, which is used to balance the expression ability and the amount of calculation, the number of attention heads is 4, the dropout rate is 0.5, the output dimension of the fully connected layer is 31, the Adam optimizer is used, the initial learning rate is , the weight decay is , the training batch size BatchSize=32, and the iteration number Epoch=200.

[0090] The loss function adopts the node feature reconstruction error (MSE), and the calculation formula of the loss function is:

[0091] , wherein, is the total loss of the model, is the training batch size, is the number of production nodes, is the number of time steps in each batch, is the real feature vector of the bth batch, tth time step and node i, is the reconstructed feature vector output by the model, is the L2 norm, which is used to measure the difference between the real and reconstructed vectors.

[0092] In real-time monitoring, the real-time feature vector is input into the trained GNN model, the current time is taken as the end point, the sequence data of the previous 4 time steps is intercepted to construct the real-time feature vector of each node, which includes process parameter sequence, image feature sequence and preliminary qualification index, the real-time graph structure is input into GNN, and the reconstructed feature vector of each node is output by the model Then, the real-time reconstruction error of each node is calculated. The formula for calculating the real-time reconstruction error is:

[0093] in, Let be the real-time reconstruction error of the i-th node. This represents the number of time steps in the real-time sequence. Let be the real-time feature vector of node i at time step t. Let be the reconstructed feature vector of node i at time step t.

[0094] The anomaly threshold is calculated by taking the 95th percentile of the reconstruction error set of all nodes and all time steps in historical normal production using a quantile function. If any node's > If all nodes... If the error is found to be normal, it is considered normal. The error percentage of each node is calculated, and the node with the largest error percentage is the root cause node of the anomaly. The warning information is output to the industrial control computer, including the abnormal node, the current error value, and the warning level. An error > 0.2 is a level 1 warning, and an error < 0.12 ≤ 0.2 is a level 2 warning.

[0095] Step S5: Based on the preliminary qualification index and real-time reconstruction error, the analytic hierarchy process is used to perform weighted summation to calculate the anomaly score. Based on the anomaly score, the overall qualification index is calculated and the risk level is divided to realize the monitoring of the cable production process.

[0096] The real-time reconstruction error calculated in step 4 is a non-negative value and needs to be mapped to the [0,1] interval to correspond with... , Using a unified dimension and employing "reverse normalization" (the smaller the error, the larger the normalized value), the formula for calculating the reconstruction error index after normalization is as follows:

[0097] in, This represents the normalized reconstruction error; a larger value indicates that the sequence is closer to the normal pattern. Let be the real-time reconstruction error of the i-th node. The minimum reconstruction error in normal historical production. This represents the maximum reconstruction error from historical abnormal production. Ensure that the normalized value is not less than 0.

[0098] Based on the preliminary pass index and real-time reconstruction error, the analytic hierarchy process (AHP) is used for weighted summation to calculate the anomaly score. Specific steps are as follows: The analytic hierarchy process (AHP) is used to determine the weight of the three types of indexes, and subjective assignment is avoided. First, a hierarchical structure is constructed: the target layer is the "node abnormal score", the criterion layer is the three types of indexes (the preliminary pass index, the normalized real-time reconstruction error), and the scheme layer is each production node (extrusion, insulation, braiding, and sheath). preliminary pass index, normalized real-time reconstruction error), and the scheme layer is each production node (extrusion, insulation, braiding, and sheath).

[0099] Five cable process experts are invited to compare the criterion layer indexes according to the "importance of indexes to quality evaluation". A 1-9 scale method is used (1 indicates that the two indexes are equally important, 3 indicates that the former index is slightly more important than the latter index; a scale of 5 indicates that the former index is obviously more important than the latter index; a scale of 7 indicates that the former index is strongly more important than the latter index; and a scale of 9 indicates that the former index is extremely more important than the latter index; and scales of 2, 4, 6, and 8 correspond to intermediate states between adjacent scales). The scores of the two-by-two comparison of the criterion layer indexes are obtained, and the average of the expert scores is taken to construct the judgment matrix A:

[0100] wherein, is the average of the scale. is the average of the scale.

[0101] It should be noted that, =1.5 indicates that the preliminary pass index is slightly more important than the normalized reconstruction error.

[0102] The index weight is calculated by the eigenvector method. The sum of each column element is calculated, each element of the judgment matrix A is divided by the sum of the column in which it is located, and the normalized matrix is obtained. The preliminary eigenvector is calculated by the normalized matrix, and the calculation formula of the preliminary eigenvector is: The sum of all elements of the preliminary eigenvector is divided by the sum to achieve normalization. The calculation formula of the weight of the fth index is:

[0103] wherein, is the weight of the fth index, is the sum of the preliminary eigenvector, is the preliminary eigenvector.

[0104] The calculation formula of the abnormal score is:

[0105] wherein, is the abnormal score, and the larger the value, the better the node quality, ​​The index weight of the preliminary qualified index, The index weight of the normalized reconstruction error.

[0106] The overall qualified index is calculated based on the abnormal score, and the specific steps are as follows: The preliminary qualified index does not consider the sequence dynamic trend, and needs to be corrected combined with the abnormal score, and the overall qualified index is calculated to provide basis for production adjustment.

[0107] According to the quality requirements of industrial cables, the abnormal score is divided into three levels, corresponding to different correction coefficients, to ensure that the corrected index matches the actual quality level: The weight of each monitoring node is calculated and determined based on the "fault influence degree weighting method": according to the production fault records in the past three years, the fault proportion of each link is recorded as , for example, based on the production fault records in the past three years of a cable factory, the statistical data shows that the fault proportion of each link is: insulation layer 42%, sheath 30%, extrusion 18%, and braiding 10%. The fault proportion is converted into the monitoring weight of each link, and the relationship formula between the fault proportion and the monitoring weight is: , wherein is the monitoring weight of the i th node, is the fault proportion of the i th link, the calculation result of the monitoring weight is rounded to one decimal place, and finally the monitoring weight of each production node is: extrusion node =0.2, insulation node =0.4, braiding node =0.1, and sheath node =0.3. This weight distribution intuitively reflects the importance of each node in the cable production quality monitoring, among which the insulation node is the core monitoring object due to the highest fault proportion (0.4), followed by the sheath node (0.3), and the extrusion and braiding nodes are respectively given corresponding weights according to the fault influence degree.

[0108] The calculation formula of the corrected node qualified index is:

[0109] , wherein is the corrected node qualified index, is the preliminary qualified index of the i th node, is the monitoring weight of the i th node, is the abnormal score, is the correction coefficient.

[0110] ≥0.95 is the excellent level, representing that the node static quality is excellent, the dynamic trend is normal, and the visual features match, 0.8≤ <0.95, representing that the node has no obvious abnormalities but has slight fluctuations, <0.8, representing that the node has an abnormal risk.

[0111] The calculation formula of the overall qualified index is:

[0112] wherein, is the overall qualified index, is the monitoring weight of the ith node, is the corrected node qualified index.

[0113] According to is divided into three levels, corresponding to different cable application scenarios, and the division is based on the reference industry quality standards and customer demand: The preliminary qualified index is corrected by the abnormal score to obtain the corrected qualified index: ≥0.95, suitable for high-voltage power transmission scenarios, requiring all nodes ≥0.9, without any abnormal items; 0.85≤ <0.95, suitable for low-voltage power distribution scenarios, allowing 1 low-weight node, but without serious abnormalities; <0.85, prohibited from leaving the factory, and needs to be reworked until ≥0.85.

[0114] If the previous node <0.8 (unqualified), the industrial control computer sends a "shutdown signal" to the production equipment, prohibiting the execution of the next node production, after shutdown, the worker adjusts the process according to the abnormal positioning result, re-executes the production and monitoring of the unqualified node, until ≥0.8, if a node is reworked for 3 times in a row <0.8, triggering a "device maintenance warning", prompting to check the equipment to avoid continuous unqualified.

[0115] The application proposes a power cable production link monitoring method, which splits the production nodes, collects multi-modal data, corrects the CNN image distortion, extracts the semi-supervised semantic segmentation features, identifies the GNN sequence abnormalities, and scores the multi-modal fusion abnormalities, to build a full-process quality monitoring system, solving the problems of data distortion, inaccurate feature extraction, missed judgment of abnormalities and one-sided scoring in traditional monitoring, and providing comprehensive and accurate technical support for cable production quality control.

[0116] The image distortion correction algorithm based on the convolutional neural network specifically solves the image radial and tangential distortion caused by lens process error and camera installation deviation in the power cable production scene, accurately predicts the distortion parameters and completes the correction, eliminates the image distortion of the conductor, insulating layer and other key areas, ensures the authenticity and reliability of the image information in the multi-modal monitoring data, and lays a high-quality data foundation for the whole process monitoring of cable production.

[0117] The semi-supervised semantic segmentation method relies on the corrected distortion-free cable image, can accurately extract key features such as conductor surface state, insulating layer thickness and braided layer density without a large amount of manual annotation, greatly reduces the data annotation labor cost of cable production monitoring, and because the corrected image retains complete process details, the feature extraction is more in line with the actual quality state of cable production, making the calculation of the preliminary qualified index of each production node more valuable.

[0118] Through the graph neural network model, the cable production sequence is identified for abnormalities, breaking through the limitations of traditional single-node threshold judgment, accurately modeling the process dependency between nodes such as extrusion, insulation, braiding and sheath, and can sensitively capture implicit abnormal trends such as slow drift of production parameters, and early warning of potential quality risks such as insufficient cross-linking of the insulating layer, avoiding missed judgment of abnormalities caused by isolated monitoring of a single node, and improving the comprehensiveness and forward-looking nature of cable production abnormality identification.

[0119] Through the analytic hierarchy process combined with the dynamic abnormal trend identified by the graph neural network, the static quality indicators and process trend information of cable production are fused to accurately correct the preliminary qualified index, avoiding the one-sidedness of single-dimensional scoring, and making the cable whole-link quality level division more in line with the needs of different application scenarios such as high-voltage and low-voltage, significantly improving the reliability of the monitoring results, providing accurate guidance for cable production process adjustment and unqualified product rework, and ensuring the stability of power cable production quality.

[0120] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "includes" or "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations. The statement "includes a limited element" does not exclude the presence of additional identical elements in the process, method, article or equipment that includes the element.

[0121] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for monitoring a power cable production stage, characterized by: The method comprises the following steps: Step S1: split the production link of the cable to be produced according to the execution sequence to obtain a plurality of production nodes, and collect process parameters and image data of each production node in real time; Step S2: correct the image data through an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image; Step S3: obtain pre-stored reference image features and reference process parameters, extract features of the corrected image through a semi-supervised image semantic segmentation method to obtain corrected image features, compare the corrected image features with the reference image features to calculate an image comparison score, compare the process parameters with the reference process parameters to calculate a parameter comparison score, and calculate the image comparison score and the parameter comparison score by using a weighted summation method to obtain a preliminary qualification index of each production node; Step S4: construct a real-time feature vector based on the process parameters, the corrected image and the preliminary qualification index, input the real-time feature vector into a graph neural network model, and output a reconstructed feature vector to calculate a real-time reconstruction error based on the reconstructed feature vector; Step S5: calculate an abnormal score by using an analytic hierarchy process for weighted summation based on the preliminary qualification index and the real-time reconstruction error, calculate an overall qualification index based on the abnormal score, and divide a risk level to realize monitoring of the cable production link.

2. A method of monitoring a power cable production stage according to claim 1, characterized in that: The production link of the cable to be produced is split according to the execution sequence to obtain a plurality of production nodes, which comprises the following specific steps: First, the core process links of the industrialized production of power cables are combed, and four continuous production links of "conductor extrusion, insulation layer coating, shielding layer braiding and outer sheath coating" are determined. The process characteristics and quality influence weights of each link are directly related, and the links are split according to the execution sequence to finally obtain an extrusion node, an insulation node, a braiding node and a sheath node.

3. A method of monitoring a power cable production stage according to claim 2, characterized in that: The image data is corrected through the image distortion correction algorithm based on the convolutional neural network to obtain the corrected image, which comprises the following steps: The preprocessed single-channel image is input into the CNN model, the model outputs five predicted distortion parameters, a nonlinear distortion model of the cable image is constructed based on the predicted distortion parameters, the mapping relationship between the distorted image pixel coordinates and the ideal non-distorted coordinates is described, and the calculation formula is: ； wherein, is the coordinate of a pixel in the ideal undistorted image, corresponding to the pixel position in the real shape of the cable, is the coordinate of a pixel in the distorted image, r is the distance of the pixel in the ideal undistorted image to the center of the image, is a low order radial distortion parameter, and is a high order radial distortion parameter, is a horizontal tangential distortion parameter, is a vertical tangential distortion parameter; The distortion correction adopts the reverse mapping method, traverses each pixel of the ideal non-distorted image first, calculates the corresponding coordinate in the distorted image through the above formula, and then extracts the gray value from the distorted image to fill in to obtain the corrected image. If is a non-integer coordinate, the gray value is calculated by using the quadratic linear interpolation method. ； wherein is the gray value of the pixel in the ideal undistorted image is the gray value of the pixel in the ideal undistorted image is the gray value of the pixel in the distorted image is the gray value of the pixel in the distorted image is the interpolation weight coefficient , {0, 1} represents the offset of the integer part of the target coordinate with respect to the 4 integer coordinate pixels of the 2x2 pixel neighborhood in which the target coordinate is located ; The corrected image is verified for morphological consistency, the cable cross-section profile is extracted through a Hough circle detection algorithm, the circularity of the profile is calculated, and if the circularity is greater than or equal to 0.95, the correction is determined to be qualified; if the circularity is less than 0.95, the distortion parameters are predicted again and the correction is performed until the correction is qualified.

4. A method of monitoring a power cable production stage according to claim 3, characterized in that: The corrected image features are obtained by extracting features of the corrected image through the semi-supervised image semantic segmentation method, which comprises the following steps: The corrected image features are obtained by extracting features of the corrected image through the semi-supervised image semantic segmentation method, and the specific steps are as follows: A confidence threshold of 0.8 is set, and the area with a pixel value greater than or equal to 0.8 is selected as a candidate prompt area; five prompt points are sampled from the candidate prompt area; The corrected image is input into the SAM model, the SAM model is optimized through weight adjustment, and finally three candidate masks are output, and the final mask calculation formula is as follows: ； wherein, is a key region pseudo-label mask for the final output, is a fine scale candidate mask output by the SAM, is a medium scale candidate mask output by the SAM, is a coarse scale candidate mask output by the SAM, and is a mask control weight; The generated pseudo-label mask A region integrity index is calculated, if the mask coverage rate of the key region is ≥ 90%, it is determined that the mask is qualified; if the coverage rate is < 90%, the prompt point is regenerated and fine-tuned again until it is qualified, ensuring that there is no region missing in the mask; Based on the qualified pseudo-label mask, features directly related to cable quality are extracted, including visible light image features and infrared image features, and the feature definition, calculation method and value are as follows: For visible light images, two features, insulation layer thickness and surface defect density, need to be calculated. The formula for calculating the insulation layer thickness is: wherein, is the number of radial pixels of the insulation layer in the mask, is the camera calibration coefficient, which is 0.01 mm / pixel, is the image zoom factor, is the insulation layer thickness; and the formula for calculating the surface defect density is: wherein, is the number of defect pixels in the mask, is the area of the key region, is the surface defect density. For the infrared image, the internal temperature standard deviation and two types of features need to be calculated, the calculation formula of the internal temperature standard deviation is: Wherein, is the pixel temperature of the infrared image, is the average temperature, is the total number of pixels, is the internal temperature standard deviation, ≦3° is qualified, avoiding local overheating leading to insulation aging; the calculation formula of the infrared image hotspot proportion is: Wherein, is the infrared image hotspot proportion, is the number of pixels with temperature>220℃.

5. A method of monitoring a power cable production stage according to claim 4, characterized in that: The image comparison score is: The image comparison score is calculated by comparing the corrected image features with the reference image features, and the calculation formula of the image feature similarity score is: ； wherein, is a similarity score for a single image feature, is a feature value of a current image, is a feature reference value, is a decay coefficient, and takes a value of 5.

6. A method of monitoring a power cable production stage according to claim 5, characterized in that: The parameter comparison score is: The parameter comparison score is calculated by comparing the process parameters with the reference process parameters, and the calculation formula of the parameter comparison score is as follows: ； wherein, is a parameter matching score representing an overall qualification level of the process parameters of the node, is the number of process parameters of the node, is the normalized value of the n-th parameter, is the reference value of the k-th parameter, is the deviation tolerance threshold of the n-th parameter.

7. A method of monitoring a power cable production stage according to claim 6, characterized in that: The image comparison score and the parameter comparison score are calculated by using the weighted summation method to obtain the preliminary qualified index of each production node, including the following specific steps: The image comparison score and the parameter comparison score are calculated by using the weighted summation method to obtain the preliminary qualified index of each production node, and the calculation formula of the preliminary qualified index is: ； wherein, is a preliminary qualification index for the production node, is a parameter comparison weight, having a value of 0.6, is an image comparison weight, is a parameter comparison score, is an image comparison score.

8. A method of monitoring a power cable production stage according to claim 7, characterized in that: The real-time reconstruction error is calculated based on the reconstructed feature vector, including the following specific steps: In real-time monitoring, the real-time feature vector is input into the trained graph neural network model, the sequence data of the previous 4 time steps is intercepted with the current time as the endpoint, and the real-time feature vector of each node is constructed , including process parameter sequence, image feature sequence, preliminary qualified index, input real-time graph structure into graph neural network model to output reconstructed feature vector of each node , then calculate the real-time reconstruction error of each node, the calculation formula of real-time reconstruction error is: ； wherein, is the real-time reconstruction error for the i-th node, is the number of time steps of the real-time sequence, is the real-time feature vector for node i at time step t, is the reconstructed feature vector for node i at time step t.

9. A method of monitoring a power cable production stage according to claim 8, characterized in that: Based on the preliminary qualified index and the real-time reconstruction error, the analytic hierarchy process is used for weighted summation to calculate the abnormal score, including the following specific steps: The analytic hierarchy process is used to calculate the index weight, the sum of each column element is calculated, each element of the judgment matrix A is divided by the sum of the column to obtain a normalized matrix , the preliminary characteristic vector is calculated through the normalized matrix, and the calculation formula of the preliminary characteristic vector is: , all elements of the preliminary characteristic vector are divided by the sum to realize normalization, and the calculation formula of the weight of the fth index is: ； wherein, is the weight of the fthindex, is the sum of the preliminary eigenvectors, is the preliminary eigenvector; The calculation formula of the abnormal score is: ； wherein, is an anomaly score, with a larger value representing a better node quality, is an index weight of the preliminary pass index, is an index weight of the normalized reconstruction error.

10. A method of monitoring a power cable production stage according to claim 9, characterized in that: The overall qualified index is calculated based on the abnormal score, including the following specific steps: The calculation formula of the corrected node qualified index is: ； wherein, is the revised node eligibility index, is the preliminary eligibility index for the i-th node, is the monitoring weight for the i-th node, is the anomaly score, is the revision coefficient; The calculation formula of the overall qualified index is: ； wherein, is the overall eligibility index, is the monitoring weight of the ith node, is the modified node eligibility index.

Images