A method for monitoring a power cable production step

Abstract

The application discloses a kind of power cable production link monitoring methods, it is related to power cable technical field.The method steps include: according to execution order split production node, real-time acquisition each production node process parameter and image data;Using convolutional neural network correction image data, obtain correction image;Based on pre-stored reference feature, by semi-supervised semantic segmentation method extraction correction image feature, respectively correction image feature and process parameter feature with pre-set reference value are compared, and preliminary qualified index is calculated;Real-time feature vector is constructed and is input into graph neural network model, and output reconstruction feature vector and calculate real-time reconstruction error;By analytic hierarchy process, preliminary qualified index and normalized reconstruction error are weighted to obtain abnormal score, based on abnormal score, obtain overall qualified index and divide risk level.

Description

Technical Field

[0001] This invention relates to the field of power cable technology, specifically a monitoring method for the power cable production process. Background Technology

[0002] As the core carrier of power transmission systems, the production quality of power cables directly determines the safety and stability of power transmission. With the promotion of high-voltage and ultra-high-voltage power projects, the cable production process needs to cover multiple processes such as conductor extrusion, insulation wrapping, shielding braiding, and sheath forming. Fluctuations in the process at each stage may cause quality problems such as insulation breakdown and shielding failure.

[0003] The traditional monitoring method widely used in the industry is the power cable production monitoring method based on single-node process parameter thresholds. This method first breaks down the cable production process into independent production nodes, presets a benchmark threshold for process parameters for each production node, collects the process parameters of each production node in real time through sensors, and compares the real-time parameters 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 dependencies between production nodes. As a result, it cannot identify the hidden quality problems of subsequent production nodes caused by the slow drift of parameters of preceding production nodes, ultimately leading to the omission of potential quality risks in the production process and making it difficult to ensure the quality stability of the entire cable production process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a monitoring method for the power cable production process, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the production process of power cables, comprising the following steps: Step S1: The production process of the cable to be produced is broken down according to the execution sequence to obtain multiple production nodes, and the process parameters and image data of each production node are collected in real time. Step S2: Perform distortion correction on the image data using an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image; Step S3: Obtain the pre-stored benchmark image features and benchmark process parameters. Extract features from the calibration image using a semi-supervised image semantic segmentation method to obtain calibration image features. Compare the calibration image features with the benchmark image features to calculate the image comparison score. Compare the process parameters with the benchmark process parameters to calculate the parameter comparison score. Use a weighted summation method to calculate the image comparison score and parameter comparison score to obtain the preliminary qualification index for each production node. Step S4: Construct a real-time feature vector based on process parameters, calibration image and preliminary pass index, input the real-time feature vector into the graph neural network model, output the reconstructed feature vector, and calculate the real-time reconstruction error based on the reconstructed feature vector; 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.

[0007] The calculation of the overall pass rate based on the anomaly score includes the following specific steps: The revised formula for calculating the node qualification index is as follows:

[0008] in, This is the corrected node qualification index. Let be the preliminary qualification index for the i-th node. Let i be the monitoring weight of the i-th node. This is an anomaly score; a higher value indicates a better node quality. Correction coefficient

[0009] The formula for calculating the overall pass rate is:

[0010] in, The overall pass rate index. Let i be the monitoring weight of the i-th node. This is the corrected node qualification index.

[0011] Preferably, the production process of the cable to be produced is broken down into multiple production nodes according to the execution sequence, including the following specific steps: First, the core process chain of industrial production of power cables was sorted out and identified as four continuous production links: "conductor extrusion → insulation layer coating → shielding layer braiding → outer sheath coating". The process characteristics and quality impact weight of each link are directly related. According to the execution order, the links are broken down to the extrusion node, insulation node, braiding node, and sheath node.

[0012] Preferably, the step of correcting the image data using an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image includes the following steps: 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:

[0013] 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. This represents the tangential distortion parameter in the horizontal direction. This refers to the tangential distortion parameter in the vertical direction; 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, the grayscale value is calculated using quadratic linear interpolation.

[0014] 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; The morphological consistency of the corrected image is verified by extracting the cross-sectional contour of the cable using the Hough circle detection algorithm and calculating the roundness of the contour. If the roundness is ≥0.95, the correction is deemed qualified; if the roundness is <0.95, the distortion parameters are re-input into the CNN model for secondary prediction and correction until qualified.

[0015] Preferably, the step of extracting features from the corrected image using a semi-supervised image semantic segmentation method to obtain the corrected image features includes the following steps: Feature extraction of the corrected image is performed using a semi-supervised image semantic segmentation method to obtain the corrected image features. The specific steps are as follows: Set the confidence threshold to 0.8, and screen out regions with pixel values ≥ 0.8 as candidate prompt regions; sample 5 prompt points from the candidate prompt regions; Input the corrected image into the SAM model, optimize the SAM model through weight adjustment, and finally output 3 candidate masks. The calculation formula for the final mask is as follows:

[0016] where, is the final output pseudo-label mask of the key region, 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 are mask control weights; For the generated pseudo-label mask calculate the regional integrity index. If the mask coverage of the key region is ≥90%, the mask is determined to be qualified; if the coverage is <90%, regenerate the prompt points and perform secondary fine-tuning until it is qualified, so as to ensure that there is no region missing in the mask; Based on the qualified pseudo-label mask, extract features directly related to cable quality, which are divided into two categories: "visible light image features" and "infrared image features". The feature definitions, calculation methods and value bases are as follows: For visible light images, two types of features, namely insulation layer thickness and surface defect density, need to be calculated. The calculation formula for the insulation layer thickness is: , where, is the number of radial pixels of the insulation layer in the mask, is the camera calibration coefficient, with a value of 0.01mm / pixel, is the image scaling factor, is the insulation layer thickness; the calculation formula for the surface defect density is: , where, is the number of defect pixels in the mask, is the area of the key region, is the surface defect density. Defect pixels are identified by edge detection, ≤0.1 piece / cm² is qualified; For infrared images, the internal temperature standard deviation and another type of feature need to be calculated. The calculation formula for the internal temperature standard deviation is: , where, 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 acceptable; avoid localized overheating that could lead to insulation aging. The formula for calculating the percentage of hotspots in infrared images is: ,in, The percentage of hotspots in infrared images. The number of pixels at temperatures >220℃.

[0017] Preferably, the image comparison score is: The corrected image features are compared with the reference image features to calculate the image comparison score. The formula for calculating the image feature similarity score is as follows:

[0018] in, Similarity scores for individual image features. The feature values ​​of the current image. As the characteristic reference value, This is the attenuation coefficient, with a value of 5.

[0019] Preferably, the parameter comparison score is: The process parameters are compared with the baseline process parameters, and the parameter comparison score is calculated. The formula for calculating the parameter comparison score is as follows:

[0020] in, The parameter comparison score represents the overall qualification level of the process parameters at that node. This represents the number of process parameters for this node. The standardized value of the nth parameter. This is the baseline value for the nth parameter. This is the deviation tolerance threshold for the nth parameter.

[0021] Preferably, the step of calculating the image comparison score and parameter comparison score using a weighted summation method to obtain the preliminary qualification index for each production node includes the following specific steps: The image matching score and parameter matching score are calculated using a weighted summation method to obtain the preliminary pass index for each production node. The formula for calculating the preliminary pass index is as follows:

[0022] in, This is the preliminary qualification index for the production node. The parameter comparison weight is set to 0.6. For image comparison weights, For parameter comparison scores, Image comparison score.

[0023] Preferably, the calculation of real-time reconstruction error based on the reconstructed feature vector includes the following specific steps: During real-time monitoring, the real-time feature vector is input into the trained graph neural network model. Taking the current time as the endpoint, the sequence data of the previous four time steps is extracted to construct the real-time feature vector for each node. It includes process parameter sequences, image feature sequences, and preliminary pass indices. The real-time graph structure is input into the graph neural network model, which outputs the reconstructed feature vector for each node. Then, the real-time reconstruction error of each node is calculated. The formula for calculating the real-time reconstruction error is:

[0024] 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.

[0025] Preferably, the step of calculating the anomaly score by weighted summation using the analytic hierarchy process (AHP) based on the preliminary qualification index and real-time reconstruction error includes the following specific steps: First, the real-time reconstruction error is reverse-normalized to obtain the normalized reconstruction error, calculated using the following formula:

[0026] in, The normalized reconstruction error, 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.

[0027] The weights of the indicators are calculated using the analytic hierarchy process (AHP). First, the sum of the elements in each column is calculated. Then, each element of the judgment matrix A is divided by the sum of its column to obtain the normalized matrix. The preliminary eigenvectors are calculated using a normalized matrix. The formula for calculating the preliminary eigenvectors is as follows: , to the initial feature vector Divide all elements by their sum to achieve normalization. The formula for calculating the weight of the f-th indicator is:

[0028] in, Let f be the weight of the f-th indicator. This is the sum of the initial eigenvectors. This is the initial feature vector; The formula for calculating the anomaly score is:

[0029] in, This is an anomaly score; a higher value indicates a better node quality. The indicator weights for the preliminary qualification index, The index weights are used to normalize the reconstruction error.

[0030] This invention provides a monitoring method for the production process of power cables, which involves machine learning and deep learning technologies, and has the following beneficial effects: (1) An image distortion correction algorithm based on convolutional neural network is used to specifically solve the radial and tangential distortion of images caused by lens process errors and camera installation deviations in the power cable production scenario. By accurately predicting distortion parameters and completing the correction, the image distortion of key areas such as conductors and insulation layers is eliminated, ensuring the authenticity and reliability of image information in multimodal monitoring data, and laying a high-quality data foundation for the monitoring of the entire cable production process.

[0031] (2) The semi-supervised semantic segmentation method relies on the corrected distortion-free cable image to accurately extract key features such as conductor surface condition, insulation layer thickness, and braiding layer density without a large amount of manual annotation. This not only greatly reduces the manpower cost of data annotation for cable production monitoring, but also makes the feature extraction more in line with the actual quality status of cable production because the corrected image retains complete process details, making the calculation of the preliminary qualification index of each production node more valuable.

[0032] (3) By using a graph neural network model to identify anomalies in the cable production sequence, the limitations of traditional single-node threshold judgment are overcome. The process dependency relationship between nodes such as extrusion, insulation, braiding, and sheathing is accurately modeled. It can keenly capture hidden abnormal trends such as slow drift of production parameters, and provide early warning of potential quality risks such as insufficient cross-linking of insulation layer. It avoids the omission of anomalies caused by isolated monitoring of a single node, and improves the comprehensiveness and foresight of cable production anomaly identification.

[0033] (4) By combining the dynamic abnormal trends identified by the analytic hierarchy process with the graph neural network, and integrating the static quality indicators and process trend information of cable production, the preliminary qualification index is accurately corrected. This not only avoids the one-sidedness of single-dimensional scoring, but also makes the quality grade classification of the entire cable process more in line with the needs of different application scenarios such as high voltage and low voltage, significantly improving the reliability of monitoring results, providing accurate guidance for cable production process adjustment and rework of unqualified products, and ensuring the stability of power cable production quality. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of the steps of a monitoring method for the production process of power cables proposed in this invention; Figure 2 This is a step hierarchy diagram of obtaining the preliminary qualification index in a power cable production monitoring method proposed in this invention; Figure 3 This is a step hierarchy diagram of obtaining anomaly scores in a power cable production monitoring method proposed in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figures 1-3 The present invention provides a technical solution: a method for monitoring the production process of power cables.

[0038] Step S1: The production process of the cable to be produced is broken down according to the execution sequence to obtain multiple production nodes, and the process parameters and image data of each production node are collected in real time.

[0039] The production process of the cable to be produced is obtained and broken down according to the execution order to obtain multiple production nodes. The specific steps are as follows: First, the core process chain of industrial production of power cables was identified as four continuous production links: conductor extrusion, insulation layer coating, shielding layer braiding, and outer sheath coating. The process characteristics and quality impact weight of each link are directly related. According to the execution order, multiple production nodes were obtained.

[0040] Secondly, for each production node, multimodal monitoring items are assigned based on its process objectives: Extrusion node: The target is the conductor forming quality; the parameter monitoring items include extrusion temperature (180~220°), die head pressure (5~8MPa, to avoid dents on the conductor surface), and traction speed (0.3~0.5m / s, to match the extrusion efficiency); the image monitoring item is the visible light image of the conductor (resolution of 1920*1080, to capture the oxide layer and scratches on the conductor surface). Insulation nodes: The goal is to achieve the required insulation performance; monitoring parameters include coating temperature (200~230°C, to ensure full curing of the insulation material), insulation layer thickness (0.8-1.2mm), and curing time (30-40min), ensuring an insulation cross-linking degree ≥70%; image monitoring parameters include visible light images of the insulation layer (monitoring surface bubbles and impurities) and infrared images (resolution 384*288, monitoring internal temperature uniformity to avoid local overheating leading to insulation aging). Braided nodes: The goal is to ensure stable shielding performance; the parameters to be monitored include braided wire diameter (0.15-0.2mm, to ensure shielding effect) and braiding density (≥90%); the image monitoring item is the visible light image of the braided layer (to monitor broken wires and missing braids); Sheath details: The objectives are weather resistance and protection; monitoring parameters include sheath thickness (1.0-1.5mm) and extrusion temperature (190-210°); image monitoring includes visible light images of the sheath layer (monitoring color uniformity and surface cracks). Real-time acquisition of process parameters and image data at each production node, specifically implemented as follows: The temperature in the extrusion / sheathing stage is measured using a PT100 platinum resistance sensor (measurement range -50-300°C, accuracy ±0.5°C), the die head pressure is measured using a diffused silicon pressure sensor (measurement range 0~15MPa, accuracy ±0.1MPa), and the insulation / sheath thickness is measured using a laser thickness gauge (measurement range 0~5mm, accuracy ±0.01mm). All sensors are connected to the edge gateway via an RS485 bus, with a sampling frequency set to 1 time / second. Based on the cable production speed (0.3~0.5m / s), the cable movement distance within 1 second is ≤0.5m, ensuring that at least 2 parameter data points are collected per meter of cable to avoid data omission. Image acquisition equipment: Visible light cameras are 2-megapixel CMOS cameras, with a total of 4 units. One unit is deployed 0.5m downstream of the conductor extrusion die outlet, one unit is deployed 0.5m downstream of the insulation coating equipment outlet, one unit is deployed 0.5m downstream of the braiding machine outlet, and one unit is deployed 0.5m downstream of the sheath forming equipment outlet. Infrared cameras are uncooled infrared focal plane cameras (resolution 384*288, temperature measurement range -20-500°C), with a total of 2 units. One unit is deployed at the extrusion node outlet (to monitor the conductor extrusion temperature distribution), and one unit is deployed at the insulation node outlet. The camera deployment positions are as follows: installation height 1.5m above the cable surface (based on a lens focal length of 16mm, to ensure that a single frame image covers a 1m long cable area), camera lens axis perpendicular to the cable axis (to avoid tangential distortion), acquisition frequency set to 1 frame / 5 seconds, each of the 4 visible light cameras is assigned a unique hardware ID, and after the camera starts, the node ID of the corresponding node is automatically embedded in the acquired image file, 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 data cleaning: the parameter data uses the "sliding window filtering method" (window size 5, that is, the average of 5 consecutive sampling points is taken to replace the original point to eliminate impulse noise), and the image data uses "size normalization" (uniformly scaled to 360*360 pixels to reduce the amount of calculation for subsequent distortion correction). The preprocessed data is temporarily stored in the computer's local database.

[0041] Step S2: The image data is distorted using an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image.

[0042] In industrial settings, distortion in power cable images primarily stems from two factors: radial distortion (such as the "fisheye effect," manifested as stretched or compressed pixels at the cable edges) caused by lens manufacturing errors, and tangential distortion (manifested as an elliptical rather than circular cable cross-section) caused by the camera axis not being perpendicular to the cable axis. To specifically correct these two types of distortion, it is necessary to first construct a training set of cable images covering different degrees of distortion using MATLAB, and then train a dedicated CNN model. Specific steps are as follows: Using standard cable samples as the subject, 500 real distortion images were captured by adjusting the lens focal length (4~300mm, covering the commonly used focal length range in industry) and the camera tilt angle (0°~15°, simulating actual installation deviations). At the same time, based on the machine vision imaging principle, 11,500 synthetic images with different distortion parameters were generated by MATLAB simulation. The two types of images, totaling 12,000, constituted the training set (training set: validation set: test set = 8:1:1).

[0043] The distortion parameters of the synthesized image follow the distribution of actual industrial scenarios, and the specific value range is as follows: radial distortion parameter, , [ ], These are low-order radial distortion parameters. For higher-order radial distortion parameters, tangential distortion parameters [ ], Affects the horizontal tangential offset. It affects the vertical tangential offset.

[0044] Design a CNN model with 4 convolutional layers, 4 pooling layers, and 2 fully connected layers to adapt to the cable image distortion parameter prediction task. The specific structure and parameter values ​​are based on the following: The input layer is 1 layer with an input size of 360*360*1; the convolutional layer is 4 layers with a 3×3 kernel, which can reduce parameters while preserving edge features. The stride is 1, and the padding is "SAME" to avoid loss of edge pixels. It is adapted to the long strip features of cables. The number of output channels is 32→64→128→256.

[0045] The activation function consists of 4 layers, using the ReLU function to solve the gradient vanishing problem. The pooling layer also consists of 4 layers, using max pooling to preserve key image edge information. The pooling kernel is 2×s² with a stride of 2. The fully connected layer consists of 2 layers, with the first layer having 512 neurons and the second layer having 5 neurons. The 512-neuron layer adapts to the high-dimensional feature mapping of the image, while the 5-neuron layer corresponds to 5 distortion parameters output.

[0046] The loss function is the mean squared error (MSE) function. The distortion parameter prediction is a regression task, and MSE can accurately measure the deviation between the predicted value and the true value.

[0047] 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.

[0048] 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:

[0049] 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.

[0050] 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.

[0051] 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:

[0052] 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.

[0053] 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.

[0054] The morphological consistency of the corrected image is verified by extracting the cross-sectional profile of the cable using the Hough circle detection algorithm and calculating the roundness of the profile (roundness = 4π × 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.

[0055] 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.

[0056] 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.

[0057] Step S3: Obtain the pre-stored benchmark image features and benchmark process parameters. Extract features from the calibration image using a semi-supervised image semantic segmentation method to obtain calibration image features. Compare the calibration image features with the benchmark image features to calculate the image comparison score. Compare the process parameters with the benchmark process parameters to calculate the parameter comparison score. Use a weighted summation method to calculate the image comparison score and parameter comparison score to obtain the preliminary qualification index for each production node.

[0058] Process parameters were collected from 1,000 batches of qualified cables produced within the past year (3 samples were randomly selected from each batch, for a total of 3,000 samples). The "3σ principle" was used to construct the baseline process parameters for each parameter. This principle can cover 99.73% of normal production data and eliminate interference from extreme outliers. All baseline process parameters are stored in the baseline database of the industrial control computer.

[0059] 10% of the qualified cable images (300 images in total, covering four major production nodes) were selected, and senior quality inspectors manually labeled key areas such as conductors and insulation layers to generate realistic label masks as seed data for semi-supervised learning. The remaining 90% of unlabeled images were input into the SAM model, and class activation maps (confidence thresholds) were generated through Grad-CAM. =0.8), after filtering high-contribution pixel regions, five cue points are extracted using farthest-point sampling, input into SAM to generate pseudo-label masks, and masks with confidence scores ≥0.9 are selected to ensure accuracy: Based on real label masks and qualified pseudo-label masks, core image features are extracted, including the smoothness of the focused edge and defect density of visible light images, and the standard deviation of the focused internal temperature and the proportion of hot spots in infrared images. All features are stored in the format of "node-feature name-mean-standard deviation" to form a benchmark image feature library.

[0060] Feature extraction of the corrected image is performed using a semi-supervised image semantic segmentation method to obtain the corrected image features. The specific steps are as follows: The corrected image is input into a pre-trained image classification model (ResNet50, fine-tuned on a cable image dataset). The output is four categories: conductor, insulation, braid, and sheath. Grad-CAM is used to calculate the gradient weights of the target categories and generate class activation maps (CAMs). The higher the pixel value in the CAM, the greater the contribution of that location to the identification of the target category. For example, in the CAM of the insulation category, the pixel value of the insulation region is significantly higher than that of other regions (mean ≥ 180, background region ≤ 50).

[0061] Set 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".

[0062] 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:

[0063] 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.

[0064] 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.

[0065] It should be noted that the significance of this algorithm lies in maximizing the point spacing to ensure that the cue points cover the edges of key areas (such as the junction between the insulation layer and the conductor, and the surface of the insulation layer), providing the SAM model with "comprehensive and non-redundant" positioning information. Compared with randomly sampled cue points, the edge matching of the pseudo-label mask is improved.

[0066] The original SAM model has insufficient segmentation accuracy for "long, strip-shaped, highly textured" targets such as cables (edge ​​false negative rate ≥15%). SAM needs to be fine-tuned for cable image features. Specific implementation details are as follows: The SAM model consists of three parts: an image encoder (ViT-H), a cue encoder, and a mask decoder. The pre-trained weights of the image encoder (trained on the SA-1B dataset) are frozen, and only the mask control weights are trained. The three candidate masks output by the SAM are optimized through weight adjustment (the original SAM outputs three masks at different scales). The final mask calculation formula is as follows:

[0067] in, This serves as a pseudo-label mask for the key regions in the final output. For fine-scale candidate masks output by SAM, For medium-scale candidate masks output by SAM, For the coarse-scale candidate mask output by SAM, and Weights are controlled by the mask.

[0068] It should be noted that the pseudo-label mask is a binary image, where 1 represents the key region and 0 represents the background. The initial values ​​of the mask control weights are all 1 / 3, and they are obtained by minimizing the "intersection over union (IoU) loss between the mask and the artificial seed label".

[0069] For the generated pseudo-label mask Calculate the region integrity index. If the mask coverage of the critical region (such as the insulating layer) is ≥90% (coverage = number of 1 pixels in the mask / number of pixels in the actual critical region), the mask is deemed qualified. If the coverage is <90%, regenerate the prompt point (add 1 prompt point) and fine-tune it a second time until it is qualified, ensuring that the mask has no "regional missing" (such as missed detection of broken filament areas in the braided layer).

[0070] Based on a qualified fake tag mask, features directly related to cable quality are extracted, categorized into "visible light image features" and "infrared image features." The feature definitions, calculation methods, and value criteria are as follows: For visible light images, two types of features need to be calculated: insulation layer thickness and surface defect density. The formula for calculating insulation layer thickness is: ,in, This represents the radial pixel count of the insulating layer in the mask. This is the camera calibration factor, with a value of 0.01 mm / pixel. Image scaling factor Where is the thickness of the insulation layer; the formula for calculating the surface defect density is: ,in, This represents the number of defective pixels in the mask. For the key area area, The surface defect density is determined by edge detection (using the Canny operator). ≤0.1 pieces / It is acceptable.

[0071] For infrared images, it is necessary to calculate the internal temperature standard deviation and two types of features. The formula for calculating the internal temperature standard deviation is: ,in, For infrared image pixel temperature, The average temperature. Total number of pixels For internal temperature standard deviation, ≤3° is acceptable; avoid localized overheating that could lead to insulation aging. The formula for calculating the percentage of hotspots in infrared images is: ,in, The percentage of hotspots in infrared images. The number of pixels at temperatures >220℃.

[0072] It should be noted that the significance of feature extraction lies in transforming the "qualitative features" of an image into "quantitative indicators," providing a unified dimension for subsequent calculation of the qualification index by integrating with process parameters.

[0073] The corrected image features are compared with the reference image features to calculate the image comparison score. The formula for calculating the image feature similarity score is as follows:

[0074] in, Similarity scores for individual image features. The feature values ​​of the current image. As the characteristic reference value, This is the attenuation coefficient, with a value of 5.

[0075] The process parameters are compared with the baseline process parameters, and the parameter comparison score is calculated. The formula for calculating the parameter comparison score is as follows:

[0076] in, The parameter comparison score represents the overall qualification level of the process parameters at that node. This represents the number of process parameters for this node. The standardized value of the nth parameter. This is the baseline value for the nth parameter. This is the deviation tolerance threshold for the nth parameter.

[0077] It should be noted that the benchmark value is the median of the benchmark interval, and the deviation tolerance threshold is 10% of the width of the benchmark interval.

[0078] For all image features at each production node, the average value is taken as the image comparison score. .

[0079] The image matching score and parameter matching score are calculated using a weighted summation method to obtain the preliminary pass index for each production node. The formula for calculating the preliminary pass index is as follows:

[0080] in, This is the preliminary qualification index for the production node. The parameter comparison weight is set to 0.6. For image comparison weights, For parameter comparison scores, Image comparison score.

[0081] It should be noted that, ≥0.8 indicates preliminary pass. A value less than 0.8 indicates that the test result is initially unqualified.

[0082] Step S4: Construct a real-time feature vector based on process parameters, calibration image and preliminary pass index, input the real-time feature vector into the graph neural network model, and output the reconstructed feature vector. Calculate the real-time reconstruction error based on the reconstructed feature vector.

[0083] Cable production is a continuous process of "extrusion → insulation → braiding → sheathing". The process status of the preceding node (such as extrusion) directly affects the quality of the subsequent node (such as insulation) (e.g., low extrusion temperature will lead to insufficient cross-linking 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: "node feature definition" and "edge weight definition". GNN nodes need to contain dynamic sequence information (capturing trends) and static feature information (reflecting the current state). Real-time feature vectors are constructed based on process parameters, image data, and preliminary pass / fail indices. The calculation formula is as follows:

[0084] in, Let be the real-time feature vector of the i-th production node. This is a sequence fragment of process parameters for the i-th node. For the corrected image feature sequence segment of the i-th node, Let be the preliminary qualification index for the i-th node. This is a vector concatenation operation.

[0085] It should be noted that the process parameter sequence segment and the correction image feature sequence segment are taken from the standardized parameter values ​​and image feature values ​​of the current time and the previous four time steps, respectively, for a total of five time steps. The time step interval is 1 second. With a time step interval of 5 seconds, the sequence length of 5 time steps can cover the range of 2.5 to 5 meters in cable production, which is sufficient to capture the trend of slow parameter drift.

[0086] Based on the power cable production process, a topological graph representing the technological relationships between production nodes is constructed. The edges of the Generative Neural Network (GNN) represent the technological dependencies between production nodes. Only directed edges between "preceding nodes → subsequent nodes" (e.g., extrusion → insulation, insulation → braiding, braiding → sheathing) are retained, with no reverse edges (since subsequent nodes do not affect preceding nodes). The formula for calculating edge weights is as follows:

[0087] in, Let be the edge weight from the previous node i to the next node j. This represents the number of times in historical production that node j became abnormal due to node i's abnormality. Let be the total number of anomalies for node j.

[0088] Furthermore, to adapt to process fluctuations, the edge weights during GNN training can be adaptively fine-tuned through the graph attention layer (the initial weights are the calculated values ​​mentioned above, and are updated during training through attention coefficients). The fine-tuning formula for the attention coefficients is as follows:

[0089] in, The attention coefficients are fine-tuned, reflecting the strength of the influence of node i on j during the model's adaptive learning. The characteristic transformation matrix, This is the attention vector, used to calculate the similarity between node pairs. Let j be the set of neighboring nodes. For activation function, Let represent the index of the neighboring node of node j.

[0090] It should be noted that, through the attention mechanism, the model can dynamically adjust the edge weights, avoiding misjudgments of anomalies caused by fixed weights, thereby improving the robustness of anomaly identification.

[0091] The core objective of the GNN model is to learn the feature association patterns of nodes under normal production conditions. Anomalies are then identified by the deviation between the current sequence and the normal pattern. Therefore, the training data must be historical, fault-free production sequence data. The specific training process is as follows: We collected cable sequence data from 500 batches of normally produced cables within the past year (each batch containing 4 nodes, with 500 time steps per node as feature vectors), and divided them into training, validation, and test sets in a 7:2:1 ratio. Each batch of data was then converted into the input graph structure of a GNN. ,in, This indicates 4 production nodes. Let represent 3 directed edges, and the feature vector of each node i is the feature vector at that time step. .

[0092] We selected a GNN structure consisting of two graph attention layers and one fully connected layer. The specific configuration and hyperparameter values ​​are based on the following: The input dimension of the input layer is 31, used for matching. To ensure untruncated features, the graph attention layer has two layers, and the hidden layer has a dimension of 64 to balance expressive power and computational cost. There are four attention heads, a dropout rate of 0.5, and the fully connected layer has an output dimension of 31. The Adam optimizer is used, with an initial learning rate of [missing value]. The weight decays to The training batch size is 32, and the number of iterations is 200.

[0093] The loss function uses the Node Feature Reconstruction Error (MSE), and the formula for calculating the loss function is as follows:

[0094] in, The total loss of the model, For training batch size, For the number of production nodes, The number of time steps per batch. Let i be the true feature vector of the b-th batch, the t-th time step, and node i. The reconstructed feature vector output by the model. It is the L2 norm, used to measure the difference between the real and reconstructed vectors.

[0095] During real-time monitoring, the real-time feature vector is input into the trained GNN model. Taking the current time as the endpoint, the sequence data of the previous four time steps are extracted to construct the real-time feature vector for each node. It includes process parameter sequences, image feature sequences, and preliminary pass indices. The real-time graph structure is input into the GNN, and the model outputs the reconstructed feature vector of each node. Then, the real-time reconstruction error of each node is calculated. The formula for calculating the real-time reconstruction error is:

[0096] 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.

[0097] 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.

[0098] 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.

[0099] 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:

[0100] 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.

[0101] 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 weights of the three types of indicators, avoiding subjective assignment. First, a hierarchical structure is constructed: the target layer is "node anomaly scoring," and the criterion layer consists of the three types of indicators (…). Preliminary pass index (Normalized real-time reconstruction error), the solution layer consists of each production node (extrusion, insulation, braiding, sheathing).

[0102] Five cable process experts were invited to conduct pairwise comparisons of the criteria-level indicators based on the "importance of indicators to quality assessment." A 1-9 scale was used (1 indicates equal importance, 3 indicates slightly more important than the latter; scale 5 indicates significantly more important than the latter; scale 7 indicates strongly more important than the latter; scale 9 indicates extremely more important than the latter; scales 2, 4, 6, and 8 correspond to intermediate states between adjacent scales). The average of the expert scores was used to construct a judgment matrix A.

[0103] in, for relatively The scale mean.

[0104] It should be noted that, =1.5 indicates a preliminary pass rate. Normalized reconstruction error Slightly important.

[0105] The index weights are calculated using the eigenvector method. First, the sum of the elements in each column is calculated. Then, each element of the judgment matrix A is divided by the sum of its column to obtain the normalized matrix. The preliminary eigenvectors are calculated using a normalized matrix. The formula for calculating the preliminary eigenvectors is as follows: , to the initial feature vector Divide all elements by their sum to achieve normalization. The formula for calculating the weight of the f-th indicator is:

[0106] in, Let f be the weight of the f-th indicator. This is the sum of the initial eigenvectors. These are the initial feature vectors.

[0107] The formula for calculating the anomaly score is:

[0108] in, This is an anomaly score; a higher value indicates a better node quality. The indicator weights for the preliminary qualification index, The index weights are used to normalize the reconstruction error.

[0109] The overall pass rate is calculated based on the anomaly score. The specific steps are as follows: The preliminary qualification index does not take into account the dynamic trend of the sequence and needs to be corrected by combining anomaly scoring. At the same time, the root cause of the anomaly should be located and the overall qualification index should be calculated to provide a basis for production adjustments.

[0110] Based on the quality requirements of industrial cables, the anomaly scores are divided into three levels, each corresponding to a different correction factor, to ensure that the corrected index closely matches the actual quality level. The weights of each monitoring node are determined based on the "fault impact weighting method": the proportion of faults in each stage is recorded according to the statistical data of production fault records over the past 3 years. For example, based on the production failure records of a cable factory over the past three years, statistical data shows that the failure percentages for each stage are as follows: insulation layer 42%, sheath 30%, extrusion 18%, and braiding 10%. Converting these failure percentages into monitoring weights for each stage, the formula relating failure percentages to monitoring weights is: ,in, Let i be the monitoring weight of the i-th node. To determine the failure rate of the i-th stage, the monitoring weight calculation is rounded to one decimal place, resulting in the monitoring weight for each production node: Extrusion Node =0.2, Insulating Node =0.4, weaving nodes =0.1, Sheath Node =0.3, this weight distribution intuitively reflects the importance of each node in the monitoring of cable production quality. Among them, the insulation node (0.4) is the core monitoring object because of the highest proportion of faults, followed by the sheath node (0.3), and the extrusion and braiding nodes are assigned corresponding weights according to the degree of fault impact.

[0111] The revised formula for calculating the node qualification index is as follows:

[0112] in, This is the corrected node qualification index. Let be the preliminary qualification index for the i-th node. Let i be the monitoring weight of the i-th node. For abnormal scoring, This is a correction factor.

[0113] A score of ≥0.95 is considered excellent, representing good static node quality, normal dynamic trends, and matching visual features. A score of 0.8≤ A value <0.95 indicates that the node shows no obvious abnormalities but exhibits slight fluctuations. A value less than 0.8 indicates that the node has an abnormal risk.

[0114] The formula for calculating the overall pass rate index is:

[0115] in, The overall pass rate index. Let i be the monitoring weight of the i-th node. This is the corrected node qualification index.

[0116] according to Cables are classified into three grades, corresponding to different application scenarios, based on industry quality standards and customer needs: The preliminary pass index is corrected by adjusting the anomaly scores to obtain the corrected pass index: ≥0.95, suitable for high-voltage transmission scenarios, requiring all nodes ≥0.9, no abnormalities; 0.85≤ <0.95, suitable for low-voltage power distribution scenarios, allows one low-weight node, but requires no serious anomalies; <0.85, prohibited from leaving the factory, requires rework until... ≥0.85.

[0117] If the previous node If the value is less than 0.8 (unacceptable), the industrial control computer sends a "stop signal" to the production equipment, prohibiting the execution of the next production node. After the stop, workers adjust the process based on the anomaly location results and re-execute the production and monitoring of the unacceptable node until... ≥0.8, if a node is reworked 3 times consecutively If the value is less than 0.8, an "equipment maintenance warning" will be triggered, prompting the user to check the equipment and prevent continued non-compliance.

[0118] This application proposes a monitoring method for the production process of power cables. By breaking down production nodes and collecting multimodal data, a full-process quality monitoring system is constructed through CNN image distortion correction, semi-supervised semantic segmentation feature extraction, GNN sequence anomaly identification, and multimodal fusion anomaly scoring. This solves the problems of data distortion, inaccurate feature extraction, anomaly omission, and biased scoring in traditional monitoring, and provides comprehensive and accurate technical support for the quality control of cable production.

[0119] The image distortion correction algorithm based on convolutional neural networks is designed to address radial and tangential distortions in power cable production caused by lens manufacturing errors and camera installation deviations. By accurately predicting distortion parameters and performing corrections, it eliminates image distortion in key areas such as conductors and insulation layers, ensuring the authenticity and reliability of image information in multimodal monitoring data and laying a high-quality data foundation for monitoring the entire cable production process.

[0120] Semi-supervised semantic segmentation relies on the corrected, distortion-free cable images to accurately extract key features such as conductor surface condition, insulation layer thickness, and braid density without extensive manual annotation. This significantly reduces the manpower cost of data annotation for cable production monitoring. Furthermore, because the corrected images retain complete process details, the feature extraction is more closely aligned with the actual quality status of cable production, making the calculation of the preliminary qualification index for each production node more valuable.

[0121] By using a graph neural network model to identify anomalies in the cable production sequence, this method overcomes the limitations of traditional single-node threshold judgment. It accurately models the process dependencies between nodes such as extrusion, insulation, braiding, and sheathing, and can keenly capture hidden abnormal trends such as slow drift of production parameters. It can also provide early warnings of potential quality risks such as insufficient cross-linking of the insulation layer, avoiding missed anomalies caused by isolated monitoring of a single node, and improving the comprehensiveness and foresight of anomaly identification in cable production.

[0122] By combining the analytic hierarchy process (AHP) with graph neural networks to identify dynamic anomalies, and integrating static quality indicators and process trend information from cable production, the preliminary qualification index is precisely corrected. This avoids the one-sidedness of single-dimensional scoring and makes the quality grade classification of cables at all stages more in line with the needs of different application scenarios such as high voltage and low voltage. This significantly improves the reliability of monitoring results, provides precise guidance for adjusting cable production processes and reworking non-conforming products, and ensures the stability of power cable production quality.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A monitoring method for the production process of power cables, characterized in that: Includes the following steps: Step S1: The production process of the cable to be produced is broken down according to the execution sequence to obtain multiple production nodes, and the process parameters and image data of each production node are collected in real time. Step S2: Perform distortion correction on the image data using an image distortion correction algorithm based on a convolutional neural network to obtain a corrected image; Step S3: Obtain the pre-stored benchmark image features and benchmark process parameters. Extract features from the calibration image using a semi-supervised image semantic segmentation method to obtain calibration image features. Compare the calibration image features with the benchmark image features to calculate the image comparison score. Compare the process parameters with the benchmark process parameters to calculate the parameter comparison score. Use a weighted summation method to calculate the image comparison score and parameter comparison score to obtain the preliminary qualification index for each production node. Step S4: Construct a real-time feature vector based on process parameters, calibration image and preliminary pass index, input the real-time feature vector into the graph neural network model, output the reconstructed feature vector, and calculate the real-time reconstruction error based on the reconstructed feature vector; 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. The calculation of the overall pass rate based on the anomaly score includes the following specific steps: The revised formula for calculating the node qualification index is as follows: ； in, This is the corrected node qualification index. Let be the preliminary qualification index for the i-th node. Let i be the monitoring weight of the i-th node. This is an anomaly score; a higher value indicates a better node quality. This is a correction factor; The formula for calculating the overall pass rate is: ； in, The overall pass rate index. Let i be the monitoring weight of the i-th node. This is the corrected node qualification index.

2. The monitoring method for the production process of power cables according to claim 1, characterized in that: The production process of the cable to be produced is broken down into multiple production nodes according to the execution sequence, including the following specific steps: First, the core process chain of industrial production of power cables was identified as four continuous production links: conductor extrusion, insulation layer coating, shielding layer braiding, and outer sheath coating. The process characteristics and quality impact weight of each link are directly related. The links were broken down according to the execution order, resulting in the extrusion node, insulation node, braiding node, and sheath node.

3. The monitoring method for the production process of power cables according to claim 2, characterized in that: The process of correcting image data using a convolutional neural network-based image distortion correction algorithm to obtain a corrected image includes the following steps: 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: ； 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. This represents the tangential distortion parameter in the horizontal direction. This refers to the tangential distortion parameter in the vertical direction; 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, the grayscale value is calculated using quadratic linear interpolation. ； 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; The morphological consistency of the corrected image is verified by extracting the cross-sectional contour of the cable using the Hough circle detection algorithm and calculating the roundness of the contour. If the roundness is ≥0.95, the correction is deemed qualified; if the roundness is <0.95, the distortion parameters are re-input into the CNN model for secondary prediction and correction until qualified.

4. The monitoring method for the production process of power cables according to claim 3, characterized in that: The step of extracting features from the corrected image using a semi-supervised image semantic segmentation method to obtain the corrected image features includes the following steps: Feature extraction of the corrected image is performed using a semi-supervised image semantic segmentation method to obtain the corrected image features. The specific steps are as follows: Set a confidence threshold of 0.8 and filter out regions with pixel values ​​≥ 0.8 as candidate prompt regions; sample 5 prompt points from the candidate prompt regions; The corrected image is input into the SAM model, and the SAM model is optimized by weight adjustment. Finally, three candidate masks are output. The final mask calculation formula is as follows: ； in, This serves as a pseudo-label mask for the key areas in the final output. For fine-scale candidate masks output by SAM, For medium-scale candidate masks output by SAM, For the coarse-scale candidate mask output by SAM, and Weights are controlled by the mask; For the generated pseudo-label mask Calculate the region integrity index. If the mask coverage rate of the critical region is ≥90%, the mask is deemed qualified. If the coverage rate is <90%, the prompt points are regenerated and fine-tuned again until qualified, ensuring that there are no missing areas in the mask. Based on a qualified pseudo-tag mask, features directly related to cable quality are extracted, categorized into two types: visible light image features and infrared image features. The feature definitions, calculation methods, and value criteria are as follows: For visible light images, two types of features need to be calculated: insulation layer thickness and surface defect density. The formula for calculating insulation layer thickness is: ,in, This represents the radial pixel count of the insulating layer in the mask. This is the camera calibration factor, with a value of 0.01 mm / pixel. Image scaling factor Where is the thickness of the insulation layer; the formula for calculating the surface defect density is: ,in, This represents the number of defective pixels in the mask. For the key area area, Surface defect density; For infrared images, it is necessary to calculate the internal temperature standard deviation and two types of features. The formula for calculating the internal temperature standard deviation is: ,in, For infrared image pixel temperature, The average temperature. Total number of pixels For internal temperature standard deviation, ≤3° is considered acceptable to avoid localized overheating that could lead to insulation aging; the formula for calculating the percentage of hotspots in infrared images is: ,in, The percentage of hotspots in infrared images. The number of pixels at temperatures >220℃.

5. A monitoring method for the production process of power cables according to claim 4, characterized in that: The image comparison score is: The corrected image features are compared with the reference image features to calculate the image comparison score. The formula for calculating the image feature similarity score is as follows: ； in, Similarity scores for individual image features. The feature values ​​of the current image. As the characteristic reference value, This is the attenuation coefficient, with a value of 5.

6. A monitoring method for the production process of power cables according to claim 5, characterized in that: The parameter comparison score is: The process parameters are compared with the baseline process parameters, and the parameter comparison score is calculated. The formula for calculating the parameter comparison score is as follows: ； in, The parameter comparison score represents the overall qualification level of the process parameters at that node. This represents the number of process parameters for this node. The standardized value of the nth parameter. This is the baseline value for the nth parameter. This is the deviation tolerance threshold for the nth parameter.

7. A monitoring method for the production process of power cables according to claim 6, characterized in that: The process of calculating the image matching score and parameter matching score using a weighted summation method to obtain the preliminary qualification index for each production node includes the following specific steps: The image matching score and parameter matching score are calculated using a weighted summation method to obtain the preliminary pass index for each production node. The formula for calculating the preliminary pass index is as follows: ； in, This is the preliminary qualification index for the production node. The parameter comparison weight is set to 0.

6. For image comparison weights, For parameter comparison scores, Image comparison score.

8. A monitoring method for the production process of power cables according to claim 7, characterized in that: The calculation of real-time reconstruction error based on the reconstructed feature vector includes the following specific steps: During real-time monitoring, the real-time feature vector is input into the trained graph neural network model. Taking the current time as the endpoint, the sequence data of the previous four time steps is extracted to construct the real-time feature vector for each node. It includes process parameter sequences, image feature sequences, and preliminary pass indices. The real-time graph structure is input into the graph neural network model, which outputs the reconstructed feature vector for each node. Then, the real-time reconstruction error of each node is calculated. The formula for calculating the real-time reconstruction error is: ； 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.

9. A monitoring method for the production process of power cables according to claim 8, characterized in that: The anomaly score is calculated by weighted summation using the analytic hierarchy process (AHP) based on the preliminary qualification index and real-time reconstruction error, including the following specific steps: First, the real-time reconstruction error is reverse-normalized to obtain the normalized reconstruction error, calculated using the following formula: ； in, The normalized reconstruction error, 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. The weights of the indicators are calculated using the analytic hierarchy process (AHP). First, the sum of the elements in each column is calculated. Then, each element of the judgment matrix A is divided by the sum of its column to obtain the normalized matrix. The preliminary eigenvectors are calculated using a normalized matrix. The formula for calculating the preliminary eigenvectors is as follows: , to the initial feature vector Divide all elements by their sum to achieve normalization. The formula for calculating the weight of the f-th indicator is: ； in, Let f be the weight of the f-th indicator. This is the sum of the initial eigenvectors. This is the initial feature vector; The formula for calculating the anomaly score is: ； in, For abnormal scoring, The indicator weights for the preliminary qualification index, The index weights are used to normalize the reconstruction error.

Images