A cable identification method and system based on image processing

Abstract

The present invention relates to a cable identification method and system based on image processing. The cable identification method based on image processing includes: collecting and processing cross-sectional images of cable insulation materials; performing image segmentation and defect detection on the processed cross-sectional images, identifying and marking defects in the images, and locating the starting point of water tree phenomena according to the marked defects; establishing a water tree phenomenon generation and development prediction model according to the starting point of the water tree phenomenon; using adaptive optical compensation methods and image enhancement methods to dynamically adjust imaging parameters in real time; using a multi-level resolution image acquisition strategy to dynamically adjust image resolution; collecting original cross-sectional image data and performing preprocessing; real-time monitoring and analysis of image data and electrical stress data to identify potential water tree phenomena and insulation material defects. The cable identification method based on image processing of the present invention greatly improves the accuracy and efficiency of defect detection, and also greatly enhances the ability to prevent and deal with water tree phenomena.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a cable identification method and system based on image processing. Background Art

[0002] There is a key technical contradiction in the online monitoring of cable insulation materials. The water tree phenomenon refers to the phenomenon that a dendritic conductive channel is gradually formed in the insulating medium of high-voltage power equipment due to the combined effects of factors such as uneven local electric field, moisture, impurities, and bubbles. This process is often slow. As time goes by, the water tree will gradually grow and spread, which may eventually lead to a decrease in insulation performance or even insulation breakdown, causing equipment failure.

[0003] In order to detect potential causes of water treeing such as tiny bubbles and impurities at an early stage, high-resolution infrared imaging technology is required for close-range and precise scanning. This requires the equipment to have extremely high imaging accuracy and be able to clearly distinguish micron-level details, but this also puts enormous pressure on data processing and storage capabilities. Secondly, monitoring also requires simultaneous real-time monitoring of the electric field inside the cable to evaluate the concentration of electrical stress at defect points, which increases the complexity of the monitoring system. Furthermore, considering the wide geographical distribution of cable laying, large-scale monitoring tasks conflict with the inefficiency of close-range and precise inspections. Finally, the impact of external environmental factors (such as haze and dust) and the equipment's own interference on imaging quality further restricts the accuracy and reliability of monitoring.

[0004] Therefore, how to comprehensively improve the performance of monitoring equipment, optimize data processing procedures, improve image quality and effectively suppress environmental interference to achieve a high-precision and high-efficiency online cable insulation monitoring system is a key technical issue that urgently needs to be broken through. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a cable identification method and system based on image processing. The cable identification method based on image processing greatly improves the accuracy and efficiency of defect detection, and also greatly enhances the ability to prevent and deal with water tree phenomena; it not only helps to protect the safe operation of the power system, but also reduces the maintenance costs and potential economic losses caused by defects.

[0006] The present invention provides a cable identification method and system based on image processing, comprising:

[0007] S1, collecting and processing cross-sectional images of cable insulation materials;

[0008] S2, performing image segmentation and defect detection on the processed cross-sectional image, identifying and marking defects in the image, and locating the starting point of the water tree phenomenon according to the marked defects;

[0009] S3. According to the starting point of the water tree phenomenon, combined with the electrical stress dynamic monitoring data and the influence of various factors on the development of the water tree phenomenon, a water tree phenomenon generation and development prediction model is established;

[0010] S4, using adaptive optical compensation method and image enhancement method to dynamically adjust imaging parameters in real time;

[0011] S5. Adopt a multi-resolution image acquisition strategy, combine the importance of the cable area and historical water tree risk data, and dynamically adjust the image resolution;

[0012] S6, collecting original cross-sectional image data according to real-time dynamic adjustment of imaging parameters and dynamic adjustment of image resolution, and performing preprocessing;

[0013] S7. Real-time monitoring and analysis of the image data and electrical stress data, combined with the cross-sectional image data pre-processed in step S6, to identify potential water tree phenomena and insulation material defects.

[0014] Preferably, the step S1 specifically includes:

[0015] Using an infrared image acquisition device, setting and adjusting parameters of the infrared image acquisition device, and acquiring cross-sectional image data covering the entire cross-sectional area of ​​the cable insulation material;

[0016] Processing the collected cross-sectional image data, including removing salt and pepper noise and removing Gaussian noise in the image;

[0017] According to the characteristics of the cross section of the cable insulation material, the cross section image is segmented and the image data of the area of ​​interest is extracted;

[0018] The segmented cross-sectional images are corrected to eliminate image distortion.

[0019] Preferably, the step S2 specifically includes:

[0020] A deep learning algorithm is used to perform semantic segmentation on the processed cross-sectional images and a segmentation model is established;

[0021] By extracting features and classifying pixels, the image is divided into defective areas and normal insulating material areas. The defect types in the defective areas include bubbles, impurities, cracks and deformations.

[0022] Using data enhancement and transfer learning methods, the segmentation model is pre-trained according to the existing defect sample data set;

[0023] According to the segmentation results, the connected domain analysis algorithm based on region growing is used to mark and number various defect regions. The contour curve of each defect region is obtained by boundary tracing, and the area, perimeter and roundness shape feature parameters of the defect region are calculated using geometric shape descriptors.

[0024] According to the defect type identification rules, the defect area is classified, the defect type is identified, and the severity of the defect is evaluated and graded according to the area threshold and roundness threshold;

[0025] The DBSCAN density clustering algorithm is used to perform cluster analysis on the marked defect areas. Adjacent defects are classified into defect groups according to the spatial distance and distribution density between defects. The morphological skeleton extraction algorithm is used to obtain the skeleton curve of each defect group as the center line of the defect group. The topological network diagram of defect distribution is constructed to reveal the expansion trend and connection path of defects on the cross section of the insulating material.

[0026] Graph theory analysis is performed on the defect distribution topology map. All defect extension paths are traversed using a depth-first search algorithm. The defect area growth rate, path length, and number of bifurcations are used to score the degree of defect evolution of each path, thereby locating the starting point of the water tree phenomenon.

[0027] Preferably, the step S3 specifically includes:

[0028] Various factors affect the development of water tree phenomena including different voltages, frequencies and ambient temperatures;

[0029] Collect electrical stress data during cable operation, combine it with static data of cable lines, and use data cleaning and feature engineering methods to preprocess the data to obtain a multi-source heterogeneous monitoring data set.

[0030] Obtain the temperature data of the cable environment, combine the cable's geographical location information and meteorological data, build a spatiotemporal distribution model of the cable environment temperature, and incorporate the temperature distribution characteristics into the feature extraction process of the water tree defect area;

[0031] The water tree starting point identification result is associated with the electrical stress monitoring data and the ambient temperature data, the time-frequency domain characteristics of the defect area centered on the water tree starting point are extracted, and the multi-dimensional feature vector of the water tree defect area is constructed;

[0032] By using time series analysis and association rule mining algorithm, combined with the association law and evolution trend of the characteristic parameters of the water tree defect area, the quantitative relationship between voltage, frequency, temperature and water tree development is obtained, and a water tree phenomenon generation and development prediction model based on time series regression is established.

[0033] Preferably, the step S3 further includes:

[0034] Electrical stress sensors are set up to monitor stress changes inside and outside the cable in real time, and the stress change data is correlated with the identified defect locations for analysis, a stress distribution model is established, and predetermined risk stress characteristics that lead to water treeing are identified;

[0035] A parallel computing analysis method is used to evaluate the potential risk of each defect location in combination with the defect location and the predetermined risk stress characteristics. A water tree risk assessment model is constructed to divide the water tree phenomenon in each area into risk levels.

[0036] Preferably, the step S4 further comprises:

[0037] According to the adaptive optical compensation method, the exposure time, aperture size and image enhancement parameters are automatically adjusted in combination with changes in ambient light and imaging conditions.

[0038] Preferably, the step S5 specifically includes:

[0039] The cable line nodes are divided into three levels: important nodes, general nodes and auxiliary nodes. The hierarchical analysis method is used to comprehensively evaluate the importance of nodes. A node importance evaluation system is constructed, which includes operating voltage level, transmission power, power supply range, geographical location and water tree fault history indicators. The importance weight of each node is obtained through indicator weight solution and consistency test. The clustering algorithm is used to classify the nodes to form the node importance division result.

[0040] Preferably, the step S6 specifically includes:

[0041] An edge computing device is arranged at a position matching the infrared image acquisition device;

[0042] The edge computing device acquires the original cross-sectional image data acquired by the infrared image acquisition device, and preprocesses the original cross-sectional image data, wherein the preprocessing includes denoising, image enhancement, and resolution adjustment;

[0043] Defect detection is performed on the preprocessed cross-sectional images to identify the starting point and extension direction of the water tree phenomenon in the insulating material.

[0044] Preferably, the step S7 further includes:

[0045] Dynamically assess the development trend of the water tree phenomenon based on the water tree risk assessment model and real-time monitoring data;

[0046] Adopt an early warning strategy and set an early warning threshold. If the real-time monitoring data does not exceed the early warning threshold, the original state is maintained. If the real-time monitoring data exceeds the early warning threshold, an early warning signal is issued.

[0047] The risk level is obtained according to the water tree risk assessment model, and combined with the development trend of the water tree phenomenon, it is determined whether there is a problem with the water tree phenomenon. If not, the original state is maintained. If so, the corresponding emergency response measures are initiated according to the risk level.

[0048] The present invention further proposes a cable identification system based on image processing, which adopts the cable identification method based on image processing as described above, and comprises:

[0049] An image processing detection module, which is used for denoising, adjusting resolution and identifying defects in cable insulation materials;

[0050] An image analysis module, which is used to segment the processed image and mark the defect location, size and shape;

[0051] A water tree phenomenon prediction and risk assessment module, which is used to predict and assess the water tree risk level in combination with dynamic monitoring data;

[0052] The monitoring and emergency response module is used to issue early warnings based on risk levels, dynamically manage the cable identification system and trigger emergency measures.

[0053] The cable identification method and system based on image processing described in the present invention have the following advantages:

[0054] The cable identification method based on image processing described in the present invention includes reducing the interference of environmental factors on image quality in real time, and dynamically adjusting the image acquisition strategy to cope with the importance and historical risk data of different areas; adopting a multi-level resolution image acquisition strategy, by dynamically adjusting the imaging parameters in real time, effectively reducing the interference of haze and smoke on image quality, and dynamically adjusting the image resolution according to the importance of the cable area and the historical water tree risk data, optimizing the image acquisition effect; by preprocessing the collected image, the preprocessed image clearly shows the defects in the insulating material; using big data analysis and machine learning algorithms, not only can the characteristics of the defect area be analyzed, but also the water tree risk level of the area can be predicted, and the development trend of the water tree phenomenon can be dynamically evaluated according to the prediction model and real-time monitoring data. Therefore, the cable identification method based on image processing greatly improves the accuracy and efficiency of defect detection, and also greatly enhances the ability to prevent and cope with the water tree phenomenon; it not only helps to protect the safe operation of the power system, but also reduces the maintenance cost and potential economic losses caused by defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a cable identification method based on image processing described in the present invention. DETAILED DESCRIPTION

[0056] like Figure 1 As shown, a cable identification method based on image processing according to the present invention comprises:

[0057] S1, collecting and processing cross-sectional images of cable insulation materials;

[0058] S2, performing image segmentation and defect detection on the processed cross-sectional image, identifying and marking defects in the image, and locating the starting point of the water tree phenomenon according to the marked defects;

[0059] S3. According to the starting point of the water tree phenomenon, combined with the electrical stress dynamic monitoring data and the influence of various factors on the development of the water tree phenomenon, a water tree phenomenon generation and development prediction model is established;

[0060] S4, using adaptive optical compensation method and image enhancement method to dynamically adjust imaging parameters in real time;

[0061] S5. Adopt a multi-resolution image acquisition strategy, combine the importance of the cable area and historical water tree risk data, and dynamically adjust the image resolution;

[0062] S6, collecting original cross-sectional image data according to real-time dynamic adjustment of imaging parameters and dynamic adjustment of image resolution, and performing preprocessing;

[0063] S7. Real-time monitoring and analysis of image data and electrical stress data, combined with the cross-sectional image data pre-processed in step S6, to identify potential water tree phenomena and insulation material defects.

[0064] Furthermore, in this embodiment, step S1 specifically includes:

[0065] Using an infrared image acquisition device, setting and adjusting parameters of the infrared image acquisition device, and acquiring cross-sectional image data covering the entire cross-sectional area of ​​the cable insulation material;

[0066] Processing the collected cross-sectional image data includes removing salt and pepper noise and Gaussian noise in the image; using a median filter algorithm to remove the salt and pepper noise in the image, and using a wavelet transform to remove the Gaussian noise, so as to improve the signal-to-noise ratio of the image;

[0067] According to the characteristics of the cross section of the cable insulation material, the cross section image is segmented and the image data of the area of ​​interest is extracted; the image is segmented using a threshold segmentation algorithm, and the segmented image is corrected using an image geometry correction algorithm;

[0068] Correct the segmented cross-sectional image to eliminate image distortion and ensure that the geometric accuracy of the image meets the requirements of subsequent analysis;

[0069] Furthermore, the corrected image is sharpened by using the Laplace operator to enhance image details and improve image clarity and contrast;

[0070] Perform SIFT feature extraction on the sharpened image to obtain the scale-invariant feature information of the image, providing data support for subsequent defect detection;

[0071] The extracted image features are input into the YOLOv3 defect detection model based on convolutional neural network to accurately identify and locate the types and locations of defects such as cracks and foreign matter in cable insulation materials, and the defect detection results are output to provide a decision-making basis for subsequent defect repair and treatment;

[0072] Here is an example:

[0073] When setting up the infrared image acquisition device, the infrared image acquisition device can be installed at a distance of 0.5 meters from the cross section of the cable insulation material, with the acquisition angle perpendicular to the cross section, the focal length set to 50 mm, and the aperture set to F / 2.8 to obtain a high-quality infrared image with a resolution of 1024×768 pixels;

[0074] For the acquired infrared image, a 5×5 median filter is used to remove the salt and pepper noise, and then the db4 wavelet in the wavelet transform is used to perform a three-layer decomposition of the image to remove the Gaussian noise and increase the signal-to-noise ratio to more than 45dB;

[0075] According to the gray histogram distribution characteristics of the cross section of the cable insulation material, the image is segmented using the Otsu method adaptive threshold segmentation algorithm to extract the region of interest;

[0076] The segmented image is corrected using a geometric correction algorithm based on a polynomial model, and the distortion parameters are estimated by matching feature points to control the image distortion within 0.1 pixels.

[0077] The corrected image is sharpened using a 3×3 Laplacian operator to enhance the edge and detail information of the image, thus improving the image clarity by more than 30%;

[0078] On the sharpened image, SIFT feature extraction algorithm is used to extract scale- and rotation-invariant local features, and 128-dimensional feature vectors are extracted from each region of interest to generate a feature database;

[0079] The extracted SIFT features are input into a convolutional neural network model based on the YOLOv3 architecture for defect detection. The model’s weights are pre-trained on the ImageNet dataset and fine-tuned on the cable insulation material defect dataset to achieve accurate positioning and classification of defects such as cracks and foreign matter, with an mAP index of more than 95%.

[0080] Furthermore, in this embodiment, step S2 specifically includes:

[0081] A deep learning algorithm is used to perform semantic segmentation on the processed cross-sectional images and establish a segmentation model; the deep learning algorithm uses a deep convolutional neural network based on the U-Net architecture;

[0082] By extracting features and classifying pixels, the image is divided into defective areas and normal insulating material areas. The defect types in the defective areas include bubbles, impurities, cracks and deformations.

[0083] Using data enhancement and transfer learning methods, the segmentation model is pre-trained based on the existing defect sample data set to improve segmentation accuracy and generalization performance;

[0084] According to the segmentation results, the connected domain analysis algorithm based on region growing is used to mark and number various defect regions. The contour curve of each defect region is obtained by boundary tracing, and the area, perimeter and roundness shape feature parameters of the defect region are calculated using geometric shape descriptors.

[0085] According to the defect type discrimination rules, the defect areas are classified, the defect types are discriminated, and the severity of the defects is evaluated and graded according to the area threshold and roundness threshold. Specifically, the decision tree algorithm is selected to classify the defect areas, discriminate the defect types, and the severity of the defects is evaluated and graded according to the area threshold and roundness threshold.

[0086] The DBSCAN density clustering algorithm is used to perform cluster analysis on the marked defect areas. Adjacent defects are classified into defect groups according to the spatial distance and distribution density between defects. The morphological skeleton extraction algorithm is used to obtain the skeleton curve of each defect group as the center line of the defect group. The topological network diagram of defect distribution is constructed to reveal the expansion trend and connection path of defects on the cross section of the insulating material.

[0087] Graph theory analysis is performed on the defect distribution topology map. All defect extension paths are traversed using a depth-first search algorithm. The defect area growth rate, path length, and number of forks are used to score the defect evolution degree of each path, thereby locating the starting point of the water tree phenomenon.

[0088] Furthermore, the identified water tree starting point area is mapped to the 3D geometric model and material structure model of the cable. The electric field distribution and insulation performance around the defect are analyzed through finite element simulation, the cause mechanism of the defect is studied, and the formula and process parameters of the insulation material are optimized to guide the subsequent cable manufacturing and quality control, thereby comprehensively improving the cable insulation performance and reducing the risk of water tree failure.

[0089] Here is an example:

[0090] When performing semantic segmentation on the cross-sectional images of cable insulation materials, the U-Net network structure is used, the input image size is 256×256, the number of convolution layers is 5, the number of convolution kernels in each layer is 64, 128, 256, 512 and 1024 respectively, the convolution kernel size is 3×3, the activation function is ReLU, the maximum pooling is used in the downsampling path, the transposed convolution is used in the upsampling path, and the jump connection is used between the encoder and the decoder to achieve fine pixel-level segmentation. When marking the defective area, the 8-connected region growing algorithm is used, the grayscale threshold is set to 50, the number of growths is set to 10, and the contours of each defective area are extracted;

[0091] When classifying defects, a decision tree classifier is selected, and input features include 10 shape descriptors such as area, perimeter, and roundness. The maximum depth and minimum number of leaf nodes of the decision tree are optimized through grid search to achieve multi-classification of defect types, with a classification accuracy of more than 95%;

[0092] In defect clustering analysis, the DBSCAN algorithm is used, the radius parameter eps is set to 10 pixels, the minimum number of points minPts is set to 5, the defect points with a distance less than eps and a number of points greater than minPts are classified into one cluster, and the skeleton extraction algorithm is used to extract the center line of each cluster to generate a defect distribution topology map;

[0093] When judging the starting point of the water tree, the Dijkstra shortest path algorithm is used to search for the defective path with the largest area growth rate and the shortest path length in the topological graph. The area where the path starting point is located is identified as the starting point of the water tree. Combined with three-dimensional electric field simulation analysis and insulation material formula optimization, the causes of defects and improvement measures are studied, and ultimately the cable insulation failure rate is reduced by more than 20%.

[0094] Furthermore, in this embodiment, step S3 specifically includes:

[0095] Various factors affect the development of water tree phenomena including different voltages, frequencies and ambient temperatures;

[0096] Collect electrical stress data during cable operation, combine it with static data of cable lines, use data cleaning and feature engineering methods to preprocess data, and obtain a multi-source heterogeneous monitoring data set; electrical stress data includes voltage, current, and frequency; static data of cable lines includes geometric parameters and insulation material performance parameters;

[0097] The temperature data of the cable environment is obtained, and the temporal and spatial distribution model of the cable environment temperature is constructed by combining the cable's geographical location information and meteorological data. The temperature distribution characteristics are then integrated into the feature extraction process of the water tree defect area. The temperature data of the cable environment is obtained by using a temperature sensor.

[0098] The water tree starting point identification results are associated with electrical stress monitoring data and ambient temperature data, the time-frequency domain features of the defect area centered on the water tree starting point are extracted, and the multi-dimensional feature vector of the water tree defect area is constructed; the time-frequency domain features include statistical features, wavelet transform coefficients, and Hilbert-Huang transform parameters;

[0099] By using time series analysis and association rule mining algorithms, combined with the association laws and evolution trends between the characteristic parameters of the water tree defect area, the quantitative relationship between voltage, frequency, temperature and water tree development is obtained, and a water tree phenomenon generation and development prediction model based on time series regression is established;

[0100] Furthermore, machine learning algorithms, such as support vector machines, random forests, and neural networks, are used to take the feature vector of the water tree defect area as input and the water tree defect development level as output. Multiple models are combined through an ensemble learning method to train a water tree phenomenon generation and development prediction model, and the model hyperparameters are optimized through cross-validation and grid search to improve the prediction accuracy and robustness.

[0101] Here is an example:

[0102] When collecting cable operation data, a data logger with a sampling frequency of 10kHz was used to record voltage, current, and frequency parameters every 1 second. At the same time, the geometric parameters of the cable conductor cross-sectional area and insulation thickness, and the dielectric constant and loss factor of the insulation material and other performance parameters were collected through the RS485 bus to form a spatiotemporal data set containing 1000 observation points.

[0103] The collected raw data are subjected to outlier detection and smoothing filtering, and the Z-score normalization method is used to standardize the data. The Kriging interpolation algorithm is used to construct an environmental temperature distribution model along the cable, generate a temperature distribution map with a resolution of 10 meters, and extract the temperature gradient characteristics within a radius of 50 meters around the starting point of the water tree;

[0104] The wavelet transform was used to perform time-frequency analysis on the electrical signal near the starting point of the water tree. The Daubechies4 wavelet basis function was selected, the decomposition scale was 5, and 8 characteristic parameters such as the mean and variance of the wavelet coefficients were extracted. The Apriori association rule mining algorithm was used with a threshold of minimum support of 0.2 and minimum confidence of 0.7 to find the association pattern of parameters such as voltage and temperature when the starting point of the electrical tree appeared, and a multivariate linear regression model of water tree development was constructed. The determination coefficient R² after the model adjustment reached more than 0.85;

[0105] A BP neural network model with two hidden layers and 50 neurons in each layer was adopted. The 50 characteristic parameters of the water tree area were taken as input and the water tree risk level was taken as output. The model was trained using 10-fold cross validation and Adam optimization algorithm with a learning rate of 0.01. The F1 score of the model on the test set reached .

[0106] Furthermore, in this embodiment, step S3 also includes:

[0107] Electrical stress sensors are set up to monitor stress changes inside and outside the cable in real time, and the stress change data is correlated with the identified defect locations for analysis, a stress distribution model is established, and predetermined risk stress characteristics that lead to water treeing are identified;

[0108] Furthermore, by optimizing the sensor layout scheme, combined with the cable geometry and the heterogeneity of the insulation material, a genetic algorithm-based layout density and spatial distribution optimization method is used to maximize the stress distribution measurement accuracy in the sensor coverage area;

[0109] The stress monitoring data is spatially and temporally aligned with the identified cable defect location information, the cable geometry parameters and material property parameters are introduced, and the stress distribution model of the cable insulation layer is established using the three-dimensional finite element method. The static and dynamic correlation between the defect location and the stress distribution is studied to improve the accuracy of stress distribution prediction.

[0110] The key characteristic parameters of stress distribution, including stress mean, peak value, gradient, and frequency domain characteristics, were extracted from the simulation model. The maximum entropy principle was used to determine the number of principal components, and the first N principal components with a cumulative contribution rate of more than 95% were selected as water tree risk assessment indicators.

[0111] The support vector machine algorithm is used to build a water tree risk prediction model, with principal component characteristics as input and water tree occurrence probability as output. Hyperparameters such as balance factor and kernel function parameters are introduced, and the grid search method is used to optimize the model performance, so that the evaluation indicators such as accuracy, precision, and recall rate are optimized.

[0112] Here is an example:

[0113] A stress sensor is arranged every 1 meter on the outer surface of the cable, and a fiber Bragg grating stress sensor is arranged every 5 meters in the insulation layer inside the cable to form a three-dimensional sensing network. The sensor measurement uncertainty is better than 1% FS and the service life exceeds 15 years.

[0114] The Non-dominated Sorting Genetic Algorithm II (NSGA-II) algorithm is used to perform multi-objective optimization on the sensor layout scheme. The optimization objectives are to minimize the monitoring blind area and the total number of sensors. The population size is 50, the evolutionary generation is 100, the crossover probability is 0.8, and the mutation probability is 0.1. The Pareto optimal sensor layout scheme set is obtained.

[0115] The stress data is uploaded to the monitoring main station in real time at a frequency of 1kHz through the IEC61850 protocol. The FPGA-based data compression algorithm is used to achieve a 10-fold data compression ratio and a transmission delay of less than 10 milliseconds. The defect position is converted into a three-dimensional space coordinate through Gaussian projection, and forms a space-time sequence with the stress distribution data. The radial basis function neural network (RBFNN) algorithm is used to build a space-time mapping model, and the differential evolution algorithm is used to optimize the number of hidden layer nodes and smoothing factor of the RBFNN to reduce the root mean square error to less than 0.5%;

[0116] The stress distribution characteristics were analyzed by wavelet transform method in time-frequency mode, and eight sensitive features such as wavelet energy spectrum entropy and wavelet singular value were extracted. After feature selection, the kernel principal component analysis (KPCA) method was used to reduce the dimension of the data from 1000 to 10.

[0117] The water tree risk prediction model was trained using the least squares support vector machine (LSSVM) algorithm, with a radial basis kernel function. The regularization parameter ranged from 0.1 to 100, and the kernel function parameter ranged from 0.01 to 10. The optimal parameter combination was determined through 5-fold cross validation, so that the accuracy of the test set reached 95%;

[0118] The long short-term memory (LSTM) network is used to extract the time-dependent features of stress signals. Each LSTM unit contains 64 neurons. The network is trained using the Adam optimizer. The initial learning rate is 0.001, and the learning rate is reduced to 0.1 times the original rate every 10 epochs. After training for 100 epochs, the optimal water tree risk time series prediction model is obtained.

[0119] The simulated annealing algorithm is used to perform multi-objective optimization of cable maintenance strategy. The objective function is to minimize the water tree risk index and the maintenance cost. The constraints are the upper limit of maintenance resources and the continuity requirement of cable operation. The initial temperature is 100, and the cooling ratio is 0.9 each time. The optimal maintenance plan is obtained to guide cable operation and maintenance management.

[0120] The parallel computing analysis method is used to evaluate the potential risk of each defect location by combining the defect location and the predetermined risk stress characteristics, build a water tree risk assessment model, and classify the risk level of water tree phenomena in each area. The parallel computing analysis method uses a Spark-based parallel computing analysis method.

[0121] Specifically, the Spark distributed computing framework is used to load the cable defect location data and stress distribution data into the HDFS distributed file system. The data format is a CSV file, which contains fields such as defect ID, X coordinate, Y coordinate, Z coordinate, defect type, and stress value;

[0122] SparkSQL was used to clean and preprocess the data, and abnormal data with missing defect types and stress values ​​were filtered out. The stress values ​​were normalized using the Z-score method. Based on the cleaned data, the DBSCAN density clustering algorithm was used to perform cluster analysis on the defect location data. The clustering radius was set to 50 meters, and the minimum number of included points was set to 5. The defect-intensive areas were extracted, and the characteristic parameters such as the number of defects, average defect spacing, and defect area in each intensive area were calculated.

[0123] Perform spatial interpolation analysis on stress distribution data, use Kriging interpolation algorithm to generate a stress distribution grid map with a resolution of 1 meter, and extract characteristic parameters such as gradient and curvature of stress distribution. Input the defect-intensive area characteristics and stress distribution characteristics into the random forest model, build a water tree risk assessment model, use grid search to optimize model hyperparameters, and evaluate model performance through 5-fold cross validation, so that the ROCAUC value of the test set reaches above 0.95;

[0124] The trained risk assessment model is used to predict the water tree risk of each defect location. The risk level is divided into three levels according to the risk probability value: high, medium and low. High risk means the probability is greater than 0.7, medium risk means the probability is between 0.3 and 0.7, and low risk means the probability is less than 0.3.

[0125] The Apriori association rule mining algorithm is used to extract frequent item sets and association rules between different risk levels and voltage levels, ambient temperature, humidity and other factors, calculate the support, confidence and lift of each rule, screen out strong association rules, and build a quantitative relationship model of risk influencing factors;

[0126] Based on the risk level prediction results and the influencing factor analysis results, differentiated risk control measures are formulated. For high-risk areas, a monthly regular inspection and partial discharge detection plan is formulated, and stress control and load adjustment are carried out. For medium-risk areas, a quarterly regular testing plan is formulated, focusing on insulation resistance and dielectric loss. For low-risk areas, regular annual inspections and tests are maintained.

[0127] Here is an example:

[0128] Apache Spark was used to efficiently read defect location data and stress distribution data stored in Parquet format. This data was stored in a columnar format and compressed using Snappy technology, which saved more than 50% of storage space while ensuring data reading efficiency.

[0129] SparkSQL was used to perform ETL processing on the data, regular expressions were used to clean string fields, non-standard ASCII characters were removed, and the random forest algorithm was used to intelligently fill in missing values, effectively controlling the missing data ratio to less than 1%, thus enhancing data integrity and analysis reliability.

[0130] The DBSCAN clustering algorithm was used to analyze the defect location data. The grid search technology was combined to finely adjust the parameters (eps was set to 30 meters and minPts was 10). Five high-density defect clusters were successfully identified. These areas contained an average of 3 defects per square meter. eps (epsilon) is a distance parameter that determines the maximum distance between two points that are considered neighbors. minPts specifies the minimum number of points in the neighborhood required for a point to be classified as a core point. Setting it to 10 means that a point is considered a core point of a cluster only when there are at least 10 other points around it.

[0131] Gaussian process regression (GPR) and exponential quadratic kernel function were used for fine interpolation of stress distribution. The optimal kernel function configuration of l=10 and sigma_f=1.5 was obtained through parameter optimization. The generated 50-meter resolution stress distribution map clearly shows the stress change trend, with an average stress gradient of 0.3MPa / m. l controls the smoothness of the function change, and a larger l means a smoother surface. sigma_f reflects the degree of variation between data points. A water tree risk assessment model based on random forest was constructed, which comprehensively considers 8 key features including defect density and stress gradient;

[0132] The model hyperparameters were carefully adjusted to n_estimators=100 and max_depth=10 through the grid search strategy. The model was rigorously cross-validated with 10 folds, and the ROC-AUC score on the test set was as high as 0.98, indicating that the model has extremely high classification prediction accuracy and generalization ability. n_estimators=100 represents the number of decision trees in the random forest, that is, the model contains 100 trees, and max_depth=10 represents the maximum depth limit of a single decision tree, which helps to prevent overfitting and maintain the generalization ability of the model.

[0133] The FP-Growth algorithm was used to mine association rules, with the minimum support set to 0.05 and the minimum confidence set to 0.8. Five strong association rules were obtained, including high risk and high humidity environment, medium risk and medium temperature environment. A Logistic regression model was constructed to quantify the strength of the rules, and the model coefficients passed the significance test.

[0134] Based on the results of risk prediction and rule mining, a three-level risk control plan is formulated. Ultrasonic partial discharge detection is carried out once every two weeks in high-risk areas, and the external temperature and humidity are continuously monitored to dynamically adjust the cable load curve; insulation resistance tests are carried out once a month in medium-risk areas, focusing on changes in insulation levels when the temperature changes suddenly; routine tests are carried out once every six months in low-risk areas to form a water tree risk management file;

[0135] In the WebGIS system, the distribution of water tree risk levels is displayed in the form of a heat map, and inspection and maintenance tasks are automatically triggered according to the risk warning level. The task content is pushed to the mobile terminal of the operation and maintenance personnel in the form of a work order. The three-dimensional water tree risk distribution visualization system is developed using WebGL technology. The risk level is intuitively displayed with a red, yellow and green color scheme. Click the defect location to pop up the risk assessment report and control suggestions, generate a water tree risk management ledger, and seamlessly integrate with the operation and maintenance management system. The task progress and completion status are tracked in real time to form a closed loop of water tree risk management.

[0136] Furthermore, in this embodiment, step S4 also includes:

[0137] According to the adaptive optical compensation method, the exposure time, aperture size and image enhancement parameters are automatically adjusted in combination with changes in ambient light and imaging conditions;

[0138] Adaptive optical compensation and image enhancement methods are used to dynamically adjust imaging parameters in real time to reduce the interference of haze and smoke on image quality, including:

[0139] The deep learning-based image enhancement algorithm was used, and a convolutional neural network with 8 convolutional layers and 3 fully connected layers was used to extract features and restore infrared images under the interference of haze and smoke. End-to-end network training was performed on a training set of 10,000 simulated degraded images and high-quality infrared images. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators were used to evaluate the network performance, which improved the image PSNR by more than 5dB.

[0140] Introducing an atmospheric correction algorithm based on a physical model, the infrared image is calibrated for radiation and spectrum restoration based on the MODTRAN atmospheric transmission model and real-time meteorological parameters, eliminating atmospheric absorption and scattering effects, and improving the image signal-to-noise ratio by more than 8dB;

[0141] Adaptive histogram equalization algorithm is used to automatically determine the optimal histogram stretching interval according to the grayscale distribution of the image. The mapping relationship between different grayscale intervals is adaptively optimized using piecewise linear stretching. Gradient weighted histogram is used as the target distribution function of equalization, taking into account the contrast and detail information of the image, increasing the image entropy by more than 1.5 and improving the edge sharpness by 30%.

[0142] For image denoising in dynamic scenes, a multi-frame denoising algorithm based on motion compensation is used to perform motion vector estimation and motion compensation on continuously acquired multi-frame images, and then adaptive weighted fusion is performed in the time domain and spatial domain. The adaptive adjustment range of the fusion weight is set to 0.2 to 0.8, which improves the image signal-to-noise ratio by more than 4dB and avoids the introduction of motion blur.

[0143] Here is an example:

[0144] Integrate an adaptive optical compensation module in the optical system of the infrared image acquisition device, use the Shack-Hartmann wavefront sensor to detect the wavefront distortion caused by atmospheric turbulence, uneven refractive index of infrared window materials and lens aberration in real time, and use multiple deformable mirror cascades or spatial light modulators to compensate for the wavefront distortion in real time. The compensation frequency reaches 500Hz, which improves the image quality by more than 25%;

[0145] In the infrared imaging system, a Shack-Hartmann wavefront sensor composed of 289 microlens arrays is used to detect the wavefront gradient of 32x32 sub-apertures in real time. The sampling frequency is 1kHz, and the wavefront reconstruction accuracy is better than λ / 20. Through closed-loop feedback control, two 69-channel deformable mirrors and a 512x512 pixel liquid crystal spatial light modulator are driven to achieve dynamic compensation of the first 40 orders of Zernike polynomials distortion, and the compensation bandwidth reaches 150Hz.

[0146] Based on the ImageNet pre-trained ResNet-50 as the backbone network, an end-to-end infrared image dehazing and dust removal network was built. 6,000 simulated degraded images with a resolution of 960x720 and high-quality images were used for supervised training. After 500 epochs of iterative optimization, the image PSNR was improved from 24.6dB to 29.8dB, and the SSIM was improved from 0.61 to 0.85.

[0147] According to the MODTRAN6.0 atmospheric transmission model, combined with the visibility, temperature, humidity, air pressure and other parameters provided by the meteorological station, the radiation transmission equation solver DISORT is used to quantitatively calculate the infrared radiation, and the atmospheric transmittance and uplink radiation in the 414μm band are hyperspectral modeled. The 320x256 hyperspectral image data cube is inverted using the least squares method to eliminate the atmospheric effect and increase the signal-to-noise ratio to 55dB.

[0148] Adaptive histogram equalization of infrared images is realized by using cumulative distribution function transformation method. The gray probability density function in the range of 0.1 is piecewise linearly fitted. The optimal stretching coefficient of 8 sub-intervals is obtained by solving the linear programming problem. The gradient weighted histogram is calculated using the gray co-occurrence matrix, and KL divergence is used as the optimization objective function. After 10 iterations of optimization, the image information entropy is increased from 4.62 to 7.25, and the edge sharpness gradient is increased from 52 to 78.

[0149] The Lucas-Kanade pyramid optical flow method is used to estimate the motion vector of 5 consecutive frames of 720P@50fps infrared video sequences, and a three-dimensional Feldkalman filter is constructed. The motion speed and acceleration parameters in the 3x3 pixel neighborhood are estimated through iterative optimization to achieve sub-pixel motion compensation. The wavelet domain 3D denoising filter is used to perform adaptive weighted fusion of inter-frame differential signals. The weight adaptive adjustment range is 0.2-0.6, which improves the image signal-to-noise ratio from 35.2 to 41.4, effectively suppressing motion blur.

[0150] The above image processing algorithm is accelerated in parallel on the FPGA hardware acceleration platform, and the pipeline architecture and distributed storage optimization are used to achieve a real-time processing speed of 100 frames per second, with the delay controlled within 50 milliseconds. On the Xilinx Virtex UltraScale + XCVU9P FPGA platform, a 5-level pipeline architecture is used to achieve full frame rate real-time image enhancement, and 64-way SIMD parallel acceleration is used to reduce the algorithm processing delay to within 20ms, and the power consumption is controlled within 12W;

[0151] Establish an infrared imaging quality evaluation system that includes five quantitative indicators, including signal-to-noise ratio, contrast, clarity, and edge sharpness, as an objective basis for algorithm optimization and system debugging, and adaptively adjust the indicator weights according to the needs of the application scenario to achieve scene-adaptive optimization of imaging quality;

[0152] An infrared imaging quality evaluation system was designed with five built-in indicators: signal-to-noise ratio, contrast, entropy, gradient, and clarity. The fuzzy comprehensive evaluation method was used to determine the scene adaptive weighting coefficient. The Pearson correlation coefficient between the evaluation results and the subjective score of the human eye reached 0.95, which can be used as an objective basis for imaging optimization.

[0153] The adaptive optical compensation method specifically includes:

[0154] An electrically controlled liquid lens and a variable aperture are introduced into the optical imaging system. The digital control signal is converted into an analog voltage of 0-5V through a 12-bit DAC chip, so that the focal length of the control lens can be continuously adjusted from 1m to infinity, and the aperture can be continuously adjusted from F / 1.8 to F / 16. At the same time, a parameter optimization strategy based on random forest regression is adopted, and a mapping model between environmental factors and optimal exposure parameters is established through offline learning of historical data, so as to realize intelligent prediction and adaptive setting of exposure time and aperture value, and continuously optimize the mapping model through online incremental learning. An image enhancement convolutional neural network based on the U-Net structure is built, with the original image as input and the feature map of the middle layer as the supervision signal. The network parameters are trained end-to-end, and the brightness and contrast loss functions of multi-scale distribution fitting are used to train on a data set of 5,000 images in extreme scenes such as low illumination and backlighting, so as to achieve image brightness balance, contrast enhancement, and detail sharpening;

[0155] In view of the radial distortion and tangential distortion of the optical lens, a distortion correction module based on the Brooks model is built in the FPGA. The 3x3 radial distortion coefficient matrix and the 2x3 tangential distortion coefficient matrix are obtained through calibration. The bicubic interpolation algorithm is used to realize image resampling and real-time correction. The linear distortion rate of the image after distortion correction is reduced to less than 0.1%.

[0156] The collected original image is decomposed into 5 levels using Haar wavelet, and the LL5 low-frequency subband coefficient is extracted as the image clarity evaluation index D, and the HL1 high-frequency subband coefficient is extracted as the image contrast evaluation index C. They are compared with the empirical thresholds Dmin=50 and Cmin=100 respectively, and the convolution kernel size and quantization bit number of the image enhancement network are adaptively adjusted. Image gradient G and information entropy E are introduced as supplementary evaluation indicators, and a weighted comprehensive quality evaluation function Q=0.4D+0.3C+0.2G+0.1E is constructed as the objective function of adaptive image enhancement.

[0157] Here is an example:

[0158] The acquisition device integrates ambient light sensor, spectrum sensor, focus status sensor and jitter sensor to detect the parameters of ambient brightness, spectrum distribution, imaging distance and jitter amplitude in real time. The sensor analog signal is converted into a digital signal through a dedicated ADC chip and transmitted to the FPGA image processing module through the I2C bus. The integrated OmniVision OV12890 sensor has a pixel size of 1.55μm and a photosensitive area of ​​1 / 2.3 inches. It is matched with an 850nm long-wave infrared filter to achieve low-light imaging of 0.01lux.

[0159] An array of 6 OSRAM3030 lamp beads is used to provide auxiliary light source. The intelligent control uses the ambient brightness sensor data as the input of the PID controller to achieve continuous adjustment of the auxiliary light intensity from 0 to 200 lm.

[0160] The sensor data acquisition module based on the STM32L433 microcontroller is used to synchronously sample the outputs of the RohmBH1749 spectral sensor, STVL53L1X time-of-flight sensor, and ADXL355 three-dimensional accelerometer at a frequency of 100 Hz. The environmental spectral vector, target distance, and jitter vector are obtained through Kalman filtering and coordinate transformation.

[0161] Build a servo control system based on the Optotune EL-10-30-C electronically controlled liquid lens, convert the PWM signal into a 0-5V DC control voltage through the TIOPA2376 operational amplifier, and achieve continuous adjustment of the lens focal length within the range of 50mm to 300mm with a response time of 20ms;

[0162] The adaptive imaging hardware platform is built using RitPix CMOS image sensor and XILINX Zynq XC7Z045 FPGA. The number of sensor pixels is 2000x2000, the pixel size is 5.5μm, and the FSIV006A 12-bit successive approximation ADC chip is used to sample the sensor analog signal, with a maximum sampling rate of 10Msps.

[0163] Implement the random forest regression algorithm in the Linux system on the FPGAPS side, extract 120-dimensional environmental feature vectors and optimal imaging parameters from 50,000 frames of images collected by the sensor historically, build a random forest mapping model with a decision tree depth of 12, set the number of support vectors for online incremental learning to 1000, set the U-Net network activation function to LeakyReLU, and the loss function to a weighted combination of L1 and SSIMloss. Set the learning rate of the Adam optimizer to 0.00001, the training batch size to 16, and the number of iterations to 200.

[0164] The image quality was initially screened using the SVM classifier, and 2,000 high-quality images were selected to form the training set. The training set was expanded to 20,000 using data enhancement methods such as rotation, scaling, and Gaussian noise. The scale coefficients of the Haar wavelet transform were set to 2, , ,The clarity, contrast, gradient and entropy weights in the image quality evaluation function are optimized by genetic algorithm;

[0165] We built an FPGA-based adaptive imaging hardware platform, using high-speed CMOS sensors and LVDS interfaces to achieve low-latency acquisition and transmission of image data, introduced an ISP image signal processing pipeline based on the Bayer array, improved the visual quality of the image through local color mapping and gamma correction, and adopted an OpenCL-based heterogeneous computing framework to transplant the core algorithm modules to the FPGA for implementation. We used on-chip BRAM and DSP resources to optimize the critical path and achieve real-time processing of the image enhancement algorithm, with a frame rate of 120fps and a latency of less than 10ms.

[0166] On the image signal processing FPGAPL side, the image enhancement network function level is parallelized based on HLS high-level synthesis, the BRAM and DSP resource usage is optimized, and a 16-bit fixed-point CNN accelerator is implemented with a clock frequency of 200MHz, a peak speed of 1.2TOPS, and a power consumption of 12W frame rate.

[0167] Furthermore, in this embodiment, step S5 specifically includes:

[0168] The cable line nodes are divided into three levels: important nodes, general nodes and auxiliary nodes. The hierarchical analysis method is used to comprehensively evaluate the importance of nodes. A node importance evaluation system is constructed, including operating voltage level, transmission power, power supply range, geographical location and water tree fault history indicators. The importance weight of each node is obtained through indicator weight solution and consistency test. The clustering algorithm is used to classify the nodes to form the node importance division result.

[0169] Furthermore, a water tree fault risk prediction model was constructed, using the Gradient Boosting Decision Tree algorithm, with characteristic parameters such as node importance, regional ambient temperature and humidity, and cable insulation aging as input, and the probability of water tree fault occurrence as output. The model hyperparameters were optimized through five-fold cross-validation and grid search, so that the prediction accuracy reached more than 85%, and the water tree risk was graded according to the fault probability threshold, where the fault probability greater than 0.6 was high risk, 0.3-0.6 was medium risk, and less than 0.3 was low risk.

[0170] For nodes of different levels, a multi-resolution image acquisition scheme is designed. For important nodes, 4096×2160 resolution full-frame acquisition is used, for general nodes, 1920×1080 resolution region of interest acquisition is used, and for auxiliary nodes, 1280×720 resolution interval sampling is used. The reasonable control of image data volume is achieved through resolution adaptive configuration.

[0171] The water tree risk assessment model is integrated into the image acquisition equipment. The image acquisition parameters are adjusted in real time according to the node importance level and risk prediction value. When the node water tree risk level is high, the image acquisition resolution of the area is increased by one level, and the sampling frequency is increased to once every 15 minutes; when the risk level is medium, the original resolution is maintained, and the sampling frequency is increased to once every 30 minutes; when the risk level is low, the resolution is reduced by one level, and the sampling frequency is reduced to once every 60 minutes. An FPGA-based multi-resolution image encoding module is used. JPEG 2000 wavelet transform encoding is used for 4K resolution images, and the compression ratio is set to 4:1; H.265 intra-frame encoding is used for full HD resolution images, and the compression ratio is set to 40:1; WebP encoding is used for 720P resolution images, and the compression ratio is set to 60:1, to achieve efficient transmission and storage of image data;

[0172] A multi-scale water tree defect detection framework was built, and the FPN feature pyramid network was used to extract multi-scale features. Channel attention and spatial attention modules were introduced to enhance defect feature expression. The Mask R-CNN instance segmentation algorithm was used for high-resolution images, the YOLOv5 target detection algorithm was used for medium-resolution images, and the MobileNet-SSD lightweight detection algorithm was used for low-resolution images. The generalization performance of the detection model was improved through transfer learning and difficult sample mining, and the mAP on the water tree defect test set reached more than 90%;

[0173] The multi-view and multi-resolution water tree defect detection results are graded and integrated. The Dempster-Shafer evidence theory is used to comprehensively consider factors such as node importance, defect detection confidence, and defect size. A basic probability assignment function is constructed. The confidence of the defect existence is calculated through evidence combination rules, and a water tree fault probability distribution map is generated. The confidence threshold is adaptively adjusted according to the current water tree development trend, and the water tree fault warning level and maintenance priority are dynamically updated to provide a quantitative basis for water tree defect management.

[0174] Here is an example:

[0175] Establish an evaluation index system for the importance of cable line nodes, including operating voltage level (5 points for 110kV and above, 4 points for 35kV-110kV, 3 points for 10kV-35kV, 2 points for 380V-10kV, and 1 point for below 380V), annual average transmission power (5 points for above 1000MW, 4 points for 500MW-1000MW, 3 points for 100MW-500MW, 2 points for 10MW-100MW, and 1 point for below 10MW), cable length (5 points for above 10km, 4 points for 5km-10km, 1km-5 km is 3 points, 500m-1km is 2 points, and below 500m is 1 point). The indicator weights are calculated by the analytic hierarchy process (AHP), the consistency ratio CR=0.056<0.1, the weight vector W={0.25, 0.20, 0.15, 0.10, 0.10, 0.08, 0.07, 0.05}, and the weighted sum is used to get the comprehensive score of node importance. The K-means clustering algorithm is used to divide the score into three levels, forming the clustering results of {important nodes (25), general nodes (150), and auxiliary nodes (400)};

[0176] The XGBoost algorithm is used to build a model for predicting water tree faults. The model is trained based on the operating data of the past ten years, which covers meteorological conditions, electrical parameters, and defect information. Whether a node has a water tree fault is marked as a binary label, that is, the occurrence of the fault is recorded as (1) and the non-occurrence is recorded as (0). The importance of the features obtained through training is ranked as follows: the insulation aging index is the most critical, with a contribution of 0.35, followed by the average humidity of the node, with a contribution of 0.25, the average temperature of the node, with a contribution of 0.15, the contribution of the node importance, with a contribution of 0.10, the contribution of the phase-to-phase insulation resistance, with a contribution of 0.08, the contribution of the outer semi-conductive layer shedding, with a contribution of 0.06, and the contribution of the number of cable joints is the smallest, at 0.01. The grid search method is used for parameter tuning, and the final learning rate is determined to be 0.1, the number of decision trees is set to 100, and early_stopping_rounds is set to 10 to prevent overfitting. The accuracy of the model on the test set reaches 0.92; if the evaluation index of the model on the validation set does not improve for 10 consecutive rounds, then the training will automatically stop;

[0177] According to the ROC curve, the classification thresholds were determined to be 0.6 and 0.3 respectively to distinguish high, medium and low levels of water tree risks. A 12-bit CMOS sensor and a 10-bit ADC were used to collect 4K, full HD and 720P resolutions respectively. The frame rate of 4K images was 1fps, the frame rate of HD images was 10fps, and the frame rate of 720P images was 30fps. An adaptive coding framework for regions of interest based on attention mapping was designed. The quality level control strategy in scalable video coding was used for key node areas. The quantization parameter QPL=8, the general area QPL=12, the non-key area QPL=20, and the total bit rate of the video stream was controlled within 80Mbps.

[0178] A real-time video defect analysis pipeline was designed, using the RetinaNet framework, equipped with a powerful ResNeXt-101 backbone network, incorporating 3x3 deformable convolution and ASPP feature fusion modules, specifically for ultra-high-definition defect detection for 4K images, maintaining a frame rate of 15fps, and achieving an average precision (mAP) of 0.95 on a specific data set after fine-tuning; for high-definition video streams, a solution combining the SSD meta-architecture and the DeepSORT algorithm was adopted to perform continuous defect tracking at a speed of 60fps, achieving efficient real-time monitoring, with an mAP of 0.92; for 720P images, a lightweight TinyYOLO model was selected, which significantly compressed the parameters to 4M to achieve high-speed defect location at 120 frames per second, with an mAP of 0.88, and average detection delays of 100ms, 30ms, and 10ms respectively;

[0179] The defect analysis results are further integrated, including the probability of occurrence of defect areas, the importance of key nodes, and the resolution of the image itself are extracted as inputs, and the Dempster-Shafer (DS) evidence reasoning model is constructed. Through the carefully designed evidence combination strategy function of quality, node, and time priority, the detection results from different sources are comprehensively evaluated and the comprehensive confidence is calculated;

[0180] According to the set threshold, a confidence level higher than 0.85 is considered a high-risk water tree defect, 0.6 to 0.85 is considered a medium risk, and below 0.6 is considered a low risk. The sliding average algorithm can also be used to dynamically adjust the distribution factor of evidence conflicts and continuously optimize the inference threshold to ensure the accuracy and adaptability of risk assessment.

[0181] Furthermore, the defect locations, stress distribution characteristics and historical water tree risk data are evaluated, and the monitoring areas of the cables are divided according to the importance of the cable areas. The monitoring areas are classified into important areas and non-important areas according to whether they need protection;

[0182] Specifically, the cable line is divided into a monitoring unit every 1 km, and the topological structure of the cable line, the location of branch nodes, and the geographical environment along the line are comprehensively considered. The density clustering algorithm is used to adaptively divide the cable line monitoring units. The fault density within the unit is similar, and the fault characteristics between units are greatly different;

[0183] Extract multi-source heterogeneous features such as fault occurrence time, fault location, defect type, load level, and operating years for each monitoring unit, and explore environmental factors affecting water tree occurrence, such as temperature, humidity, and vibration, to build a comprehensive water tree risk characteristic indicator system;

[0184] The multi-view learning method is used to group and fuse different types of features, and the Kendall rank correlation coefficient is used to measure the correlation between each group of features and water tree risk. The comprehensive evaluation index of water tree risk is obtained through weighted fusion.

[0185] According to the importance of different monitoring units, a comprehensive evaluation is conducted from two dimensions: the importance of the cable itself and the danger of the external environment. The fuzzy hierarchical analysis method is used to construct an evaluation matrix. The indicators include the cable line level, the scope of fault impact, the importance of the load, and the complexity of the environment. The importance of each monitoring unit is determined by the maximum membership principle.

[0186] The comprehensive evaluation index of water tree risk and the importance degree membership vector are synthesized by fuzzy multiplication to obtain the risk-importance two-dimensional evaluation matrix of each monitoring unit. The evaluation matrix is ​​divided into five monitoring zones: important high risk, important low risk, general high risk, general low risk, and non-important area by cluster analysis method.

[0187] Develop differentiated monitoring strategies for different monitoring zones, implement adaptive monitoring solutions with status assessment as the goal;

[0188] Online monitoring is used in important high-risk areas. The measurement points cover important nodes and fault-prone locations along the cable. Physical quantities including voltage, current, and partial discharge are collected once per minute. High-precision fiber optic sensors are used, with amplitude resolution better than 0.5% and frequency resolution better than 1kHz.

[0189] Important low-risk areas adopt a combination of online monitoring and regular inspections. The online measurement points are arranged as above, and offline manual inspections are conducted once a month, focusing on weak points such as cable heads and intermediate joints;

[0190] In general areas, the monitoring method is adjusted according to the risk status level. When the risk is high, the density of online measurement points and sampling frequency are increased. When the risk is low, the monitoring intensity is reduced, and offline regular inspection is mainly used. Non-important areas use offline manual sampling once a quarter to collect typical parameters such as cable insulation resistance;

[0191] Combining the internal factors such as insulation material aging, local electric field concentration, impurity ionization, and the external factors such as load fluctuation, switching inrush current, and environmental stress, a multi-factor coupled water tree growth dynamics equation is established.

[0192] Based on the mechanism modeling, the long short-term memory neural network is used to extract the time-dependent characteristics of the water tree fault sequence, explore the water tree development mode on different time scales, and realize the multi-step prediction of water tree defects.

[0193] An integrated learning method is used to integrate multiple prediction models such as support vector machines, extreme gradient boosting trees, and convolutional neural networks to form a multi-modal early warning system for water tree faults. The warning is triggered by fault precursor information, and the warning time scale is extended from 1 day to 1 month.

[0194] Develop proactive prevention strategies for high-risk areas identified by early warning, such as optimizing cable insulation structure, improving joint waterproof sealing, implementing cathodic protection of metal shielding layers, and conducting partial discharge detection and location to suppress water tree initiation;

[0195] According to the development trend of water trees in different monitoring areas, differentiated condition maintenance strategies are formulated. For example, in high-risk areas, the maintenance cycle is shortened and condition-based corrective maintenance is carried out; in medium- and low-risk areas, the maintenance cycle is extended and condition-based preventive tests are carried out; in non-critical areas, post-replacement is the main method. Through risk classification management and condition-based maintenance strategy optimization, the water tree governance cost can be minimized under reliability constraints.

[0196] Here is an example:

[0197] The DBSCAN density clustering algorithm is used to divide the 956 monitoring units of the cable line, setting the radius parameter ε=500m and the density threshold minPts=5, and finally 78 fault-intensive areas and 45 fault-sparse areas are obtained;

[0198] The 12-dimensional feature vectors of each monitoring unit in the past year were extracted, including the average ambient temperature (0.2), relative humidity (0.15), partial discharge frequency (0.3), maximum partial discharge amplitude (0.1), cable operating years (0.1), and average load level (0.05). The numbers in brackets are the significance coefficients of each feature. The features are divided into three groups based on the multi-view learning of PowerIterationClustering: environment, state, and working condition. The average values ​​of the Kendall rank correlation coefficients within the groups are 0.75, 0.82, and 0.69, respectively, and the average value of the correlation coefficients between the groups is 0.33. The weighted fusion coefficients of each group are determined by alternating least squares regression. The comprehensive risk index of water trees is calculated by Score10.5+Score20.35+Score3*0.15.

[0199] The importance evaluation matrix was constructed by fuzzy analytic hierarchy process, and the linguistic variable "relatively important" was expressed by triangular fuzzy numbers. The relative importance of each evaluation object was quantified by means of Sigmoid membership function (its parameter p was set to 1.2). Specifically, 220kV line was given the highest membership of 0.7, followed by 110kV line (0.65), 35kV line (0.55), 10kV line (0.45) and distribution network branch line (0.3). The evaluation matrix was constructed by fuzzy analytic hierarchy process, and the consistency test (the consistency ratio CR=0.016 was obtained, which was much smaller than the critical value 0.1) ensured the rationality of the evaluation structure. Based on this, the weight vector W={0.35,0.25,0.2,0.12,0.08} of each voltage level line and distribution network branch line was obtained, reflecting its relative importance in the system. Through the fuzzy synthesis operator θ, the importance of the monitoring unit (expressed by the membership degree I) is combined with the water tree risk index S to form a comprehensive two-dimensional evaluation matrix. In the specific calculation, the importance of each monitoring unit I is obtained by its weight W and risk level R through fuzzy operation;

[0200] Using the K-means clustering algorithm, all monitoring units were divided into five categories according to their positions in the two-dimensional matrix. When I was greater than 0.5 and S was greater than 0.4, it was defined as an “important high-risk area”, and 15 such areas were identified. When I was also greater than 0.5 but S was less than 0.4, it was classified as an “important low-risk area”, and 26 such areas were identified.

[0201] A four-level detection strategy is formulated for different monitoring zones. In high-risk areas, multi-parameter cross-checks such as temperature measurement, humidity measurement, ultrasound, electromagnetic, CT / PT, etc. are used. The sampling rate is 1kHz, and the distance between important node measurement points is 50 meters. The sensitivity of the optical fiber vibration sensor is better than 0.01g, and the frequency response range is 3Hz to 2500Hz.

[0202] In general areas, only online monitoring of temperature and humidity (1 unit / km) and partial discharge (2 units / km) is deployed. Offline inspection personnel carry infrared thermometers (temperature measurement accuracy 0.1°C) and digital ultrasonic testers (amplification factor 90dB) to conduct monthly inspections. In non-critical areas, 500V megohmmeters (accuracy level 0.5) are used to measure insulation resistance. A value below 2MΩ is considered a serious defect.

[0203] A three-dimensional random Monte Carlo water tree growth simulation model is constructed. The dielectric breakdown field strength obeys the Weibull distribution, the shape coefficient is 2.15, the scale coefficient is 35kV / mm, the water dendrite growth obeys the Logistic curve, the growth coefficient is 0.12, the ambient temperature acceleration coefficient is 0.06 / ℃, the frequency acceleration coefficient is 5.4×10^(-9) / Hz^2, the load acceleration coefficient is 2.7×10^(-4) / A^2, and the average water tree length for 500 days is predicted to be 0.75mm. The bidirectional LSTM network is used as the core, aiming to predict the possible future power system water tree faults by analyzing historical data. The model training configuration includes the use of the Adam optimization algorithm for efficient iteration, and the time step of time series data processing is set to 50 to capture sufficient historical dynamic information;

[0204] In terms of network structure design, a 1024-dimensional hidden layer is configured to enhance the learning ability of the model, and a 512-dimensional fully connected layer is configured to integrate features to ensure that the model has good generalization performance. The learning rate is set to 0.0008, balancing the convergence speed and learning accuracy. At the same time, regularization (coefficient 0.02) and dropout (probability 0.5) strategies are added to prevent overfitting and improve the stability and generalization ability of the model. The data set is scientifically divided into training set, validation set and test set, with a ratio of 6:2:2, to ensure effective verification and evaluation of the model;

[0205] On the test set, the model showed an 85% prediction accuracy, which means that the prediction of fault events is highly reliable. For the anomaly detection capability one week before the fault occurs, the model achieved a 52% detection rate. The water tree fault alarm system based on ensemble learning integrates three powerful and complementary base learners: support vector machine (SVM), gradient boosted decision tree (GBDT) and LightGCN. By integrating their prediction results, the system can analyze the fault mode more comprehensively.

[0206] Under cross-validation evaluation, the integrated system showed excellent performance indicators, with a hamming loss of only 0.12, a one-error of 0.09, and a ranking loss of 0.15. These values ​​all indicate that the system has a high degree of accuracy in fault classification and ranking tasks, and can effectively reduce false positives and false negatives;

[0207] Condition repair tests are carried out in high-risk areas, including power frequency withstand voltage (1.5 times rated voltage, lasting 1 minute), DC leakage current (<0.1mA), dielectric loss tangent (<0.5%), and cables with serious defects are replaced or cut off (<1% / year). For general areas, the maintenance cycle is extended, and preventive insulation resistance tests and partial discharge tests are carried out every 2 years. Cables in non-critical areas are replaced every 6 years on average (<0.1% / year). Through zoning monitoring and hierarchical management and control, the cable water tree failure rate is reduced by 15%, saving annual maintenance costs;

[0208] Furthermore, according to the classification of the monitoring area, personalized monitoring measures are implemented. For predetermined important areas, predetermined defects are captured using image acquisition equipment with predetermined resolutions, and for non-important areas, predetermined resolution image acquisition methods are used for monitoring;

[0209] Specifically, based on the classification results of the importance of the monitoring areas, a hierarchical and layered personalized monitoring strategy is adopted to implement high-resolution, high-frequency, multi-modal refined monitoring for important areas, and low-resolution, low-frequency, single-modal simplified monitoring for non-important areas, so as to achieve the rational allocation and optimal utilization of monitoring resources;

[0210] Deploy high-definition industrial cameras in important monitoring areas, using Gigabit Ethernet interfaces, supporting 1920×1200 pixel resolution, 60 frames / second video acquisition rate, equipped with 50-200mm telephoto zoom lenses, and a field of view covering 5-50 meters, to meet the needs of fine-grained observation of key defects;

[0211] According to the defect imaging requirements under different lighting conditions, the camera's exposure time, gain, white balance and other parameters are adaptively adjusted. Through time-sharing exposure and multi-frame fusion, the dynamic range of the image is expanded to improve the defect imaging quality and signal-to-noise ratio.

[0212] For water tree defect images collected in important areas, a deep learning detection model based on convolutional neural network was constructed. ResNet-50 was selected as the backbone network, pre-trained on the ImageNet dataset, and then fine-tuned on water tree defect samples to optimize the cross entropy loss function. The training samples were expanded through data enhancement to improve the generalization ability of the model.

[0213] Design a cascade classifier that integrates handcrafted features such as texture, edge, and shape to complement the deep learning model and further improve the accuracy and recall of defect detection;

[0214] The FPGA-based parallel image processing module is used to perform real-time analysis on the collected high-resolution video stream. Through traditional image processing algorithms such as morphological filtering, edge detection, and region growing, the geometric parameters of water tree defects, including length, width, area, and perimeter, are extracted to achieve accurate positioning and quantitative characterization of defects.

[0215] The defect detection and segmentation results are transmitted back to the data center in real time, and the H.265 encoding algorithm is used to compress the video data to reduce the transmission bandwidth and storage space occupied;

[0216] Low-cost industrial cameras are deployed in non-critical monitoring areas. They use USB2.0 interfaces, support 1280×720 pixel resolution, 30 frames / second video acquisition rate, and are equipped with 4-8mm wide-angle fixed-focus lenses. The field of view covers 5-20 meters, meeting the needs of macro inspection of general defects.

[0217] Offline batch analysis of video images collected in non-critical areas, using support vector machine (SVM) classifiers based on histogram of oriented gradients (HOG) features for preliminary defect screening and location, and distributed parallel computing through server clusters to improve the processing efficiency of massive data;

[0218] Perform preliminary analysis and feature extraction on monitoring data of non-critical areas at the edge computing node, upload only defect statistics and key features to the cloud, and package them in JSON format to reduce bandwidth and latency of data transmission;

[0219] Design a user-friendly visual human-machine interface, use electronic maps as the base map, overlay power GIS data, and intuitively display the density distribution and risk level of water tree defects in the form of heat maps. It also supports online browsing, query, and retrieval of defect images, making it easier for operation and maintenance personnel to understand the health status of equipment and conduct defect diagnosis and risk warning in a timely manner.

[0220] Here is an example:

[0221] Deploy 2-megapixel industrial cameras based on CMOS sensors in important monitoring areas. The cameras use SonyIMX3271 / 2.8'' sensors with a pixel size of 2.9μm, support 10-bitADC, and a maximum frame rate of 60fps. HR industrial lenses with a focal length of 50mm and an aperture of F1.4 are used. The field of view is about 2.5 meters wide at a distance of 10 meters, the resolution is better than 0.5mm / pixel, and the exposure time is adjustable from 10μs to 1s. The image brightness can be adaptively adjusted through the automatic exposure algorithm.

[0222] According to the distribution characteristics of water tree defects in the shooting area, a training sample library was built, with 2,000 positive samples and 5,000 negative samples. Five data enhancement methods (rotation, translation, scaling, noise, blur) were used to expand the number of samples to 100,000. The ResNet-50 pre-trained on ImageNet was used as the backbone network, and the FPN structure was added to achieve multi-scale feature fusion. At the end of the network, manual features such as texture, edge, and shape were connected in parallel to build a cascade classifier of mixed features. The loss function includes cross entropy loss, IoU loss, and L1 regularization loss. The weight ratio is 1:0.5:0.1, the BatchNorm momentum coefficient is set to 0.9, the initial value of the learning rate is 0.01, and the StepLR strategy is adopted, with a decay of 50% every 10 epochs.

[0223] Intel Arria 10 SoC FPGA is used, equipped with a 925MHz dual-core ARMA9 processor, 2GB DDR4 video memory, and an external M.2 solid-state drive, which can support real-time defect segmentation of 1080P@60fps high-definition video;

[0224] Divide the image into 512×512 pixel blocks, use morphological Gray-scale reconstruction and H-maxima transformation method to adaptively extract the foreground area, calculate the I-order moment and II-order moment of the defect area, extract 12 shape features such as area, aspect ratio, roundness, and concavity of the defect, and achieve a detection speed of 250us / frame;

[0225] The video data is encoded in YUVSemiPlanar4:2:0 format, with constant bit rate factor CRF=23, spatial complexity SC_THRESHOLD=40, minimum quantization factor QP_MIN=10, maximum quantization factor QP_MAX=45, average bit rate controlled at around 3Mbps, and subjective video quality DMOS score of 4.5 points;

[0226] The metadata such as defect coordinates, size, shape features, etc. are packaged in XML format, multiplexed with H.265 raw stream, and transmitted through RTPoverRTMP protocol, with delay jitter less than 50ms;

[0227] The front-end AI acceleration module uses Rockchip RV1126 CPU (4×Cortex-A7@1.5GHz) + 0.8TOPSNPU, runs the lightweight MobileNetV2-SSDLite defect detection model, and can achieve 20fps video analysis. The uploaded data is compressed in ProtocolBuffer hexadecimal encoding and uses the TLS transport layer security protocol. The end-to-end latency is better than 10ms and the packet loss rate is less than one ten-thousandth.

[0228] The front-end interface uses Vue framework, and the back-end uses flask+redis. It uses Baidu Echarts chart library + Amap API to achieve interactive integration of GIS, images, and algorithms.

[0229] Defect information is superimposed on the online electronic map in the form of point, line and surface vector elements, and the defect risk level is represented by a red-yellow-green color scheme. A large database of defects over the entire life cycle of the equipment is generated, supporting three-dimensional retrieval in time, space and semantics.

[0230] Furthermore, in this embodiment, step S6 specifically includes:

[0231] An edge computing device is arranged at a location matching the infrared image acquisition device;

[0232] The edge computing device obtains the original cross-sectional image data collected by the infrared image acquisition device, and preprocesses the original cross-sectional image data. The preprocessing includes denoising, image enhancement and resolution adjustment. Specifically, the denoising algorithm is used to reduce the noise of the original image data to eliminate noise interference in the image and improve the image quality. The denoised image is enhanced according to the image enhancement algorithm to improve the contrast and clarity of the image and highlight the defect characteristics of the insulating material. The resolution of the enhanced image is adjusted by the resolution adjustment algorithm so that the resolution of the image meets the requirements of subsequent defect detection.

[0233] Perform defect detection on the pre-processed cross-sectional images to identify the starting point and extension direction of water treeing in insulating materials;

[0234] If the starting point of the water tree phenomenon is detected, the position coordinates of the starting point are determined and the characteristic information of the starting point is recorded. If the expansion direction of the water tree phenomenon is identified, the angle information of the expansion direction is obtained to determine the degree of danger of the expansion direction;

[0235] Generate a water tree defect report based on the location coordinates of the water tree starting point and the angle information of the expansion direction;

[0236] Here is an example:

[0237] According to the location information of the image acquisition device, an edge computing device is set at a predetermined location 20 meters away from the infrared image acquisition device to achieve position matching between the infrared image acquisition device and the edge computing device;

[0238] Acquire the original cross-sectional image data of the infrared image acquisition device through the edge computing device, and perform preprocessing operations on the original cross-sectional image data;

[0239] The median filter denoising algorithm is used to reduce the noise of the original image data, removing 80% of the Gaussian noise, increasing the signal-to-noise ratio of the image by 5dB, effectively eliminating the noise interference in the image and improving the image quality. The denoised image is enhanced according to the histogram equalization image enhancement algorithm, which increases the image contrast by 30% and the clarity by 20%, and further highlights the defect characteristics of the insulating material;

[0240] The resolution of the enhanced image is adjusted through the bilinear interpolation resolution adjustment algorithm, and the image resolution is adjusted from 1024×768 to 2048×1536, so that the image resolution meets the requirements of subsequent defect detection. Defect detection is performed on the preprocessed image to identify the starting point and expansion direction of the water tree phenomenon in the insulating material. If the starting point of the water tree phenomenon is detected, the position coordinates of the starting point are determined, for example, the starting point coordinates are (120,250), and the characteristic information such as the image grayscale value of 180 and the gradient direction of 45 degrees at the starting point are recorded. If the expansion direction of the water tree phenomenon is identified, the angle information of the expansion direction is obtained, for example, the angle between the expansion direction and the horizontal direction is 30 degrees. According to experience, the expansion direction less than 45 degrees is less dangerous, and the expansion direction greater than or equal to 45 degrees is more dangerous;

[0241] Based on the location coordinates of the starting point of the water tree phenomenon (120, 250) and the angle information of the expansion direction of 30 degrees, a defect report of the water tree phenomenon is generated to provide a decision-making basis for subsequent insulation material maintenance.

[0242] Furthermore, in this embodiment, step S7 also includes:

[0243] Dynamically evaluate the development trend of water tree phenomenon based on water tree risk assessment model and real-time monitoring data;

[0244] Adopt early warning strategies and set early warning thresholds. If the real-time monitoring data does not exceed the early warning threshold, the original state is maintained. If the real-time monitoring data exceeds the early warning threshold, an early warning signal is issued.

[0245] The risk level is obtained according to the water tree risk assessment model, and combined with the development trend of the water tree phenomenon, it is determined whether there is a problem with the water tree phenomenon. If not, the original state is maintained. If so, the corresponding emergency response measures are initiated according to the risk level;

[0246] Specifically, the electrical stress data of the cable is obtained, including voltage, current, and frequency, and the electric field strength and heat loss power of the cable insulation material are calculated based on the electrical stress data to determine whether the cable insulation state is abnormal;

[0247] The pre-processed cross-sectional image data and electrical stress data are input into the deep learning algorithm model, the feature information of the image is extracted through the convolutional neural network, and the time series characteristics of the electrical stress data are analyzed using the recursive neural network to comprehensively judge whether the cable insulation material has water tree phenomenon and defects;

[0248] If the deep learning algorithm identifies the presence of water tree phenomena or defects in the cable insulation material, the water tree risk assessment model is used to calculate the risk level of the water tree phenomenon based on the type, location and severity of the defect, and the development trend prediction result of the water tree phenomenon is obtained;

[0249] According to the risk level and development trend prediction results of water tree phenomenon, the risk warning threshold is adjusted dynamically. When the real-time monitoring data exceeds the warning threshold, the warning signal is triggered to notify the operation and maintenance personnel to conduct on-site inspection and maintenance;

[0250] If the risk level of the water tree phenomenon is high and the development trend prediction results show that the water tree phenomenon may cause cable insulation breakdown, the emergency response measures will be automatically initiated to adjust the cable load, reduce the electrical stress of the cable, and slow down the development of the water tree phenomenon;

[0251] If the risk level of water tree phenomenon is still high after the cable load is adjusted, the backup line will be automatically switched to disconnect the faulty cable to avoid insulation breakdown accidents caused by water tree phenomenon;

[0252] Based on the water tree phenomenon and defect location information identified by the deep learning algorithm, a cable defect distribution map is generated to guide operation and maintenance personnel to conduct targeted on-site inspections and repairs, thereby improving maintenance efficiency and accuracy;

[0253] Here is an example:

[0254] Examples of image preprocessing algorithms for denoising and enhancing cross-sectional images:

[0255] The median filter algorithm is used to denoise the infrared image of 256x256 pixels, and the filter window size is 5x5. The signal-to-noise ratio of the processed image is improved by 3dB;

[0256] Example of calculating the electric field strength and heat loss power of a cable insulation:

[0257] The working voltage of the cable is 10kV, the conductor cross-sectional area is 50 square millimeters, and the thickness of the insulating material is 5 millimeters, then the electric field strength is about 2kV / mm. If the cable current is 100A and the dielectric loss factor of the insulating material is 0.1, then the heat loss power is about 10W;

[0258] A convolutional neural network consisting of three convolutional layers and two fully connected layers was used to train 1,000 infrared images, identifying water trees in 80 of them with a positioning accuracy of 90%.

[0259] Based on the area, shape, location and other characteristics of the water tree phenomenon, the risk level is calculated to be 8, which is high risk. The probability of insulation breakdown in the next year is predicted to be 60%;

[0260] When the risk level of water tree phenomenon exceeds 7, the warning threshold is adjusted from 8 to 75, and a warning signal is issued 3 days in advance;

[0261] When the risk level of water tree phenomenon reaches 85 and still exceeds 8 after load adjustment, 80% of the load will be automatically switched to the backup line and the faulty cable will be disconnected within 30 minutes;

[0262] A 10-kilometer-long cable was scanned, 20 water tree phenomena and defects were identified, and a defect distribution map was generated based on their locations. The defect type and risk level were marked with an accuracy of 1 meter to guide operation and maintenance personnel in repairs;

[0263] 1,000 historical water tree phenomenon cases were input into the decision tree algorithm for training to generate optimized risk assessment rules, reducing the false alarm rate of the original model from 5% to 2%, thereby improving the accuracy of the early warning.

[0264] The present invention further proposes a cable identification system based on image processing, which adopts the cable identification method based on image processing as described above, and comprises:

[0265] An image processing detection module, which is used for denoising, adjusting resolution and identifying defects in cable insulation materials;

[0266] Image analysis module: The image analysis module is used to segment the processed image and mark the defect location, size and shape;

[0267] Water tree phenomenon prediction and risk assessment module: The water tree phenomenon prediction and risk assessment module is used to combine dynamic monitoring data to predict and assess the water tree risk level;

[0268] Monitoring and emergency response module,The monitoring and emergency response module is used to issue warnings based on risk levels, dynamically manage the cable identification system and trigger emergency measures.

[0269] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention.

[0270] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. A cable identification method based on image processing, characterized in that: include: S1, collecting and processing cross-sectional images of cable insulation materials; S2, performing image segmentation and defect detection on the processed cross-sectional image, identifying and marking defects in the image, and locating the starting point of the water tree phenomenon according to the marked defects; S3. According to the starting point of the water tree phenomenon, combined with the dynamic monitoring data of electrical stress and the influence of various factors on the development of the water tree phenomenon, a water tree phenomenon generation and development prediction model is established; S4, using adaptive optical compensation method and image enhancement method to dynamically adjust imaging parameters in real time; S5. Adopt a multi-resolution image acquisition strategy, combine the importance of the cable area and historical water tree risk data, and dynamically adjust the image resolution; S6, collecting original cross-sectional image data according to real-time dynamic adjustment of imaging parameters and dynamic adjustment of image resolution, and performing preprocessing; S7, real-time monitoring and analysis of image data and electrical stress data, combined with the cross-sectional image data pre-processed in step S6, to identify potential water tree phenomena and insulation material defects; The step S3 specifically includes: Various factors affect the development of water tree phenomena including different voltages, frequencies and ambient temperatures; Collect electrical stress data during cable operation, combine it with static data of cable lines, and use data cleaning and feature engineering methods to preprocess the data to obtain a multi-source heterogeneous monitoring data set. Obtain the temperature data of the cable environment, combine the cable's geographical location information and meteorological data, build a spatiotemporal distribution model of the cable environment temperature, and incorporate the temperature distribution characteristics into the feature extraction process of the water tree defect area; The water tree starting point identification result is associated with the electrical stress monitoring data and the ambient temperature data, the time-frequency domain characteristics of the defect area centered on the water tree starting point are extracted, and the multi-dimensional feature vector of the water tree defect area is constructed; By using time series analysis and association rule mining algorithms, combined with the association laws and evolution trends between the characteristic parameters of the water tree defect area, the quantitative relationship between voltage, frequency, temperature and water tree development is obtained, and a water tree phenomenon generation and development prediction model based on time series regression is established; Electrical stress sensors are set up to monitor stress changes inside and outside the cable in real time, and the stress change data is correlated with the identified defect locations for analysis, a stress distribution model is established, and predetermined risk stress characteristics that lead to water treeing are identified; By adopting parallel computing analysis methods, combined with defect locations and predetermined risk stress characteristics, the potential risk of each defect location is evaluated, a water tree risk assessment model is constructed, and the water tree phenomenon in each area is divided into risk levels; The step S7 further comprises: Dynamically assess the development trend of the water tree phenomenon based on the water tree risk assessment model and real-time monitoring data; Adopt an early warning strategy and set an early warning threshold. If the real-time monitoring data does not exceed the early warning threshold, the original state is maintained. If the real-time monitoring data exceeds the early warning threshold, an early warning signal is issued. The risk level is obtained according to the water tree risk assessment model, and combined with the development trend of the water tree phenomenon, it is determined whether there is a problem with the water tree phenomenon. If not, the original state is maintained. If so, the corresponding emergency response measures are initiated according to the risk level.

2. The cable identification method based on image processing according to claim 1, characterized in that: The step S1 specifically includes: Using an infrared image acquisition device, setting and adjusting parameters of the infrared image acquisition device, and acquiring cross-sectional image data covering the entire cross-sectional area of ​​the cable insulation material; Processing the collected cross-sectional image data, including removing salt and pepper noise and removing Gaussian noise in the image; According to the characteristics of the cross section of the cable insulation material, the cross section image is segmented and the image data of the area of ​​interest is extracted; The segmented cross-sectional images are corrected to eliminate image distortion.

3. The cable identification method based on image processing according to claim 1, characterized in that: The step S2 specifically includes: A deep learning algorithm is used to perform semantic segmentation on the processed cross-sectional images and a segmentation model is established; By extracting features and classifying pixels, the image is divided into defective areas and normal insulating material areas. The defect types in the defective areas include bubbles, impurities, cracks and deformations. Using data enhancement and transfer learning methods, the segmentation model is pre-trained according to the existing defect sample data set; According to the segmentation results, the connected domain analysis algorithm based on region growing is used to mark and number various defect regions. The contour curve of each defect region is obtained by boundary tracing, and the area, perimeter and roundness shape feature parameters of the defect region are calculated using geometric shape descriptors. According to the defect type identification rules, the defect area is classified, the defect type is identified, and the severity of the defect is evaluated and graded according to the area threshold and roundness threshold; The DBSCAN density clustering algorithm is used to perform cluster analysis on the marked defect areas. Adjacent defects are classified into defect groups according to the spatial distance and distribution density between defects. The morphological skeleton extraction algorithm is used to obtain the skeleton curve of each defect group as the center line of the defect group. The topological network diagram of defect distribution is constructed to reveal the expansion trend and connection path of defects on the cross section of the insulating material. Graph theory analysis is performed on the defect distribution topology map. All defect extension paths are traversed using the depth-first search algorithm. The defect area growth rate, path length, and number of forks are used to score the degree of defect evolution of each path, thereby locating the starting point of the water tree phenomenon.

4. The cable identification method based on image processing according to claim 1, characterized in that: The step S4 further comprises: According to the adaptive optical compensation method, the exposure time, aperture size and image enhancement parameters are automatically adjusted in combination with changes in ambient light and imaging conditions.

5. The cable identification method based on image processing according to claim 1, characterized in that: The step S5 specifically includes: The cable line nodes are divided into three levels: important nodes, general nodes and auxiliary nodes. The hierarchical analysis method is used to comprehensively evaluate the importance of nodes. A node importance evaluation system is constructed, which includes operating voltage level, transmission power, power supply range, geographical location and water tree fault history indicators. The importance weight of each node is obtained through indicator weight solution and consistency test. The clustering algorithm is used to classify the nodes to form the node importance division result.

6. The cable identification method based on image processing according to claim 2, characterized in that: The step S6 specifically includes: An edge computing device is arranged at a position matching the infrared image acquisition device; The edge computing device acquires the original cross-sectional image data acquired by the infrared image acquisition device, and preprocesses the original cross-sectional image data, wherein the preprocessing includes denoising, image enhancement, and resolution adjustment; Defect detection is performed on the preprocessed cross-sectional images to identify the starting point and extension direction of the water tree phenomenon in the insulating material.

7. A cable identification system based on image processing, using a cable identification method based on image processing as claimed in any one of claims 1 to 6, characterized in that: include: An image processing detection module, which is used for denoising, adjusting resolution and identifying defects in cable insulation materials; An image analysis module, which is used to segment the processed image and mark the defect location, size and shape; A water tree phenomenon prediction and risk assessment module, which is used to predict and assess the water tree risk level in combination with dynamic monitoring data; The monitoring and emergency response module is used to issue early warnings based on risk levels, dynamically manage the cable identification system and trigger emergency measures.

Images