CN118898475A - A heuristic algorithm-based method for generating maintenance strategies for distribution networks - Google Patents

Abstract

The present invention discloses a distribution network maintenance strategy generation method based on a heuristic algorithm, including multi-dimensional health status indicators such as the aging degree of the surface of the insulator in the distribution network, the crystal structure of the glaze layer, the geometric size deviation, the surface cracks and the temperature distribution, etc., to evaluate the aging degree of the coating and judge whether the coating needs to be replaced or the glaze layer needs to be repaired, and by analyzing the geographical location of the pole tower where the insulator is located, the surrounding vegetation coverage and the electromagnetic interference level, the maintenance area is divided and the optimal maintenance path is determined; during the ground inspection, the three-dimensional point cloud data of the transmission line is obtained to judge whether the insulator is damaged or contaminated; the width and length of the surface cracks are measured, the crack expansion trend and the remaining life are predicted, the maintenance time window is formulated and the maintenance resources are dispatched; after the insulator is replaced or repaired, its temperature distribution is detected to judge the installation quality; finally, through multi-dimensional data analysis, the optimal maintenance decision sequence for the entire life cycle of the insulator is generated.

Description

A heuristic algorithm-based method for generating maintenance strategies for distribution networks

Technical Field

The present invention relates to the field of information technology, and in particular to a method for generating a distribution network maintenance strategy based on a heuristic algorithm.

Background Art

There is a technical contradiction in the maintenance process of insulators in the distribution network, that is, how to minimize the negative impact of the maintenance process on the environment and human health while ensuring the quality and efficiency of maintenance. In order to accurately identify the hot spots, potential fault locations and corona discharge of insulators, high-precision temperature sensors and corona discharge detection equipment are needed to monitor the status of insulators in real time, and conduct fine-grained analysis and evaluation of the maintenance needs of each insulator based on the monitoring data. When implementing power outage maintenance, dismantling old insulators, and replacing new insulators, noise and electromagnetic radiation emissions will inevitably be generated, which will have a certain impact on the surrounding environment and human health. When formulating maintenance plans, it is also necessary to comprehensively consider the dependencies between various maintenance strategies, the urgency of maintenance tasks, and the availability of maintenance resources, and reasonably arrange the execution order and resource allocation of maintenance tasks to achieve a balance between maintenance efficiency and cost.

Summary of the invention

In order to solve the above-mentioned technical problems, the present invention provides a method for generating a distribution network maintenance strategy based on a heuristic algorithm.

The technical solution of the present invention is implemented as follows: a method for generating a distribution network maintenance strategy based on a heuristic algorithm, comprising:

Analyze the aging degree of the coating on the surface of the insulator in the distribution network, including using Fourier transform infrared spectroscopy analysis technology to obtain the chemical bond vibration characteristic data of the coating material, build an aging degree assessment model, determine whether the aging degree of the insulator is greater than a threshold, and if so, identify that the coating needs to be replaced, enter the insulator number and location information into the fault maintenance database, and determine the maintenance priority based on the aging degree of the coating;

An X-ray diffractometer is used to analyze the physical composition of the glaze layer of the insulator, and the diffraction spectrum data about the crystal structure of the glaze layer is obtained. The feature extraction and classification of the diffraction spectrum data are performed through the convolutional neural network model to determine whether the glaze layer is distorted. If it is identified as distorted, the insulator number and location information are entered into the fault maintenance database, and the glaze repair or replacement plan is determined according to the type of distortion;

For each insulator to be repaired in the fault maintenance database, by analyzing the geographical location of the tower and the electromagnetic interference level, the insulator to be repaired is divided into multiple maintenance areas through the fuzzy C-means clustering module, and the optimal maintenance path for each maintenance area is determined. The optimal maintenance path is sent to the maintenance personnel's terminal, and the maintenance personnel are located and tracked in real time through the location perception module to ensure that maintenance is carried out along the optimal maintenance path;

During the maintenance process, the photogrammetry module is used to obtain image data of the transmission line, extract the geometric model of the insulator, and determine whether the insulator has defects, including damage or contamination. If so, the defect type and location coordinates are uploaded to the fault maintenance database, and the local cleaning or replacement process is triggered;

For defective insulators, the width and length of the cracks on the surface of the insulator are measured, and the crack propagation path is determined through the ant colony search module. The crack propagation trend and the remaining life of the insulator are predicted. If the remaining life is lower than the threshold, the insulator number and location information are entered into the fault maintenance database, and the maintenance time window is determined according to the crack propagation rate. The maintenance resources are scheduled through the heuristic rule module to minimize the risks brought by crack propagation.

After the insulator is replaced or repaired, the temperature distribution on the insulator surface is detected through the thermal image data on the insulator surface, the frequency domain features of the thermal image data are extracted, and the current quality of the insulator is judged using a convolutional autoencoder network. If there is an abnormal temperature distribution, the thermal image data is uploaded to the fault maintenance database, and the repair process is triggered to dynamically adjust the maintenance strategy.

Based on the multidimensional data of insulators, including coating aging degree, glaze distortion, surface cracks, and temperature anomaly, the comprehensive health index of insulators is obtained. The Markov chain prediction model is used to estimate the state transition probability under different maintenance strategies, and the multi-objective particle swarm optimization module is used to generate the optimal maintenance decision sequence for the entire life cycle of insulators.

Beneficial Effects

The invention provides a method for generating a distribution network maintenance strategy based on a heuristic algorithm. Through Fourier transform infrared spectroscopy analysis and X-ray diffraction technology, the aging degree of insulator coating and glaze distortion can be non-destructively evaluated, potential problems can be identified early, and power outages caused by insulator failures can be reduced. The application of fuzzy C-means clustering and location sensing technology can optimize maintenance path planning, ensure the effective allocation of maintenance resources, and monitor the activities of maintenance personnel in real time to improve work efficiency and safety. Insulator defects can be automatically detected by combining photogrammetry and deep learning technology, which reduces omissions in manual inspections and shortens response time. The use of ant colony search and multi-objective optimization models can further improve the The intelligent level of complex crack processing ensures the scientific and economical maintenance decisions; through the simulation of crack propagation and the prediction of remaining life, combined with the Markov chain model, the transformation from passive response to faults to active preventive maintenance is realized, effectively extending the service life of insulators and reducing maintenance costs; the application of multi-objective particle swarm optimization technology comprehensively considers various factors to formulate the optimal maintenance decision sequence, ensuring the optimal execution of maintenance work under limited resources, while reducing the system risks caused by insulator failures; integrating multi-dimensional insulator data, constructing a comprehensive health index, providing each insulator with a full-cycle health management plan from installation to retirement, and enhancing the comprehensive management level of power grid assets.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a structural block diagram of a method for generating a distribution network maintenance strategy using a heuristic algorithm in an embodiment of the present invention;

FIG2 is a flowchart of a method for generating a distribution network maintenance strategy using a heuristic algorithm in an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

The preferred implementation methods of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element, or connected to the other element through an intermediate element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc. if there is transmission of electrical signals or data between the connected objects.

When used herein, the singular forms "a", "an", and "said/the" may also include plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "include/comprise" or "have" and the like specify the presence of stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. At the same time, the term "and/or" used in this specification includes any and all combinations of the relevant listed items.

Referring to FIG. 1-2 , a method for generating a distribution network maintenance strategy using a heuristic algorithm in this embodiment specifically includes:

The S101 step analyzes the aging degree of the coating on the surface of the insulator in the distribution network, including using Fourier transform infrared spectroscopy analysis technology to obtain chemical bond vibration characteristic data of the coating material, constructing an aging degree assessment model, and judging whether the aging degree of the insulator is greater than a threshold value. If so, it is identified as needing to replace the coating, and the number and location information of the insulator are entered into a fault maintenance database, and the maintenance priority is determined according to the aging degree of the coating.

Specifically, the Fourier transform infrared spectroscopy analysis technology is used to analyze the chemical composition of the insulator surface coating material, obtain the chemical bond vibration characteristic data of the coating material, take a small amount of sample from the insulator surface coating, grind it into powder with a mortar, press the powder sample into a tablet, and put it into the FTIR instrument for testing. The test data is baseline corrected and smoothed using FTIR spectrum analysis software, and then compared with the data in the standard spectrum library to calculate the similarity. The similarity is measured using Euclidean distance and correlation coefficient indicators. The degree of aging of the coating material is judged based on the similarity. The smaller the similarity, the greater the change in the chemical composition of the material and the higher the degree of aging. A threshold is set. When the similarity is low, the value is calculated. When the threshold is exceeded, the coating is considered to be seriously aged and needs to be replaced. By analyzing the chemical composition and aging characteristics of different materials, a coating material performance prediction model is established to select the optimal coating material. A coating aging degree assessment model based on a convolutional neural network is constructed. The chemical bond vibration characteristics obtained by FTIR analysis are spliced with the surface image features of the insulator coating to form a complete sample feature input. The aging degree category is used as the output. The cross entropy loss function and Adam optimizer are used for training. A large number of insulator coating surface images and corresponding aging degree annotation data are collected and divided into training set, validation set and test set. The pre-trained CNN model is used to extract features from the images to obtain high-dimensional feature vectors.

Build a CNN model, continuously optimize the model performance by adjusting the hyperparameters, evaluate the model's accuracy, recall rate, and F1 value indicators on the validation set, and perform error analysis to find out the shortcomings of the model. Use the early stopping method to prevent overfitting, apply the trained model to the test set, output the aging degree prediction results of each sample, compare them with the manual annotation results, calculate the accuracy of the model, set the coating aging degree threshold, compare the aging degree output by the evaluation model with the threshold, and if the aging degree exceeds the threshold, enter the number and location information of the insulator that needs to be replaced into the fault maintenance database, determine the maintenance priority according to the aging degree of the coating, and the higher the aging degree, the higher the priority of the insulator. The earliest deadline first (EDF) algorithm based on deadline is used to prioritize and schedule maintenance tasks. According to the coating aging degree, the priority of each maintenance task is calculated. A mapping function is set to map the aging degree to a priority value. According to the priority and duration of the maintenance task, its deadline is calculated. The deadline = current time + duration * (1 + priority factor). The priority factor is set according to the specific situation, such as 0.2. According to the deadline, all maintenance tasks are sorted in ascending order. According to the sorted order, the maintenance tasks are assigned to maintenance personnel in turn. After the maintenance personnel complete a task, they request the next task to be assigned until all tasks are completed.

After determining the maintenance priority, robotic technology is used to automatically complete the insulator coating replacement operation, visual guidance and force control technology are used to achieve precise cleaning and re-spraying of the coating, the automatic proportioning system of the coating material is used to realize the configuration of the new coating material, and the quality of the new coating is inspected by the quality inspection system. An insulator full life cycle management system is established to record the entire process data of the insulator from installation, operation, maintenance to scrapping.

In one embodiment, when performing chemical composition analysis on the surface coating material of the insulator, a Fourier transform infrared spectrometer Nicoleti S50 can be used, and the scanning wave number range is set to 400-4000 cm-1, the resolution is 4 cm-1, and the number of scans is 32 times. The test data is processed using OMNIC software, and the similarity index is calculated by comparing with the built-in polymer additive spectrum library. When the similarity is lower than 0.8, it is determined that the coating is severely aged and needs to be replaced. At the same time, a coating material performance prediction model is established using principal component analysis (PCA) and support vector machine (SVM) algorithms. By analyzing the infrared spectral characteristics and aging test data of different materials, the epoxysilicone coating material with the best comprehensive performance is selected. When constructing a convolutional neural network model, ResNet-50 is used as the backbone network, pre-trained on the ImageNet dataset, and then transfer learning is performed on the coating surface image dataset. By setting the learning rate to 0.001 and the batch size to 0.01, the performance prediction model is established. The number of training epochs is 32, and 100 epochs are trained. The accuracy rate is 98% on the validation set. The model is deployed on the embedded device Jetson Xavier NX to monitor the coating aging in real time. When the aging degree exceeds 0.7, the maintenance task is automatically triggered and the task information is written into the MySQL database. When scheduling tasks, the priority and deadline of the maintenance task are calculated according to the coating aging rate and the estimated remaining life. The improved EDF algorithm is used to sort and allocate tasks. The reasonable allocation of maintenance resources is achieved through the dynamic priority adjustment mechanism. For coating replacement operations, the ABBIRB1200 robot is used, equipped with a high-precision laser scanner and a spraying end effector. Automatic positioning and trajectory planning are achieved through machine vision guidance. The closed-loop control algorithm is used to achieve precise control of coating thickness and uniformity. The spraying accuracy reaches ±0.1mm. In the insulator full life cycle management system, distributed data acquisition and edge computing technology are used to achieve real-time monitoring of insulator status and fault warning.

The step S102 uses an X-ray diffractometer to analyze the physical composition of the glaze layer of the insulator, obtains diffraction spectrum data about the crystal structure of the glaze layer, extracts and classifies the diffraction spectrum data through a convolutional neural network model, and determines whether the glaze layer is distorted. If it is identified as distorted, the insulator number and location information are entered into a fault maintenance database, and a glaze repair or replacement plan is determined according to the type of distortion.

Specifically, an X-ray diffractometer is used to analyze the phase composition of the glaze layer of the insulator to obtain the diffraction spectrum data of the crystal structure of the glaze layer. The setting parameters of the X-ray diffractometer include X-ray wavelength, incident angle, scanning step and scanning speed. Through data preprocessing and feature engineering, the key characteristic parameters of the diffraction spectrum are extracted, such as the diffraction peak position corresponding to the crystal plane spacing, the peak intensity corresponding to the crystal plane orientation, the peak width corresponding to the grain size and the background intensity corresponding to the amorphous phase content. The extracted features are deeply learned and classified using a convolutional neural network model. By training normal glaze diffraction spectrum samples, a multi-classification distortion recognition model is established to determine whether the glaze layer is distorted. The convolutional neural network model adopts the LeNet-5 or AlexNet structure, the input layer is the diffraction spectrum image, the convolution layer and the pooling layer are used to extract the local features of the image, and the fully connected layer is used for feature integration and classification judgment. For the main distortion type of the glaze layer, the mullite crystal phase , quartz crystal phase, crystallization, residual stress, set corresponding labels, and perform multi-classification training. The model's hyperparameters including convolution kernel size, number of convolution layers, and number of nodes are optimized through the gridsearch method. When the glaze layer is identified to be deformed, the insulator number and location information are automatically obtained and entered into the fault maintenance database. At the same time, according to the type and severity of the distortion, a glaze layer repair or replacement plan is intelligently generated. The association rule mining Apriori algorithm, decision tree ID3, and C4.5 algorithm data mining methods are used to analyze historical maintenance data, find out the correspondence between different distortion types and maintenance measures, and form an IF-THEN rule base. When generating a repair and replacement plan, the historical case most similar to the current distortion type is found through rule matching and reasoning, and its maintenance measures are associated. For the glaze layer repair plan, image segmentation algorithms including threshold segmentation and edge detection are used. The system automatically detects and locates the surface defects of the glaze layer through measurement and regional growth, separates the defective area from the normal area, extracts the geometric features of shape, size and direction of the segmented defective area, and judges the types of defects including cracks, bubbles and peeling through decision tree or support vector machine classification algorithm. According to the defect type and characteristics, the process parameters of laser cladding or plasma spraying including power, speed and powder feeding rate are automatically generated, and the robot is used to accurately clad or spray the defective area. For the glaze replacement plan, big data analysis technology is used to comprehensively consider the operating conditions and environmental conditions of the insulator to select the best glaze material formula. The intelligent glaze configuration system is used to automatically batch, mix and prepare the glaze according to the preferred formula to ensure the consistency of the glaze performance. The old glaze layer is removed by laser cleaning and water jet cleaning, and the new glaze layer is coated on the insulator by robot automatic spraying technology. , the motion trajectory and spraying parameters of the spray gun are controlled by computer to achieve precise control of the thickness and uniformity of the glaze layer. The insulators after spraying need to be sintered in a tunnel kiln. The glaze layer and the porcelain body are fully combined by controlling the temperature and atmosphere parameters. In the process of glaze repair and replacement, online monitoring and quality inspection technology are used to perform real-time detection of key indicators of glaze thickness, uniformity, and insulation resistance. Once deviations are found, the process parameters are optimized and adjusted immediately to ensure that the quality of repair and replacement meets the standard requirements. An insulator online monitoring system based on the Internet of Things is established. By installing temperature, vibration, and strain wireless sensors on the insulators, the state parameters of the glaze layer are collected in real time. The monitoring data is analyzed in real time through edge computing nodes, and a correlation model between temperature, vibration, strain and glaze distortion is established. When the monitoring parameters exceed the normal range or abnormal fluctuations occur, timely warning of the risk of glaze distortion is issued to trigger maintenance tasks.

In one embodiment, the setting parameters of the X-ray diffractometer include X-ray wavelength CuKα,λ=0.15406nm, incident angle 10°-90° scanning step length 0.02° and scanning speed 5°/min, and the relevant information of the insulator to be repaired is extracted from the fault maintenance database, including the geographical location coordinates of the tower, such as longitude 1123 degrees, latitude 312 degrees, altitude 25 meters, and the power frequency electric field strength of the surrounding environment is 35V/m, the radio field strength is 2V/m, and the magnetic field strength is 8A/m. The ArcGIS system is used to perform spatial visualization analysis on the tower position, and the distance matrix between the towers is calculated by the Geodesic distance calculation function. For example, the distance between tower A and tower B is 153 meters, and the fuzzy C-means clustering algorithm is used to divide the insulator to be repaired into maintenance areas. , we select the geographical location of the tower and the electromagnetic interference intensity as clustering features, and use the Mahalanobis distance to calculate the similarity between insulators. The Mahalanobis distance takes into account the correlation between features and can better reflect the real differences between insulators. We use the Gaussian membership function, and the membership value range is from 0 to 1. The larger the membership, the more likely the insulator belongs to this category. We calculate the membership matrix based on the Mahalanobis distance between the insulator and the cluster center. We determine the region to which each insulator belongs by the maximum membership principle, and calculate the center coordinates of each region. We use the FHI index to evaluate the quality of the clustering results. The FHI index comprehensively considers the cluster compactness and cluster separation. The larger the value, the better the clustering effect. We compare the results of different cluster numbers using the FHI index. When the number of clusters is 5, the FHI index reaches the maximum value of 92, so the insulator to be repaired is selected. The system is divided into five maintenance areas. In each maintenance area, the improved ant colony optimization algorithm is used to search for the best maintenance path. The maintenance area is abstracted into an undirected weighted graph, with insulators as nodes and the distances between insulators as edge weights. At the same time, the electromagnetic interference intensity between insulators is used as heuristic information. The greater the interference intensity, the greater the edge weight, which attracts ants to search. Global pheromones and local pheromones are introduced. The initial value of the global pheromone concentration is set to 2, the initial value of the local pheromone concentration is set to 5, the pheromone volatility coefficient is set to 1, and the local pheromone attenuation coefficient is set to 7. The ant population diversity evaluation index is introduced, and the Simpson index is used to quantify the population diversity. When the Simpson index is lower than 6, 20% of the ant individuals are regenerated through mutation operations. The ant colony algorithm converges to the global optimal solution after 500 iterations, and each The best maintenance path in the maintenance area is converted into JSON format data, including the maintenance area ID, path node longitude and latitude, node distance, and estimated time-consuming attributes. The JSON data is pushed to the /api/inspection/tasks interface of the mobile terminal APP through an HTTPPOST request, and different maintenance areas and paths are marked with different colors on the AutoNavi Map API. Maintenance personnel query task details through voice interaction functions, such as "query today's maintenance tasks" and "navigate to the next maintenance point". The mobile terminal APP integrates a location perception module, uses GPS/Beidou dual-mode positioning, and combines the Chinese precision CORS system for RTK differential positioning. The horizontal positioning accuracy is better than 5cm, and the vertical positioning accuracy is better than 10cm. At the same time, the extended Kalman filter algorithm is used , integrating GPS positioning data and mobile phone IMU inertial navigation data, the position update frequency reaches 20Hz, the position perception module collects on-site images through the mobile phone camera, and uses the YOLOv5s target detection algorithm based on MobileNet-SSD to detect and locate insulators in real time. The average detection accuracy mAP reaches 95%, and the detection speed reaches 25FPS. The front image of the insulator is extracted through affine transformation and perspective transformation, and sent to the CNN-based Tesseract0OCR model for nameplate information recognition. The average recognition accuracy reaches 95%. Finally, the recognition result is compared with the asset management system data to verify the correctness of the maintenance object. The system matches the trajectory data with the optimal maintenance path, and uses the dynamic time warping (DTW) algorithm to calculate the trajectory deviation. Once the deviation exceeds 1 0 meters, and promptly send early warning prompts to maintenance personnel, prompting them to adjust the maintenance path until all maintenance tasks are completed. The verification results will also be fed back to maintenance personnel in real time to avoid maintenance errors. When performing X-ray diffraction analysis on the insulator glaze layer, a BrukerD8Advance diffractometer is used, and the CuKα radiation source wavelength is set to 0.15406nm, the scanning range is 10°-90°, the scanning step is 0.02°, and the scanning speed is 5°/min. The JADE6.5 software is used to perform background subtraction, peak search and crystal phase calibration on the diffraction spectrum. By comparing with the ICDDPDF-2 database, the phase composition of the glaze layer is determined, and the half-height width of the diffraction peak is calculated using the Scherrer formula to estimate the grain size of the glaze layer. The diffraction spectrum data is converted into a grayscale image of 1024×1024 pixels. , and sent it to the pre-trained AlexNet network for feature extraction and classification judgment. The input layer of the network is 227×227 pixels. The convolution kernel sizes of the five convolutional layers are 11×11×3, 5×5×64, 3×3×192, 3×3×384, and 3×3×256, respectively. The number of nodes in the three fully connected layers is 4096, 4096, and 8, respectively, corresponding to eight typical types of glaze distortion. The network is trained using the Adam optimizer and the cross entropy loss function. The batch size is 128, the initial learning rate is 0.001, the momentum factor is 0.9, and the regularization coefficient is 0.0005. When the accuracy on the validation set does not improve for five consecutive epochs, the early stopping mechanism is triggered, and the current optimal model parameters are saved. The SEM image of the glaze surface is segmented by Ostu threshold to extract the morphological features of the defective area. The defect types are classified into five categories including cracks, bubbles, peeling, pinholes and impurities using the CART decision tree algorithm, and the processing parameters of laser cladding are automatically generated according to the defect size and quantity, such as laser power 300W, spot diameter 0.5mm, scanning speed 800mm/min, and powder feeding rate 12g/min. The AdaBoost algorithm is used to evaluate the comprehensive performance of different material formulations, and the optimal glaze composition is automatically screened out to achieve intelligent optimization of the glaze formulation. The time domain signal collected by the online monitoring system is converted to the frequency domain through fast Fourier transform, and the characteristic parameters of the spectrum, such as root mean square value, kurtosis and skewness, are extracted. The wavelet neural network is used to establish the mapping relationship between the characteristic parameters and the degree of damage of the glaze layer, so as to achieve real-time evaluation of the health status of the glaze layer and prediction of the remaining life.

S103. For each insulator to be repaired in the fault maintenance database, by analyzing the geographical location of the tower and the electromagnetic interference level, the insulator to be repaired is divided into multiple maintenance areas through the fuzzy C-means clustering module, and the optimal maintenance path for each maintenance area is determined. The optimal maintenance path is sent to the maintenance personnel's terminal, and the maintenance personnel is located and tracked in real time through the position perception module to ensure that maintenance is carried out along the optimal maintenance path.

Specifically, the relevant information of the insulator to be repaired is extracted from the fault maintenance database, including the geographical coordinates of the tower including longitude and latitude, elevation, and the electromagnetic interference intensity of the surrounding environment including power frequency electric field, radio field, and magnetic field. The GIS system is used to perform spatial visualization analysis of the tower position, and the distance matrix between the towers is calculated as the input of fuzzy C-means clustering. The fuzzy C-means clustering algorithm is used to divide the maintenance area of the insulator to be repaired. The geographical location of the tower and the electromagnetic interference intensity are selected as clustering features. The Euclidean distance or Mahalanobis distance is used to calculate the similarity between the insulators. The exponential membership function is used to calculate the membership matrix according to the distance between the insulator and the cluster center, and the maximum membership principle is used to determine the distance matrix of each insulator. The maintenance area is assigned to each maintenance area, and the center coordinates of each area are calculated. The Xie-Beni index or FHI fuzzy hyperspherical shell index is used to evaluate the quality of the clustering results, and the optimal number of clusters is selected. In each maintenance area, the improved ant colony optimization algorithm is used to search for the best maintenance path. The maintenance area is abstracted as an undirected weighted graph, with insulators as nodes and the distances between insulators as edge weights. At the same time, the electromagnetic interference intensity between insulators is used as heuristic information. The greater the interference intensity, the greater the edge weight, which attracts ants to search. Global pheromones and local pheromones are introduced to enhance the global search ability and local development ability of the algorithm. Ant population diversity evaluation indicators are introduced. When the population tends to be homogenized, a part of the ant individuals are regenerated through mutation operations. The ant colony algorithm continuously optimizes the search path through the pheromone update mechanism until it converges to the global optimal solution, and converts the best maintenance path into JSON format data, including the maintenance area ID, path node longitude and latitude, node distance, and estimated time-consuming attributes. The JSON data is pushed to the specified API interface of the mobile terminal APP through the HTTP POST request, and different maintenance areas and paths are marked with different colors on the electronic map. Maintenance personnel query the task details through the voice interaction function and perform on-site operations according to the navigation prompts. The mobile terminal APP integrates a location perception module, adopts GPS/Beidou dual-mode positioning, and combines base station differential positioning (RTK) technology to achieve centimeter-level high-precision positioning. At the same time, it adopts card Man filter algorithm integrates GPS positioning data and inertial navigation data to achieve smooth trajectory tracking. The position perception module collects on-site images through the mobile phone camera, and uses the YOLOv5 target detection algorithm based on deep learning to detect and locate the insulator in real time. The front image of the insulator is extracted through affine transformation and perspective transformation, and sent to the OCR model for nameplate information recognition. Finally, the recognition result is compared with the asset management system data to verify the correctness of the maintenance object. The system matches the trajectory data with the optimal maintenance path and calculates the trajectory deviation. Once the deviation exceeds the threshold, an early warning prompt is issued to the maintenance personnel in time until all maintenance tasks are completed. The verification results will also be fed back to the maintenance personnel in real time to avoid maintenance errors.

In one embodiment, relevant information of the insulator to be repaired is extracted from the fault maintenance database, including the geographical coordinates of the tower where it is located, such as longitude of 1123 degrees, latitude of 312 degrees, altitude of 25 meters, and the power frequency electric field strength of the surrounding environment of 35V/m, radio field strength of 2V/m, and magnetic field strength of 8A/m. The ArcGIS system is used to perform spatial visualization analysis on the tower position, and the distance matrix between the towers is calculated by the Geodesic distance calculation function. For example, the distance between tower A and tower B is 153 meters. The fuzzy C-means clustering algorithm is used to divide the insulators to be repaired into maintenance areas, and the geographical location of the tower and the electromagnetic interference intensity are selected as clustering features. The Mahalanobis distance is used to calculate the similarity between insulators. The Mahalanobis distance takes into account the correlation between features and can better reflect the real differences between insulators. The Gaussian membership function is used, and the membership value range is 0 to 1. The larger the membership degree, the more likely the insulator belongs to this category. The membership matrix is calculated based on the Mahalanobis distance between the insulator and the cluster center. The region to which each insulator belongs is determined by the maximum membership principle, and the center coordinates of each region are calculated. The FHI index is used to evaluate the quality of the clustering results. The FHI index comprehensively considers the cluster compactness and cluster separation. The larger the value, the better the clustering effect. The results of different cluster numbers are compared by the FHI index. When the number of clusters is 5, the FHI index reaches a maximum value of 92. Therefore, the insulators to be repaired are divided into 5 maintenance areas. In each maintenance area, the improved ant colony optimization algorithm is used to search for the best maintenance path. The maintenance area is abstracted into an undirected weighted graph with insulators as nodes and the distance between insulators as edge weights. At the same time, the electromagnetic interference intensity between insulators is used as heuristic information. The greater the interference intensity, the greater the edge weight, which attracts ants to search;

Global pheromones and local pheromones were introduced. The initial value of global pheromone concentration was set to 2, the initial value of local pheromone concentration was set to 5, the pheromone volatility coefficient was set to 1, and the local pheromone attenuation coefficient was set to 7. The ant population diversity evaluation index was introduced, and the Simpson index was used to quantify the population diversity. When the Simpson index was lower than 6, 20% of the ant individuals were regenerated through mutation operations. The ant colony algorithm converged to the global optimal solution after 500 iterations, and the optimal maintenance path for each maintenance area was obtained. The optimal maintenance path was converted into JSON format data, including the maintenance area ID, path node longitude and latitude, node distance, and estimated time-consuming attributes. The JSON data was pushed to the /api/inspection/tasks interface of the mobile terminal APP through an HTTP POST request, and different maintenance areas and paths were marked with different colors on the AutoNavi map API. Maintenance personnel queried task details through the voice interaction function, such as "query today's maintenance tasks" and "navigate to the next maintenance point". The mobile terminal APP integrated a location perception module, using GPS/Beidou dual-mode positioning, combined with the Chinese precision CORS system for RTK differential positioning, horizontal positioning accuracy is better than 5cm, vertical positioning accuracy is better than 10cm. At the same time, the extended Kalman filter algorithm is used to integrate GPS positioning data and mobile phone IMU inertial navigation data, and the position update frequency reaches 20Hz. The position perception module collects on-site images through the mobile phone camera, and uses the YOLOv5s target detection algorithm based on MobileNet-SSD to detect and locate insulators in real time. The average detection accuracy mAP reaches 95%, and the detection speed reaches 25FPS. The front image of the insulator is extracted through affine transformation and perspective transformation, and sent to the CNN-based Tesseract0OCR model for nameplate information recognition, with an average recognition accuracy of 95%. Finally, the recognition result is compared with the asset management system data to verify the correctness of the maintenance object. The system matches the trajectory data with the optimal maintenance path, and uses the dynamic time warping (DTW) algorithm to calculate the trajectory deviation. Once the deviation exceeds 10 meters, an early warning prompt is issued to the maintenance personnel in time, prompting the maintenance personnel to adjust the maintenance path until all maintenance tasks are completed. The verification results will also be fed back to the maintenance personnel in real time to avoid maintenance errors.

During the maintenance process, the S104 step uses a photogrammetry module to obtain image data of the transmission line, extract the geometric model of the insulator, and determine whether the insulator has defects, including damage or contamination. If so, the defect type and location coordinates are uploaded to the fault maintenance database, and a local cleaning or replacement process is triggered.

Specifically, based on the transmission line image data obtained by the photogrammetry module, an image segmentation algorithm is used to extract the insulator area. Through the three-dimensional reconstruction technology, a three-dimensional geometric model of the insulator is constructed according to the insulator area image. A deep learning algorithm is used to perform defect judgment on the insulator geometric model to identify whether the insulator is damaged or contaminated. If the insulator is judged to be damaged, an edge detection algorithm is used to determine the position coordinates of the damaged area. If the insulator is judged to be contaminated, a color segmentation algorithm is used to determine the position coordinates of the contaminated area. Based on the defect judgment result, the defect type is obtained, and the defect type and position coordinate information are uploaded to the fault maintenance database. Based on the defect information in the fault maintenance database, it is determined whether a local cleaning or replacement process needs to be triggered. If local cleaning is required, robot automatic cleaning technology is used to perform positioning cleaning based on the position coordinates of the contaminated area. If the insulator needs to be replaced, robot automatic replacement technology is used to perform positioning replacement based on the position coordinates of the damaged area.

In one embodiment, a high-definition camera mounted on a drone is first used to photograph the transmission line to obtain a high-definition image with a resolution of 4000×3000. Then, a region-based watershed segmentation algorithm is used to segment the image and extract the insulator region. The segmentation accuracy can reach the pixel level. The insulator region is reconstructed in three dimensions using multi-view images. A three-dimensional geometric model of the insulator is obtained by constructing a triangular mesh model. The reconstruction accuracy can reach the millimeter level. Then, a deep learning algorithm based on a convolutional neural network is used to determine defects in the insulator geometric model. Through the training sample data set, it is possible to identify whether the insulator is damaged or contaminated. The judgment accuracy can reach more than 95%. If the insulator is judged to be damaged, the Canny edge detection algorithm is used to determine the damaged area. The position coordinates of the domain can be determined with a positioning accuracy of up to the centimeter level. If it is determined that the insulator is contaminated, a color segmentation algorithm based on the HSV color space is used to determine the position coordinates of the contaminated area, with a segmentation accuracy of up to the pixel level. According to the defect judgment result, the defect type and position coordinate information are uploaded to the fault maintenance database, and the historical maintenance data is mined and analyzed using big data analysis technology to determine whether a local cleaning or replacement process needs to be triggered. If local cleaning is required, the robot automatic cleaning technology is used for positioning cleaning based on the position coordinates of the contaminated area, and the cleaning efficiency can be increased by more than 50%. If the insulator needs to be replaced, the robot automatic replacement technology is used for positioning replacement based on the position coordinates of the damaged area, with a replacement accuracy of up to the millimeter level.

The step S105 measures the width and length of the cracks on the surface of the defective insulators, determines the crack propagation path through the ant colony search module, predicts the crack propagation trend and the remaining life of the insulator, and enters the insulator number and location information into the fault maintenance database if the remaining life is lower than the threshold, determines the maintenance time window according to the crack propagation rate, and schedules maintenance resources through the heuristic rule module to minimize the risk caused by crack propagation.

Specifically, a 3D laser scanner is used to scan the defective surface of the insulator, and the point cloud data obtained by the scan is denoised, meshed, and reconstructed to obtain a 3D model of the insulator surface with an accuracy of 0.1 mm. The geometric morphological analysis module is then used to extract crack features from the 3D model to obtain the length, width, and depth geometric parameters of the crack. The geometric parameters of the crack are input into the ant colony search module, which uses the ant colony optimization algorithm (Ant Colony Optimization, ACO) based on graph theory to divide the insulator surface into N×N grids, each grid corresponding to a node, and there are connections between adjacent grids. The connection weight is proportional to the probability of crack propagation. M ants are randomly released, and each ant moves between nodes according to the probability transfer rule based on the pheromone concentration and the heuristic factor. The pheromone is updated after completing a cycle, and the iteration is repeated K times until the global pheromone concentration converges to obtain the path with the highest probability of crack propagation. According to the crack propagation probability distribution map, combined with the material of the insulator, the ant colony optimization algorithm is used to find the crack propagation path. Material properties, load spectrum, environmental temperature and humidity factors, the Paris formula based on fracture mechanics is used to predict the crack expansion trend in the future. The Paris formula describes the relationship between the crack growth rate da/dN and the stress intensity factor range ΔK: da/dN=C(ΔK)^m, where C and m are constants related to the material, obtained by fitting the crack growth test. The stress intensity factor range ΔK is related to the load, crack size, and specimen geometry, obtained by finite element analysis. The Paris formula is discretized and the crack size is recursively calculated: a_(i+1)=a_i+C(ΔK_i)^m×ΔN_i, where ΔN_i is the number of cycles of the i-th cyclic loading, determined according to the set time interval and cycle frequency. The crack growth trend data is input into the remaining life prediction model. The model uses Weibull distribution to describe the fatigue life distribution of the insulator, and its probability density function is: f(t)=(m/η)(t/η)^(m-1)exp, where t is the fatigue life, m is the shape parameter, η is the scale parameter, which is obtained by fitting the test data. Based on the critical crack size a_c predicted by the crack growth model, the α quantile t_α of the Weibull distribution is taken as the remaining life prediction value under the α confidence level, and the threshold of the remaining life of the insulator is set. The predicted remaining life is compared with the threshold. If the remaining life is lower than the threshold, the number and location information of the insulator are automatically entered into the fault maintenance database, and the latest time window for maintenance is determined according to the crack growth model. The insulator information and maintenance time window in the fault maintenance database are input into the heuristic rule module. The module adopts the multi-objective particle swarm optimization algorithm (Multi-objective Particle Swarm Optimization, MOPSO), takes the maintenance task as the particle, and the maintenance efficiency, maintenance cost, and risk level as the optimization objectives. The Pareto optimal solution set is searched in the solution space, and a group of random particles are initialized. Each particle contains the priority, maintenance time, and maintenance of the maintenance task. Resource allocation decision variables, particles update speed and position according to their own historical optimal position and global optimal position, iterate T times until convergence, and select the final maintenance plan according to the decision preference for the final Pareto frontier solution set, generate the priority sorting and resource allocation plan of the maintenance task, and achieve the balance of maintenance efficiency, maintenance cost and risk control objectives under the premise of meeting the maintenance time window constraint. The maintenance task plan is automatically sent to the maintenance personnel's smart terminal through the mobile collaborative office platform to guide the maintenance personnel to carry out maintenance work according to the optimal plan, collect on-site information, and associate the on-site information with the digital insulator ledger and electronic work ticket to generate a structured maintenance report. Through big data analysis technology, text mining and semantic analysis are performed on historical maintenance reports to extract fault modes and maintenance experience strategies for optimizing maintenance decision models and crack extension prediction models. At the same time, blockchain technology is used to trace and audit the maintenance process to ensure the authenticity and traceability of maintenance data and realize closed-loop management of the entire process of fault maintenance.

In one embodiment, the latest time window for maintenance is determined. For example, if the crack growth rate is 0.1 mm/month and the critical crack size is 10 mm, the maintenance time window is (10-current crack length)/0.1 month. The surface of the insulator is scanned using a Creaform Go SCAN50 3D laser scanner. The scanning distance is 30 cm, the scanning angle is 45°, the point cloud density is set to 0.05 mm, and the scanning speed is 4.8 million points/second. The point cloud data is processed using Geomagic Studio software, and the moving least squares method is used for noise reduction. The curvature sampling rate is set to 30%, and the grid density is set to 0.1 mm. A high-precision 3D model is generated, and cracks are extracted from the 3D model using MATLAB's Image Processing Toolbox. The Roberts operator was used for edge detection, and the threshold was set to 0.02. The point cloud subset of the crack area was extracted, and the length, width, and depth geometric parameters of the crack were calculated. The crack parameters were input into the COMSOL Multiphysics finite element analysis software, and the crack extension module was used. The mesh type was tetrahedron, the maximum mesh size was 0.5 mm, the load was set to cyclic tensile load, the stress ratio R=0.1, and the frequency was 10 Hz. The calculated stress intensity factor range ΔK was substituted into the Paris formula to obtain the crack extension rate. The crack extension rate was substituted into the Weibull distribution fitting program written in Python, and the maximum likelihood estimation method was used. After 50 iterations, the shape parameter m=2.3 and the scale parameter η=2.5e6 were obtained. The remaining life of the insulator at the 95% confidence level was calculated to be 1. In 2 years, the insulator number, crack parameters, and remaining life data were uploaded to Alibaba Cloud's HBase distributed database. The MapReduce parallel computing framework was used to clean and aggregate the data to generate a maintenance task priority list. The MOPSO algorithm was implemented using MATLAB's GlobalOptimizationToolbox. The number of particles was set to 50, the maximum number of iterations was 200, the inertia weight was 0.8, the acceleration constant c1=c2=2.0, and the search dimension was 5, including task priority, maintenance time window, required personnel, spare parts, and tool resources. The convergence judgment threshold was set to 1e-6, and the Pareto optimal solution set was obtained. The task scheduling optimization model was built using the Huawei Cloud ModelArts platform, and the deep reinforcement learning algorithm DDPG was selected. The state quantity includes task tightness. Urgency, resource occupancy, action volume including task start time, resource allocation plan, reward function comprehensively considers maintenance timeliness, cost savings, risk aversion factors, and adds soft constraint penalty items, trains 500 episodes, with a learning rate of 0.001 and a ReplayMemory size of 10,000, outputs the optimal scheduling strategy, converts the strategy into a task card in JSON format and sends it to the maintenance personnel's smart terminal, displays it using the mobile collaborative office App developed with the Flutter framework, uses the mobile phone camera to scan the QR code on the insulator, associates the defect information with the task card, uses voice input to fill in the maintenance process record and completion confirmation, and uploads it to the knowledge graph platform of Huawei Cloud, uses natural language processing technology for semantic understanding and extraction, and updates the fault diagnosis rule library and maintenance standard library;

According to the width and length of the crack, the multi-scale image segmentation module is used to extract the crack area on the insulator surface, build a three-dimensional crack model, and simulate the crack expansion process;

A digital microscope is used to perform high-magnification imaging on the surface of the insulator to obtain crack image data with micron-level resolution. The image is denoised by wavelet transform, and the multi-scale texture features of the crack are extracted. The crack edge is then enhanced by Gabor filtering. The image is binarized by setting a dynamic threshold. Finally, the morphological closing operation is used to eliminate noise and obtain the complete crack area. The preprocessed crack image is segmented using a semantic segmentation algorithm based on deep learning, such as U-Net and DeepLab. The convolutional neural network is trained on a large number of labeled samples to automatically learn the multi-scale features of the crack, and accurate segmentation of cracks of arbitrary width and length is achieved to obtain the two-dimensional contour of the crack. The extracted crack contour data is input into the three-dimensional reconstruction module. The structured light three-dimensional scanning technology is used to obtain the three-dimensional morphology information of the surface by projecting coded stripes. The stripe phase is decoded using the phase shift unwrapping algorithm. The surface three-dimensional point cloud is reconstructed through the phase unwrapping and triangulation principles, and the point cloud is meshed and texture mapped. A three-dimensional crack model of the insulator with millimeter-level accuracy was constructed, and the crack area was accurately mapped to the three-dimensional crack model to obtain the true crack morphology. According to the three-dimensional crack model, the key geometric parameters of the crack were extracted, and the crack depth was calculated by measuring the maximum height difference of the crack area on the three-dimensional model; the radius of curvature of the crack tip was obtained by fitting the spline curve of the crack tip and taking the radius of curvature of the curve at the tip. According to the material type and load condition of the insulator, the fracture criterion and stress intensity factor calculation formula were selected. Linear elastic fracture mechanics and KI stress intensity factor were used for ceramic insulators, and elastic-plastic fracture mechanics and J integral were used for composite insulators. The stress and strain distribution at the crack tip was analyzed to determine the crack initiation and propagation conditions. The geometric parameters and material properties of the crack were input into the ABAQUS finite element analysis software, and the extended finite element method was used to numerically simulate the crack propagation process. When modeling, the insulator was first divided into a hexahedral structured grid as a whole, and the grid density at the crack tip and its extension direction was It should be higher than other areas. A crack extension unit is introduced at the crack tip, a crack extension criterion is set, the crack extension direction and incremental step size are defined, the coupling effect of electrical stress, mechanical stress and thermal stress on the insulator during operation is considered, the boundary conditions and loading methods of multi-physical fields are set, and the Newton-Raphson iterative algorithm is used for nonlinear solution. When the stress intensity factor or J integral meets the fracture criterion, the crack extension simulation is started until the crack extends to the specified length or the insulator fails. The long short-term memory neural network time series prediction model is used to perform machine learning on the crack extension data obtained by finite element simulation. The crack extension length, extension direction and extension rate are used as sample features, and the corresponding load condition parameters are used as sample labels to construct a training data set. By setting the number of neurons and activation functions in the input layer, hidden layer and output layer, the depth and width of the network are optimized. Dropout regularization and early stopping mechanism are introduced to prevent overfitting. The random gradient descent method is used to train the network parameters, and the optimization is continuously iterated. , until the prediction accuracy or convergence speed requirements are met, the trained prediction model is applied, the load parameters under actual operating conditions are input, and the intelligent prediction of the crack growth trend and remaining life of the insulator is realized. The crack growth prediction results are combined with the electrical and mechanical performance degradation models of the insulator, and the coupling relationship between crack growth and breakdown probability and strength reduction is established. The Weibull distribution is used to describe the relationship between the breakdown probability of ceramic insulators and the electric field strength and temperature factors. The Paris formula is used to describe the relationship between the strength reduction and crack growth rate of composite insulators. A physical mechanism model for insulator status assessment and life prediction is established, the key factors affecting the reliability of insulators are identified, and a risk-based status maintenance strategy is formulated. Big data on insulator crack growth and performance degradation is accumulated, and machine learning algorithms are used for knowledge mining to refine the laws of crack occurrence and development, form a knowledge base for intelligent crack diagnosis and early warning, guide the hierarchical control of insulator defects, and provide a decision-making basis for the status maintenance and replacement of insulators.

In one embodiment, a Keyence VHX-7000 series digital microscope is used, the optical zoom is set to 200 times, the resolution is set to 0.1 μm, and a 10 cm×10 cm area on the surface of the insulator is scanned and imaged to generate a high-definition image containing 20 million pixels. The image is decomposed into 5 levels based on the Daubechies 8th order basis function of wavelet transform, and soft threshold denoising is applied to the high-frequency sub-band. The threshold is determined by 3 times the noise variance estimate. A Gabor filter with a scale of 5×5 and directions of 0°, 45°, 90°, and 135° is used to extract crack texture. The differential Gaussian algorithm is used to detect the crack edge. The threshold is 1.5 times the mean of the edge gradient. The U-Net convolutional neural network adopts a 4-layer downsampling and 4-layer upsampling structure, the convolution kernel size is 3×3, and the number of feature map channels is 64, 128, 256, and 512, respectively. The Adam optimizer was trained for 100 epochs with the ReLU activation function and BatchNorm normalization. The initial learning rate was 0.001 and decayed by 10% every 10 epochs. The crack contour was segmented at the pixel level with an mIoU of 0.98. The structured light 3D scanning system used a DLP projector to generate sinusoidal grating fringes with a period of 1920μm and a phase shift step of 4. The industrial camera had a resolution of 1280×1024 and a field of view of 50mm×40mm. The calibration reprojection error was less than 0.05mm. The point cloud was solved using phase profilometry with a density of 100 points per square millimeter. The Poisson reconstruction obtained a 0.5mm thick insulator surface mesh model. The model accuracy was evaluated by coverage, consistency and integrity with an error of less than 0.2mm. The 3D crack contour was matched with the insulator model to extract the crack depth, surface area and volume geometric parameters. The third-order B-spline curve is fitted by the mesh nodes at the crack tip, and the curvature radius is obtained by differentiation. The ABAQUS finite element model is based on the eight-node hexahedral unit C3D8R, the crack tip mesh size is 0.1mm, the energy integral convergence control parameter is 0.05, the material parameters are isotropic elastic, the elastic modulus is 110GPa, the Poisson's ratio is 0.28, the Paris formula C=1.8×10^-10, m=3.2, the crack extension adopts the maximum energy release rate criterion, the extension direction is perpendicular to the maximum tensile stress plane, the extension step is 0.05mm, the electric-mechanical-thermal multi-physics field coupling applies 150kV voltage, 500 Newton tension and 80℃ steady-state temperature, the iteration error is less than 10^-3, and the calculation is performed for 12 hours until the insulator fails. The LSTM network input layer has 56 neurons, corresponding to 7 features of crack extension data × 8 time steps, and the hidden layer has 12 8 neurons, 2 stacking layers, 1 neuron in the output layer, predict the crack extension length of the next time step, learning rate 0.01, verification every 10 epochs, earlystop patience 5 times, prediction error converges to within 3%, Weibull distribution shape parameter m=7.85, scale parameter η=15.2, 63.2% breakdown voltage is 140kV, predict the remaining life when more than 1000 hot and cold cycles, the crack size should be controlled within 4mm, Paris formula combined with crack extension prediction model, infer that the critical crack depth of insulator strength degradation to 70% is 8mm, association rules mining of insulator crack parameters and defect modes, support 0.4, confidence 0.85, lift 2.5, found that glaze cracks and pores are the main causes of flashover, it is recommended to optimize the glaze formula and firing process in a targeted manner;

Combined with the simulation data of crack propagation, a multi-objective optimization model for crack repair is established to determine the best repair strategy and resource allocation plan;

According to the simulation data of cracks in the simulation expansion process, a mathematical model of crack repair is established by using a multi-objective optimization method. By defining the repair strategy and resource allocation scheme as optimization variables, the crack growth rate and repair cost are obtained as optimization targets, and the objective function and constraints of the multi-objective optimization model are obtained. The non-dominated sorting genetic algorithm NSGA-II is used to solve the above multi-objective optimization model. By randomly generating an initial population of repair strategies and resource allocation schemes, the target values of crack growth rate and repair cost corresponding to each individual are obtained. According to the non-dominated sorting and crowding calculation, the fitness and sorting results of the population are obtained. According to the fitness and sorting results of the population, a binary tournament selection operator is used to select excellent individuals from the current population. By simulating binary crossover and polynomial mutation operators, crossover and mutation operations are performed on the selected individuals to obtain new repair strategies and resource allocation plans. The newly generated individuals are merged with the current population, and non-dominated sorting and crowding calculation are used to obtain a new generation of population. By iterating the above selection, crossover mutation and population update process until the maximum number of iterations or convergence conditions are reached, the Pareto frontier of the optimal repair strategy and resource allocation plan is obtained. According to the decision maker's preference, a satisfactory solution is selected from the Pareto optimal solution set as the best crack repair strategy and resource allocation plan. By analyzing the crack growth rate and repair cost of the plan, its effect in extending the service life of the component and saving maintenance costs is judged, and the optimal decision for crack repair is determined.

In one embodiment, according to the simulation data of crack propagation, the NSGA-II algorithm is used to perform multi-objective optimization on the crack repair strategy and resource allocation scheme. First, 100 initial populations are randomly generated, each individual contains 10 repair strategy variables and 5 resource allocation variables, the strategy variable value range is 0 to 1, and the resource variable value range is 0 to 1 million yuan. Then, each individual is substituted into the crack propagation model to obtain the crack propagation rate and repair cost. Next, the population is non-dominated sorted, and the crowding degree is calculated between individuals to obtain the fitness and sorting. After that, the binary tournament selection operator is used with a crossover probability of 8 and a mutation probability of 1. The individuals are cross-mutated to obtain a new solution, which is merged with the current population, and the population is updated. After 500 iterations, the Pareto frontier is obtained. Finally, according to the decision maker's preference, an optimal solution that balances the propagation rate and cost is selected: the crack propagation rate is reduced by 20%, and the repair cost is 800,000 yuan, which can effectively extend the life of the component by more than 10 years, save about 30% of the maintenance cost, and realize the scientific optimization of crack repair decision.

After the insulator is replaced or repaired, the step S106 detects the temperature distribution on the surface of the insulator through the thermal image data of the insulator surface, extracts the frequency domain features of the thermal image data, and uses a convolutional autoencoder network to judge the current quality of the insulator. If there is an abnormal temperature distribution, the thermal image data is uploaded to the fault maintenance database, and the rework process is triggered to dynamically adjust the maintenance strategy.

Specifically, an infrared thermal imager is used to collect thermal images of the insulator surface after the insulator is replaced or repaired. The shooting distance and angle are selected, and the large insulator is photographed in different areas to obtain thermal image data at different viewing angles and distances. Then, an image stitching algorithm based on SIFT feature point matching is used to generate a full-range, high-resolution surface temperature distribution map of the insulator. The two-dimensional Fourier transform 2D-FFT and wavelet transform mathematical tools are used to perform time-frequency domain analysis on the thermal image data. The spectral energy distribution characteristics of the thermal image are extracted in the frequency domain, such as the low-frequency, medium-frequency, and high-frequency energy ratios, as well as the spectral entropy and spectral uniformity texture characteristics. The multi-scale and multi-directional features of the thermal image are extracted in the wavelet domain, such as wavelet energy and wavelet coefficient mean. The extracted frequency domain and wavelet domain features are constructed into a feature vector of the insulator surface temperature distribution, and the feature vector of the thermal image is input into a U-Net-based thermal image processing system. The quality judgment is carried out in the convolutional autoencoder network of the structure. The encoder part consists of 4 convolutional layers and 4 maximum pooling layers. The convolution kernel size is 3×3 and the activation function is ReLU. The decoder part consists of 4 deconvolution layers and 4 upsampling layers. The detailed information of the original thermal image is gradually restored. The network training adopts Adam optimizer, the initial learning rate is 0.001, the loss function is mean square error (MSE), and the evaluation index is peak signal-to-noise ratio (PSNR). The quality of the insulator is judged by the size of the reconstruction error. The higher the PSNR value, the better the reconstruction quality and the higher the insulator quality. A standard template library for insulator temperature distribution is constructed, and thermal image samples of insulators of different models, materials, and defect modes are collected. The K-means clustering algorithm is used to group the samples. The HOG and LBP texture features of the thermal image samples are first extracted, and then the PCA algorithm is used to classify the samples. The features are compressed to 20 dimensions, and then K-means clustering is performed. The Silhouette coefficient determines that the optimal number of clusters is 5. The central thermal image of each cluster is calculated as the standard template, and its frequency domain and wavelet domain features are extracted. When judging whether the newly collected thermal image is abnormal, the Euclidean distance and cosine similarity between it and each standard template are calculated respectively. If the minimum distance is greater than the threshold (such as 10) or the maximum similarity is less than the threshold 0.6, it is judged as abnormal temperature distribution. When the insulator is monitored to have abnormal temperature distribution, the edge upload process of thermal image data is triggered, and the collected thermal image data is cropped and normalized for preprocessing, and then compressed by H.265 encoding and encrypted by AES-256 algorithm. The compressed and encrypted data is packaged into JSON format and sent to the fault maintenance database in the cloud in real time through the MQTT protocol, and abnormal alarm information is generated. According to the abnormal temperature distribution of the insulator, the maintenance strategy and rework process are dynamically adjusted. The method combining Q-learning and simulated annealing algorithm is adopted. The abnormal degree and frequency of insulator temperature are used as the state space, the detection frequency, replacement of equipment and process parameters are used as the action space, and the reward function is used to improve the qualified rate and reduce the cost. The optimal maintenance strategy is learned. At the same time, a multi-agent reinforcement learning framework is adopted to realize the collaborative maintenance optimization of multiple insulators. The operation and maintenance level of the entire transmission line is improved through the game between intelligent agents. A quality management platform covering the entire life cycle of insulators is established, including raw material quality inspection, production process monitoring, operation status monitoring, defect intelligent diagnosis, status assessment and prediction, maintenance strategy optimization, and scrap disposal traceability functional modules. Blockchain technology is used to record and store quality data of each link in an unalterable manner, realizing a closed loop of insulator quality management.

In one embodiment, the temperature measurement range of the thermal imager is -40℃~2000℃, the thermal sensitivity is 0.02℃, the spatial resolution is 1024×768, and the frame rate is 30Hz. The FLIRT1020 infrared thermal imager is used to collect thermal images at a distance of 2 meters from the insulator and an angle of 30°. The resolution is set to 1024×768, the frame rate is 30Hz, the temperature measurement range is -20℃150℃, and the thermal sensitivity is 0.02℃@30℃. For large insulators with a diameter of more than 50 cm, they are divided into 4 quadrants, and thermal images are collected separately, with each quadrant overlapping by 20%. Then, the SIFT algorithm is used to extract feature points, and the homography matrix is estimated by the RANSAC algorithm. Finally, they are spliced into a complete thermal image with a splicing accuracy better than 0.5 pixels. The thermal image is analyzed in the frequency domain using the SignalProcessing Toolbox of MATLAB, and 2D-FFT is used The amplitude spectrum and phase spectrum are obtained by transformation, and the energy ratio features of low frequency 00.1π, medium frequency 0.1π, and high frequency 0.5π are extracted. The gray-level co-occurrence matrix is calculated to obtain 14 texture features including contrast, correlation, energy, and entropy. The db4 wavelet basis function is used to perform 5-level wavelet decomposition on the thermal image, and 9 wavelet domain features including the mean, variance, and energy of the HL, LH, and HH subband coefficients are extracted. The U-Net convolutional autoencoder network is trained, the encoder is downsampled 4 times, the decoder is upsampled 4 times, the convolution kernel size is 3×3, the step size is 1, the padding method is the same, the activation function is ReLU, the maximum pooling size is 2×2, and the Adam optimizer is used. The initial learning rate is 0.001, which is decayed by 50% every 10 epochs, the minimum learning rate is 1e-6, the batch_size is 16, and 100 epochs are trained. The early stopping method is used to prevent overfitting. The best model has a P value of 0. The SNR reaches 35.2dB. The K-means clustering algorithm is used to cluster the HOG and LBP feature vectors of 500 normal insulators and 200 abnormal insulators. The optimal cluster number k=5 is determined by the elbow rule. The average PSNR of each cluster center and the sample is calculated, and the threshold is 34dB. When the PSNR of the reconstructed image of the test insulator is lower than the threshold, it is judged as temperature abnormality and uploaded to the Huawei cloud server. It is first compressed by 50% using H.265 encoding, and then encrypted using AES-256. It is transmitted through the MQTT protocol, and the network bandwidth is reduced by 75%. The Q-learning algorithm is used for strategy learning. The state space is the insulator health (0~100), and the action space is the detection frequency 1 day/time, 7 days/time, and 30 days/time. The reward function is 100 points for a 10% reduction in the fault rate, and 500 points for a 3% reduction. The learning rate α=0.1, the discount factor γ=0.9, exploration probability ε=0.7, maximum number of rounds 1000, average reward after convergence is 85 points. On this basis, simulated annealing strategy is added, initial temperature T0=100, decay rate λ=0.95, inner loop L=50, outer loop D=20, search for the global optimal solution, average reward increased to 90 points, for three 100 km long 500 kV transmission lines, 30 fault monitoring agents are deployed, multi-agent-Q-learning framework is adopted, through 20,000 rounds of training, the fault rate is reduced by 15%, and the operation and maintenance cost is reduced by 12%. In this process, the Hyperledger Fabric blockchain platform is used to record 30 million insulator quality data, the block size is 2MB, the consensus mechanism is PBFT, the throughput is 1500TPS, the shared ledger storage capacity is 6TB, and the data traceability and privacy protection level meet the HIPAA standard.

The step S107 obtains the comprehensive health index of the insulator based on the multi-dimensional data of the insulator, including coating aging degree, glaze layer distortion, surface cracks, and temperature anomaly, and uses a Markov chain prediction model to estimate the state transition probability under different maintenance strategies, and uses a multi-objective particle swarm optimization module to generate the optimal maintenance decision sequence for the entire life cycle of the insulator.

Specifically, the multi-dimensional data of insulator coating aging degree, glaze distortion, surface crack parameters, and temperature anomaly monitoring are extracted from the multi-source heterogeneous systems of fault diagnosis system, condition monitoring system, and environmental monitoring system. The ontology mapping method is used to construct the insulator health assessment domain ontology, define the semantic association of each data source, and realize automatic association and conversion of data semantic level through ontology reasoning. For data quality issues, noise identification, missing value filling, and outlier detection data cleaning technology are used. Combined with electrical field knowledge, data correction rules are constructed to form a structured and standardized health record database for the entire life cycle of insulators. According to the insulator health status assessment standard, key indicators reflecting the health level of insulators are extracted from multi-dimensional monitoring data. The subjective and objective combined weights of each indicator are determined by combining the Delphi method and the entropy weight method. The comprehensive health index of the insulator is calculated by weighted average. Based on the health index, the support vector machine algorithm is used to construct a multi-classification model of the insulator health status. Multi-classification is achieved by finding the maximum margin hyperplane, and the kernel function is introduced to solve the nonlinear problem. The precision, recall and F1 value evaluation indicators are used, and the optimal model is selected through cross-validation. The hidden Markov model is used to construct the Markov chain of the evolution of the insulator health state, and the comprehensive health index of the insulator is discretized into normal, sub-healthy, slightly degraded, moderately degraded, and severely degraded states. The state transition probability matrix is learned from the monitoring data through the Baum-Welch algorithm, and the observation probability matrix is estimated through the correspondence between the insulator health state and multidimensional data. A reduction algorithm is introduced to reduce the dimension of the state space. Thus, the dynamic evolution trend and steady-state distribution of the insulator health state under different maintenance strategies are predicted, and a multi-objective optimization model for insulator maintenance decision-making is constructed. The objective function includes three aspects: improving the health level of insulators, minimizing the maintenance cost, and maximizing the remaining life of insulators. The constraints include the transfer of insulator health state, the upper limit of maintenance resources, and the attenuation of the remaining life of insulators. The decision variable is the insulation. The objective function and constraint conditions are quantitatively expressed through mathematical formulas. The improved multi-objective particle swarm optimization algorithm (MOPSO) is used to solve the maintenance decision optimization model. The Pareto optimal solution set archive and congestion measurement are introduced to ensure the convergence and distribution diversity of the solution. The position and velocity of the particles are updated according to the standard formula through inertia weight, acceleration constant, particle historical optimal position and global optimal position update, and the key parameters of the algorithm are optimized through sensitivity analysis and experimental design. Thus, the optimal maintenance decision sequence of the insulator's entire life cycle is generated, and the optimal maintenance decision sequence is decoded into the specific maintenance time, maintenance method, maintenance parameters and maintenance measures of each stage of the insulator's entire life cycle, forming a dynamic maintenance schedule covering the entire life cycle. When dynamically adjusting the maintenance plan, a rolling optimization strategy is adopted. According to the insulator state monitoring data and changes in environmental factors, the state transition probability, maintenance cost, and remaining life parameters in the multi-objective optimization model are periodically updated. Generate the optimal maintenance plan within the rolling period, and dynamically adjust the weight coefficient of the optimization target according to the execution feedback, realize the adaptive optimization of maintenance decisions, realize real-time monitoring and early warning of insulator status, develop an insulator asset management system based on big data analysis and artificial intelligence, integrate the full life cycle data of insulator design, manufacturing, operation, maintenance, and scrapping, and use machine learning and deep learning algorithms such as random forest, XGBoost, and LSTM to establish a prediction model and abnormal diagnosis model for the insulator degradation mechanism. The knowledge base adopts a hybrid reasoning method based on ontology and rules, integrates equipment nameplates, defect cases, and diagnostic test structured and unstructured knowledge to form a knowledge graph for the full life cycle management of insulators. The knowledge graph is used to realize intelligent diagnosis, prediction, and decision-making, and continuously optimize the insulator health status assessment model and maintenance decision optimization model. It provides intelligent analysis and auxiliary decision-making functions for key decisions on insulator selection, transformation, maintenance, and replacement, and improves the management level of power equipment and asset utilization efficiency.

In one embodiment, the aging degree data of the insulator coating is extracted from the fault diagnosis system. For example, the average temperature of the coating area in the infrared image is 68°C, which is 13°C higher than the normal value, and the aging degree is serious. The glaze layer distortion is judged by X-ray imaging, and the image is subjected to Canny edge detection. The distortion area ratio is calculated to be 2%, which exceeds the warning value of 5%. The surface crack parameters are obtained by laser three-dimensional scanning, and the cracks are segmented by regional growth algorithm. The length, width and depth of the cracks are extracted to be 5mm, 8mm and 2mm respectively. The environmental monitoring system collects the temperature, humidity, wind speed and pollution degree data around the insulator. The temperature data is judged to be abnormal through the anomaly detection model, which deviates from the normal value by 5 standard deviations. The above multi-source heterogeneous data are mapped through the ontology to construct the insulator health assessment domain ontology, including coating aging, glaze distortion, crack risk and environmental anomaly categories, and the semantic associations between the concepts are defined, such as the functional relationship between coating aging and infrared temperature. Noise identification is performed on the monitoring data. For example, if the high-frequency component of the temperature data exceeds the threshold, it is judged as noise. The missing values are filled by the nearest neighbor interpolation method, and the outliers are detected by the isolation forest algorithm. Combined with electrical field knowledge, data correction rules are constructed. For example, when the wind speed exceeds 30m/s, the pollution degree data is multiplied by a correction coefficient of 5. Finally, a structured health record database for the entire life cycle of insulators is formed. Key indicators are extracted from the health record database. The Delphi method is used to consult 10 experts to obtain the subjective weight vector [3, 25, 2, 25]. The objective weight vector [28, 32, 15, 25] is calculated by the entropy weight method. The combined weight vector is [29, 285, 175, 25]. The weighted average obtains the comprehensive health index of the insulator as 75 points. Based on the health index, a multi-classification model of health status is constructed using support vector machine. The state space is "normal, sub-healthy, mildly deteriorated, moderately deteriorated, and severely deteriorated". The sample size is 1000 and the feature dimension is 10. The radial basis kernel function is used, the penalty coefficient C=5, the kernel function parameter g=08, and through 5-fold cross validation, the model accuracy is 92%, the recall rate is 95%, and the F1 value is 95%. The comprehensive health index of the insulator is discretized into 5 states, and a hidden Markov model is constructed. The initial state probability vector is [7, 2, 06, 03, 01]. The Baum-Welch algorithm is used to iteratively optimize the state transition probability matrix and the observation probability matrix, and the dimension of the observation data is reduced from 10 to 6 through the reduction algorithm. The evolution trend of the health status of insulators in the next 10 years is predicted. After 5 years, the health status is stable at a sub-healthy level. A multi-objective optimization model for insulator maintenance decision-making is constructed. The MOPSO algorithm is used to solve it, 200 particles are initialized, the inertia weight w=7, the acceleration constant c1=c2=5, the maximum number of iterations is 500, and the archive capacity is 50. Through the Pareto optimal solution set and congestion measurement, the maintenance cost of 200,000 yuan and the health level improvement are obtained. The optimal maintenance decision sequence for increasing the service life by 10% and extending the remaining service life by 5 years includes preventive maintenance in the third year and replacing the aged coating, condition maintenance in the fifth year and repairing the glaze defects, periodic maintenance in the eighth year and tightening the hardware. An insulator asset management system is developed, and a degradation mechanism prediction model is established using the random forest algorithm. The feature importance is ranked as coating aging degree 35, glaze distortion degree 28, crack risk degree 22, and environmental abnormality degree 15. The XGBoost algorithm is used to establish an abnormal diagnosis model, and the fault type is multi-classified with an accuracy of 95%. The knowledge graph contains 500 concepts and 1,500 associations. Based on ontology and rule reasoning, intelligent diagnosis and prediction are realized. The knowledge system contains 300 diagnostic rules and 150 successful cases. Through case reasoning and rule reasoning, intelligent analysis and auxiliary decision-making are provided for key decisions such as insulator selection, transformation, maintenance, and replacement. The average time for each decision is shortened from 2 hours to 10 minutes.

So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present invention shall be included in the protection scope of the present invention.

Claims (7)

1. A method for generating a distribution network maintenance strategy based on a heuristic algorithm, characterized in that: S101. Analyze the aging degree of the coating on the surface of the insulator in the distribution network, including using Fourier transform infrared spectroscopy analysis technology to obtain chemical bond vibration characteristic data of the coating material, construct an aging degree assessment model, determine whether the aging degree of the insulator is greater than a threshold, and if so, identify that the coating needs to be replaced, enter the number and location information of the insulator into the fault maintenance database, and determine the maintenance priority according to the aging degree of the coating; S102, using an X-ray diffractometer to analyze the physical composition of the glaze layer of the insulator, obtaining diffraction spectrum data about the crystal structure of the glaze layer, extracting features and classifying the diffraction spectrum data through a convolutional neural network model, and determining whether the glaze layer is distorted. If it is determined that the glaze layer is distorted, the insulator number and location information are entered into a fault maintenance database, and a glaze repair or replacement plan is determined according to the type of distortion; S103, for each insulator to be repaired in the fault maintenance database, by analyzing the geographical location of the tower and the electromagnetic interference level, the insulator to be repaired is divided into multiple maintenance areas through the fuzzy C-means clustering module, and the optimal maintenance path for each maintenance area is determined, the optimal maintenance path is sent to the terminal of the maintenance personnel, and the maintenance personnel are located and tracked in real time through the position sensing module; S104. During the maintenance process, the photogrammetry module is used to obtain image data of the transmission line, extract the geometric model of the insulator, and determine whether the insulator has defects, including damage or contamination. If so, the defect type and location coordinates are uploaded to the fault maintenance database, and a local cleaning or replacement process is triggered; S105. For defective insulators, measure the width and length of cracks on the surface of the insulators, determine the crack expansion path through the ant colony search module, predict the crack expansion trend and the remaining life of the insulators, and enter the insulator number and location information into the fault maintenance database if the remaining life is lower than the threshold. Determine the maintenance time window according to the crack expansion rate, and schedule maintenance resources through the heuristic rule module; S106. After the insulator is replaced or repaired, the temperature distribution on the surface of the insulator is detected through the thermal image data on the surface of the insulator, the frequency domain features of the thermal image data are extracted, and the current quality of the insulator is judged using a convolutional autoencoder network. If there is an abnormal temperature distribution, the thermal image data is uploaded to the fault maintenance database, and the repair process is triggered to dynamically adjust the maintenance strategy. S107. Based on the multi-dimensional data of the insulator, including coating aging degree, glaze distortion, surface cracks, and temperature anomaly, the health index of the insulator is obtained, and the state transition probability under different maintenance strategies is estimated using the Markov chain prediction model. The multi-objective particle swarm optimization module is used to generate the optimal maintenance decision sequence for the entire life cycle of the insulator. 2. A method for generating a distribution network maintenance strategy based on a heuristic algorithm according to claim 1, characterized in that: the S101 step analyzes the aging degree of the coating on the surface of the insulator in the distribution network by using Fourier transform infrared spectroscopy analysis technology, performs chemical composition analysis on the coating material on the surface of the insulator, obtains chemical bond vibration characteristic data of the coating material, removes a sample from the coating on the surface of the insulator, grinds it into powder with a mortar, presses the powder sample into a tablet, and puts it into an FTIR instrument for testing, uses FTIR spectrum analysis software to perform baseline correction and smoothing preprocessing on the test data, and then compares it with the data in the standard spectrum library to calculate the similarity, which is measured by Euclidean distance and correlation coefficient indicators. According to the size of the similarity, the aging degree of the coating material is judged, the chemical bond vibration characteristics obtained by FTIR analysis are spliced with the surface image characteristics of the insulator coating to form a sample feature input, and the aging degree category is used as output, and training is performed using a cross entropy loss function and an Adam optimizer. 3. According to the heuristic algorithm generation method of distribution network maintenance strategy in claim 1, it is characterized in that: the S102 uses an X-ray diffractometer to perform phase composition analysis on the glaze layer of the insulator to obtain diffraction spectrum data of the crystal structure of the glaze layer, and the setting parameters of the X-ray diffractometer include X-ray wavelength, incident angle, scanning step and scanning speed. Through data preprocessing and feature engineering, key characteristic parameters of the diffraction spectrum are extracted, such as 0 diffraction peak position corresponding to crystal plane spacing, peak intensity corresponding to crystal plane orientation, peak width corresponding to grain size and background intensity corresponding to amorphous phase content. 4. The method for generating a distribution network maintenance strategy based on a heuristic algorithm according to claim 1 is characterized in that: in the step S103, for each insulator to be repaired in the fault repair database, the information of the insulator to be repaired is extracted from the fault repair database, the position of the tower is spatially visualized and analyzed using the ArcGIS system, and the distance matrix between the towers is calculated using the Geodesic distance calculation function. 5. According to the method for generating a distribution network maintenance strategy based on a heuristic algorithm of claim 1, it is characterized in that: the step S104 uses a three-dimensional laser scanner to scan the surface of the defective insulator, denoises, grids, and reconstructs the point cloud data obtained by the scan to obtain a three-dimensional model of the insulator surface with an accuracy of 0.1 mm, and then uses a geometric morphological analysis module to extract crack features from the three-dimensional model to obtain geometric parameters of the length, width, and depth of the crack, and inputs the geometric parameters of the crack into an ant colony search module, which uses an ant colony optimization algorithm based on graph theory to divide the insulator surface into N×N grids, each grid corresponds to a node, there are connections between adjacent grids, the connection weight is proportional to the probability of crack extension, and m ants are randomly released. 6. A method for generating a distribution network maintenance strategy based on a heuristic algorithm according to claim 1, characterized in that: the step S105 adopts the Paris formula based on fracture mechanics to predict the crack growth trend, the Paris formula describes the relationship between the crack growth rate da/dN and the stress intensity factor range ΔK: da/dN=C(ΔK)^m, wherein C and m are material constants obtained by fitting the crack growth test, the stress intensity factor range ΔK is associated with the load, crack size, and sample geometry, obtained by finite element analysis, and P The aris formula is discretized and the crack size is calculated recursively: a_(i+1)=a_i+C(ΔK_i)^m×ΔN_i, where ΔN_i is the number of cycles of the i-th cyclic loading, which is determined according to the set time interval and cycle frequency. The crack extension trend data is input into the remaining life prediction model. The model uses Weibull distribution to describe the fatigue life distribution of the insulator, and its probability density function is: f(t)=(m/η)(t/η)^(m-1)exp, where t is the fatigue life, m is the shape parameter, and η is the scale parameter. 7. According to the heuristic algorithm for generating a distribution network maintenance strategy according to claim 1, it is characterized in that: in the step S106, for large insulators with a diameter exceeding 50 cm, it is divided into 4 quadrants, and thermal images are collected separately, with each quadrant overlapping by 20%, and then the SIFT algorithm is used to extract feature points, the homography matrix is estimated by the RANSAC algorithm, and the thermal images are spliced together, and the thermal images are analyzed in the frequency domain by using MATLAB's Signal Processing Toolbox, and the amplitude spectrum and phase spectrum are obtained by 2D-FFT transformation.