A collaborative control method for ring main unit, branch box and substation based on AR

Abstract

The present application provides a collaborative management and control method for ring network cabinets, branch boxes, and substations based on AR, including: determining maintenance tasks based on the output of a fault identification model, and allocating maintenance tasks based on the skills mastered by the current on-duty personnel, and obtaining equipment data of corresponding equipment through an on-site AR system when the current on-duty personnel perform maintenance tasks; obtaining the idle status of the current expert, and combining with the personnel management system, sending the idle status of the current expert to the terminal of the current on-duty personnel or displaying it on the current on-duty public screen; when multiple maintenance tasks are performed simultaneously, adjusting the priority of information transmission between the on-site AR system and the expert AR system according to the processing priority of the maintenance tasks.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an AR-based collaborative management and control method for a ring main unit, a branch box, and a substation. Background Art

[0002] In the collaborative management and control of ring main units, branch boxes and substations, the use of AR technology has brought significant improvements to the maintenance and management of industrial facilities, but it has also introduced technical challenges, especially the contradiction between real-time performance and data update frequency. In order to ensure timely response to equipment failures, the system needs to continuously collect and analyze equipment operation data. This includes intelligently assigning maintenance tasks to appropriate maintenance personnel based on the type and severity of equipment failures, and updating data during the maintenance process in real time so that experts in the corresponding fields can be summoned for remote assistance at any time. When experts provide remote assistance, frequent data transmission may burden system performance and stability. After all, the communication resources for synchronizing data in the AR system are limited, and it is difficult to meet the real-time data transmission. In the process of remote collaborative maintenance, there are many technical challenges in the real-time synchronization of maintenance data between the on-site AR system and the expert's personal AR system. The network environment of the two AR systems may be very different. The on-site environment may be a mobile network or even no network coverage, while the expert's location may be a stable high-speed network. The difference in network environment leads to uncontrollable data synchronization delay, which affects the maintenance synchronization rate. The types of data generated during the maintenance process are diverse, including small data such as text and images, as well as large data such as videos and three-dimensional models. The requirements for transmission bandwidth and real-time performance of different types of data vary greatly, further increasing the difficulty of data synchronization. In addition, the rendering capabilities of the AR system are limited. How to dynamically adjust the data synchronization frequency and data accuracy to maximize the synchronization rate while ensuring the rendering quality is also a major challenge. Therefore, in the application of AR systems for the coordinated management and control of ring network cabinets, branch boxes, and substations, factors such as the current status of the equipment, the type of fault, and the information required for maintenance operations need to be considered. Summary of the invention

[0003] The present invention provides a collaborative management and control method for a ring main unit, a branch box, and a substation based on AR, which mainly includes:

[0004] Obtain common faults and corresponding abnormal power data of ring main units, branch boxes and substations, extract fault features from the abnormal power data and associate them with corresponding fault types, establish a fault identification model for identifying corresponding fault types based on the power data of power facilities and calculating the corresponding fault occurrence probability;

[0005] Determine maintenance tasks based on the output of the fault identification model, and assign maintenance tasks based on the skills of the current on-duty personnel. When the current on-duty personnel perform maintenance tasks, they obtain the equipment data of the corresponding equipment through the on-site AR system;

[0006] Obtain the idle status of the current expert, and combine with the personnel management system to send the idle status of the current expert to the terminal of the current on-duty personnel or display it on the current on-duty public screen;

[0007] Classify the maintenance tasks of ring main units, branch boxes and substations, including analyzing the urgency of maintenance tasks based on whether the maintenance tasks are assigned to the task list of the current on-duty personnel. The urgency of maintenance tasks includes high urgency and low urgency. The maintenance tasks are divided into corresponding processing priorities according to the importance and impact scope of the corresponding faulty equipment.

[0008] For low-urgency maintenance tasks, if an expert request from the current on-duty personnel is detected, the device data of the corresponding device is synchronized to the expert AR system of the corresponding expert, and maintenance suggestions from the expert AR system are received in real time and synchronized to the on-site AR system. The maintenance suggestions are associated with the corresponding device in the on-site AR system and displayed in the display area of ​​the corresponding device;

[0009] For high-urgency maintenance tasks, the current equipment data with timestamps is sent to the expert AR system of the corresponding expert, and the maintenance plan with timestamps is received from the expert AR system. When the high-urgency maintenance task is assigned to the current on-duty personnel, the maintenance plan is mapped to the on-site AR system;

[0010] When multiple maintenance tasks are performed simultaneously, the priority of information transmission between the on-site AR system and the expert AR system is adjusted according to the processing priority of the maintenance tasks;

[0011] Based on the equipment data of ring main units, branch boxes and substations, a fault prediction model is trained to identify changes before and after a fault. Based on the trained fault prediction model, the equipment data of ring main units, branch boxes and substations are monitored in real time to predict potential faults of the corresponding equipment.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] The present invention discloses a collaborative management and control method for ring main units, branch boxes, and substations based on AR. When multiple maintenance tasks occur at the same time, it can ensure that tasks with higher priority can be processed in a timely manner, thereby improving the maintenance and fault handling efficiency of power facilities. The present invention includes collecting equipment data of ring main units, branch boxes, and substations, analyzing these data using machine learning models to identify potential equipment failures, and monitoring the status of these devices in real time. In short, the technical effects of the present invention include improving the operational safety and reliability of power facilities, optimizing maintenance response time and quality, and enhancing the work efficiency and decision-making quality of on-site maintenance personnel through real-time data analysis and expert system support. In addition, the present invention can also reduce economic losses and equipment damage caused by improper fault handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flowchart of a collaborative control method of a ring main unit, a branch box, and a substation based on AR.

[0015] Figure 2 It is a schematic diagram of a collaborative control method of a ring main unit, a branch box, and a substation based on AR according to the present invention.

[0016] Figure 3 It is another schematic diagram of a collaborative control method of a ring main unit, a branch box, and a substation based on AR according to the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1-3 In this embodiment, a collaborative management and control method of a ring main unit, a branch box, and a substation based on AR may specifically include:

[0019] S101. Acquire common faults and corresponding abnormal power data of ring main units, branch boxes and substations, extract fault features from the abnormal power data and associate them with corresponding fault types, and establish a fault identification model for identifying corresponding fault types based on the power data of power facilities and calculating the corresponding fault occurrence probability.

[0020] According to the historical operation records and maintenance logs, the fault types of ring network cabinets, branch boxes and substations and the corresponding abnormal power data features are collected to establish a database of association mapping relationships between fault types and abnormal power data; the collected abnormal power data is preprocessed to remove noise and outliers, and key parameters that can reflect fault characteristics are extracted. The key parameters are processed by feature engineering to obtain fault feature vectors with discrimination; according to the database of association mapping relationships between fault types and abnormal power data, the extracted fault feature vectors are labeled to construct a training data set of fault feature vectors and fault types; a machine learning algorithm is used to train the fault recognition model The hyperparameters of the fault identification model are optimized through cross-validation and grid search; the trained fault identification model is deployed to the online environment, and the power data of the ring main unit, branch box and substation are received in real time. The power data is extracted and preprocessed, and input into the fault identification model for prediction to identify potential fault types; for the identified fault type, the probability of occurrence of the fault type under the current operating conditions is calculated based on historical data and expert knowledge, and a fault risk assessment report is generated; new operating data and fault records of the ring main unit, branch box and substation are collected, and the fault identification model is retrained and optimized.

[0021] Specifically, according to the historical operation records and maintenance logs, the common fault types of ring network cabinets, branch boxes and substations and the corresponding abnormal power data features are collected to establish a database of the association mapping relationship between fault types and abnormal power data. The collected abnormal power data is preprocessed, and the median filter, moving average and other methods are used to remove noise and outliers, and key parameters that can reflect fault characteristics are extracted, such as voltage, current, power, temperature, etc. For time series data such as voltage and current, the sliding window method is used to extract statistical features within a period of time, such as mean, variance, maximum value, minimum value, etc.; for numerical data such as power and temperature, the binning method is used to convert continuous variables into discrete variables; for the extracted features, feature selection algorithms such as filtering, wrapping, embedding, etc. are used to select the most discriminative feature subsets to form fault feature vectors. According to the association mapping relationship database between fault types and abnormal power data, the extracted fault feature vectors are labeled to construct a training data set of fault feature vectors and fault types. In order to deal with the problem of unbalanced label data, data balancing technologies such as oversampling and undersampling can be used, such as SMOTE algorithm, random undersampling, etc. Using machine learning algorithms, such as support vector machines, random forests, neural networks, etc., multiple different fault identification models are trained on the training set. Through cross-validation, the performance of the model is evaluated on the validation set, and the model with the best performance is selected; for the selected model, grid search or random search is used to optimize the model's hyperparameters, such as the kernel function type and penalty coefficient of the support vector machine, the number of trees of the random forest, the maximum depth of the tree, etc., to improve the performance of the model. The trained fault identification model is deployed to the online environment, and the power data of the ring network cabinet, branch box and substation are received in real time. The data is feature extracted and preprocessed, and input into the fault identification model for prediction to identify potential fault types. In order to ensure the performance and efficiency of online prediction, model compression, model acceleration and other technologies, such as model pruning, quantization, knowledge distillation, etc., can be used to reduce the computational complexity and storage overhead of the model. At the same time, it is necessary to establish a monitoring and alarm mechanism for the model to monitor the performance indicators of the model in real time, such as accuracy, recall rate, delay, etc. When the performance indicators are lower than the preset threshold, the alarm is triggered in time to notify the operation and maintenance personnel to check and tune. For the identified fault types, based on historical data and expert knowledge, an association model between the fault type and the influencing factors is established, such as a fault tree model, a Bayesian network model, etc., to calculate the probability of occurrence of the fault type under the current operating conditions. The output results of the fault identification model are combined with the output results of the association model to generate a fault risk assessment report to provide decision support for operation and maintenance personnel.Continuously collect the operation data and fault records of ring main units, branch boxes and substations, regularly retrain and optimize the fault identification model to improve the adaptability and robustness of the model, and update the associated mapping relationship database between fault types and abnormal power data to ensure that the model can adapt to the ever-changing power grid environment and equipment status, and continuously improve the effect of fault identification and risk assessment. When collecting common faults and abnormal power data of ring main units, branch boxes and substations, extract the fault type and abnormal data features from the online fault case library provided by the equipment manufacturer with their consent, such as the current waveform distortion of the closing coil corresponding to the operating mechanism failure of the medium-voltage switchgear, the increase of the motor current caused by the jamming of the energy storage spring, etc. Taking 100 fault cases of each device as an example, an associated mapping database containing 10,000 records is constructed. In the data preprocessing stage, for voltage and current data with a sampling frequency of 100Hz, a sliding window of 50 data points can be used to extract 20 statistical features such as the mean and root mean square value in each window; for temperature data, it can be divided into three intervals: "low temperature (<-10℃), normal temperature (-10℃~40℃), and high temperature (>40℃)"; the chi-square test is used to select 30 optimal features with a p value less than 0.01 from all features to form a feature vector. In order to solve the problem of uneven distribution of the number of fault type labels, the SMOTE algorithm is used to oversample the severe fault types with a small sample size, such as fault arc and switch refusal, by 3 times to balance the number of samples of each type. Then, the data set is randomly divided into training set and test set by 10-fold cross validation. Models such as support vector machine (SVM), random forest (RF), and gradient boosted decision tree (GBDT) are trained on the training set. The penalty coefficient C of SVM and the number of decision trees of RF are optimized by grid search. The RF model with a fault recognition accuracy of 95% on the test set is selected for deployment in the online system. To ensure the efficiency of real-time prediction, the RF model is subjected to feature selection based on L1 regularization, and the feature dimension is compressed from 30 to 20. At the same time, asynchronous threads are used to process data to control the average prediction delay within 50ms. When the fault underreporting rate of the model is monitored to exceed 0.5% for three consecutive days, the model retraining process is automatically triggered to update the model parameters. Combined with the fault identification results, the Bayesian network is used to establish an association model between the fault cause and the equipment operating conditions. The inference calculation shows that under the current ambient temperature of 38.5℃, humidity of 75%, and equipment operation time of 18 months, the probability of an operating mechanism failure in a certain switchgear is 1.2%. Finally, a risk assessment report with information such as the cause of the fault, probability of occurrence, and countermeasures is generated to guide preventive maintenance on site. By collecting information on components replaced during each maintenance, the fault case library and association model can be continuously optimized to continuously improve the accuracy of fault identification and risk assessment.

[0022] S102, determining maintenance tasks based on the output of the fault identification model, and allocating maintenance tasks based on the skills mastered by the current on-duty personnel, and obtaining equipment data of corresponding equipment through the on-site AR system when the current on-duty personnel perform maintenance tasks.

[0023] Based on the output results of the fault identification model, combined with the fault type, fault severity, and fault location information, a corresponding equipment maintenance task list is generated; a skill knowledge graph of on-duty personnel is constructed, which contains the basic information, skill level, skill type, proficiency attributes of each on-duty personnel, and the mutual evaluation relationship between employees. Through the connection relationship between employee nodes and skill nodes in the knowledge graph, the comprehensive ability index of each employee to master each skill is calculated; each task in the equipment maintenance task list is matched with the skill node in the on-duty personnel skill knowledge graph, and the skills required for the task and the skills of the on-duty personnel are calculated. similarity between them; according to the task allocation plan, a work schedule for each on-duty personnel is generated, and the work schedule contains task details, time arrangements, and route planning information; when the employee arrives at the site, the target equipment is identified and located through the AR device; according to the real-time status data of the equipment and the employee's operating behavior, the employee's operating process is analyzed and judged whether it is standardized, and early warning prompts are given for potential dangerous operations or improper behaviors. At the same time, the employee's operating process and results are recorded, and a work report is generated; after the employee completes the work, the content of the work report is reviewed and analyzed, abnormal situations are marked and alarmed, and the on-duty personnel's skill knowledge map is updated at the same time.

[0024] Specifically, according to the output results of the fault identification model, combined with information such as fault type, fault severity, and fault location, the corresponding equipment maintenance task list is automatically generated. The task list includes attributes such as task number, task description, priority, and expected completion time. The knowledge graph technology is used to construct the knowledge graph of the skills of the on-duty personnel. The graph contains the basic information, skill level, skill type, proficiency, and other attributes of each on-duty personnel, as well as objective data such as the employee's historical work performance and training assessment records. Through the connection relationship between the employee node and the skill node in the graph, the comprehensive ability index of each employee to master each skill is calculated. Each task in the equipment maintenance task list is matched with the skill node in the knowledge graph, and the similarity between the skills required for the task and the skills of the on-duty personnel is calculated. The personnel with the highest similarity and the lowest work saturation are selected as candidates. At the same time, factors such as the priority of the task and the geographical location of the employee are considered. The optimal task allocation plan is obtained through a multi-objective optimization algorithm, such as a genetic algorithm. The objective function includes maximizing the matching degree between the task and the employee's skills, minimizing the employee's work saturation, and minimizing the employee's moving distance. The constraints include that each task must be assigned to an employee and the workload of each employee cannot exceed his or her capacity limit. According to the task allocation plan, a work schedule is generated for each on-duty personnel. The schedule contains task details, time arrangements, route planning and other information, which are pushed to employees' mobile terminals through mobile apps, work emails, text messages and other methods, so that employees can understand and perform tasks in a timely manner. After arriving at the site, employees can quickly identify and locate target equipment by wearing AR glasses or handheld AR devices and scanning the QR code or label on the equipment. The AR system collects equipment status parameters, alarm information and other data in real time through IoT sensors and equipment interfaces, and displays the data on the physical equipment. At the same time, according to the equipment's historical maintenance records, operation manuals and other information, employees are guided to perform standardized operations through virtual annotations, voice prompts and other methods, so that employees can intuitively understand equipment information and operation procedures. The on-site AR system collects employee video images through cameras, and uses pre-trained deep neural network models, such as C3D networks and dual-stream networks, to analyze and judge the employee's operating behavior in real time, identify behavioral categories such as "standard operation" and "dangerous operation", and then use rule-based expert systems to determine whether the employee's operation process complies with predefined operating specifications. For possible dangerous operations or improper behaviors, early warning prompts are given in a timely manner. At the same time, the employee's operating behavior category and standardization score are recorded, and an operation report is automatically generated. After the employee completes the work, the report content is automatically reviewed and analyzed, and abnormal situations are marked and alarmed. At the same time, the employee's skill knowledge map and personal skill model are updated, task allocation and personnel scheduling strategies are optimized, and the intelligent level of operation and maintenance management is improved.Continuously collect employees' operation data and feedback, iteratively optimize knowledge graphs, task allocation algorithms, AR systems, etc., continuously improve the practicality and accuracy of the system, and provide comprehensive, efficient and intelligent decision-making support for the company's equipment maintenance work. The fault identification model is trained by the support vector machine algorithm based on equipment operation data such as temperature, pressure, vibration, etc. and historical fault cases. The fault identification accuracy of the model on the test set reaches 95%, and it can identify 20 common fault types, such as motor bearing wear and poor gearbox lubrication. When the equipment is found to have signs of failure, the model will automatically generate maintenance tasks and assign priorities according to the severity of the fault. For example, minor faults or mild wear have a priority of 3 and need to be handled within 1 week; serious faults such as bearing burnout have a priority of 1 and need to be handled within 24 hours. The skill knowledge graph of on-duty personnel is built using the Neo4j graph database, which includes 120 employees and 30 skills. Each employee has an average of 5 skills, and the proficiency is quantified by 0-100. The performance score and training test results of each employee in the past year are extracted from the OA system to form a 50-dimensional skill feature vector. When assigning tasks, the Word2Vec algorithm is first used to calculate the cosine similarity between the skills required for the task and the skills of the employees, and then the genetic algorithm is used for combined optimization. The fitness function is set to the similarity mean / variance + employee idleness and busyness balance, which converges at the 200th generation, takes 5 seconds to solve, and the matching accuracy is improved by 20%. The video images collected by the AR device are input into the C3D network at a frequency of 10 frames per second. Through 3 3D convolutional layers and 2 fully connected layers, real-time recognition of 6 common dangerous operations, such as not wearing a safety helmet or insulating clothing, is achieved, and the frame-level recognition accuracy is more than 90%. The expert system contains 200 safety rules, such as high-voltage operations must be carried out by more than two people, and the operation process is checked frame by frame to determine whether it is in violation of the rules. In the one year since the expert system was put into operation, the standard operation rate of employees has increased from 85% to 95%, and the equipment failure rate has decreased.

[0025] S103: Obtain the idle status of the current expert, and in combination with the personnel management system, send the idle status of the current expert to the terminal of the current on-duty personnel or display it on the current on-duty public screen.

[0026] An expert idle time prediction model is established based on the expert's historical work log and scheduling plan. The expert idle time prediction model inputs the expert's historical work log data, including working time, location, type and other attributes, and outputs the probability that the expert is idle within a certain time period; based on the expert's location information obtained in real time, whether the expert is in the work area and whether he is idle is determined by the expert's location information; based on the expert's physiological and behavioral data, whether he is in a working state is determined by the expert's physiological state; the expert's call records and instant messaging records are analyzed to extract call time, object, and keyword information to determine whether the expert is conducting work exchanges or remote guidance; the expert idle time prediction model, real-time location information, physiological behavior data, and communication record analysis results are comprehensively judged to obtain the expert's real-time idle status, and matched with the personnel management system to determine the expert's idle time period in the scheduling plan; the expert's idle time period in the scheduling plan is sent to the on-duty personnel's mobile terminal.

[0027] Specifically, according to the expert's historical work log and shift plan, an expert idle time prediction model is established. The LSTM algorithm is used. The input is the multi-dimensional time series data of the expert's historical work log, such as working time, work location, work type, etc., and the output is the probability of the expert being idle in the next hour. The LSTM network consists of an input layer, an LSTM layer, and the LSTM layer contains 20 neurons, a fully connected layer, and an output layer. It is trained and predicted through 20 time step sequence data, and the prediction accuracy can reach more than 85%. Obtain the expert's behavioral data, including location, physiological characteristics, call records, etc., and judge the expert's working status through comprehensive analysis. Based on the expert's location information, it can be collected through the positioning devices worn by the expert, such as UWB and RFID. The positioning accuracy can reach the centimeter level. The location information can be used to judge whether the expert is in the working area and whether he is idle. Based on the expert's physiological data, such as heart rate, body temperature, number of steps, etc., it is collected through wearable devices worn by the expert, such as smart watches and smart clothes. According to the expert's physiological state, it is judged whether he is in a working state. For example, a high heart rate and a high body temperature may indicate that the expert is working at a high intensity and is not idle. Analyze the expert's call records and instant messaging records, extract call time, object, keywords and other information, and determine whether the expert is conducting work communication or remote guidance. The call duration threshold can be set to an artificial experience value, such as 5 minutes, and then dynamically adjusted through data analysis. Comprehensively judge the expert's idle time prediction model, real-time location information, physiological behavior data, communication record analysis results and other multi-dimensional data, and use the weighted average method. The weight can be determined by expert scoring or machine learning algorithms, such as Adaboost. The result of the judgment can be a quantitative indicator of idleness, such as 0-100 points, to obtain the expert's real-time idle status, and match it with the personnel scheduling system to determine the expert's idle time in the scheduling plan. Send the expert's real-time idle status information to the on-duty personnel's mobile terminal, such as a mobile phone or tablet computer, and remind them in the form of pop-ups, messages, voice, etc. At the same time, the idle status is displayed in a visual way on the large screen of the duty room, including the real-time location distribution of experts, the proportion of experts in different idle states, and the time series trend of experts' idle time. Use intuitive visualization methods such as maps, pie charts, and line charts to facilitate on-duty personnel to view and analyze in real time. The on-duty personnel can use the mobile terminal to provide feedback on the actual idleness of the experts. For example, if the expert is shown as idle but is actually handling other work tasks, the idleness of the expert will be automatically updated after receiving the feedback to improve the accuracy of the status judgment. At the same time, the idleness prediction model can also be optimized in real time through reinforcement learning algorithms such as Q-Learning. The state space is the expert's historical work log and real-time behavior data, the action space is the predicted idle probability, and the reward is the difference between the predicted result and the actual feedback. The real-time and accuracy of the prediction can be improved through continuous exploration and learning.By collecting the work log data of experts in the past three months, including daily work hours, locations, types, etc., each expert has an average of 500 records, totaling 2000 data. The data are arranged in chronological order, with every 30 minutes as a time step, and 10 features such as work status and number of collaborative experts are extracted to form a 50-dimensional input vector. TensorFlow is used to build an LSTM network, which includes 1 50-neuron input layer, 2 LSTM layers with 20 neurons in each layer, 1 10-neuron fully connected layer and 1 1-neuron output layer representing the idle probability. The cross entropy loss function and Adam optimizer are used, the learning rate is set to 0.01, the batch size is 32, and 100 epochs are trained. The prediction accuracy on the test set reaches 88%. The location data of the UWB positioning tag worn by the expert is matched with the electronic map of the office area, laboratory, conference room and other areas once per second with an accuracy of 5cm to determine whether the expert is in the work area, and the expert's work status is inferred based on the expert's schedule. At the same time, according to the heart rate data collected by the smart bracelet worn by the expert, the heart rate threshold is set to 100 beats / minute. If the heart rate exceeds this threshold for 5 minutes, the expert is considered to be in working state. The call records of the experts are semantically analyzed to extract keywords, such as "plan discussion", "technical consultation", "on-site guidance", etc. If the cumulative call duration exceeds 1 hour in a day and the keyword frequency exceeds 10 times, the expert is considered to be in working state. The Adaboost algorithm is used to weightedly fuse the judgment results of the above three dimensions to obtain the comprehensive idleness score of the expert. The weights are obtained through grid search, which are 0.5, 0.3, and 0.2 respectively. When the idleness score is >80 points, a push message "XXX expert is currently idle and can be contacted to assist in solving the problem" will pop up on the mobile App of the on-duty personnel, and the expert's name, photo, idle duration, idle reason and other information will be displayed on the large screen. After each call with the expert, the on-duty personnel must feedback the contact effect on the App, such as "connected", "no one answered", and "busy", and automatically record the feedback log for correcting the idle state. At the same time, the Q-Learning algorithm is run every day, with the expert's hourly work status and physiological data as the state space, the idle probability prediction value as the action space, the feedback log as the reward, the learning rate is set to 0.1, the discount factor is set to 0.9, and after 10,000 iterations, the prediction accuracy is improved from 88%.

[0028] S104. Classify the maintenance tasks of the ring main unit, branch box and substation, including analyzing the urgency of the maintenance task based on whether the maintenance task is assigned to the task list of the current on-duty personnel. The urgency of the maintenance task includes high urgency and low urgency. The maintenance tasks are divided into corresponding processing priorities according to the importance and impact scope of the corresponding faulty equipment.

[0029] Obtain all pending maintenance tasks for the ring main unit, branch box, and substation from the maintenance management system, extract the equipment type, fault type, fault level, occurrence time, and impact range attributes of the task, construct a task feature vector, use the KNN algorithm to cluster the tasks, and divide them into task sets of different urgency according to the clustering results; track the task list of each on-duty personnel, obtain the planned completion time, actual completion progress, and completion quality score information of the task, and analyze the correlation between task characteristics and completion performance; model the topological structure of the ring main unit, branch box, and substation, construct a power equipment connection diagram, and use the PageRank algorithm to calculate the weight of each device node. The system uses task urgency and equipment importance as two dimensions to sort maintenance tasks in two dimensions, build a task priority matrix, and divide tasks into different processing priorities according to their positions in the matrix. For high-priority maintenance tasks, a centralized scheduling strategy is adopted to optimize task allocation. For medium and low-priority maintenance tasks, a distributed scheduling strategy is adopted, and each on-duty personnel takes the task independently according to their own situation. The processing of maintenance tasks is monitored in real time, and on-site operation data is collected through sensors and monitoring equipment. Compliance checks are carried out on operation behaviors, and abnormal situations that occur during the operation are warned and recorded in the task operation log.

[0030] Specifically, all pending maintenance tasks of ring main units, branch boxes and substations are obtained from the maintenance management system, and the attributes of the tasks such as equipment type, fault type, fault level, occurrence time, and impact range are extracted. The task feature vector is constructed, and the K-means algorithm is used to cluster the tasks. The sum of squared Euclidean distances between tasks within all clusters is minimized through iterative optimization until the cluster center no longer changes. The tasks are divided into categories with different degrees of urgency, such as urgent, important, and minor. The response time for urgent tasks is required to be within 1 hour, for important tasks within 24 hours, and for minor tasks within 7 days. The task list of each on-duty personnel is tracked in real time to obtain information such as the planned completion time, actual completion progress, and completion quality score of the task. The Pearson correlation coefficient is used to measure the linear correlation between task characteristics and completion performance. For discrete variables, such as fault type, One-Hot encoding is first used to convert them into numerical variables, and then the correlation coefficient matrix between each feature and completion performance is calculated. According to the significance test, such as t-test, the significantly correlated features are determined, and these features are used as input variables, and the completion performance is used as the output variable to build a linear regression model. When a new task arrives, it is substituted into the model for completion performance prediction, and the tasks with predicted completion performance below the average level are focused on and resources are tilted to improve the task completion rate. The topological structure of the ring network cabinet, branch box and substation is modeled, and a power equipment connection diagram is constructed. The graph database Neo4j is used to store device nodes and connection relationships. Attributes such as failure rate, load rate, number of power supply households, and affected range are added to each device node. The PageRank algorithm is used to calculate the importance weight of each device node. The higher the weight, the more critical the role of the device node in the entire power supply network. Once a fault occurs, the larger the impact range, and it needs to be handled first. The AHP hierarchical analysis method is used to construct the task priority matrix. The task urgency and equipment importance are used as two criterion layers, and weights are assigned respectively, such as 0.6 and 0.4. Three indicator layers, such as urgent, important, and secondary, are set under each criterion layer. Pairwise comparisons are made to form a judgment matrix. The weight vector of the indicator layer is calculated and a consistency test is performed. Then the tasks are scored according to the priorities of the two criterion layers, and the total priority score is obtained by weighted summation. The higher the score, the higher the priority. According to the priority, the tasks are divided into three levels: high, medium, and low, and the corresponding scheduling strategy is established.For high-priority maintenance tasks, a centralized scheduling strategy is adopted. The scheduling center uniformly balances the task and resource allocation, and uses genetic algorithms to optimize task scheduling, with the optimization goals of minimizing task response time and maximizing equipment availability. The task execution order, the matching degree between personnel and equipment, etc. are encoded, and the global optimal scheduling scheme is obtained through iterative optimization of genetic operations such as selection, crossover, and mutation. For medium and low-priority maintenance tasks, a distributed scheduling strategy is adopted. Each on-duty personnel takes the task independently according to their own position, task familiarity and other factors. The scheduling center is responsible for task status tracking and priority dynamic adjustment. Through IoT sensors, monitoring equipment and other real-time collection of on-site operation data, such as personnel location, equipment status, tool use, etc., combined with operation standards, the operation process is monitored in real time using anomaly detection algorithms, such as the isolated forest algorithm, to find abnormal situations, such as equipment alarms, personnel standing still for a long time, etc., and timely warnings are issued. Real-time message push, sound and light alarms and other methods are used to notify relevant personnel to deal with them, and emergency measures such as reassigning personnel and stopping operations are taken according to the severity of the abnormality. The abnormal situation is recorded in the operation log as a basis for subsequent performance evaluation and process optimization. Using natural language processing technology, semantic analysis is performed on unstructured data such as historical maintenance work orders, fault reports, and operation tickets to extract key elements such as equipment type, fault phenomenon, processing method, and processing time to form structured knowledge units; then, the ontology construction method is used to define core concepts such as faults, equipment, methods, and tools to form a maintenance domain ontology, and knowledge units are stored based on the graph database Neo4j to build a knowledge graph. Through the graph algorithms of the graph database, such as the shortest path and clustering, association reasoning and semantic search are realized. When a new maintenance task arrives, it automatically matches similar historical tasks, and recommends processing methods and experiences that can be used for reference by the on-duty personnel to improve processing efficiency and quality. A total of 5,230 maintenance task data for ring network cabinets, branch boxes, and substations in 2022 are extracted from the maintenance management system. Each task contains 12 attributes such as equipment type, fault type, number of affected users, and duration. Min-Max normalization is used to preprocess the data and construct a task feature vector. Then, the silhouette coefficient method is used to determine that the optimal number of clusters is 4, and the K-means algorithm is used to iteratively optimize the cluster center until the center point changes less than 1e-4, and the tasks are divided into 4 categories: processing within 1 hour, processing within 24 hours, processing within 7 days, and processing more than 7 days, accounting for 5%, 20%, 45% and 30% respectively. A correlation analysis is conducted on a total of 120,000 task processing records of 1,000 operation and maintenance personnel in 2022. The Pearson correlation coefficient is calculated to show that the correlation coefficients of task urgency, equipment importance and completion time are 0.78 and 0.65 respectively, which are significant at the 0.01 confidence level through the t-test. Based on this, a multivariate linear regression model for task completion time prediction is constructed, and the average prediction error of 5 cross-validations is 1.5 hours.The NetworkX library is used to model the primary and secondary system topology of the substation. A total of 532 equipment nodes such as transformers, switches, busbars, and feeders are extracted. The node attributes are mapped, including 8 indicators such as rated capacity, overload rate, and failure rate. The importance weight of each node is obtained by PageRank iteration. Among them, the 220kV main transformer and 35kV busbar section switch have the highest weights, which are 0.15 and 0.12 respectively. The task priority evaluation matrix is ​​constructed based on the task urgency and the equipment importance weight. The weights of the two criterion layers are determined to be 0.6 and 0.4 respectively through the hierarchical analysis method. The weights of the indicator layer are calculated by pairwise comparison. The consistency ratio is less than 0.1, and the task priority score is obtained. The natural breakpoint method is used to divide the tasks into three priorities: high, medium, and low, with a proportion of 15%, 50%, and 35% respectively. For high-priority tasks, a genetic algorithm is used for centralized scheduling optimization, with the minimum task response time and the maximum equipment availability as the optimization goals. The encoding method is real number encoding, the population size is 100, the crossover probability is 0.8, the mutation probability is 0.1, and the iteration is 100 times. The average response time of the optimal scheduling scheme is reduced by 30% compared with manual scheduling, and the equipment availability is increased by 5%. The operation process data is sampled every 30 seconds, and a 15-dimensional feature vector is constructed to input the isolation forest anomaly detection model for discrimination. The abnormal state includes 12 types such as timeout, out of bounds, and static. The abnormal score threshold is set to 0.7, and the abnormal coverage rate of the model detection reaches 95%. The abnormal state is pushed to the duty supervisor in real time, and corresponding measures are taken according to the severity of the abnormality. The BERT pre-trained model was used to extract 10 key elements, including faulty parts, fault phenomena, processing methods, and tools used, from a total of 200,000 maintenance work orders in five years, with an average extraction accuracy of 92%. Then, an ontology model was built, 138 core concepts and 612 semantic relationships were defined, and 170,000 entities and 6.38 million relationship edges were stored based on the Neo4j graph database. When a new task arrived, the TransE-based knowledge representation learning model was used to calculate the semantic similarity between the task description and each task node in the ontology knowledge base, and the Top 3 tasks were taken as the recommendation results, with an average recommendation accuracy of 85%, which can quickly match relevant processing experience.

[0031] S105. For low-urgency maintenance tasks, if an expert request from the current on-duty personnel is detected, the equipment data of the corresponding equipment is synchronized to the expert AR system of the corresponding expert, and maintenance suggestions from the expert AR system are received in real time and synchronized to the on-site AR system. The maintenance suggestions are associated with the corresponding equipment in the on-site AR system and displayed in the display area of ​​the corresponding equipment.

[0032] By analyzing the task priority matrix of task urgency and equipment importance weight, combined with the task response time requirement, the scope of the low-urgency maintenance task is determined; according to the scope of the low-urgency task, a low-urgency label is marked in the maintenance management system; for the low-urgency task, when the on-duty personnel initiates an expert request, the location information, task information and equipment information of the on-duty personnel are obtained; semantic search is performed based on the equipment type and fault type keywords in the equipment information to obtain a list of experts related to the task; based on the idle status information of the experts in the expert list, the expert list is matched with the idle expert list to determine the most matching and idle expert; and a request for assistance is sent to the most matching and idle expert. When the most suitable and free expert receives a request for assistance, the 3D model, specification parameters, historical maintenance records and equipment parameter information of real-time operation data of the equipment are loaded through the expert AR system; the equipment parameter information is spatially aligned with the real-scene image of the equipment and displayed in the expert AR system; the health status and fault cause of the equipment are determined by analyzing the real-time parameter curve and alarm information of the equipment; according to the health status and fault cause of the equipment, equipment maintenance plans and equipment maintenance recommendation reports in similar cases are extracted; semantic understanding and key information extraction of equipment maintenance recommendation reports are performed; key information extracted from the equipment maintenance recommendation reports are pushed to the AR system of on-site duty personnel; after responding to the key information, the duty personnel carry out maintenance operations; the maintenance operation process is monitored, and when illegal operations are detected, early warnings are issued.

[0033] Specifically, by analyzing the task priority matrix of task urgency and equipment importance weight, combined with the task response time requirements, the scope of low-urgency maintenance tasks is determined, and tasks with response time requirements of more than 7 days are classified as low-urgency tasks and are labeled as low-urgency tasks in the maintenance management system; when the on-duty personnel receive tasks on the mobile terminal, the urgency of the task can be automatically judged. For low-urgency tasks, the "Expert Assistance" button is displayed on the task details page, allowing the on-duty personnel to initiate expert requests. After the on-duty personnel clicks the "Expert Assistance" button, the location information, task information, equipment information, etc. of the on-duty personnel are automatically obtained. According to keywords such as equipment type and fault type, semantic search is performed using the knowledge graph to obtain a list of experts related to the current task, and the idle status information and skill expertise of the experts are obtained at the same time. The collaborative filtering algorithm is used to calculate the similarity score between the expert and the task, and the top N experts with the highest scores and idle time are selected as candidates, and assistance requests are sent to the candidate experts. After receiving the request for assistance, the expert receives the request through the expert-side AR system, automatically loads the current equipment's 3D model, specification parameters, historical maintenance records, real-time operation data and other information, and spatially registers this information with the equipment's real-life image, and displays it in the AR glasses. The expert can control the display and hiding of information through voice commands, and control the rotation, scaling, and decomposition of the 3D model through gestures, and interact with the equipment information. With the help of the case-based reasoning (CBR) intelligent decision support system, the expert inputs the equipment fault characteristics, quickly retrieves historical cases similar to the current fault, analyzes the diagnostic ideas and maintenance methods in the cases, and optimizes the maintenance recommendations for the current equipment by combining real-time data and professional experience. A multimedia solution package containing the cause of the fault, processing steps, tool list, precautions and other contents is formed and uploaded to the cloud. After receiving the expert's maintenance recommendations, the cloud uses natural language processing technology to understand the semantics of the recommendations and extract key information. Based on the field ontology of equipment components, functions, faults, maintenance, etc., the extracted keywords such as equipment name, fault mode, and operation steps are mapped to the ontology concept, and the precise association between the recommendations and the equipment is determined through ontology reasoning, and the associated structured recommendation information is pushed to the on-site on-duty personnel AR system. On-site staff on duty wear AR glasses. After receiving maintenance suggestions from experts, they match the suggestions with the three-dimensional model of the current equipment and attach the suggestions to the corresponding parts of the model in the form of AR labels. The staff on duty stare at the labels to view detailed information such as text, images, and voice of the suggestions in a pop-up window. The labels can also display different colors and flashing frequencies according to the importance of the suggestions, reminding the staff on duty to pay priority attention.When carrying out specific maintenance operations, according to the operation steps in the expert's suggestions and the pre-recorded 3D animation clips, the clips are spliced ​​and combined into a complete 3D operation guide. Each animation clip in the guide corresponds to a specific step one by one. The on-duty personnel can control the playback, pause, and replay of the animation through voice commands, and can superimpose the animation with the real equipment to intuitively understand the position, strength, angle and other essentials of each operation step, effectively guide the on-site operation and improve the operation quality. After the operation is completed, the operation process data is automatically recorded, and the on-duty personnel are prompted to rate the guidance effect of the expert's suggestions and fill in brief feedback. The cloud performs sentiment analysis on the feedback, and statistically obtains indicators such as the adoption rate, availability, and satisfaction of the expert's suggestions. Combined with objective indicators such as fault diagnosis accuracy and average response time, the expert's assistance performance is comprehensively evaluated, and early warning prompts are issued to experts whose performance is continuously below the average level, and their priority is reduced in subsequent task allocation, so as to encourage experts to continuously improve the level of remote assistance. After statistically analyzing the historical response time of 1,352 low-urgency maintenance tasks, it was found that the average response time was 13.5 days. Tasks with a response time greater than one standard deviation from the average, that is, 17.2 days, were defined as low-urgency tasks, accounting for 18% of the total tasks. When the on-duty personnel receive tasks through the mobile terminal, the planned response time of the task is compared with the low-urgency task threshold. If the threshold is exceeded, the "Expert Assistance" button is highlighted on the page. When the button is clicked, the feature vector of the current task is extracted from 12 dimensions, such as professional field, professional title level, and geographical location, and semantic similarity matching and clustering are performed in the expert knowledge graph to select the Top 5 related experts; at the same time, the expert schedule is accessed to obtain the continuous idle time of each expert in the next 4 hours. The intersection of the two is taken and sorted by idle time, and the Top 3 experts are recommended and assistance requests are sent. After the expert puts on the AR glasses, the 16 built-in cameras in the glasses collect high-definition images of the equipment at a speed of 30 frames per second in real time, and synchronize the images with the 3D model of the equipment, historical maintenance records and other data streams, and then overlay and render them to the binocular display screen after alignment to achieve virtual-real fusion; the expert controls the calling out / hiding of information by tracking the gaze point coordinates of the eye movement, and controls the perspective and details of the 3D model by identifying the gestures of both hands, and the interactive response delay is less than 100ms. At the same time, the intelligent strategy recommendation engine retrieves 12 cases that are most relevant to the current fault from 372 historical maintenance cases, extracts key sentences and pictures from them to form a recommendation list, and presents it to the expert for reference. With the help of virtual maintenance simulation tools, experts drag and select relevant parts, tools, materials, etc., build a realistic on-site environment, set environmental parameters such as temperature and humidity, disassemble, test, calibrate, and replace parts, simulate and analyze the fault mechanism and maintenance effect, and describe the maintenance steps in natural language to form a report.The AR cloud platform performs semantic analysis on the unstructured maintenance reports uploaded by experts, and extracts key elements such as fault location, fault cause, maintenance tools, and operation steps by matching them with the concepts and relationships in the ontology knowledge base. Then, the semantic relationship of the elements in the ontology is used to verify and optimize the report logic, correct professional terminology, measurement units, causal logic errors, etc., and generate a highly structured maintenance plan. The structured maintenance plan is parsed into AR tags associated with specific equipment and pushed to the AR model of the on-site equipment; at the same time, the text description in the maintenance plan is mapped with the maintenance animation resources to automatically generate a three-dimensional visual maintenance guide. On-site technicians can enjoy immersive auxiliary guidance by staring at the AR tags, significantly improving maintenance efficiency and accuracy. After the plan is applied, the operation data of each link is recorded, including 17 indicators such as maintenance time, tool utilization rate, number of misoperations, and fault recurrence rate. Combined with the subjective evaluation of technicians and experts, the effectiveness of the maintenance plan is evaluated in multiple dimensions, and the comprehensive benefit improvement rate is calculated. If it is less than 5%, optimization suggestions are sent to experts.

[0034] S106. For high-urgency maintenance tasks, the current equipment data with timestamps are sent to the expert AR system of the corresponding expert, and a maintenance plan with timestamps is received from the expert AR system. When the high-urgency maintenance task is assigned to the current on-duty personnel, the maintenance plan is mapped to the on-site AR system.

[0035] According to the urgency of the high-urgency task, extract the key operating parameters of the equipment; perform data preprocessing on the key operating parameters of the equipment to form time series data; wherein the data preprocessing includes adding a clock-synchronized timestamp to the key operating parameters of the equipment; send the time series data to the expert AR system; integrate multi-source equipment data into the time series database; visualize the time series data; mark abnormal data points and alarm events by comparing the time series data with the rated parameter threshold of the equipment. According to the expert's line of sight position and duration, identify the equipment and measurement points that the expert is currently paying attention to; dynamically adjust the display order and detail level of the visualization parameters, including adjusting the display mode of the parameters according to the expert's attention; associate the parameters with the historical trend curves and associated measurement point data corresponding to the parameters to assist the expert in diagnosing the equipment status; combine the real-time status data of the equipment and the equipment parameters to perform attribution analysis and severity assessment of the equipment failure; the equipment parameters include the nameplate parameters, design drawings, operating procedures and fault manuals of the equipment; match the standard handling process for the failure mode.

[0036] Specifically, according to the urgency of the high-urgency task, the real-time data stream processing technology is used to continuously collect the key operating parameters of the equipment, such as voltage, current, power, temperature, pressure, etc., with a period of 100ms, and the data is quality checked to filter out abnormal values ​​and outliers, extract statistical features of the data, such as mean, variance, peak, etc., construct a multi-dimensional time series feature vector, attach a timestamp synchronized with the system clock, and send it to the expert AR system through industrial real-time communication protocols such as OPC UATSN to meet the millisecond-level low-latency transmission requirements. After receiving the equipment operating parameter stream, the expert AR system follows the information model of OPC UA, maps the static attributes of the equipment, such as nameplate parameters, and dynamic attributes, such as real-time measurement points, to unified node objects and variables, and uses a unified address space model for semantic representation, and then cleans and normalizes the data, integrates it into the time series database, and intuitively displays the equipment status through real-time curves and threshold comparisons. When experts wear AR glasses to view the status of equipment, the attention mechanism is used to dynamically adjust the display content according to the position and duration of the expert's line of sight, and gesture interaction is supported, allowing experts to zoom in, disassemble, and perform other operations on the equipment model in the area of ​​interest. Combined with expert experience and typical fault cases, the rule-based expert system makes a preliminary judgment on the fault mode. After the expert completes the equipment status analysis in the AR interface, the standard handling process for the fault mode is matched, and the specific equipment precautions supplemented by the voice are converted into structured data using speech recognition and natural language processing technology, which are automatically filled into the corresponding fields of the process template according to grammatical rules to form a personalized maintenance plan. The plan is semantically modeled, and key elements such as fault characteristics, severity, and handling steps are extracted. The case-based reasoning (CBR) method is used to calculate the semantic similarity between the current fault and historical cases, and the most similar K cases are selected. Considering factors such as case similarity, expert performance, and equipment health, the weighted average is used to estimate the fault handling time, generate a Gantt chart, and mark the links with high progress risks in red. The personalized plan is timestamped and stored in the plan library. The on-site AR system obtains incremental updates of the solution library through the edge gateway, dynamically loads new solutions to the AR terminal, and considers factors such as personnel's professional ability, geographical distance, and busyness when pushing them. It weights and scores each candidate, and selects the one with the highest comprehensive score to push. After the on-site personnel receive the solution, the AR terminal renders the solution workflow on the glasses screen in real time. The personnel can call up the device information and operation steps by staring at the AR label to obtain immersive guidance.During the operation, the operation behavior is continuously recorded in the form of video stream, and the structured information such as personnel location, action, tool use, etc. is extracted through target detection and behavior recognition, and compared with the solution workflow in real time. If the personnel are found to have illegal operations such as missing steps and wrong sequence, a voice prompt warning will be issued. At the same time, AR glasses also integrate environmental sensors to collect personnel's physiological state data, evaluate personnel fatigue and emotional state, and timely warn and replace high-risk personnel to double guarantee operation safety. The operation process data is transmitted back to the expert AR system through the edge gateway. Using the AR remote collaboration function, experts can establish a video connection with on-site personnel for real-time guidance and improve the timeliness of fault handling. Taking a transformer failure in a 220kV substation as an example, the concentration of dissolved gas in the transformer oil suddenly increased, and the acetylene concentration rose sharply from 2ppm to 150ppm. Data is collected every 100ms and cleaned and filtered, and transmitted to the remote expert system within 20ms through the OPCUATSN protocol. The expert observed the 3D visualization simulation of key parameters such as transformer oil temperature, pressure, and gas concentration through AR glasses, and called out the diagnostic rules of the dissolved gas three-ratio method to quickly determine that the fault type was high-energy discharge. Subsequently, the expert called out the standard processing flow of the "internal discharge of the transformer" fault, which has a total of 5 steps, and supplemented it through voice input with "Note that the power frequency withstand voltage test time should not exceed 1 minute to avoid damaging the winding". After semantic analysis, the entry was inserted into the third step of the process in JSON format. The cosine similarity between this fault and 3250 faults in the historical case library was calculated, and the most similar Top5 cases were extracted. The average handling time of the cases was 26 hours. Combined with the expert's historical performance of an average handling time of 21 hours, an overtime rate of 5%, and the transformer health of 12 years of remaining life, it was estimated that the handling time of this fault was 24±3 hours, and the fourth step with the risk of exceeding the time limit was marked in red, and the Gantt chart was generated and inserted into the end of the process to form a complete personalized "CT001" handling plan and save it. The AR on-site system comprehensively analyzes the feature vectors of the five operation and maintenance personnel, assigns weights of 0.5, 0.3, and 0.2 to the three indicators of skill matching, geographical distance, and busy and idle status, respectively, and calculates that the best candidate is Zhang San, and pushes the "CT001" plan to Zhang San's AR glasses. Zhang San arrives at the site and scans the QR code on the transformer nameplate. The five-step workflow and key parameters are immediately presented in front of him. Zhang San follows the process, and the glasses collect the operation video in real time and perform behavior recognition. When Zhang San misses the "disconnect the transformer connection line" in step 2, the voice reminder is timely and the warning is pushed to the expert end. The expert finds that Zhang San's pupils are dilated and his heart rate is increased. He judges that he is in a state of fatigue and recommends suspending the operation and letting Li Si take over until the fault is resolved. The whole process took 22.5 hours, which is 6.25% shorter than the estimated time. The expert added this case to the case database and marked the typical fatigue operation risks. The early warning of behavior is optimized by transfer learning. If similar features appear in the future, the early warning trigger time can be advanced by 5 minutes.The successful handling of this transformer failure has accumulated valuable practical experience for the company. Through review and summary, optimization suggestions were made for the existing two management systems and three operating procedures to further improve the level of equipment operation and maintenance management.

[0037] A future maintenance period is configured for the high-urgency maintenance task, so that the high-urgency maintenance task is assigned to the current on-duty personnel who have mastered the corresponding skill matching in the future maintenance period.

[0038] Based on the attributes of each on-duty personnel in the shift schedule, according to the shift time and shift duration, the available time window of each on-duty personnel is calculated, and the available time window is split into multiple time slots; based on the attributes of the high-urgency maintenance tasks, different priorities are assigned according to the urgency of the high-urgency maintenance tasks, and the high-urgency maintenance tasks are sorted by priority and deadline; the sorted high-urgency maintenance tasks are traversed, and the personnel time slots that match the skill level requirements of the high-urgency maintenance tasks and are free are searched from the available time windows of the on-duty personnel, and according to the attributes of the personnel time slots, the personnel time slots are sorted. Calculate the comprehensive score, obtain the preset personnel time slots with the highest comprehensive score, and generate a task-time slot-personnel triple; group the task-time slot-personnel triples by task, analyze the optimal personnel allocation plan for each task, and obtain the optimal executor and the earliest start time for each task; update the work schedule and statistical indicators of the on-duty personnel, update the status and execution plan of the high-urgency maintenance task, and generate a task execution work order; track the execution progress and personnel presence of the high-urgency maintenance task, and dynamically optimize task execution and personnel scheduling according to actual conditions; record the personnel performance after the completion of the high-urgency maintenance task.

[0039] Specifically, the shift schedule for the coming week is extracted from the personnel management system. The table contains the ID, name, shift, skill level, busy and idle status of each on-duty personnel. According to the shift schedule and shift duration, the available time window of each personnel is calculated. The time window is divided into multiple time slots with a minimum granularity of 15 minutes. Each time slot contains the start time and end time attributes. All high-urgency maintenance tasks are extracted from the equipment management system. Each task contains the task ID, equipment ID, fault type, fault severity, handling deadline, required skill level and other attributes. Different priority levels are assigned according to the urgency of the task. The task list is sorted from high to low by priority. Tasks with the same priority are sorted from near to far by deadline. Traverse the sorted task list. For each task, query the personnel time slots that match the task skill level requirements and are free in time from the personnel time window. Calculate the comprehensive score of each time slot based on the four factors of Priority, FreeTime, SkillMatch, and Workload. The weights of each factor are 0.3, 0.3, 0.2, and 0.2 respectively. The values ​​are normalized. Select the top 3 time slots with the comprehensive score and the corresponding personnel ID as candidates to form (task, time slot, personnel) triples. Group the above triples by task ID. For multiple candidate personnel for the same task, use the integer programming model to solve the global optimal allocation. The objective function is to minimize the total delay time. The constraints include that only one person is assigned to each task, each person only performs one task in the same time slot, and the continuous working time of the personnel does not exceed MaxWorkTime, such as 8 hours. Use 0-1 decision variables xi,j to indicate whether task i is assigned to personnel time slot j. The solution results in the optimal executor and the earliest start time of the task. The above allocation results are synchronized to the personnel management system and the equipment management system, and the allocated time slots are marked in the personnel's available time window, and the statistical indicators such as the personnel's skill utilization rate and working hours are updated in conjunction; the corresponding task status is updated to "assigned" in the equipment management system, and the executor ID, planned start and end time are recorded, and the order is dispatched to the corresponding personnel's work order pool. During the task execution process, the task progress and personnel location are tracked in real time. If the task progress is ahead of schedule, it is evaluated whether the personnel are allowed to receive new tasks after completing them in advance to avoid wasting time; if the task progress delay exceeds the preset threshold, such as 30 minutes, the rescheduling process is triggered: evaluate the impact of task delay on subsequent related tasks, such as power recovery, update the priority and deadline of subsequent tasks, re-execute the task sorting and allocation algorithm, adjust the task schedule, and arrange backup personnel or spare parts for key tasks in advance to ensure the smooth execution of the task chain.After the task is completed, the personnel management system automatically records the actual working hours, efficiency, quality and other performance indicators of the personnel, and associates the specific task type, severity, completion means and other multi-dimensional labels. The completion results are accepted and confirmed from the perspectives of personnel self-evaluation, leadership rating, user satisfaction, etc., and the results are entered into the task management system as the basis for subsequent task allocation and performance appraisal. At the same time, the collaborative filtering algorithm is used to continuously model the skill profile of the personnel, collect the efficiency and quality ratings of each person performing different types of tasks, and construct a "personnel-task" rating matrix. The matrix decomposition technology is used to decompose it into a personnel matrix and a task matrix, reconstruct the prediction personnel's rating of the unperformed tasks, and characterize the types of tasks that the personnel are good at or not good at. When the subsequent tasks are assigned, the matching degree is directly calculated using the learned personnel profile and task profile, and the task allocation is optimized to achieve the collaborative management of skills and performance and stimulate the enthusiasm of employees. Taking the distribution operation and maintenance scenario of a power company as an example, the scheduling plan for the next 7 days is extracted from the human resources management system, involving 18 operation and maintenance personnel, each shift lasts 8 hours, and the total time window is 1008 hours, which is divided into 4032 time slots by 15 minutes. At the same time, 9 high-urgency maintenance tasks are extracted, with priorities ranging from 1 (highest) to 5, and the handling deadline spans from 2 hours to 48 hours. After sorting the task priority and deadline combination, each task is traversed and the matching time slot is searched in the personnel time window. For example, if the priority of task A is 2 and high-voltage maintenance skills are required, the idle time slots of personnel with this skill and priority less than or equal to 2 are screened out, and then the normalized scores of time slot priority of 0.8, idle time of 120 minutes, i.e. 0.6, skill matching degree of 0.9, and task saturation of 0.3 are evaluated respectively. The weighted summation obtains a comprehensive matching degree of 0.68, and finally the top 3 personnel time slots with the highest matching degree are selected. The integer programming software CPLEX is used to solve the task allocation, and the objective function is set to minimize the total delay time. Constraints such as task uniqueness, time slot mutual exclusivity, and working time upper limit, such as 8 hours, are added. After 30 seconds of iteration, the optimal solution is obtained. Task A is assigned to Zhang San, and the planned start time is 8 am the next day. Automatically update Zhang San's work calendar, lock the corresponding time slot, and synchronize the task dispatch information. During the task execution process, obtain Zhang San's GPS location and work order progress every 5 minutes. If it is completed 1 hour in advance, a new task will be pushed immediately; if it is overdue by 30 minutes, analyze the subsequent impact of task A, such as raising the priority of power supply restoration task B from 3 to 2, re-execute the Hungarian algorithm to reallocate task B, and reserve spare transformers and other materials for it. After completing the task, Zhang San submitted a self-assessment score of 4.5 out of 5 points, and the leader scored 4.8 points. Combined with the user satisfaction of 98%, the weighted average score was calculated to be 4.7 points and recorded in the employee performance file.At the same time, the score is updated to the corresponding elements of Zhang San and "high-voltage electricity maintenance" in the "personnel-task" matrix. The gradient descent method is used for matrix decomposition, with a learning rate of 0.01 and 500 iterations. The reconstructed prediction shows that Zhang San's score for the "low-voltage electricity maintenance" task is 4.1, and Zhang San will be recommended first in subsequent related task matching.

[0040] An equipment prediction model is established in combination with historical equipment data, and future equipment data in future maintenance periods is predicted based on current equipment data, and the future equipment data is sent to the expert AR system.

[0041] Based on historical equipment data, a machine learning algorithm is used to train the equipment prediction model; the real-time operation data of the current equipment is obtained, and it is converted into the input format of the equipment prediction model through data preprocessing. The trained equipment prediction model is used for prediction reasoning to obtain the predicted values ​​of various parameters of the equipment in the future. According to the data interface requirements of the expert AR system, the predicted future equipment data is converted into a format and sent to the expert AR system through the data transmission interface to realize the visualization of the future operation status of the equipment in AR. The evaluation results of the prediction results fed back by the experts through the AR system are obtained, and the equipment prediction model is optimized using the evaluation results to make the prediction results fit the actual operation status of the equipment. According to the equipment prediction results, the optimal maintenance period of the equipment is determined, and the maintenance period recommendations are sent to the operation and maintenance personnel through the expert AR system. The maintenance period is dynamically adjusted based on the feedback from the operation and maintenance personnel to obtain the optimal maintenance plan that takes into account both equipment operation and personnel arrangement.

[0042] Specifically, based on the equipment's operating data over the past three months, including 50 parameters such as temperature and rotation speed, the LSTM neural network algorithm is used to train the equipment prediction model on the Tesla V100 GPU. The Adam optimization algorithm is used to set the model's learning rate to 001, and the training is iterated for 500 epochs to obtain an equipment prediction model with a prediction accuracy of 95%. The real-time operation data of the current equipment is obtained through an industrial Internet of Things platform, such as the Huawei IoT platform. After data cleaning and normalization, it is converted into a 50-dimensional input vector and input into the trained LSTM model to predict the trend of each parameter of the equipment in the next 24 hours. According to the JSON data interface specification of the expert AR system, the predicted parameter values ​​of the equipment temperature, rotation speed, etc. in the next 24 hours are converted into a standard JSON format, transmitted to the expert AR system through the MQTT protocol, and intuitively displayed in the form of a curve chart in the AR scene. The prediction accuracy score fed back by the expert through the AR system is obtained. When the accuracy is less than 90%, the incremental learning method is used to retrain the LSTM model using the latest equipment operation data to increase the model prediction accuracy to more than 95%. According to the future operation trend of the equipment predicted by the LSTM model, the optimal equipment maintenance time is calculated by the evolutionary algorithm as the 20th hour in the future, and the expert AR system pushes the maintenance period suggestion to the operation and maintenance personnel. If the operation and maintenance personnel report that there is a conflict in the personnel arrangement during this period, the particle swarm optimization algorithm is used to recalculate the maintenance period, and the optimal period that takes into account both equipment operation and personnel arrangement is the 18th hour in the future.

[0043] S107. When multiple maintenance tasks are performed simultaneously, the priority of information transmission between the on-site AR system and the expert AR system is adjusted according to the processing priority of the maintenance tasks.

[0044] Based on the execution status data of high-urgency maintenance tasks and the location data of the on-duty personnel; if the execution progress of the high-urgency maintenance task meets the preset early completion conditions, determine whether the on-duty personnel is allowed to complete the task in advance and receive a new task; if the execution progress of the high-urgency maintenance task meets the preset delay conditions, analyze the impact of the task delay on subsequent related tasks, determine whether it is necessary to trigger the task rescheduling process, re-assign tasks, and at the same time adopt emergency resource allocation for the related tasks of the high-urgency maintenance task.

[0045] Specifically, all ongoing maintenance tasks are extracted from the task management system. Each task contains attributes such as task ID, device ID, task type, priority, start time, and planned completion time. The task list is sorted according to the priority of the task. Tasks with the same priority are sorted from early to late according to the start time to generate an ordered task queue. In addition to the task attributes, each task element in the queue also contains the network communication parameters such as the IP address, port number, and protocol type of the on-site AR device corresponding to the task. For each task in the task queue, check the network connection status between the corresponding on-site AR device and the MEC edge gateway, use network detection tools such as ping and traceroute to obtain performance indicators such as network delay, jitter, and packet loss rate, and monitor network topology changes and load levels in real time to dynamically evaluate the available bandwidth and service quality level of the current network; if the network status cannot meet the data transmission requirements of the task, send an alarm to the task management system to trigger the network optimization process, such as adding backup links, switching network standards, and increasing transmission power. According to the task priority and network status, the data transmission strategy between the on-site AR device and the MEC edge gateway is dynamically adjusted. For high-priority tasks, multi-path transmission, link aggregation and other technologies are used to allocate exclusive high-quality data channels for them, and the data packet expedited marking and priority forwarding mechanism are enabled; for medium-priority tasks, differential coding, data compression and other methods are used to reduce the amount of transmitted data while ensuring data quality; for low-priority tasks, a best-effort transmission strategy is adopted to reduce the data sampling frequency, increase the cache area, transmit key data, reduce the requirements for transmission quality and real-time performance, and avoid occupying too many network resources. On the MEC edge gateway side, according to the task ID and priority tag of the received data packet, a multi-queue weighted polling algorithm is used to perform differentiated traffic scheduling and forwarding for data packets of different tasks. The queue weight is determined by the priority and delay sensitivity of the task. The weight calculation formula is Weight = 0.6*Priority+0.4*Sensitivity, where Priority represents the task priority of 1-5, and Sensitivity represents the delay sensitivity of 1-3. The two indicators are normalized and the weighted sum is obtained to obtain the queue weight. By dynamically adjusting the queue weight, the relative priority control of different task data flows is achieved, and data packets that exceed the delay threshold are discarded or retransmitted to ensure that the data of high-priority tasks are processed and forwarded first.For maintenance tasks that require expert remote guidance, the edge gateway packages the video, audio, sensor data, etc. collected by the on-site equipment according to priority, and transmits them to the expert AR system in the cloud through the backbone network. Based on real-time interactive protocols such as WebRTC, a point-to-point communication link is established between the on-site AR device and the expert AR device to ensure the real-time and synchronization of the audio and video data required for remote guidance. If there are abnormal conditions such as network jitter and excessive delay, it will automatically switch to the backup link and adjust the video encoding rate to ensure the smoothness of guidance when necessary. After receiving and processing the task data uploaded by the on-site AR device, the expert AR system on the cloud will package the expert's diagnosis results, maintenance plan, operation guidance and other feedback information with metadata such as task ID, and send them to the MEC edge gateway through high-speed cache components such as redis message queues. The gateway distributes the feedback information to the corresponding on-site AR device according to the task ID. The on-site device prompts the on-site staff with the feedback information in different forms such as floating windows, voice reminders, and vibration warnings according to the priority of the task. After the expert AR system receives the audio and video data stream from the scene, it flexibly selects edge rendering or cloud rendering according to the task priority and data characteristics. Edge rendering uses the AR device's own GPU and other computing resources to render 3D models locally, and cloud rendering offloads the rendering task to the cloud server's GPU cluster for parallel computing. For rendering tasks with low priority and high computational complexity, the cloud rendering mode is adopted to use the powerful parallel computing capabilities of the cloud to accelerate the rendering process; for tasks with high priority and high rendering real-time requirements, the edge rendering mode is adopted to use the local GPU of the expert AR device terminal to complete the rendering nearby, reduce transmission delay and bandwidth occupancy, and improve the interactive experience. In the expert AR system, experts can centrally manage and collaborate on the AR digital twin environments of multiple on-site maintenance tasks. Based on the experts' eye movement and gesture trajectories, they can use a deep learning algorithm based on the attention mechanism to predict the AR areas of interest that the experts may focus on. Combined with prior knowledge such as equipment topology and expert operation logs, a Bayesian network is constructed to predict expert intentions, and intelligently focus on the three-dimensional models of key equipment and the visualization panel of key data. Experts use the interactive tools of AR devices to simulate operations, annotations, and discussions in a virtual environment to form decision-making instructions, which are pushed to the corresponding on-site AR devices. Based on the results of the instruction execution, the corresponding AR twin environment status is updated in real time, supporting seamless collaboration between on-site and rear experts and efficiently completing multiple maintenance tasks.It also provides AR collaboration process customization functions for different tasks, allowing experts or on-site staff to flexibly configure the order, synchronization method, drill scenario, data presentation method, etc. of each subtask according to the characteristics of the task, forming a set of best practices, solidifying it into the task knowledge base, for subsequent similar tasks to refer to and reuse, continuously accumulating and optimizing AR collaboration mode, improving multi-task collaboration efficiency, and efficiently delivering expert experience to the front line, providing strong support for smart power grid operation and maintenance and safe and stable operation. A power grid company uses the task management system to prioritize 30 maintenance tasks of 10 substations, with priorities ranging from 1 to 5, including 5 level 1 tasks, 8 level 2 tasks, 10 level 3 tasks, 5 level 4 tasks, and 2 level 5 tasks. The ping tool is used every 5 minutes to test the latency between each on-site AR device and the MEC edge gateway. When the latency of the level 1 task exceeds 100ms and the packet loss rate exceeds 1%, it automatically switches to the backup 5G link. For the data transmission of level 1 tasks, MPQUIC protocol is used to achieve multi-path aggregation to ensure that the bandwidth is greater than 10Mbps; H.265 encoding is used for level 3 tasks, and the compression ratio is increased by 30%; the data sampling interval of level 5 tasks is enlarged to 1 minute, and the local cache capacity is increased to 1GB. The edge gateway sets the weight of level 1 tasks to 0.9, level 3 tasks to 0.6, and level 5 tasks to 0.3, and schedules them every 100ms. It retransmits the level 1 task data packets that have not been transmitted for more than 200ms and forwards them to the cloud first. The cloud server uses WebRTC to establish a two-way real-time audio and video channel with each on-site device, and the round-trip delay is controlled within 500ms and the jitter is less than 50ms. For the AR rendering task of level 1 tasks, the edge gateway dynamically applies for 5 VR hardware accelerators to render 3D models and data visualization interfaces locally in parallel, and the frame rate is maintained at 60FPS; for level 5 tasks, the rendering tasks are offloaded to 12 GPU servers in the cloud, and the frame rate is reduced to 30FPS. Experts observe the on-site situation through AR glasses, and the eye-tracking camera collects the line of sight position. The area with a line of sight overlap of more than 90% in 20 consecutive frames is extracted as ROI. The key equipment in the ROI is identified through the YOLOv5 model, and then combined with the knowledge graph to reason the expert's diagnostic intention, intelligently focus on the 3D model of the faulty equipment refined to the component level, highlight the fault parameter trend chart, and the expert fills in the diagnosis report using voice input. The voice recognition accuracy rate reaches 98%, and the diagnosis plan is pushed to the site in real time through the MQTT protocol, with a response delay of less than 1 second. Throughout the process, 10% of the expert diagnosis plans were solidified into standard operating specifications, the task knowledge base was expanded by 15%, and 12 similar tasks were completed with assistance, reducing the average time.

[0046] If the on-duty personnel are following up on a low-urgency maintenance task and an expert request is detected, the information transmission between the on-site AR system and the corresponding expert AR system is adjusted to the highest level to maintain real-time communication between the current on-duty personnel and the corresponding expert.

[0047] Obtain information on the urgency of the maintenance task currently being performed by the on-duty personnel to determine whether it is a low-urgency maintenance task; monitor in real time whether there is an AR remote assistance request from an expert; if the current on-duty personnel is performing a low-urgency maintenance task and an AR remote assistance request initiated by an expert is detected, set the priority of the information transmission channel between the on-duty personnel's on-site AR system and the expert AR system requesting assistance to the highest level; establish a real-time audio and video communication link between the on-duty personnel and the expert, and transmit the equipment operating parameters and fault information data collected by the on-site AR system to the expert AR system in real time; after receiving the on-site information through the AR system, the expert provides remote guidance to assist the on-duty personnel in completing the equipment maintenance task.

[0048] Specifically, the urgency of the current maintenance task is obtained through the task management system on the on-duty personnel's AR glasses, and whether it is a low-urgency task is determined according to the preset urgency threshold, such as 2 or below in the 5-level scale; the face recognition and voice recognition capabilities of the Huawei HiLens platform are used to detect in real time whether there is an AR remote assistance request initiated by an expert; once an expert request is detected and the current maintenance task is a low-urgency maintenance task, the priority of the information transmission channel between the on-duty personnel's on-site AR system, such as HoloLens, and the expert AR system is set to the highest level 7 according to the IEEE801p standard; H.265 video encoding and AES-256 encryption algorithm are used to improve communication efficiency while ensuring data security during the AR assistance process; WebRTC technology is used The technology establishes a real-time audio and video communication link between the on-duty personnel and the experts, and transmits the on-site equipment operating parameters, such as temperature, speed, fault information, such as the alarm code defined by the OPC UA standard, to the expert AR system in real time through the MQTT protocol; the experts provide remote guidance through AR projection, 3D annotation and other methods to assist the on-duty personnel in completing equipment maintenance, such as using the PTC Videoforia platform to identify specific equipment and superimpose relevant operation information; during the AR remote assistance process, the iperf tool is continuously used to monitor the bandwidth, delay, jitter and other quality indicators of the communication link. When the experience quality is lower than the MOS value of 5, the adaptive transmission strategy is triggered to dynamically adjust the video encoding parameters and data transmission frequency to maintain a good real-time interactive experience until the current equipment maintenance task is successfully completed.

[0049] S108. Based on the equipment data of the ring main unit, branch box and substation, a fault prediction model is trained to identify changes before and after a fault, and the equipment data of the ring main unit, branch box and substation are monitored in real time based on the trained fault prediction model to predict potential faults of the corresponding equipment.

[0050] The equipment historical operation data of the ring main unit, branch box and substation are obtained, and the equipment historical operation data includes temperature, voltage, current and distribution direction parameters. The equipment historical operation data is preprocessed and classified and stored according to the equipment type and parameter category; the time series analysis method is used to extract features of the equipment historical operation data to obtain key feature vectors reflecting the equipment operation status; the equipment historical operation data is annotated according to the equipment fault history records, and the training data set and test data set of the machine learning model are constructed; the fault prediction model is trained by the machine learning algorithm, and the fault prediction model parameters are optimized by the cross-validation method; the equipment operation data of the ring main unit, branch box and substation are collected online in real time, and the equipment operation data is cleaned, normalized and feature extracted to obtain real-time feature vectors; the real-time feature vectors are input into the trained fault prediction model to determine whether the equipment has potential fault risks. If the prediction result exceeds the preset threshold, a fault warning is triggered; a work order is generated according to the fault warning information, and the operation and maintenance personnel are notified to conduct on-site inspection and repair.

[0051] Specifically, for example, the temperature data of the past year was collected from 10 ring network cabinets, and 24 data points were collected from each device every day, totaling 87,600 temperature data. Similar collection and storage were performed for other parameters such as voltage and current. Through data cleaning, 2% of outliers and invalid data were removed. Then all data were normalized to the maximum and minimum values, and the values ​​were mapped to the range of 0 to 1. Using the time series analysis method, 10 key features of the temperature data of the ring network cabinet, such as mean, variance, kurtosis, etc., were extracted to form a 10-dimensional feature vector to characterize the operating status of the equipment. According to the fault history of the ring network cabinet in the past three years, the 87,600 collected temperature data points were marked as normal or faulty, of which fault data accounted for 8%. Then they were divided into training set and test set in a ratio of 8:2. The support vector machine algorithm was selected, and the optimal model parameters were obtained using grid search and 5-fold cross validation. The fault prediction accuracy of the model on the test set reached 95%. A set of temperature data is collected from a ring main unit in real time, and the corresponding 10-dimensional feature vector is extracted [25,78,...,14]. The feature vector is input into the trained SVM fault prediction model, and the fault probability output by the model is 75%, which exceeds the preset warning threshold of 60%, thus triggering a fault warning. The task management system automatically generates a fault work order and sends it to the nearest operation and maintenance personnel Xiao Wang to go to the site for inspection. After arriving at the site, Xiao Wang found that the ring main unit did have an overheating problem, and then replaced the faulty component to eliminate the hidden danger. The processing results are fed back to the task management system to complete the work order closure. 21,900 new operation data and 5 fault processing records generated by 10 ring main units in the past three months are collected to update and optimize the fault prediction model, and the performance of the model is further improved, so that the fault prediction accuracy is further increased from 95% to.

[0052] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A collaborative control method for ring main unit, branch box and substation based on AR, characterized in that: The method comprises: Obtain common faults and corresponding abnormal power data of ring main units, branch boxes and substations, extract fault features from the abnormal power data and associate them with corresponding fault types, establish a fault identification model for identifying corresponding fault types based on the power data of power facilities and calculating the corresponding fault occurrence probability; Determine maintenance tasks based on the output of the fault identification model, and assign maintenance tasks based on the skills of the current on-duty personnel. When the current on-duty personnel perform maintenance tasks, they obtain the equipment data of the corresponding equipment through the on-site AR system; Obtain the idle status of the current expert, and combine with the personnel management system to send the idle status of the current expert to the terminal of the current on-duty personnel or display it on the current on-duty public screen; Classify the maintenance tasks of ring main units, branch boxes and substations, including analyzing the urgency of maintenance tasks based on whether the maintenance tasks are assigned to the task list of the current on-duty personnel. The urgency of maintenance tasks includes high urgency and low urgency. The maintenance tasks are divided into corresponding processing priorities according to the importance and impact scope of the corresponding faulty equipment. For low-urgency maintenance tasks, if an expert request from the current on-duty personnel is detected, the device data of the corresponding device is synchronized to the expert AR system of the corresponding expert, and maintenance suggestions from the expert AR system are received in real time and synchronized to the on-site AR system. The maintenance suggestions are associated with the corresponding device in the on-site AR system and displayed in the display area of ​​the corresponding device; For high-urgency maintenance tasks, the current equipment data with timestamps is sent to the expert AR system of the corresponding expert, and the maintenance plan with timestamps is received from the expert AR system. When the high-urgency maintenance task is assigned to the current on-duty personnel, the maintenance plan is mapped to the on-site AR system; When multiple maintenance tasks are performed simultaneously, the priority of information transmission between the on-site AR system and the expert AR system is adjusted according to the processing priority of the maintenance tasks; Based on the equipment data of ring main units, branch boxes and substations, a fault prediction model is trained to identify changes before and after a fault. Based on the trained fault prediction model, the equipment data of ring main units, branch boxes and substations are monitored in real time to predict potential faults of the corresponding equipment. The maintenance tasks of the ring main unit, branch box and substation are classified, including analyzing the urgency of the maintenance task based on whether the maintenance task is assigned to the task list of the current on-duty personnel, the urgency of the maintenance task includes high urgency and low urgency, and according to the importance and impact scope of the corresponding faulty equipment, the maintenance task is divided into corresponding processing priorities, including: Obtain all pending maintenance tasks of the ring main unit, branch box and substation from the maintenance management system, extract the equipment type, fault type, fault level, occurrence time and impact range attributes of the task, construct a task feature vector, cluster the tasks using the KNN algorithm, and divide them into task sets of different urgency according to the clustering results; Track the task list of each on-duty personnel, obtain the planned completion time, actual completion progress and completion quality score information of the task, and analyze the correlation between task characteristics and completion performance; Model the topological structure of the ring main unit, branch box and substation, construct the power equipment connection diagram, and use the PageRank algorithm to calculate the importance weight of each equipment node; Taking the task urgency and equipment importance weight as two dimensions, the maintenance tasks are sorted in two dimensions, and a task priority matrix is ​​constructed. According to the position of the task in the matrix, different processing priorities are divided; For high-priority maintenance tasks, a centralized scheduling strategy is adopted to optimize the task allocation plan; For maintenance tasks of medium and low priority, a distributed scheduling strategy is adopted, and each on-duty personnel takes on tasks independently according to their own situation; The maintenance task processing process is monitored in real time, on-site operation data is collected through sensors and monitoring equipment, compliance checks are performed on operation behaviors, and abnormal situations that occur during the operation are warned and recorded in the task operation log.

2. The method according to claim 1, wherein: The method of obtaining common faults and corresponding abnormal power data of ring main units, branch boxes and substations, extracting fault features from the abnormal power data and associating corresponding fault types, and establishing a fault identification model for identifying corresponding fault types and calculating corresponding fault occurrence probabilities according to power data of power facilities includes: According to historical operation records and maintenance logs, the fault types of ring main units, branch boxes and substations and the corresponding abnormal power data characteristics are collected to establish a database of associated mapping relationships between fault types and abnormal power data; Preprocess the collected abnormal power data to remove noise and outliers, extract key parameters that can reflect fault characteristics, and perform feature engineering on the key parameters to obtain fault feature vectors with distinguishability; According to the associated mapping relationship database between fault types and abnormal power data, the extracted fault feature vectors are labeled to construct a training data set of fault feature vectors and fault types; Using machine learning algorithms to train a fault identification model, and optimizing hyperparameters of the fault identification model through cross-validation and grid search; Deploy the trained fault identification model to an online environment, receive power data from ring main units, branch boxes and substations in real time, extract and preprocess the power data, and input it into the fault identification model for prediction to identify potential fault types; For the identified fault type, the occurrence probability of the fault type under the current operating conditions is calculated based on historical data and expert knowledge, and a fault risk assessment report is generated; The operation data and fault records of new ring main units, branch boxes and substations are collected, and the fault identification model is retrained and optimized.

3. The method according to claim 1, wherein: The maintenance task is determined based on the output of the fault identification model, and the maintenance task is assigned based on the skills mastered by the current on-duty personnel. When the current on-duty personnel performs the maintenance task, the equipment data of the corresponding equipment is obtained through the on-site AR system, including: Based on the output results of the fault identification model, combined with the fault type, fault severity and fault location information, a corresponding equipment maintenance task list is generated; Construct a skill knowledge graph of on-duty personnel, which contains basic information, skill level, skill type and proficiency attributes of each on-duty personnel, as well as the mutual evaluation relationship between employees. Through the connection relationship between employee nodes and skill nodes in the knowledge graph, calculate the comprehensive ability index of each employee to master each skill; Match each task in the equipment maintenance task list with a skill node in the on-duty personnel skill knowledge graph, and calculate the similarity between the skills required for the task and the skills of the on-duty personnel; Generate a work schedule for each on-duty personnel according to the task allocation plan, wherein the work schedule includes task details, time arrangement and route planning information; When employees arrive at the site, they use AR devices to identify and locate target equipment; Analyze and judge whether the employee's operation process is standardized based on the real-time status data of the equipment and the employee's operation behavior. Give early warning prompts for potential dangerous operations or improper behaviors. Record the employee's operation process and results and generate operation reports. After the employee completes the work, the work report content is reviewed and analyzed, abnormal situations are marked and alarmed, and the skill knowledge map of the on-duty personnel is updated.

4. The method according to claim 1, wherein: The method of obtaining the idle state of the current expert, combining with the personnel management system, sending the idle state of the current expert to the terminal of the current on-duty personnel or displaying it on the current on-duty public screen includes: According to the expert's historical work log and shift schedule, an expert idle time prediction model is established. The expert idle time prediction model inputs the expert's historical work log data, including working time, location and type attributes, and outputs the probability of the expert being idle within a certain period of time; Based on the expert location information obtained in real time, it is determined whether the expert is in the work area or is idle. Based on the expert's physiological and behavioral data, determine whether the expert is in working condition through the expert's physiological state; Analyze the expert's call records and instant messaging records, extract call time, object and keyword information, and determine whether the expert is conducting work communication or remote guidance; Comprehensively judge the expert's idle time prediction model, real-time location information, physiological behavior data and communication record analysis results to obtain the expert's real-time idle status, and match it with the personnel management system to determine the expert's idle time in the scheduling plan; The expert's free time in the shift schedule is sent to the on-duty personnel's mobile terminal.

5. The method according to claim 1, wherein: For low-urgency maintenance tasks, if an expert request from the current on-duty personnel is detected, the device data of the corresponding device is synchronized to the expert AR system of the corresponding expert, and maintenance suggestions from the expert AR system are received in real time and synchronized to the on-site AR system. The maintenance suggestions are associated with the corresponding device in the on-site AR system and displayed in the display area of ​​the corresponding device, including: By analyzing the task priority matrix of task urgency and equipment importance weight, combined with the task response time requirements, the scope of the low-urgency maintenance task is determined; According to the scope of the low-urgency tasks, mark them with low-urgency tags in the maintenance management system; For the low-urgency task, when the on-duty personnel initiates an expert request, the location information, task information and equipment information of the on-duty personnel are obtained; Perform semantic search based on the keywords of equipment type and fault type in the equipment information to obtain a list of experts related to the task; Based on the idle status information of the experts in the expert list, matching the expert list with the idle expert list to determine the most matching and idle expert; Sending a request for assistance to the best matching and available expert; When the best-matched and available expert receives a request for assistance, the expert AR system loads the equipment parameter information including the three-dimensional model, specification parameters, historical maintenance records, and real-time operation data of the equipment; Spatially registering the device parameter information with the device real scene image and displaying them in the expert AR system; Determine the health status and fault cause of the equipment by analyzing the real-time parameter curve and alarm information of the equipment; Extract equipment maintenance plans and equipment maintenance recommendation reports from similar cases based on the health status and failure causes of the equipment; Perform semantic understanding and key information extraction on equipment maintenance recommendation reports; Push key information extracted from equipment maintenance recommendation reports to the on-site duty personnel AR system; After responding to the critical information, the on-duty personnel carry out maintenance work; Monitor the maintenance process and issue an early warning when any illegal operation is detected.

6. The method according to claim 1, wherein: For the high-urgency maintenance task, the current equipment data with timestamp is sent to the expert AR system of the corresponding expert, and the maintenance plan with timestamp is received from the expert AR system. When the high-urgency maintenance task is assigned to the current on-duty personnel, the maintenance plan is mapped to the on-site AR system, including: Extracting key operating parameters of the equipment according to the urgency of the high-urgency task; Preprocess the key operating parameters of the equipment to form time series data; Among them, data preprocessing includes adding clock-synchronized timestamps to key operating parameters of the equipment; Sending the time series data to an expert AR system; Integrate multi-source equipment data into a time series database; Visualize time series data; By comparing time series data with the device rated parameter thresholds, abnormal data points and alarm events are marked; Identify the equipment and measurement points that the expert is currently focusing on based on the expert's line of sight position and duration; Dynamically adjust the display order and detail level of visualization parameters, including adjusting the display of parameters based on expert attention; Display the parameters and the historical trend curves corresponding to the parameters and the associated measurement point data to assist experts in diagnosing the equipment status; Combine the real-time status data and equipment parameters of the equipment to conduct attribution analysis and severity assessment of equipment failures; Equipment parameters include equipment nameplate parameters, design drawings, operating procedures and fault manuals; Match standard handling procedures for failure modes; It also includes: configuring a future maintenance period for a high-urgency maintenance task, so as to assign the high-urgency maintenance task to a current on-duty personnel with matching skills in the future maintenance period; establishing an equipment prediction model in combination with historical equipment data, predicting future equipment data in the future maintenance period based on current equipment data, and sending the future equipment data to an expert AR system; The method configures a future maintenance period for the high-urgency maintenance task so as to allocate the high-urgency maintenance task to the current on-duty personnel who have the corresponding skill matching in the future maintenance period, specifically comprising: based on the attributes of each on-duty personnel in the shift schedule, according to the shift time and the shift duration, calculating the available time window of each on-duty personnel, and splitting the available time window into multiple time slots; based on the attributes of the high-urgency maintenance task, assigning different priorities according to the urgency of the high-urgency maintenance task, and sorting the high-urgency maintenance tasks according to priority and deadline; traversing the sorted high-urgency maintenance tasks, and querying the skill matching corresponding to the high-urgency maintenance task from the available time window of the on-duty personnel. The personnel time slots that match the level requirements and are free are calculated according to the attributes of the personnel time slots, and the first preset personnel time slots with the highest comprehensive scores are obtained to generate task-time slot-personnel triples; the task-time slot-personnel triples are grouped by task, and the optimal personnel allocation plan for each task is analyzed to obtain the optimal executor and the earliest start time for each task; the work arrangements and statistical indicators of the on-duty personnel are updated, the status and execution plan of the high-urgency maintenance tasks are updated, and task execution work orders are generated; the execution progress of the high-urgency maintenance tasks and the presence of personnel are tracked, and task execution and personnel scheduling are dynamically optimized according to actual conditions; and the performance of personnel after the completion of the high-urgency maintenance tasks is recorded; The method of establishing an equipment prediction model in combination with historical equipment data, predicting future equipment data in future maintenance periods based on current equipment data, and sending the future equipment data to an expert AR system specifically includes: training the equipment prediction model using a machine learning algorithm based on historical equipment data; obtaining the real-time operation data of the current equipment, converting it into the input format of the equipment prediction model through data preprocessing, and using the trained equipment prediction model to perform predictive reasoning to obtain predicted values ​​of various parameters of the equipment in the future period of time; converting the predicted future equipment data into a format according to the data interface requirements of the expert AR system, and sending it to the expert AR system through the data transmission interface to achieve a visual display of the future operation status of the equipment in AR; obtaining the evaluation results of the prediction results fed back by the experts through the AR system, and optimizing the equipment prediction model using the evaluation results to make the prediction results fit the actual operation status of the equipment; determining the best maintenance period for the equipment based on the equipment prediction results, sending the maintenance period suggestions to the operation and maintenance personnel through the expert AR system, and dynamically adjusting the maintenance period based on the feedback from the operation and maintenance personnel to obtain the optimal maintenance plan that takes into account both equipment operation and personnel scheduling.

7. The method according to claim 1, wherein: When multiple maintenance tasks are performed simultaneously, adjusting the priority of information transmission between the on-site AR system and the expert AR system according to the processing priority of the maintenance tasks includes: Based on the execution status data of high-urgency maintenance tasks and the location data of on-duty personnel; If the execution progress of the high-urgency maintenance task meets the preset early completion condition, it is determined whether the on-duty personnel are allowed to complete the task in advance and receive a new task; If the execution progress of the high-urgency maintenance task meets the preset delay condition, the impact of the task delay on subsequent related tasks is analyzed to determine whether it is necessary to trigger the task rescheduling process and reallocate tasks, and at the same time, emergency resource allocation is adopted for the related tasks of the high-urgency maintenance task; It also includes: if the on-duty personnel is following up on a low-urgency maintenance task and an expert request is detected, the information transmission between the on-site AR system and the corresponding expert AR system is adjusted to the highest level to maintain real-time communication between the current on-duty personnel and the corresponding expert, specifically including: obtaining the urgency information of the maintenance task currently being performed by the on-duty personnel to determine whether it is a low-urgency maintenance task; real-time monitoring to see whether there is an AR remote assistance request from an expert; if the current on-duty personnel is performing a low-urgency maintenance task and an AR remote assistance request initiated by an expert is detected, the priority of the information transmission channel between the on-duty personnel's on-site AR system and the expert AR system requesting assistance is set to the highest level; a real-time audio and video communication link is established between the on-duty personnel and the expert, and the equipment operating parameters and fault information data collected by the on-site AR system are transmitted to the expert AR system in real time; after the expert receives the on-site information through the AR system, it provides remote guidance to assist the on-duty personnel in completing the equipment maintenance task.

8. The method according to claim 1, wherein: The method includes training a fault prediction model for identifying changes before and after a fault based on the equipment data of the ring main unit, branch box, and substation, and monitoring the equipment data of the ring main unit, branch box, and substation in real time based on the trained fault prediction model to predict potential faults of the corresponding equipment, including: Obtaining historical equipment operation data of ring main units, branch boxes and substations, the equipment historical operation data including temperature, voltage, current and power distribution direction parameters, performing data preprocessing on the equipment historical operation data, and classifying and storing the data according to equipment type and parameter category; Using a time series analysis method, feature extraction is performed on the historical operation data of the equipment to obtain a key feature vector reflecting the operation status of the equipment; Based on the equipment failure history, the equipment historical operation data is annotated to build the training data set and test data set of the machine learning model; Train the fault prediction model through machine learning algorithms and optimize the fault prediction model parameters through cross-validation methods; Collect the equipment operation data of ring main cabinets, branch boxes and substations online in real time, clean, normalize and extract features of the equipment operation data to obtain real-time feature vectors; The real-time feature vector is input into the trained fault prediction model to determine whether the equipment has potential fault risks. If the prediction result exceeds the preset threshold, a fault warning is triggered. Generate a work order based on the fault warning information and notify the operation and maintenance personnel to conduct on-site inspection and repair.

Images