Regional remote intelligent patrol method and system for transformer substation with synergy of storage, calculation, iteration and transmission

Through the 'storage-calculation-idling-issuance' collaborative method, the resource limitation and efficiency problems of the remote intelligent inspection system of the substation in multi-site data processing and algorithm model update are solved, and efficient data storage and computing resource allocation are achieved, which improves the The accuracy of equipment defect identification and model update efficiency reduce operation and maintenance costs.

CN118944282APending Publication Date: 2024-11-12NARI NANJING CONTROL SYSTEM CO LTD +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410890628.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-04
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing remote intelligent inspection system of substations has resource limitations and efficiency problems in multi-site data processing and algorithm model update, resulting in high operation and maintenance costs, lagging fault handling, and the frequency and efficiency of algorithm model updates cannot be guaranteed.

Method used

Using the 'storage-calculation-idle-distribution' collaborative method, by building virtual storage devices and intelligent computing allocation, combining dual-model fusion algorithm and incremental sample transfer learning, the efficient utilization of data storage and computing resources is achieved, and quantitative evaluation standards are passed and parameter search for model adaptive iterative update.

Benefits of technology

It realizes efficient data storage and real-time image recognition of more than 100 substations, reduces operation and maintenance costs, improves equipment defect recognition accuracy and algorithm model update efficiency, supports adaptive model adjustment, and improves operation and maintenance quality of inspection system .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118944282A_ABST
    Figure CN118944282A_ABST
Patent Text Reader

Abstract

The invention discloses a storage-calculation-iteration-transmission coordinated regional remote intelligent patrol method and system for a transformer substation. The method comprises the following steps: carrying out distributed storage on daily data; an inspection system operator selects a required inspection point position according to in-station equipment and inspection requirements, and establishes and executes an inspection task; carrying out real-time image intelligent calculation; defect alarm information is uploaded, wherein defect images and alarm text information obtained through an intelligent analysis algorithm are uploaded to an in-station artificial intelligence platform according to a fixed format to be managed; carrying out one-key updating on the algorithm model; carrying out adaptive iteration on the model; and issuing and updating the high-quality model to realize automatic restart of the station end identification service. According to the method, a recognition algorithm based on double-model fusion and a discrimination algorithm for improving the twin network are designed, the recognition accuracy of equipment appearance defects, operation states and the like is improved, the recognition types are expanded, and the practicability of an artificial intelligence algorithm in the aspect of improving the inspection quality and efficiency of operation and maintenance personnel is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image processing and substation inspection, and in particular to a "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method and system. Background Art

[0002] For a long time, the daily work of substation equipment operation and maintenance has been carried out in the form of manual on-site operation, manual transcription, frequent on-site travel, etc. With the rapid growth of social electricity consumption, the number of substations has increased year by year. The State Grid Corporation has more than 40,000 substations, with 0.8 substations operated and maintained per capita and 3.9 substations monitored per capita. On average, each substation needs to be inspected hundreds of times a year, and it takes 4 people 4 hours to inspect each time. The operation and maintenance radius is large, the time taken for personnel to arrive at the station is long, and the planned maintenance is highly delayed. The degree of damage to faulty equipment is further increased, and even explosions occur, resulting in an increase in the equipment damage rate and the rate of personal injury and death, which has an adverse impact on the safe and stable operation of my country's power grid. Therefore, the traditional method of relying on "manual inspection and manual recording" can no longer meet the needs of the development of the modern power industry.

[0003] Existing substations have gradually adopted remote intelligent patrol systems to replace traditional manual patrols. The system can access patrol equipment such as cameras and robots within the station, issue patrol tasks, complete intelligent analysis, and upload defect alarms. The single-station remote intelligent patrol system can reduce the work cost of operation and maintenance personnel to a certain extent, but substations of various voltage levels are deployed with a separate patrol system. The software and server costs and supplier maintenance costs are relatively high, and data islands are easily generated.

[0004] Therefore, the regional remote intelligent patrol system that can access multi-site data and perform tasks at the same time is gradually applied to the management and monitoring of multiple medium and low voltage substations. The patrol system and server are deployed at the main station of the regional patrol. The substations under its jurisdiction are connected to the main station directly or through edge hosts. The main station realizes the control of each substation monitoring equipment, task issuance and execution, and report transmission. Compared with the single-station patrol system, it has the advantages of low operation and maintenance costs, fast fault handling, and simultaneous patrol of multiple substations. However, the current regional patrol system only uses a single patrol host and an analysis host. The number of accessible sites and server computing resources are limited. The artificial intelligence algorithm model carried by the system is iteratively updated by the system supplier. The frequency and efficiency of the update iteration cannot be guaranteed. The specific performance of the model is not known to the system operator, and it is even more impossible to decide whether to update it. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a regional remote intelligent inspection method and system for substations with "storage-computation-iteration-dispatch" coordination.

[0006] Technical solution: The "storage-computation-iteration-transmission" coordinated substation regional remote intelligent patrol method of the present invention comprises the following steps:

[0007] (1) Daily data distribution storage: The regional remote intelligent patrol system needs to access data from hundreds of substations at most, which requires a high storage space. The storage resources on each server are used to form a virtual storage device for unified call by the patrol host.

[0008] (2) Establishment and execution of patrol tasks: Patrol system operators select required patrol points, establish and execute patrol tasks based on station equipment and patrol requirements;

[0009] (3) Intelligent computing of real-time images: The inspection system can complete the intelligent allocation of computing services based on the current resource usage of each computing node. The target detection algorithm based on dual model fusion can recognize the real-time images taken during the inspection task to complete tasks such as equipment appearance defects and operating status identification. Based on the improved twin target detection algorithm, the difference between the real-time image and the reference image can be compared to complete the defect identification task.

[0010] (4) Upload defect alarm information: The defect image and alarm text information obtained by the intelligent analysis algorithm are uploaded to the on-site artificial intelligence platform in a fixed format for management;

[0011] (5) One-click update of algorithm model: Based on incremental sample transfer learning, efficient training of algorithm model is achieved;

[0012] (6) Model adaptive iteration: Based on field implementation experience, quantify model evaluation standards, establish a model update process, and implement model adaptive adjustment and parameter search until the training termination condition is reached;

[0013] (7) High-quality model distribution and update: The operator can distribute the best model in the current or historical training process with one click, so that the station-side recognition service can be automatically restarted.

[0014] Furthermore, the daily data in step (1) is divided into basic data services of the patrol system and patrol point configuration files. According to the number of on-site server deployments, the distributed storage includes:

[0015] When the number of deployed servers is less than 4, when the basic data services and patrol point configuration files change, incremental updates and backup transmissions are performed through the LAN;

[0016] When the number of deployed servers is greater than or equal to 4, a master-slave service cluster architecture is constructed to form a virtual storage space. The patrol system can add, delete, modify and query basic data services and configuration files through the virtual network address.

[0017] Furthermore, the step (2) comprises:

[0018] (2.1) Enter the patrol system task compilation interface and select the points to be patrolled according to the equipment and point information;

[0019] (2.2) Fill in the basic information of task name, task type, execution method, execution time, and execution period;

[0020] (2.3) After the new creation is completed, the patrol task can be completed immediately or periodically through the task management interface.

[0021] Furthermore, the step (3) includes: (3.1) the patrol system (patrol host) calls the process management interface of each server (hereinafter referred to as the analysis host in the present invention) equipped with intelligent reasoning hardware to check basic information such as available computing resources, supported algorithm model capabilities, and reasoning task progress;

[0022] (3.2) When it is found that the server is in an idle state, the patrol host task scheduling service sends an analysis request to the analysis host in a fixed format;

[0023] (3.3) If all servers are not idle, the server with the least computing resource pressure will be automatically selected to queue and wait for the algorithm to be called;

[0024] (3.4) Parsing the received request parameters;

[0025] (3.5) Dual-model reasoning: The preprocessed image is used to identify important equipment and coarse-grained defect features in the image through the improved first-order target detection model YOLOV5; the prediction box of important equipment is extracted according to the prediction results, and some pixels are expanded outward as the input image of the second-order model; the fine-grained defect features in the image are identified through the Swin Transformer model;

[0026] (3.6) Frame fusion processing: Through the weighted model fusion algorithm, the coarse and fine granularity recognition results are integrated to improve the algorithm detection rate;

[0027] (3.7) Design logical post-processing strategies to explore the symbiotic and exclusive relationships between equipment and defects and suppress the algorithm false detection rate;

[0028] (3.8) Using the identification result as the information body carried by the client to call the patrol host server interface;

[0029] (3.9) Parse the received request parameters;

[0030] (3.10) Reading the normal image of the point saved in the configuration file described in step 1 according to the point information;

[0031] (3.11) Task real-time image registration and alignment with normal image;

[0032] (3.12) The aligned image and the normal image are input into the twin object detection network to extract the features of both, and the feature vectors are fused;

[0033] (3.13) The fused features are input into the detection head to perceive the areas with differences between the two images;

[0034] (3.14) The client uses the identification result as the information body to call the patrol host server interface.

[0035] Furthermore, the step (4) comprises the following steps:

[0036] (4.1) After receiving the algorithm recognition results, the patrol host task scheduling service reports the defective points to the artificial intelligence platform. The reported information mainly includes the defect type, recognition frame coordinates, original image and recognition result image;

[0037] (4.2) Operators can log in to the AI ​​platform interface through a browser to view and manually modify, add, and delete defect identification boxes and types;

[0038] (4.3) After confirming that the location and type are correct, save the results in the xml file format required by the PASCAL VOC dataset and manage them uniformly on the training server.

[0039] Furthermore, the step (5) comprises the following steps:

[0040] (5.1) The user selects the category label file required for this model training through the artificial intelligence platform;

[0041] (5.2) Select the training and validation data sets required for this update process;

[0042] (5.3) Parse the annotation file and generate the required label documents in txt, json and other formats;

[0043] (5.4) Check whether the selected category labels are consistent with the pre-trained weight labels;

[0044] (5.5) If the labels are consistent, load the pre-trained model and perform transfer learning;

[0045] (5.6) If the labels are inconsistent, a new task branch is automatically generated for incremental learning;

[0046] (5.7) The loss value during training and the performance indicators on the validation set are transmitted to the backend service through the interface and displayed in real time on the frontend interface.

[0047] Furthermore, the step (6) comprises the following steps:

[0048] (6.1) Build a parameter retrieval library containing information such as pre-trained models, network structures, functional modules, and hyperparameters;

[0049] (6.2) When the training process described in step (5) reaches 70% of the set number of rounds, compare the difference in the evaluation criteria between this model and the field application model;

[0050] (6.3) If the updated model is better than the applied model or has similar performance to the optimal model, continue the training process to the end and compare the performance again;

[0051] (6.4) If the final performance is still better than the updated model, save the model file and end the training process;

[0052] (6.5) If the performance of the trained model is significantly different from that of the on-site model, stop training the model;

[0053] (6.6) Select different parameters from the parameter retrieval library and repeat the model update process of steps (6.2)-(6.5).

[0054] The "storage-computation-iteration-transmission" coordinated substation regional remote intelligent patrol system of the present invention comprises:

[0055] The data storage module is used for daily data distribution storage. It uses the storage resources on each server to form a virtual storage device for unified call by the patrol host.

[0056] The intelligent computing module is used in the patrol system to complete the intelligent allocation of computing services according to the current resource usage of each computing node. The target detection algorithm based on dual model fusion is used to identify the real-time images taken during the patrol task, and complete tasks such as equipment appearance defects and operating status identification. Based on the improved twin target detection algorithm, the difference between the real-time image and the reference image is compared to complete the defect identification task.

[0057] Defect alarm module, which is used to upload defect images and alarm text information obtained through intelligent analysis algorithms to the on-site artificial intelligence platform in a fixed format for management;

[0058] The adaptive iteration module is used to quantify the model evaluation criteria based on field implementation experience, establish a model update process, and implement model adaptive adjustment and parameter search iteration until the training termination condition is reached.

[0059] Furthermore, when the processor executes the program, the "storage-computation-iteration-transmission" coordinated regional remote intelligent inspection method for substations is implemented.

[0060] Furthermore, the computer program is designed to implement the "storage-computation-iteration-dispatch" coordinated substation regional remote intelligent inspection method when running.

[0061] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: in terms of daily data storage, the present invention constructs a virtual storage device to maximize the use and sharing of storage resources of each server, and can support the access of more than 100 substations of various voltage levels; in terms of real-time image intelligent computing, the patrol system can schedule artificial intelligence computing tasks according to the usage of computing resources, realize intelligent allocation of computing resources, design a recognition algorithm based on dual model fusion and an improved twin network discrimination algorithm, improve the recognition accuracy of equipment appearance defects, operating status, etc., expand the recognition types, and promote the practical implementation of artificial intelligence algorithms in improving the quality and efficiency of operation and maintenance personnel's patrols; in terms of algorithm model iteration, firstly, based on pre-training technology, an incremental sample transfer learning method is designed to realize efficient algorithm training, and secondly, through quantitative evaluation standards and parameter search, the model is automatically distributed. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a system architecture diagram of the present invention;

[0063] Figure 2 This is the architecture diagram of the YOLO metric learning algorithm based on feature layer transformation;

[0064] Figure 3 This is the technical architecture diagram of the model adaptive evaluation. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0066] In the regional remote intelligent inspection method for substations with the "storage-computation-iteration-dispatch" collaboration described in the present invention, operators can remotely monitor the equipment operation status of the substations under their jurisdiction at the regional master station, and replace manual on-site inspections with remote inspections. The present invention constructs a patrol system server cluster to expand the number of sites connected to the master station; automatically allocates computing resources according to the server status to realize task scheduling; based on the recognition algorithm of dual-model fusion and the improved twin network discrimination algorithm, improves the accuracy of equipment appearance recognition and expands the recognition types; builds an in-station artificial intelligence platform, based on incremental sample transfer learning and other technologies, promotes efficient algorithm training; designs model evaluation standards to realize model adaptive iteration. The overall system architecture diagram is shown below. Figure 1 shown.

[0067] A set of "storage-computation-iteration-transmission" coordinated substation regional remote intelligent patrol device is configured with a patrol host (named as server A) and four analysis hosts with intelligent identification functions (named as servers BE) as computing nodes. The "storage-computation-iteration-transmission" coordinated substation regional remote intelligent patrol method of the present invention specifically includes the following steps:

[0068] (1) Daily data distribution storage

[0069] The regional remote intelligent patrol system needs to access the data of hundreds of substations and has high requirements for storage space. The present invention constructs different distributed storage strategies according to the number of on-site servers.

[0070] When there are less than 4 deployed servers, when the basic data service and patrol point configuration files change, incremental updates and backup transmissions are performed through the LAN. The specific steps are as follows:

[0071] (1.1) All servers are located in the local area network and adopt a master-slave architecture. The master server is deployed with a full set of patrol system programs, and the slave server is only deployed with data services, intelligent recognition algorithms and other services;

[0072] (1.2) Operators can modify patrol points, cameras and other system data through the database configuration interface, complete meter, infrared recognition and other point configurations, and store the modified or added data on the main server;

[0073] (1.3) Click the Synchronize button on the page to send data services such as database and meter configuration files to the slave server via FTPS and automatically restart related mirror services.

[0074] When the number of deployed servers is greater than or equal to 4, a master-slave service cluster architecture is constructed to form a virtual storage space. The patrol system can realize the addition, deletion, modification and query of basic data services and configuration files through the virtual network address. The present invention now assumes that the data size generated by a single patrol is 3G, and the distributed storage data block size is set to 512MB, which needs to be stored in the virtual storage space composed of 5 servers. The specific steps are as follows:

[0075] (1.4) The client initiates a request to create a file to the primary node server A, and the request information includes the file name, size, etc.

[0076] (1.5) After receiving the request, Server A determines whether the file exists. If not, it creates the file.

[0077] (1.6) Calculate the number of blocks according to the file and the data block size set by rounding up. For 3G patrol data, the data block size of 512MB needs to be divided into 6 blocks;

[0078] (1.7) Server A determines the storage location of the data block based on the remaining disk space of server BE, and feeds the above information back to the client;

[0079] (1.8) After receiving the feedback information, the client cuts the data into blocks and writes the data blocks to the corresponding server node through Socket;

[0080] (1.9) When all data blocks are stored, a successful storage message is sent to the client. After receiving the message, the client closes the data stream and ends the distributed file storage.

[0081] The specific steps for distributed file reading are as follows:

[0082] (1.10) The client initiates a request to the primary node server A to read the file, including the file name, etc.

[0083] (1.11) Determine whether the file exists based on the file name. If it does not exist, return to the client that the file does not exist; if it does exist, read the file's metadata information, including the number of data blocks, the storage location of each data block, etc.;

[0084] (1.12) Server A returns the metadata to the client. The client reads the data blocks from the corresponding servers respectively and then combines the data blocks to form the entire file.

[0085] (2) Establishment and implementation of inspection tasks

[0086] During the debugging and deployment phase of the patrol system, the locations of the main equipment in the station, configuration files, etc. have been recorded in the system database in the form of patrol points, and saved in the server through the distributed storage strategy described in step 1. Operators need to select the required patrol points, establish patrol tasks, manage the execution plan of patrol tasks, and execute patrol tasks manually or on a scheduled basis according to the patrol standard requirements of the "five-way" equipment in the substation and the actual deployment of equipment in the station. The specific steps are as follows:

[0087] (2.1) Enter the "Patrol Plan Management" interface of the patrol system device, select the name of the substation for which the patrol plan needs to be managed on the left side of the interface, and the list on the right side of the interface will automatically switch to all the current patrol plans of the station;

[0088] (2.2) Select "New Patrol Plan" to create a new task. In the patrol point list tree, check or select all the required patrol points according to the four-level structure of level-interval-equipment-point;

[0089] (2.3) Fill in the basic information such as the patrol plan name and execution type. The execution types are divided into scheduled execution, periodic execution, interval execution and immediate execution. After filling in the information, click "Submit" to complete the creation of the patrol plan. The specific information for the task execution type is as follows:

[0090] The scheduled execution mode means that the patrol task is executed at a fixed time point;

[0091] The periodic execution mode means that after the plan is successfully released, the task will be executed once at the time filled in the task execution time, and then the task will be executed according to the daily, weekly or monthly cycle;

[0092] Interval execution mode refers to periodic execution at fixed intervals of days or hours;

[0093] Immediate execution mode means that the task is executed immediately after the plan is successfully released;

[0094] (2.4) The "Edit" option in the operation bar can be used to edit patrol points, basic information, and plan execution methods;

[0095] (2.5) Click the "Delete" option in the operation bar to delete the plan;

[0096] (2.6) The "Disable" option in the operation bar can cancel all settings of the plan;

[0097] (2.7) The "Execute Now" option in the operation bar can execute the plan once immediately;

[0098] (2.8) After the task is published and successfully executed, the operation status and results of the plan can be viewed in the "Patrol Task Management" interface.

[0099] (3) Real-time image intelligent computing

[0100] The patrol system can complete the intelligent allocation of computing services based on the current resource usage of each computing node: the target detection algorithm based on dual model fusion can identify the real-time images taken during the patrol mission to complete tasks such as equipment appearance defects and operating status identification; based on the improved twin network discrimination algorithm, the difference between the real-time image and the reference image can be compared to complete the comparison and discrimination task. The specific implementation steps are as follows:

[0101] (3.1) After the task is established, released and executed, a real-time screenshot to be analyzed will be generated for each patrol point, and a URL format link that can be directly read will be generated;

[0102] (3.2) The master node server A actively calls the resource query service on the computing node server BE, and the server returns the current heartbeat, GPU occupancy, usage rate, whether there is still inference, task queue status, etc.;

[0103] (3.3) Server A sorts the priority of the inference server call according to the query results. The sorting principle is as follows:

[0104] If there are idle servers, they will be called first;

[0105] If all servers are performing test tasks, the server with the shorter current queue process will be called first;

[0106] If some servers are currently performing model update training tasks, these servers will not be called for point reasoning in this task;

[0107] If servers BE are all performing model update training tasks, the update service of one server will be randomly terminated, and the server will perform the reasoning work of this patrol task;

[0108] After the termination, the model training task is automatically started and the model training is resumed from the breakpoint.

[0109] (3.4) Server A intelligently distributes each patrol screenshot generated by this task according to the sorting results, and calls the algorithm reasoning service of server BE through the client. The information contained in the interface call is as follows:

[0110] (3.4.1) If the identification type configured in the patrol point database is defect identification, the interface information includes the IP address, port, analysis point identifier, analysis type, and URL information of the image to be analyzed for feedback of the analysis results;

[0111] (3.4.2) If the identification type configured in the inspection point database is defect identification, the interface information includes the IP address, port, analysis point identifier, analysis type, identification reference map URL information, and image URL information for analysis result feedback;

[0112] (3.5) The server BE algorithm inference server will automatically call the defect appearance recognition and discrimination service according to the analysis type, and feedback the analysis results to server A through the client. The information sent includes whether there are abnormalities during the algorithm operation, analysis results, result image URL information, defect location in the image, result description, etc.;

[0113] (3.6) The inspection system matches the defect alarm strategy in the database based on the above information, and displays the result graph and defect type after reasoning in the "Task Details" interface. If there is an alarm, a real-time pop-up window will be displayed;

[0114] (3.7) The defect identification service specifically adopts a defect identification algorithm based on "intelligent identification-mechanism confirmation". Intelligent identification uses a two-level deep learning model to perform coarse and fine granularity identification of equipment and its appearance defects; mechanism confirmation is to double-confirm the identification results by mining the relationship between equipment and defects, thereby ensuring the accuracy and rationality of the output results. The specific steps are as follows:

[0115] (3.7.1) Construct a training set and a validation set containing all categories, and convert the annotated XML file into the format required by the two-level recognition model according to the recognition algorithm's requirements for input labels, as follows:

[0116] For the first-level YOLO model, the annotation file needs to be converted into a TXT file with the same prefix name. Each annotation box corresponds to a line in the TXT file, which consists of category and normalized annotation box information.

[0117] For the second-level Swin Transformer model, all annotation files need to be converted into a JSON file, including the image path, annotation box information, etc. of the entire dataset;

[0118] (3.7.2) Preprocess and enhance the images in the training set. The main methods include random scaling, cropping, brightness adjustment, GridMask masking, Mosaic splicing, etc.

[0119] (3.7.3) Perform adaptive anchor frame calculation and size scaling on the preprocessed image;

[0120] (3.7.4) The improved YOLOV5 is used as the first-level model to intelligently identify important equipment and coarse-grained defect features in the image. YOLOV5 consists of backbone network, neck network, loss function and other components, as follows:

[0121] The backbone network is a Darknet53 network that combines the Focus and CSP (Cross Stage Partial) structures to extract image features;

[0122] The neck network can fuse information from different feature layers. The present invention uses the structure of FPN (Feature Pyramid Network) combined with PAN (Pyramid Attention Network);

[0123] The loss function consists of Focal Loss and CIoU Loss. Focal Loss can alleviate the problem of class imbalance in target detection and improve the performance of the model. CIoU Loss takes into account the overlap between the detection box and the annotation box (Ground Truth), and also takes into account the scale information of the aspect ratio of the bounding box.

[0124] The improvements made to YOLOV5 in the present invention are as follows:

[0125] In view of the complex scenarios of power equipment, the large differences in the size of the same target, and the low detection rate of some small-sized defects, the SK Attention (Selective) structure is introduced. Multiple convolution kernels of different sizes are used for calculation to improve the expression ability of the model. At the same time, the channel weight obtained by each convolution kernel is dynamically calculated, and a channel attention mechanism is constructed to perform weighted fusion of the results.

[0126] Due to the excellent performance of the self-attention mechanism in Transformer in computer vision and natural language processing, the present invention introduces a recursive gating module (GnConv Block) in YOLOV5 to enhance the perception ability of the model. This module extends the second-order interaction in the self-attention mechanism to any order without introducing a large amount of additional calculations. In the present invention, all the residual structures in the YOLOV5 backbone network are replaced by recursive gating modules.

[0127] (3.7.5) Perform post-processing such as non-maximum suppression on the model output results, and only retain the boxes with the highest scores belonging to the same category in a certain area. The Cluster NMS algorithm is used in the present invention, which can utilize the matrix operation performance of the GPU to improve the NMS speed;

[0128] (3.7.6) The final output format of the model is an N*6 matrix, where N represents the number of prediction boxes in the image and 6 represents the output vector dimension, which includes the prediction box category, confidence, and coordinate position information respectively;

[0129] (3.7.7) If the prediction result is defect type, it is temporarily saved for the subsequent model fusion process; if the prediction result is equipment type, the predicted coordinate position is expanded outward by 30 pixels as the input of the second-level detection model;

[0130] (3.7.8) The second-level detection model is Swin Transformer, which performs fine-grained recognition of appearance defects. The model structure specifically includes:

[0131] Block feature generation: The input image is divided into multiple fixed-size blocks, and each block is mapped to an embedding vector through a linear transformation;

[0132] Shifted Window Attention: In each segmented block, a local attention mechanism is used to interact the block in the local window with other blocks. The local window is implemented by adding a circle of extra blocks around the block. In order to avoid overlapping calculations between blocks, the position of the local window will shift with the depth of the layer.

[0133] Transformer encoder: The block feature vector after the local attention mechanism is input into the encoder layer to extract high-level feature representation;

[0134] Transformer decoder: generates output predicted categories and boxes.

[0135] (3.7.9) For a single target detection model, the NMS series of technologies are often used to screen the prediction boxes. However, for a model fusion algorithm based on ensemble learning, NMS cannot fully utilize the information of all prediction boxes. Therefore, the present invention uses a box weighted fusion algorithm to construct the final prediction box. The specific steps are as follows:

[0136] The predicted boxes of each model are added to list B and sorted in descending order of confidence;

[0137] All the boxes belonging to the same target form a list L, and the result boxes after the set fusion form a list F;

[0138] Traverse the predicted boxes in list B and find the corresponding matching boxes in F according to the intersection-over-union ratio threshold;

[0139] If no match is found, add the predicted box as a new box to the lists L and F, and continue traversing the next predicted box;

[0140] After the traversal is completed, for each set in the list L, the coordinates and confidence of all boxes are fused and calculated. The fusion formula is shown in Formula 1:

[0141]

[0142] Among them, T is the number of frames to be fused in each set, c i is the confidence of the i-th box in the set, conf is the fused confidence, X1,2 and X1,2 i are the x-coordinate values ​​after fusion and the x-coordinate values ​​of the i-th box in the set, Y1,2 and Y1,2 i They are the y coordinate value after fusion and the y coordinate value of the i-th box in the set respectively;

[0143] The fused prediction frames are stored in the list F according to the corresponding positions and merged into one frame;

[0144] Recalculate the confidence score in F, as shown in formula 2:

[0145]

[0146] Among them, N is the number of fused models. If the number of boxes T in the set is small, it means that only a few models predict the target, which can reduce the confidence score.

[0147] (3.7.10) Mechanism confirmation is performed on the results of model intelligent identification. The specific steps of mechanism confirmation include:

[0148] Customize thresholds for different categories;

[0149] Mutually exclusive states retain states with high confidence, such as the separation and combination of the pressure plate will not exist at the same time. If they do exist, the category with high confidence will be retained;

[0150] Filter out the defect types that do not conform to the equipment mechanism. For example, if there is a large overlap between the oil stain on the ground and the respirator, the prediction box will be filtered out;

[0151] Merge prediction boxes of the same category that have inclusion relations;

[0152] The prediction boxes are merged according to the overlapping area. If the defects of the same category intersect with the same equipment parts, the defect prediction boxes are merged according to the maximum circumscribed rectangle.

[0153] (3.7.11) Arrange the output results into the parameter format in step 35 and transmit them through the interface in the form of a JSON string;

[0154] (3.8) The defect identification service specifically uses the YOLO metric learning algorithm based on feature layer transformation. The algorithm principle diagram is as follows: Figure 2 As shown, it is necessary to compare the real-time images taken during the patrol process with the normal images of the same location to identify the areas where the equipment operating status has changed significantly. The specific steps are as follows:

[0155] (3.8.1) Image registration and alignment. Since the two images are taken at different times, the mechanical error of the camera and other reasons may cause displacement deviation between the images. It is necessary to register the real-time image during the inspection (hereinafter referred to as the image to be detected) with respect to the normal image in the library (hereinafter referred to as the reference image). The present invention adopts a feature matching algorithm based on Homography. The specific steps include:

[0156] Use the FAST algorithm to calculate the areas with rapid grayscale changes in the two images and extract feature points;

[0157] Based on a set of key feature points, the BRIEF algorithm is used to create a binary feature descriptor;

[0158] Use the nearest neighbor matching algorithm to find the best matching point in another image for each feature point in the image to be detected;

[0159] Use the selected feature point matching pairs to calculate the Homography matrix, perform perspective transformation, map the points in the image to be detected to the reference image, and achieve image alignment;

[0160] (3.8.2) Image annotation. In the image to be tested, mark the areas that have significant changes compared to the reference image, and name the category "diff". If there is no obvious change between the two, no annotation is required. Changes caused by natural conditions such as weather, lighting, and shadows do not need to be marked;

[0161] (3.8.3) Model training. The present invention uses the YOLO model based on feature layer transformation. The main network adopts the YOLOV5-6.0 version, the model scale is L, and the improvements made include:

[0162] Image preprocessing adjustment: The identification task has clear requirements on the position of defects in the two images, so methods that easily cause position changes, such as random flipping, translation, angle change, and mosaic, are discarded;

[0163] The twin network structure design requires two images as input to the algorithm model, so the YOLOV5 model is changed to the traditional twin network structure to achieve parameter sharing in the front part of the feature extraction layer;

[0164] The feature layer transformation method subtracts the feature vectors of the two images to obtain the absolute value, multiplies it with the mask matrix of the annotated area, and then performs a weighted summation with the feature vector of the reference image. The calculation method is as shown in Formula 3:

[0165] f=w1*(f 待检测 -f 基准 )·MASK+w2*f 基准 (3)

[0166] Among them, f 待检测 and f 基准 It represents the feature maps of different scales extracted by the YOLOV5 model from two images. MASK represents the mask matrix. The matrix size is the same as the corresponding feature map size. In the matrix, the mapping area value of the area marked as diff in the feature map is 1, and the rest of the area is 0. w1 and w2 are the weights of the two feature vectors.

[0167] Improved loss function: In addition to the original target detection loss for box classification and regression, the present invention also adds metric learning loss, compares the differences between feature maps, and optimizes the metric space through parameter backpropagation, so that similar feature maps are close to each other and heterogeneous feature maps are far away from each other in this space;

[0168] (3.8.4) Model application. After the algorithm training is completed, it needs to be deployed in the remote intelligent patrol system for application. The specific steps are as follows:

[0169] Configure preset positions for the camera as patrol points of the patrol system;

[0170] Create a folder with the unique uuid of the patrol point in the database, and capture the real-time position of the camera as a reference image;

[0171] When the patrol task reaches the point, a real-time screenshot is taken as the image to be detected. Both images generate URL links for the analysis host to read;

[0172] The specific parameters of the interface sent are shown in step (3.4.2);

[0173] Register and align the image to be detected with the reference image;

[0174] Input the two images into the algorithm model to predict the different areas, confidence and feature map matching degree;

[0175] Filter by feature map degree. If the matching degree of the four size feature maps of the twin YOLO model is greater than 0.9, it is considered that there is no significant difference between the two images regardless of whether the detection model is output;

[0176] Filter the prediction boxes according to the confidence level, the confidence level set in the present invention is 0.3;

[0177] Merge the prediction boxes and output only one prediction box for each image;

[0178] (3.8.5) Arrange the output results into the parameter format in step 35 and transmit them through the interface in the form of a JSON string.

[0179] (4) Confirmation and upload of alarm information

[0180] The defect information found during the inspection process will be automatically uploaded or manually uploaded to the in-station artificial intelligence platform after confirmation. In order to avoid data jitter caused by the data flow to the main server A, the front-end interface and database involved in the in-station artificial intelligence platform are deployed on the slave server B. The specific steps include:

[0181] (4.1) After receiving the algorithm analysis results, the patrol system backend scheduling module server will save the results in the historical database and message queue, and the image with the prediction box will be stored in the analysis host, generating a URL link that can be directly used for the front-end interface to save, call and display;

[0182] (4.2) The task details interface of the inspection system can manually review the defect alarm information found in the task. The specific functions of the interface include:

[0183] (4.2.1) The operator clicks the manual review button, and the interface will display the current defect type, category, alarm level, threshold, etc.;

[0184] (4.2.2) The operator can change whether it is an alarm point, whether the alarm level is correct, the manual identification result, the cause of the abnormal identification, etc.;

[0185] (4.2.3) If the manual review confirms that it is a defect alarm, click Confirm to save, and the record will be entered into the alarm database and can be browsed in the alarm information query interface. At the same time, the original point map, identification map, and reference map are all saved to a fixed path;

[0186] (4.3) Click Upload on the alarm query interface. The scheduling module will push the defect information, alarm image files, description information, etc. to the artificial intelligence platform on the station to realize the structured upload of alarm data. The specific steps include:

[0187] (4.3.1) Structured data is sent to the specified TOPIC through the MQTT protocol. The data body mainly includes the plant name, voltage level, equipment, interval, point, time, etc. where the alarm occurred, the original image of the defect, the identification result image, the reference image path, etc., the specific location of the defect in the image, the alarm type, description, etc.;

[0188] (4.3.2) Upload the corresponding file to the artificial intelligence platform according to the alarm type. If the alarm type is defect recognition, the inspection system needs to confirm that both the recognition result image and the original image exist in the fixed directory before uploading. The result image must contain the main information such as the recognition box and defect type. If the alarm type is defect identification, the system needs to confirm whether the original image, result image, and reference image files all exist before uploading.

[0189] (4.3.3) The above files are transferred via FTPS protocol, and TLS / SSL encryption adopts implicit encryption;

[0190] (4.4) The operator can log in to the artificial intelligence platform interface through the browser. Using the platform data management function, the operator can view and manually modify the defect frame coordinate position and type. The specific steps are as follows:

[0191] (4.4.1) The left side of the interface lists the alarm information that the platform has received but has not yet confirmed into the training library;

[0192] (4.4.2) Clicking each record will enter the drawing mode, where you can modify the annotation information. When you hover the mouse in the image, you can see the coordinate information of the current mouse position in the image;

[0193] (4.4.3) If defects are missed in the image, you can create a labeling box by dragging;

[0194] (4.4.4) After determining the position of the rectangular frame, you can click to select or manually enter the category information corresponding to the current frame;

[0195] (4.4.5) After the box position and category are determined, click the Save option, and the annotation record will be listed and viewable in the image annotation information column on the right side of the interface;

[0196] (4.4.6) If you accidentally touch the left mouse button during the annotation process, press the right button to cancel the current drawing;

[0197] (4.4.7) If you want to delete the annotation box, select the annotation box with the left mouse button and delete it directly using the keyboard or by right-clicking and selecting the delete option, or you can click on the image annotation information bar to delete it;

[0198] (4.4.8) Simply drag the vertices of the rectangular frame to modify the frame position;

[0199] (4.4.9) Double-click a record in the image annotation information column to modify the category name corresponding to the box;

[0200] (4.5) After the operator confirms the defect type and location, select the server-side save path and click the Save button at the bottom of the annotation information bar. The platform can save the annotation results in the XML annotation file format specified by the PASCAL VOC dataset and save them together with the original image. Figure 1 The alarm is saved in the selected path, with the same label as the image file name, consisting of the substation name, alarm type, and defect discovery time;

[0201] (4.6) The operator also needs to click the option to select the storage option. If the storage option is selected, the record can be queried and modified again in the platform data set and historical storage record interface; if the option is not to store, it means that the image does not have the defects that need to be identified and does not need to be added to the current training set. The record can only be queried and modified in the historical storage record interface;

[0202] (4.7) Due to the large number of equipment and business scenarios in the substation, operators can modify the image annotation again through the historical storage record interface and re-select whether to store it;

[0203] (4.8) After refreshing or re-entering the interface, the processed records can no longer be viewed, annotated, or otherwise operated on the data governance interface.

[0204] (5) One-click update of algorithm model

[0205] When users need to update and iterate the algorithm model on the site, they can enter the artificial intelligence algorithm management platform, select the corresponding data set, model structure, pre-trained model, etc., and issue the algorithm training task.

[0206] (5.1) Filter the iterative training set. Since there are limited patrol points in the station, patrol images with the same angle and similar content will be generated. If a large number of similar images are input into the algorithm model, the model performance will be degraded. Therefore, the data after manual review and storage in step 4 must be screened first. The specific steps are as follows:

[0207] (5.1.1) If multiple images are stored at the same location on the same day, only one image will be retained;

[0208] (5.1.2) The retained image is similar to other images of the same point in the existing iteration library. If the similarity is higher than 0.95, it will not be placed in the iteration library.

[0209] (5.2) Users can view the model network structure and identifiable categories currently in use on the site through the artificial intelligence platform;

[0210] (5.3) The user can choose whether to add a new category. If the user chooses not to add a new category, the specific steps are as follows:

[0211] (5.3.1) Click on the interface to select the training set and validation set for this iteration;

[0212] (5.3.2) Select the algorithm model and pre-trained weights. Users can view the performance of the weights on the validation set;

[0213] (5.3.3) According to the algorithm training requirements, convert the XML image annotation files that have been reviewed and stored into the database into TXT, JSON and other format configuration files required for model training;

[0214] (5.3.4) Configure model training iterations, learning rate, batch size, image size and other hyperparameters;

[0215] (5.3.5) Before model training begins, the categories selected by the user will be checked to see if they are consistent with the categories included in the pre-trained model;

[0216] (5.3.6) After successful configuration and verification, the model will start transfer learning;

[0217] (5.3.7) If the interval from the last full data training is three months, or the performance score of a certain category drops by more than 5% after the model is updated, it indicates that the algorithm has been degraded in this iteration. The system will prompt you to update the full data. The number of rounds of full data update is 5-10. The optimal number of rounds is taken as the final application weight. The algorithm performance score formula is as follows:

[0218] Score=0.4×P+0.4×R+0.2×AP (4)

[0219] Among them, the value range of P, R and AP is [0, 1]. P represents the accuracy (Precison), 1-P represents the false detection rate for a certain type of target, R represents the recall rate (Recall), which is the detection rate of the algorithm for a certain type of target, and AP represents the average precision, which is the result of the comprehensive ranking of the recall rate and the detection rate.

[0220] (5.4) If the user wants to add a new category, the model will start incremental learning. The first five steps are the same as step 53. The other specific steps are as follows:

[0221] (5.4.1) The model freezes the main network layer, uses the full dataset, and only trains the model head network, so that the model has the ability to classify unknown categories;

[0222] (5.4.2) Fine-tune the parameters of the overall model on the iterative dataset;

[0223] (5.4.3) According to Formula 4, if performance degradation occurs on non-extended categories, the overall model parameters are fine-tuned on the full dataset as needed.

[0224] (6) Model Adaptive Iteration

[0225] The present invention proposes a model adaptive evaluation technology, such as Figure 3 As shown, it is possible to realize model adaptive adjustment and parameter search, establish a model update process, and quantify model evaluation standards.

[0226] The iterative process architecture consists of the priori layer, training layer, and testing layer. The priori layer includes the model version currently used on site and the evaluation results in the current test set; the testing layer performs reasoning on models at different stages and completes performance comparison to determine whether model structure adjustments are needed; the training layer is mainly used for model training, training parameter library establishment and retrieval, etc.

[0227] (6.1) Build a training parameter retrieval library. Before the iteration starts, the user selects the initialization parameters from the retrieval library, and the interface will also push the current default training parameters. The parameter retrieval library specifically includes:

[0228] (6.1.1) Pre-trained models: the current model in operation on the site, the historically optimal model of the network structure, the official pre-trained model based on the COCO dataset, etc.;

[0229] (6.1.2) Network structure: The platform supports iterative training of single-stage and dual-stage target detection models. The single-stage model supports YOLO V5-V8, YOLOair, YOLOX, etc., and the dual-stage model supports Cascade RCNN, Swin Transformer V1-V2, CO-DETR, etc.

[0230] (6.1.3) Structural modules: For single-stage adjustable modules, they include attention mechanism, Neck layer, training loss function, upsampling method, etc.; for dual-stage adjustable modules, they include data enhancement method, loss function, input image size, etc.;

[0231] (6.1.4) Training hyperparameters: number of training rounds, learning rate and its decay, number of warmup rounds, etc.

[0232] (6.1.5) Dataset: training and validation dataset;

[0233] (6.1.6) Training method: whether it is incremental training, etc.

[0234] (6.2) Performance evaluation of the model in operation. If the operational model has been evaluated on the current validation set, the evaluation index is called according to the MD5 value of the model; if no evaluation record is found, the performance of the model is tested at the prior layer, and the index is recorded according to Formula 4 as the basis for iteration, and is set as Score 站内 ;

[0235] (6.3) Model training. In the training layer of the system, model training begins according to the selected initialization parameters. The system will record the loss function value during the training process and the evaluation performance on the validation set after each round of training. The scheduling module will push the data in the algorithm training to the front-end interface for real-time display;

[0236] (6.4) Adaptive iteration. When the model is trained to 70% of the set number of rounds, compare the best evaluation index (set as Score 训练 ) and Score 站 According to the difference between the two, adaptive iteration will be performed according to the evaluation results. The specific steps include:

[0237] (6.4.1) If Score 训练 >Score 站内 , then continue training until the end, evaluate and record the best performance Score during the process 最优 , and save the weight file, and this model iteration ends;

[0238] (6.4.2) If Score 训练 <Score 站内, calculate the performance difference between the two models, as shown in Equation 5:

[0239]

[0240] If Score_diff>3%, it means that there is a gap between the model obtained by the current training parameters and methods and the model in the station. The current parameter training stops, returns to the training parameter library, updates the hyperparameters or network modules, and repeats step 64 until the Score is found. 最优 The corresponding model or iteration process is completed;

[0241] If Score_diff < 3%, train the model to the set number of rounds, evaluate and record the best score during the training process. 最优 , the performance comparison is as follows:

[0242] (1) If Score 最优 >Score 站内 , then the model is recorded as the optimal model, and this model iteration ends;

[0243] (2) If Score 最优 <Score 站内 , then return to the training parameter library, update the hyperparameters or modules, and repeat step 64 until the optimal model is found or the iterative process ends.

[0244] (6.5) Iteration record storage. The best evaluation performance during the iteration process will be written into the historical database for the front-end interface to call and display.

[0245] (7) Model distribution

[0246] After the iteration is completed, the model can be sent to the site with one click, and the algorithm application service can be restarted and updated.

[0247] (7.1) The AI ​​platform interface will display the various indicators of the optimal model on the validation set during the current iteration, the model structure, the optimal indicators of the validation set before this iteration, and other information;

[0248] (7.2) The weight provides a downloadable link, which can be copied via portable transmission media, or one-click distribution can be selected;

[0249] (7.3) Click the Send button. The system background will determine whether the network structure of the currently sent model is the same as that of the model on the site, and use MD5 to determine whether the inference code and container environment are compatible with the model. The specific operations after the judgment are as follows:

[0250] (7.3.1) If the model to be sent has the same structure as the model on the site or the inference code can be used universally, only the weighted model will be sent;

[0251] (7.3.2) If the container environment is compatible but the inference code has been changed, the weights and code need to be delivered;

[0252] (7.3.3) If the container environment and code are changed or the network structure is different from that on the site, the docker image package, code and weight need to be distributed.

[0253] (7.4) The platform backend will send an update request to the algorithm update interface service of each analysis host of the remote intelligent patrol system, including the path, algorithm type, unique algorithm number and other information of the content to be updated this time;

[0254] (7.5) After receiving the request, the analysis service will determine whether the sent file path is on the local machine. If not, the file needs to be transferred via FTP;

[0255] (7.6) After the file transfer is completed, an update completion request will be sent to the platform and patrol host;

[0256] (7.7) The algorithm update service will call the resource information interface of each server to check whether there are any inference tasks. If there are no inference tasks, the current algorithm inference service will be terminated, the updated content will be copied to the corresponding path, and the algorithm service will be restarted; if the current inference task has not yet ended, the algorithm update process will be executed after the task is completed;

[0257] (7.8) After the algorithm is successfully restarted, the unique algorithm number of the current algorithm version will be pushed to the patrol host and displayed on the patrol system service version information interface.

[0258] The "storage-computation-iteration-transmission" coordinated substation regional remote intelligent patrol system of the present invention comprises:

[0259] The data storage module is used for daily data distribution storage. It uses the storage resources on each server to form a virtual storage device for unified call by the patrol host.

[0260] The intelligent computing module is used in the patrol system to complete the intelligent allocation of computing services according to the current resource usage of each computing node. The target detection algorithm based on dual model fusion is used to identify the real-time images taken during the patrol task, and complete tasks such as equipment appearance defects and operating status identification. Based on the improved twin target detection algorithm, the difference between the real-time image and the reference image is compared to complete the defect identification task.

[0261] Defect alarm module, which is used to upload defect images and alarm text information obtained through intelligent analysis algorithms to the on-site artificial intelligence platform in a fixed format for management;

[0262] The adaptive iteration module is used to quantify the model evaluation criteria based on field implementation experience, establish a model update process, and implement model adaptive adjustment and parameter search iteration until the training termination condition is reached.

Claims

1. A regional remote intelligent inspection method for substations based on the "storage-computation-iteration-distribution" collaboration. It is characterized in that The steps include: (1) Daily data distribution storage: The regional remote intelligent patrol system needs to access data from hundreds of substations at most, which requires a high storage space. The storage resources on each server are used to form a virtual storage device for unified call by the patrol host. (2) Establishment and execution of patrol tasks: Patrol system operators select required patrol points, establish and execute patrol tasks based on station equipment and patrol requirements; (3) Real-time image intelligent computing: The patrol system can complete the intelligent allocation of computing services based on the current resource usage of each computing node; The target detection algorithm based on dual-model fusion is used to identify the real-time images taken during the inspection mission, and complete tasks such as equipment appearance defects and operating status identification; Based on the improved twin target detection algorithm, the difference between the real-time image and the reference image is compared to complete the defect identification task; (4) Upload defect alarm information: The defect image and alarm text information obtained by the intelligent analysis algorithm are uploaded to the on-site artificial intelligence platform in a fixed format for management; (5) One-click update of algorithm model: Based on incremental sample transfer learning, efficient training of algorithm model is achieved; (6) Model adaptive iteration: Based on field implementation experience, quantify model evaluation standards, establish a model update process, and implement model adaptive adjustment and parameter search until the training termination condition is reached; (7) High-quality model distribution and update: The operator can distribute the best model in the current or historical training process with one click, so that the station-side recognition service can be automatically restarted.

2. According to the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method of claim 1, It is characterized in that In step (1), the daily data is divided into the basic data service of the patrol system and the patrol point configuration file. According to the number of on-site server deployments, the distributed storage includes: When the number of deployed servers is less than 4, when the basic data services and patrol point configuration files change, incremental updates and backup transmissions are performed through the LAN; When the number of deployed servers is greater than or equal to 4, a master-slave service cluster architecture is constructed to form a virtual storage space. The patrol system can add, delete, modify and query basic data services and configuration files through the virtual network address.

3. According to the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method of claim 1, It is characterized in that The step (2) comprises: (2.1) Enter the patrol system task compilation interface and select the points to be patrolled according to the equipment and point information; (2.2) Fill in the basic information of task name, task type, execution method, execution time, and execution period; (2.3) After the new creation is completed, the patrol task can be completed immediately or periodically through the task management interface.

4. According to the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method of claim 1, It is characterized in that The step (3) comprises: (3.1) The patrol system (patrol host) calls the process management interface of each server equipped with intelligent reasoning hardware (hereinafter referred to as the analysis host in the present invention) to view basic information such as available computing resources, supported algorithm model capabilities, and reasoning task progress; (3.2) When it is found that the server is in an idle state, the patrol host task scheduling service sends an analysis request to the analysis host in a fixed format; (3.3) If all servers are not idle, the server with the least computing resource pressure will be automatically selected to queue and wait for the algorithm to be called; (3.4) Parsing the received request parameters; (3.5) Dual-model reasoning: The preprocessed image is used to identify important equipment and coarse-grained defect features in the image through the improved first-order target detection model YOLOV5; the prediction box of important equipment is extracted according to the prediction results, and some pixels are expanded outward as the input image of the second-order model; the fine-grained defect features in the image are identified through the Swin Transformer model; (3.6) Frame fusion processing: Through the weighted model fusion algorithm, the coarse and fine granularity recognition results are integrated to improve the algorithm detection rate; (3.7) Design logical post-processing strategies to explore the symbiotic and exclusive relationships between equipment and defects and suppress the algorithm false detection rate; (3.8) Using the identification result as the information body carried by the client to call the patrol host server interface; (3.9) Parse the received request parameters; (3.10) Reading the normal image of the point saved in the configuration file described in step 1 according to the point information; (3.11) Task real-time image registration and alignment with normal image; (3.12) The aligned image and the normal image are input into the twin object detection network to extract the features of both, and the feature vectors are fused; (3.13) The fused features are input into the detection head to perceive the areas with differences between the two images; (3.14) The client uses the identification result as the information body to call the patrol host server interface.

5. According to the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method of claim 1, It is characterized in that The step (4) comprises the following steps: (4.1) After receiving the algorithm recognition results, the patrol host task scheduling service reports the defective points to the artificial intelligence platform. The reported information mainly includes the defect type, recognition frame coordinates, original image and recognition result image; (4.2) Operators can log in to the AI ​​platform interface through a browser to view and manually modify, add, and delete defect identification boxes and types; (4.3) After confirming that the location and type are correct, save the results in the xml file format required by the PASCAL VOC dataset and manage them uniformly on the training server.

6. According to the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method of claim 1, It is characterized in that The step (5) comprises the following steps: (5.1) The user selects the category label file required for this model training through the artificial intelligence platform; (5.2) Select the training and validation data sets required for this update process; (5.3) Parse the annotation file and generate the required label documents in txt, json and other formats; (5.4) Check whether the selected category labels are consistent with the pre-trained weight labels; (5.5) If the labels are consistent, load the pre-trained model and perform transfer learning; (5.6) If the labels are inconsistent, a new task branch is automatically generated for incremental learning; (5.7) The loss value during training and the performance indicators on the validation set are transmitted to the backend service through the interface and displayed in real time on the frontend interface.

7. According to claim 1, the "storage-computation-iteration-transmission" coordinated substation regional remote intelligent inspection method, It is characterized in that The step (6) comprises the following steps: (6.1) Build a parameter retrieval library containing information such as pre-trained models, network structures, functional modules, and hyperparameters; (6.2) When the training process described in step (5) reaches 70% of the set number of rounds, compare the difference in the evaluation criteria between this model and the field application model; (6.3) If the updated model is better than the applied model or has similar performance to the optimal model, continue the training process to the end and compare the performance again; (6.4) If the final performance is still better than the updated model, save the model file and end the training process; (6.5) If the performance of the trained model is significantly different from that of the on-site model, stop training the model; (6.6) Select different parameters from the parameter retrieval library and repeat the model update process of steps (6.2)-(6.5).

8. A regional remote intelligent inspection system for substations with "storage-computation-iteration-transmission" coordination. It is characterized in that include: The data storage module is used for daily data distribution storage. It uses the storage resources on each server to form a virtual storage device for unified call by the patrol host. Intelligent computing module, used in patrol system, can complete intelligent allocation of computing services according to the current resource usage of each computing node; The target detection algorithm based on dual-model fusion is used to identify the real-time images taken during the inspection mission, and complete tasks such as equipment appearance defects and operating status identification; Based on the improved twin target detection algorithm, the difference between the real-time image and the reference image is compared to complete the defect identification task; Defect alarm module, which is used to upload defect images and alarm text information obtained through intelligent analysis algorithms to the on-site artificial intelligence platform in a fixed format for management; The adaptive iteration module is used to quantify the model evaluation criteria based on field implementation experience, establish a model update process, and implement model adaptive adjustment and parameter search iteration until the training termination condition is reached.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, It is characterized in that When the processor executes the program, it implements the "storage-computation-iteration-dispatch" coordinated substation regional remote intelligent inspection method according to any one of claims 1-7.

10. A storage medium storing a computer program, It is characterized in that The computer program is designed to implement the "storage-computation-iteration-distribution" coordinated substation regional remote intelligent inspection method according to any one of claims 1 to 7 when running.

Citation Information

Cited By

  • Automatic intelligent operation and maintenance method and system for power grid dispatching based on multi-modal large model

    CN119379263A

  • Automatic sample collection method based on artificial intelligence cloud edge collaboration

    CN120355964A