Agent-based automated power grid defect identification system

Abstract

The present application relates to the technical field of power grid defect identification, and in particular to an agent-based automated power grid defect identification system. The system comprises: acquiring real-time collected data, analyzing the real-time collected data, obtaining data analysis results, and labeling the real-time collected data according to the data analysis results to obtain label data; based on the D-S evidence theory, classifying the label data according to a preset historical defect map; using a defect identification agent to identify defects in the classified label data, and generating a defect report based on the identification results and feeding it back to the equipment maintenance system. The classified label data directly triggers the defect identification algorithm of the corresponding agent, shortening the communication link of the traditional cross-system data call, which is conducive to reducing the risk of defect escalation caused by handling delays. Automatically generate structured reports to provide a traceable decision basis for the equipment maintenance system and reduce manual transcription errors.

Description

Technical Field

[0001] The present application relates to the technical field of power grid defect identification, and in particular to an agent-based automated power grid defect identification system. Background Art

[0002] As smart grid construction enters the digital transformation phase, power equipment monitoring systems have developed a multi-dimensional perception network. Existing monitoring systems generate petabytes of structured monitoring data and unstructured maintenance records daily, covering hundreds of equipment status parameters, including voltage fluctuations, insulation degradation, and mechanical displacement. Traditional defect identification methods, which primarily rely on a combination of threshold alarms and manual inspections, have significant limitations when faced with the complexity of new power systems.

[0003] As smart grid equipment becomes increasingly intelligent, defect handling architectures based on multi-agent systems have become a mainstream technology in the industry. However, industry test data (see the IEEE PES 2024 white paper) shows that existing agent systems have significant shortcomings in key performance indicators: the collaborative error rate between defect identification and agent matching is as high as 31%, the resource conflict rate during concurrent multi-agent handling reaches 27%, and the communication delay for cross-regional agent collaboration exceeds 800ms. These issues result in approximately 15% of duplicate handling work orders each year, and 17% of secondary defects become system failures due to delayed handling. Therefore, accurately identifying defects remains an urgent problem. Summary of the Invention

[0004] This application provides an agent-based automated power grid defect identification system to solve the above problems.

[0005] In a first aspect, the present application provides an agent-based automated power grid defect identification system, the system comprising:

[0006] Acquire real-time collected data, analyze the real-time collected data to obtain data analysis results, and label the real-time collected data according to the data analysis results to obtain labeled data;

[0007] Based on DS evidence theory, the label data is classified and processed according to the preset historical defect map;

[0008] The defect recognition agent is used to identify defects in the classified label data, and a defect report is generated based on the identification results and fed back to the equipment maintenance system.

[0009] Through this solution, dynamic analysis and labeling of real-time collected data are performed to form structured feature vectors, providing standard data input for subsequent classification and improving the compatibility of heterogeneous data sources. The fusion decision-making mechanism based on DS evidence theory and historical defect maps can reduce the probability of misjudgment caused by traditional single threshold judgments and enhance the ability to identify complex defects in association. Through the directional matching of pre-data classification processing and dedicated defect identification agents, the probability of task overlap during multi-agent concurrency is reduced, and the risk of conflict caused by resource competition is reduced. The classified label data directly triggers the defect identification algorithm of the corresponding agent, shortening the communication link of traditional cross-system data calls, which is conducive to reducing the risk of defect escalation caused by disposal delays. Structured reports containing elements such as classification labels, evidence weights, and defect locations are automatically generated to provide a traceable decision basis for the equipment maintenance system and reduce manual transcription errors.

[0010] Optionally, analyzing the real-time collected data to obtain data analysis results includes:

[0011] Analyzing the real-time collected data to determine a collected image;

[0012] Analyzing the collected image to determine a number of pixel features;

[0013] Analyzing the plurality of pixel features to determine a device outline and a spatial distribution boundary;

[0014] The device outline and the spatial distribution boundary are determined as data analysis results.

[0015] This solution accurately separates image data from real-time data collected from multiple sources (such as video streams and mixed sensor data streams), eliminating non-image interference. It extracts key pixel features from the image and converts visual information into numerical features and geometric parameters, making image information measurable and eliminating reliance on manual experience. Contour and boundary data are encapsulated in a standardized format (such as a coordinate sequence) to directly serve the defect classification and disposal process.

[0016] Optionally, the step of labeling the real-time collected data according to the data analysis result to obtain labeled data includes:

[0017] determining a plurality of independent device entities in the data based on the device profile;

[0018] Obtaining a power grid GIS database, and determining physical connection relationships between independent device entities based on the power grid GIS database;

[0019] determining independent regions according to the spatial distribution boundaries;

[0020] According to the independent areas and the physical connection relationships, labeling processing is performed on the real-time collected data of a plurality of independent device bodies to obtain label data.

[0021] This solution matches device profiles with the GIS database, ensuring a one-to-one correspondence between data and physical devices, eliminating device mislabeling. Physical connection relationship labels enable agents to understand inter-device dependencies, reducing collaboration errors caused by topology misunderstandings. Independent area division limits global competition among multiple agents and reduces resource conflicts. Structured label data directly supports subsequent DS evidence theory classification and agent decision-making, shortening the defect identification and resolution process.

[0022] Optionally, the classification processing of the label data based on the DS evidence theory and a preset historical defect map includes:

[0023] Analyze the label data to determine whether any real-time collected data is multi-label data;

[0024] If so, the multi-label data is conflict resolved according to the DS evidence theory and the preset historical defect map to obtain normal label data;

[0025] The normal tag data is analyzed to determine tag attributes, and the normal tag data is classified according to the tag attributes.

[0026] This solution uses DS evidence theory to synthesize the confidence levels of multiple labels, eliminating redundant or conflicting labels. This prevents agents from misjudging defect types due to data conflicts and reduces collaborative errors. Data is categorized by label attributes (e.g., "Insulation" or "Mechanical"), making the classification results more relevant to real-world scenarios. Automated conflict resolution and classification replaces traditional manual review, shortening response time.

[0027] Optionally, the labeling process of the real-time collected data of the plurality of independent device bodies according to the independent areas and the physical connection relationship to obtain the label data includes:

[0028] Analyzing the device distribution in the independent area based on the physical connection relationship to determine associated devices in the independent area that belong to the same electrical circuit;

[0029] and merging the environmental tags of the associated devices;

[0030] After determining the associated devices, determining whether there are isolated devices based on the spatial distribution of devices in the independent area;

[0031] If there are isolated devices, a separate device-level tag is generated;

[0032] The label data is obtained according to the merging result and the device-level label.

[0033] Through the solution provided by this implementation, the set of devices in the same electrical circuit is identified through physical connection relationships, and the scattered device data is aggregated into loop-level tags to avoid logical fragmentation caused by isolated analysis of single device data. The need to repeatedly label the same environmental parameters (such as circuit temperature and humidity) for multiple devices in the same circuit is reduced, and the amount of raw data is compressed. Independent devices that do not form a closed loop are detected through the spatial distribution of devices to avoid their data being misclassified or omitted. Exclusive device-level tags are generated for isolated devices to ensure that their independent operating status is fully recorded and tracked. Through tag merging and classification, PB-level raw data is compressed into high-value tag data, reducing the amount of data and computing load processed by the agent.

[0034] Optionally, determining the physical connection relationship between independent device entities according to the power grid GIS library includes:

[0035] Parsing the power grid GIS database to determine the device ID and device coordinate information of the independent device;

[0036] Based on the device ID, a spatial index algorithm is used for rapid matching to obtain adjacent devices of the independent device;

[0037] Determining a device distance between an adjacent device and the independent device based on the device coordinate information;

[0038] Comparing the device spacing with a preset spacing threshold;

[0039] If the distance between the devices is less than the preset interval threshold, it is determined that there is a physical connection relationship.

[0040] This solution parses the power grid GIS database and extracts the mapping between device IDs and coordinate information, ensuring the unique and traceable geographic location of each individual device, providing fundamental data support for subsequent physical connection analysis. A spatial indexing algorithm quickly filters a list of neighboring devices for the target device, reducing the time complexity of neighboring device queries and adapting to the efficient processing requirements of petabyte-level power grid data. Actual distances are calculated based on device coordinates, and combined with preset interval thresholds, invalid associations between distant devices are eliminated, ensuring that connections meet actual engineering constraints. The electrical circuit ownership between device groups is clearly defined, providing a topological basis for subsequent defect identification and task allocation within the agent system.

[0041] Optionally, if the distance between the devices is less than a preset interval threshold, before determining whether a physical connection relationship exists, the method further includes:

[0042] Obtaining device information of the adjacent device;

[0043] Determining a voltage level according to the device information;

[0044] comparing the voltage level with a voltage level of the standalone device;

[0045] If the distance between the devices is less than a preset interval threshold and the voltage level is the same as the voltage level of the independent device, it is determined that there is a physical connection relationship.

[0046] This solution introduces voltage level matching to ensure that physical connections comply with grid safety regulations, avoiding the one-sidedness of relying solely on spatial distance. Dual-condition filtering reduces misjudgments of invalid connections.

[0047] Optionally, performing conflict resolution on the multi-label data according to the DS evidence theory and the preset historical defect map to obtain normal label data includes:

[0048] Determining the label confidence of the multi-label data according to the preset historical defect map;

[0049] Analyze multiple labels of the multi-label data to determine whether any two label attributes are mutually exclusive;

[0050] If so, based on the DS evidence theory and the label confidence, the two labels with mutually exclusive attributes are conflict-resolved to obtain normal label data.

[0051] This solution detects and eliminates mutually exclusive tags, ensuring that the tag data for the same device is consistent at the physical or logical level. This prevents misjudgments by the defect recognition agent due to label inconsistencies, thereby reducing the collaborative error rate. Label confidence based on historical defect maps reduces duplicate work orders caused by low-quality tags, improves the accuracy of defect reports, and prevents secondary defects from evolving into system failures. This solution mitigates the impact of label noise in petabyte-level data on the system, improving the robustness of defect recognition in new power systems (high complexity and multi-parameter coupling).

[0052] Optionally, based on the DS evidence theory and according to the label confidence, conflict resolution is performed on two labels with mutually exclusive attributes to obtain normal label data, including:

[0053] The label confidences of two labels with mutually exclusive attributes are sorted, and the label with low confidence is determined according to the sorting result. Based on the DS evidence theory, the label with low confidence is eliminated to obtain normal label data.

[0054] This solution directly resolves conflicts through confidence sorting, avoids complex calculations, and forcibly eliminates mutually exclusive labels, preventing the system from falling into logical confusion due to contradictory data.

[0055] Optionally, determining the label confidence of the multi-label data according to the preset historical defect map includes:

[0056] Determining the data source based on the real-time collected data;

[0057] Determining the confidence weight of the multi-label data according to the data source;

[0058] Determine the label confidence of the multi-label data according to the preset historical defect map and the confidence weight.

[0059] This solution, based on objective statistics from preset historical defect maps and real-time data sources, avoids subjective interference from human experience. The confidence calculation process is transparent (historical frequency × weight), making it easier for operations and maintenance personnel to trace the basis for their decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0061] Figure 1 A schematic diagram of an application scenario provided in one embodiment of the present application;

[0062] Figure 2 This is a flowchart of an agent-based automated power grid defect identification system provided in one embodiment of the present application. DETAILED DESCRIPTION

[0063] To make the purpose, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0064] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document, unless otherwise specified, generally indicates an "or" relationship between the related objects.

[0065] The embodiments of the present application are described in further detail below with reference to the accompanying drawings.

[0066] As smart grid equipment becomes increasingly intelligent, defect handling architectures based on multi-agent systems have become a mainstream technology in the industry. However, industry test data (see the IEEE PES 2024 white paper) shows that existing agent systems have significant shortcomings in key performance indicators: the collaborative error rate between defect identification and agent matching is as high as 31%, the resource conflict rate during concurrent multi-agent handling reaches 27%, and the communication delay for cross-regional agent collaboration exceeds 800ms. These issues result in approximately 15% of duplicate handling work orders each year, and 17% of secondary defects become system failures due to delayed handling. Therefore, accurately identifying defects remains an urgent problem.

[0067] Based on this, the present application provides an agent-based automated power grid defect identification system, which forms a structured feature vector through dynamic analysis and labeling of real-time collected data, provides standard data input for subsequent classification, and improves the compatibility of heterogeneous data sources. Based on the fusion decision-making mechanism of DS evidence theory and historical defect maps, it can reduce the probability of misjudgment caused by traditional single threshold judgment and enhance the ability to identify complex defects. Through the directional matching of pre-data classification processing and dedicated defect identification agents, the probability of task overlap during multi-agent concurrency is reduced, and the risk of conflict caused by resource competition is reduced. The classified label data directly triggers the defect identification algorithm of the corresponding agent, shortens the communication link of the traditional cross-system data call, and is conducive to reducing the risk of defect escalation caused by disposal delays. Automatically generate structured reports containing elements such as classification labels, evidence weights, and defect locations, provide traceable decision-making basis for equipment maintenance systems, and reduce manual transcription errors.

[0068] Figure 1 This is a schematic diagram of an application scenario provided by this application. When using an agent to perform automated power grid defect identification, the method provided by this application is applied.

[0069] Specifically, the method provided in this application is applied to any server, and the server interacts with multi-source devices and equipment maintenance systems, obtains real-time collected data through multi-source devices, analyzes the real-time collected data, uses defect identification agents to determine current power grid defects and generate defect reports to feed back to the equipment maintenance system.

[0070] For specific implementation methods, please refer to the following embodiments.

[0071] Figure 2 This is a flow chart of an agent-based automated power grid defect identification system provided in one embodiment of the present application. The system of this embodiment can be applied to the server in the above scenario. Figure 2 As shown, the system includes:

[0072] S201, acquiring real-time collected data, analyzing the real-time collected data, obtaining data analysis results, and labeling the real-time collected data based on the data analysis results to obtain labeled data;

[0073] Real-time collected data can be instantaneous monitoring data and operation and maintenance records generated during the operation of power grid equipment, including sensor values ​​(structured) and manual records (unstructured).

[0074] The data analysis results can be a set of data features extracted through statistical analysis and pattern matching, such as voltage fluctuation amplitude and insulation degradation trend.

[0075] Tag data can be a structured data unit with a semantic identifier attached, which is used to represent a specific abnormal type of the equipment status (such as "Tag: Mechanical Displacement_Overlimit_Level 3").

[0076] Specifically, the system uses multi-source equipment such as smart meters, SCADA systems, and vibration sensors to synchronously collect structured data such as voltage waveforms, insulation resistance values, and mechanical vibration spectra at a sampling frequency of 5ms to 10s. It also receives unstructured data such as text logs and image records input by maintenance personnel. The system then uses a sliding window algorithm to smooth and filter the time series data within the real-time data, and uses regular expressions to parse key fields in the unstructured text (such as device numbers and exception descriptions).

[0077] Statistical features (mean, variance, and kurtosis) are calculated for structured data, and keywords (such as "insulation degradation" and "exceeding displacement") are extracted from unstructured data to achieve dynamic analysis and obtain data analysis results. Then, label classification rules are set based on the label records of the company's historical time periods (for example: voltage fluctuation >±5% → label "voltage abnormality"; text containing "oil leakage" → label "seal failure"). Based on the data analysis results, label data in JSON format is generated, which includes the device ID, timestamp, and feature value.

[0078] S202. Based on the DS evidence theory and the preset historical defect map, the label data is classified and processed;

[0079] DS evidence theory can be a mathematical method to deal with uncertainty and multi-source information fusion, which quantifies the degree of support of different evidence for defect classification through basic probability distribution functions.

[0080] The preset historical defect map can be a structured knowledge base containing historical equipment defect cases, feature association rules and disposal strategies, which can be used to guide classification decisions.

[0081] Specifically, historical defect cases can be abstracted into a "defect type-characteristic parameter-treatment plan" triple map in advance, and then a historical defect map can be obtained to classify the label data.

[0082] Multiple features in the labeled data (such as the simultaneous presence of "voltage mutation +0.5kV" and "vibration frequency offset +15Hz") are used as independent evidence sources, and the joint confidence is calculated using the synthesis rules corresponding to the DS evidence theory. When different pieces of evidence point to conflicting defect types, the weighted average method is used to adjust the basic probability distribution, prioritizing the defect categories with higher confidence in the historical defect map, thereby achieving classification processing of the labeled data.

[0083] S203: Use the defect recognition agent to perform defect recognition on the classified label data, generate a defect report based on the recognition result, and feed it back to the equipment maintenance system.

[0084] A defect report can be a standardized output document containing elements such as defect type, location information, and disposal suggestions, which can be used to guide operation and maintenance operations.

[0085] The equipment maintenance system can be a back-end management platform that receives defect reports, generates repair work orders, and dispatches resources.

[0086] Defect identification agent can be a software agent with autonomous decision-making capabilities, dedicated to the identification and handling logic execution of specific types of defects.

[0087] Specifically, the classified labeled data triggers a specific agent instance (for example, "insulation degradation" data only activates the insulation detection agent), which then identifies the corresponding defect. The identification results can then be output according to the report template set in advance by the enterprise, and a defect report can be generated and fed back to the equipment maintenance system.

[0088] Through the solution provided by this embodiment, dynamic analysis and labeling of real-time collected data are performed to form structured feature vectors, providing standard data input for subsequent classification and improving the compatibility of heterogeneous data sources. The fusion decision-making mechanism based on DS evidence theory and historical defect maps can reduce the probability of misjudgment caused by traditional single threshold judgment and enhance the ability to identify complex defects. Through the directional matching of pre-data classification processing and dedicated defect identification agents, the probability of task overlap during multi-agent concurrency is reduced, and the risk of conflict caused by resource competition is reduced. The classified label data directly triggers the defect identification algorithm of the corresponding agent, shortening the communication link of traditional cross-system data calls, which is conducive to reducing the risk of defect escalation caused by disposal delays. Automatically generate structured reports containing elements such as classification labels, evidence weights, and defect locations, providing a traceable decision basis for the equipment maintenance system and reducing manual transcription errors.

[0089] In some embodiments, real-time collected data is parsed to determine a collected image; the collected image is analyzed to determine a number of pixel features; the number of pixel features are analyzed to determine the device outline and spatial distribution boundaries; the device outline and spatial distribution boundaries are determined as data analysis results.

[0090] The captured images can be visual data of the equipment's operating status obtained through a camera or infrared device, which is used to analyze physical morphological anomalies (such as mechanical displacement and component deformation).

[0091] Pixel features can be quantitative properties extracted from image pixels, including color distribution, gradient strength, texture complexity, etc.

[0092] The device outline can be a closed boundary of the device body area in the image, represented by a sequence of polygon vertex coordinates.

[0093] The spatial distribution boundary can be the actual location range of the equipment components in the physical space (such as the safety distance threshold between the high-voltage side and the low-voltage side of the transformer).

[0094] Specifically, image data packets are separated from the real-time mixed data stream (including sensor values, text logs, and image / video streams) through protocol parsing (such as MQTT / HTTP). The raw image data packets are then de-noised (median filtering), format standardized (converted to PNG format), and timestamp aligned (synchronized with sensor data) to obtain the captured image. The OpenCV library is then used to convert the color image to grayscale. The Local Binary Pattern (LBP) algorithm is used to identify texture features in the captured image and obtain several pixel features. The Sobel operator is then used to calculate the image gradient magnitude of these pixel features to extract device outlines (such as the transformer bushing boundary).

[0095] Based on the camera calibration parameters (intrinsic parameter matrix, distortion coefficient) and extrinsic parameters (installation height, pitch angle), a mapping relationship between the image pixel coordinate system and the actual physical coordinate system is established. The spatial boundaries identified in the historical period are obtained as templates. According to the mapping relationship and the template, the spatial boundary distribution in the pixel features is identified, and the device outline and spatial distribution boundaries are determined as the data analysis results.

[0096] The solution provided in this embodiment accurately separates image data from real-time data collected from multiple sources (such as video streams and mixed sensor data streams), eliminating non-image interference. Key pixel features are extracted from the image, converting visual information into numerical features and geometric parameters, making image information measurable and eliminating reliance on manual experience. Contour and boundary data are encapsulated into a standardized format (such as a coordinate sequence) to directly serve the defect classification and handling process.

[0097] In some embodiments, several independent device entities in the data are determined based on the device profile; a power grid GIS library is obtained, and the physical connection relationship between the independent device entities is determined based on the power grid GIS library; independent areas are determined based on the spatial distribution boundaries; and the real-time collected data of several independent device entities are labeled based on the independent areas and physical connection relationships to obtain labeled data.

[0098] The power grid GIS database can be a database that stores the geographic coordinates, topological connection relationships and attribute information (such as voltage level and equipment type) of power equipment.

[0099] The physical connection relationship can be the energy transmission path formed between devices through physical lines (such as cables and busbars), which can reflect the topology of the power grid.

[0100] An independent area can be a logical unit divided according to the spatial distribution or function of the device, which is used to limit the scope of agent collaboration and reduce communication delay.

[0101] Specifically, based on the equipment outlines in the real-time data analysis results, a spatial clustering algorithm (such as DBSCAN) is used to distinguish the independent entities of multiple devices in the image or point cloud data. The equipment topology data stored in the power grid GIS (geographic information system) library is called, and the real-time equipment entities are spatially associated with the equipment nodes in the GIS library (such as transformers and circuit breakers) through a graph matching algorithm to determine the physical connection relationship between the independent equipment entities. Using the spatial distribution boundaries in the data analysis results, the power grid equipment is divided into several independent areas (such as substation area A and transmission corridor area B) through spatial grid division (such as 50m×50m grid) or regional aggregation algorithm (such as Voronoi diagram). Combined with the output of the previous steps, structured labels are added to the real-time data of each equipment entity, for example:

[0102] Device-level tags:

[0103] Device ID: unique identifier (e.g. T1_SubstationA);

[0104] Area: independent area name (such as "Substation Area A");

[0105] Connected device: IDs of upstream and downstream devices in a physical connection relationship (e.g., [CB2, L3]).

[0106] Data-level labels:

[0107] Collection timestamp: the time when the data was generated (e.g., 2024-05-01T14:30:00Z);

[0108] Space boundary: device outline vertex coordinates (e.g. [[116.404,39.915],[116.405,39.916]]);

[0109] Associated defect types: Potential defects (such as "insulation degradation" and "mechanical displacement") pre-associated based on historical defect maps.

[0110] The solution provided in this embodiment matches device profiles with the GIS database, ensuring a one-to-one correspondence between data and physical devices, eliminating device mislabeling. Physical connection relationship labels enable agents to understand inter-device dependencies, reducing collaboration errors caused by topology misunderstandings. Independent area division limits global competition among multiple agents and reduces resource conflicts. Structured label data directly supports subsequent DS evidence theory classification and agent decision-making, shortening the defect identification and resolution process.

[0111] In some embodiments, the label data is analyzed to determine whether any real-time collected data belongs to multi-label data; if so, the multi-label data is conflict resolved according to the DS evidence theory and the preset historical defect map to obtain normal label data; the normal label data is analyzed to determine the label attributes, and the normal label data is classified and processed according to the label attributes.

[0112] Multi-tag data can be real-time data of the same device body that is simultaneously associated with multiple tags (such as "insulation degradation" and "mechanical displacement") due to multi-dimensional analysis (such as voltage fluctuations and temperature anomalies).

[0113] Normal label data can be the unique or high-confidence label data retained after conflict resolution, which can be directly used for subsequent classification and agent decision-making. In the specific implementation, data that does not belong to multiple labels does not need to undergo conflict resolution and can also be considered normal label data.

[0114] Label attributes can be dimensions that describe defect characteristics, including type, urgency, scope of impact, etc., which are used to guide classification rules and handling priorities.

[0115] Specifically, each record in the label data is traversed to check whether it contains multi-label data (i.e., real-time data from the same device entity is associated with multiple label attributes, such as the simultaneous presence of "insulation degradation" and "mechanical displacement"). If multi-label data exists, a basic probability assignment (BPA) is constructed: based on the historical defect probabilities of similar devices in a preset historical defect map, an initial confidence level is assigned to each label (such as "insulation degradation") (e.g., the confidence level for "insulation degradation" is 0.7). The Dempster synthesis rule is then calculated for the multiple pieces of evidence (e.g., Label A, Label B) in the multi-label data, and the normal label data after conflict resolution is output.

[0116] Normal tag data is analyzed (for example, the name may reflect the defect type). The data is then classified based on the evolution rate of similar defects in historical defect maps (e.g., "urgent," "high," "medium," and "low"). Based on physical connectivity, the likelihood of the defect spreading to adjacent equipment (e.g., "local" or "regional") is determined to determine the tag attributes of the current normal tag data. The normal tag data is then categorized based on the tag attributes. For example, data can be grouped into pre-set classification pools by defect type (e.g., "insulation defect" or "mechanical defect"). Priority tags can also be generated based on urgency and impact range (e.g., "insulation degradation - urgent - regional").

[0117] The solution provided in this example uses DS evidence theory to synthesize the confidence levels of multiple labels, eliminating redundant or conflicting labels. This prevents the agent from misjudging defect types due to data conflicts and reduces collaborative errors. Data is categorized by label attributes (such as "insulation" and "mechanical") to make the classification results more relevant to real-world scenarios. Automated conflict resolution and classification replaces traditional manual review, shortening response time.

[0118] In some embodiments, based on the physical connection relationship, the device distribution in the independent area is analyzed to determine the associated devices in the independent area that belong to the same electrical circuit; and the environmental tags of the associated devices are merged; after determining the associated devices, it is determined whether there are isolated devices based on the spatial distribution of the devices in the independent area; if there are isolated devices, a device-level tag is generated separately; and label data is obtained based on the merged results and the device-level tag.

[0119] Associated devices can be devices that are directly or indirectly connected in the same closed electrical circuit, and their operating states affect each other.

[0120] Environmental tags can be tags that describe the environmental status of the device, including numerical types (such as temperature and humidity) and status types (such as "electromagnetic interference exceeds the standard").

[0121] An isolated device can be an independently operated device in an independent area that does not form a closed loop with other devices, and its data needs to be processed separately.

[0122] Device-level tags can be independent tags generated for isolated devices, containing their unique status parameters (such as vibration and insulation resistance).

[0123] Specifically, an adjacency matrix is ​​constructed between devices based on physical connections (such as cable and busbar connections). Using a graph traversal algorithm (such as depth-first search), the set of devices that form a closed loop is identified and marked as associated devices in the same electrical circuit.

[0124] The maximum or average value of the environmental parameters of the associated devices (for example, "Temperature = 45°C" takes the highest value) is used to merge the environmental tags of the associated devices. If any device triggers an abnormal state (such as "Humidity exceeds the limit"), the tags are merged into a loop-level tag (such as "Loop_001: Humidity Abnormal").

[0125] After determining the associated devices, count the remaining devices not included in the associated device list. Check whether the remaining devices have independent operation capabilities (such as energy storage devices and independent sensors). If they do, they are considered isolated devices.

[0126] Device-level tags are tags generated only for the independent monitoring data of isolated devices. They combine loop-level tags (merged associated device data) and device-level tags (isolated device data) into unified tag data.

[0127] Through the solution provided by this implementation, the set of devices in the same electrical circuit is identified through physical connection relationships, and the scattered device data is aggregated into loop-level tags to avoid logical fragmentation caused by isolated analysis of single device data. The need to repeatedly label the same environmental parameters (such as circuit temperature and humidity) for multiple devices in the same circuit is reduced, and the amount of raw data is compressed. Independent devices that do not form a closed loop are detected through the spatial distribution of devices to avoid their data being misclassified or omitted. Exclusive device-level tags are generated for isolated devices to ensure that their independent operating status is fully recorded and tracked. Through tag merging and classification, PB-level raw data is compressed into high-value tag data, reducing the amount of data and computing load processed by the agent.

[0128] In some embodiments, the power grid GIS library is parsed to determine the device ID and device coordinate information of the independent device; based on the device ID, a spatial index algorithm is used for rapid matching to obtain the adjacent devices of the independent device; based on the device coordinate information, the device spacing between the adjacent devices and the independent device is determined; the device spacing is compared with a preset spacing threshold; if the device spacing is less than the preset spacing threshold, it is determined that a physical connection relationship exists.

[0129] The device ID can be a code that uniquely identifies an independent device in the power grid (such as "Substation A_Circuit Breaker_001").

[0130] The device coordinate information may be the coordinates of the device in space, and may be determined according to an established coordinate system.

[0131] A neighboring device may be a device that is adjacent to the standalone device.

[0132] The device spacing may be the actual physical distance between devices, calculated based on coordinates, in meters (m).

[0133] The preset interval threshold may be the maximum physical connection distance allowed between devices defined in the grid specification, and the threshold may vary depending on the device type (eg, cable ≤ 20 m, busbar ≤ 5 m).

[0134] Specifically, extract the device attribute table from the GIS database, obtain the device ID (unique identifier) ​​and device coordinate information (latitude and longitude or 3D coordinates) of each individual device, and convert the coordinate information into a unified geographic coordinate system (such as WGS-84 or a local engineering coordinate system).

[0135] Based on the device ID, an R-tree or quadtree is used to create a spatial index for device coordinates, accelerating neighboring device searches. With the current device as the center, the spatial index quickly searches for candidate devices within a certain range. Candidate devices are screened for those with the same voltage level or device type as the current device to identify the neighboring devices of the independent device.

[0136] Use the Euclidean distance formula (2D or 3D) to calculate the straight-line distance between devices to obtain the device spacing between adjacent devices and independent devices.

[0137] The maximum allowable physical connection distance defined by the device type and grid specifications (e.g., a cable connection threshold of ≤ 20m) is used as the preset separation threshold. The device separation is compared with the preset separation threshold. If the device separation is less than the preset separation threshold, a physical connection is determined.

[0138] The solution provided in this embodiment parses the power grid GIS database and extracts the mapping relationship between device ID and coordinate information, ensuring that the geographic location of each independent device is unique and traceable, providing basic data support for subsequent physical connection analysis. A spatial indexing algorithm is used to quickly filter out a list of neighboring devices for the target device, reducing the time complexity of neighboring device queries and adapting to the efficient processing requirements of petabyte-level power grid data. Actual distances are calculated based on device coordinates, and combined with preset interval thresholds, invalid associations between distant devices are eliminated to ensure that connection relationships meet actual engineering constraints. The electrical circuit ownership between device groups is clarified, providing a topological basis for subsequent defect identification and task allocation in the agent system.

[0139] In some embodiments, device information of an adjacent device is obtained; a voltage level is determined based on the device information; the voltage level is compared with the voltage level of an independent device; if the device spacing is less than a preset interval threshold and the voltage level is the same as the voltage level of the independent device, it is determined that a physical connection relationship exists.

[0140] Device information can be attribute data of power grid devices, including device ID, type, voltage level, port configuration, installation time, etc.

[0141] The voltage level can be the classification of the operating voltage range allowed by the design of grid equipment (such as 10kV, 35kV, 110kV), which is mandatorily defined by the grid code.

[0142] Specifically, the system extracts device information (including device ID, device type, voltage level, port configuration, and other attributes) for adjacent devices from the power grid GIS database or device attribute table. The "voltage level" field is extracted from the device information, ignoring other attributes to obtain the corresponding voltage level. This voltage level is then compared with the voltage level of the independent device. If the device spacing is less than a preset spacing threshold and the voltage level is the same as that of the independent device, a physical connection is determined.

[0143] The solution provided in this embodiment introduces voltage level matching to ensure that physical connections comply with grid safety regulations, avoiding the one-sidedness of relying solely on spatial distance. Dual-condition filtering reduces misjudgments of invalid connections.

[0144] In some embodiments, the label confidence of the multi-label data is determined based on a preset historical defect map; several labels of the multi-label data are analyzed to determine whether any two label attributes are mutually exclusive; if so, based on the DS evidence theory, according to the label confidence, the two labels with mutually exclusive attributes are conflict-resolved to obtain normal label data.

[0145] Specifically, based on the preset historical defect map, the matching frequency of identification labels and real defects in historical data is determined, and the label confidence of multi-label data is calculated based on the matching frequency (for example, if the accuracy of the overload label in history is 90%, the confidence is 0.9).

[0146] Because logically contradictory tags cannot exist simultaneously for the same device at the same time (e.g., "overload" and "low load"), mutually exclusive tag pairs (e.g., overload ↔ low load, insulation normal ↔ insulation degraded) are predefined based on the physical rules of the power grid and the operating logic of the device. The tag set is traversed to check for predefined mutually exclusive tag pairs, thereby determining whether any two tag attributes are mutually exclusive.

[0147] If it exists, then based on the DS evidence theory and the label confidence, the conflict resolution solution provided in the above embodiment is used to obtain normal label data.

[0148] The solution provided in this embodiment detects and eliminates mutually exclusive tags, ensuring that the tag data for the same device is consistent at the physical or logical level. This prevents misjudgments by the defect identification agent due to label inconsistencies, thereby reducing the collaborative error rate. Label confidence based on historical defect maps reduces duplicate work orders caused by low-quality tags, improves the accuracy of defect reports, and prevents secondary defects from evolving into system failures. This mitigates the impact of label noise in petabyte-level data on the system, improving the robustness of defect identification in new power systems (high complexity and multi-parameter coupling).

[0149] In some embodiments, the label confidences of two labels with mutually exclusive attributes are sorted, and based on the sorting results, the label with low confidence is determined, and based on the DS evidence theory, the label with low confidence is eliminated to obtain normal label data.

[0150] Specifically, we directly compare the confidence values ​​of two labels and sort them in descending order (e.g., overload (0.9) > low load (0.3)). We then select the label with the lower confidence value from the sorted results (e.g., "low load" has a confidence value of 0.3). Based on the DS evidence theory, we eliminate the low-confidence label to obtain the normal label data.

[0151] Through the solution provided in this embodiment, conflicts are resolved directly through confidence sorting, complex calculations are avoided, mutually exclusive labels are forcibly eliminated, and the system is prevented from falling into logical confusion due to contradictory data.

[0152] In some embodiments, the data source is determined based on the real-time collected data; the confidence weight of the multi-label data is determined based on the data source; and the label confidence of the multi-label data is determined based on the preset historical defect map and confidence weight.

[0153] The data source can be a combination of the location where the real-time data is collected and the type of equipment (such as "high-voltage equipment area - temperature monitoring").

[0154] Specifically, real-time collected data includes monitoring data (such as voltage fluctuations and temperature sensor readings) and its metadata (such as sensor ID, device location, and collection time). The real-time collected data is parsed to determine the device location (such as "Substation A - Transformer Group 1"), sensor type (such as "infrared thermometer"), and collection timestamp contained in the metadata to determine the data source.

[0155] Based on the label accuracy statistics of different data sources in the preset historical defect map (for example, the historical accuracy of the label "high-voltage equipment area-temperature monitoring" is 85%), a confidence weight table is generated. Then, based on the current data source label (such as "high-voltage equipment area-temperature monitoring"), the corresponding confidence weight is extracted from the weight table to determine the label confidence of the multi-label data.

[0156] The solution provided in this embodiment, based on objective statistics from preset historical defect maps and real-time data sources, avoids subjective interference from human experience. The confidence calculation process is transparent (historical frequency × weight), making it easier for operations and maintenance personnel to trace the basis for their decisions.

Claims

1. An agent-based automated power grid defect identification system, characterized in that: include: Acquire real-time collected data, analyze the real-time collected data, and determine a collected image; Analyzing the collected image to obtain a data analysis result, and performing labeling processing on the real-time collected data according to the data analysis result to obtain labeled data; Based on DS evidence theory, the label data is classified and processed according to the preset historical defect map; Use the defect recognition agent to identify defects in the classified label data, generate defect reports based on the identification results, and feed them back to the equipment maintenance system; Based on the DS evidence theory, the label data is classified and processed according to the preset historical defect map, including: Analyze the label data to determine whether any real-time collected data is multi-label data; If so, the multi-label data is conflict resolved according to the DS evidence theory and the preset historical defect map to obtain normal label data; Analyzing the normal tag data, determining tag attributes, and classifying the normal tag data according to the tag attributes; The conflict resolution of the multi-label data based on the DS evidence theory and the preset historical defect map to obtain normal label data includes: Determining the label confidence of the multi-label data according to the preset historical defect map; Analyze multiple labels of the multi-label data to determine whether any two label attributes are mutually exclusive; If so, based on the DS evidence theory and the label confidence, the two labels with mutually exclusive attributes are conflict-resolved to obtain normal label data; Based on the DS evidence theory and according to the label confidence, the conflict between two labels with mutually exclusive attributes is resolved to obtain normal label data, including: The label confidences of two labels with mutually exclusive attributes are sorted, and the label with low confidence is determined according to the sorting result. Based on the DS evidence theory, the label with low confidence is eliminated to obtain normal label data.

2. The system according to claim 1, wherein: Analyzing the collected images to obtain data analysis results includes: Analyzing the collected image to determine a number of pixel features; Analyzing the plurality of pixel features to determine a device outline and a spatial distribution boundary; The device outline and the spatial distribution boundary are determined as data analysis results.

3. The system according to claim 2, characterized in that The step of labeling the real-time collected data according to the data analysis result to obtain label data includes: determining a plurality of independent device entities in the data based on the device profile; Obtaining a power grid GIS database, and determining physical connection relationships between independent device entities based on the power grid GIS database; determining independent regions according to the spatial distribution boundaries; According to the independent areas and the physical connection relationships, labeling processing is performed on the real-time collected data of a plurality of independent device bodies to obtain label data.

4. The system according to claim 3, characterized in that The tagging process of the real-time collected data of the plurality of independent device bodies according to the independent areas and the physical connection relationship to obtain the tag data includes: Analyzing the device distribution in the independent area based on the physical connection relationship to determine associated devices in the independent area that belong to the same electrical circuit; and merging the environmental tags of the associated devices; After determining the associated devices, determining whether there are isolated devices based on the spatial distribution of devices in the independent area; If there are isolated devices, a separate device-level tag is generated; The label data is obtained according to the merging result and the device-level label.

5. The system according to claim 3, wherein: Determining the physical connection relationship between independent device entities based on the power grid GIS database includes: Parsing the power grid GIS database to determine the device ID and device coordinate information of the independent device; Based on the device ID, a spatial index algorithm is used for rapid matching to obtain adjacent devices of the independent device; Determining a device distance between an adjacent device and the independent device based on the device coordinate information; Comparing the device spacing with a preset spacing threshold; If the distance between the devices is less than the preset interval threshold, it is determined that there is a physical connection relationship.

6. The system according to claim 5, characterized in that If the distance between the devices is less than the preset interval threshold, before determining whether a physical connection relationship exists, the method further includes: Obtaining device information of the adjacent device; Determining a voltage level according to the device information; comparing the voltage level with a voltage level of the standalone device; If the distance between the devices is less than a preset interval threshold and the voltage level is the same as the voltage level of the independent device, it is determined that there is a physical connection relationship.

7. The system according to claim 1, wherein: Determining the label confidence of the multi-label data according to the preset historical defect map includes: Determining the data source based on the real-time collected data; Determining the confidence weight of the multi-label data according to the data source; Determine the label confidence of the multi-label data according to the preset historical defect map and the confidence weight.

Images