Measurement asset fault active troubleshooting method based on knowledge graph and rule engine

Abstract

The invention discloses a measuring asset fault active troubleshooting method based on a knowledge graph and a rule engine, and relates to the related technical field of data processing, and the method comprises the steps: collecting the image data of a measuring device, carrying out the analysis processing of the image data through a visual model, and extracting the structural features, including the defect type, position coordinates, confidence, and the subordinate component; mapping the defect nodes into defect nodes in the knowledge graph, and establishing an association relationship between the defect nodes and associated nodes; based on a dynamic rule engine, generating a graded work order by combining the confidence coefficient and the risk level; and dynamically updating the knowledge graph through graded work order execution result feedback. The technical problems that in the prior art, manual inspection efficiency is low, the fault omission ratio is high, static rule adaptability is poor, defect risks lack dynamic evaluation, and operation and maintenance feedback cannot be subjected to closed-loop optimization are solved. The technical effects of improving the troubleshooting efficiency, realizing risk grading early warning, dynamically optimizing the rule engine and forming a knowledge graph self-learning closed loop are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a metering asset fault active troubleshooting method based on a knowledge graph and a rule engine. BACKGROUND

[0002] With the deepening of the construction of smart grid, as an important part of the power system, the running state of the metering equipment directly affects the accuracy and reliability of power metering. However, due to the long-term exposure of metering equipment in complex environment, it is easy to be affected by environmental erosion, mechanical wear, electrical aging and other factors, resulting in frequent occurrence of various faults. The traditional fault troubleshooting is low in efficiency, high in cost and high in missed detection rate, which is difficult to meet the needs of intelligent and lean management of modern power system. Computer vision technology can automatically identify defect features such as cracks, rust and deformation from equipment images, but the existing method lacks comprehensive evaluation of defect associated risks, and the rule engine is difficult to adapt to the dynamic changes of equipment running state, resulting in insufficient accuracy and timeliness of fault troubleshooting; in addition, there are still deficiencies in the collaborative application of knowledge graph and rule engine, especially in the aspects of dynamic risk grading and work order generation, and a closed-loop optimized fault troubleshooting mechanism has not been formed, making it difficult to realize dynamic updating and rule optimization of knowledge graph.

[0003] In the related art at present, there are technical problems such as low efficiency of artificial inspection, high fault missed detection rate, poor adaptability of static rules, lack of dynamic evaluation of defect risks, and non-closed-loop optimization of operation and maintenance feedback. SUMMARY

[0004] The present application provides a metering asset fault active troubleshooting method based on a knowledge graph and a rule engine, which solves the technical problems of low efficiency of artificial inspection, high fault missed detection rate, poor adaptability of static rules, lack of dynamic evaluation of defect risks and non-closed-loop optimization of operation and maintenance feedback in the prior art, and achieves the technical effects of improving fault troubleshooting efficiency, realizing risk grading and early warning, dynamically optimizing the rule engine and forming a knowledge graph self-learning closed loop.

[0005] The present application provides a metering asset fault active troubleshooting method based on a knowledge graph and a rule engine, which includes: collecting image data of metering equipment, analyzing and processing the image data through a visual model, extracting structured features, the structured features including defect types, position coordinates, confidence levels and belonging components; mapping the extracted structured features to defect nodes in the knowledge graph, establishing an association relationship between the defect nodes and associated nodes, wherein the risk nodes in the associated nodes are bound to preset risk levels; generating a graded work order based on a dynamic rule engine, combining confidence levels and risk levels according to the knowledge graph nodes and the association relationship; and dynamically updating the knowledge graph through the feedback of the execution result of the graded work order.

[0006] In a possible implementation, the metering asset fault active troubleshooting method based on the knowledge graph and the rule engine further performs the following processing: replacing a backbone network of a YOLOv5 model architecture with a ConvNeXt-Tiny; adding a fault feature enhancement layer to the YOLOv5 model architecture, inputting standard positions and size parameters of key components of a device; weighting a key area through a feature attention mechanism, and amplifying a feature signal of a size defect, to build a visual model framework structure; and training and converging the visual model framework structure through a training data set, to obtain the visual model.

[0007] In a possible implementation, the metering asset fault active troubleshooting method based on the knowledge graph and the rule engine further performs the following processing: inputting the image data into the visual model, outputting a basic feature map through a ConvNeXt-Tiny backbone network; generating three scale fusion feature maps through a feature pyramid; inputting the fusion feature map into a fault feature enhancement layer, injecting a standard position coordinate of a key component of a device into the feature map, and weighting a feature of a door lock and a terminal area through a feature attention mechanism, to amplify a preset small size feature weight; and when an overlapping degree of a detection frame and the key component reaches a threshold value, outputting a structured feature vector by a detection head.

[0008] In a possible implementation, the metering asset fault active troubleshooting method based on the knowledge graph and the rule engine further performs the following processing: adopting a self-adaptive light compensation algorithm to perform light compensation on input image data, to eliminate reflection interference.

[0009] In a possible implementation, the metering asset fault active troubleshooting method based on the knowledge graph and the rule engine further performs the following processing: performing multi-source data standardization processing on visual extraction features and device operation parameters; performing multi-modal feature correlation on standardized data, and mapping to four types of nodes, the four types of nodes including a defect node, a cause node, a risk node, and a strategy node; calculating correlation strength based on a graph neural network, connecting the four types of nodes, setting weights between nodes, and constructing the knowledge graph.

[0010] In a possible implementation, the metering asset fault active troubleshooting method based on the knowledge graph and the rule engine further performs the following processing: integrating the visual extraction features and the device operation parameters, wherein the visual extraction features include a defect type, a position coordinate, and a belonging component, and the device operation parameters include a temperature, a current fluctuation, and a running duration, to obtain multi-source data; converting text features in the multi-source data into 128-dimensional vectors through BERT word embedding; and mapping numerical features in the multi-source data into 128-dimensional vectors through a full connection layer after normalization.

[0011] In a possible implementation, the method further performs the following processing: starting from the defect node, tracing the cause node, the risk node and the strategy node along the associated edges to generate a complete path; multiplying all edge weights on the complete path to calculate a path weight; performing path screening based on the path weight to determine a main treatment scheme and a backup scheme; calculating an urgency based on the main treatment scheme and the backup scheme according to the confidence and the risk level in combination with a diffusion probability; and triggering work order grading to generate the graded work order according to matching of the urgency and a rule threshold.

[0012] In a possible implementation, the method further performs the following processing: starting from the defect node, tracing the cause node, the risk node and the strategy node along the associated edges to generate a complete path, including: an industrial constraint mechanism; when a deviation between the fault position coordinates and a standard position of the component is greater than 15%, multiplying all edge weights of the current path by a weight reduction factor 0.8; for a power core fault feature, multiplying the weight of an associated path including a connection terminal defect and a current fluctuation by a reinforcement factor 1.2; and outputting a compatibility of the strategy node and the equipment model in front of the path.

[0013] In a possible implementation, the method further performs the following processing: when a feedback of the graded work order execution result is a work order success, reinforcing the associated edge weight; when the work order fails and a root cause is not covered, adding a graph node and a relationship chain, and removing an inefficient edge below a weight threshold based on edge weights in the knowledge graph.

[0014] The method can collect image data of the metering equipment, analyze and process the image data through a visual model, extract structured features including a defect type, position coordinates, a confidence and a component to which the defect belongs, map the features into a defect node in a knowledge graph, establish an association relationship between the defect node and associated nodes, generate a graded work order based on a dynamic rule engine in combination with the confidence and the risk level, and dynamically update the knowledge graph through a feedback of the graded work order execution result. The method solves technical problems in the prior art, such as low artificial inspection efficiency, high fault missing rate, poor adaptability of static rules, lack of dynamic evaluation of defect risks and failure of operation and maintenance feedback to form a closed-loop optimization, and achieves technical effects of improving fault troubleshooting efficiency, realizing risk grading early warning, dynamically optimizing the rule engine and forming a knowledge graph self-learning closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. In the present application, flowcharts are used to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the foregoing or the following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously according to needs. Meanwhile, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.

[0016] Figure 1 A flowchart of the method for actively troubleshooting metering asset faults based on a knowledge graph and a rule engine provided by the embodiments of the present application is shown.

[0017] Figure 2 A flowchart of the method for extracting structured features in the method for actively troubleshooting metering asset faults based on a knowledge graph and a rule engine provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0018] The foregoing description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the embodiments of the present application can be implemented according to the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described as follows.

[0019] In order to make the purposes, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings, and the described embodiments should not be regarded as limiting the present application. All other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application.

[0020] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict, and the term "first\second" referred to only distinguishes similar objects, and does not represent a specific order of the objects. The terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, product or server including a series of steps does not have to be limited to those steps clearly listed, but can include other steps not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art of the technology to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application.

[0021] The embodiments of the present application provide a method for actively troubleshooting metering asset faults based on a knowledge graph and a rule engine, as shown inFigure 1 The method comprises the following steps: In step S100, image data of the metering equipment is collected, and the image data is analyzed and processed by a visual model to extract structured features including defect types, position coordinates, confidence levels, and component types.

[0022] Step S100 further comprises the following steps: S101, replacing the backbone network of the YOLOv5 model architecture with a ConvNeXt-Tiny network; S102, adding a fault feature enhancement layer in the YOLOv5 model architecture, and inputting standard positions and size parameters of key components of the equipment; S103, weighting key regions by a feature attention mechanism and amplifying feature signals of size defects to build a visual model framework structure; and S104, training and converging the visual model framework structure by a training data set to obtain the visual model.

[0023] Preferably, the original CSPDarknet backbone network of YOLOv5 is replaced with a ConvNeXt-Tiny network. ConvNeXt is a convolution-based neural network architecture that draws on the design ideas of Transformers and performs well in image classification and feature extraction tasks. Using ConvNeXt-Tiny as the backbone network can more efficiently extract the basic features of metering equipment images. A fault feature enhancement layer is added to the YOLOv5 model architecture. This layer receives standard positions and size parameters of key components of the equipment as prior knowledge, so that the model can more specifically focus on key parts of the equipment prone to failure, enhance the extraction ability of features in these regions, and improve the accuracy of fault detection. A feature attention mechanism is introduced to weight process key regions of the equipment (such as wiring terminals and display screens), so that the model pays more attention to these important regions during feature processing. Meanwhile, for size-related defects (such as component deformation and looseness), the feature signals are amplified, and a visual model framework structure is built to identify such faults. A training data set containing various fault conditions of metering equipment is used to train and converge the visual model framework, that is, by continuously iterating and optimizing model parameters, the model gradually converges during the training process, and finally a visual model that can accurately detect faults of metering equipment is obtained, which can extract structured features such as defect types, position coordinates, confidence levels, and component types from images.

[0024] Preferably, through the camera, unmanned aerial vehicle, inspection robot and other equipment, the image of the metering equipment in the actual running environment is shot, which may include the whole equipment, local details, pictures of different angles, etc., to form an original image data set, which may contain normal equipment or equipment with faults and defects, such as damaged shell, loose wiring, abnormal display screen, etc.; the collected original image is input into the customized visual model, and the image is subjected to feature extraction, target detection and defect identification through a deep learning algorithm, including locating the key components of the metering equipment in the image, such as the terminal, shell, display screen, button, etc.; identifying whether the components have abnormalities, such as defects such as damage, deformation, stains, misplacement, etc.; calculating the reliability of the identification result, i.e. the confidence; then extracting structured features with fixed formats from the image analysis results, including defect type, position coordinates, confidence and belonging component, wherein the defect type is the identified fault category, such as "damaged shell", "loose wiring", "display screen black screen", "fuzzy mark", etc.; the position coordinates are the specific positions of the defects marked in the image (usually represented by pixel coordinates or relative coordinates, such as "x1=120, y1=80, x2=180, y2=140"), which are used to accurately locate the fault position; the confidence is the credibility of the model to the identification result, for example, 0.92 indicates that the model has 92% confidence to determine that the defect is "loose wiring"; and the belonging component indicates which specific part of the metering equipment the defect belongs to, such as "belongs to the incoming terminal", "belongs to the liquid crystal display screen", "belongs to the shell upper cover", etc.

[0025] Further, as shown in Figure 2 S100 further comprises the step S110 of inputting the image data into the visual model and outputting basic feature maps through a ConvNeXt-Tiny backbone network; the step S120 of generating three scale fusion feature maps through a feature pyramid; the step S130 of inputting the fusion feature maps into a fault feature enhancement layer, injecting standard position coordinates of key components of the equipment into the feature maps, and then weighting and amplifying the preset small size feature weight through a feature attention mechanism; and the step S140 of outputting a structured feature vector by the detection head when the overlap degree of the detection box and the key components reaches a threshold.

[0026] Preferably, after the collected metering equipment image is input into the visual model, the image is first processed through the ConvNeXt-Tiny backbone network, multi-scale convolution, normalization and other operations are performed on the image, and bottom-level visual features such as edges, textures and component outlines are gradually extracted from the original pixels, and finally a basic feature map containing overall feature information of the equipment is output; the defects of the metering equipment may have size differences (such as small terminal loosening and large shell damage), and different scale feature maps can be adapted to detect large, medium and small defects. The basic feature map is input into the feature pyramid network to generate three fusion feature maps of different scales. By fusing features of different levels, the model can not only retain global structure information (suitable for detecting large defects), but also capture local details (suitable for detecting small defects), thereby improving the recognition ability of various size faults.

[0027] Preferably, the fusion feature map is input into the fault feature enhancement layer, and the feature map is injected with standard position coordinates of key components of the equipment, such as the preset installation position of the door lock and the terminal row; then the feature attention mechanism is used to weight the features of the high-frequency fault areas such as the door lock and the terminal row, so that the model automatically focuses on these key areas when analyzing, reduces the interference of irrelevant backgrounds, and at the same time, the weight of the preset small size features (such as screw loosening and terminal misalignment) is amplified, solving the problem that small fault features are easily ignored, and improving the detection sensitivity; the enhanced feature map finally enters the detection head, which can identify the defects in the image through boundary box prediction and other operations, and when the overlap degree between the predicted defect detection box and the standard position area of the key component reaches a set threshold (such as 70%), it is determined that the defect belongs to the key component, and at this time the detection head outputs a structured feature vector containing the defect type, position coordinates, confidence and belonging component.

[0028] Further, step S110 further includes step S111 of using an adaptive light compensation algorithm to compensate for the input image data, and eliminating the interference of reflection.

[0029] Preferably, when shooting metering equipment (such as electric meters, transformers, etc.), due to the complex on-site lighting conditions, such as direct sunlight, backlight, local reflection, etc., the image may appear overexposed areas (glare), uneven brightness, or details are lost, which interferes with the subsequent visual model to recognize device defects, for example, glare may cover loose wiring faults, and overexposed areas may cause component outlines to be blurred; an adaptive lighting compensation algorithm is used to compensate for the input image data, that is, the brightness and contrast are dynamically adjusted according to the local lighting conditions of the image. Specifically, analyze the lighting intensity of different areas in the image, such as identifying glare areas, shadow areas, and normal brightness areas. The overexposed glare area is suppressed, such as reducing brightness and enhancing details. The overexposed shadow area is brightened to improve brightness while suppressing noise, and the visual features of the normal lighting area remain unchanged. Through adaptive adjustment, the glare, shadow and other lighting interference are eliminated or weakened, and the texture, outline, color and other features of the components of the metering equipment in the image, such as terminal rows, door locks, and display screens, are clearly identifiable.

[0030] In step S200, the extracted structured features are mapped to defect nodes in the knowledge graph, and an association relationship between the defect nodes and associated nodes is established, wherein the risk nodes in the associated nodes are bound to a preset risk level.

[0031] Preferably, the structured features output by the visual model (including defect type, position coordinates, confidence, and belonging component) are mapped to defect nodes in the knowledge graph, and each defect node represents a specific fault instance, for example, "terminal row loosening" is an independent node in the knowledge graph, which stores all attribute information of the defect. In addition to defect nodes, the knowledge graph also includes other associated nodes, including device nodes, cause nodes, strategy nodes, and risk nodes. The connection between the defect nodes and these associated nodes is established, for example, the "terminal row loosening" defect node is associated with the "terminal row component node", the "XX electric meter device node", and the "outdoor environment node" to form a "device-component-defect-environment" relationship network. In addition, the risk nodes in the associated nodes are bound to a preset risk level, which is usually divided into three levels: high, medium, and low. When the defect node is associated with the risk node, it directly inherits the preset level of the risk node, for example, the shell damage defect may be associated with the device water ingress risk (medium level), and the wiring terminal overheating defect may be associated with the short circuit fire risk (high level).

[0032] Further, step S200 further includes step S210 of performing multi-source data standardization processing on the visual extraction features and device operation parameters; step S220 of associating the standardized data with multi-modal features and mapping to four types of nodes, including defect nodes, cause nodes, risk nodes, and strategy nodes; and step S230 of calculating the association strength based on a graph neural network, connecting the four types of nodes, setting the weight between the nodes, and constructing the knowledge graph.

[0033] Preferably, the visual extraction features are structured features identified from images by a visual model, such as defect types, locations, confidence levels, etc., the device operation parameters refer to real-time or historical operation data of the metering device, such as voltages, currents, temperatures, communication states, power consumptions, etc., the standardized processing refers to converting data of different sources and formats into a unified format and standard, to obtain standardized data; then the standardized data is subjected to multi-modal feature correlation, i.e., mining the internal relationship between the visual features and the operation parameters, such as the defect of "loose wiring terminal" may be associated with operation parameter abnormalities such as "large current fluctuation" and "local temperature rise"; subsequently, the correlated features are classified and mapped into four types of core nodes of the knowledge graph, including defect nodes, cause nodes, risk nodes and strategy nodes, wherein the defect nodes correspond to specific fault manifestations, the cause nodes correspond to possible reasons for the fault, the risk nodes correspond to possible consequences caused by the fault, and the strategy nodes correspond to handling schemes for the fault; finally, the correlation strength is calculated based on the graph neural network, i.e., the dependence relationship between nodes is analyzed by an algorithm to determine the closeness, and then weights are set for the connections between nodes based on the correlation strength, to finally construct a complete knowledge graph.

[0034] Further, step S210 further includes step S211 of integrating the visual extraction features and the device operation parameters, wherein the visual extraction features include defect types, location coordinates, and belonging components, the device operation parameters include temperatures, current fluctuations, and operation durations, to obtain multi-source data; step S212 of converting text features in the multi-source data into 128-dimensional vectors by BERT word embedding; and step S213 of mapping numerical features in the multi-source data into 128-dimensional vectors by a fully connected layer after normalization.

[0035] Preferably, two types of data, visual extraction features (from image analysis) and equipment operation parameters (from sensors or monitoring systems), are identified and integrated, the visual extraction features including defect types (such as terminal loosening) identified from images, position coordinates (such as the specific position of the defect on the equipment), and the component to which the defect belongs (such as the incoming terminal component), and the equipment operation parameters including temperature (such as 35°C) reflecting the state of the equipment, current fluctuation (such as ±5A), and operation time (such as 1800 hours); then for the text type features in the multi-source data, a BERT model is used for word embedding conversion processing, wherein the BERT model is a pre-trained language model based on Transformer, which can understand the semantic information of the text, i.e., converting the text information into a fixed-dimensional numerical vector, and each dimension in the vector represents a specific semantic feature of the text, and the fixed dimension is 128; for the numerical type features in the multi-source data, normalization processing is first performed to map numerical features of different magnitudes to a unified interval, eliminating the dimensional difference; then through the full connection layer in the neural network, it is mapped into a 128-dimensional vector, so that the numerical features and the text features have the same dimension and data distribution.

[0036] Step S300, according to the knowledge graph nodes and the associated relationship, generating a hierarchical work order based on a dynamic rule engine, combining confidence and risk level.

[0037] Step S300 further includes step S310, starting from the defect node, tracing the cause node, the risk node, and the strategy node along the associated edge to generate a complete path; step S320, multiplying all edge weights on the complete path to calculate the path weight; step S330, based on the path weight, performing path screening to determine the main disposal scheme and the backup scheme; step S340, based on the main disposal scheme and the backup scheme, calculating the urgency according to the confidence and the risk level in combination with the diffusion probability; step S350, according to the matching of the urgency and the rule threshold, triggering the work order grading to generate the hierarchical work order.

[0038] Preferably, the knowledge graph stores various nodes (such as defect nodes, equipment nodes, cause nodes, risk nodes, and strategy nodes) related to faults and the associated relationship between the nodes, and the dynamic rule engine is a flexible logic rule with built-in decision logic based on operation and maintenance experience and business specifications, which can be adjusted in real time according to actual scenarios such as equipment importance, seasonal factors, and historical processing effects, for example, the preset rules may include: if the risk level is high and the confidence is ≥0.8, a first-level urgent work order is generated; if the risk level is medium and the confidence is ≥0.6, a second-level regular work order is generated; if the confidence is <0.5, regardless of the risk level, a third-level review work order is generated.

[0039] Preferably, starting from the defect node in the knowledge graph, other nodes related to it are traced along the associated edges (weighted connections) between the nodes to form a complete fault analysis path, the cause node is traced to find possible reasons leading to the defect, the risk node is traced to analyze the consequences that the defect may cause, and the strategy node is traced to match the treatment scheme for the defect; the weights of all associated edges on each complete path are multiplied to obtain the path weight, and the edge weight represents the reliability of the association between nodes. The higher the path weight, the more reliable the cause-and-effect relationship reflected by the path and the more credible the logical chain. Then, all possible paths are sorted and filtered according to the path weight, the highest path weight corresponds to the main disposal scheme, and the next highest path weight serves as a backup scheme. Then, based on the main disposal scheme and the backup scheme, the urgency is calculated according to the confidence and risk level in combination with the diffusion probability, that is, the urgency of fault disposal is calculated in combination with the risk node level corresponding to the main / backup scheme, the confidence of the defect identified by the visual model, and the diffusion probability, and the higher the value, the more immediate the processing is needed. Finally, the calculated urgency is matched with the preset rule threshold, and when the corresponding threshold is reached, the work order of the corresponding level is triggered, such as the emergency repair order corresponding to the high urgency and the routine inspection order corresponding to the low urgency, and finally a classified work order is generated, each work order containing clear execution information, including work order level, fault details, treatment suggestions and priority explanation, to ensure that maintenance resources are preferentially invested in high-risk and high-confidence faults, and to improve overall maintenance efficiency.

[0040] Further, step S310 further includes including an industrial constraint mechanism; wherein when the fault position coordinate deviates from the standard position of the component by > 15%, multiply all edge weights of the current path by a weight reduction factor 0.8; for power core fault features, including the association path between terminal defects and current fluctuations, multiply the weight by a reinforcement factor 1.2; output the compatibility of the strategy node and the equipment model before the path.

[0041] Preferably, the industrial constraint mechanism is a restrictive and adjustable rule set according to the actual needs and business rules of the industrial scene, which is used to optimize the accuracy of knowledge graph path reasoning and the rationality of work order generation. When the deviation between the fault location coordinates identified by the visual model and the standard position parameters of the key components of the equipment exceeds 15%, multiply all edge weights of the fault-related path in the current knowledge graph by the weight reduction factor 0.8. When the deviation between the fault location and the standard position of the component is too large, it means that the defect may not belong to the component or the identification result has errors, so the credibility of the associated relationship on the path is reduced. For the core fault features in the power system that directly affect the safety and accuracy of metering, including the association path of terminal defects and current fluctuations, multiply all edge weights of the path by the reinforcement factor 1.2 to highlight the relevance of key faults and ensure that faults related to the core functions of the power system are prioritized for identification and processing. Before outputting the final path, verify the compatibility of the strategy nodes and the fault device model in the knowledge graph to avoid invalid operations caused by mismatch between the scheme and the device model, and ensure the actual executability of the operation and maintenance strategy. Different types of metering equipment, such as smart meters from different manufacturers and transformers of different specifications, may differ in structure and component size, so the corresponding repair or replacement scheme needs to be adapted.

[0042] Step S400, feedback through hierarchical work order execution results, dynamically update the knowledge graph.

[0043] Step S400 further includes step S410, when the feedback of the hierarchical work order execution result is that the work order is successful, the associated edge weight is reinforced; step S420, when the work order fails and the root cause is not covered, new graph nodes and relationship chains are added, and inefficient edges below the weight threshold are periodically removed based on the edge weight in the knowledge graph.

[0044] Preferably, the execution result feedback of the hierarchical work order is used to update the knowledge graph dynamically. Specifically, when the execution result feedback of the hierarchical work order is success, i.e., the fault is solved after processing according to the work order scheme, it indicates that the association relationship of the path corresponding to the fault in the knowledge graph, such as the association relationship of the defect node→cause node→strategy node, is accurate and effective, and then the weights of all associated edges in the path are reinforced, such as increasing a certain proportion or a fixed value. When the work order execution fails, i.e., the fault is not solved, and the analysis finds that the root cause of the failure is information not contained in the current knowledge graph, such as unrecorded fault causes, unassociated risk points, etc., the graph nodes and relationship chains are added, for example, if the execution of the electric meter black screen work order fails, it is found that the mainboard chip aging causes the failure, but the knowledge graph originally does not have the cause node of mainboard chip aging, so the cause node is added. The association between the new node and the original node is established, such as the association edge between the electric meter black screen and the mainboard chip aging, and an initial weight is assigned. The weights of all associated edges in the knowledge graph are checked regularly, and the edges below the preset weight threshold are determined as inefficient edges and removed, wherein the inefficient edge refers to an edge with weak association relationship or long-term ineffective use, such as a very low weight, which indicates that the association is rarely established in historical verification.

[0045] The above detailed description does not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application. In some cases, the actions or steps described in the present application can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

Claims

1. A method for active troubleshooting of metrology assets based on a knowledge graph and a rule engine, characterized in that, The method comprises the following steps: Collecting image data of a metering device, analyzing and processing the image data through a visual model, extracting structured features, including defect type, position coordinates, confidence, and belonging component; Mapping the extracted structured features to defect nodes in a knowledge graph, and establishing the association relationship between the defect nodes and the associated nodes, wherein the risk nodes in the associated nodes are bound to a preset risk level; According to the knowledge graph nodes and the association relationship, generating a hierarchical work order based on a dynamic rule engine, combining the confidence and the risk level; Feedback through the execution result of the hierarchical work order to dynamically update the knowledge graph.

2. The knowledge graph and rule engine-based metering asset fault proactive troubleshooting method according to claim 1, characterized in that, The method further comprises the following steps before analyzing and processing the image data through the visual model and extracting the structured features: Replacing the backbone network of the YOLOv5 model architecture with ConvNeXt-Tiny; Adding a fault feature enhancement layer to the YOLOv5 model architecture, and inputting the standard position and size parameters of the key components of the device; Building a visual model framework structure by weighting the key areas through a feature attention mechanism and amplifying the feature signals of size defects; Training and converging the visual model framework structure through a training data set to obtain the visual model.

3. The knowledge graph and rule engine-based metering asset fault proactive troubleshooting method according to claim 2, characterized in that, The method further comprises the following steps before analyzing and processing the image data through the visual model and extracting the structured features: Inputting the image data into the visual model, outputting a basic feature map through the ConvNeXt-Tiny backbone network; Generating three scale fusion feature maps through a feature pyramid; The fusion feature map enters the fault feature enhancement layer, injects the standard position coordinates of the key components of the device into the feature map, and then weights the lock and terminal area features through a feature attention mechanism to amplify the preset small size feature weight; When the overlap degree of the detection box and the key components reaches a threshold, the detection head outputs a structured feature vector.

4. The knowledge graph and rule engine-based metering asset fault proactive troubleshooting method according to claim 3, characterized in that, The method further comprises the following steps before inputting the image data into the visual model, outputting a basic feature map through the ConvNeXt-Tiny backbone network: Using an adaptive light compensation algorithm to compensate the input image data, and eliminating the interference of reflection. 5.The knowledge graph and rule engine based metrology asset fault proactive troubleshooting method of claim 1, wherein, The method further comprises the following steps before mapping the extracted structured features to defect nodes in a knowledge graph: Standardizing the visual extraction features and the device operation parameters through multi-source data; Correlating the standardized data through multi-modal features, and mapping them to four types of nodes, including defect nodes, cause nodes, risk nodes, and strategy nodes; Calculating the correlation strength based on a graph neural network, connecting the four types of nodes, setting the weight between the nodes, and constructing the knowledge graph.

6. The knowledge graph and rule engine-based metering asset fault proactive troubleshooting method according to claim 5, characterized in that, The method further comprises the following steps before standardizing the visual extraction features and the device operation parameters through multi-source data: Integrating the visual extraction features and the device operation parameters, wherein the visual extraction features include defect type, position coordinates, and belonging component, and the device operation parameters include temperature, current fluctuation, and operation time, to obtain multi-source data; Converting the text features in the multi-source data into 128-dimensional vectors through BERT word embedding; After normalizing the numerical features in the multi-source data, mapping them to 128-dimensional vectors through a fully connected layer.

7. The knowledge graph and rule engine based proactive troubleshooting method of metrology assets according to claim 1, characterized in that, According to the knowledge graph nodes and the associated relationship, a hierarchical work order is generated based on a dynamic rule engine, combined with confidence and risk level, including: Starting from the defect node, tracing the cause node, risk node and strategy node along the associated edge to generate a complete path; Multiply all edge weights on the complete path to calculate the path weight; Based on the path weight, the path is screened to determine the main disposal scheme and the backup scheme; Based on the main disposal scheme and the backup scheme, the urgency is calculated according to the confidence and the risk level combined with the diffusion probability; According to the matching of the urgency and the rule threshold value, the work order grading is triggered to generate the hierarchical work order.

8. The knowledge graph and rule engine-based metering asset fault proactive troubleshooting method according to claim 7, characterized in that, Starting from the defect node, tracing the cause node, risk node and strategy node along the associated edge to generate a complete path, including an industrial constraint mechanism; When the fault position coordinates deviate from the standard position of the component by more than 15%, multiply all edge weights on the current path by a weight reduction factor of 0.8; For power core fault characteristics, including the associated path of terminal defects and current fluctuations, multiply the weight by a reinforcement factor of 1.2; Verify the compatibility of the strategy node and the equipment model before outputting the path. 9.The knowledge graph and rule engine based metrology asset fault proactive troubleshooting method of claim 1, wherein, Through the feedback of the hierarchical work order execution result, the knowledge graph is dynamically updated, including: When the feedback of the hierarchical work order execution result is that the work order is successful, the associated edge weight is reinforced; When the work order fails and the root cause is not covered, new graph nodes and relationship chains are added, and based on the edge weight in the knowledge graph, inefficient edges below the weight threshold value are removed regularly.

Images