Operation safety management and control system and method based on risk grading
By integrating multiple modal data into the knowledge graph and using the map attention network, the problem of insufficient single modal data in power field operation safety is solved, the accuracy of risk assessment and targeted control are improved, and the intelligence is achieved risk assessment and multi-dimensional analysis.
Patent Information
- Application Number
- CN202411272197.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the safety control of power field operations, single modal data is difficult to provide complete information, resulting in inaccuracy in the risk assessment of artificial intelligence models and the inability to effectively identify complex risk factors and influencing mechanisms.
By integrating power field operation data of multiple modes in the knowledge graph, using the map attention network to supplement and optimize the knowledge graph, identify key risk factors and impact mechanisms, build an operation risk assessment model, extract response measures and suggestions, and determine The control priority of entity nodes.
It improves the accuracy and reliability of risk grading assessment, realizes intelligent risk assessment and control, can automatically learn and explore complex relationships in the knowledge graph, provide multi-dimensional risk analysis indicators, and reasonably allocate management and control resources.
Smart Images

Figure CN119047842A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power field operation safety technology, and in particular to an operation safety management and control system and method based on risk classification. Background Art
[0002] There are many and wide-ranging power site operations, many construction workers, and complex environments. Various safety risks and potential accidents are intertwined and superimposed, making safety management and control difficult. In view of the scattered operation sites, insufficient personnel and equipment control, and insufficient safety supervision manpower in on-site operation safety management and control, on-site operation safety management and control based on various monitoring terminals is currently widely carried out. Based on the joint use of various sensors, video image recognition systems, wearable equipment, and tools, illegal behaviors in on-site operations are automatically identified and alarmed, improving the level of operation safety supervision.
[0003] The Chinese patent with announcement number CN113254594B discloses a method and system for constructing a safety knowledge graph for smart power plants. The method includes first collecting data and constructing a knowledge model, classifying, cleaning and storing the data according to the knowledge model; then extracting knowledge from the data without entity separation, and using the results of knowledge extraction and the data with entity separation to construct a knowledge graph; finally, the constructed knowledge graph is applied to different scenarios. However, the above scheme does not take into account the complex environment of the power site, the variety of management and control types, and the overlapping of various risks and hidden dangers. It is difficult to provide complete information about a phenomenon using single modal data, which can easily cause operational risks in artificial intelligence models. Therefore, it is very necessary to provide an operational safety management and control system and method based on risk grading to improve the accuracy and reliability of risk grading assessment. Summary of the invention
[0004] In view of this, the present invention proposes a risk-based safety management and control system and method. By integrating multiple modalities of power field operation data in the knowledge graph to fully reflect the various entities, attributes and relationships in the operation process, and supplementing and optimizing the knowledge graph based on the graph attention network, the key risk factors and influencing mechanisms can be more accurately identified, thereby continuously improving the accuracy and reliability of risk grading assessment.
[0005] The present invention provides a method for operation safety management and control based on risk classification, the method comprising: Acquire power field operation data of multiple modes, and construct a knowledge graph based on the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes, and relationship nodes; Based on the graph attention network, each node and relationship edge in the knowledge graph is supplemented to obtain a job risk assessment model, wherein the relationship edge represents an edge connecting nodes; Inputting the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level; Based on the risk level and the risk probability, extract relevant nodes and relationship edges from the knowledge graph to form a countermeasure suggestion subgraph; According to the countermeasure recommendation values of the countermeasure recommendation subgraph, the risk levels, entity nodes and countermeasures of the power field operations are sorted to obtain the control priority corresponding to each entity node, and the operation safety control of all entity nodes is performed based on the control priority.
[0006] On the basis of the above technical scheme, preferably, the entity nodes include power operation categories, power operation projects, power operation personnel, power operation equipment, and power operation environment; the attribute nodes include the regulations and systems of power operation categories, the risk levels of power operation projects, the qualifications of power operation personnel, the status of power operation equipment, and the conditions of the power operation environment; the relationship nodes include the inclusion relationship between power operation categories and power operation projects, the allocation relationship between power operation projects and power operation personnel, the usage relationship between power operation projects and power operation equipment, and the influence relationship between power operation projects and the power operation environment.
[0007] On the basis of the above technical solution, preferably, the operation risk assessment model includes a relationship quantification layer, an entity attention network layer, a relationship attention network layer and an entity classification layer, wherein: The relationship quantization layer is connected to the entity attention network layer, and the relationship quantization layer is used to quantify the relationship between different entity nodes, and output the quantized relationship value between different entity nodes to the entity attention network layer according to the quantized relationship function; The entity attention network layer is connected to the relationship attention network layer, and the entity attention network layer is used to construct an entity attention network according to the attribute nodes and the quantized relationship values to obtain normalized attention coefficients between entity nodes, and output an updated embedding vector corresponding to each entity node to the relationship attention network layer based on the normalized attention coefficients and the vector update function; The relationship attention network layer is connected to the entity classification layer, and the relationship attention network layer is used to construct a relationship attention network according to the quantized relationship value, the updated embedding vector and the relationship node, and output an adjusted attention coefficient to the entity classification layer according to the updated attention function; The entity classification layer is used to classify all entity nodes according to adjusting the attention coefficient and updating the embedding vector, and obtain the risk category probability corresponding to each entity node according to the risk classification function.
[0008] More preferably, the expression of the quantization relationship function is: ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The quantitative relationship value under the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, M k Indicates k A relationship matrix, k represents the number of relationship matrices, Indicated in k The mapping function under the relationship matrix is Indicates i The embedding vector corresponding to the entity node f The value of the dimension, F represents the total number of dimensions of the embedding vector, Indicates j The embedding vector corresponding to the entity node f The value of the dimension.
[0009] More preferably, the expression of the vector update function is: ; ; ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The attention coefficient in the relationship matrix, represents the mapping function under the kth relationship matrix, Indicates k The weight matrix of the entity attention network layer in the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, ej Indicates i The embedding vector corresponding to the entity node, || represents the connection symbol, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The normalized attention coefficient in the relationship matrix, LeakyReLU ( ) represents the first nonlinear activation function, s Represents and except i Any entity node other than entity nodes, N Represents the total number of entity nodes, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, s represents the second nonlinear activation function, J Indicates k All neighboring points whose relationship quantization values under the relationship matrix are greater than the relationship quantization threshold j , G represents the total number of attention units, g Indicates g Attention units, W g Indicates g The feature transformation matrix corresponding to each attention unit.
[0010] More preferably, the expression of updating the attention function is specifically: ; in, Indicated in k The adjusted attention coefficient in the relationship matrix, N Represents the total number of entity nodes, K represents the total number of relationship matrices, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, W represents the parameterized matrix, b represents the offset vector, a represents the attention mechanism vector, sigmoid ( ) represents the third nonlinear activation function.
[0011] More preferably, the risk classification function is expressed as: ; ; ; in, represents the integrated feature vector set output by the operational risk assessment model, sigmoid ( ) represents the third nonlinear activation function, W C represents the risk weight, s represents the second nonlinear activation function, G represents the total number of attention units, g Indicates g Attention units, K represents the total number of relationship matrices, Indicated in K The adjusted attention coefficient in the relationship matrix, Indicated in K The relationship matrix i The updated embedding vector of entity nodes, Indicates i The entity node belongs to x The probability of risk category, D represents the total number of integrated feature vectors in the integrated feature vector set, d Indicates d The integrated feature vector, f d Indicates the first d The integrated feature vector, f x Represents the first risk category vector in the risk category set x risk category vectors, and C represents the total number of risk categories.
[0012] More preferably, the method further comprises: Based on the true category label of each entity node and the risk category probability of the corresponding entity node, the entity node loss function is obtained; Obtaining a regularization term function based on the embedding vector of the entity node and the positive sample embedding vector of the corresponding entity node; The entity node loss function and the entity node loss function are summed to obtain a total loss function, and the operation risk assessment model is compensated based on the total loss function.
[0013] More preferably, the expression of the total loss function is: ; ; ; in, L task represents the entity node loss function, N represents the total number of entity nodes, C represents the total number of risk categories,y i,c Indicates i The true category labels of entity nodes, Indicates i The entity node belongs to x The probability of risk category, L KG represents the regularization function of the knowledge graph, c represents the interval hyperparameter, d e ( ) represents the distance metric function, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, Indicates i The positive sample embedding vector of entity nodes in the knowledge graph, L is the total loss function, l represents the regularization weight coefficient.
[0014] In a second aspect of the present application, a risk-based operation safety management and control system is provided, the operation safety management and control system comprising a knowledge graph construction module, an operation risk assessment module and an operation safety management and control module, wherein: The knowledge graph construction module is used to obtain power field operation data of multiple modes, and construct a knowledge graph according to the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes and relationship nodes; The operation risk assessment module is used to supplement each node and relationship edge in the knowledge graph based on the graph attention network to obtain an operation risk assessment model, wherein the relationship edge represents the edge connecting the nodes, and the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node are input into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level; The operation safety management and control module is used to extract relevant nodes and relationship edges from the knowledge graph based on the risk level and the risk probability to form a countermeasure suggestion subgraph, and sort the risk level, entity nodes and countermeasures of the power field operation according to the countermeasure suggestion value of the countermeasure suggestion subgraph to obtain the management and control priority corresponding to each entity node, and perform operation safety management and control on all entity nodes based on the management and control priority.
[0015] The risk classification-based operation safety management system and method provided by the present invention has the following beneficial effects compared with the prior art: (1) By integrating multi-modal power field operation data into the knowledge graph to comprehensively reflect the various entities, attributes and relationships in the operation process, data and knowledge support are provided for operation risk assessment and control. In addition, the knowledge graph is supplemented and optimized based on the graph attention network, which can more accurately identify key risk factors and influencing mechanisms, and can automatically learn and mine the complex relationships implicit in the knowledge graph to achieve intelligent risk assessment, so as to continuously improve the accuracy and reliability of risk classification assessment and the pertinence of management and control. According to the risk level and risk probability, a subgraph of countermeasures suggestions is constructed, which can put forward targeted management and control measures for various entity nodes, so as to determine the management and control priority of each entity node, which is conducive to the reasonable allocation of limited management and control resources. (2) By quantifying the relationship between entity nodes and constructing entity attention networks and relationship attention networks, the mutual influence and risk propagation mechanism between nodes can be better captured. Through the vector update function and attention mechanism, the operation risk assessment model can adaptively adjust the embedding vectors of entity nodes and relationship attention coefficients to continuously optimize the accuracy of risk assessment. The operation risk assessment model can not only output the risk category probability of each entity node, but also provide multi-dimensional risk analysis indicators such as quantified relationship values, normalized attention coefficients and adjusted attention coefficients, thereby realizing intelligent perception and multi-dimensional analysis of operation risks. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0017] Figure 1 A schematic diagram of the process of the operation safety management and control method based on risk classification provided by the present invention; Figure 2 A schematic diagram of the framework of the operation safety management and control system provided by the present invention.
[0018] Explanation of the accompanying drawings: 1. Operation safety management and control system; 11. Knowledge graph construction module; 12. Operation risk assessment module; 13. Operation safety management and control module. DETAILED DESCRIPTION
[0019] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0020] The present invention discloses a method for operation safety management and control based on risk classification, referring to Figure 1 , the steps of the method include S1~S5.
[0021] Step S1, obtaining power field operation data of multiple modes, and constructing a knowledge graph based on the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes, and relationship nodes.
[0022] In this step, the power field operation data includes structured data, unstructured data and multi-sensor data. Among them, structured data includes operator information, such as job position, work experience, training records and health status; operation equipment information, such as equipment model, equipment status and maintenance records; operation environment information, such as temperature, humidity, wind speed and illumination and other environmental parameters; operation process information, such as operation flow, operation instructions and operation time; historical accident information, such as accident cause, accident loss and accident investigation report; safety management system, such as operating procedures, safety training requirements and emergency plans. Unstructured data includes operation site photos / videos; operation process documents, such as operation logs, operation reports and emergency handling records; external data sources, such as meteorological data, geographic environment data and industry standards. Multi-sensor data includes IoT sensors deployed at the operation site, such as equipment vibration, current and temperature and other operating status data; on-site image data collected by drones / robots, such as operation environment and equipment status. After collecting the above structured data, unstructured data and multi-sensor data, the above data are sequentially subjected to operations such as data cleaning, data conversion, data integration, data standardization and data enrichment.
[0023] Furthermore, named entity recognition (NER) technology is used to extract entities such as operators, equipment, and environment from unstructured text data, and combined with knowledge bases and domain dictionaries, more abundant entity types such as types of operations and accident types are identified. Deep learning models such as BERT and SpaCy are used to improve the accuracy and coverage of entity recognition, and the identified entities are normalized to eliminate the ambiguity of synonyms and abbreviations. Dependency syntactic analysis or pattern matching is used to identify semantic relationships between entities from texts. For example, common relationship types include operation-equipment, monitoring-environment, construction-process, and report-accident, etc. The extracted relationships are selected and classified to build a standardized relationship pedigree, combined with domain knowledge, to supplement those relationships that are implicit or not explicitly stated in the text.
[0024] The identified entities and relationships are mapped to nodes and edges of the knowledge graph, where entity nodes represent specific things such as operators, equipment, and environment, and relationship edges represent logical associations between entities, such as "operation-equipment" or "monitoring-environment". Attribute information is added to the nodes, such as name, position, work experience, and training status for operators; model, status, and maintenance records for operating equipment; and temperature, humidity, wind speed, and illumination for the operating environment. At the same time, relationship types and relationship weights are added to the relationship edges to indicate the strength of association between entities. Relationship types include operation, monitoring, reporting, and construction, and relationship weights are used to reflect the importance of the relationship, which can be set according to the actual scenario.
[0025] Align and fuse knowledge graphs from different data sources, identify and eliminate ambiguity and duplication of entities and relationships in the graph, and establish mapping relationships based on the semantic similarity of entities and relationships. Use graph embedding technology to discover hidden semantic associations in the knowledge graph, capture implicit relationships between entity nodes through node representation learning, discover potential associations between entity nodes based on relationship representation learning, and introduce domain expert knowledge to manually correct and optimize the knowledge graph, thereby supplementing missing entities and relationships in the graph and correcting erroneous or unreasonable content in the graph.
[0026] Furthermore, entity nodes include power operation categories, power operation projects, power operation personnel, power operation equipment, and power operation environments; attribute nodes include regulations and systems for power operation categories, risk levels of power operation projects, qualifications of power operation personnel, status of power operation equipment, and conditions of the power operation environment; relationship nodes include the inclusion relationship between power operation categories and power operation projects, the allocation relationship between power operation projects and power operation personnel, the usage relationship between power operation projects and power operation equipment, and the influence relationship between power operation projects and the power operation environment.
[0027] In one example, the power operation category includes different types of power operations such as line erection, transformer maintenance and line inspection, power operation projects include specific operation tasks, such as erection of 110kV transmission lines, maintenance of 10MVA transformers, etc., power operation personnel include workers, technicians and managers involved in the operation, power operation equipment includes various tools and equipment used for the operation, such as insulating operating rods, thermal imagers, etc., and the power operation environment includes the geographical location, climatic conditions and surrounding environment of the operation site. The regulations and systems of the power operation category include the operation standards and safety regulations of the State Grid Corporation, the risk levels of power operation projects include high, medium, low and no risk levels divided according to the degree of danger of the operation, the qualifications of power operation personnel include professional skills, safety training and operation certificates, the status of power operation equipment includes equipment model, performance parameters and maintenance status, and the conditions of the power operation environment include environmental parameters such as temperature, humidity, wind speed and illumination. The inclusion relationship between power operation categories and power operation projects includes that a certain type of operation includes multiple specific operation projects. The allocation relationship between power operation projects and power operation personnel includes that different operation projects need to be completed by personnel with specific qualifications. The usage relationship between power operation projects and power operation equipment includes that operation projects need to use specific operation equipment. The impact relationship between power operation projects and the power operation environment includes that the operation environment will affect the implementation and risks of operation projects.
[0028] By covering key elements such as operation categories, operation items, operators, operation equipment and operation environment, a three-dimensional and multi-dimensional power operation knowledge system has been constructed, and logical relationships have been established between operation categories and operation items, and between operation items and operators / equipment / environment. The inherent connections between the various elements can be deeply analyzed. At the same time, the knowledge graph also incorporates the risk level information of the operation items, and can combine information such as operator qualifications, equipment status, and environmental conditions to achieve intelligent operation risk warnings. By incorporating historical accident information into the graph and establishing connections with operation elements, the cause of the accident can be quickly located, and accurate accident tracing analysis can be performed.
[0029] Step S2, based on the graph attention network, supplement each node and relationship edge in the knowledge graph to obtain a job risk assessment model, where the relationship edge represents the edge connecting the nodes.
[0030] In this step, the graph attention network includes an input layer, a graph attention layer, a graph convolution layer, and an output layer. The input layer receives node feature input and relationship feature input. The node feature includes the feature vector of each entity node in the knowledge graph, such as node type, attribute information, etc. These features can be obtained by performing feature engineering on the power field operation data. The relationship feature includes the relationship matrix of each relationship node, which is used to describe the connection relationship between nodes. The relationship matrix can be represented as an adjacency matrix or other forms of relationships. The graph attention layer adopts a multi-head attention mechanism to allow the operation risk assessment model to learn the importance of nodes and relationships from different perspectives, and each attention head can capture different types of associated features. The graph convolution layer includes a graph convolution based on the attention mechanism. Using the learned attention weights, the node features and relationship features are convoluted to capture the multi-level associations between nodes and neighboring nodes and relationships. At the same time, multiple graph convolution layers are stacked to gradually learn the graph representation from local to global, and extract the high-order semantic features of nodes and relationships. The output layer includes inputting the node features output by the graph convolution layer into the fully connected layer or other prediction heads, and outputting the risk level and risk probability of each entity node. The graph attention network is used to automatically supplement the nodes and relationships in the knowledge graph, discover missing entity nodes and relationship edges, and infer and complete them based on the characteristics of surrounding nodes and edges, and update the attribute information of nodes and edges.
[0031] In this embodiment, the operation risk assessment model includes a relationship quantification layer, an entity attention network layer, a relationship attention network layer and an entity classification layer, wherein: The relationship quantization layer is connected to the entity attention network layer. The relationship quantization layer is used to quantify the relationship between different entity nodes and output the quantized relationship values between different entity nodes to the entity attention network layer according to the quantized relationship function.
[0032] The expression of the quantified relationship function is: ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The quantitative relationship value under the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, M k Indicates k Relationship matrices, each relationship corresponds to a relationship matrix M ,M ={ M 1, M 2, M 3,..., M k}, k represents the number of relationship matrices, Indicated in k The mapping function under the relationship matrix is Indicates i The embedding vector corresponding to the entity node f The value of the dimension, F represents the total number of dimensions of the embedding vector, Indicates j The embedding vector corresponding to the entity node f The value of the dimension.
[0033] The entity attention network layer is connected to the relationship attention network layer. The entity attention network layer is used to construct an entity attention network based on attribute nodes and quantized relationship values to obtain the normalized attention coefficient between entity nodes, and output the updated embedding vector corresponding to each entity node to the relationship attention network layer based on the normalized attention coefficient and the vector update function.
[0034] The expression of the vector update function is: ; ; ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The attention coefficient in the relationship matrix, represents the mapping function under the kth relationship matrix, Indicates k The weight matrix of the entity attention network layer in the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, || represents the connection symbol, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The normalized attention coefficient in the relationship matrix, LeakyReLU ( ) represents the first nonlinear activation function, s Represents and excepti Any entity node other than entity nodes, N Represents the total number of entity nodes, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, s represents the second nonlinear activation function, J Indicates k All neighboring points whose relationship quantization values under the relationship matrix are greater than the relationship quantization threshold j , G represents the total number of attention units, g Indicates g Attention units, W g Indicates g The feature transformation matrix corresponding to each attention unit.
[0035] The relationship attention network layer is connected to the entity classification layer. The relationship attention network layer is used to build a relationship attention network based on the quantized relationship values, updated embedding vectors and relationship nodes, and output the adjusted attention coefficient to the entity classification layer according to the updated attention function.
[0036] The expression for updating the attention function is: ; in, Indicated in k The adjusted attention coefficient in the relationship matrix, N Represents the total number of entity nodes, K represents the total number of relationship matrices, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, W represents the parameterized matrix, b represents the offset vector, a represents the attention mechanism vector, sigmoid ( ) represents the third nonlinear activation function.
[0037] The entity classification layer is used to classify all entity nodes according to the adjusted attention coefficient and updated embedding vector, and obtain the risk category probability corresponding to each entity node according to the risk classification function.
[0038] The expression of risk classification function is: ; ; ; in, represents the integrated feature vector set output by the operational risk assessment model, sigmoid ( ) represents the third nonlinear activation function, W C represents the risk weight, s represents the second nonlinear activation function, G represents the total number of attention units, g Indicates g Attention units, K represents the total number of relationship matrices, Indicated in K The adjusted attention coefficient in the relationship matrix, Indicated in K The relationship matrix i The updated embedding vector of entity nodes, Indicates i The entity node belongs to x The probability of risk category, D represents the total number of integrated eigenvectors in the integrated eigenvector set, d Indicates d The integrated feature vector, f d Indicates the first d The integrated feature vector, f x Represents the first risk category vector in the risk category set x risk category vectors, and C represents the total number of risk categories.
[0039] In one example, C=1 means no risk, C=2 means low risk, C=3 means medium risk, and C=4 means high risk.
[0040] Step S3, inputting the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level.
[0041] In one example, assume that there is a knowledge graph of a power operation site, which includes operation category nodes, operation project nodes, operation personnel nodes, operation equipment nodes, and operation environment nodes.
[0042] The node characteristics of the operation category node include: operation type (construction, maintenance, overhaul, etc.), operation scene (line, substation, etc.).
[0043] The node characteristics of the operation project node include: operation content description, operation risk level assessment results, and historical accident information.
[0044] The node characteristics of the operator node include: personal information, safety qualification certification, safety training experience, and work experience.
[0045] The node characteristics of the operating equipment node include: equipment model, status parameters, and maintenance records.
[0046] The node characteristics of the working environment node include: climate conditions, geographical location, and working site conditions.
[0047] The feature vectors of these entity nodes and the relationship matrix between nodes are input into the trained graph attention network model. The output of the operation risk assessment model is as follows: Operation category node "Line construction" Risk probability: The probability is 0.7 when C=3, and the probability is 0.3 when C=4.
[0048] Risk level: 3 (medium risk) Operation project node "110kV line maintenance": Risk probability: The probability is 0.9 when C=4, and the probability is 0.1 when C=3.
[0049] Risk level: 4 (high risk) Operator node "XX": Risk probability: The probability is 0.8 when C=2, and the probability is 0.2 when C=3.
[0050] Risk level: 2 (low risk) Job equipment node "Circuit Breaker X-100": Risk probability: The probability is 0.6 when C=3, and the probability is 0.4 when C=4.
[0051] Risk level: 3 (medium risk) Operating environment node "plateau area": Risk probability: The probability is 0.8 when C=4 and 0.2 when C=3.
[0052] Risk level: 4 (high risk) Furthermore, the method further comprises: Based on the true category label of each entity node and the risk category probability of the corresponding entity node, the entity node loss function is obtained; Obtain a regularization term function based on the embedding vector of the entity node and the positive sample embedding vector of the corresponding entity node; The entity node loss function and the entity node loss function are summed to obtain the total loss function, and the operation risk assessment model is compensated based on the total loss function.
[0053] The expression of the total loss function is ; ; ; in, L task represents the entity node loss function, N represents the total number of entity nodes, C represents the total number of risk categories, y i,c Indicates i The true category labels of entity nodes, Indicates i The entity node belongs to x The probability of risk category, L KG represents the regularization function of the knowledge graph, c represents the interval hyperparameter, d e ( ) represents the distance metric function, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, Indicates i The positive sample embedding vector of entity nodes in the knowledge graph, L is the total loss function, l represents the regularization weight coefficient.
[0054] Step S4, based on the risk level and risk probability, extract relevant nodes and relationship edges from the knowledge graph to form a countermeasure suggestion subgraph.
[0055] In this step, all entity nodes in the knowledge graph are traversed to obtain the risk level and risk probability distribution information of each node. This information may come from the output of the previous risk assessment model. According to the preset risk level and risk probability threshold, the entity nodes that meet the "high risk" condition are screened out. For example, nodes with a risk level of 3 or 4 and a risk probability greater than 0.5. These entity nodes identified as "high risk" are used as the core nodes of the countermeasure suggestion subgraph. For each identified high-risk entity node, search and extract along the various relationship edges in the knowledge graph. The following types of relationship edges are mainly concerned: job category relationship edge (extracting the job category node related to the high-risk entity); job project relationship edge (extracting the job project node related to the high-risk entity); operator relationship edge (extracting the operator node related to the high-risk entity); operation equipment relationship edge (extracting the operation equipment node related to the high-risk entity); operation environment relationship edge (extracting the operation environment node related to the high-risk entity). Through the search of these relationship edges, other entity nodes related to the high-risk entity node can be obtained. These nodes may be the key factors leading to high risk or the target objects of countermeasures.
[0056] In one example, corresponding emergency responses can be made for different risk levels, among which, Level I emergency response: for high-risk situations with the highest risk level (such as level 4) and a high risk probability (such as >0.7). In this case, serious consequences such as personal injury, equipment damage or environmental pollution may occur. The most stringent and comprehensive emergency response measures need to be taken, including immediate suspension of operations, activation of emergency plans, emergency evacuation of personnel, and mobilization of a large number of emergency resources. Level II emergency response: for medium-risk situations with a high risk level (such as level 3) and a high risk probability (such as >0.5). In this case, a certain degree of personal injury, equipment damage or environmental pollution may occur. Relatively strict emergency response measures need to be taken, including temporary suspension of operations, partial activation of emergency plans, and mobilization of necessary emergency resources. Level III emergency response: for low-risk situations with a low risk level (such as level 2) and a high risk probability (such as >0.6). In this case, minor consequences such as personal injury, equipment damage or environmental pollution may occur. Certain emergency response measures need to be taken, including temporary adjustment of the operation plan, strengthening on-site monitoring, and preparing necessary emergency resources.
[0057] Furthermore, the entity nodes identified as "high risk" are used as the core nodes of the countermeasure suggestion subgraph. For each high-risk entity node, other entity nodes related to it are extracted from the knowledge graph, including job categories, job items, operators, operating equipment, and operating environments. These related entity nodes are added to the subgraph to form a node set related to the high-risk entity nodes. After adding the related entity nodes, corresponding relationship edges are established between these nodes. The types of relationship edges may include job category relationships, job item relationships, operator relationships, operating equipment relationships, and operating environment relationships. After the subgraph framework is constructed, the expert experience, historical accident data, and other information related to the high-risk entities are further supplemented. Based on the above information, targeted countermeasure suggestions are proposed, such as adjusting the operation plan, strengthening training, and optimizing equipment maintenance. These countermeasure suggestions are attached to the subgraph in the form of annotations and comments to form a complete knowledge graph enhanced risk management and control plan. In order to facilitate understanding and application, the countermeasure suggestion subgraph can also be structured. For example, the high-risk entity nodes in the subgraph can be classified according to the emergency response level (I, II, and III), and the corresponding emergency plan information can be attached to each of them. The sub-maps can also be partitioned or displayed in layers according to the type of response measures (such as adjustment of work plan, personnel training, equipment maintenance, etc.).
[0058] Step S5, according to the countermeasure recommendation values of the countermeasure recommendation subgraph, the risk level, entity nodes and countermeasures of the power field operation are sorted to obtain the control priority corresponding to each entity node, and the operation safety control of all entity nodes is performed based on the control priority.
[0059] In this step, the countermeasure recommendation value can be obtained according to the weighted algorithm of risk level, risk probability, key influencing factors and expert experience, wherein different weights are given according to the risk level (such as 1-4) of each entity node obtained in the previous step, and the higher the risk level, the higher the countermeasure recommendation value. At the same time, the risk probability distribution of each entity node is considered, and the higher the risk probability, the higher the countermeasure recommendation value. Analyze other entity nodes related to high-risk entity nodes in the subgraph (such as operation category, operator, operation equipment, etc.) to identify key influencing factors. Combined with the experience and suggestions of experts in related fields and the analysis of historical accident data, the importance of countermeasures is evaluated.
[0060] The expression of the recommended value of the countermeasure is: Recommended value of countermeasures = α*risk level + β*risk probability + γ*key influencing factors + δ*expert experience Among them, α, β, γ and δ are weight coefficients of different factors, which can be adjusted according to actual conditions.
[0061] In an example, suppose we have the following five entity nodes: high-voltage line maintenance operations (risk level 4, recommended response value is 8.5); transformer maintenance operations (risk level 3, recommended response value is 7.2); distribution box maintenance operations (risk level 3, recommended response value is 6.8); cable laying operations (risk level 2, recommended response value is 5.9); meter installation operations (risk level 2, recommended response value is 5.2).
[0062] First, the risk levels are ranked: high-voltage line maintenance work (level 4); transformer maintenance work (level 3); distribution box maintenance work (level 3); cable laying work (level 2); meter installation work (level 2).
[0063] Sort entity nodes: At risk level 4, high-voltage line maintenance work (8.5); At risk level 3, transformer maintenance work (7.2), distribution box maintenance work (6.8); At risk level 2, cable laying operations (5.9), meter installation operations (5.2); Rank the responses: High-voltage line maintenance operations: strengthen operator training (9.0), optimize operation plans (8.5), and strengthen equipment inspections (7.8); Transformer maintenance work: optimize operation plan (7.5), strengthen equipment maintenance (7.2), strengthen on-site monitoring (6.8); Distribution box maintenance work: strengthen equipment maintenance (7.0), optimize operation plan (6.8), strengthen on-site monitoring (6.5); Cable laying operations: optimize operation plans (6.2), strengthen operator training (5.9), and strengthen on-site supervision (5.6); Electric meter installation operations: strengthen operator training (5.5), optimize operation plans (5.2), and strengthen on-site monitoring (5.0); Through the above ranking, it can be clearly understood that high-voltage line maintenance operations are the highest-risk operations and need to be controlled first. Among high-risk operations, the key influencing factors of high-voltage line maintenance are the most important, and relevant countermeasures should be implemented first. For each operation, the countermeasures with the highest recommended value should be implemented first. Based on this ranking result, the control priority of each entity node can be determined, and the operation safety control of all entity nodes can be carried out accordingly.
[0064] By integrating multi-modal power field operation data into the knowledge graph, it can comprehensively reflect the various entities, attributes and relationships in the operation process, providing data and knowledge support for operation risk assessment and control. Moreover, by supplementing and optimizing the knowledge graph based on the graph attention network, it can more accurately identify key risk factors and influencing mechanisms, and can automatically learn and mine the complex relationships implicit in the knowledge graph to achieve intelligent risk assessment, so as to continuously improve the accuracy of risk grading assessment and the targeted management and control. According to the risk level and risk probability, a response measure recommendation subgraph can be constructed, which can put forward targeted management and control measures for various entity nodes, so as to determine the management and control priority of each entity node, which is helpful to reasonably allocate limited management and control resources.
[0065] Based on the above method, the embodiment of the present application discloses a work safety management and control system based on risk classification, referring to Figure 2 The operation safety management and control system 1 includes a knowledge graph construction module 11, an operation risk assessment module 12 and an operation safety management and control module 13, wherein: The knowledge graph construction module 11 is used to obtain power field operation data of multiple modes, and construct a knowledge graph according to the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes and relationship nodes; The operation risk assessment module 12 is used to supplement each node and relationship edge in the knowledge graph based on the graph attention network to obtain an operation risk assessment model, wherein the relationship edge represents the edge connecting the nodes, and the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node are input into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level; The operation safety management and control module 13 is used to extract relevant nodes and relationship edges from the knowledge graph based on risk level and risk probability, and form a countermeasure suggestion subgraph. According to the countermeasure suggestion value of the countermeasure suggestion subgraph, the risk level, entity node and countermeasure of the power field operation are sorted to obtain the control priority corresponding to each entity node, and the operation safety management and control of all entity nodes is performed based on the control priority.
[0066] In an example, entity nodes include power operation categories, power operation projects, power operation personnel, power operation equipment, and power operation environments; attribute nodes include regulations and systems for power operation categories, risk levels for power operation projects, qualifications of power operation personnel, status of power operation equipment, and conditions for the power operation environment; relationship nodes include the inclusion relationship between power operation categories and power operation projects, the allocation relationship between power operation projects and power operation personnel, the usage relationship between power operation projects and power operation equipment, and the influence relationship between power operation projects and the power operation environment.
[0067] In one example, the operation risk assessment model includes a relationship quantification layer, an entity attention network layer, a relationship attention network layer, and an entity classification layer, wherein: The relationship quantization layer is connected to the entity attention network layer. The relationship quantization layer is used to quantify the relationship between different entity nodes and output the quantized relationship values between different entity nodes to the entity attention network layer according to the quantized relationship function; The entity attention network layer is connected to the relationship attention network layer. The entity attention network layer is used to construct an entity attention network according to the attribute nodes and the quantized relationship values to obtain the normalized attention coefficients between the entity nodes, and output the updated embedding vector corresponding to each entity node to the relationship attention network layer based on the normalized attention coefficients and the vector update function; The relation attention network layer is connected to the entity classification layer. The relation attention network layer is used to construct a relation attention network according to the quantized relation values, updated embedding vectors and relation nodes, and output the adjusted attention coefficient to the entity classification layer according to the updated attention function; The entity classification layer is used to classify all entity nodes according to the adjusted attention coefficient and updated embedding vector, and obtain the risk category probability corresponding to each entity node according to the risk classification function.
[0068] In one example, the expression of the quantized relationship function is: ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The quantitative relationship value under the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, M k Indicates k A relationship matrix, k represents the number of relationship matrices, Indicated in k The mapping function under the relationship matrix is Indicates i The embedding vector corresponding to the entity node f The value of the dimension, F represents the total number of dimensions of the embedding vector, Indicates j The embedding vector corresponding to the entity node f The value of the dimension.
[0069] In one example, the expression of the vector update function is: ; ; ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The attention coefficient in the relationship matrix, represents the mapping function under the kth relationship matrix, Indicates k The weight matrix of the entity attention network layer in the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, || represents the connection symbol, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The normalized attention coefficient in the relationship matrix, LeakyReLU ( ) represents the first nonlinear activation function, s Represents and except i Any entity node other than entity nodes, N Represents the total number of entity nodes, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, s represents the second nonlinear activation function, J Indicates k All neighboring points whose relationship quantization values under the relationship matrix are greater than the relationship quantization threshold j , G represents the total number of attention units, g Indicates g Attention units, W g Indicates g The feature transformation matrix corresponding to each attention unit.
[0070] In one example, the expression for updating the attention function is: ; in, Indicated ink The adjusted attention coefficient in the relationship matrix, N Represents the total number of entity nodes, K represents the total number of relationship matrices, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, W represents the parameterized matrix, b represents the offset vector, a represents the attention mechanism vector, sigmoid ( ) represents the third nonlinear activation function.
[0071] In one example, the risk classification function is expressed as: ; ; ; in, represents the integrated feature vector set output by the operational risk assessment model, sigmoid ( ) represents the third nonlinear activation function, W C represents the risk weight, s represents the second nonlinear activation function, G represents the total number of attention units, g Indicates g Attention units, K represents the total number of relationship matrices, Indicated in K The adjusted attention coefficient in the relationship matrix, Indicated in K The relationship matrix i The updated embedding vector of entity nodes, Indicates i The entity node belongs to x The probability of risk category, D represents the total number of integrated eigenvectors in the integrated eigenvector set, d Indicates d The integrated feature vector, f d Indicates the first d The integrated feature vector, f x Represents the first risk category vector in the risk category set x risk category vectors, and C represents the total number of risk categories.
[0072] In one example, the method also includes: obtaining an entity node loss function based on the true category label of each entity node and the risk category probability of the corresponding entity node; obtaining a regularization term function based on the embedding vector of the entity node and the positive sample embedding vector of the corresponding entity node; summing the entity node loss function and the entity node loss function to obtain a total loss function, and compensating the job risk assessment model based on the total loss function.
[0073] In one example, the expression of the total loss function is ; ; ; in, L task represents the entity node loss function, N represents the total number of entity nodes, C represents the total number of risk categories, y i,c Indicates i The true category labels of entity nodes, Indicates i The entity node belongs to x The probability of risk category, L KG represents the regularization function of the knowledge graph, c represents the interval hyperparameter, d e ( ) represents the distance metric function, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, Indicates i The positive sample embedding vector of entity nodes in the knowledge graph, L is the total loss function, l represents the regularization weight coefficient.
[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for operation safety management and control based on risk classification, It is characterized in that The method comprises: Acquire power field operation data of multiple modes, and construct a knowledge graph based on the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes, and relationship nodes; Based on the graph attention network, each node and relationship edge in the knowledge graph is supplemented to obtain a job risk assessment model, wherein the relationship edge represents an edge connecting nodes; Inputting the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level; Based on the risk level and the risk probability, extract relevant nodes and relationship edges from the knowledge graph to form a countermeasure suggestion subgraph; According to the countermeasure recommendation values of the countermeasure recommendation subgraph, the risk levels, entity nodes and countermeasures of the power field operations are sorted to obtain the control priority corresponding to each entity node, and the operation safety control of all entity nodes is performed based on the control priority.
2. The method according to claim 1, It is characterized in that The entity nodes include power operation categories, power operation projects, power operation personnel, power operation equipment, and power operation environment; the attribute nodes include the regulations and systems of power operation categories, the risk levels of power operation projects, the qualifications of power operation personnel, the status of power operation equipment, and the conditions of the power operation environment; the relationship nodes include the inclusion relationship between power operation categories and power operation projects, the allocation relationship between power operation projects and power operation personnel, the usage relationship between power operation projects and power operation equipment, and the influence relationship between power operation projects and the power operation environment.
3. The method according to claim 1, It is characterized in that The operation risk assessment model includes a relationship quantification layer, an entity attention network layer, a relationship attention network layer and an entity classification layer, wherein: The relationship quantization layer is connected to the entity attention network layer, and the relationship quantization layer is used to quantify the relationship between different entity nodes, and output the quantized relationship value between different entity nodes to the entity attention network layer according to the quantized relationship function; The entity attention network layer is connected to the relationship attention network layer, and the entity attention network layer is used to construct an entity attention network according to the attribute nodes and the quantized relationship values to obtain normalized attention coefficients between entity nodes, and output an updated embedding vector corresponding to each entity node to the relationship attention network layer based on the normalized attention coefficients and the vector update function; The relationship attention network layer is connected to the entity classification layer, and the relationship attention network layer is used to construct a relationship attention network according to the quantized relationship value, the updated embedding vector and the relationship node, and output an adjusted attention coefficient to the entity classification layer according to the updated attention function; The entity classification layer is used to classify all entity nodes according to adjusting the attention coefficient and updating the embedding vector, and obtain the risk category probability corresponding to each entity node according to the risk classification function.
4. The method according to claim 3, It is characterized in that The expression of the quantized relationship function is: ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The quantitative relationship value under the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, M k Indicates k A relationship matrix, k represents the number of relationship matrices, Indicated in k The mapping function under the relationship matrix is Indicates i The embedding vector corresponding to the entity node f The value of the dimension, F represents the total number of dimensions of the embedding vector, Indicates j The embedding vector corresponding to the entity node f The value of the dimension.
5. The method according to claim 3, It is characterized in that The expression of the vector update function is: ; ; ; in, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The attention coefficient in the relationship matrix, represents the mapping function under the kth relationship matrix, Indicates k The weight matrix of the entity attention network layer in the relationship matrix, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, || represents the connection symbol, Indicates i The embedding vector corresponding to the entity node and the j The embedding vector corresponding to the entity node is in k The normalized attention coefficient in the relationship matrix, LeakyReLU ( ) represents the first nonlinear activation function, s Represents the difference between i Any entity node other than entity nodes, N Represents the total number of entity nodes, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, σ represents the second nonlinear activation function, J Indicates k All neighboring points whose relationship quantization values under the relationship matrix are greater than the relationship quantization threshold j , G represents the total number of attention units, g Indicates g Attention units, W g Indicates g The feature transformation matrix corresponding to each attention unit.
6. The method according to claim 3, It is characterized in that The expression of the updated attention function is specifically: ; in, Indicated in k The adjusted attention coefficient in the relationship matrix, N Represents the total number of entity nodes, K represents the total number of relationship matrices, Indicated in k The relationship matrix i The updated embedding vector of entity nodes, W represents the parameterized matrix, b represents the offset vector, a represents the attention mechanism vector, sigmoid ( ) represents the third nonlinear activation function.
7. The method according to claim 3, It is characterized in that The expression of the risk classification function is: ; ; ; in, represents the integrated feature vector set output by the operational risk assessment model, sigmoid ( ) represents the third nonlinear activation function, W C represents the risk weight, σ represents the second nonlinear activation function, G represents the total number of attention units, g Indicates g Attention units, K represents the total number of relationship matrices, Indicated in K The adjusted attention coefficient in the relationship matrix, Indicated in K The relationship matrix i The updated embedding vector of entity nodes, Indicates i The entity node belongs to x The probability of risk category, D represents the total number of integrated feature vectors in the integrated feature vector set, d Indicates d The integrated feature vector, f d Indicates the first d The integrated feature vector, f x Represents the first risk category vector in the risk category set x risk category vectors, and C represents the total number of risk categories.
8. The method according to claim 1, It is characterized in that The method further comprises: Based on the true category label of each entity node and the risk category probability of the corresponding entity node, the entity node loss function is obtained; Obtaining a regularization term function based on the embedding vector of the entity node and the positive sample embedding vector of the corresponding entity node; The entity node loss function and the entity node loss function are summed to obtain a total loss function, and the operation risk assessment model is compensated based on the total loss function.
9. The method according to claim 8, It is characterized in that The expression of the total loss function is: ; ; ; in, L task represents the entity node loss function, N represents the total number of entity nodes, C represents the total number of risk categories, y i,c Indicates i The true category labels of entity nodes, Indicates i The entity node belongs to x The probability of risk category, L KG represents the regularization function of the knowledge graph, γ represents the interval hyperparameter, d e ( ) represents the distance metric function, e i Indicates i The embedding vector corresponding to the entity node, e j Indicates i The embedding vector corresponding to the entity node, Indicates i The positive sample embedding vector of entity nodes in the knowledge graph, L is the total loss function, λ represents the regularization weight coefficient.
10. A work safety management and control system based on risk classification, It is characterized in that The operation safety management and control system (1) comprises a knowledge graph construction module (11), an operation risk assessment module (12) and an operation safety management and control module (13), wherein: The knowledge graph construction module (11) is used to obtain power field operation data of multiple modes, and construct a knowledge graph according to the power field operation data, wherein the knowledge graph includes entity nodes, attribute nodes and relationship nodes; The operation risk assessment module (12) is used to supplement each node and relationship edge in the knowledge graph based on the graph attention network to obtain an operation risk assessment model, wherein the relationship edge represents an edge connecting nodes, and the feature vector set corresponding to all entity nodes and the relationship matrix corresponding to each relationship node are input into the operation risk assessment model to obtain the risk level of each entity node and the risk probability corresponding to the risk level; The operation safety management module (13) is used to extract relevant nodes and relationship edges from the knowledge graph based on the risk level and the risk probability to form a countermeasure suggestion subgraph, and to sort the risk level, entity nodes and countermeasures of the power field operation according to the countermeasure suggestion value of the countermeasure suggestion subgraph to obtain the management and control priority corresponding to each entity node, and to perform operation safety management and control on all entity nodes based on the management and control priority.
Citation Information
Patent Citations
A method and system for constructing a safety knowledge graph for smart power plants
CN113254594B
Food safety event risk research and judgment method and device based on knowledge graph reasoning
CN111832922A
Medical knowledge graph relation prediction method based on graph attention mechanism
CN114610897A
Image classification robustness improvement method based on metric learning
CN114723990A
Equipment health state assessment method based on knowledge graph attention network
CN116737949A
Cited By
Sensitive data identification protection method and system based on intelligent matching
CN119577815A
Industrial field safety management and control method and system combined with multi-modal model
CN120296159A
A method and system for industrial field safety management and control combined with multimodal models
CN120296159B
Whole-process standardized closed-loop safety management and control method and system for power enterprise
CN120706879A
Knowledge graph-based risk early warning method and system for medical insurance settlement data
CN120725806A
