Knowledge graph-based power grid metadata deep semantic mining and intelligent management method and device

Abstract

The application discloses a kind of based on knowledge graph's power grid metadata deep layer semantic mining and intelligent management method and device, it is related to artificial intelligence and power system cross field, including: based on the standardization business metadata of multiple-source heterogeneous data of power grid being generated by pre-processing to cross-modal fusion framework;Through knowledge enhanced pre-training model, semantic coding is carried out and combined bidirectional causal inference network extracts implicit correlation characteristics;Metadata is decoupled into multidimensional features and based on scene perception attention mechanism dynamically allocates weight to calculate dynamic similarity;Based on graph neural network and dynamic similarity, knowledge graph is constructed and is updated incrementally by flow processing engine;Based on knowledge graph reasoning and through federated learning iterative optimization model parameters.The application improves the accuracy of implicit correlation identification and cross-scene similarity determination precision, realizes knowledge graph real-time dynamic updating and closed-loop optimization.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and power systems, specifically relating to a method and device for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs. Background Technology

[0002] With the deepening of the digital transformation of the power grid, business metadata such as equipment ledgers, maintenance work orders, fault waveforms, and voice recordings are experiencing explosive growth. Existing technologies typically employ general natural language processing models or static knowledge graph construction methods for semantic mining and knowledge management of this multi-source heterogeneous data. However, power grid operations are highly specialized and involve complex cross-domain relationships. Existing general models struggle to capture implicit causal relationships that transcend physical phenomena and electrical behaviors, such as "icing-tripping," resulting in insufficient semantic parsing depth. Furthermore, existing similarity assessment methods often use fixed feature weights, failing to dynamically adjust focus based on the needs of different business scenarios such as fault diagnosis and process optimization, leading to the failure of scenario-based assessments. In addition, static knowledge graphs lack incremental update mechanisms for power grid topology changes and real-time business flows, resulting in low knowledge reuse efficiency and difficulty in supporting the digital operation and maintenance needs of smart grids. Summary of the Invention

[0003] To address the shortcomings of existing technologies, one of the objectives of this invention is to provide a method for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs. This method enables accurate mining of deep semantics in power grid metadata, dynamic similarity analysis across scenarios, and real-time incremental updates and closed-loop optimization of the knowledge graph.

[0004] The second objective of this invention is to provide an apparatus for implementing the knowledge graph-based deep semantic mining and intelligent management method for power grid metadata.

[0005] A method for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs includes:

[0006] Preprocessing of multi-source heterogeneous data from the power grid based on a cross-modal fusion framework generates standardized business metadata;

[0007] The standardized business metadata is semantically encoded using a knowledge-enhanced pre-trained model, and latent association features are extracted using a bidirectional causal reasoning network to obtain deep semantic features.

[0008] The standardized business metadata is decoupled into multi-dimensional features, and the weights of each feature in the similarity calculation are dynamically allocated based on the scene-aware attention mechanism to calculate dynamic similarity.

[0009] A business metadata knowledge graph is constructed based on graph neural networks and the dynamic similarity, and incremental updates are performed through a stream processing engine.

[0010] Knowledge reasoning is performed based on the business metadata knowledge graph, and model parameters are iteratively optimized through a federated learning mechanism.

[0011] The method of extracting implicit association features by combining bidirectional causal reasoning networks includes: based on a historical causal chain database, calculating the conditional probability of the result event occurring by assuming that the cause event does not occur through counterfactual intervention, in order to complete the semantic logic chain of fragmented text and extract the implicit association features.

[0012] The calculation of the conditional probability of the outcome event includes: determining a weighted causal strength score based on the co-occurrence frequency of the causal event and the outcome event, and the expert confidence level. The calculation logic of the causal strength score is: adding the product of the co-occurrence frequency and the first weight coefficient to the product of the expert confidence level and the second weight coefficient; calculating the counterfactual conditional probability of the outcome event occurring under the condition that the intervention variable occurs, based on the joint probability of the outcome event and the intervention variable, and the marginal probability of the intervention variable. The calculation logic of the counterfactual conditional probability is: dividing the joint probability by the marginal probability; wherein, the intervention variable is the assumption that the causal event does not occur.

[0013] The step of decoupling the standardized business metadata into multi-dimensional features and dynamically allocating the weights of each feature in similarity calculation based on a scene-aware attention mechanism to calculate dynamic similarity includes: decoupling the standardized business metadata into semantic features, structural features, and business features; determining the current business scenario type based on the business tags of the standardized business metadata; and dynamically allocating the weights of the semantic features, structural features, and business features in similarity calculation based on the current business scenario type through the scene-aware attention mechanism to calculate the dynamic similarity.

[0014] The calculation of the dynamic similarity includes: calculating the semantic features using cosine similarity, calculating the structural features using edit distance, and calculating the business features using ontology tag matching degree; when the current business scenario type is a fault diagnosis scenario, the weight of the semantic features is not lower than a first threshold, and the first threshold is greater than or equal to 0.7; when the current business scenario type is a process optimization scenario, the weight of the structural features is not lower than a second threshold, and the second threshold is greater than or equal to 0.6.

[0015] The construction of the business metadata knowledge graph based on graph neural networks and dynamic similarity includes: defining core nodes, attribute nodes, and scene tag nodes; defining synonymous equivalence edges, causal inference edges, temporal dependency edges, and scene binding edges; and determining the edge weights of each relation edge based on the dynamic similarity and expert confidence weights to construct the business metadata knowledge graph.

[0016] The incremental update via the stream processing engine includes: accessing new business data through the stream processing engine and incrementally updating the business metadata knowledge graph; introducing a time decay function to decay the weights of the associated edges of silent nodes that have not been updated for a preset time. The calculation logic of the time decay function is: multiplying the baseline weight at the initial moment by an exponential function of the natural constant, where the exponent of the exponential function is the negative value of the product of the decay coefficient and the time interval; wherein, the decay coefficient of device nodes in the core nodes is less than the decay coefficient of work order nodes.

[0017] The preprocessing of multi-source heterogeneous data from the power grid based on the cross-modal fusion framework to generate standardized business metadata includes: converting speech data in the multi-source heterogeneous data into text data through a speech recognition model, and converting image data in the multi-source heterogeneous data into text data through an optical character recognition model; constructing a dynamic vocabulary for the power grid domain based on byte-pair encoding word segmentation and a new word discovery model, wherein the new word discovery model identifies new terms in the power grid domain based on mutual information between words; and aligning terms through knowledge distillation technology combined with expert rules and data tags to achieve terminology standardization and generate the standardized business metadata.

[0018] The knowledge reasoning based on the business metadata knowledge graph and the iterative optimization of model parameters through a federated learning mechanism include: performing relationship prediction and path reasoning on the business metadata knowledge graph based on a graph attention network to obtain knowledge reasoning results; evaluating the application effect of the knowledge reasoning results in actual business scenarios based on accuracy, efficiency and stability metrics; and aggregating local training parameters from multiple terminals through a federated averaging algorithm to iteratively optimize model parameters.

[0019] This invention also provides a knowledge graph-based device for deep semantic mining and intelligent management of power grid metadata, comprising: a preprocessing module for preprocessing multi-source heterogeneous data of the power grid based on a cross-modal fusion framework to generate standardized business metadata; a semantic mining module for semantically encoding the standardized business metadata using a knowledge-enhanced pre-trained model and extracting implicit association features by combining a bidirectional causal reasoning network to obtain deep semantic features; a similarity analysis module for decoupling the standardized business metadata into multi-dimensional features and dynamically allocating the weights of each dimension feature in similarity calculation based on a scene-aware attention mechanism to calculate dynamic similarity; a graph construction module for constructing a business metadata knowledge graph based on a graph neural network and the dynamic similarity, and incrementally updating it through a stream processing engine; and a closed-loop optimization module for performing knowledge reasoning based on the business metadata knowledge graph and iteratively optimizing model parameters through a federated learning mechanism.

[0020] The aforementioned device implements the technical functions corresponding to the method through a modular architecture, ensuring the efficient execution and implementation of deep semantic mining and intelligent management processes.

[0021] This invention discloses a method and device for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs. By using a bidirectional causal reasoning network and a counterfactual intervention mechanism, it breaks through the limitations of traditional methods that rely solely on explicit text matching. It can reconstruct a complete causal logic chain from fragmented data, significantly improving the accuracy and recall of identifying cross-professional implicit associations (such as those between environment and equipment failures).

[0022] This invention employs a multi-dimensional feature decoupling and scene-aware attention mechanism, which changes the evaluation mode of a single fixed weight. This enables the system to adaptively focus on semantic or structural features according to specific scenarios such as fault diagnosis or process optimization, effectively solving the problem of large fluctuations in similarity judgment accuracy in multiple scenarios and improving the targeting of business matching.

[0023] This invention combines a stream processing engine with a differentiated time decay function to achieve real-time incremental updates and lifecycle management of the knowledge graph. While ensuring the stability of knowledge of core devices, it eliminates outdated noise, significantly improving the efficiency of knowledge reuse and the system's response speed to dynamic changes in the power grid.

[0024] The federated learning closed-loop optimization system constructed in this invention achieves global collaborative iteration of model parameters while ensuring data privacy and security in multiple regions. This enables the system to continuously absorb business feedback from various regions and continuously enhance its adaptability and intelligence level to complex power grid business scenarios. Attached Figure Description

[0025] Figure 1This is a flowchart of the knowledge graph-based deep semantic mining and intelligent management method for power grid metadata according to an embodiment of the present invention;

[0026] Figure 2 This is a flowchart of the power grid metadata preprocessing and standardization business process according to an embodiment of the present invention;

[0027] Figure 3 This is a flowchart of deep semantic mining and causal reasoning of power grid metadata according to an embodiment of the present invention;

[0028] Figure 4 This is a flowchart illustrating the construction and dynamic updating of the power grid metadata knowledge graph according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0031] like Figure 1 As shown, this embodiment provides a method for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs. This method achieves a complete closed loop from multi-source heterogeneous data to structured knowledge and then to intelligent applications through the coordinated execution of five core steps.

[0032] In step S100, multi-source heterogeneous data from the power grid is preprocessed based on a cross-modal fusion framework to generate standardized business metadata. Specifically, power grid business data comes from a wide range of sources and has diverse formats, including structured data such as equipment ledgers, as well as unstructured data such as maintenance voice recordings, on-site infrared images, and work order texts. This step uses the cross-modal fusion framework to convert unstructured data into text form using technologies such as speech recognition and optical character recognition. Combined with a domain-dynamic thesaurus and terminology alignment mechanism, semantic ambiguity and expression differences between data from different sources are eliminated, ultimately outputting standardized business metadata with a unified format and semantically consistent meaning. This process lays a high-quality data foundation for subsequent deep semantic mining, ensuring consistency across the entire processing chain.

[0033] In step S200, standardized business metadata is semantically encoded using a knowledge-enhanced pre-trained model, and implicit correlation features are extracted using a bidirectional causal reasoning network to obtain deep semantic features. Specifically, simple text matching is insufficient to capture the complex relationships spanning physical phenomena and electrical behaviors in power grid operations. For example, equipment failures caused by environmental factors often present fragmented distributions in the text. This step first uses a pre-trained model incorporating power grid domain knowledge to map standardized metadata into high-dimensional semantic vectors, capturing explicit semantic information. Then, a bidirectional causal reasoning network is introduced to perform counterfactual deduction and completion of missing logical links in the text, thereby uncovering implicit causal relationships that are not directly stated but objectively exist. This combined approach significantly improves the completeness of semantic features and the explanatory power of the business.

[0034] In step S300, standardized business metadata is decoupled into multi-dimensional features, and the weights of each feature in similarity calculation are dynamically allocated based on a scenario-aware attention mechanism to calculate dynamic similarity. Specifically, power grid metadata contains various information such as semantic descriptions, process topology structures, and business attribute tags. Different business scenarios have significantly different levels of attention to this information. For example, fault diagnosis scenarios rely more on matching semantic descriptions, while process optimization scenarios focus more on the similarity of topology structures. This step eliminates mutual interference between features by orthogonally decoupling metadata into three independent dimensions: semantic, structural, and business. Simultaneously, it adaptively adjusts the contribution weight of each dimension in similarity calculation according to the current business scenario type, thereby achieving accurate adaptation to diverse scenarios.

[0035] In step S400, a business metadata knowledge graph is constructed based on graph neural networks and dynamic similarity, and incremental updates are performed using a stream processing engine. Specifically, to organize discrete metadata into reasonable structured knowledge, this step utilizes graph neural networks to model metadata entities and their relationships as graph nodes and edges, and uses the aforementioned dynamic similarity as an important calculation factor for edge weights, enabling the graph structure to reflect the true strength of business relationships. Simultaneously, considering the real-time nature of power grid business, a stream processing engine is used to process newly added data in real time and fuse it with the graph. A time decay mechanism is introduced to differentiate the weights of long-unupdated, silent nodes, ensuring the knowledge base remains constantly updated.

[0036] In step S500, knowledge reasoning is performed based on the business metadata knowledge graph, and model parameters are iteratively optimized through a federated learning mechanism. Specifically, the ultimate goal of knowledge graph construction is to support business applications. This step performs path reasoning and relationship prediction based on the graph structure, outputting reasoning results such as fault root cause analysis and process optimization suggestions, and applying them to actual business scenarios. Simultaneously, application effect feedback data is collected. While ensuring the privacy of data across all terminals, the federated learning mechanism aggregates local training parameters from multiple parties to globally and collaboratively optimize the upstream semantic mining, similarity analysis, and graph construction models, forming a continuous evolutionary closed loop of "mining-reasoning-application-feedback".

[0037] like Figure 3 As shown, in some embodiments, implicit correlation features are extracted by combining a bidirectional causal reasoning network. This includes: based on a historical causal chain database, calculating the conditional probability of the outcome event by assuming the causal event does not occur through counterfactual intervention, in order to complete the semantic logic chain of fragmented text and extract implicit correlation features. Specifically, power grid operation and maintenance records often exhibit highly fragmented characteristics. For example, a fault report may only record the result of "line tripping," while missing preceding causes such as "insulator aging," or the two may be scattered across different work orders and monitoring data. In order to reconstruct complete causal logic from this incomplete information, this embodiment constructs a historical causal chain database containing timestamps and confidence fields as the foundation for reasoning. This database not only stores explicit causal relationship pairs but also records the temporal context of the relationship and the credibility of expert annotations. Based on this, the bidirectional causal reasoning network does not simply perform keyword matching, but introduces causal inference theory. By simulating the counterfactual scenario of "if the cause did not occur", it quantifies the dependence of the outcome event on a specific cause. Thus, even in the absence of direct textual evidence, it can still complete the semantic logic chain with high confidence and achieve accurate extraction of implicit association features.

[0038] Furthermore, to ensure the reliability of the inference results, the strength of the association between the cause and the effect needs to be quantified before calculating the conditional probability of the outcome event. Specifically, the process of calculating the conditional probability of the outcome event includes: determining a weighted causal strength score based on the co-occurrence frequency of the cause and effect events and the expert confidence level. The calculation logic for the causal strength score is: adding the product of the co-occurrence frequency and the first weight coefficient to the product of the expert confidence level and the second weight coefficient. In this embodiment, this calculation process can be expressed by the following formula: ;in, Representative causal entity, Represents the result entity; This indicates the co-occurrence frequency of the two in historical data. This value is normalized and mapped to the [0,1] interval, reflecting the correlation at the statistical level of the data. This represents the expert confidence level, with a value range of [0,1]. It comes from the knowledge annotations of domain experts or authoritative procedure libraries and represents the certainty at the level of prior knowledge. The weighting coefficient is set to 0.7 in the preferred configuration of this embodiment. Taking the causal relationship of "insulator aging → tripping" as an example, if the co-occurrence frequency of the two is high in historical work orders but some records contain noise, and experts confirm that the causal relationship has a very high degree of physical inevitability based on the equipment operation procedures, then through the above formula, even if... Slight fluctuations due to data quality issues, higher It can still support a stable causal strength score. Setting it to 0.7 instead of higher or lower is a trade-off based on the characteristics of power grid operations: it ensures the model's sensitivity to massive amounts of operation and maintenance data (data-driven as the main approach) while retaining sufficient expert knowledge to correct biases (knowledge-driven as a secondary approach), effectively preventing spurious associations that may arise from pure data mining.

[0039] After identifying potential causal candidates, their logical necessity needs to be further verified through counterfactual reasoning. Specifically, based on the joint probability of the outcome event and the intervention variable, as well as the marginal probability of the intervention variable, the counterfactual conditional probability of the outcome event occurring given the occurrence of the intervention variable is calculated. The calculation logic for the counterfactual conditional probability is: divide the joint probability by the marginal probability; where the intervention variable is the assumed causal event that does not occur. This calculation process can be expressed by the following formula: ;in, This refers to counterfactual intervention, which logically forces a presupposed cause. This does not occur, which differs from ordinary conditional observation. It eliminates the interference of confounding factors and simulates the real causal blocking effect; Indicates the result event Intervention variables The joint probability of simultaneous occurrence; This represents the marginal probability of the intervention variable occurring. Taking "the probability of tripping when the insulator is crack-free" as an example, the system first searches the database or estimates the probability of the "line tripping" event under the intervention condition of "insulator in good condition (crack-free)". If the calculated probability is... The significantly lower probability of tripping compared to actual observations conversely proves that "insulator cracks" are the key implicit cause of this tripping, even though the original text does not directly mention the connection between cracks and tripping. Through this counterfactual probability calculation, the system can automatically complete a complete logical chain spanning multiple intermediate states, such as "environmental stress → insulator micro-cracks → flashover → tripping," greatly improving the explanatory power and accuracy of deep semantic mining.

[0040] It should be understood that although this embodiment uses a counterfactual reasoning framework based on the Do-calculus (Do operator) as a preferred implementation, this is not the only limitation of the present invention. In other optional embodiments, a structural causal model (SCM) can also be used to achieve the same technical functionality. SCM explicitly describes the functional dependencies between variables by constructing a system of structural equations, and can also support counterfactual queries and causal effect identification. Regardless of the specific mathematical tools used, as long as the core idea is to assess causal dependencies by simulating scenarios where causes are missing, thereby completing the logical chain of fragmented text, it should be considered to be covered within the scope of protection of the present invention.

[0041] In some embodiments, standardized business metadata is decoupled into multi-dimensional features, and the weights of each feature in similarity calculation are dynamically allocated based on a scene-aware attention mechanism to calculate dynamic similarity. Specifically, power grid business metadata naturally possesses multimodal and multi-granular characteristics, and a single vector representation cannot simultaneously take into account the semantic connotation of textual descriptions, the topological structure of business processes, and the classification information of equipment attributes. If directly mixed, features of different dimensions often interfere with each other, causing key information to be submerged by noise in specific scenarios. This embodiment uses feature decoupling technology to orthogonally decompose a unified metadata object into three independent feature subspaces: semantic, structural, and business, enabling each dimension to focus on expressing a specific type of information; furthermore, a scene-aware attention mechanism is introduced to adaptively adjust the contribution weight of each dimension in similarity evaluation according to the business attributes of the current task, thereby achieving precise adaptation to diversified business needs.

[0042] In the specific implementation of feature decoupling, standardized business metadata is first decoupled into semantic features, structural features, and business features. This decoupling process can be expressed by the following general formula: ;in, Represents the set of features after decoupling. , , These correspond to semantic features, structural features, and business features, respectively. Standardized business metadata for input; These are the learnable parameters for the decoupling network. The decoupling function is not a simple slicing operation, but rather uses a multi-branch neural network to selectively extract and enhance different types of information.

[0043] For semantic features It is mainly used to capture the deep semantic relationships hidden in metadata, and the extraction formula is as follows: ;in, The contextualized sentence vectors output by the knowledge-enhanced pre-trained model in Example 2 are typically 2304-dimensional; Proj is a linear projection function used to map the high-dimensional vectors to a 1024-dimensional target semantic space. This dimensionality reduction projection not only reduces the complexity of subsequent calculations, but more importantly, through supervised fine-tuning, it eliminates general language noise unrelated to power grid business in the original vectors, while retaining semantic representations with implicit cross-professional connections, such as "ice covering" and "dancing".

[0044] For structural features It is mainly used to depict the topological form of business processes or device connection relationships, and the extraction formula is: ;in, The adjacency matrix of the process topology graph is derived from the node connection relationships in the business metadata; GCN is a graph convolutional network operation. To effectively aggregate neighbor node information and preserve topological structure features, the inter-layer propagation formula is as follows: In the formula, To add the adjacency matrix after adding self-loops, Its degree matrix, For the first The learnable weight matrix of the layer, The ReLU activation function is used. Structural features are obtained through multi-layer graph convolution. It can encode temporal dependencies and hierarchical relationships such as "dispatch → maintenance → acceptance". Even if the text descriptions of two work orders are very different, as long as their flow topology is consistent, their structural features can still maintain a high degree of similarity.

[0045] For business characteristics It is mainly used to explicitly express the category attributes and domain tags of metadata, and the encoding formula is: ;in, This is a set of domain ontology tags, such as "voltage level = 220kV", "fault type = short circuit", etc. For a predefined label space, the OneHot function converts discrete labels into sparse binary vectors, ensuring the accuracy and interpretability of business attribute matching and avoiding semantic drift that may occur in continuous vector spaces.

[0046] After feature decoupling, the system determines the current business scenario type based on business tags from standardized business metadata. Scenario tags are typically derived directly from the business attribute fields of the metadata itself, or automatically identified through a lightweight classifier. Based on the current business scenario type, a scenario-aware attention mechanism dynamically allocates weights for semantic features, structural features, and business features in similarity calculation, calculating dynamic similarity. The core of this mechanism is to acknowledge the differences in feature sensitivity among different business objectives and to achieve adaptive optimization of weights through reinforcement learning or meta-learning strategies.

[0047] In calculating dynamic similarity, differentiated measurement algorithms were employed for different feature dimensions. Cosine similarity was used for semantic features, with the following formula: ;in, The semantic features of standard business metadata 'a'; The semantic features of standard business metadata b; The semantic cosine similarity is used to measure the directional consistency between two semantic vectors. In the preferred configuration of this embodiment, when the cosine similarity is greater than or equal to 0.85, the two are considered to be highly related at the semantic level. Edit distance is used to calculate structural features, and the formula is: ;in The representative will draw the picture Convert to The required basic operations (such as adding and deleting nodes, replacing edges) are represented by Cost, which is the cost of these operations. In this embodiment, if the number of difference steps is less than or equal to 2, the structures are considered similar. This metric based on transformation cost reflects the essential differences in topology more effectively than simple vector distance. The ontology label matching degree is used to calculate business features, using the following formula: ;in, The business characteristics of standard business metadata a; The business characteristics of standard business metadata b; The set of tags corresponding to the business characteristics of standard business metadata 'a'; The set of tags corresponding to the business characteristics of standard business metadata b; Semantic-related tags for the business characteristics of standard business metadata 'a'; Semantic-related tags for the business characteristics of standard business metadata b; For tags With tags Matching weights; This represents the business ontology matching degree. The formula introduces semantic relevance weights on top of the traditional Jaccard coefficient. For example, although "220kV" and "110kV" are not exactly the same, they belong to the same voltage level category and can achieve partial matching scores. In this embodiment, a tag overlap rate of 60% or more is considered as business attribute compatibility.

[0048] To enable the model to quickly adapt to weight allocation requirements in different scenarios, this embodiment uses the MAML (Model-Agnostic Meta-Learning) framework for meta-learning optimization, with the following loss function: ;in, For scene distribution; In the scene The parameters are updated via an inner loop, with an inner loop learning rate of 0.01. For the scene The loss;

[0049] The final formula for calculating the overall similarity is: ;in, The final similarity score; For semantic feature weights, For structural feature weights, The weights are assigned dynamically by reinforcement learning, representing business feature weights. ; This is the normalized result of the structural edit distance;

[0050] To more clearly illustrate the creativity and practicality of dynamic weight allocation, two typical scenarios are compared below. In the fault diagnosis scenario, the system assigns weights to semantic features. The threshold is set to be no lower than a first threshold, and this first threshold is greater than or equal to 0.7. This is because fault diagnosis relies heavily on textual understanding of phenomenon descriptions; for example, the semantic association between "white powder on the insulator surface" and "flashover discharge" is key to locating the root cause, while minor differences in process structure are relatively minor at this point. Conversely, in process optimization scenarios, the system assigns weights to structural features. The weight is set to be no lower than a second threshold, and this second threshold is greater than or equal to 0.6. This is because the core objective of process optimization is to identify redundant steps or violations, focusing on whether the sequence and connection mode between nodes are compliant and efficient. Even if the text descriptions of two work orders are quite different, as long as their topological paths show unusual similarity, they should be identified as similar processes for analysis. Through this differentiated weight constraint, this invention effectively overcomes the performance fluctuation problem of traditional fixed-weight methods in cross-scenario applications, and significantly improves the efficiency of metadata management in complex power grid business environments. It should be understood that the above weight threshold is only a preferred example, and can be fine-tuned according to specific business feedback in actual deployment. As long as the dynamic adjustment trend of "emphasizing semantics for faults and emphasizing structure for processes" is maintained, it falls within the protection scope of this invention.

[0051] like Figure 4 As shown, in some embodiments, a business metadata knowledge graph is constructed based on graph neural networks and dynamic similarity. Specifically, in order to transform discrete metadata into a computable and reasonable structured knowledge network, this embodiment adopts the RDF-STAR model idea to define the graph structure, which is formally represented as follows: Among them, the node set It contains three types of semantically defined nodes: the core node set The entities within Standardized business metadata directly derived from step S100, such as core business objects like main transformers, circuit breakers, or work order records in the equipment ledger; attribute node sets. key-value pairs The triples extracted from the knowledge-enhanced pre-trained model in step S200 are used to describe the state parameters or specifications of entities in a fine-grained manner; scene label node set Scene tags The scenario classification results derived from step S300, such as fault diagnosis and process optimization, are used to provide business context anchors for knowledge retrieval. This three-level node design not only preserves the ontological attributes of entities but also explicitly associates them with their business application scenarios, laying a structural foundation for subsequent cross-scenario reasoning.

[0052] After defining the nodes, it is necessary to further quantify the strength of the associations between them. This embodiment calculates the node association strength using a graph neural network, with the following formula: .in, and They are nodes and The feature vector, with a dimension of 2304, is obtained from the semantic encoding output in step S200; The dynamic similarity calculated in step S300 is used as prior knowledge to inject into the graph convolution process; GCN adopts a 2-layer network structure and performs symmetric normalization on the adjacency matrix to stabilize gradient propagation; The Sigmoid activation function maps association strength to the [0,1] interval. By using semantic similarity as the input feature of graph convolution instead of relying solely on topological connectivity, this mechanism can effectively identify latent association nodes that have similar textual descriptions but are physically distant, thus enhancing the semantic connectivity of the graph.

[0053] Based on this, synonymous equivalence edges, causal inference edges, temporal dependency edges, and scenario binding edges are defined. The edge weights of each edge are determined by weighting based on dynamic similarity and expert confidence, thus constructing a business metadata knowledge graph. The formula for calculating edge weights is as follows: .in, The comprehensive dynamic similarity output in step S300 represents data-driven evidence of association. Expert confidence level represents the certainty of prior knowledge in the domain; Setting it to 0.6 indicates that the system prioritizes associations discovered through data mining while retaining 40% of the expert knowledge weight to prevent spurious associations generated by purely data-driven approaches. For different types of relationship edges, Adopt a differentiated fine-tuning strategy: for synonymous equivalent edges, The value is increased to 0.7 because terminological equivalence relies more on semantic statistical regularities; for causal inference edges, the causal strength score from step S200 is introduced as... The addition of this feature gives higher weight to strong causal relationships; for time-dependent edges, the weight is dynamically adjusted based on the time difference between process nodes. The shorter the time interval, the stronger the data dependency; for scene binding edges, the weight allocation result of the scene in step S300 is dynamically adjusted to ensure that the binding relationship of high-frequency core scenes is more stable.

[0054] To address real-time changes in power grid business data, a stream processing engine is used to access newly added business data and incrementally update the business metadata knowledge graph. The specific execution logic can be represented as follows: .in, For newly connected business data streams; The entire process of preprocessing, semantic mining, and similarity calculation from steps S100 to S300 is reused to ensure that new data and existing knowledge are aligned in the same semantic space. This represents stream processing operations based on the Apache Flink framework, including node creation, edge weight updates, and subgraph merging. In the preferred configuration of this embodiment, the end-to-end latency from new data access to graph update completion is controlled within 1 minute. For high-priority data such as fault alarms, the latency can be further compressed to within 30 seconds by setting a smaller processing window, thereby ensuring that the knowledge base can reflect the latest changes in the power grid's operating status in real time.

[0055] Furthermore, a time decay function is introduced to attenuate the weights of associated edges of silent nodes that have not been updated for a preset time. The calculation logic of the time decay function is as follows: the baseline weight at the initial moment is multiplied by an exponential function of the natural constant, where the exponent of the exponential function is the negative of the product of the decay coefficient and the time interval, i.e. .in, For the edge At the initial moment The weight, This is the time interval since the last update. This refers to the attenuation coefficient. The key innovation of this embodiment lies in the use of differentiated attenuation coefficients: for core equipment nodes (such as main transformers, GIS equipment, etc.) Setting it to 0.005 / day exhibits a slow decay characteristic; however, for temporary work order nodes or short-term event nodes, Setting the attenuation rate to 0.05 / day exhibits rapid decay characteristics. The business rationale for this design lies in the fact that the full lifecycle knowledge of core equipment assets has long-tail value. Even if no new work orders are generated for a long time, its historical failure modes and maintenance experience still have important reference value for future status assessments and should not be quickly forgotten over time. Conversely, temporary maintenance records or short-term environmental events often have strong time sensitivity; retaining outdated temporary information for a long time will interfere with the judgment of the current status. Through this differentiated attenuation mechanism that conforms to the cognitive patterns of power grid business, this invention effectively suppresses the accumulation of outdated noise while ensuring the stable inheritance of core knowledge, maintaining a high signal-to-noise ratio and query efficiency of the knowledge graph. It should be understood that the specific values ​​of the above attenuation coefficients are only examples. In actual deployment, adaptive calibration can be performed according to the importance level of equipment type or the update frequency distribution of business data. As long as the overall principle of keeping the attenuation rate of core assets lower than that of temporary information is maintained, it falls within the protection scope of this invention.

[0056] like Figure 2 As shown, in some embodiments, multi-source heterogeneous data from the power grid is preprocessed based on a cross-modal fusion framework to generate standardized business metadata. Specifically, this preprocessing process is a crucial step in converting unstructured signals such as sound and images from the physical world into standardized semantic symbols that can be understood by computers. Its execution quality directly determines the upper limit of subsequent deep semantic mining. This embodiment will focus on describing the specific implementation logic and optimal parameter configuration of cross-modal conversion and terminology standardization.

[0057] In step S101, the speech data from the multi-source heterogeneous data is converted into text data using a speech recognition model. Considering the noisy and technically complex environment of power grid operation and maintenance sites, this embodiment employs an automatic speech recognition model based on the Transformer architecture. Its core computational logic can be expressed as follows: ;in, The input is the speech feature vector. The text sequence is for prediction. In the preferred configuration of this embodiment, the number of heads in the multi-head attention mechanism is set to 8, and the hidden layer dimension is set to 512. Choosing 8 heads, rather than more or fewer, is a trade-off based on the characteristics of power grid speech: 8 heads can capture semantic features of different subspaces in parallel, sufficient to distinguish between professional instructions such as "sluice gate" and "close," which have similar tones but different meanings. Furthermore, actual testing shows a 12% improvement in terminology discrimination accuracy compared to a 4-head model, while avoiding the redundant overhead of a 16-head model in terms of inference speed. The 512-dimensional hidden layer ensures the ability to capture the contextual associations of long terms such as "SF6 gas leak" while maintaining the lightweight requirement for deployment at the edge. During the training phase, the learning rate is set to 1e-4, with 30 iterations. The optimizer is Adam, and power grid environmental noise with a weight of 0.3 is introduced for mixed training. The sampling rate is locked at 16kHz. This combination of parameters ensures that the model maintains a word error rate of less than 3% even under real-world conditions such as wind noise and equipment hum, effectively supporting the accurate textification of unstructured speech data.

[0058] In step S102, image data from multi-source heterogeneous data is converted into text data using an optical character recognition model. Addressing challenges such as worn nameplates on power grid equipment and diverse handwritten characters in infrared image annotations, this embodiment employs a recognition architecture combining CNN and RNN, with the probability distribution calculation formula as follows: ;in, For input image data, To predict the probability distribution of text sequences, a ResNet50 CNN backbone was selected, leveraging its deep residual structure to effectively extract subtle features such as metal frame reflections and blurred text. The RNN portion employed a two-layer bidirectional LSTM with 128 hidden units per layer to balance contextual dependency capture and inference efficiency for long strings like "220kV Donghong Line #1 Tower". During training, the learning rate was set to 0.001, the batch size to 32, and the input image resolution was uniformly adjusted to 320×320. This parameter configuration ensured a stable character-level recognition accuracy of over 99% while keeping the processing time per image within 50ms, meeting the timeliness requirements of real-time inspection scenarios.

[0059] In step S103, a dynamic vocabulary for the power grid domain is constructed based on byte-pair encoding word segmentation and a new word discovery model. To address the rapid iteration of power grid technology and the frequent emergence of new terminology, this embodiment designs a dynamically updated vocabulary construction mechanism. First, the byte-pair encoding (BPE) algorithm is used to segment the text into sub-words, with the following probability model: ;in, For input text, For sub-word units, the preferred parameters are: minimum sub-word granularity of 2, iteration count of 2000, and minimum merging frequency threshold of 5. This configuration can both reasonably decompose compound words such as "GIS equipment" into meaningful semantic units, avoiding single-character fragmentation, and filter out low-frequency misspellings, ensuring that the vocabulary covers more than 98% of commonly used terms. Secondly, a new word discovery model based on mutual information is introduced, with the calculation formula as follows: Among them, X and Y are candidate words in the text data when new technologies and equipment are introduced; Let X be the probability of X appearing alone. Let Y be the probability of Y appearing alone; The probability of word X co-occurring with word Y is set; the co-occurrence window is set to 5 words, the mutual information threshold is 0.8, and the length of new words ranges from 2 to 6 characters. This parameter combination can accurately capture the strong correlation between emerging terms such as "integrated optical storage and charging" and their context, while eliminating the interference of ordinary phrases, so that the accuracy of new word recognition reaches more than 92%.

[0060] In step S104, terminology alignment is performed using knowledge distillation technology combined with expert rules and data tags to achieve terminology standardization and generate standardized business metadata. To address the semantic fragmentation caused by synonyms (such as "knife switch" and "isolating switch"), this embodiment constructs an alignment model that integrates expert knowledge and data statistics, with the following loss function: ;in, This is the total loss function; This is a soft-label loss, stemming from expert rules; This is hard-label loss, originating from data annotation; This is the weighting coefficient; the initial value is set to 0.6, and it is dynamically adjusted according to the maturity of the term within the range of [0.4, 0.8]: new terms are given more emphasis on expert rules ( Approaching 0.8), the focus is on data statistics for mature terms ( (Approaching 0.4). Combined with a low learning rate of 0.0001 for fine-tuning, this mechanism keeps the accuracy of new term recognition stable above 98%, ensuring the consistency of semantics across the entire data domain.

[0061] Furthermore, this embodiment establishes strict conflict resolution rules during multimodal data fusion. When the speech recognition confidence level is below 0.6, the system automatically marks the text as a low-confidence sample and performs secondary verification by associating it with contextual information such as the work order title. When conflicts occur between different modal recognition results (e.g., speech transcription is "knife switch" while OCR recognition is "isolating switch"), if the difference in confidence levels is greater than or equal to 0.2, the higher-confidence result is directly adopted; otherwise, a manual review process is triggered. This mechanism effectively ensures the reliability of standardized business metadata, providing a high-quality data foundation for subsequent knowledge graph construction. It should be understood that the above model architecture and parameters are only preferred examples adapted to the current power grid business characteristics. In practical applications, they can be adaptively adjusted according to factors such as hardware computing power and data scale. As long as the core ideas of cross-modal fusion and dynamic standardization are followed, they all fall within the protection scope of this invention.

[0062] In some embodiments, knowledge reasoning is performed based on a business metadata knowledge graph, and model parameters are iteratively optimized through a federated learning mechanism. Specifically, to enable the system to continuously evolve and adapt to the dynamic changes in power grid business, this embodiment constructs a complete closed-loop optimization system of "mining-reasoning-application-feedback". This system not only focuses on the accuracy of single reasoning but also emphasizes the global collaborative updating of model parameters driven by actual business feedback. The entire closed-loop iterative process can be expressed by the following general formula: ;in, for The set of model parameters at each time step covers the Grid-BERTv2 semantic encoding parameters, meta-learner weights, and graph neural network parameters involved in the aforementioned embodiments; The global learning rate is set to 0.001 in the preferred configuration of this embodiment to balance convergence speed and stability; , , These represent three functional modules: knowledge mining, knowledge reasoning, and knowledge application. This data is collected from application feedback in real-world business scenarios. The formula shows that each parameter update of the model stems directly from the gradient feedback of the business application's effectiveness, thus ensuring that the direction of technical optimization always aligns with actual business needs.

[0063] During the knowledge mining phase, the system continuously absorbs new data to enrich the knowledge base; its execution logic is represented as follows: .in, For newly connected standardized business metadata, for A knowledge graph of moments. For the core node set, This is the function for calculating dynamic similarity. This step is essentially an online execution instance of the incremental update mechanism, ensuring that the inference engine always runs based on the latest knowledge state.

[0064] In the knowledge reasoning stage, a graph attention network is used to perform relationship prediction and path reasoning on the business metadata knowledge graph to obtain the knowledge reasoning result. The specific calculation formula is as follows: .in, and As the core node, Given a set of candidate relation types, with probabilities The calculation was performed using a graph attention network. To meet the urgent need for root cause analysis in power grid operation and maintenance, this embodiment artificially increased the weight coefficient of the causal inference edge by 20% when calculating the attention weight. This bias strategy enables the model to prioritize exploring causal paths such as "environmental stress → equipment defect → fault tripping" during the inference process, rather than merely staying on semantically similar superficial associations, thereby significantly improving the inference recall and interpretability in fault diagnosis scenarios.

[0065] In the application effectiveness evaluation phase, the application effectiveness of knowledge reasoning results in actual business scenarios is evaluated based on accuracy, efficiency, and stability metrics. The comprehensive performance evaluation formula is as follows: .in, This indicates the consistency ratio between the reasoning result and the expert review conclusion; This indicates the percentage reduction in business process time or the reduction in the number of manual interventions after adopting the reasoning suggestions; This represents the reciprocal of the system's anomaly rate or the stability of the device's state after applying the inference results. Under the default configuration, the weighting coefficients... , , The weights are 0.5, 0.3, and 0.2 respectively, reflecting the principle of prioritizing accuracy. However, the system supports dynamic fine-tuning of the weights for different business scenarios: for example, in fault repair scenarios, where time sensitivity is extremely high, the system will... Increased to 0.4, correspondingly reduced The value is set to 0.4 to encourage models to output emergency strategies that, while not absolutely accurate, can quickly restore power; however, in long-term equipment operation and maintenance planning scenarios, the value will be... The value was increased to 0.3 to ensure the robustness of decision-making. This scenario-aware evaluation mechanism avoids the deviation in business value that may be caused by a single indicator.

[0066] Based on the above evaluation results, the system constructs a feedback loss function. To guide parameter optimization. Among them, This is a performance loss item; the worse the application effect, the greater the loss. For L2 regularization terms, The value is set to 0.001. The key reason for introducing the regularization term is that business feedback data is often sparse and noisy. If only the maximization of feedback effect is pursued, the model is prone to overfitting to individual extreme cases. By limiting the parameter norm, it is ensured that the model maintains its ability to remember general knowledge while absorbing new feedback.

[0067] Considering that power grid data involves critical infrastructure security and user privacy, raw work orders and monitoring data from each power supply bureau's terminals are strictly prohibited from leaving the domain. Therefore, a federated averaging algorithm is used to aggregate local training parameters from multiple terminals to iteratively optimize the model parameters. The parameter aggregation formula is as follows: At the execution level, this process strictly follows the principle of "data remains stationary while the model moves": First, each participating node ( Each power supply station terminal uses at least 100,000 historical work order data entries locally to fine-tune the model and obtain locally updated parameters. Subsequently, only the parameter gradients, which have undergone homomorphic encryption, are uploaded to the central aggregation server; the server then assigns weights based on the amount of data from each node. Perform a weighted average to generate new global parameters. The data is then distributed to all terminals. The entire iteration cycle is controlled within 14 days, ensuring that the model can promptly absorb the latest business experience from various regions (such as the unique icing failure mode of a certain region) while avoiding excessively frequent communication overhead. It should be understood that although this embodiment uses the FedAvg algorithm as the preferred algorithm, in other embodiments, federated learning algorithms such as FedProx or SCAFFOLD, which are better able to handle data heterogeneity, can also be used. As long as their core is to achieve multi-party collaborative parameter optimization under privacy protection, they should all be considered to be covered within the scope of protection of this invention.

[0068] This embodiment provides a knowledge graph-based deep semantic mining and intelligent management device for power grid metadata. The device adopts a software functional module architecture and is used to execute any of the methods described in the foregoing embodiments.

[0069] Specifically, the device encapsulates the complex semantic mining and knowledge management process into five highly cohesive and loosely coupled functional modules. It can be deployed as an independent software system on the central cloud server, or it can be deployed as a lightweight, customized edge computing node in a substation to adapt to the computing resources and real-time requirements of different levels of the power grid.

[0070] The preprocessing module is used to preprocess multi-source heterogeneous data from the power grid based on a cross-modal fusion framework, generating standardized business metadata. This module corresponds to the specific implementation logic of step S100. In its implementation, the preprocessing module integrates a speech recognition unit, an optical character recognition unit, and a terminology standardization unit. The speech recognition unit loads a Transformer model fine-tuned for the power grid domain to convert maintenance recordings into text; the optical character recognition unit uses a combination of ResNet50 and bidirectional LSTM networks to extract label information from equipment nameplates and infrared images; and the terminology standardization unit eliminates semantic ambiguity between multi-source data through a dynamic vocabulary and knowledge distillation alignment mechanism. Through this modular encapsulation, the preprocessing module can shield the format differences of the underlying heterogeneous data and output unified and standardized business metadata to the upper layer, laying a high-quality data foundation for subsequent in-depth data mining.

[0071] The semantic mining module is used to semantically encode standardized business metadata using a knowledge-enhanced pre-trained model and extract implicit association features using a bidirectional causal inference network to obtain deep semantic features. This module corresponds to the specific implementation logic of step S200. The semantic mining module incorporates a Grid-BERTv2 semantic encoder and a counterfactual inference engine. The semantic encoder maps standardized text into a high-dimensional vector representation containing domain knowledge; the counterfactual inference engine, based on a historical causal chain database, automatically completes the missing logical links in fragmented text by simulating conditional probability calculations under scenarios where causes are missing. The output of this module not only includes explicit semantic vectors but also implicit association features verified by causal inference, significantly improving the downstream tasks' ability to understand complex business logic.

[0072] The similarity analysis module decouples standardized business metadata into multi-dimensional features and dynamically allocates the weights of each feature dimension in similarity calculation based on a scene-aware attention mechanism, thus calculating dynamic similarity. This module corresponds to the specific implementation logic of step S300. The similarity analysis module includes a feature decoupling submodule and a scene-adaptive weight allocation submodule. The feature decoupling submodule orthogonally decomposes the input metadata into three independent feature spaces: semantic, structural, and business. The scene-adaptive weight allocation submodule dynamically adjusts the contribution ratio of each feature dimension in the comprehensive similarity calculation based on the business tags of the current metadata (such as fault diagnosis or process optimization) through a reinforcement learning strategy. This design enables the device to automatically adapt to diverse business evaluation needs without retraining the model, solving the problem of large accuracy fluctuations in traditional fixed-weight methods when applied across scenarios.

[0073] The graph construction module is used to build a business metadata knowledge graph based on graph neural networks and dynamic similarity, and to perform incremental updates through a stream processing engine. This module corresponds to the specific implementation logic of step S400. The graph construction module integrates a graph structure generator and a streaming update engine. The graph structure generator uses graph neural networks to aggregate node features and dynamic similarity to build a structured knowledge network containing core nodes, attribute nodes, scene label nodes, and four types of relationship edges; the streaming update engine interfaces with real-time computing frameworks such as Apache Flink to perform millisecond-level processing and graph fusion on newly accessed data, and execute differentiated time decay strategies. This module ensures that the knowledge base can reflect the latest changes in the power grid's operating status in real time, and maintains a balance between the long-term stability of core equipment knowledge and the timeliness of temporary information through differentiated decay mechanisms.

[0074] The closed-loop optimization module is used for knowledge reasoning based on the business metadata knowledge graph and iteratively optimizes model parameters through a federated learning mechanism. This module corresponds to the specific implementation logic of step S500. The closed-loop optimization module includes an inference engine, an effect evaluator, and a federated aggregator. The inference engine outputs fault root causes or process optimization suggestions based on a graph attention network; the effect evaluator quantifies application feedback from three dimensions: accuracy, efficiency, and stability; and the federated aggregator, while ensuring that the data privacy of each terminal does not leave the domain, aggregates local training parameters from multiple parties through the FedAvg algorithm, driving the continuous evolution of the models of each upstream module. Through this closed-loop mechanism, the device possesses self-learning and adaptive capabilities, and can continuously improve its intelligence level as the power grid business develops.

[0075] It should be understood that the above module division is only a logical functional definition. In actual engineering implementation, the modules can be merged, split, or deployed in a distributed manner according to hardware resources and application scenarios. For example, in a lightweight deployment scenario on the edge side, the preprocessing module and the semantic mining module can be merged into a compact inference service, while the federated aggregator in the closed-loop optimization module can be kept on the central cloud for execution. As long as the functions performed by each module cover the core steps of the method of this invention, regardless of how their physical deployment form changes, they should be considered to be covered within the protection scope of this invention.

[0076] This embodiment provides a comprehensive application scenario for power grid fault diagnosis and process optimization. It aims to intuitively demonstrate the execution logic and technical effects of the present invention in actual power grid operation and maintenance by connecting the technical features of the aforementioned embodiments through a real business flow. Specifically, this embodiment uses the entire process of a power supply bureau handling a "220kV line icing trip" fault as an example to demonstrate the complete closed loop from multi-source data access to knowledge reasoning feedback.

[0077] In the data preprocessing stage of step S100, the system receives voice repair reports from on-site maintenance personnel. Based on a cross-modal fusion framework, the system preprocesses the multi-source heterogeneous data of the power grid to generate standardized business metadata. Specifically, when maintenance personnel report "220 kV line icing trip" via mobile terminal, the system calls the automatic speech recognition model for processing. According to the formula... The input speech feature vector After being encoded and decoded using an 8-head attention mechanism and a 512-dimensional hidden layer, it is precisely converted into a text sequence. In this process, the model not only completed the transcription from speech to text, but also combined with a dynamic vocabulary for the power grid field to automatically standardize the colloquial "220 kV" into the standardized term "220 kV" in the equipment ledger, and identified "ice-induced tripping" as a composite fault event entity. Finally, it output the standardized business metadata text "220 kV line ice-induced tripping", which eliminated the noise interference of unstructured data for subsequent semantic mining.

[0078] In the deep semantic mining stage of step S200, to address potential logical gaps in the standardized text, a knowledge-enhanced pre-trained model is used to semantically encode the standardized business metadata, and a bidirectional causal reasoning network is combined to extract implicit association features, resulting in deep semantic features. In this scenario, although the text mentions "icing" and "tripping," it lacks a description of the intermediate physical process. The system first uses a causal strength scoring formula to quantify the association between the two: Substitute the actual parameters, and the cause entity. For "ice covering", the resulting entity The term is "power outage". Assume the historical data is normalized to reflect the co-occurrence frequency. Expert confidence level Weighting coefficient The causal strength score is then calculated as follows: This high score indicates a strong potential causal relationship between the two. To further verify and complete the logical chain, the system performs counterfactual reasoning to calculate the tripping probability under the assumption that icing does not occur: Based on historical database statistics, the joint probability of this line tripping under non-icing weather conditions. The marginal probability is only 0.05, while the marginal probability without icing conditions is... The value is 0.8, and the calculated counterfactual conditional probability is 0.0625. Compared to the actual observed tripping probability of 0.85 under icing weather conditions, this significant difference confirms that "icing" is the key trigger for this tripping. Based on this, the system automatically completes the implicit causal logic chain of "icing → conductor galloping → mechanical stress overload → tripping," solving the problem that traditional text matching cannot capture the intermediate physical mechanism.

[0079] In the similarity analysis phase of step S300, to quickly locate historical handling experience, standardized business metadata is decoupled into multi-dimensional features, and the weights of each feature in the similarity calculation are dynamically allocated based on a scene-aware attention mechanism to calculate dynamic similarity. The system automatically identifies the current scenario as "fault diagnosis" based on the business tags in the metadata. According to the strategy, the weight of semantic features is dynamically elevated to a dominant position in this scenario. The specific calculation uses a comprehensive similarity formula: In this scenario, the reinforcement learning agent assigns semantic weights. The allocation is 0.75 (satisfies the requirement). (constraints), structural weights Set to 0.15, business weight The weight was set to 0.1. This weight configuration allows the system to focus highly on matching semantic descriptions such as "icing," "galloping," and "tripping" when retrieving historical cases, while appropriately downplaying the impact of differences in process topology. Calculations showed that the current fault had a comprehensive similarity of 0.92 with a "220kV North Line Icing and Galloping Trip" case from three years ago, successfully matching historical knowledge including effective handling measures such as "installing anti-galloping spikes" and "DC de-icing," thus avoiding missed detections due to minor differences in process nodes.

[0080] In the knowledge graph construction and update phase of step S400, a business metadata knowledge graph is constructed based on graph neural networks and dynamic similarity, and incremental updates are performed through a stream processing engine. New nodes and edges generated during this fault are written to the graph in real time. Simultaneously, the system introduces a time decay function to dynamically manage existing knowledge. For the core equipment nodes of the "220kV line" involved in this incident, the attenuation coefficient The value is set to 0.005 / day to ensure that its operation and maintenance knowledge, spanning up to ten years, is preserved long-term; while for the associated "temporary emergency repair work order" node, The value is set to 0.05 / day, which allows the weight of expired temporary dispatch instructions to naturally decrease after several months. This differentiated attenuation mechanism ensures that the fault diagnosis can access the anti-icing modification records of the line from ten years ago (the long-tail value of core knowledge), while avoiding the interference of outdated information from a temporary power supply task five years ago with current decision-making, thus maintaining a high signal-to-noise ratio in the spectrum.

[0081] In the closed-loop optimization phase of step S500, knowledge reasoning is performed based on the business metadata knowledge graph, and model parameters are iteratively optimized through a federated learning mechanism. Based on the matched historical cases and the current graph status, the system infers disposal suggestions of "activating the DC de-icing device" and "installing interphase spacers." The application effect is evaluated using the following formula: In this emergency repair scenario, the system will assign weights to efficiency metrics. Dynamically adjusted to 0.4, accuracy weight. Adjusted to 0.4, stability weight The efficiency score remained at 0.2. In practical applications, this suggestion reduced fault handling time by 30% (Eff=0.8), the reasoning conclusion was consistent with expert review (Acc=0.9), and no secondary anomalies were triggered (Stab=0.95), resulting in an overall performance score of 0.89. This feedback data was then used in the iterative optimization of global model parameters through a federated learning mechanism, while protecting local data privacy, enabling the system to output the optimal strategy more quickly when facing similar icing faults in the future.

[0082] As demonstrated in the above full-process demonstration, this invention exhibits significant advantages in handling complex power grid faults compared to traditional methods based on keyword matching or static rules. On one hand, by supplementing implicit causal chains through counterfactual reasoning, the implicit association identification rate is increased from less than 60% in traditional methods to over 92%, effectively solving the problem of cross-disciplinary semantic gaps. On the other hand, the scene-aware dynamic weighting and differentiated graph update mechanism improves fault diagnosis response speed and knowledge reuse efficiency by over 60%, truly realizing deep semantic mining and intelligent full lifecycle management of power grid metadata. It should be understood that this embodiment is merely a typical application example of the technical solution of this invention. In actual deployment, the specific fault types, parameter values, and business processes can be adjusted according to the actual needs of different power supply bureaus. As long as they adhere to the core technical concept of this invention, they should all be covered within the protection scope of this invention.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0084] It should be specifically noted that the specific numerical parameters (such as weight coefficients, thresholds, decay rates, number of model layers and dimensions, etc.), specific algorithm models (such as counterfactual reasoning frameworks, meta-learning strategies, federated aggregation algorithms, etc.) and module division methods mentioned in the foregoing embodiments are all preferred examples listed to clearly and completely disclose the technical solution of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the core technical concept of the present invention, equivalent transformations, simple substitutions, adjustments to the order of steps, merging or splitting of module functions based on the teachings of this specification, and differentiated deployments on different computing power nodes (such as cloud, edge, and terminal devices) should all be considered to be covered within the scope of protection of the present invention. Those skilled in the art should understand that as long as the technical solution substantially achieves the technical effects of deep semantic mining of power grid metadata based on knowledge graphs, cross-scenario dynamic similarity analysis, and incremental updates and closed-loop optimization, regardless of its specific implementation form, it falls within the scope of protection claimed by the claims of the present invention.

Claims

1. A method for deep semantic mining and intelligent management of power grid metadata based on knowledge graphs, characterized in that, include: Preprocessing of multi-source heterogeneous data from the power grid based on a cross-modal fusion framework generates standardized business metadata; The standardized business metadata is semantically encoded using a knowledge-enhanced pre-trained model, and latent association features are extracted using a bidirectional causal reasoning network to obtain deep semantic features. The standardized business metadata is decoupled into multi-dimensional features, and the weights of each feature in the similarity calculation are dynamically allocated based on the scene-aware attention mechanism to calculate dynamic similarity. A business metadata knowledge graph is constructed based on graph neural networks and dynamic similarity, and incremental updates are performed through a stream processing engine. Knowledge reasoning is performed based on the business metadata knowledge graph, and model parameters are iteratively optimized through a federated learning mechanism.

2. The method according to claim 1, characterized in that, The extraction of latent association features using a bidirectional causal reasoning network includes: Based on a historical causal chain database, by assuming the causal event does not occur through counterfactual intervention, the conditional probability of the result event occurring is calculated to complete the semantic logic chain of fragmented text and extract the implicit association features.

3. The method according to claim 2, characterized in that, The calculated conditional probability of the event occurring includes: Based on the co-occurrence frequency of the causal event and the result event, as well as the expert confidence level, a weighted causal strength score is determined. The calculation logic of the causal strength score is as follows: the product of the co-occurrence frequency and the first weight coefficient is added to the product of the expert confidence level and the second weight coefficient. Based on the joint probability of the outcome event and the intervention variable, and the marginal probability of the intervention variable, the counterfactual conditional probability of the outcome event occurring under the condition that the intervention variable occurs is calculated. The calculation logic of the counterfactual conditional probability is: dividing the joint probability by the marginal probability; wherein, the intervention variable is the assumption that the causal event does not occur.

4. The method according to claim 1, characterized in that, The process of decoupling the standardized business metadata into multi-dimensional features and dynamically allocating the weights of each feature in similarity calculation based on a scene-aware attention mechanism to calculate dynamic similarity includes: The standardized business metadata is decoupled into semantic features, structural features, and business features; The current business scenario type is determined based on the business tags derived from the standardized business metadata. Based on the current business scenario type, the semantic features, structural features, and business features are dynamically assigned weights in similarity calculation through the scenario-aware attention mechanism, and the dynamic similarity is calculated.

5. The method according to claim 4, characterized in that, The calculation of the dynamic similarity includes: The semantic features are calculated using cosine similarity, the structural features are calculated using edit distance, and the business features are calculated using ontology tag matching degree. When the current business scenario type is a fault diagnosis scenario, the weight of the semantic feature is not less than a first threshold, and the first threshold is greater than or equal to 0.

7. When the current business scenario type is a process optimization scenario, the weight of the structural feature is not lower than the second threshold, and the second threshold is greater than or equal to 0.

6.

6. The method according to claim 1, characterized in that, The construction of the business metadata knowledge graph based on graph neural networks and dynamic similarity includes: Define the core node, attribute node, and scene label node; Define synonymous equivalence edges, causal inference edges, temporal dependency edges, and scenario binding edges; The edge weights of each relation edge are determined by weighting the dynamic similarity and expert confidence, and the business metadata knowledge graph is constructed.

7. The method according to claim 6, characterized in that, The incremental update via the stream processing engine includes: New business data is accessed through the stream processing engine, and the business metadata knowledge graph is incrementally updated. A time decay function is introduced to decay the weights of the associated edges of silent nodes that have not been updated for a preset time. The calculation logic of the time decay function is as follows: multiply the base weight at the initial time by an exponential function of the natural constant, where the exponent of the exponential function is the negative value of the product of the decay coefficient and the time interval. Among them, the attenuation coefficient of the device type nodes in the core nodes is smaller than that of the work order type nodes.

8. The method according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous power grid data based on the cross-modal fusion framework generates standardized business metadata, including: The speech data in the multi-source heterogeneous data is converted into text data through a speech recognition model, and the image data in the multi-source heterogeneous data is converted into text data through an optical character recognition model. A dynamic vocabulary for the power grid field is constructed based on byte-pair encoding word segmentation and a new word discovery model. The new word discovery model identifies new terms in the power grid field based on mutual information between words. By combining knowledge distillation technology with expert rules and data tags to align terms, terminology standardization is achieved, generating the standardized business metadata.

9. The method according to claim 1, characterized in that, The knowledge reasoning based on the business metadata knowledge graph, and the iterative optimization of model parameters through a federated learning mechanism, includes: Based on graph attention networks, relation prediction and path reasoning are performed on the business metadata knowledge graph to obtain knowledge reasoning results; The application effect of the knowledge reasoning results in actual business scenarios is evaluated based on accuracy, efficiency and stability metrics. The model parameters are iteratively optimized by aggregating local training parameters from multiple terminals using a federated averaging algorithm.

10. An apparatus for implementing the knowledge graph-based deep semantic mining and intelligent management method for power grid metadata as described in any one of claims 1 to 9, characterized in that, include: The preprocessing module is used to preprocess multi-source heterogeneous data of the power grid based on the cross-modal fusion framework to generate standardized business metadata. The semantic mining module is used to semantically encode the standardized business metadata through a knowledge-enhanced pre-trained model and extract implicit association features by combining a bidirectional causal reasoning network to obtain deep semantic features. The similarity analysis module is used to decouple the standardized business metadata into multi-dimensional features, and dynamically allocate the weights of each dimension feature in the similarity calculation based on the scene-aware attention mechanism to calculate dynamic similarity. The graph construction module is used to construct a business metadata knowledge graph based on graph neural networks and the dynamic similarity, and to perform incremental updates through a stream processing engine. The closed-loop optimization module is used to perform knowledge reasoning based on the business metadata knowledge graph and to iteratively optimize the model parameters through a federated learning mechanism.

Images