Power grid standard analysis method, device and equipment and storage medium

Abstract

The power grid standard analysis method, device, equipment and storage medium provided by the application relate to the field of power systems. The method comprises the following steps: obtaining a power grid standard document, wherein the power grid standard document comprises standard text clauses, circuit diagram images and parameter table data; performing semantic labeling on the standard text clauses to identify power terminology entities; performing disambiguation processing on the power terminology entities to obtain disambiguated standard text clauses; generating standardized knowledge units based on the disambiguated standard text clauses, the circuit diagram images and the parameter table data; and performing logical reasoning on the standardized knowledge units to determine a content consistency checking result of the power grid standard document. Through the application, the drawbacks of low efficiency, strong subjectivity and unstable quality of manual analysis are overcome, and efficient, objective and standardized processing of the power grid standard is realized.

Description

Technical Field

[0001] This application relates to the field of power systems, and more particularly to a power grid standard analysis method, apparatus, equipment, and storage medium. Background Technology

[0002] Power grid standards are the technological cornerstone supporting the safe, stable, and economical operation of power systems, covering all aspects from planning and design, equipment manufacturing, engineering construction, dispatching, operation, and maintenance. Accurate and efficient parsing and structured processing of power grid standard documents are crucial for improving the level of intelligent power grid management and ensuring the security of the energy network.

[0003] Currently, the industry's interpretation of power grid standards mainly relies on traditional manual methods. This process typically involves technical experts with deep expertise manually reading and understanding the standard text, extracting key information such as terminology definitions, technical parameters, logical relationships, constraints, and normative clauses, and then organizing it into a structured format for reference or subsequent use. This method not only results in an inefficient and time-consuming interpretation process, but also makes the quality of the interpretation susceptible to the influence of individual experience and subjective judgment. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, and storage medium for analyzing power grid standards, in order to overcome the drawbacks of low efficiency, strong subjectivity, and unstable quality of manual analysis, so as to achieve efficient, objective, and standardized processing of power grid standards.

[0005] Firstly, this application provides a method for resolving power grid standards, including:

[0006] Obtain power grid standard documents, which include standard text clauses, circuit diagrams, and parameter tables;

[0007] Semantic annotation is performed on standard text clauses to identify electrical terminology entities;

[0008] Disambiguation of electrical terminology entities is performed to obtain the disambiguated standard text clauses.

[0009] Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, standardized knowledge units are generated.

[0010] Logical reasoning is performed on standardized knowledge units to determine the consistency verification results of the power grid standard documents.

[0011] In one possible implementation, standardized knowledge units are generated based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, including:

[0012] Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, a multimodal knowledge graph containing multiple nodes is constructed.

[0013] Embedded representation learning is performed on the nodes of the multimodal knowledge graph to obtain the feature vectors of each node;

[0014] Based on the similarity between feature vectors, nodes are associated and mapped to generate standardized knowledge units.

[0015] In one possible implementation, embedding representation learning is performed on the nodes of the multimodal knowledge graph to obtain the feature vector of each node, including:

[0016] Extract the node features corresponding to the node;

[0017] Multimodal feature alignment is performed on the node features to obtain the aligned features;

[0018] The aligned features are fused to obtain the feature vectors of each node.

[0019] In one possible implementation, the standard text terms are semantically annotated to identify electrical term entities, including:

[0020] Calculate the vector representation of term tags in standard text clauses;

[0021] Extract features from the vector representation to obtain vector features;

[0022] The vector features are sampled twice to obtain the target vector features;

[0023] The target vector features are input into the conditional model to obtain power terminology;

[0024] Extract phrase information from electrical terminology to obtain electrical term entities.

[0025] In one possible implementation, the electrical terminology entities are disambiguated to obtain disambiguated standard text clauses, including:

[0026] Power term entities are mapped to a set of candidate terms in a pre-defined database to calculate the similarity between power term entities and candidate terms in the set of candidate terms.

[0027] Based on the similarity, the disambiguated standard text clauses are obtained.

[0028] In one possible implementation, obtaining power grid standard documents includes:

[0029] Obtain multimodal power grid standard documents containing standard text, circuit diagrams, and parameter tables;

[0030] Extract the standard text clauses, initial circuit diagram images, and initial parameter table data from the multimodal power grid standard document;

[0031] A pre-trained neural network is used to predict the connection relationship between the standard text clauses and the initial circuit diagram image, generating a set of candidate connections between the nodes of the standard text clauses and the initial circuit diagram image.

[0032] Optimize neural networks using the link prediction loss function;

[0033] The candidate connection set is filtered based on the optimized neural network to determine the semantic mapping relationship between the standard text terms and the nodes of the initial circuit diagram image;

[0034] Based on semantic mapping relationships, the standard text terms, circuit diagram images, and parameter table data are determined from the standard text terms, initial circuit diagram images, and initial parameter table data.

[0035] In one possible implementation, logical reasoning is performed on standardized knowledge units to determine the content consistency verification result of the power grid standard document, including:

[0036] Based on standardized knowledge units, determine the ambiguity vector;

[0037] A cross-attention mechanism is used to calculate the similarity of entity relationships in standardized knowledge units, resulting in entity pair vectors.

[0038] Based on the ambiguous vector and the entity pair vector, the disambiguation vector is obtained;

[0039] The disambiguation vector and the syntactic tree vector are fused to generate a fused vector. The syntactic tree vector is obtained by tree encoding the disambiguated standard text clauses.

[0040] Pre-mapped fusion vectors yield pre-mapped vectors;

[0041] Input the fused vector and pre-mapped vector into the rule-based reasoning model to generate and output the content consistency verification result.

[0042] Secondly, this application provides a power grid standard analysis device, comprising:

[0043] The acquisition module is used to acquire power grid standard documents, which include standard text clauses, circuit diagrams, and parameter table data.

[0044] The identification module is used to perform semantic annotation on standard text clauses to identify electrical terminology entities;

[0045] The determination module is used to disambiguate electrical terminology entities to obtain disambiguated standard text clauses.

[0046] The generation module is used to generate standardized knowledge units based on the disambiguated standard text clauses, circuit diagram images, and parameter table data.

[0047] The determination module is also used to perform logical reasoning on standardized knowledge units and determine the content consistency verification results of power grid standard documents.

[0048] In one possible implementation, the generation module is specifically used for:

[0049] Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, a multimodal knowledge graph containing multiple nodes is constructed.

[0050] Embedded representation learning is performed on the nodes of the multimodal knowledge graph to obtain the feature vectors of each node;

[0051] Based on the similarity between feature vectors, nodes are associated and mapped to generate standardized knowledge units.

[0052] In one possible implementation, the determining module is specifically used for:

[0053] Extract the node features corresponding to the node;

[0054] Multimodal feature alignment is performed on the node features to obtain the aligned features;

[0055] The aligned features are fused to obtain the feature vectors of each node.

[0056] In one possible implementation, the identification module is specifically used for:

[0057] Calculate the vector representation of term tags in standard text clauses;

[0058] Extract features from the vector representation to obtain vector features;

[0059] The vector features are sampled twice to obtain the target vector features;

[0060] The target vector features are input into the conditional model to obtain power terminology;

[0061] Extract phrase information from electrical terminology to obtain electrical term entities.

[0062] In one possible implementation, the determining module is specifically used for:

[0063] Power term entities are mapped to a set of candidate terms in a pre-defined database to calculate the similarity between power term entities and candidate terms in the set of candidate terms.

[0064] Based on the similarity, the disambiguated standard text clauses are obtained.

[0065] In one possible implementation, the acquisition module is specifically used for:

[0066] Obtain multimodal power grid standard documents containing standard text, circuit diagrams, and parameter tables;

[0067] Extract the standard text clauses, initial circuit diagram images, and initial parameter table data from the multimodal power grid standard document;

[0068] A pre-trained neural network is used to predict the connection relationship between the standard text clauses and the initial circuit diagram image, generating a set of candidate connections between the nodes of the standard text clauses and the initial circuit diagram image.

[0069] Optimize neural networks using the link prediction loss function;

[0070] The candidate connection set is filtered based on the optimized neural network to determine the semantic mapping relationship between the standard text terms and the nodes of the initial circuit diagram image;

[0071] Based on semantic mapping relationships, the standard text terms, circuit diagram images, and parameter table data are determined from the standard text terms, initial circuit diagram images, and initial parameter table data.

[0072] In one possible implementation, the determining module is specifically used for:

[0073] Based on standardized knowledge units, determine the ambiguity vector;

[0074] A cross-attention mechanism is used to calculate the similarity of entity relationships in standardized knowledge units, resulting in entity pair vectors.

[0075] Based on the ambiguous vector and the entity pair vector, the disambiguation vector is obtained;

[0076] The disambiguation vector and the syntactic tree vector are fused to generate a fused vector. The syntactic tree vector is obtained by tree encoding the disambiguated standard text clauses.

[0077] Pre-mapped fusion vectors yield pre-mapped vectors;

[0078] Input the fused vector and pre-mapped vector into the rule-based reasoning model to generate and output the content consistency verification result.

[0079] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0080] The memory stores the instructions that the computer executes;

[0081] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0082] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0083] Fifthly, this application provides a computer program product, including a computer program that, when executed, implements the first aspect and / or various possible implementations of the first aspect.

[0084] This application provides a method, apparatus, equipment, and storage medium for parsing power grid standards, relating to the field of power systems. The method includes: acquiring a power grid standard document, which includes standard text clauses, circuit diagram images, and parameter table data; semantically annotating the standard text clauses to identify power term entities; disambiguating the power term entities to obtain disambiguated standard text clauses; generating standardized knowledge units based on the disambiguated standard text clauses, circuit diagram images, and parameter table data; and performing logical reasoning on the standardized knowledge units to determine the content consistency verification result of the power grid standard document. This application, by acquiring standard text clauses, circuit diagram images, and parameter table data, provides a multimodal data foundation for the parsing process, ensuring the comprehensiveness of information sources and overcoming parsing biases caused by missing or incomplete data in traditional manual methods. It converts power terminology in the standard text clauses into computer-processable numerical forms, realizing the digital representation of power terminology, facilitating subsequent automated processing, and reducing the burden of manual reading and understanding. Disambiguation processing of electrical terminology entities enables the automatic and accurate identification and unification of different meanings of the same electrical term or the differentiation of similar meanings among different electrical terms. This process effectively overcomes the inconsistencies and subjectivity caused by traditional methods that rely entirely on expert personal experience. Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, standardized knowledge units are generated, achieving a dynamic association between text semantics and circuit diagram topology, breaking down information silos. Finally, logical reasoning is performed on the standardized knowledge units to determine the content consistency verification results of the power grid standard document. Thus, this method replaces traditional manual parsing with automated steps, significantly improving the efficiency of power grid standard parsing and ensuring the objectivity and reliability of the parsing results. Attached Figure Description

[0085] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0086] Figure 1 A flowchart illustrating the power grid standard parsing method provided in this application embodiment. Figure 1 ;

[0087] Figure 2 This application provides an intelligent analysis system for power grid standards based on a multimodal knowledge graph.

[0088] Figure 3 This is a schematic diagram of the structure of the power grid standard analysis device provided in the embodiments of this application;

[0089] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0090] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0091] In the power system field, power grid standards are the core guidelines for the design, construction, operation, and maintenance of power equipment. With the rapid development of new power systems, the number of power grid standards is growing exponentially, and their content is becoming increasingly complex, containing a large amount of technical terminology, technical parameters, specifications, and multimodal information. Traditional power grid standard analysis mainly relies on manual reading and annotation, which is inefficient, subjective, and prone to errors. For example, in substation design, engineers need to handle dozens of standards simultaneously, such as the "Design Code for 110kV and Above Substations" and the "Technical Standard for Power Transformers." Manually verifying the matching of standard clauses with circuit diagram parameters is time-consuming and labor-intensive, and it is difficult to detect logical contradictions or parameter inconsistencies between clauses. Furthermore, the multimodal nature of power grid standards further exacerbates the difficulty of analysis. For example, the statement in the standard text that "the rated voltage of the transformer should be 110kV" needs to be verified for consistency with the voltage parameters in the circuit diagram and the values ​​in the parameter tables. Manual operation is prone to misjudgment due to fragmented information.

[0092] To address the aforementioned issues, this application provides a method for parsing power grid standards. This method involves acquiring a power grid standard document, which includes standard text clauses, circuit diagrams, and parameter tables; semantically annotating the standard text clauses to identify electrical terminology entities; disambiguating the electrical terminology entities to obtain disambiguated standard text clauses; generating standardized knowledge units based on the disambiguated standard text clauses, circuit diagrams, and parameter tables; and performing logical reasoning on the standardized knowledge units to determine the content consistency verification result of the power grid standard document.

[0093] This application applies to standard compliance verification scenarios in power systems, such as substation design, new energy grid connection, and smart grid operation and maintenance.

[0094] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0095] The execution entity of the power grid standard parsing method provided in this application embodiment can be a computing device such as a server or server cluster. The server can be a mobile phone, computer, tablet, or other device. This application embodiment does not impose any particular restrictions on the implementation method of the execution entity.

[0096] Figure 1 A flowchart illustrating the power grid standard parsing method provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0097] S101. Obtain the power grid standard documents, which include standard text clauses, circuit diagrams, and parameter table data.

[0098] This step is the initial and fundamental step in the intelligent analysis process of power grid standards, namely the raw data acquisition stage. The core of this step lies in systematically collecting all the raw materials required for subsequent analysis.

[0099] Specifically, acquiring power grid standard documents refers to collecting electronic versions of target documents from standards management agencies, internal enterprise knowledge bases, or related platforms. These documents are not of a single type, but rather a multimodal, structured collection of information containing standard text clauses, circuit diagrams, and parameter tables.

[0100] Among them, standard text clauses can be used to specify technical requirements and standard clauses in text and tables, circuit diagram images are used to describe electrical connections, equipment layouts, schematic diagrams or engineering drawings, and parameter table data refers to the specific technical data such as voltage, current, and capacity marked on each component and line in the circuit diagram image.

[0101] This step ensures that subsequent parsing algorithms have a complete and accurate input source, which is the cornerstone for realizing the transformation from unstructured documents to structured knowledge.

[0102] S102. Perform semantic annotation on standard text clauses to identify electrical terminology entities.

[0103] This step aims to perform deep semantic analysis and digital representation of the acquired standard text clauses. First, key electrical terminology entities, such as equipment names (e.g., transformers, circuit breakers), technical parameters (e.g., rated voltage, short-circuit current), and status descriptions (e.g., normal operation, fault isolation), are automatically identified and extracted from the unstructured standard text clauses. Then, a vector representation of each identified electrical terminology entity is calculated. This process transforms discrete textual terms into continuous, high-dimensional numerical vectors, thereby capturing the deep meaning and relationships of electrical terms in the semantic space.

[0104] S103. Perform disambiguation processing on the electrical terminology entities to obtain the disambiguated standard text clauses.

[0105] Because polysemy and the repetition of the same meaning in words are common in natural language and standard text, computing devices may misunderstand electrical terminology, severely impacting the accuracy and reliability of subsequent electrical standard parsing. Existing methods typically employ simple dictionary matching or rule-based shallow semantic analysis to address this issue, but these approaches fail to combine context and domain knowledge for accurate disambiguation, still resulting in a high error rate in electrical terminology recognition.

[0106] To address this issue, the embodiments of this application quantify the semantic differences of electrical terms in different contexts, thereby automatically and accurately identifying and unifying different meanings of the same electrical term or distinguishing similar meanings of different electrical terms. This process effectively overcomes the inconsistency and subjectivity caused by relying entirely on the personal experience of experts in traditional methods.

[0107] S104. Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, generate standardized knowledge units.

[0108] This step involves constructing a standardized engineering knowledge system that can be deeply understood and processed by computers. It involves determining and confirming whether information extracted from disambiguated standard text clauses, circuit diagrams, and parameter tables points to the same entity or concept. For example, it associates the protection circuits described in the standard text clauses, the corresponding modules in the circuit diagrams, and the device specifications in the parameter tables as the same object. Finally, this associated information is encapsulated into structured standardized knowledge units. Each standardized knowledge unit, centered on an entity, integrates its type, composition, functional description, quantitative attributes, and system relationships to form standardized knowledge nodes that can be parsed and reasoned about by machines.

[0109] This step constructs computable, standardized knowledge units, realizes a formal representation of power grid standards, and lays the foundation for the intelligent application of standards.

[0110] S105. Perform logical reasoning on standardized knowledge units to determine the content consistency verification results of power grid standard documents.

[0111] This step involves standardizing knowledge units to form a knowledge set to be verified. Subsequently, logical reasoning is initiated on the knowledge set to be verified, which generates content consistency verification results for power grid standard documents. This transforms the traditional qualitative review that relies on expert experience into an efficient and accurate quantitative logical judgment based on structured knowledge.

[0112] This application provides a multimodal data foundation for the parsing process by acquiring standard text clauses, circuit diagram images, and parameter table data, ensuring the comprehensiveness of information sources and overcoming the parsing bias caused by missing or incomplete data in traditional manual methods. The electrical terminology in the standard text clauses is converted into a computer-processable numerical form, realizing the digital representation of electrical terminology, facilitating subsequent automated processing, and reducing the burden of manual reading and understanding. Disambiguation processing of electrical terminology entities enables the automatic and accurate identification and unification of different meanings of the same electrical term or the differentiation of similar meanings of different electrical terms. This process effectively overcomes the inconsistency and subjectivity caused by relying entirely on expert personal experience in traditional methods. Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, standardized knowledge units are generated, realizing the dynamic association between text semantics and circuit diagram topology, breaking down information silos. Finally, logical reasoning is performed on the standardized knowledge units to determine the content consistency verification result of the power grid standard document. Therefore, this method replaces traditional manual parsing with automated steps, significantly improving the efficiency of power grid standard parsing and ensuring the objectivity and reliability of the parsing results.

[0113] Currently, existing technologies process standard text, circuit diagrams, and corresponding parameters independently, lacking a cross-modal semantic mapping mechanism. For example, short-circuit impedance in text needs to be associated with impedance parameters and their values ​​in a circuit diagram, but traditional methods cannot establish a dynamic mapping relationship between text semantics and circuit topology, leading to the problem of information silos.

[0114] To address this issue, the embodiments of this application fully consider the relationship between the disambiguated standard text clauses, circuit diagram images, and parameter table data, and generate standardized knowledge units based on the disambiguated standard text clauses, circuit diagram images, and parameter table data.

[0115] For example, based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, standardized knowledge units are generated, including: constructing a multimodal knowledge graph containing multiple nodes based on the disambiguated standard text clauses, circuit diagram images, and parameter table data; performing embedding representation learning on the nodes of the multimodal knowledge graph to obtain the feature vectors of each node; and performing association mapping on the nodes based on the similarity between the feature vectors to generate standardized knowledge units.

[0116] Based on the disambiguated standard text clauses, circuit diagram images, and their associated parameter tables obtained in the above embodiments, this application aims to construct a unified multimodal knowledge graph and utilize graph neural networks for deep representation learning.

[0117] First, the nodes of the disambiguated standard text clauses, circuit diagram images, and parameter table data are abstracted into nodes in a multimodal knowledge graph. The edges between nodes constitute semantic mapping relationships and parameter association relationships. Thus, the originally heterogeneous standard text clauses, circuit diagram images, and parameter table data are integrated into a structured semantic network, forming a unified knowledge representation that supports deep analysis.

[0118] Subsequently, the multimodal knowledge graph is processed. Specifically, a message-passing mechanism enables each node to aggregate feature information from its neighbors. After multiple iterations, each node learns a low-dimensional, dense feature vector (i.e., an embedding representation). This feature vector not only encodes the node's own attributes but also implies its contextual semantic relationships within the overall graph structure.

[0119] This process transforms raw, multimodal data into a unified vector representation that is machine-understandable and computable. The resulting feature vectors can significantly improve the performance of downstream tasks because they capture, in a computable form, the complex professional knowledge relationships across text, vision, and data in power grid engineering.

[0120] In some examples, embedding representation learning is performed on the nodes of a multimodal knowledge graph to obtain the feature vectors of each node. This includes: extracting the node features corresponding to the node; aligning the node features with multimodal features to obtain aligned features; and fusing the aligned features to obtain the feature vectors of each node. The extracted features will differ depending on the data type of the node.

[0121] In this example, firstly, the parameter table data is input into the graph neural network to obtain the node features of the corresponding parameter nodes. The method for calculating the node features of the parameter nodes is as follows: ,in: This represents the initial hidden state matrix of the graph neural network, i.e., the node features of the parameter nodes, with dimension 1. ; This represents the eigenvector matrix, with dimension 1. ,in Indicates the first The first parameter node Dimensional features; This represents the weight matrix, with dimension 1. This is used for feature transformation. In this embodiment, the dimension of the feature vector matrix can be set to... ,in, The number of nodes is a parameter. The feature dimension can be set to 128. The weight matrix has the following dimensions. ,in The output feature dimension can be set to 256.

[0122] Secondly, the features of the above parameter nodes are updated using a normalized activation function and a weighted summation and propagation algorithm to obtain the target feature representation. The parameter node update formula is as follows:

[0123] .

[0124] in: Indicates the first Parameter nodes in the layer Target feature representation; Indicates the first Parameter nodes in the layer Target feature representation; This represents the activation function, which can be set to the ReLU function; The weight coefficient represents the information of the parameter node itself, and its value range is... The weight coefficients representing the neighbor parameter node information take values ​​in the range [0,1] and satisfy the following conditions: Indicates parameter node The set of neighbor parameter nodes; Indicates parameter node For parameter nodes The influence weights are trainable parameters; Indicates the parameter node All neighbor parameter nodes Perform a summation operation. In practical applications, optionally, Set to 0.7, Set to 0.3 to balance the influence of the parameter node's own information and the parameter's neighbor node information.

[0125] Then, the circuit diagram image is input into the graph convolutional network (GCNN), which extracts features from the circuit diagram image to obtain the features of the corresponding image nodes. This embodiment also uses a cross-entropy loss function to calculate the loss value of the GCNN for learning the network. The formula for the cross-entropy loss function is as follows:

[0126] .

[0127] in, This represents the cross-entropy loss value; This represents the total number of samples (node ​​pairs) participating in the training. This represents the true label of the i-th sample, used to indicate whether there is a connection between the two corresponding nodes: if a connection exists, the value is 1, otherwise it is 0; denoted as the prediction output of the graph convolutional network for the i-th sample, that is, the probability that the graph convolutional network judges that there is a connection between the node pairs, and the value range is [0,1]. Let represent the natural logarithm function. Optionally, the Adam optimizer is used with a learning rate of 0.001, 100 training epochs, and a batch size of 64. The Adam optimizer can adaptively adjust the learning rate of each parameter, making it suitable for handling sparse gradients and non-stationary objectives, and efficiently optimizing the parameters of graph convolutional models.

[0128] Furthermore, by combining the features of image nodes, the target feature representations corresponding to parameter nodes, and the features of text nodes, a multimodal vector representation is obtained. It should be noted that the feature determination of text nodes is similar to the feature determination of the other types of nodes mentioned above.

[0129] The formula for determining the multimodal vector representation is as follows:

[0130] .

[0131] Where v represents the multimodal vector representation, with dimension 1. ; The feature of the i-th image node has a dimension of . ; This represents the target feature representation of the i-th parameter node, with dimension . ; The feature of the i-th text node has dimension 1. ; express Weighting coefficients; express Weighting coefficients; express The weight coefficients, and satisfying In practical applications, optionally, , and The initial values ​​were all set to 0.33 and optimized through model training.

[0132] The embodiments of this application can effectively combine information from different modalities through this weighted fusion method, thereby improving the comprehensiveness and accuracy of feature representation.

[0133] The formula for calculating the similarity between feature vectors can be expressed as follows:

[0134] .

[0135] in, express and The cosine similarity between them ranges from 1 to 10. ; The feature vector representing the standard text clause has a dimension of . ; The feature vector representing a circuit diagram image or parameter table, with dimension 1. ; Represents the dot product of two vectors. ; Representing vectors The Euclidean norm, ; Representing vectors The Euclidean norm of the feature vector is used. The dimension of the feature vector can be set to 512, and cosine similarity is used to calculate the similarity. When the similarity exceeds a preset threshold, such as 0.7, the table-graph mapping relationship is considered valid. The threshold is set to 0.7 to ensure a high-confidence mapping relationship and reduce the possibility of false matches.

[0136] Next, based on the constructed table-diagram mapping relationship, the parameter relationships between the circuit diagram image and the parameter table are determined, and a standard spectrum is generated. The formula for calculating the standard spectrum is:

[0137] .

[0138] in, The standard graph is a directed attribute graph; Represents a set of nodes. It contains parameter nodes and text nodes; Represents an edge set. This indicates the relationship between parameters and text; Represents a set of attributes. , representing the attributes of nodes and edges. Standard graphs use a directed attribute graph structure, with node types including parameter nodes and clause nodes, and edge types including constraint, reference, definition, and other relation types.

[0139] Finally, computable standardized knowledge units are formed, realizing the formal representation of power grid standards. Standardized knowledge units consist of three parts: entities (electrical terminology, parameters), relationships (constraints, references, etc.), and attributes (parameter values, units, etc.). The formal representation can be defined as:

[0140] .

[0141] in, Represents standardized knowledge units; Represents a set of entities. ; Represents a set of relations. ; Represents a collection of attributes. This formal representation enables the calculation and reasoning of power grid standards, laying the foundation for subsequent compliance checks.

[0142] In some embodiments, semantic annotation of standard text clauses is performed to identify electrical term entities, including: calculating the vector representation of term tags in the standard text clauses; extracting features from the vector representations to obtain vector features; performing secondary sampling on the vector features to obtain target vector features; inputting the target vector features into a conditional model to obtain electrical terms; and extracting phrase information of electrical terms to obtain electrical term entities.

[0143] As an example, a pre-trained language model, BERT (Bidirectional Encoder Representations from Transformers), is used to compute vector representations of term tags in standard text clauses. The process begins by inputting standard text clauses containing electrical term entities into the BERT model. The model encodes the input sequence using its bidirectional encoder architecture, generating a vector representation for each entity that incorporates its own semantics and its complete contextual information.

[0144] Specifically, the BERT model first segments the input standard text clause into a series of term tags. For each standard text clause, the BERT model calculates the vector representation of its context for each term tag using the following formula: .in: For the first The vector representation of each term tag, with a dimension of 768; BERT is the BERT pre-trained model function; For the first Embedsion vectors of term tags, with a dimension of 768; For the first A segmentation embedding vector of term tags, used to distinguish different sentences, with a dimension of 768; The location embedding vector for the i-th term label represents the location information and has a dimension of 768. In practical applications, the BERT model can be configured with a hidden layer dimension of 768, containing 12 encoding layers and 12 attention heads to balance the BERT model's representational power and computational efficiency.

[0145] The vector representation is fed into a Bidirectional Long Short-Term Memory (BiLSTM) layer for feature extraction. BiLSTM is a bidirectional long short-term memory network capable of capturing long-range dependencies in sequence data. The formula for feature extraction using BiLSTM is: Where: H is the output sequence of BiLSTM with a dimension of n×2d, where n is the sequence length and d is the hidden layer dimension of the unidirectional LSTM; BiLSTM is the BiLSTM model function; X is the input sequence, i.e., the vector representation determined in the above embodiment, with a dimension of n×768. In some embodiments, the hidden layer dimension of BiLSTM is preferably set to 256, the number of layers is 2, and the dropout rate is 0.3, in order to balance the performance and complexity of BiLSTM.

[0146] After capturing the bidirectional long-range dependency features of the input sequence using a BiLSTM layer, a CNN layer performs convolution operations on this feature sequence to further extract and condense key local features. Finally, the output features of the CNN layer are input into a CRF (Conditional Random Fields) model. The CRF model learns the transition constraints between labels and outputs a globally optimal predicted label sequence, thereby determining the term category (e.g., equipment, parameter) to which each term tag belongs. Based on this label sequence, the type of power term entities and their boundary positions in the text (e.g., start and end indices) can be parsed, completing the recognition of power term entities.

[0147] During the training of a CRF model, the CRF layer optimizes its parameters by minimizing its loss function. The formula for calculating the loss function of a CRF model is as follows: Where L is the CRF loss function; y represents the sequence of true class labels, y={y1,y2,...,y...} n}, y i This represents the actual label of the i-th term tag; This indicates that the CRF model predicts the category sequence. , This represents the predicted label for the i-th term tag; s(y,θ) represents the model parameters; s(y,θ) represents the label sequence. In parameters The score below; Z represents the normalization factor, Y represents the set of all possible label sequences. In practical applications, the learning rate of the CRF layer is preferably set to 0.001, the number of training epochs is 50, and the batch size is 32.

[0148] Furthermore, the phrase information of the power terminology is extracted to obtain the power term entity. It should be noted that the phrase information is composed in BIO format, where B represents the beginning of a phrase, I represents the inside of the phrase, and O represents the marker outside the phrase.

[0149] This application embodiment can accurately identify electrical terminology entities by semantically annotating standard text clauses.

[0150] Based on the above embodiments, disambiguation processing is performed on the electrical term entities to obtain disambiguated standard text clauses, including: mapping the electrical term entities to a candidate term set in a preset database to calculate the similarity between the electrical term entities and the candidate terms in the candidate term set; and obtaining the disambiguated standard text clauses based on the similarity.

[0151] In this embodiment, it can be understood that after obtaining the power term entity, the power term entity is mapped to a candidate term set in a preset database to calculate the similarity between the power term entity and the candidate terms in the candidate term set. The preset database contains commonly used terms in the power field and their possible meanings. For example, the power term "protection" in the power field may refer to different meanings such as protection device, protection function, or protection operation. The candidate term set is defined as follows: Where: C(t) represents the set of candidate terms for electrical term t, which is a set containing n elements; c i Let t represent the i-th possible meaning of the electrical term t, where i ranges from 1 to n; n represents the number of possible meanings in the candidate term set, which is usually between 2 and 5 depending on the complexity of the electrical term.

[0152] As an example, the default database is a collection of synonyms and generalized synonyms from WordNet. WordNet is a lexical database that contains semantic feature-level information, providing semantic relationships between words. The synonym collection is defined as follows: .in, Electricity terms The set of synonyms is a collection of... A collection of elements; Electricity terms The Synonyms, The value range is 1 to ; This indicates the number of synonyms in the synonym set, usually determined by the number of synonyms for that power term in WordNet.

[0153] Then, the semantic similarity of electricity terms is calculated using the distance at the semantic feature level in WordNet. A semantic feature is the basic semantic unit of word meaning; calculating the distance at the semantic feature level allows for a more accurate assessment of semantic similarity between words. Specifically, the formula for calculating semantic similarity is:

[0154] .

[0155] in: Candidate meanings for indicating electrical terms Synonyms The semantic similarity between them, with values ​​ranging from [0,1]; This is the weight of the frequency ratio of lexical semantic features, with a value range of [0,1]. This represents the frequency of occurrence of the current semantic element in the database; The total frequency of all semantic elements; This represents the weight of the position ratio of lexical semantic features, with a value range of [0,1]. This represents the position number of the current semantic element in the word's meaning; The total number of positions in the word's meaning; This is the lexical semantic feature weighting coefficient, with a value range of [0,1], and is pre-set based on the importance of the semantic features. In practical applications, optionally... Set to 0.5. Set to 0.3, Set it to 0.2 to balance the effects of the three factors: frequency, location, and weight.

[0156] Finally, based on the calculated semantic similarity, this embodiment selects the most matching meaning of the electrical term to eliminate ambiguity and obtain the disambiguated standard text clause, which includes the disambiguated electrical term. The selection rule is as follows:

[0157] .

[0158] in: This indicates the meaning of the selected, most matching electrical term; This indicates obtaining the largest value in the following expression. value; express It is an electrical term. Candidate set One of the elements; Indicates the use of electrical terminology All synonyms Calculation and candidate meaning The sum of similarity values. When the highest similarity value exceeds a preset threshold, for example, when the highest similarity value exceeds 0.75, disambiguation is considered successful; otherwise, the original terminology is retained and marked as unsuccessful disambiguation.

[0159] This application embodiment utilizes a semantic similarity calculation method to further verify the domain rationality of candidate terms, significantly improving the accuracy of power term disambiguation and reducing standard parsing errors caused by misjudgment of power terms. Furthermore, the dynamic nature of the semantic similarity calculation method allows this application embodiment to adapt to the polysemy of power domain terms and the rapid iteration of new terms, providing a more accurate terminological foundation for subsequent multimodal information fusion.

[0160] Based on the above embodiments, obtaining a power grid standard document includes: obtaining a multimodal power grid standard document containing standard text, circuit diagrams, and parameter tables; extracting standard text clauses, initial circuit diagram images, and initial parameter table data from the multimodal power grid standard document; using a pre-trained neural network to predict the connection relationship between the standard text clauses and the initial circuit diagram images, generating a candidate connection set between the standard text clauses and the nodes of the initial circuit diagram images; optimizing the neural network using a link prediction loss function; filtering the candidate connection set based on the optimized neural network to determine the semantic mapping relationship between the standard text clauses and the nodes of the initial circuit diagram images; and determining the standard text clauses, circuit diagram images, and parameter table data from the standard text clauses, the initial circuit diagram images, and the initial parameter table data based on the semantic mapping relationship.

[0161] After extracting the standard text clauses, the initial circuit diagram image, and the initial parameter table data, this embodiment of the application constructs a neural network model to calculate the correlation strength between two types of heterogeneous nodes—standard text clause nodes and initial circuit diagram image element nodes. The neural network model calculates the relationship score using the following formula:

[0162] .

[0163] in: Represents a node and nodes The relationship score is calculated, with values ​​ranging from [-1, 1]. Represents a node The feature vector has a dimension of . ; Represents a node The feature vector has a dimension of . ; and Denotes a trainable weight matrix, with dimensions 1. ; To represent a non-linear activation function, one can use... function, The value range is [-1, 1]. Optionally, and The dimension is ,in This is the feature dimension (set to 256). One point needs to be noted... There are two types of heterogeneous nodes.

[0164] Based on the relationship scores calculated above, the neural network further employs an attention mechanism to refine the prediction and aggregation of relationships between nodes. Specifically, for a target node t (e.g., an initial circuit diagram image element), the importance of its association with each neighboring node i (e.g., related standard text clauses) is determined by attention weights. To quantify it, the calculation formula is as follows:

[0165] .

[0166] in: Representing neighboring nodes For the target node The attention weights take values ​​in the range [0,1] and satisfy the following conditions: Represents the target node eigenvectors; Represents the target node The set of neighboring nodes; Represents the natural constant exponential function Indicates the target node All neighboring nodes The summation is then performed. It should be noted that this attention mechanism can automatically learn the importance weights between different nodes, highlighting the influence of key nodes and reducing the interference of secondary nodes.

[0167] In summary, it can be understood that the embodiments of this application first utilize a pre-trained neural network to calculate the basic relationship scores between all node pairs in the standard text clauses of the power grid and the initial circuit diagram image. Subsequently, based on these basic relationship scores, the aforementioned attention mechanism is used to aggregate the neighbor information of each target node and generate a weight distribution. Finally, the neural network model filters out high-confidence associations according to the weights, forming a candidate connection set between the standard text clauses and the nodes of the initial circuit diagram image. This process achieves refined relationship construction from basic association prediction to importance perception.

[0168] Furthermore, embodiments of this application also use a link prediction loss function to learn and optimize the neural network. The calculation formula for the link prediction loss function is as follows: .in, This represents the loss value of the link prediction loss function; , γ and γ are the weight coefficients of each part of the link prediction loss function, and satisfy γ = γ + γ. This represents the binary cross-entropy loss function. This indicates the actual relationship, usually 0 or 1 to indicate whether the relationship exists; Indicates the probability of predicting a relationship; This represents the L2 regularization term of the weight matrix, used to prevent overfitting. ; This represents a relational regularization term, used to encourage neural networks to predict more meaningful relations. ,in Represents a known set of edges. Optionally, and The values ​​were set to 0.6, 0.2, and 0.2 respectively to balance the effects of the main loss and the regularization term.

[0169] Finally, the generated candidate connection set is filtered based on the optimized neural network to ultimately determine the precise semantic mapping relationship between the standard text terms and the nodes of the initial circuit diagram image. The specific filtering rules are as follows: when the relationship score of any pair of nodes... If the confidence threshold is exceeded, the semantic relationship is considered valid and included in the final mapping set; otherwise, it is considered that there is no direct and valid semantic association between the node pairs. The confidence threshold can be set to 0.65, a value determined based on experimental results, which can achieve a sufficiently high recall rate while ensuring accuracy. This score-based screening mechanism can effectively reduce the false detection rate and improve the accuracy of relationship identification. It should be noted that the confidence threshold value can be modified according to actual conditions.

[0170] In some embodiments, it is necessary to filter non-compliant, disambiguated standard text clauses based on a power terminology rule base. The power terminology rule base contains usage specifications, standard expressions, and constraints for power-related terminology. For example, it may stipulate that rated current must be used with specific numerical values ​​and units. Specifically, the filtering rules can be expressed as follows:

[0171] .

[0172] in, The terminology rule function represents the input power terms. The evaluation is performed, and the output is a binary value; 1 represents the electrical term to be evaluated; 1 represents the electrical term. It conforms to the rules defined in the rule base; 0 indicates an electricity term. This does not comply with the rules. This indicates that the use of electrical terminology in the disambiguated standard text is non-compliant and needs to be corrected.

[0173] When non-compliant use of electricity terminology is detected in the disambiguated standard text clauses, error correction samples are automatically generated. These samples are then used for supervised learning of the syntax-semantic analysis error correction rules. The supervised learning process uses correctly labeled samples as training data. The loss function for supervised learning is defined as:

[0174] .

[0175] in, The loss function representing syntax and semantic analysis error correction; Indicates the number of samples; Indicates the number of error types; Indicates sample Does an error type exist? The value can be 0 or 1, where 1 indicates that the error of this type exists and 0 indicates that it does not exist; Indicates the predicted sample Error type exists The probability of is in the range of [0,1]. The loss function represents the natural logarithm. Optionally, error types include subject-verb disagreement, missing components, and redundant components. This multi-class classification loss function can guide supervised learning to identify different types of syntactic and semantic errors.

[0176] After supervised learning is completed, the optimized syntax and semantic analysis error correction rules are applied to automatically judge the standard text clauses and identify potential grammatical or semantic errors. The judgment rules can be expressed as the following formula:

[0177] .

[0178] in, This represents an error detection function that processes the standard text clauses after input disambiguation. The evaluation is performed, and the output is a binary value; This indicates the disambiguated standard text terms to be evaluated; Represents a set of rules. ,Include A different detection rule; Representation rules Standard text clauses after disambiguation The judgment result is 1, indicating that an error exists, and 0, indicating that no error exists; This indicates the existence of a certain rule. Belongs to the rule set ;1 indicates the standard text clause after disambiguation. There is at least one syntactic or semantic error; 0 indicates the standard text clause after disambiguation. There are no grammatical or semantic errors.

[0179] If grammatical or semantic errors are found, they will be corrected; otherwise, the process will proceed to the table. Figure 1 The consistency check step involves verifying whether the circuit diagram image matches the parameter table data. The aforementioned error correction process employs a rule-based approach, applying appropriate correction rules based on the error type.

[0180] In the table Figure 1 During the consistency check, the named entities of the parameter data are replaced with electrical terms, and a table is generated based on this. Figure 1 Consistency samples for the consistency check process. The format of the consistency sample is defined as follows:

[0181] .

[0182] in, Represents a consistent set of samples; Represents a consistent sample triple; Indicates electrical terms, such as rated voltage; Indicates the corresponding parameter, such as Un; Indicates a parameter value, such as 110kV; Indicates the number of samples; Indicates the sample index, from 1 to .

[0183] Next, set up the table. Figure 1 Consistency testing rules are used to extract standard text clauses with inconsistent table parameters and generate a consistency testing rule sample. The testing rules can be expressed as:

[0184] .

[0185] in, This represents the detection function, which is applied to the input electrical terms. ,parameter and parameter data The evaluation is performed, and the output is a binary value; Indicates electrical terminology; Indicates parameters; Indicates parameter data; This refers to the parameter data mentioned in the standard text clause after disambiguation; This represents the parameter data in the parameter table; This indicates that the parameter data mentioned in the standard text clause after disambiguation is inconsistent with the parameter data in the parameter table; This indicates that the parameter data mentioned in the disambiguated standard text clause is consistent with the parameter data in the parameter table. When the two are inconsistent, it is considered that there is an inconsistency problem and correction is required.

[0186] Then, a supervised learning approach is adopted, using labeled consistent samples as training data. Finally, the disambiguated standard text clauses are checked to determine whether the table and graph parameters are consistent. If inconsistencies are found, corrections are made.

[0187] Furthermore, this application embodiment also performs separate checks on the parameter table data, that is, it uses inter-table calculation rules and intra-table calculation rules to check the parameter table data separately. The inter-table calculation rules are used to check for inconsistencies in parameters between parameter tables and parameter tables after association, and the calculation formula is:

[0188] ,

[0189] in: This indicates the calculation error between parameter tables, used to determine if there are inconsistencies in the calculations between parameter tables; Indicates the target parameter value; Indicates parameters based on other parameter tables The calculation results The calculation function is defined according to the formulas in electrical engineering. This represents the absolute error between the actual value and the calculated value; This indicates the tolerance threshold. Optionally, The threshold is set to 1% of the parameter value to account for the error tolerance in actual engineering. When the error exceeds the threshold, it is considered that there is an inconsistency in the calculation between tables.

[0190] The calculation rules within the table are used to check for errors in the calculation results of the parameters within the parameter table. The calculation formula is as follows:

[0191] ,

[0192] in: This indicates the calculation error within the parameter table, used to determine if there are inconsistencies in the calculations within the parameter table; This indicates the parameter value calculated based on other parameters in the parameter table; This indicates the parameter value explicitly declared in the parameter table; This represents the absolute error between the calculated value and the declared value; This indicates the tolerance threshold. Optionally, Set it to 0.5% of the parameter value to ensure the accuracy of calculations within the table. This threshold is lower than the threshold for calculations between tables because calculations within the parameter table usually require higher precision.

[0193] The above process is equivalent to preprocessing the disambiguated standard text clauses, circuit diagrams, and parameter tables, and then using rule-based reasoning algorithms to extract the dynamic constraints between the disambiguated standard text clauses, determining whether the logical relationships in the disambiguated standard text clauses are correct. The determination process first establishes a dependency graph between the standard text clauses. ,in This represents the standard text clause node set. The dependency edge set is represented, and then the existence of circular dependencies or contradictory constraints is checked. Circular dependency detection uses a depth-first search algorithm, and contradictory constraint detection uses a constraint satisfaction problem (CSP) solution method. If logical errors are found, an iterative optimization strategy based on reinforcement learning is used to update the standard text clauses. The iterative optimization process of reinforcement learning can be represented as follows: .in, The Q-value function represents the state. Take action Expected cumulative reward; This represents the learning rate, which controls the update step size, and its value ranges from [value range missing]. ; This indicates an immediate reward, depending on the effectiveness of the action; This represents the discount factor, which controls the importance of future rewards, and its value ranges from [value range missing]. ; Indicates the execution of actions The next state that is reached later; Indicates all possible next actions Select The action with the greatest value; Representing state Take action of Value. In the embodiments of this application, state The current state of the standard text clauses after disambiguation, action Rewards for possible modifications The degree of improvement in logical consistency corresponding to the modification. Optionally, Set as , Set as To balance current learning with long-term planning.

[0194] The iterative process generates iterative samples, adds them to the sample set, and repeatedly executes the syntax and semantic analysis error correction rules and tables. Figure 1Consistency checks, inter-table calculation rules, intra-table calculation rules, and logical contradiction correction rules are applied until the logical relationships in the standard text clauses are correct. The iteration terminates when the number of errors no longer decreases after five consecutive iterations, or when the maximum number of iterations (which can be set to 20) is reached. This iterative optimization mechanism enables the system to adaptively learn, continuously improving the accuracy and effectiveness of rule-based reasoning.

[0195] Optionally, the iterative optimization process based on reinforcement learning can employ the Q-learning algorithm. The state space includes the current state of the standard text clause, the action space includes possible modification operations, and the reward function is designed based on the improvement in logical consistency after modification. The update formula for the Q-learning algorithm is:

[0196] ,

[0197] in: This represents the learning rate, which controls the update speed, and its value range is... This indicates an immediate reward, depending on the effectiveness of the action; This represents the discount factor, which controls the importance of future rewards, and its value ranges from [value range missing]. This indicates selecting the next action that maximizes the Q-value. Optionally, the learning rate is set to 0.1, the discount factor is set to 0.9, and the exploration rate is initially set to 0.5 and gradually reduced to 0.1 to balance exploration and exploitation.

[0198] Furthermore, semantic templates can be used to mine latent semantic relationships within the disambiguated standard text clauses, enabling comprehensive syntactic and semantic analysis of these clauses. During semantic template mining, various template types can be defined, including conditional templates, constraint templates, and reference templates. These template types are used to match specific semantic structures within the disambiguated standard text clauses. For example, a conditional template can be represented as "if condition, then result," used to capture conditional constraint relationships. Template matching employs a combination of rule-based and statistical methods to improve matching accuracy and coverage.

[0199] Based on the above embodiments, logical reasoning is performed on standardized knowledge units to determine the content consistency verification result of power grid standard documents. This includes: determining ambiguity vectors based on standardized knowledge units; calculating the similarity of entity relationships in standardized knowledge units using a cross-attention mechanism to obtain entity pair vectors; obtaining disambiguation vectors based on ambiguity vectors and entity pair vectors; fusing the disambiguation vectors and syntactic tree vectors to generate a fused vector, where the syntactic tree vector is obtained by tree encoding the disambiguated standard text clauses; pre-mapping the fused vector to obtain a pre-mapped vector; and inputting the fused vector and the pre-mapped vector into a rule reasoning model to generate and output the content consistency verification result.

[0200] This application's embodiments extract entities and entity relationships based on standardized knowledge units, and use entity relationships as disambiguation tags. Further, by analyzing the topological features of power terminology entities in the metadata set—including indicators describing the structural characteristics of entities in a multimodal knowledge graph such as degree, centrality, and clustering coefficient—positive, negative, and neutral ambiguity vectors are calculated and determined for each entity relationship. The ambiguity vector can be formally defined as a function mapping or feature combination related to topological features, used to quantify the semantic tendency and degree of ambiguity of the relationship in different contexts. Specifically, the ambiguity vector is defined as:

[0201] , , .

[0202] in, This represents a positive ambiguity vector, indicating the positive or affirmative meaning of electrical terms, with a dimension of 768. This represents a negative ambiguity vector, indicating the negative or negative meaning of electrical terms, with a dimension of 768. This represents the neutral ambiguity vector, indicating the neutral meaning of electrical terms, with a dimension of 768. Indicates the size of the set of positive relations; Indicates the size of the set of negative relations; Indicates the size of the set of indirect relations; Represents a set of positive relations All relationships Perform summation; Representing relations The vector representation has a dimension of 768. The vector dimension can be set to 768, consistent with the BERT output dimension, to facilitate subsequent fusion with BERT features.

[0203] Next, entity pairs containing electricity terms are obtained from standardized knowledge units. Using a cross-attention mechanism, the semantic similarity between the electricity term relation vectors within the entity pairs is calculated based on a pre-trained model, thus obtaining the corresponding positive, negative, and neutral entity pair vectors. The cross-attention mechanism calculation formula is expressed as follows: .in, The output of the attention function is represented by the dimension and same; Represents a query matrix with dimensions of Represents the key matrix, with dimension . Represents a value matrix with dimension 1. This represents matrix multiplication of the query matrix and the transpose of the key matrix, with the result having dimension 1. The dimension of the key vector is used for normalization; softmax represents the softmax function. This converts the input into a probability distribution. This represents the scaling factor, used to prevent the dot product result from being too large, which would lead to an excessively small softmax gradient. In practical applications, optionally, the number of attention heads is set to 8 and the hidden layer dimension is set to 512 to balance computational complexity and model expressiveness.

[0204] As an example, the attention mechanism described above can be a multi-head attention mechanism, and the calculation formula for the multi-head attention mechanism is as follows: .in, This represents the output of multi-head attention; Indicates the first The output of each attention head; , , Indicates the first The parameter matrix of each attention head; This indicates the output linear transformation matrix; Concat indicates a concatenation operation. Optionally, the number of heads... Set to 8, each head has a dimension of 64, and the total dimension is 512.

[0205] The formula for calculating the similarity of entity relationships is as follows: .in, Representing vectors and The cosine similarity between them ranges from 1 to 10. ; and This represents the vector whose similarity needs to be calculated; Representing vectors and The dot product; Representing vectors The Euclidean norm; Representing vectors The Euclidean norm of vectors. Cosine similarity can effectively measure the similarity of vector directions and is not affected by vector length.

[0206] Then, the positive entity pair vector and the negative entity pair vector are averaged and then added to the positive ambiguity vector to obtain the disambiguation vector. The formula for calculating the disambiguation vector is as follows:

[0207] .

[0208] in, This represents the disambiguation vector, with a dimension of 768; This represents a positive ambiguity vector with a dimension of 768; This represents a positive entity pair vector with a dimension of 768; This represents a negative entity pair vector with a dimension of 768. This represents the average of the positive entity pair vector and the negative entity pair vector. This summation operation can combine ambiguous information and entity relationship information to generate a more comprehensive disambiguation representation.

[0209] Next, the disambiguation vector and the syntax tree vector are fused to generate a fused vector. The fusion formula is: .in, This represents the fusion vector, with a dimension of 768; This represents the disambiguation vector, with a dimension of 768; The vector representing the syntax tree has a dimension of 768; This represents the fusion weight coefficient, with a value range of [0,1]. The fusion weight coefficient can be set to 0.6 to balance the influence of disambiguation information and syntactic structure information.

[0210] Furthermore, a syntactic tree vector is a vector representation obtained by constructing a syntactic dependency tree from the disambiguated standard text terms and encoding that tree. Syntactic tree vectors can represent the syntactic structure information of a statement, which helps in understanding the grammatical roles of terms within a sentence.

[0211] Then, the fused vectors are pre-mapped using a pre-trained model to obtain the pre-mapped vectors. The pre-mapped formula is: .in, This represents a pre-mapped vector with a dimension of 1024; This represents the fusion vector, with a dimension of 768; This represents the weight matrix, with dimension 1. This represents the bias vector, with a dimension of 1024. Optionally, the weight matrix and bias vector are obtained through pre-training, with the pre-mapped vector dimension set to 1024 to improve representational power.

[0212] Finally, the fused vector and pre-mapped vector are input into the rule-based reasoning model to generate compliance verification results. The rule-based reasoning model, based on a predefined rule base, infers whether the input vectors comply with power grid standards. The content consistency verification results include information such as compliance with specifications, the type and severity of existing problems, and suggested corrective actions.

[0213] The practical application effect of the embodiments of this application is illustrated below through a specific example.

[0214] This embodiment selects the transformer-related standards from the "Design Code for 110kV and Above Substations" for analysis. First, it obtains the standard text clauses, circuit diagrams, and parameter tables. Then, it preprocesses the standard text, including sentence segmentation and removal of special characters.

[0215] Next, the BERT-BiLSTM-CRF model is used to perform multi-task deep semantic annotation on the standard text clauses to identify electrical term entities, such as transformer, rated voltage, and short-circuit impedance. Then, disambiguation processing of electrical term entities is performed; for example, disambiguation of "protection" in transformer protection is performed as "protection device" rather than "protection operation".

[0216] Then, a graph neural network is used to encode the transformer parameters in the circuit diagram image, obtaining a parameter vector representation. Through cross-modal mapping and link prediction, the association between standard text clauses and circuit diagram parameters is established, such as associating the rated voltage of the transformer being 110kV with the voltage parameters in the circuit diagram image.

[0217] Next, the standard text clauses are mapped to circuit diagrams and parameter table data to form standardized knowledge units. For example, the short-circuit impedance of the transformer should be in the range of 4% to 8% and the short-circuit impedance value in the parameter table is correlated.

[0218] Finally, rule-based reasoning is used to obtain the dynamic constraint relationships between standard text clauses and detect potential problems. For example, if the standard text clause states that the transformer's rated voltage is 110kV, while the parameter table data gives a value of 115kV, the system detects this inconsistency and provides correction suggestions.

[0219] Figure 2 This application provides an intelligent analysis system for power grid standards based on a multimodal knowledge graph. For example... Figure 2 As shown, the system includes:

[0220] The model training module is used to construct model training samples, that is, to perform multi-task deep semantic annotation on standard power grid text.

[0221] The knowledge fusion module is used to achieve knowledge fusion through the power terminology disambiguation module and the knowledge graph fusion module, wherein:

[0222] The power term disambiguation module maps power terms to the corresponding power term candidate set, thereby constructing word graph features. In the deep learning algorithm, multimodal feature fusion is achieved through the attention mechanism.

[0223] The knowledge graph fusion module includes a cross-modal mapping unit and a link prediction unit. The cross-modal mapping unit uses graph neural networks to align multimodal knowledge features, and the link prediction unit uses an attention mechanism to filter relationships.

[0224] The rule-based reasoning module is used to obtain dynamic constraint relationships between standard text clauses using rule-based reasoning, determine logical relationships within standard text clauses, including syntax and semantic analysis error correction rules, and tables. Figure 1Consistency check rules, inter-table calculation rules, intra-table calculation rules, and logical contradiction correction rules;

[0225] Among them, the rule reasoning module uses semantic templates to mine potential semantic relationships in standard power grid texts and performs iterative optimization based on reinforcement learning algorithms.

[0226] In summary, the beneficial effects of the embodiments of this application include:

[0227] 1. By using multi-level semantic similarity calculation and context-aware word graph feature construction, the ambiguity problem of power terminology is effectively solved, and the accuracy of terminology recognition is improved.

[0228] 2. By employing graph neural networks and attention mechanisms, an effective mapping between text terms and circuit diagrams and parameter tables was achieved, overcoming the challenge of multimodal information fusion;

[0229] 3. The rule reasoning mechanism based on semantic template mining and reinforcement learning iterative optimization can automatically detect and correct syntactic and semantic errors, parameter inconsistencies and logical contradictions in standard clauses;

[0230] 4. It has improved the automation, accuracy and efficiency of power grid standard analysis, and has important practical value for the compilation, review and application of power grid standards.

[0231] Figure 3 This is a schematic diagram of the structure of the power grid standard analysis device provided in the embodiments of this application, as shown below. Figure 3 As shown, the power grid standard analysis device provided in this embodiment includes:

[0232] The acquisition module 301 is used to acquire power grid standard documents, which include standard text clauses, circuit diagrams, and parameter table data.

[0233] The identification module 302 is used to perform semantic annotation on standard text clauses to identify electrical terminology entities;

[0234] The determination module 303 is used to perform disambiguation processing on electrical terminology entities to obtain disambiguated standard text clauses.

[0235] The generation module 304 is used to generate standardized knowledge units based on the disambiguated standard text clauses, circuit diagram images, and parameter table data.

[0236] The determination module 303 is also used to perform logical reasoning on standardized knowledge units to determine the content consistency verification results of power grid standard documents.

[0237] In one possible implementation, the generation module 304 is specifically used for:

[0238] Based on the disambiguated standard text clauses, circuit diagram images, and parameter table data, a multimodal knowledge graph containing multiple nodes is constructed.

[0239] Embedded representation learning is performed on the nodes of the multimodal knowledge graph to obtain the feature vectors of each node;

[0240] Based on the similarity between feature vectors, nodes are associated and mapped to generate standardized knowledge units.

[0241] In one possible implementation, the determining module 303 is specifically used for:

[0242] Extract the node features corresponding to the node;

[0243] Multimodal feature alignment is performed on the node features to obtain the aligned features;

[0244] The aligned features are fused to obtain the feature vectors of each node.

[0245] In one possible implementation, the identification module is specifically used for:

[0246] Calculate the vector representation of term tags in standard text clauses;

[0247] Extract features from the vector representation to obtain vector features;

[0248] The vector features are sampled twice to obtain the target vector features;

[0249] The target vector features are input into the conditional model to obtain power terminology;

[0250] Extract phrase information from electrical terminology to obtain electrical term entities.

[0251] In one possible implementation, the determining module 303 is specifically used for:

[0252] Power term entities are mapped to a set of candidate terms in a pre-defined database to calculate the similarity between power term entities and candidate terms in the set of candidate terms.

[0253] Based on the similarity, the disambiguated standard text clauses are obtained.

[0254] In one possible implementation, the acquisition module 301 is specifically used for:

[0255] Obtain multimodal power grid standard documents containing standard text, circuit diagrams, and parameter tables;

[0256] Extract the standard text clauses, initial circuit diagram images, and initial parameter table data from the multimodal power grid standard document;

[0257] A pre-trained neural network is used to predict the connection relationship between the standard text clauses and the initial circuit diagram image, generating a set of candidate connections between the nodes of the standard text clauses and the initial circuit diagram image.

[0258] Optimize neural networks using the link prediction loss function;

[0259] The candidate connection set is filtered based on the optimized neural network to determine the semantic mapping relationship between the standard text terms and the nodes of the initial circuit diagram image;

[0260] Based on semantic mapping relationships, the standard text terms, circuit diagram images, and parameter table data are determined from the standard text terms, initial circuit diagram images, and initial parameter table data.

[0261] In one possible implementation, the determining module 303 is specifically used for:

[0262] Based on standardized knowledge units, determine the ambiguity vector;

[0263] A cross-attention mechanism is used to calculate the similarity of entity relationships in standardized knowledge units, resulting in entity pair vectors.

[0264] Based on the ambiguous vector and the entity pair vector, the disambiguation vector is obtained;

[0265] The disambiguation vector and the syntactic tree vector are fused to generate a fused vector. The syntactic tree vector is obtained by tree encoding the disambiguated standard text clauses.

[0266] Pre-mapped fusion vectors yield pre-mapped vectors;

[0267] Input the fused vector and pre-mapped vector into the rule-based reasoning model to generate and output the content consistency verification result.

[0268] The power grid standard analysis device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0269] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into an integrated circuit within the above device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0270] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented by calling program code through a processing element, that processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together to implement a System-On-a-Chip (SOC).

[0271] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 provided in this application embodiment may include: a processor 401, and a memory 402 communicatively connected to the processor, wherein:

[0272] The memory stores the instructions that the computer executes;

[0273] The processor executes computer execution instructions stored in memory to implement the method described in the foregoing method embodiments.

[0274] It should be understood that processor 401 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor. Memory 402 may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.

[0275] Optionally, the electronic device 400 may also include a communication interface 403. In specific implementations, if the communication interface 403, memory 402, and processor 401 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0276] Optionally, in a specific implementation, if the communication interface 403, memory 402 and processor 401 are integrated on a single integrated circuit, then the communication interface 403, memory 402 and processor 401 can communicate through an internal interface.

[0277] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described in any of the foregoing embodiments.

[0278] It is understood that the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0279] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the computer-readable storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an ASIC. Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic device.

[0280] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a computer-readable storage medium, include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application.

[0281] This application also provides a computer program product, including a computer program that, when executed, implements the method described in any of the foregoing embodiments.

[0282] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0283] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0284] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as these combinations of technical features do not contradict each other, they should be considered within the scope of this specification.

[0285] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0286] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for analyzing power grid standards, characterized in that, include: Obtain power grid standard documents, which include standard text clauses, circuit diagrams, and parameter table data; Semantic annotation is performed on the standard text clauses to identify electrical terminology entities; The electrical terminology entities are disambiguated to obtain the disambiguated standard text clauses; Based on the disambiguated standard text clauses, the circuit diagram image, and the parameter table data, standardized knowledge units are generated. Logical reasoning is performed on the standardized knowledge units to determine the content consistency verification result of the power grid standard document.

2. The method according to claim 1, characterized in that, The process of generating standardized knowledge units based on the disambiguated standard text clauses, the circuit diagram image, and the parameter table data includes: Based on the disambiguated standard text terms, the circuit diagram image, and the parameter table data, a multimodal knowledge graph containing multiple nodes is constructed. Embedding representation learning is performed on the nodes of the multimodal knowledge graph to obtain the feature vector of each node; Based on the similarity between the feature vectors, the nodes are associated and mapped to generate standardized knowledge units.

3. The method according to claim 2, characterized in that, The embedding representation learning of nodes in the multimodal knowledge graph to obtain feature vectors for each node includes: Extract the node features corresponding to the node; Multimodal feature alignment is performed on the node features to obtain aligned features; By fusing the aligned features, the feature vectors of each node are obtained.

4. The method according to any one of claims 1 to 3, characterized in that, The semantic annotation of the standard text clauses to identify electrical terminology entities includes: Calculate the vector representation of the term tags in the standard text clauses; Extract the features of the vector representation to obtain vector features; The vector features are sampled twice to obtain the target vector features; The target vector features are input into the conditional model to obtain power terminology; Extract the phrase information of the power term to obtain the power term entity.

5. The method according to any one of claims 1 to 3, characterized in that, The disambiguation process performed on the electrical terminology entity to obtain the disambiguated standard text clauses includes: The power term entities are mapped to a set of candidate terms in a preset database to calculate the similarity between the power term entities and the candidate terms in the set of candidate terms. Based on the similarity, the disambiguated standard text terms are obtained.

6. The method according to any one of claims 1 to 3, characterized in that, The acquisition of power grid standard documents includes: Obtain multimodal power grid standard documents containing standard text, circuit diagrams, and parameter tables; Extract the standard text clauses, initial circuit diagram images, and initial parameter table data from the aforementioned multimodal power grid standard document; A pre-trained neural network is used to predict the connection relationship between the standard text clauses and the initial circuit diagram image, generating a candidate connection set between the nodes of the standard text clauses and the initial circuit diagram image; The neural network is optimized using a link prediction loss function; The candidate connection set is filtered based on the optimized neural network to determine the semantic mapping relationship between the standard text terms and the nodes of the initial circuit diagram image; Based on the semantic mapping relationship, the standard text terms, the circuit diagram image, and the parameter table data are determined from the standard text terms, the initial circuit diagram image, and the initial parameter table data.

7. The method according to any one of claims 1 to 3, characterized in that, The step of performing logical reasoning on the standardized knowledge units to determine the content consistency verification result of the power grid standard document includes: Based on the standardized knowledge units, determine the ambiguity vector; A cross-attention mechanism is used to calculate the similarity of entity relationships in the standardized knowledge unit, resulting in an entity pair vector. Based on the ambiguity vector and the entity pair vector, a disambiguation vector is obtained; The disambiguation vector and the syntactic tree vector are fused to generate a fused vector, wherein the syntactic tree vector is obtained by tree encoding the disambiguated standard text clauses; The fusion vector is pre-mapped to obtain a pre-mapped vector; The fusion vector and the pre-mapped vector are input into the rule-based reasoning model to generate and output the content consistency verification result.

8. A power grid standard analysis device, characterized in that, include: The acquisition module is used to acquire power grid standard documents, which include standard text clauses, circuit diagram images, and parameter table data. The identification module is used to perform semantic annotation on the standard text clauses to identify electrical terminology entities; The determination module is used to disambiguate the electrical terminology entities to obtain the disambiguated standard text clauses. The generation module is used to generate standardized knowledge units based on the disambiguated standard text clauses, the circuit diagram image, and the parameter table data; The determining module is further configured to perform logical reasoning on the standardized knowledge units to determine the content consistency verification result of the power grid standard document.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1-7.

Images