Construction method and device of power inspection scene knowledge graph, equipment and storage medium

Abstract

The application relates to a power operation and inspection scene knowledge graph construction method and device, equipment and a storage medium, multi-modal data in a power operation and inspection scene is acquired; the multi-modal data comprises image data, text data, table data and time series data; text data is subjected to feature extraction according to a preset large model, and a target triple sequence is obtained; the target triple sequence comprises multiple triples; each triple comprises entity information and relationship information. An initial power operation and inspection knowledge graph is constructed according to the target triple sequence, and other modal data in the multi-modal data is dynamically associated with the initial power operation and inspection knowledge graph, and a multi-modal knowledge graph is obtained. Through acquisition of the multi-modal data, a data basis for fault reasoning is provided, and through accurate description of entity attributes and relationships by the triple sequence, the multi-modal data is associated with the power operation and inspection knowledge graph, and the accuracy of a fault reasoning result is improved.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method, apparatus, equipment and storage medium for constructing a knowledge graph for power operation and maintenance scenarios. Background Technology

[0002] In the operation and maintenance of modern power systems, power operation and maintenance information, such as transformer operation logs, relay protection monitoring setting sheets, manufacturer technical manuals, PMS equipment ledgers, SCADA equipment data, and drone inspection thermal imaging maps, are core information sources for ensuring the safe and stable operation of the power grid. This information covers key power aspects such as equipment status, consequences of faults and accidents, and hazard planning, and has irreplaceable value for decision-making by dispatchers, operation and maintenance specialists, and field workers.

[0003] Existing power operation and maintenance information is mainly obtained from different systems and processed according to the "phenomenon-principle-process" rule to diagnose specific power equipment faults. Because unstructured texts such as regulations and drawings are scattered across different systems with formatted or visual data such as SCADA interfaces and infrared thermal images, severe data silos have formed, making it difficult to unify power operation and maintenance information. Traditional keyword-based retrieval systems lack logical reasoning capabilities and struggle to solve complex fault tracing problems.

[0004] In recent years, large language models have performed well in the general cognitive domain. By inputting power operation and maintenance information into large language models for recognition, power operation and maintenance detection results can be obtained.

[0005] However, the aforementioned large language models lack rigorous domain logic constraints, are prone to giving incorrect suggestions that do not comply with safety regulations, and cannot deeply correlate multi-source information, thus failing to support highly reliable power operation and maintenance work. Summary of the Invention

[0006] Therefore, it is necessary to provide a method, apparatus, equipment, and storage medium for constructing a knowledge graph of power operation and maintenance scenarios that can accurately infer faults in power systems, in order to address the above-mentioned technical problems.

[0007] Firstly, this application provides a method for constructing a knowledge graph for power operation and maintenance scenarios, including:

[0008] Acquire multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data;

[0009] Based on a pre-defined large model, feature extraction is performed on the text data to obtain a target triplet sequence; the target triplet sequence includes multiple triplets; each triplet includes entity information and relation information.

[0010] An initial power operation and maintenance knowledge graph is constructed based on the target triple sequence, and other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0011] In one embodiment, feature extraction is performed on the text data according to a preset large model to obtain a target triplet sequence, including:

[0012] Text data is divided into domain semantic units based on a fine-grained partitioning strategy;

[0013] The domain semantic units are input into a pre-defined large model for processing, generating an initial triplet sequence and a reference tag for each triplet;

[0014] For each triplet in the initial triplet sequence, the power terminology of the triplet is compared with the standard power dictionary, and the relationship information of the triplet is checked for consistency with the original electrical logic corresponding to the reference mark.

[0015] If the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and its corresponding reference tag are identified as target triples, and a target triple sequence is generated based on multiple target triples.

[0016] In one embodiment, the text data includes runtime logs, procedure data, and ledger data; the text data is divided into domain semantic units according to a fine-grained partitioning strategy, including:

[0017] The operation logs, procedure data, and ledger data are preprocessed to identify electrical boundary identifiers in the text data. Electrical boundary identifiers include at least one of the following: equipment ledger start mark, fault clause number, timestamp separator, and procedure chapter title.

[0018] Using electrical boundary identifiers as segmentation anchors, and specific equipment ledger paragraphs or fault handling clauses as independent units, text data is segmented into multiple candidate semantic units.

[0019] Perform electrical logic integrity verification on each candidate semantic unit, and determine the candidate semantic units that pass the integrity verification as domain semantic units.

[0020] In one embodiment, other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph, including:

[0021] Extracting image data and time series data from multimodal data;

[0022] Image data and time series data are mounted onto the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0023] In one embodiment, the method further includes:

[0024] Vector transformation is performed on the multimodal knowledge graph to generate a knowledge graph vector library; the vectors in the knowledge graph vector library uniquely correspond to the nodes and relations in the multimodal knowledge graph.

[0025] For any two vectors in the knowledge graph vector database, calculate the similarity between the two vectors; the similarity includes entity similarity and attribute description similarity.

[0026] Vectors with a similarity greater than a preset similarity threshold are merged.

[0027] In one embodiment, the method further includes:

[0028] The nodes in the power operation and maintenance knowledge graph are clustered based on the graph community detection algorithm to generate multiple subsystem communities;

[0029] For each subsystem community, extract all nodes and their associated factual claims within the subsystem community to generate a context information package for the subsystem community;

[0030] The context information package is evaluated based on the preset health assessment model to obtain a structured health assessment summary of the subsystem community;

[0031] The structured health assessment summary is attached as an attribute of the subsystem community to the knowledge graph of the power operation and maintenance scenario.

[0032] Secondly, this application also provides a device for constructing a knowledge graph for power operation and maintenance scenarios, including:

[0033] The acquisition module is used to acquire multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data.

[0034] The extraction module is used to extract features from text data according to a preset large model to obtain a target triple sequence; the target triple sequence includes multiple triples; each triple includes entity information and relation information.

[0035] The construction module is used to build an initial power operation and maintenance knowledge graph based on the target triple sequence, and dynamically associate other modal data in the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0036] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0037] Acquire multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data;

[0038] Based on a pre-defined large model, feature extraction is performed on the text data to obtain a target triplet sequence; the target triplet sequence includes multiple triplets; each triplet includes entity information and relation information.

[0039] An initial power operation and maintenance knowledge graph is constructed based on the target triple sequence, and other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0041] Acquire multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data;

[0042] Based on a pre-defined large model, feature extraction is performed on the text data to obtain a target triplet sequence; the target triplet sequence includes multiple triplets; each triplet includes entity information and relation information.

[0043] An initial power operation and maintenance knowledge graph is constructed based on the target triple sequence, and other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0044] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0045] Acquire multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data;

[0046] Based on a pre-defined large model, feature extraction is performed on the text data to obtain a target triplet sequence; the target triplet sequence includes multiple triplets; each triplet includes entity information and relation information.

[0047] An initial power operation and maintenance knowledge graph is constructed based on the target triple sequence, and other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0048] The aforementioned method, apparatus, equipment, and storage medium for constructing a knowledge graph for power operation and maintenance scenarios acquire multimodal data from power operation and maintenance scenarios. This multimodal data includes image data, text data, tabular data, and time-series data. Feature extraction is performed on the text data according to a pre-defined large model to obtain a target triplet sequence. Each target triplet sequence includes multiple triples, with each triple containing entity and relation information. An initial power operation and maintenance knowledge graph is constructed based on the target triplet sequence, and other modal data from the multimodal data are dynamically associated with this initial knowledge graph to obtain the multimodal knowledge graph. Acquiring multimodal data provides the data foundation for fault reasoning using a large model, effectively solving the severe data silo problem in power systems. The precise description of entity attributes and relationships through the triplet sequence ensures the integrity of the fault deduction chain. Finally, associating multiple modal data with the power operation and maintenance knowledge graph forms a deep cross-modal anchoring and association of data in the power system, improving the accuracy of fault reasoning results. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is an internal structural diagram of a computer device in one embodiment.

[0051] Figure 2 This is a flowchart illustrating a method for constructing a knowledge graph for a power operation and maintenance scenario in one embodiment.

[0052] Figure 3 This is a schematic diagram illustrating the process of feature extraction from multimodal data based on a preset large model in one embodiment.

[0053] Figure 4 This is a flowchart illustrating the process of dividing text data into domain semantic units according to a fine-grained partitioning strategy in one embodiment.

[0054] Figure 5 This is a flowchart illustrating the process of dynamically associating other modal data in multimodal data with an initial power operation and maintenance knowledge graph in one embodiment.

[0055] Figure 6 This is a flowchart illustrating the process of merging vectors in one embodiment;

[0056] Figure 7This is a flowchart illustrating the process of clustering nodes in a power operation and maintenance knowledge graph in one embodiment.

[0057] Figure 8 This is a flowchart illustrating the method for constructing a knowledge graph for a power operation and maintenance scenario in another embodiment;

[0058] Figure 9 This is a structural block diagram of a device for constructing a knowledge graph for power operation and maintenance scenarios in one embodiment. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0061] In the operation and maintenance of modern power systems, power operation and maintenance information, such as transformer operation logs, relay protection monitoring setting sheets, manufacturer technical manuals, PMS equipment ledgers, SCADA equipment data, and UAV inspection thermal imaging images, are core information sources for ensuring the safe and stable operation of the power grid. This information covers key power aspects such as equipment status, consequences of faults and accidents, and hazard planning, and has irreplaceable value for decision-making by dispatchers, operation and maintenance specialists, and field workers. Currently, power operation and maintenance information is mainly obtained from different systems and processed using the "phenomenon-principle-process" rule to diagnose specific power equipment faults. Because unstructured texts such as regulations and drawings are scattered across different systems with formatted or visual data such as SCADA interfaces and infrared thermal images, serious data silos have formed, making it difficult to unify power operation and maintenance information. Traditional keyword-based retrieval systems lack logical reasoning capabilities and struggle to solve complex fault tracing problems. In recent years, large language models have performed well in the general cognitive domain. By inputting power operation and maintenance information into large language models for recognition, power operation and maintenance detection results can be obtained. However, the aforementioned large language models lack rigorous domain logic constraints, are prone to giving incorrect suggestions that do not comply with safety regulations, and have serious safety misconceptions, making it impossible to accurately reason about and resolve faults in power systems.

[0062] In view of the above-mentioned technical problems, this application provides a method for constructing a power operation and maintenance scenario knowledge graph that can accurately infer power system faults. The following embodiments will specifically illustrate the method for constructing the power operation and maintenance scenario knowledge graph.

[0063] The method for constructing a knowledge graph for power operation and maintenance scenarios provided in this application embodiment can be applied to, for example... Figure 1 The computer device shown may be a server, and its internal structure diagram may be as follows. Figure 1 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores multimodal data and a power operation and maintenance knowledge graph for power operation and maintenance scenarios. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for constructing a knowledge graph for power operation and maintenance scenarios.

[0064] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0065] In one exemplary embodiment, such as Figure 2 As shown, a method for constructing a knowledge graph for power operation and maintenance scenarios is provided. This embodiment illustrates the application of this method to computer equipment. In this embodiment, the method includes:

[0066] S201, acquire multimodal data in power operation and maintenance scenarios.

[0067] The multimodal data includes image data, text data, tabular data, and time series data. Image data can be infrared thermal images from UAV inspections. Text and tabular data can be any one or a combination of substation operation logs, relay protection setting sheets, manufacturer's technical manuals, and PMS equipment ledgers. Time series data can be SCADA time series data.

[0068] In the embodiments of this application, the computer device can download multimodal data in the power operation and maintenance scenario from official or academic platforms; it can also use web crawlers to extract text and image content from relevant web pages within the scope of legality and compliance, and use annotation tools to classify the data to obtain multimodal data; it can also perform keyword searches on open source platforms to obtain multimodal data in the power operation and maintenance scenario.

[0069] S202, feature extraction is performed on the text data according to the preset large model to obtain the target triplet sequence.

[0070] The target triple sequence comprises multiple triples; each triple includes two entity information items and one relation information item. Entity information can be any of the following: substation / line, main equipment / component, sensor / measuring point, fault / defect phenomenon, safety regulations / settings / contingency plans, dispatching / maintenance personnel, and remote signaling / telemetry alarm signals. Relationship information can be the relationship or attribute between entity information items. The pre-defined large model can be a large language model, constructed using a power core concept ontology, Chinese prompt word templates, and a review verification mechanism. The Chinese prompt word templates can be Few-shot templates.

[0071] In the embodiments of this application, when the computer device acquires multimodal data under power operation and maintenance, it extracts and integrates information from the multimodal data based on the phenomenon-principle-procedure logic of the power operation and maintenance scenario to obtain a knowledge graph ontology concept set for the power operation and maintenance scenario. The obtained knowledge graph ontology concept set is injected into the initial large model extraction tool, and a Chinese prompt word template and a review verification mechanism are configured. The Chinese prompt word template is designed to force the large language model to output according to the "head entity-relationship-tail entity" structure, and embeds a constraint instruction in the prompt word to prohibit the generation of illusionary output that does not comply with power safety regulations. In each step of the reasoning process of the large model in the triple extraction, the review verification mechanism forces the large model to generate reference marks pointing to specific sentences in the text data. If the generated relation logic contradicts the original electrical logic at the reference mark, the extraction result is rejected, and finally a preset large model is formed. During the inference process, text data is extracted from multimodal data and input into a pre-defined large model for feature extraction. Based on the injected power ontology concept, the large model extracts head entities, relations, and tail entities from the text data, and assigns semantic strength weights representing the degree of relational connection to each relation. All extracted triple relations are accompanied by the original source location of the original operation log or procedure. This yields an initial triple sequence. The elements in the initial triple sequence are then encoded to obtain the target triple sequence.

[0072] S203. Construct an initial power operation and maintenance knowledge graph based on the target triple sequence, and dynamically associate other modal data in the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0073] In the embodiments of this application, the computer device constructs an initial power operation and maintenance knowledge graph using entity vectors of the target triple sequence as nodes and relation vectors as edges. It then extracts other modal data besides text data from the multimodal data and attaches this other modal data to the initial power operation and maintenance knowledge graph, thus obtaining the multimodal knowledge graph. The extracted power entities and electrical topology relationships are stored in a graph database, and the procedure clauses and entity descriptions are transformed into high-dimensional semantic vectors through an embedding model and stored in a vector database for persistent storage. In some embodiments, since the node information of the power operation and maintenance knowledge graph at this time only involves the single modality of text data, in order to improve the accuracy of the power operation and maintenance fault cause reasoning results, other modal data in the multimodal data are associated with the nodes of the power operation and maintenance knowledge graph. That is, modal data other than text data in the multimodal data can be attached to the nodes of the power operation and maintenance knowledge graph, realizing deep association of mixed text and graph data.

[0074] The aforementioned method for constructing a knowledge graph for power operation and maintenance scenarios involves acquiring multimodal data within the power operation and maintenance scenario; extracting features from the text data based on a pre-defined large model to obtain a target triple sequence; constructing an initial power operation and maintenance knowledge graph based on the target triple sequence; and dynamically associating other modal data from the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph. The multimodal data includes image data, text data, tabular data, and time-series data; the target triple sequence includes multiple triples; each triple includes entity information and relational information. Acquiring multimodal data provides the data foundation for fault reasoning using a large model, effectively solving the severe data silo problem in the power system. Furthermore, the precise description of entity attributes and relationships through the triple sequence ensures the integrity of the fault deduction chain. Finally, associating multiple modal data with the power operation and maintenance knowledge graph forms a deep cross-modal anchoring and association of data in the power system, improving the accuracy of fault reasoning results.

[0075] In one exemplary embodiment, such as Figure 3 As shown, feature extraction is performed on the text data according to a pre-defined large model to obtain the target triplet sequence, including:

[0076] S301, divide text data into domain semantic units according to a fine-grained partitioning strategy.

[0077] The fine-grained segmentation strategy can be based on using complete equipment ledger paragraphs as the basic unit, or on using complete handling clauses for a type of fault that are semantically independent. Specifically, using complete equipment ledger paragraphs as the basic unit involves identifying all attribute lines describing the same equipment in the text. These attribute lines can include equipment name, model, commissioning date, rated parameters, and maintenance records. These attribute lines are then aggregated into an indivisible domain semantic unit, prohibiting the splitting of the same equipment's parameter descriptions into two different semantic units. Similarly, using complete handling clauses for a type of fault as an independent unit involves identifying consecutive sentence paragraphs describing the same fault type in the text. These consecutive sentence paragraphs can include fault phenomena, fault causes, handling steps, and preventative measures. These paragraphs are then aggregated into an indivisible domain semantic unit, ensuring the semantic integrity of the fault deduction chain.

[0078] In the embodiments of this application, the computer device can divide text data into domain semantic units by using complete equipment ledger paragraphs as basic units; alternatively, it can divide text data into domain semantic units by using complete processing clauses for a type of fault that are semantically independent from the preceding and following paragraphs. In some embodiments, the computer device can divide a portion of the text data using complete equipment ledger paragraphs as basic units and another portion of the text data using complete processing clauses for a type of fault as independent units. The use of fine-grained division strategies is set according to the actual scenario requirements and is not limited here.

[0079] S302, input the domain semantic units into the preset large model for processing, and generate an initial triple sequence and a reference tag corresponding to each triple.

[0080] In the embodiments of this application, the predefined domain semantic units are input into a preset large model. The large model is forced to generate an initial triple sequence according to the output format constraint instructions. Each triple in the initial triple sequence is represented in the form of head entity-relation-tail entity. During each step of the reasoning process in generating the initial triple sequence, a lookback verification mechanism is triggered synchronously. This lookback verification mechanism forces the large model to output a reference marker pointing to the original text position in the domain semantic unit.

[0081] S303, for each triplet in the initial triplet sequence, compare the power terminology of the triplet with the standard power dictionary, and perform a consistency check between the relational information of the triplet and the original electrical logic corresponding to the reference mark.

[0082] In the embodiments of this application, the computer device first performs terminology verification and matching on the triples, that is, matches the electrical terms in the triples with the standard electrical terms in a preset standard electrical dictionary to obtain the terminology verification result; then, it performs logical consistency verification on the triples, that is, it matches the relational information in the triples with the electrical logic in the original text segment corresponding to the reference mark. Specifically, it performs type matching verification on the extracted relational information of the triples with the relational domain and value domain in the injected ontology concept set. In some embodiments, the execution order of terminology verification and matching and logical consistency verification can be set according to the actual scenario requirements, and is not limited here.

[0083] S304, if the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and the corresponding reference tag are identified as target triples, and a target triple sequence is generated based on multiple target triples.

[0084] Among them, the relationship information can pass the consistency check if the type of the head entity belongs to the domain type allowed by the relationship information, and the type of the tail entity belongs to the value range type allowed by the relationship information.

[0085] In the embodiments of this application, if the power terminology of the triple is consistent with the standard power terminology of the standard power dictionary, or the similarity is greater than a preset similarity threshold, the triple passes the terminology verification and matching. If the power terminology of the triple is inconsistent with the standard power terminology of the standard power dictionary, or the similarity is not greater than the preset similarity threshold, the triple fails the terminology verification and matching. If the type of the head entity in the triple belongs to the domain type allowed by the relation information, and the type of the tail entity belongs to the value range type allowed by the relation information, the triple passes the logical consistency check. If the type of the head entity in the triple does not belong to the domain type allowed by the relation information, or the type of the tail entity belongs to the value range type allowed by the relation information, the triple fails the logical consistency check.

[0086] In some embodiments, triples that pass both terminology matching and logical consistency checks can be encoded to ensure rapid location within the graph. Specifically, the TransE model can be used for encoding. Given a triple (h, r, t) to be encoded, with its head entity vector h, relation vector r, and tail entity vector t, the core idea of ​​the encoding model is to ensure that... Therefore, a distance scoring function is defined. And it is expressed by the following relation (1):

[0087] (1);

[0088] In the formula, L1 and L2 represent two commonly used vector norms, and L1 and L2 are the Manhattan distance and Euclidean distance, respectively.

[0089] The encoding model is trained by minimizing an edge-based loss function to minimize the score of correct triples, thereby generating a knowledge graph vector library containing information on power equipment and regulations. In some embodiments, the encoded triples are combined to form a target triple sequence.

[0090] The above method uses domain semantic unit segmentation strategy and large language model for feature extraction, which ensures the boundary integrity of electrical logic context from the data source, greatly reduces the "illusion" phenomenon generated by large language model when processing complex procedures, and ensures the high confidence of the constructed graph as the foundation of intelligent question answering and reasoning.

[0091] In one exemplary embodiment, such as Figure 4As shown, the text data includes operation logs, procedure data, and ledger data; based on a fine-grained partitioning strategy, the text data is divided into domain semantic units, including:

[0092] S401 preprocesses the operation logs, procedure data, and ledger data to identify electrical boundary identifiers in the text data.

[0093] The electrical boundary identifiers include at least one of the following: equipment ledger start marker, fault clause number, timestamp separator, and procedure chapter title. They may also include the full equipment name, equipment number, and a fault phenomenon description phrase. For example, the equipment ledger start marker might be a field start marker such as "Equipment Name:", "Asset Number:", or "Model:". The fault clause number might be a hierarchical number such as "3.2.1", "Clause Five", or "Article 12" in the procedure text. The timestamp separator might be "2025-01-01 08:23:15" or "08:23", indicating the time boundary of different events. The procedure chapter title might be "Chapter Two Transformer Operation Procedures" or "4.3 Overcurrent Protection Setting". The full equipment name and equipment number might be "Main Transformer No. 3" or "Circuit Breaker 5021", etc., as markers for the start of a new paragraph; the fault phenomenon description phrase might be "Protection Action", "Reclosing Failure", or "CT Disconnection".

[0094] In embodiments of this application, when a computer device acquires text data such as operation logs, procedure data, and ledger data, it can identify and recognize the operation logs, procedure data, and ledger data based on identifier recognition technology to obtain electrical boundary identifiers. In some embodiments, the computer device can also identify and recognize the operation logs, procedure data, and ledger data using a regular expression model to obtain electrical boundary identifiers. In some embodiments, the computer device can also identify and recognize the operation logs, procedure data, and ledger data using a conceptual data model to obtain electrical boundary identifiers.

[0095] S402 uses electrical boundary identifiers as segmentation anchors and ledger paragraphs for specific equipment or handling clauses for a type of fault as independent units to segment text data into multiple candidate semantic units.

[0096] In the embodiments of this application, the computer device can treat a ledger segment of a specific device as an independent unit, use the electrical boundary identifier as the segmentation anchor point, identify all attribute lines describing the same device in the text data, and aggregate these consecutive attribute lines describing the same device into a candidate semantic unit. It is prohibited to split the parameter description of the same device into two different candidate semantic units. The candidate semantic units of all devices are integrated to obtain multiple candidate semantic units of the text data.

[0097] In some embodiments, the computer device may also treat the handling clauses of a type of fault as independent units, using electrical boundary identifiers as segmentation anchors, to identify consecutive sentence paragraphs in text data describing the same fault type. These consecutive sentence paragraphs should contain a complete fault derivation chain, typically including a description of the fault phenomenon, fault cause analysis, handling procedures, and preventative measures recommendations. These paragraph contents are then aggregated into a candidate semantic unit. The candidate semantic units for all fault types are integrated to obtain multiple candidate semantic units for the text data.

[0098] S403, perform electrical logic integrity verification on each candidate semantic unit, and determine the candidate semantic units that pass the integrity verification as domain semantic units.

[0099] In the embodiments of this application, for each candidate semantic unit, the computer device performs an electrical logic integrity check on the candidate semantic unit, that is, checks whether the candidate semantic unit contains at least one complete device entity identifier and whether the candidate semantic unit contains at least one predicate description of an electrical action or state change. If it is found that the candidate semantic unit contains at least one complete device entity identifier and at least one predicate description of an electrical action or state change, the candidate semantic unit passes the integrity check and is designated as a domain semantic unit.

[0100] In some embodiments, it is checked whether a contextual dependency exists between any two candidate semantic units. This contextual dependency can be a causal relationship. Specifically, it is checked whether there is a causal conjunction in the first candidate semantic unit, and the object of the conjunction points to the fault phenomenon or device state in the second candidate semantic unit. If a contextual dependency exists between any two candidate semantic units, the two candidate semantic units are merged into a single domain semantic unit.

[0101] The above method achieves the integrity of electrical logic through domain semantic unit segmentation strategy, which improves the accuracy of subsequent knowledge graph reasoning results.

[0102] In one exemplary embodiment, such as Figure 5 As shown, other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph, including:

[0103] S501 extracts image data and time series data from multimodal data.

[0104] In the embodiments of this application, the computer device constructs an initial power operation and maintenance knowledge graph using entity information of each triple as nodes and relation information as edges. The relation information can be divided into electrical topology connection relationships and semantic association relationships. A first weight is assigned to the electrical topology connection relationships, and a second weight is assigned to the semantic association relationships. The first and second weights can be set according to the actual scenario and are not limited here. Then, the computer device extracts image data and time-series data from the multimodal data to construct the initial power operation and maintenance knowledge graph. The image data and time-series data are encoded using the TransE model to obtain visual feature vectors and time-series numerical feature vectors.

[0105] S502, image data and time series data are mounted onto the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0106] In the embodiments of this application, the computer device determines the entity nodes corresponding to the image data and time series data from the initial power operation and maintenance knowledge graph according to the correspondence of multimodal data. Based on the entity nodes, the image data and time series data are dynamically attached to the entity nodes corresponding to the power operation and maintenance knowledge graph.

[0107] In practical applications, when equipment malfunctions, the system leverages multimodal data and uses the aforementioned joint feature representation to perform precise retrieval enhancement and generation, providing intelligent decision support that is free of illusions and has traceability.

[0108] In one exemplary embodiment, such as Figure 6 As shown, it also includes:

[0109] S601 performs vector transformation on multimodal knowledge graphs to generate a knowledge graph vector library.

[0110] Among them, the vectors in the knowledge graph vector library correspond uniquely to the nodes and relations in the multimodal knowledge graph.

[0111] In the embodiments of this application, the computer device first uses an embedding model to convert the multimodal knowledge graph into vector information and stores all vector information in a database to form a knowledge graph vector library. The embedding model may include a graph embedding model, an image embedding model, and a temporal embedding model. In some embodiments, the graph embedding model is used to perform vector transformation on the node data and edge data in the multimodal knowledge graph to generate graph feature vectors; the image embedding model is used to perform vector transformation on the image data in the multimodal knowledge graph to generate image feature vectors; the temporal embedding model is used to perform vector transformation on the time series data in the multimodal knowledge graph to generate temporal feature vectors; finally, the graph feature vectors, image feature vectors, and temporal feature vectors are concatenated to obtain a multimodal feature vector; after vector transformation is completed for each node in the multimodal knowledge graph, a knowledge graph vector library is generated.

[0112] In some embodiments, the underlying graph device node is set as follows: Compare it with the time-series numerical feature vector in the SCADA system and visual feature vectors in UAV infrared images Multimodal features are obtained by splicing and mapping. For example, real-time current curve features, hot spot temperature distribution features, and device nodes are spliced ​​and mapped. Multimodal features can be represented by the following relationship (2):

[0113] (2);

[0114] In practical applications, when equipment malfunctions, the system leverages multimodal data and uses the aforementioned joint feature representation to perform precise retrieval enhancement and generation, providing intelligent decision support that is free of illusions and has traceability.

[0115] S602: For any two vectors in the knowledge graph vector library, calculate the similarity between the two vectors.

[0116] Vector similarity includes entity similarity and attribute description similarity.

[0117] In the embodiments of this application, for any two vectors in the knowledge graph vector library, the similarity of the entity parts of the two vectors can be calculated to obtain entity similarity. The similarity of the relation parts of the two vectors can be calculated to obtain attribute description similarity. The entity similarity and attribute description similarity are weighted and summed to obtain the final similarity between the two vectors. In some embodiments, the similarity of the head entities of any two triples in the multimodal knowledge graph can be calculated to obtain entity similarity. The similarity of the tail entities of the two triples can also be calculated. Furthermore, the similarity of the concatenated vectors of the head and tail entities of the two triples can also be calculated to obtain entity similarity. When semantic alignment of synonymous devices is required, such as "No. 1 main transformer" and "#1 transformer", a joint evaluation using vector cosine similarity and Jaccard text similarity is employed.

[0118] In some embodiments, entity similarity can be represented by the following relation (3):

[0119] (3);

[0120] In the formula, and It is an entity vector.

[0121] In some embodiments, attribute description similarity can be represented by the following relation (4):

[0122] (4);

[0123] In the formula, and These are the text description character sets representing the two entities, i.e., the relation vectors. This indicates the number of feature words in the character set.

[0124] S603: Merge vectors with a similarity greater than a preset similarity threshold.

[0125] In the embodiments of this application, the computer device compares the similarity with a preset similarity threshold. When the similarity of the vectors is greater than the preset similarity threshold, an entity merging operation is triggered to merge the two vectors in the knowledge graph vector library. In some embodiments, corresponding nodes and edges in the multimodal knowledge graph can also be merged, that is, one node is retained and the topological relationships (edges) of other similar nodes are migrated to that node.

[0126] By using similarity judgment to disambiguate entities in the topology, the generated power operation and maintenance knowledge graph is made consistent with the actual state of power equipment, thus improving the accuracy of fault reasoning results.

[0127] In one exemplary embodiment, such as Figure 7 As shown, the method also includes:

[0128] S701 uses a graph community detection algorithm to cluster nodes in the power operation and maintenance knowledge graph, generating multiple subsystem communities.

[0129] Among them, the graph community detection algorithm can be the Leiden community algorithm.

[0130] In embodiments of this application, a computer device can recursively cluster device nodes to generate subsystem communities based on electrical topology relationships and strong correlations between triples using a graph community detection algorithm. After all nodes are partitioned, multiple subsystem communities are generated. In some embodiments, the strong correlation between triples can be determined by semantic strength weighting. If the semantic strength of any two triples exceeds a set threshold, the corresponding triples are considered strongly correlated and should be further divided into a subsystem community. For example, the circuit breakers in a certain bay and their associated CTs, PTs, and disconnectors can be divided into a subsystem community.

[0131] S702: For each subsystem community, extract all nodes and their associated factual statements within the subsystem community to generate a context information package for the subsystem community.

[0132] In the embodiments of this application, the computer device extracts factual statements of different device nodes from the operation logs in the text data. These factual statements include information such as the subject, fault category, status label, time range, and detailed explanation. The factual statements of all nodes in each subsystem community are integrated to obtain the context information package of that subsystem community.

[0133] S703, evaluate the context information package according to the preset health assessment model to obtain a structured health assessment summary of the subsystem community.

[0134] The preset health assessment model can be a large language model, configured with an operation and maintenance health assessment prompt template. The operation and maintenance health assessment prompt template must contain at least one of the following instructions: force the large language model to output a summary in JSON format, with output fields including at least: subsystem identifier, overall health level, list of major risk points, and recommended operation and maintenance measures; require the model to specify the specific device name associated with each risk point and the corresponding original source of evidence when generating each risk point; prohibit the model from generating fault predictions beyond the scope of devices included in the context information packet.

[0135] In the embodiments of this application, for each subsystem community, the computer device inputs the context information packet into the preset health assessment model, loads an operation and maintenance health assessment prompt word template, and forces the preset health assessment model to output a structured health assessment summary for that subsystem community.

[0136] S704 attaches the structured health assessment summary as an attribute of the subsystem community to the knowledge graph of the power operation and maintenance scenario.

[0137] In the embodiments of this application, the structured health assessment summary is used as an attribute of the subsystem community node and attached to the power operation and maintenance scenario knowledge graph. In some embodiments, the generated subsystem community structured health assessment summary can also be linked to all underlying device entity nodes within the subsystem community to establish a reverse index, so that any device node can trace back to its subsystem community and its assessment summary.

[0138] By outputting subsystem-level health assessment summaries, the efficiency of assessment and decision-making for roles such as operation and maintenance specialists and dispatchers when facing complex faults is greatly improved.

[0139] In addition to the methods described in all the above embodiments, a method for mitigating voltage fluctuations in distribution networks is also provided, such as... Figure 8 As shown, the method includes:

[0140] S801 acquires multimodal data in power operation and maintenance scenarios; multimodal data includes image data, text data, tabular data, and time series data;

[0141] S802 preprocesses the operation log, procedure data, and ledger data to identify electrical boundary identifiers in the text data; the electrical boundary identifiers include at least one of the following: equipment ledger start mark, fault clause number, timestamp separator, and procedure chapter title.

[0142] S803 uses electrical boundary identifiers as segmentation anchors and ledger paragraphs for specific equipment or handling clauses for a type of fault as independent units to segment text data into multiple candidate semantic units.

[0143] S804, Perform electrical logic integrity verification on each candidate semantic unit, and determine the candidate semantic unit that passes the integrity verification as the domain semantic unit;

[0144] S805, input the domain semantic unit into the preset large model for processing, and generate an initial triple sequence and a reference tag corresponding to each triple;

[0145] S806, for each triplet in the initial triplet sequence, compare the power terminology of the triplet with the standard power dictionary, and perform consistency verification between the relational information of the triplet and the original electrical logic corresponding to the reference mark;

[0146] S807, if the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and the corresponding reference tag are identified as target triples, and a target triple sequence is generated based on multiple target triples; the target triple sequence includes multiple triples; each triple includes entity information and relation information;

[0147] S808, Construct an initial power operation and maintenance knowledge graph based on the target triple sequence;

[0148] S809, extract image data and time series data from the multimodal data; mount the image data and time series data onto the initial power operation and maintenance knowledge graph to obtain the multimodal knowledge graph;

[0149] S810 performs vector transformation on the multimodal knowledge graph to generate a knowledge graph vector library; the vectors in the knowledge graph vector library uniquely correspond to the nodes and relations in the multimodal knowledge graph.

[0150] S811 calculates the similarity between any two vectors in the knowledge graph vector library; the similarity includes entity similarity and attribute description similarity.

[0151] S812 merges vectors with a similarity greater than a preset similarity threshold.

[0152] S813, cluster the nodes in the power operation and maintenance knowledge graph according to the graph community detection algorithm to generate multiple subsystem communities;

[0153] S814: For each subsystem community, extract all nodes and their associated factual statements within the subsystem community to generate a context information package for the subsystem community;

[0154] S815, evaluate the context information package according to the preset health assessment model to obtain a structured health assessment summary of the subsystem community;

[0155] S816 incorporates structured health assessment summaries as attributes of the subsystem community and mounts them into the knowledge graph of the power operation and maintenance scenario.

[0156] Each of the above steps has been described in the foregoing embodiments. For details, please refer to the foregoing content. They will not be repeated here.

[0157] It should be understood that although the steps in the flowcharts of the above embodiments 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 flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0158] Based on the same inventive concept, this application also provides a power operation and maintenance scenario knowledge graph construction apparatus for implementing the above-mentioned method for constructing a power operation and maintenance scenario knowledge graph. The solution provided by this apparatus is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the power operation and maintenance scenario knowledge graph construction apparatus provided below can be found in the limitations of the power operation and maintenance scenario knowledge graph construction method described above, and will not be repeated here.

[0159] In one exemplary embodiment, such as Figure 9 As shown, a device for constructing a knowledge graph for power operation and maintenance scenarios is provided, including: an acquisition module 91, an extraction module 92, and a construction module 93, wherein:

[0160] The acquisition module 91 is used to acquire multimodal data in power operation and maintenance scenarios; the multimodal data includes image data, text data, tabular data and time series data;

[0161] The extraction module 92 is used to extract features from text data according to a preset large model to obtain a target triple sequence; the target triple sequence includes multiple triples; each triple includes entity information and relation information.

[0162] Module 93 is used to construct an initial power operation and maintenance knowledge graph based on the target triple sequence, and to dynamically associate other modal data in the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0163] In an exemplary embodiment, the extraction module 92 described above is specifically used for:

[0164] Text data is divided into domain semantic units based on a fine-grained partitioning strategy;

[0165] The domain semantic units are input into a pre-defined large model for processing, generating an initial triplet sequence and a reference tag for each triplet;

[0166] For each triplet in the initial triplet sequence, the power terminology of the triplet is compared with the standard power dictionary, and the relationship information of the triplet is checked for consistency with the original electrical logic corresponding to the reference mark.

[0167] If the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and its corresponding reference tag are identified as target triples, and a target triple sequence is generated based on multiple target triples.

[0168] In an exemplary embodiment, the extraction module 92 described above is specifically used for:

[0169] The operation logs, procedure data, and ledger data are preprocessed to identify electrical boundary identifiers in the text data. Electrical boundary identifiers include at least one of the following: equipment ledger start mark, fault clause number, timestamp separator, and procedure chapter title.

[0170] Using electrical boundary identifiers as segmentation anchors, and specific equipment ledger paragraphs or fault handling clauses as independent units, text data is segmented into multiple candidate semantic units.

[0171] Perform electrical logic integrity verification on each candidate semantic unit, and determine the candidate semantic units that pass the integrity verification as domain semantic units.

[0172] In an exemplary embodiment, the aforementioned building module 93 is specifically used for:

[0173] Extracting image data and time series data from multimodal data;

[0174] Image data and time series data are mounted onto the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph.

[0175] In an exemplary embodiment, the aforementioned building module 93 is specifically used for:

[0176] Vector transformation is performed on the multimodal knowledge graph to generate a knowledge graph vector library; the vectors in the knowledge graph vector library uniquely correspond to the nodes and relations in the multimodal knowledge graph.

[0177] For any two vectors in the knowledge graph vector database, calculate the similarity between the two vectors; the similarity includes entity similarity and attribute description similarity.

[0178] Vectors with a similarity greater than a preset similarity threshold are merged.

[0179] In one exemplary embodiment, the above-described apparatus further includes a clustering module for:

[0180] The nodes in the power operation and maintenance knowledge graph are clustered based on the graph community detection algorithm to generate multiple subsystem communities;

[0181] For each subsystem community, extract all nodes and their associated factual claims within the subsystem community to generate a context information package for the subsystem community;

[0182] The context information package is evaluated based on the preset health assessment model to obtain a structured health assessment summary of the subsystem community;

[0183] The structured health assessment summary is attached as an attribute of the subsystem community to the knowledge graph of the power operation and maintenance scenario.

[0184] Each module in the aforementioned power operation and maintenance scenario knowledge graph construction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0185] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the steps in the above-described method embodiments.

[0186] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0187] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0188] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0189] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0190] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a knowledge graph for power operation and maintenance scenarios, characterized in that, The method includes: Acquire multimodal data in power operation and maintenance scenarios; the multimodal data includes image data, text data, tabular data, and time series data; the text data includes operation logs, procedure data, and ledger data; Based on a pre-defined large model, feature extraction is performed on the text data to obtain a target triple sequence; the target triple sequence includes multiple triples; each triple includes entity information and relation information; An initial power operation and maintenance knowledge graph is constructed based on the target triple sequence, and other modal data in the multimodal data are dynamically associated with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph. The step of extracting features from the text data according to a preset large model to obtain a target triplet sequence includes: The text data is divided into domain semantic units according to a fine-grained partitioning strategy; The domain semantic units are input into the preset large model for processing to generate an initial triplet sequence and a reference tag corresponding to each triplet. For each triplet in the initial triplet sequence, the power terminology of the triplet is compared with the standard power dictionary, and the relationship information of the triplet is checked for consistency with the original electrical logic corresponding to the reference mark. If the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and the corresponding reference tag are identified as target triples, and a sequence of target triples is generated based on multiple target triples. The step of dividing the text data into domain semantic units according to a fine-grained partitioning strategy includes: The operation logs, procedure data, and ledger data are preprocessed to identify electrical boundary identifiers in the text data; the electrical boundary identifiers include at least one of the following: equipment ledger start mark, fault clause number, timestamp separator, and procedure chapter title. Using the electrical boundary identifier as the segmentation anchor point, and the ledger paragraphs of specific equipment or the handling clauses of a type of fault as independent units, the text data is segmented into multiple candidate semantic units; Perform electrical logic integrity verification on each of the candidate semantic units, and determine the candidate semantic units that pass the integrity verification as the domain semantic units; The step of dynamically associating other modal data in the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph includes: Extract image data and time series data from the multimodal data; The image data and the time series data are mounted onto the initial power operation and maintenance knowledge graph to obtain the multimodal knowledge graph.

2. The method according to claim 1, characterized in that, The method further includes: The multimodal knowledge graph is vectorized to generate a knowledge graph vector library; the vectors in the knowledge graph vector library uniquely correspond to the nodes and relationships in the multimodal knowledge graph. For any two vectors in the knowledge graph vector library, calculate the entity similarity and attribute description similarity between the two vectors respectively, and then perform a weighted sum of the entity similarity and attribute description similarity to obtain the similarity between the two vectors. When the similarity between the two vectors is greater than a preset similarity threshold, the two vectors are merged.

3. The method according to any one of claims 1-2, characterized in that, The method further includes: The nodes in the power operation and maintenance knowledge graph are clustered according to the graph community detection algorithm to generate multiple subsystem communities; For each subsystem community, extract all nodes and their associated factual statements within the subsystem community to generate a context information package for the subsystem community; The context information package is evaluated according to a preset health assessment model to obtain a structured health assessment summary of the subsystem community; The structured health assessment summary is used as an attribute of the subsystem community and attached to the power operation and maintenance scenario knowledge graph.

4. A device for constructing a knowledge graph for power operation and maintenance scenarios, characterized in that, The device includes: The acquisition module is used to acquire multimodal data in power operation and maintenance scenarios; the multimodal data includes image data, text data, tabular data, and time series data; the text data includes operation logs, procedure data, and ledger data. An extraction module is used to extract features from the text data according to a preset large model to obtain a target triple sequence; the target triple sequence includes multiple triples; each triple includes entity information and relation information; A construction module is used to construct an initial power operation and maintenance knowledge graph based on the target triple sequence, and to dynamically associate other modal data in the multimodal data with the initial power operation and maintenance knowledge graph to obtain a multimodal knowledge graph; The extraction module is specifically used for: The text data is divided into domain semantic units according to a fine-grained partitioning strategy; The domain semantic units are input into the preset large model for processing to generate an initial triplet sequence and a reference tag corresponding to each triplet. For each triplet in the initial triplet sequence, the power terminology of the triplet is compared with the standard power dictionary, and the relationship information of the triplet is checked for consistency with the original electrical logic corresponding to the reference mark. If the power terminology of the triple is consistent with the standard power dictionary and the relation information passes the consistency check, the triple and the corresponding reference tag are identified as target triples, and a sequence of target triples is generated based on multiple target triples. The extraction module is specifically used for: The operation logs, procedure data, and ledger data are preprocessed to identify electrical boundary identifiers in the text data; the electrical boundary identifiers include at least one of the following: equipment ledger start mark, fault clause number, timestamp separator, and procedure chapter title. Using the electrical boundary identifier as the segmentation anchor point, and the ledger paragraphs of specific equipment or the handling clauses of a type of fault as independent units, the text data is segmented into multiple candidate semantic units; Perform electrical logic integrity verification on each of the candidate semantic units, and determine the candidate semantic units that pass the integrity verification as the domain semantic units; The building module is also used for: Extract image data and time series data from the multimodal data; The image data and the time series data are mounted onto the initial power operation and maintenance knowledge graph to obtain the multimodal knowledge graph.

5. The apparatus according to claim 4, characterized in that, The building module is also used for: The multimodal knowledge graph is vectorized to generate a knowledge graph vector library; the vectors in the knowledge graph vector library uniquely correspond to the nodes and relationships in the multimodal knowledge graph. For any two vectors in the knowledge graph vector library, calculate the entity similarity and attribute description similarity between the two vectors respectively, and then perform a weighted sum of the entity similarity and attribute description similarity to obtain the similarity between the two vectors. When the similarity between the two vectors is greater than a preset similarity threshold, the two vectors are merged.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images