Electric power semantic large model adaptive data query and dynamic response method and system

Abstract

The invention relates to the technical field of intelligent data processing and artificial intelligence in the electric power industry, and provides an electric power semantic large model adaptive data query and dynamic response method and system.The method comprises the steps that electric power industry natural language query and related images are obtained, and after feature extraction is conducted through a large language model and a deep learning model, the images are obtained; the knowledge graph and the visual question and answer graph neural network are utilized, features of different modes are unified into the same representation in a bidirectional fusion mode, fused features are obtained, after intelligent error correction and natural language query optimization are conducted through the hypergraph neural network and multiple rounds of interaction, an extended query statement is obtained through hypergraph reasoning, and a query result is obtained. Converting the extended query statement into a structured query statement through an autoregression model to obtain a query result; and performing dynamic permission adaptation, sensitive data protection and compliance and security check on a query result to realize dynamic response. And the query accuracy and efficiency of the power industry are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent data processing and artificial intelligence technology in the electric power industry, and in particular to a method and system for adaptive data query and dynamic response of a large electric power semantic model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In the process of digital transformation of the power industry, massive amounts of data provide important support for equipment management, fault detection, predictive analysis, etc. However, power system data is often complex in structure, involving a variety of information such as equipment type, operating parameters, fault records, etc., and the data types include multimodal information such as text, images, and time series. How to efficiently and accurately query and analyze this data and gain insights from it has always been a technical challenge facing the industry.

[0004] Traditional data query methods often require users to have professional database knowledge to construct effective SQL queries, which not only requires high skills from operators, but also takes a long time when complex queries are involved. At the same time, power data queries involve a large amount of sensitive information, especially the operating status and fault records of key equipment. Data security and permission management are also important considerations. At present, static permission control is difficult to flexibly respond to the diverse needs of users, which can easily lead to data leakage risks or affect the integrity of information acquisition due to insufficient permissions. In addition, power data queries often contain vague or imprecise descriptions, especially natural language queries entered by non-technical users, which are difficult for traditional systems to accurately understand and respond to.

[0005] In recent years, with the development of natural language processing (NLP) and deep learning (DL) technologies, intelligent query methods can automatically parse users' natural language needs to a certain extent and generate corresponding database queries. However, the application of these technologies in the power industry still faces challenges, such as how to effectively integrate multiple data such as text and images; how to build a multimodal knowledge graph that conforms to the characteristics of the power industry; how to ensure data security during the query process; and how to automatically identify and correct semantic deviations in user queries. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides an adaptive data query and dynamic response method and system for a large electric power semantic model. Through adaptive context perception, deep fusion of multimodal data and intelligent error correction mechanism, it can quickly parse and understand users' natural language queries, especially for professional terms and complex scenarios in the electric power industry, and can automatically generate accurate structured query statements. Users can quickly obtain the required information without manual screening and repeated error correction, which greatly improves the accuracy and efficiency of queries.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for adaptive data query and dynamic response of a large power semantic model, which comprises: Obtain natural language queries and related images of the power industry, extract features through a large language model and a deep learning model respectively, and then use the knowledge graph and the visual question-answering graph neural network to unify the features of different modalities into the same representation in a two-way fusion manner to obtain fused features; based on the fused features, perform intelligent error correction through a hypergraph neural network and multiple rounds of interaction, optimize the natural language query, obtain an extended query statement through hypergraph reasoning, and convert the extended query statement into a structured query statement through an autoregressive model to obtain the query result; For query results, dynamic permission adaptation, sensitive data protection, and compliance and security checks are performed to achieve dynamic response.

[0008] Furthermore, the step of generating the fusion feature includes: Based on the knowledge graph and query statement features, concepts are extracted from natural language queries to build a concept graph. Based on image features, a scene graph generator is used to extract visual information from images and build a scene graph. The nodes in the scene graph are connected to the nodes in the concept graph through super nodes. Through the visual question-answering graph neural network, the features of the nodes in the scene graph and the nodes in the concept graph are transferred through the super node, and after iteratively updating the features of the super node, the nodes in the scene graph and the nodes in the concept graph, the fused features are spliced.

[0009] Furthermore, the hypergraph neural network is used to automatically identify and correct semantic errors or logical deviations in the natural language query to obtain candidate queries.

[0010] Furthermore, in each round of interaction, after receiving the candidate query selected by the user, the candidate query selected by the user is weightedly fused with the natural language query of the previous round to obtain the natural language query of this round, and the fusion features and candidate query are regenerated until the candidate query no longer changes.

[0011] Furthermore, the dynamic adaptation of permissions includes: verifying the identity of the user, evaluating the access permissions according to the user's attributes, and performing fuzzy processing on the query results according to the access permissions.

[0012] Furthermore, the sensitive data protection includes: determining whether the data in the query result is sensitive data, and if sensitive data exists, performing differential privacy processing on the query result.

[0013] Furthermore, the compliance and security check includes: automatically reviewing the query results through a rule-based compliance engine, and if the review fails, blocking the response of the query results, or shielding or obfuscating the query results during the response process.

[0014] The second aspect of the present invention provides a power semantic big model adaptive data query and dynamic response system, which includes: The adaptive data query module is configured to: obtain natural language queries and related images of the power industry, extract features through a large language model and a deep learning model respectively, and then use a knowledge graph and a visual question-answering graph neural network to unify the features of different modalities into the same representation in a two-way fusion manner to obtain fused features; based on the fused features, perform intelligent error correction through a hypergraph neural network and multiple rounds of interaction, optimize the natural language query, obtain an extended query statement through hypergraph reasoning, and convert the extended query statement into a structured query statement through an autoregressive model to obtain a query result; The dynamic response module is configured to: dynamically adapt permissions, protect sensitive data, and perform compliance and security checks on query results to achieve dynamic response.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for adaptive data query and dynamic response of a large power semantic model as described above.

[0016] The fourth aspect of the present invention provides a computer device, including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein when the processor executes the program, the steps in the method for adaptive data query and dynamic response of a large power semantic model as described above are implemented.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention innovatively proposes an adaptive context-aware, deep fusion and intelligent error correction mechanism for multimodal data, which can quickly parse and understand users' natural language queries, especially for professional terms and complex scenarios in the power industry, and can automatically generate accurate structured query statements. Users can quickly obtain the required information without manual screening and repeated error correction, greatly improving work efficiency; and using deep reasoning of hypergraph structures, it can automatically identify complex potential association patterns between data, providing users with deep data insights beyond the query content. This automated association discovery helps users identify the causes of equipment abnormalities or potential failures, and provides preventive maintenance recommendations, thereby effectively improving the reliability and safety of the power system, and providing intelligent support for the long-term operation and operational management of equipment.

[0018] The present invention innovatively proposes to use a hypergraph neural network to automatically detect and correct semantic and logical deviations in queries, and provide users with query options that meet their intentions through multiple rounds of interactions. For users with fuzzy queries or who are not familiar with professional terms, this function helps them find the information they need more quickly, significantly improving the user experience. It is not only suitable for professional users, but also for a wider user group.

[0019] The present invention innovatively proposes a dynamic permission adaptation and sensitive data protection mechanism, which can automatically adjust the display method of the results according to the user's permission level and the sensitivity of the queried data. For example, for users with different permissions, data shielding, fuzzy display or differential privacy protection can be automatically performed, ensuring the safe use of data on the basis of compliance. This not only helps the power industry to achieve compliance management in data use, but also avoids the risk of sensitive data leakage, thereby further improving the security of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 It is a flow chart of the method for adaptive data query and dynamic response of a large power semantic model according to the first embodiment of the present invention; Figure 2 It is a structural diagram of a computer device according to a fourth embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0023] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0024] Embodiment 1 This embodiment provides a method for adaptive data query and dynamic response of a large power semantic model.

[0025] The adaptive data query and dynamic response method of the power semantic big model provided in this embodiment proposes an adaptive NLP2DL power semantic big model to realize intelligent data query and dynamic response. Through adaptive multimodal deep fusion, intelligent error correction and dynamic authority management, an intelligent query platform is constructed that can understand user natural language, protect data privacy, and realize deep correlation analysis, so as to meet the power industry's needs for intelligent, flexible and secure data query.

[0026] The power semantic big model adaptive data query and dynamic response method provided in this embodiment is an intelligent data processing solution for the power industry. It combines the reasoning capabilities of big models, multimodal fusion and hypergraph structures to realize the intelligent conversion of user natural language queries to precise SQL data queries; dynamically adapt to user permissions and data sensitivity, provide differential privacy and fuzzy display to ensure data security, and use deep correlation analysis to automatically discover the potential correlations behind the queries and generate intelligent recommendations; it not only improves the accuracy and efficiency of data queries, but also provides deep data insights and security guarantees, providing key support for the digital and intelligent transformation of the power industry.

[0027] The adaptive data query and dynamic response method of the electric power semantic big model provided in this embodiment has extremely strong adaptability and can flexibly support various application scenarios in the electric power industry, including equipment management, fault monitoring, data analysis, compliance checking, etc. This multi-scenario adaptability has become a key technology in the intelligent and digital transformation of the electric power industry, providing important support for optimizing production, management and services, and bringing substantial improvements in query accuracy, data security, correlation analysis and user experience.

[0028] The power semantic big model adaptive data query and dynamic response method provided in this embodiment is as follows: Figure 1 As shown, the specific steps are as follows: Step 1: Obtain the natural language query input by the user and the image related to the query; realize natural language understanding and reasoning through multimodal knowledge graph and graph neural network (GNN), especially for context perception in the specific context of the power industry.

[0029] In multimodal data fusion, it is necessary to process multimodal data such as text and images, and fuse these different forms of information to better understand user queries. The core of multimodal data fusion is to unify the representation of these heterogeneous data sources by combining large language models (LLM) and knowledge graphs, and using graph neural networks (GNNs) to achieve deep reasoning for complex queries.

[0030] Step 101: Receive a natural language query (text data) input by a user; in parallel, data in other modalities, such as images related to the query (such as photos of power equipment), are also received.

[0031] Among them, query-related images refer to input images associated with the user's query statement, which can usually be divided into the following categories in actual scenarios: (1) Equipment status images: For example, photos of the operating status of power equipment (such as transformers, transmission lines, and switchgear) to support queries on equipment operating conditions or problem location.

[0032] (2) On-site surveillance images: Surveillance images of construction or maintenance sites provide contextual information to assist in the query of specific objects or events involved.

[0033] (3) Sample image: A reference image uploaded by the user for comparison, search, or identification, such as “find damaged transformer equipment similar to this one.”

[0034] (4) Historical archive images: stored historical record images, combined with query statements to retrieve related records.

[0035] The relationship between the image and the query statement: (1) Image as the core object of the query: The query statement describes the specific target or event contained in the image. Query statement: "Is there anything abnormal in the transformer in the image?" Image content: Photo of the transformer taken on site.

[0036] (2) Images provide query context: The information in the query is incomplete, and images provide supplementary information. Query: "Check the maintenance records for the past three months." Image content: Transformer image with device ID. Then, extract the device ID from the image. Find relevant data for the device in the maintenance records.

[0037] (3) Images as references for search: Users want to find other instances similar to the one in the image. Query: "Find maintenance records for equipment with similar damage." Image content: Photo of a transformer showing damage. Then, extract damage features from the image. Retrieve records in the database that match these features.

[0038] In order to integrate these different modal information, a multimodal GNN model is used, which can handle the complex relationship between structured and unstructured data.

[0039] Step 102: The natural language query is parsed through a pre-trained language model (e.g., a large language model such as GPT), a contextual representation of the question is generated, encoded into a vector representation, and query statement features are obtained; and image data is extracted through a deep learning model (e.g., a VIT (vision transformer) model) to obtain image features.

[0040] Step 103: Generate scene graph and concept graph.

[0041] The input multimodal information (text and image) is converted into a unified semantic representation through a graph neural network (GNN), including: based on query statement features and knowledge graphs, the power knowledge base (knowledge graph) is used to extract relevant concepts from the query statement and construct a concept graph; based on image features, a pre-trained scene graph generator (Scene Graph Generator) is used to extract visual information from the image and construct a scene graph.

[0042] Among them, the scene graph is mainly used to describe the objects in the input image and their relationships. It is a structured expression of visual information, with nodes representing objects in the image (such as "car", "person"), and edges representing the relationships between these objects (such as "behind", "near").

[0043] For example, suppose the input image is a photo containing the following content: a car parked on the street, a pedestrian next to the car, and a tree next to the street. The scene graph represents: nodes (objects): car, pedestrian, tree, street; edges (relationships): car-parked-on-street, pedestrian-near-car, tree-next-to-street.

[0044] Among them, concept graphs are mainly used to describe background knowledge related to natural language queries. They are usually extracted from knowledge bases (such as ConceptNet). They represent the vocabulary, concepts and their relationships related to the query in a structured manner.

[0045] For example, suppose the user's natural language query is: "What are the abnormalities of high-load transformers in the past week?" The concept graph represents: nodes (concepts): high-load transformers, abnormalities, past week, power equipment; edges (relationships): high-load transformers-belong to-power equipment, abnormalities-occurred-in-the-past-week.

[0046] This structured concept graph helps extract key concepts and their contextual relationships from the language so as to perform multimodal reasoning in conjunction with the scene graph.

[0047] Step 104: In this embodiment, the natural language query input by the user is parsed through a combination of vision and language.

[0048] The VQA-GNN (Visual Question Answering Graph Neural Network) model uses a bidirectional fusion approach to unify the information of these different modalities into the same representation, and performs reasoning through the message passing mechanism of the graph neural network.

[0049] Among them, the input of VQA-GNN is divided into three main sources: images, natural language questions and knowledge base; the output includes feature vectors that can be used for intelligent query completion or fuzzy completion.

[0050] Among them, the processing steps of the VQA-GNN (Visual Question Answering Graph Neural Network) model include: (1) Construction of multimodal semantic graph.

[0051] Among them, the super node is a special node used to connect the scene graph and the concept graph: QA-context Node: Represents a natural language query and its context. Extracts natural language questions and generates semantic features through language models (such as large language models such as GPT).

[0052] QA-concept Node: represents the global visual semantics extracted from the scene graph and is related to the overall context of the image.

[0053] The connection rules include: Connection between scene graph and supernode: Each scene graph node (such as "transformer") is connected to QA-contextNode through an edge. Edge type: Image Edge: describes the relationship between the scene graph node and the query context (such as "transformer appears in the image").

[0054] Connection between concept graph and supernode: Each concept graph node (such as "fault") is connected to QA-context Node through an edge. Edge type: Question Edge: describes the key concept in the query statement (such as the association between "transformer" and "fault").

[0055] Connection between scene graph and concept graph: Nodes in the scene graph (such as "transformer") and nodes in the concept graph (such as "power equipment") are indirectly connected through supernodes. Information is transmitted across modalities through supernodes.

[0056] Therefore, the construction formula of the multimodal graph is: Where E represents a set of edges; each scene graph node vs and concept graph node vc are connected by a supernode z.

[0057] (2) Two-way information fusion.

[0058] VQA-GNN uses two modality-specific graph neural networks (GNNs) to perform message passing on the scene graph and concept graph respectively.

[0059] Node feature update formula: The feature of each node v is updated by aggregating the features of neighboring nodes: ; Where, N(v): the set of neighbor nodes of node v; AGG: aggregation function (such as summation or attention mechanism); W: learnable weight matrix; σ: activation function (such as ReLU).

[0060] The features of each node (e.g., “transformer”, “cable”) of the scene graph not only contain their visual information, but also incorporate semantic information from the concept graph.

[0061] The features of each node (e.g., “operating status,” “abnormality”) of the concept graph are also updated to integrate the contextual visual and linguistic information of the scene graph.

[0062] (3) Cross-modal messaging.

[0063] The information of scene graph and concept graph is passed through super nodes: ; ;in, and : Attention weight, used to measure the importance between super nodes and neighbor nodes. Super nodes connect scene graphs and concept graphs. Their features represent the global semantic fusion results of visual modality and language modality, and are an important summary of the entire multimodal semantic graph.

[0064] Supernode feature update formula: ; : Information aggregated from the scene graph; : information aggregated from the concept graph; ||: feature concatenation operation; fz: linear transformation or nonlinear activation function.

[0065] (4) Extract the global representation of the scene graph and concept graph through pooling operation, concatenate the feature vectors of QA-context Node and QA-concept Node, and output the fused feature , used for subsequent analysis or as a query of the associated vector to provide key field completion for fuzzy search.

[0066] Example description: Image: Contains on-site photos of a transformer equipment, with the equipment number marked in the picture (such as "Equipment Number: T12345").

[0067] Query statement: "What is the most recent operating temperature of the transformer in the figure?"

[0068] Concept graph generation: Extract concept nodes related to "operating temperature" from the knowledge graph, and retrieve the database field "temperature"; the concept graph nodes include "operating temperature" and "time point".

[0069] Super Node (QA-context Node): connects the scene graph node "transformer" and the concept graph node "operating temperature", and associates the contextual semantics in the query (such as "most recently").

[0070] Message passing: The visual information in the scene graph is passed to the concept graph through super nodes. The semantic information in the concept graph is passed back to the scene graph through super nodes to achieve cross-modal information fusion.

[0071] Through this message passing mechanism, GNN can effectively fuse the features of text, image, and time series data, so that deep reasoning can be performed based on multimodal data. For example, when a user queries a question related to a certain power equipment, not only can the specific type of the equipment be parsed from the text data, but also the image features and historical operation data of the equipment can be combined to generate a more comprehensive understanding. For example, when a user queries "abnormal operation of the transformer in the past week", GNN will fuse the equipment appearance information from the image, the equipment description from the text, and the historical operation status from the time series to determine whether there is an abnormality.

[0072] Through this deep fusion of multimodal data, stronger associations can be established between different data sources, thereby improving the accuracy of responses to queries. Especially in the power industry, data of different modalities usually contain rich background information. Through the integration and reasoning of this information by GNN, users can be provided with more accurate and comprehensive answers. This multimodal fusion mechanism enables queries to not only rely on a single data modality, but to make full use of all available data information in order to efficiently and accurately reason about complex problems.

[0073] Multimodal GNN technology is introduced to perform cross-modal information transfer between scene graphs and concept graphs to reduce the representation gap between different modalities and enhance the ability to understand context.

[0074] Through this two-way fusion and cross-modal messaging, it is not only possible to obtain information from a single text or image, but also to combine vision and language through the graph neural network model to perform more complex conceptual hierarchical reasoning. In this process, the power industry-specific knowledge in the knowledge graph (such as equipment type, fault information, etc.) is used as background information to further enhance the understanding of natural language, and can combine user queries with proprietary knowledge in the power field. For example, a user may enter "high-load transformer anomalies in the past week", which will be associated with the keywords in the query through the knowledge graph, and the specific information related to the query will be inferred based on the scene and the data in the knowledge graph.

[0075] Step 2: Intelligent error correction and query optimization mainly rely on Hypergraph Neural Networks (HGNN) to automatically identify and correct semantic errors or logical deviations in user queries.

[0076] In user queries, if there are errors in the input device name or parameters, the hypergraph neural network can detect these semantic errors or logical deviations by comparing them with known devices and parameters in the knowledge graph of the power industry. For example, when a user enters "temperature control valve status of a high-voltage transformer", it may be detected that the device name does not match the parameter configuration, because "high-voltage transformer" does not have the device component "temperature control valve". The hypergraph neural network's error correction mechanism will be used to prompt users with possible correct combinations of device names and parameters, such as "cooling system status of the transformer".

[0077] In addition, intelligent optimization is performed based on historical query records. When a user's query is too vague or logically ambiguous, similar historical queries will be referenced and combined with background knowledge of the power industry to generate optimized query options. For example, when a user enters "device abnormalities in the past week", if the query scope is too large or lacks specificity, the user will be advised to narrow the query scope, such as "temperature abnormalities of high-voltage transformers in the past week" to provide more precise results.

[0078] Through multiple rounds of interaction, queries can be further optimized. When it is detected that the query entered by the user is ambiguous, a series of candidate query options will be generated, and the user will be guided to select the query path that best meets his or her intention. For example, if the user is querying the fault information of multiple devices, the system will automatically generate multiple specific queries, such as "transformer fault information" and "switch device fault information", etc., to help users gradually optimize the query content in multiple rounds of interaction, ensuring that the final query results are more accurate.

[0079] Specifically, based on the fusion features (supernodes and related node features updated by GNN), missing field information, semantic errors, logical deviations, fuzziness or logical ambiguity (such as time range) in the query are identified, and candidate query statements are generated by aligning with the concept graph node features. For example, "find abnormal records of high-voltage equipment" can be supplemented as "abnormal records of high-voltage equipment in the past year."

[0080] Example description: Field completion for fuzzy query, image: surveillance image of a construction site. Query statement: "Find the operation log of high-voltage equipment at this site." System processing: The scene graph extracts the relevant equipment in the image and matches it to the "high-voltage equipment" field. The missing "operation log" time range in the query statement is completed by the context (such as the default of the last 30 days).

[0081] In step 3, query enhancement and multi-round interaction optimization, the focus is on generating multiple candidate queries by detecting ambiguous or conflicting information in user input, and continuously optimizing the query content through multiple rounds of interaction, ultimately ensuring that the query results meet user needs.

[0082] Specifically, when a user enters a fuzzy or ambiguous query, the query is first parsed using a natural language processing model to identify polysemous words or possible conflicts in the query. For example, when a user enters "check the status of the device", this query may be too general, the device type is unclear, or the query target is unclear. At this time, multiple candidate query paths are generated based on the ambiguity of the input. These paths may be differentiated according to different equipment types, time periods, or parameters, such as "check the status of the transformer", "check the status of the switchgear", or "check the status of the device in the past 24 hours".

[0083] In this process, a multimodal graph structure is used to capture the relationships between different entities in the query. Candidate queries are generated by automatically identifying multiple interpretation paths or contexts related to the user's initial input, thereby providing users with multiple possible choices. In each round of interaction, these possible queries are presented to the user, allowing the user to select the query path that best meets their intentions.

[0084] Multiple rounds of interactive optimization are achieved through the following feedback mechanisms: ;in, is the query in round t, is the user’s feedback on the query in round t, and α is the learning rate, which indicates the influence weight of the user’s choice. Each time the user selects a candidate query path, it returns to step 1, re-evaluates the user’s true intention based on the selection result, and further optimizes the candidate query.

[0085] This multi-round interaction method gradually guides users to clarify their needs by reducing ambiguity and vagueness in the query. Through multiple interactions, it helps users gradually converge from vague input to clear and precise query requests. For example, a user may initially query "device status". After several rounds of interaction, the user can be guided to refine the query to "view the fault records of a specific device in a certain time period."

[0086] In step 4, deep reasoning and enhancement of hypergraph structure, the hypergraph structure is used to model the potential multi-layer relationships in the complex query context to achieve refined association discovery and data insight enhancement. Different from error correction or simple query optimization, this step aims to provide users with in-depth data patterns and trends beyond the query surface results, thereby helping users gain a richer analytical perspective in the context of complex data.

[0087] The information nodes such as power equipment, operating status, historical records, environmental factors, etc. involved in the query are combined into a hypergraph structure, so that these nodes present a multi-layer association pattern in the query. Hypergraphs can support high-order connections between a node and multiple other nodes to capture implicit but critical association patterns in the data. This structure allows the system to perform deep reasoning through non-explicit conditions (such as historical co-occurrence, association conditions, etc.). For example, when a user queries "the cause of the recent transformer failure", not only the direct status information of the transformer is returned, but also the fault records and external factors under similar conditions in the past are associated through the hypergraph structure, thereby helping users identify possible trends or hidden causes.

[0088] In the multi-layer hypergraph reasoning process, feature aggregation and transfer are used to model the implicit relationships between different nodes. Each node in the hypergraph represents a specific device state or historical event, and multiple hyperedges correspond to the co-occurrence patterns of these events under specific conditions. Through multiple iterative analyses, it is inferred which relationships between nodes meet the aggregation or heterogeneity of specific conditions. For example, by analyzing the "common time period of multiple device failures under the same load conditions", potential associated risks can be predicted, prompting users to pay attention to specific operating conditions.

[0089] Hypergraph reasoning also includes cross-time series correlation analysis, which can provide multi-dimensional data insights in query results by identifying the commonalities of states under different time periods and conditions. For example, when a user queries the temperature anomaly of a certain device, Hypergraph reasoning can simultaneously analyze data on temperature changes, device load, maintenance records, and other dimensions over the past few months to infer the potential causes of temperature fluctuations. This analysis is not limited to the current query target, but can also predict problems that may arise under similar conditions in the future, providing users with preventive suggestions.

[0090] Through deep reasoning of this hypergraph structure, queries are extended to the hidden levels of the data, providing users with deep insights based on data association and pattern recognition, helping users discover potential associations in complex power data.

[0091] Specifically, based on the supernodes and related node features updated by GNN, the extended query statement is obtained through hypergraph reasoning.

[0092] Multimodal query expansion, input image: monitoring screen, showing damaged transmission lines. Query statement: "Are there any recent maintenance records for transmission lines similar to the one in the image?"; System processing: The scene graph extracts the damage features in the image. The concept graph expands the "similar transmission lines" condition in the query. Return query results and suggestions: "The recent maintenance record shows that transmission line ID 12345 completed maintenance in October 2024.

[0093] Step 5: In the natural language to SQL query conversion, it is necessary to convert the user's natural language query into a structured SQL query based on understanding the user's intention. This process uses an autoregressive model SQLformer, which can gradually generate SQL (structured query statement) statements based on natural language input. In order to accurately complete this conversion process, it not only relies on the semantic embedding in the knowledge graph, but also combines the structured information of the power equipment.

[0094] First, when receiving a user's natural language query, such as "abnormal conditions of high-load transformers in the past week", the natural language processing model will decompose the key components of the query and identify the core objects in the query (such as "transformer"), time range (such as "past week"), and query intent (such as "abnormal conditions"). In order to convert these unstructured natural language elements into structured SQL queries, the entity embedding in the knowledge graph is used to locate the specific entities and attributes of these elements in the power system. For example, "transformer" is mapped to the corresponding equipment table in the database, and "abnormal conditions" corresponds to the abnormal field in the fault record table.

[0095] In the process of generating SQL queries, the autoregressive model will gradually predict each component of the query, from SELECT, FROM to WHERE conditions. The SQLformer model is used to represent the SQL query as an abstract syntax tree (AST) and use the autoregressive model to generate each node of the SQL query. In this process, the corresponding SQL structure is generated according to the database table structure, field information and query intent.

[0096] Formulaically, the autoregressive model predicts the next SQL token at each step, and the formula is expressed as:

[0097] Among them, yt is the SQL token generated at time step t, X represents the input natural language query, ht is the hidden state of the model at time step t, and Wt is the weight matrix used to map the hidden state to the output token. The model generates the next SQL token based on the previously generated token sequence y1, y2, ..., yt-1 and the input query X through autoregression.

[0098] For example, for the query "Abnormal conditions of high-load transformers in the past week", first identify that the query needs to select data from the equipment table (SELECT), and then combine the time condition (WHERE time>= NOW() - INTERVAL '1WEEK') and the abnormal status field (WHERE status = 'Abnormal'). In the generated SQL statement, the model will gradually build each part of the query:

[0099] This method of step-by-step generation through the abstract syntax tree can ensure that the generated SQL statement structure is correct and meets the needs of multi-condition and cross-table queries. Since data in the power industry is usually distributed in multiple related tables (such as equipment tables, fault record tables, operating status tables, etc.), it is necessary to gradually generate cross-table query statements through the autoregressive mechanism of SQLformer and automatically process the associations therein.

[0100] In addition, when generating SQL queries, the relationship between tables and fields in the knowledge graph is referenced. For example, it is known in which tables the relevant information of "transformer" is stored, and the terms in the natural language query are matched with the fields in the database through semantic embedding. In this way, complex SQL queries can be automatically generated, and the query statements are guaranteed to be consistent with the database structure specific to the power industry. Through this autoregressive SQL generation model, the user's natural language query can be efficiently and accurately converted into SQL queries, ensuring that users can quickly obtain the required power data.

[0101] Step 6, dynamic adaptation of permissions, involves dynamically adjusting data access permissions based on user identity, permissions, and data sensitivity through the attribute-based access control (ABAC) model. Before executing any query, the user's identity is first verified, and their permissions to access certain data are evaluated based on their attributes (such as position, department, job role, etc.). The ABAC model plays a key role here. It not only checks the user's identity, but also comprehensively considers various attributes related to the data, such as data sensitivity, data classification level, etc., to ensure that the query results match the user's permissions.

[0102] By associating each user with specific attributes and matching these attributes with data access policies, it is determined whether the user has access to certain sensitive data. Specifically, the ABAC model can be described by the following formula:

[0103] Among them, P (access) is the result of the access right determination, subject_attributes represents the attributes of the user, object_attributes represents the attributes of the data object, and environment_conditions represents the environmental conditions for access. Function f combines these attributes and conditions to determine whether the user is allowed to access a data set. For example, when a user tries to query the operating status of a specific transformer, the user's position and permissions will be checked. If the user is a system administrator or a senior engineer, he may have the permission to view detailed operating data, while ordinary users can only view general information.

[0104] During operation, a set of user attributes and data attributes will be generated for each query request, which will be input into the ABAC policy engine, which will match them according to the predefined access control rules and generate corresponding access decisions. For example, high-privileged users may be allowed to access certain sensitive data, such as the real-time operating status and historical fault records of the equipment, while low-privileged users can only access relatively public non-sensitive information, such as basic information and profile data of the equipment.

[0105] If the object attribute that the user attempts to access indicates that it is sensitive data (for example, it contains private information or confidential device data), and the user does not have sufficient permissions, a response of access denial will be returned, or the sensitive data will be shielded and a fuzzy result will be returned. In this case, a privacy-preserving fuzzy query mechanism may be used to add noise to the results or shield some information, ensuring that even in the case of insufficient permissions, the user can still obtain some useful information without leaking sensitive data.

[0106] For example, when a user with low permissions queries the detailed operation log of a transformer, the log fields involving sensitive information may be shielded, or only the operation overview of the transformer may be returned, while the precise values ​​and sensitive time period information are hidden. Assuming that the user requests a detailed query, but his permissions do not allow viewing all data, the following fuzzy SQL query may be generated:

[0107] This SQL statement will filter the accessible data according to the user's permission level. For example, a high-privilege user may be able to obtain the precise numerical value of the running status, while a low-privilege user can only view general information such as "normal" or "abnormal".

[0108] Step 7, Sensitive data protection and fuzzy result presentation, mainly involves automatic protection of sensitive data. Differential privacy technology or fuzzy result presentation is used according to the user permission level and the sensitivity of the data to prevent unauthorized access and ensure data security and compliance. By combining user permissions, data attributes and sensitivity levels, the query result presentation method is dynamically adjusted to provide a flexible and secure access control mechanism.

[0109] When processing queries involving sensitive data, we first determine whether the data in the query is highly sensitive. If the data is marked as highly sensitive, differential privacy is applied to the results. The basic idea of ​​differential privacy is to protect user privacy and data confidentiality by adding noise to the query results to ensure that individual data entries do not have a significant impact on the overall results. The formula for differential privacy is as follows:

[0110] Where A(D) is the query algorithm running on dataset D, D' is an adjacent dataset that differs from D by only one data entry, is the privacy budget. Through this formula, appropriate noise can be added to each query result to limit the impact of a single data record on the result, ensuring that even sensitive data will not be leaked through queries.

[0111] The application of differential privacy is particularly suitable for scenarios that require accurate statistical information, such as when a user queries the usage and fault statistics of a specific device. In this scenario, noise can be added to the results so that even if sensitive data of a specific time period or a specific device is queried, too many detailed information will not be leaked. The introduction of the differential privacy protection mechanism ensures that sensitive individual data will not be inferred or identified while users query statistical information.

[0112] For users with insufficient permissions or when more sensitive data items are queried, a fuzzy display method will be used. For example, if a user attempts to query the detailed operating status or history of a key device, and his permissions are insufficient to access the complete data, the results will be automatically fuzzified so that the user can only see generalized or partially visible results. Fuzzy results can be achieved by hiding specific fields, partially masking data values, or displaying ranged data. For example, when a user queries the operating status of an electric device, the exact value may be replaced with a range (such as "50-100") instead of displaying the specific value. This process can be formalized as follows:

[0113] This fuzzy SQL query will dynamically determine the level of detail displayed based on the user's permission level. If the user's permission is high enough, the specific running status will be returned; if the permission is low, the fuzzy data will be returned, such as "normal" or "abnormal" instead of the exact value.

[0114] In addition, when processing extremely sensitive data (such as confidential power equipment or operating data of high-security systems), you can also request additional permission authorization. At this time, a permission escalation request will be sent to the user, requiring the user to obtain a higher level of access rights, or the query request will be directly rejected. Through this mechanism, it is ensured that even when facing highly sensitive data queries, data access and display can be strictly controlled to prevent the leakage of confidential information.

[0115] Step 8: Compliance and security check is crucial in the query execution process. It is necessary to ensure that the query behavior complies with the regulations and safety standards of the power industry. This step ensures the legality and security of all data access behaviors through automated compliance checks, permission management, and audit mechanisms, thereby avoiding illegal operations or unauthorized data access.

[0116] During the query execution process, the legality of the query is first automatically evaluated based on the regulations of the power industry and the company's internal security standards. This process includes permission verification, data sensitivity checks, and the matching degree of the query content with regulations. For example, if the query involves sensitive power equipment or confidential operating data, the data security requirements in the relevant regulations will be referenced to determine whether the query meets the security compliance standards.

[0117] The query is automatically reviewed through a rule-based compliance engine. The role of the compliance engine is to match the query with pre-defined regulations and company policies to ensure that all query requests comply with regulations. This process can be represented by a logical decision function:

[0118] Among them, C(Q) is the compliance judgment result of query Q, regulation indicates applicable laws and regulations, query_attributes is the relevant attributes in the query (such as the object and field of the query), and user_permissions indicates the user's permission information. Through this function, it is possible to determine whether the query meets the compliance requirements based on the query attributes, user permissions and relevant regulations. If it does not meet the requirements, the query will be blocked from being executed, or the data will be shielded or obfuscated during the execution process.

[0119] In addition to permission and compliance verification, logging is also performed during query execution to ensure that all query behaviors can be traced and audited. Logging is an important part of compliance checks, especially in the power industry, where audit requirements for query behaviors are extremely high. Detailed information about each query is recorded, including the query content, time, executor, and query results. These log data are stored in a secure database and audited regularly to ensure that there are no non-compliant query behaviors. The logging function can not only play a role in post-analysis, but also play a warning and prevention role in real-time monitoring.

[0120] In terms of data masking and permission management, certain sensitive data will be masked according to compliance requirements to prevent low-privilege users or queries that do not meet security standards from obtaining this data. For example, if the query involves the operating status of high-risk power equipment, and the user's permissions are insufficient to view this data, this information will be automatically masked to prevent unauthorized access. This masking can be expressed by the following formula:

[0121] Here, R(Q) represents the result of query Q, and mask(D) represents the masking operation on data D. When the query does not meet the compliance requirements, the system will partially or completely mask the returned data to ensure the security and privacy of the data.

[0122] Through logging, data masking, and permission management, sensitive data can also be prevented from being leaked or improperly used. In addition, possible security threats will be detected in real time at every step of query execution, including possible SQL injection attacks or unauthorized privilege escalation. Once an anomaly is detected, query execution will be terminated immediately and a security alert will be generated to remind the administrator to handle it.

[0123] The power semantic big model adaptive data query and dynamic response method provided in this embodiment has significant advantages, first of all, it improves the accuracy and efficiency of the query. Through adaptive context perception, deep fusion of multimodal data and intelligent error correction mechanism, it can quickly parse and understand the user's natural language query, especially for the professional terminology and complex scenarios of the power industry, and can automatically generate accurate SQL queries. In this way, users can quickly obtain the required information without manual screening and repeated error correction, which greatly improves work efficiency.

[0124] Another important benefit is its deep correlation analysis and intelligent recommendation capabilities. Using deep reasoning of hypergraph structures, complex potential correlation patterns between data can be automatically identified, providing users with deep data insights beyond the query content. This automated correlation discovery helps users identify the causes of equipment anomalies or potential failures and provides preventive maintenance recommendations, thereby effectively improving the reliability and safety of the power system and providing intelligent support for the long-term operation and operational management of equipment.

[0125] The intelligent error correction and optimized query path function further enhances user friendliness. It can automatically detect and correct semantic and logical deviations in queries, and provide users with query options that meet their intentions through multiple rounds of interaction; for users with fuzzy queries or unfamiliar with professional terms, this function helps them find the information they need more quickly, significantly improving the user experience and making the system suitable not only for professional users, but also for a wider range of users.

[0126] In addition, it also has excellent performance in data security and permission control. With the help of dynamic permission adaptation and sensitive data protection mechanism, the display method of results can be automatically adjusted according to the user's permission level and the sensitivity of the queried data. For example, for users with different permissions, data masking, fuzzy display or differential privacy protection can be automatically performed to ensure the safe use of data on the basis of compliance. This not only helps the power industry to achieve compliance management in data use, but also avoids the risk of sensitive data leakage, thereby further improving data security.

[0127] Finally, it has strong adaptability and can flexibly support a variety of application scenarios in the power industry, including equipment management, fault monitoring, data analysis, compliance checking, etc. This multi-scenario adaptability makes this method a key technology in the intelligent and digital transformation of the power industry, providing important support for optimizing production, management and services. Therefore, as an innovative system for power data, this method has brought substantial improvements in query accuracy, data security, correlation analysis and user experience.

[0128] Embodiment 2 The present embodiment provides an adaptive data query and dynamic response system for a large power semantic model, which specifically includes an adaptive data query module, which is configured to: obtain natural language queries and related images of the power industry, extract features through a large language model and a deep learning model respectively, and then use a knowledge graph and a visual question-answering graph neural network to unify the features of different modalities into the same representation in a bidirectional fusion manner to obtain fused features; based on the fused features, perform intelligent error correction through a hypergraph neural network and multiple rounds of interaction, optimize the natural language query, obtain an extended query statement through hypergraph reasoning, and convert the extended query statement into a structured query statement through an autoregressive model to obtain a query result; The dynamic response module is configured to: dynamically adapt permissions, protect sensitive data, and perform compliance and security checks on query results to achieve dynamic response.

[0129] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0130] Embodiment 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method for adaptive data query and dynamic response of a large power semantic model as described in the first embodiment are implemented.

[0131] Embodiment 4 This embodiment provides a computer device, such as Figure 2 As shown, it includes a display device, an input device, a computer-readable storage medium (volatile memory and non-volatile storage medium), a processor, a communication interface (i.e., a network interface), and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor, the communication interface, and the computer-readable storage medium can be connected via a bus or other means. The communication interface is used to receive and send data, and when the processor executes the program, the steps in the method for adaptive data query and dynamic response of the electric power semantic large model as described in the first embodiment above are implemented.

[0132] Among them, any reference to memory, storage, database or other media provided in the present application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0133] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0134] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. The method for adaptive data query and dynamic response of power semantic big model is characterized by: include: Obtain natural language queries and related images of the power industry, extract features through a large language model and a deep learning model, and then use the knowledge graph and visual question-answering graph neural network to unify the features of different modalities into the same representation in a bidirectional fusion manner to obtain fused features; Based on the fusion features, intelligent error correction is performed through a hypergraph neural network and multiple rounds of interaction. After optimizing the natural language query, an extended query statement is obtained through hypergraph reasoning, and the extended query statement is converted into a structured query statement through an autoregressive model to obtain a query result; For query results, dynamic permission adaptation, sensitive data protection, and compliance and security checks are performed to achieve dynamic response.

2. The method for adaptive data query and dynamic response of a large power semantic model according to claim 1, characterized in that: The step of generating the fusion feature comprises: Based on the knowledge graph and query statement features, concepts are extracted from natural language queries to build a concept graph. Based on image features, a scene graph generator is used to extract visual information from images and build a scene graph. The nodes in the scene graph are connected to the nodes in the concept graph through super nodes. Through the visual question-answering graph neural network, the features of the nodes in the scene graph and the nodes in the concept graph are transferred through the super node, and after iteratively updating the features of the super node, the nodes in the scene graph and the nodes in the concept graph, the fused features are spliced.

3. The method for adaptive data query and dynamic response of electric power semantic big model according to claim 1, characterized in that: The hypergraph neural network is used to automatically identify and correct semantic errors or logical deviations in the natural language query to obtain candidate queries.

4. The method for adaptive data query and dynamic response of a large power semantic model according to claim 3, characterized in that: In each round of interaction, after receiving the candidate query selected by the user, the candidate query selected by the user is weightedly fused with the natural language query of the previous round to obtain the natural language query of this round, and the fusion features and candidate query are regenerated until the candidate query no longer changes.

5. The method for adaptive data query and dynamic response of electric power semantic big model according to claim 1, characterized in that: The dynamic adaptation of permissions includes: verifying the identity of the user, evaluating the access permissions according to the user's attributes, and performing fuzzy processing on the query results according to the access permissions.

6. The method for adaptive data query and dynamic response of electric power semantic big model according to claim 1, characterized in that: The sensitive data protection includes: determining whether the data in the query result is sensitive data, and if sensitive data exists, performing differential privacy processing on the query result.

7. The method for adaptive data query and dynamic response of electric power semantic big model according to claim 1, characterized in that: The compliance and security check includes: automatically reviewing the query results through a rule-based compliance engine, and if the review fails, blocking the response of the query results, or shielding or obfuscating the query results during the response process.

8. The power semantic big model adaptive data query and dynamic response system is characterized by: include: The adaptive data query module is configured to: obtain natural language queries and related images of the power industry, extract features through a large language model and a deep learning model respectively, and then use a knowledge graph and a visual question-answering graph neural network to unify the features of different modalities into the same representation in a bidirectional fusion manner to obtain fused features; Based on the fusion features, intelligent error correction is performed through a hypergraph neural network and multiple rounds of interaction. After optimizing the natural language query, an extended query statement is obtained through hypergraph reasoning, and the extended query statement is converted into a structured query statement through an autoregressive model to obtain a query result; The dynamic response module is configured to: dynamically adapt permissions, protect sensitive data, and perform compliance and security checks on query results to achieve dynamic response.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the power semantic big model adaptive data query and dynamic response method as described in any one of claims 1-7 are implemented.

10. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored in the computer-readable storage medium and executable on the processor, characterized in that: When the processor executes the program, the steps in the power semantic big model adaptive data query and dynamic response method as described in any one of claims 1-7 are implemented.

Images