Method and system for automatic extraction of power equipment data quality rules and knowledge construction

Abstract

The application discloses a kind of electric power equipment data quality rule automatic extraction and knowledge construction method and system, belong to artificial intelligence and electric power data management technical field.For solving the problem of low precision in complex table analysis, lack of physical constraint verification in extraction result and model catastrophic forgetting caused by standard update in electric power technical standard document, the application uses multi-modal large language model to analyze document, obtains structured document object model through hierarchical table structure coding, cross-page context maintenance and table annotation semantic binding;Entity relationship extraction is carried out by large language model fine-tuned by domain knowledge, and consistency verification is carried out in combination with electric power domain physical constraint knowledge base;The entity relationship after verification is assembled as structured quality rule record and stored in graph database;Adaptive elastic weight reinforcement strategy is used for continuous learning, and man-machine collaborative quality inspection optimization model is used, high-precision complex table analysis, physical consistency verification and anti-forgetting continuous learning are realized.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and power data governance, specifically relating to a method and system for automatically extracting data quality rules and constructing knowledge for power equipment. Background Technology

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

[0003] Technical standards documents refer to various technical documents used to regulate the entire lifecycle of power equipment (planning and design, material procurement, engineering construction, operation and maintenance, and decommissioning), including industry standards, enterprise standards, and material procurement technical specifications. These documents are the core basis for power equipment data quality control and compliance checks. However, existing technologies face a series of problems when utilizing this standard knowledge.

[0004] First, technical standard documents are mostly in unstructured or semi-structured form, containing a large number of complex tables. These tables have unique complexities such as multi-level nested headers, irregularly merged cells, continuous tables spanning multiple pages, and semantic binding between table comments and cells. Traditional OCR-based and rule-based template methods have extremely low accuracy in parsing these tables. In recent years, although some studies have used general multimodal large models to process document tables, they lack explicit modeling of the header hierarchy, resulting in a high rate of confusion in field membership relationships under multi-level nested headers; their page-crossing processing relies on externally generated pagination templates, which have limited generalization ability; and they have failed to solve the problem of semantic binding between table comments and cells, resulting in the loss of key information such as applicable conditions.

[0005] Secondly, most existing automated extraction methods are based on keyword matching or general natural language processing techniques, lacking a deep understanding of the inherent physical relationships and engineering experience among the technical parameters of power equipment. For example, there are strict physical equation constraints between the no-load loss and excitation current, and between short-circuit loss and short-circuit impedance of a transformer, and the parameters have reasonable numerical ranges under different capacity and voltage levels. Existing solutions can only perform simple numerical range searches and judgments, and cannot use the physical equations between parameters for cross-validation. This leads to contradictory information being output as correct results, greatly increasing the workload of manual verification.

[0006] Furthermore, quality rules for the same type of equipment are scattered across dozens of independent standards, with complex relationships such as referencing, substitution, and complementarity among these standards. Currently, there is a lack of effective structured integration and correlation mining methods. Simultaneously, the standard system is constantly being updated and iterated, and existing models generally suffer from catastrophic forgetting during incremental learning: after fine-tuning on a new standard, the extraction performance on the old standard significantly decreases, while new and old standards often coexist in practical use for extended periods, requiring the model to maintain high accuracy on the old standard and rapid adaptation to the new standard simultaneously. Existing methods completely fail to consider the problem of continuous learning and cannot achieve stable model updates. Summary of the Invention

[0007] To address the problems of low parsing accuracy of power equipment technical standard documents, logical inconsistencies caused by the lack of physical constraint verification in the extracted results, and catastrophic forgetting of models due to standard updates in existing technologies, this invention provides an automatic extraction and knowledge construction method for power equipment data quality rules. It parses documents using a multimodal large language model to obtain a structured document object model; extracts entity relationships using a large language model fine-tuned with power domain knowledge, and performs verification based on a physical constraint knowledge base to obtain a set of consistency-verified triples; assembles the triples into structured quality rule records and stores them in a graph database; and iteratively optimizes the model through human-machine collaborative quality inspection and feedback data. This invention achieves high-precision parsing of complex tables, automatic physical consistency verification, and continuous learning to resist forgetting.

[0008] On the one hand, a method for automatic extraction of data quality rules and knowledge construction of power equipment is provided, including: The document obtains technical standard documents for power equipment, parses the documents using a pre-trained multimodal large language model, identifies the document structure and extracts content features to obtain a structured document object model. Based on the document object model, named entity recognition and relation extraction are performed using a large language model fine-tuned with knowledge of the power industry. After extraction, constraint verification is performed based on a pre-built physical constraint knowledge base of the power industry to obtain a set of entity and relation triples that have passed physical consistency verification. Based on a predefined rule classification system and data model, the entity and relation triple set is assembled into a structured quality rule record, and the quality rule record is stored in a graph database. The quality rule records are subjected to human-machine collaborative quality checks, and the multimodal large language model and the large language model fine-tuned with power domain knowledge are iteratively optimized using human feedback data.

[0009] Furthermore, the document parsing includes: For the table areas identified in the document, row and column coordinate detection and visual connectivity analysis are performed to identify the spatial range of merged cells. A directed acyclic graph is constructed to represent the hierarchical membership of the table headers, with each header cell as a node and parent-child edges representing inclusion or membership relationships. This directed acyclic graph is encoded into a position-aware hierarchical embedding vector, which is then fused with the cell text embedding and input into the multimodal large language model. When a table spans multiple pages, the features of the previous page's table area are cached as a short-term memory vector. When parsing subsequent pages, this memory vector is queried through a cross-page attention mechanism to automatically associate the continuity of the cross-page table. By spatially distance-weighted association between the annotation text block below the table and the table area, a triplet attention structure is constructed for annotations, tables, and cells. This enables the multimodal large language model to actively query the bound annotation set when parsing cell content and embed the conditional information in the annotations into the cell representation.

[0010] Furthermore, the physical constraint knowledge base in the power field includes: a mathematical relationship library between parameters, a reasonable range library of parameters, and a consistency rule library between parameters. The mathematical relationship library between parameters stores the physical equation expressions between power equipment parameters. The execution constraint verification includes: substituting the extracted parameters into the physical equation expression, calculating the relative deviation between the theoretical value and the actual value, and if the deviation exceeds the preset tolerance, it is determined that there is an extraction error, and the parameter most likely to be wrong is located through partial derivative sensitivity analysis. The contradictory information discovered during verification is transformed into a weak supervision signal and fed back to the large language model, which has been fine-tuned with knowledge from the power sector, to adjust the confidence level of subsequent extractions or trigger the re-extraction of specific regions.

[0011] Furthermore, the parameter mathematical relationship library includes at least the following: basic electrical engineering relationships between transformer no-load loss and excitation current, short-circuit loss and short-circuit impedance, and capacity and voltage and current; the parameter reasonable range library establishes typical value ranges for each parameter for each equipment type, capacity level, and voltage level combination; and the parameter consistency rule library stores empirical ratio constraints. The method of locating the most likely error-causing parameter through partial derivative sensitivity analysis specifically includes: calculating the sensitivity of each parameter to the deviation, marking the parameter with the highest sensitivity as the most likely error source, and triggering a re-extraction of the region where the parameter is located.

[0012] Furthermore, the predefined rule classification system includes: performance parameter class, quality requirement class, test method class, and identification specification class; the structured quality rule record adopts a unified JSON Schema format, and the fields include rule number, parameter name, parameter code, value, unit, applicable conditions, applicable equipment range, standard source, and extraction confidence level; The knowledge graph of power equipment data quality in the graph database establishes entity nodes for equipment, parameters, standards, terms, values, and conditions, and constructs multi-dimensional semantic associations through definition, reference, constraint, substitution, and computation of dependency edge types.

[0013] Furthermore, the manual feedback data includes: rule records marked as correct after experts confirm the quality rule records on the quality inspection platform, corrected rule records obtained after correcting the content that was extracted incorrectly, and rejection marks generated when rule records that cannot be confirmed or do not conform to the standard are rejected.

[0014] Furthermore, the iterative optimization includes: Learnable version identity embedding vectors, including version timestamps and standard states, are added to the inputs of the multimodal large language model and the large language model fine-tuned with knowledge of the power domain to the standard document. Automatically compare the chapter structure, parameter names, numerical ranges, and reference relationships between the old and new versions of the standard, and output a difference score; A regularization term is introduced into the loss function of incremental learning, and the loss function is designed as follows: ; in, For the loss on new data, For the new model parameters, i represents the index of the model parameter. For the old model parameters, The diagonal elements of the Fisher information matrix represent the importance of the parameters to the old task. Score the differences between the standard and standard versions. The regularization strength; Maintain a cache of old standard core samples. During incremental learning, mix the old standard samples in the cache with the new standard samples for training and replay the experience.

[0015] Furthermore, the multimodal large language model is a pre-trained model based on the visual language Transformer architecture.

[0016] On the other hand, an automatic extraction and knowledge construction system for power equipment data quality rules is also provided, including: The document parsing and feature extraction module is configured to: acquire power equipment technical standard documents, parse the documents using a pre-trained multimodal large language model, identify the document structure and extract content features to obtain a structured document object model; The entity relationship extraction and verification module is configured to: perform named entity recognition and relationship extraction based on the document object model using a large language model fine-tuned with knowledge of the power field, and perform constraint verification based on a pre-built physical constraint knowledge base of the power field after extraction, to obtain a set of entity and relationship triples that have passed physical consistency verification. The rule structuring and storage module is configured to: assemble the entity and relation triple set into a structured quality rule record according to a predefined rule classification system and data model, and store the quality rule record in a graph database; The human-machine collaborative optimization module is configured to: perform human-machine collaborative quality checks on the quality rule records, and iteratively optimize the multimodal large language model and the large language model fine-tuned with power domain knowledge using human feedback data.

[0017] Furthermore, a computer-readable storage medium is provided on which a computer program is stored, which, when executed by a processor, performs the method of the first aspect.

[0018] The above technical solution has the following advantages or beneficial effects: (1) High-precision complex table parsing capability. This invention solves the unique challenges of multi-level nested headers, irregularly merged cells, cross-page tables, and annotation associations in power standard documents by using hierarchical table structure encoding, cross-page context preservation, and semantic binding of table annotations. Experiments show that for documents containing the above-mentioned complex tables, it is significantly improved compared to the solution using a general multimodal large model.

[0019] (2) Strong physical consistency. This invention introduces a physical constraint knowledge base and verification algorithm in the power field, and uses physical equations between parameters to perform cross-verification and deviation inversion positioning, which greatly reduces the workload of manual review. The extracted results can be directly applied to intelligent quality inspection and compliance review.

[0020] (3) Anti-forgetting and continuous learning capabilities. Through standard version identity embedding, elastic weight consolidation with adaptive difference, and experience replay strategy, the catastrophic forgetting caused by standard updates is effectively suppressed. After 5 consecutive standard version iterations, balanced and high-precision support for both the old and new standards is achieved, ensuring the stability of the system in long-term operation and maintenance.

[0021] (4) High degree of automation and efficiency. It achieves high-precision end-to-end automatic extraction from unstructured documents to structured knowledge, and the processing efficiency is several orders of magnitude higher than that of manual processing.

[0022] (5) Knowledge structuring and computability. Produces machine-readable, understandable, and computable structured quality rules and interconnected knowledge graphs, providing real-time, accurate, and unified standard data services for business scenarios such as intelligent procurement, automated quality inspection, online compliance review, and standard difference analysis. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0024] Figure 1 This is a flowchart of the overall method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of a multi-level nested header in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the hierarchical table structure encoding of Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of cross-page context preservation according to Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the semantic binding of table annotations in Embodiment 1 of the present invention; Figure 6 This is a flowchart of the physical constraint verification process according to Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the structured rule assembly and knowledge graph construction of Embodiment 1 of the present invention; Figure 8 This is a comparison chart of the accuracy of complex table extraction in Embodiment 1 of the present invention; Figure 9 This is a comparison chart showing the decrease in the F1 value of the old standard after the standard update in Embodiment 1 of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Those skilled in the art should understand that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0026] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] Example 1 This embodiment provides a method for automatically extracting data quality rules and constructing knowledge for power equipment, such as... Figure 1 As shown, the method includes the following steps: S1: Obtain technical standard documents for power equipment, integrate a pre-trained multimodal large language model with a dedicated table parsing module to parse the documents, identify the document structure and extract content features to obtain a structured document object model; S2: Based on the document object model, a large language model fine-tuned with knowledge of the power industry is used to perform named entity recognition and relation extraction. After extraction, constraint verification is performed based on a pre-built physical constraint knowledge base of the power industry to obtain a set of entity and relation triples that have passed physical consistency verification. S3: Based on the predefined rule classification system and data model, assemble the entity and relation triple set into a structured quality rule record, and store the quality rule record in the graph database; S4: Perform human-machine collaborative quality checks on the quality rule records, and use human feedback data to iteratively optimize the multimodal large language model and the large language model fine-tuned with knowledge from the power field.

[0028] In step S1: For different categories of power equipment, obtain a set of technical standard documents for the target power equipment (such as transformers), including PDF files of standards such as GB, DL, and Q / GDW. Input these documents into a pre-trained multimodal large language model and integrate the dedicated table parsing module designed in this invention to parse the documents. For example... Figure 2 As shown, Figure 2 This is a diagram illustrating a multi-level nested header. In the header area of ​​this table, the "Loss" cell spans multiple columns, covering the two sub-columns "Idle Loss" and "Load Loss".

[0029] The model performs the following operations.

[0030] S1.1, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the hierarchical table structure encoding. The multimodal large language model simultaneously processes the pixel information of the scanned image of the document and the embedded text information. Through a visual-language cross-attention mechanism, it identifies the logical structural blocks of the document, with a focus on identifying table regions. For the identified table of "Technical Parameters of 10kV Oil-Immersed Distribution Transformer," the table region is first subjected to row and column coordinate detection and visual connectivity analysis to identify the spatial range of merged cells. Specifically, firstly, the row and column separators of the table are located through horizontal / vertical projection or edge detection to determine the initial grid coordinates of each cell; then, the binarized image is labeled with connected components to identify the range of multiple grid cells covered by the merged cells, thereby obtaining complete cell row and column positions and merging span information.

[0031] For example, the "Loss" cell in the header area spans multiple columns, covering the sub-columns "Idle Loss" and "Load Loss". Then, a Directed Acyclic Graph (DAG) is constructed to represent the hierarchical membership of the header: the root node is "Loss", and its child nodes are "Idle Loss" and "Load Loss"; "Idle Loss" can also have conditional branches such as "Oil-Immersed" and "Dry". Finally, this hierarchical table structure is encoded as a position-aware hierarchical embedding vector, which is then fused with the text embedding of each cell and input into a multimodal large language model. This allows the model to accurately understand that "Idle Loss" is a subordinate parameter of "Loss" rather than an independent parameter. This design solves the problem of incorrect field attribution in multi-level nested headers in general multimodal models.

[0032] S1.2, as Figure 4 As shown, Figure 4 This invention illustrates the preservation of cross-page context. A cross-page attention masking mechanism is designed. When a cross-page table is detected, the features of the previous page's table region are cached as a short-term memory vector. When parsing subsequent pages, the model queries this memory vector through cross-page attention, automatically associating the continuity of the cross-page table without relying on external RAG retrieval or pagination templates. Simultaneously, table fingerprint hashing is introduced to automatically identify the boundaries of cross-page tables through fuzzy matching of row count, column count, and header text.

[0033] For example, when a table extends from page 10 to page 12, the model detects a "continued" marker at the end of the table on page 10 and automatically identifies the boundary of the cross-page table. The features of the table area on page 10 (header structure, column width, unit information) are cached as a short-term memory vector. When parsing page 11, the model queries this memory vector through a cross-page attention mechanism, ensuring that values ​​on subsequent pages are correctly associated with the header definition on the first page, avoiding the misinterpretation of "1.72" as an independent, unitless value.

[0034] S1.3, such as Figure 5 As shown, Figure 5 This diagram illustrates the semantic binding of table annotations. The annotation text blocks below the table are spatially distance-weighted and associated with the table area (based on Y-coordinate similarity and horizontal alignment), constructing a "annotation-table-cell" triplet attention. For example, the table contains the annotation "Note: No-load loss is the measured value under rated frequency and rated voltage." This annotation is bound to the table. When the model parses the cell "1.72" under the "No-load Loss" column, it actively queries the bound annotation set and embeds the conditional information "rated frequency and rated voltage" from the annotation into the cell's representation. This achieves semantic binding between the numerical values ​​in the table and their applicable conditions, with the numerical values ​​carrying associated conditional information, thus solving the problem of lost applicable conditions.

[0035] As a further implementation method, the multimodal large language model in this step is a pre-trained model based on the vision-language Transformer architecture, and is supervised fine-tuned through massive power equipment technical standard documents and their labeled data. The dedicated table parsing module is used as a pluggable module for parallel or fusion training with the model.

[0036] Through the above steps, a structured Document Object Model (DOM) is formed. In this model, each element (including text blocks and table cells) carries semantic type, content, location information, and associated annotation conditions. For example, cell "1.72" carries the applicable environmental conditions of "rated frequency and rated voltage", thus ensuring that key annotation information is not lost in subsequent extraction steps.

[0037] Step S2: Domain-Knowledge-Enhanced Entity Relation Extraction and Physical Constraint-Based Consistency Verification. Based on the DOM parsing result from Step S1, a large language model fine-tuned with power domain knowledge is used for named entity recognition and relation extraction. After extraction, power domain physical constraint verification is performed to ensure the engineering rationality of the extraction results. For example... Figure 6 As shown, the specific steps include the following.

[0038] S2.1 Joint Extraction. Based on a pre-built dictionary of power equipment, parameter ontology, and lightweight knowledge graph, a joint learning task of sequence labeling (entity recognition) and relation classification is designed during the fine-tuning stage of the large language model. This task is used to identify entities such as equipment type, technical parameters, parameter values, units, test conditions, and performance requirements, and to extract the correspondence, constraints, and reference relationships between entities. During the inference stage, text mentions are aligned with knowledge graph nodes through entity links, and the accuracy of extraction is improved by utilizing graph relationships.

[0039] Specifically, entity types include: PARAM (parameter name), VALUE (value), UNIT (unit), CONDITION (condition), DEVICE (device type), REF_STD (reference standard), etc. For example, from the DOM, "no-load loss" is identified as PARAM, "1.72" as VALUE, "kW" as UNIT, and the "rated frequency, rated voltage" bound to the annotation as CONDITION, and relationships such as hasValue, hasUnit, and underCondition are established.

[0040] Meanwhile, during joint extraction, external power industry knowledge graphs can be queried to assist in reasoning. For example, when the model identifies "no-load loss," it can locate the standard parameter node in the knowledge base through entity links to obtain its expected unit (which should be kW or W) and typical value range (for a 10kV / 3150kVA oil-immersed transformer, the typical no-load loss value is 0.8-1.2kW), which can be used for subsequent verification.

[0041] S2.2, For the target equipment type (such as a transformer), three types of physical constraint knowledge bases are pre-built, including: Parameter mathematical relation library: Stores the physical equations between parameters in the form of computable expressions, such as transformer no-load loss. P 0 and excitation current I 0 、 Magnetizing resistor R m Approximate relation Short-circuit loss P k With short-circuit impedance U k Current I relational formula and capacity S ,Voltage U Current I The basic electrical engineering relationship between them .

[0042] Parameter reasonable range library: Based on historical qualified product data and standard regulations, typical value ranges for each parameter are established for each combination of "equipment type, capacity level and voltage level". For example, for oil-immersed distribution transformer (capacity 3150kVA, voltage level 10kV), the reasonable ranges for no-load loss are set as [0.8, 1.2]kW, short-circuit loss as [10, 15]kW, and short-circuit impedance as [4.5, 6.0]%.

[0043] A base of rules for consistency among parameters: stores empirical ratio constraints, such as the ratio of short-circuit loss to short-circuit impedance. It should be within a specific range (0.9~1.1).

[0044] S2.3, a verification algorithm based on physical constraints, performs three types of verification on the parameter set extracted from the model.

[0045] Single parameter rationality verification: The parameter value is compared with the reasonable range library. If it exceeds the threshold, it is marked as "suspected anomaly" and the anomaly type is recorded. For example, the currently extracted "no-load loss = 1.72kW" exceeds the upper limit of the reasonable range by 1.2kW, triggering the "suspected anomaly" mark, and the anomaly type is recorded as "value too high".

[0046] Parameter consistency verification: Substitute the extracted parameters into pre-stored physical relationship equations to calculate the relative deviation between theoretical and actual values. If the deviation exceeds a preset tolerance (e.g., 10%), an extraction error is determined, and the most likely erroneous parameter is located using equation inversion (through partial derivative sensitivity analysis). For example, if "rated capacity = 3150kVA" and "short-circuit impedance = 5.5%" are extracted simultaneously, the theoretical short-circuit loss is calculated by substituting it into the physical relationship and compared with the extracted value. If the deviation exceeds 10%, an extraction error is determined, and the most likely erroneous parameter is located using sensitivity analysis.

[0047] Cross-standard consistency verification: If the same parameter appears in multiple related standards, compare their differences. If the difference exceeds an engineering experience threshold, prompt the user to confirm the standard version and scope of application. For example, if the same parameter is specified in both GB / T 6451-2023 and DL / T 1264-2022, compare the differences between the two. If the difference exceeds an engineering threshold (e.g., 20%), prompt the user to confirm the standard version and scope of application.

[0048] This physical constraint verification surpasses existing technologies that rely solely on RAG retrieval of simple value ranges, enabling cross-validation and error localization by leveraging the deep physical relationships between parameters.

[0049] S2.4, Verification Result Feedback. Logical inconsistencies identified during the physical constraint verification process include, but are not limited to, single-parameter rationality contradictions (parameter values ​​exceeding the threshold of the pre-built reasonable range library), inter-parameter consistency contradictions (the relative deviation between theoretical and actual values ​​after substitution into the physical equation exceeds the preset tolerance), and cross-standard consistency contradictions (the difference in the specified values ​​of the same parameter in different standards exceeds the engineering experience threshold). These contradictions indicate that the current extraction results may contain errors. The contradictions discovered during verification need to be converted into weak supervision signals and fed back to the large language model in S2.1, which has been fine-tuned using power sector knowledge, to adjust the confidence level of subsequent extractions or trigger a re-extraction of specific areas.

[0050] For example, there is a contradiction in the rationality of a single parameter: the standard for a certain 10kV / 3150kVA oil-immersed distribution transformer specifies that the reasonable range of no-load loss is 0.8 to 1.2 kW, but the extracted result is "1.72 kW", which exceeds the upper limit and constitutes a contradiction.

[0051] Inconsistent parameters: The extraction results simultaneously record "rated capacity S=3150kVA" and "rated voltage U=10kV", but the calculated current I (I = S / ( The deviation of U) ≈ 182A from the extracted "rated current I=210A" exceeds 10%, indicating that at least one parameter was extracted incorrectly.

[0052] Inconsistent values ​​across standards: The same transformer model has a no-load loss of 1.0 kW specified in GB / T 6451-2023, 1.3 kW in DL / T 1264-2022, and a model-extracted value of 1.2 kW. Although this does not exceed the range of a single standard, the difference between the two standards exceeds the threshold of engineering experience (e.g., 0.2 kW), prompting users to verify the applicable standard version.

[0053] The aforementioned physical constraint knowledge base supports dynamic expansion. Newly discovered parameter relationships or empirical ranges can be incrementally added after expert review and fed back to the verification module through the knowledge graph.

[0054] Through the above steps, the final output is a set of entity and relation triples that have been physically verified.

[0055] In step S3, after completing the physical consistency verification in step S2 and obtaining a reliable set of entity and relation triples, this step assembles these triples into structured quality rule records and dynamically updates them to the power equipment data quality knowledge graph. For example... Figure 7 As shown, the specific steps include the following.

[0056] S3.1: Rule Structuring: Rule records are defined using a unified JSON Schema according to a predefined rule classification system (including performance parameter, quality requirement, test method, and identification specification categories). Each record's fields include: rule number, parameter name, parameter code, value, unit, applicable conditions, applicable device range, standard source (standard number, clause, version), and extraction confidence level. For example... Figure 4 As shown. For example, a performance parameter rule regarding the no-load loss of a transformer can be structured as follows: { "rule_id": "TRANSFORMER_P001", "parameter_name": "Idle load loss", "parameter_code": "GB_T_1094.1_2023_NO_LOAD_LOSS", "value": "1.72", "unit": "kW", "applicable_conditions": "at rated frequency and rated voltage", "applied_to": { "device_type": "oil-immersed distribution transformer", "voltage_level": "10kV", "capacity": "3150kVA" }, "source": { "standard_id": "GB / T 6451-2023", "clause": "version": "2023" }, "extraction_confidence": 0.96, "physical_check_status": "passed_with_warning", "warning_info": "The value is slightly higher than the upper limit of the typical range; it has been marked for manual review." }

[0057] S3.2: Knowledge Graph Update: Transform the above structured rule records into RDF triples and store them in a graph database (such as Neo4j). Establish entity nodes such as devices, parameters, standards, clauses, values, and conditions, and construct multi-dimensional semantic relationships through edge types such as definition, reference, constraint, substitution, and computational dependency; at the same time, establish substitution, reference, and difference relationships between standards to provide a foundation for subsequent cross-standard consistency verification and continuous learning.

[0058] As a further implementation, this method also supports external business systems to query quality rules and verify the compliance of equipment parameters through a RESTful API, and can transmit standard version information to achieve version-based knowledge retrieval.

[0059] In step S4, a closed-loop optimization mechanism of "initial algorithm screening, expert review, and continuous learning" is established. To address the catastrophic forgetting problem caused by frequent standard updates, an elastic weight consolidation (EWC) strategy is adopted for incremental learning. This includes the following sub-steps.

[0060] S4.1, Standard Version Identity Embedding. This step connects to publicly available channels such as the National Standards Information Public Service Platform, the National Standards Full-Text Disclosure System, the official website of the China Electricity Council, or the official website of the National Energy Administration. It automatically extracts information such as the standard number, version year, and standard status (current / repealed / transitional) from document metadata, filenames, or standard number fields. At the input of the multimodal large language model in S1 and the large language model in S2 (fine-tuned with power sector knowledge), a learnable version identity embedding vector is added to the standard document, including the version timestamp and standard status (current / repealed / transitional). This embedding vector serves as a trainable parameter matrix. As the model incrementally learns on both old and new version standard data, it provides version identification criteria for the subsequent step S4.3, enabling the model to distinguish the feature distributions of different version standards.

[0061] S4.2, Two-stage quality inspection. For the structured quality rule records assembled in step S3 (i.e., rules automatically extracted and structured from technical standard documents), first, rule-based automatic logical verification is performed, including unit consistency checks, numerical range reasonableness judgments, and standard reference validity verification. This is then combined with the physical constraint verification results from step S2.3 for a comprehensive evaluation. Results with confidence levels below a preset threshold or with verification conflicts are prioritized and pushed to the quality inspection platform for confirmation, correction, or rejection by domain experts. Experts can view the rule records on the platform in highlighted form and can perform confirmation, correction, or rejection operations.

[0062] S4.3, Elastic Weight Consolidation Adaptively Based on Differences. An improved EWC strategy is employed when incrementally learning the multimodal large language model in S1 and the large language model in S2, which has been fine-tuned using power domain knowledge, based on high-quality data that has been manually reviewed, verified, or corrected.

[0063] First, the standard difference analysis module automatically compares the chapter structure, parameter names, numerical ranges, and citation relationships between the old and new versions of the standard, outputting a difference score d, with a value range of [0,1]. Specifically, the difference score d is calculated by weighting three parts: chapter structure difference (the proportion of newly added, deleted, or adjusted chapters), parameter name difference (the proportion of terminology changes), and numerical range difference (the magnitude of parameter value changes). Finally, the weighted sum is used to obtain the comprehensive difference score d.

[0064] Then, the regularization strength of EWC is adaptively adjusted according to d. This is to balance the relationship between learning new tasks and memorizing old tasks. A first threshold and a second threshold are preset, with the first threshold being less than the second threshold. When d is less than the first threshold, the threshold is increased. Strengthen the protection of knowledge from old standards; when d is greater than the second threshold, reduce... This allows the model to be updated moderately on the key parameters of the new standard. When d is between the two, the baseline regularization strength is maintained.

[0065] A regularization term is introduced into the loss function of incremental learning, and the loss function is designed as follows: ; in, Losses due to new data; For the new model parameters, i Index representing model parameters; These are the parameters for the old model; The diagonal elements of the Fisher information matrix represent the importance of the parameters to the old task. The regularization strength is adaptively adjusted based on the difference score d. When the differences between the old and new standards are significant (e.g., substantial changes in parameter names or numerical ranges), Decrease the value to allow for larger updates to model parameters; when the difference is small... This method increases and strengthens the protection of knowledge from older standards. It addresses the catastrophic forgetting problem that is completely ignored by existing technologies.

[0066] S4.4, Experience Replay. A cache of old standard core samples is maintained to store the most representative labeled pages from each old standard (selected based on page information entropy or sample diversity). During incremental learning, old standard samples from the cache are mixed with new standard samples for training, further enhancing resistance to forgetting.

[0067] S4.5, periodically triggered updates. Supports triggering incremental learning processes for the model based on quantitative amounts (e.g., accumulating 1000 review samples) or timed intervals (e.g., monthly), enabling continuous adaptation to new standards and business feedback.

[0068] The invention will be further illustrated below with reference to a specific implementation example.

[0069] Taking "no-load loss" extracted from "GB / T 6451-2023 Technical Parameters and Requirements for Oil-Immersed Power Transformers" as an example.

[0070] First, upload the PDF document of the standard.

[0071] In step S1, during the parsing phase: the model correctly identifies multi-level table headers through the hierarchical table encoder; the cross-page context preservation module caches the first page header information; and the table-annotation binding module associates the annotations below with the table.

[0072] Step S2 Extraction and Verification Stage: The model identifies parameters, values, units, and conditions; physical verification finds that 1.72kW is slightly higher than the typical upper limit, triggering an alarm. After cross-standard verification, it is marked as "awaiting manual confirmation".

[0073] Step S3: Structuring and Storage Stage: Generate JSON records and store them in the knowledge graph.

[0074] Step S4 Quality Inspection and Learning Phase: After expert confirmation, the range library is updated, and EWC incremental learning is triggered.

[0075] To verify the effectiveness of the method of this invention, 50 national and enterprise standards PDF documents in the field of transformers (totaling approximately 2000 pages) were selected, including 100 multi-level nested header tables, 50 cross-page tables, and 80 tables with annotation binding requirements. The test set consisted of 5000 manually annotated rules accumulated over time.

[0076] The experiment uses three methods: the traditional template method based on regular expressions and fixed templates, the method of combining a general multimodal large model with conventional incremental learning, and the method of this invention.

[0077] Experimental results are as follows Figure 8 and Figure 9 As shown, Figure 8 A comparison chart of extraction accuracy for complex tables. Figure 9 The chart shows a comparison of the F1 score decrease after the standard update. Analysis reveals that the extraction accuracy of this invention exceeds 91% in complex table scenarios (multi-level headers, page spans, and annotation binding), significantly outperforming the general multimodal model (maximum 82.1%) and traditional methods (maximum 32.5%). Furthermore, after five standard updates, the F1 score of this invention only decreased by 2.1%, while conventional incremental learning decreased by 18.3%, demonstrating the effectiveness of the continuous learning strategy of this invention.

[0078] Meanwhile, physical constraint verification reduced the logical consistency error rate from 12.7% to 2.5%, and reduced the workload of expert review by 65%. A transformer quality knowledge graph containing more than 120,000 nodes and 180,000 relationships was successfully built and has been integrated into the actual business system via API, improving efficiency by more than 70%.

[0079] Example 2 This embodiment provides a system for automatically extracting data quality rules and constructing knowledge for power equipment, including: The document parsing and feature extraction module is configured to: acquire power equipment technical standard documents, parse the documents using a pre-trained multimodal large language model, identify the document structure and extract content features to obtain a structured document object model; The entity relationship extraction and verification module is configured to: perform named entity recognition and relationship extraction based on the document object model and a large language model fine-tuned with knowledge of the power field; and perform constraint verification based on a pre-built physical constraint knowledge base of the power field after extraction to obtain a set of entity and relationship triples that have passed physical consistency verification. The rule structuring and storage module is configured to: assemble the entity and relation triple set into a structured quality rule record according to the predefined rule classification system and data model, and store the quality rule record in the graph database; The human-machine collaborative optimization module is configured to: perform human-machine collaborative quality checks on quality rule records, and iteratively optimize the multimodal large language model and the large language model fine-tuned with knowledge from the power field using human feedback data.

[0080] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0081] Example 3 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method of Embodiment 1.

[0082] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatic extraction and knowledge construction of data quality rules for power equipment, characterized in that, include: The document obtains technical standard documents for power equipment, parses the documents using a pre-trained multimodal large language model, identifies the document structure and extracts content features to obtain a structured document object model. The document parsing process includes: For the table areas identified in the document, row and column coordinate detection and visual connectivity analysis are performed to identify the spatial range of merged cells. A directed acyclic graph is constructed to represent the hierarchical membership of the table headers, with each header cell as a node and parent-child edges representing inclusion or membership relationships. This directed acyclic graph is encoded into a position-aware hierarchical embedding vector, which is then fused with the cell text embedding and input into the multimodal large language model. When a table spans multiple pages, the features of the previous page's table area are cached as a short-term memory vector. When parsing subsequent pages, this memory vector is queried through a cross-page attention mechanism to automatically associate the continuity of the cross-page table. The annotation text block below the table is spatially distance-weighted and associated with the table area to construct a triplet attention of annotation, table and cell. This enables the multimodal large language model to actively query the bound annotation set when parsing cell content and embed the conditional information in the annotation into the cell representation. Based on the document object model, named entity recognition and relation extraction are performed using a large language model fine-tuned with knowledge of the power industry. After extraction, constraint verification is performed based on a pre-built physical constraint knowledge base of the power industry to obtain a set of entity and relation triples that have passed physical consistency verification. Based on a predefined rule classification system and data model, the entity and relation triple set is assembled into a structured quality rule record, and the quality rule record is stored in a graph database. The quality rule records are subjected to human-machine collaborative quality checks, and the multimodal large language model and the large language model fine-tuned with power domain knowledge are iteratively optimized using human feedback data.

2. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 1, characterized in that, The physical constraint knowledge base in the power field includes: a mathematical relationship database between parameters, a reasonable range database of parameters, and a consistency rule database between parameters. The mathematical relationship database between parameters stores the physical equation expressions between power equipment parameters. The execution constraint verification includes: substituting the extracted parameters into the physical equation expression, calculating the relative deviation between the theoretical value and the actual value, and if the deviation exceeds the preset tolerance, it is determined that there is an extraction error, and the parameter most likely to be wrong is located through partial derivative sensitivity analysis. The contradictory information discovered during verification is transformed into a weak supervision signal and fed back to the large language model, which has been fine-tuned with knowledge from the power sector, to adjust the confidence level of subsequent extractions or trigger the re-extraction of specific regions corresponding to that parameter.

3. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 2, characterized in that, The parameter mathematical relationship library includes at least the following: basic electrical engineering relationships between transformer no-load loss and excitation current, short-circuit loss and short-circuit impedance, and capacity and voltage and current; the parameter reasonable range library establishes typical value ranges for each parameter for each equipment type, capacity level, and voltage level combination; and the parameter consistency rule library stores empirical ratio constraints. The method of locating the most likely error-causing parameter through partial derivative sensitivity analysis specifically includes: calculating the sensitivity of each parameter to the deviation, marking the parameter with the highest sensitivity as the most likely error source, and triggering a re-extraction of the region where the parameter is located.

4. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 1, characterized in that, The predefined rule classification system includes: performance parameter class, quality requirement class, test method class, and identification specification class; the structured quality rule record adopts a unified JSON Schema format, and the fields include rule number, parameter name, parameter code, value, unit, applicable conditions, applicable equipment range, standard source, and extraction confidence level; The knowledge graph of power equipment data quality in the graph database establishes entity nodes for equipment, parameters, standards, terms, values, and conditions, and constructs multi-dimensional semantic associations through definition, reference, constraint, substitution, and computation of dependency edge types.

5. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 1, characterized in that, The manual feedback data includes: rule records marked as correct after experts confirm the quality rule records on the quality inspection platform, corrected rule records obtained after correcting the content that was extracted incorrectly, and rejection marks generated when rule records that cannot be confirmed or do not conform to the standard are rejected.

6. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 1, characterized in that, The iterative optimization includes: Learnable version identity embedding vectors, including version timestamps and standard states, are added to the inputs of the multimodal large language model and the large language model fine-tuned with knowledge of the power domain to the standard document. Automatically compare the chapter structure, parameter names, numerical ranges, and reference relationships between the old and new versions of the standard, and output a difference score; A regularization term is introduced into the loss function of incremental learning, and the loss function is designed as follows: ； in, For the loss on new data, For the new model parameters, i represents the index of the model parameter. For old model parameters, The diagonal elements of the Fisher information matrix represent the importance of the parameters to the old task. Score the differences between the standard and standard versions. The regularization strength; Maintain a cache of old standard core samples. During incremental learning, mix the old standard samples in the cache with the new standard samples for training and replay the experience.

7. The method for automatic extraction and knowledge construction of power equipment data quality rules according to claim 1, characterized in that, The multimodal large language model is a pre-trained model based on the visual language Transformer architecture.

8. A system for automatic extraction and knowledge construction of data quality rules for power equipment, characterized in that, include: The document parsing and feature extraction module is configured to: acquire power equipment technical standard documents, parse the documents using a pre-trained multimodal large language model, identify the document structure and extract content features to obtain a structured document object model; The document parsing process includes: For the table areas identified in the document, row and column coordinate detection and visual connectivity analysis are performed to identify the spatial range of merged cells. A directed acyclic graph is constructed to represent the hierarchical membership of the table headers, with each header cell as a node and parent-child edges representing inclusion or membership relationships. This directed acyclic graph is encoded into a position-aware hierarchical embedding vector, which is then fused with the cell text embedding and input into the multimodal large language model. When a table spans multiple pages, the features of the previous page's table area are cached as a short-term memory vector. When parsing subsequent pages, this memory vector is queried through a cross-page attention mechanism to automatically associate the continuity of the cross-page table. The annotation text block below the table is spatially distance-weighted and associated with the table area to construct a triplet attention of annotation, table and cell. This enables the multimodal large language model to actively query the bound annotation set when parsing cell content and embed the conditional information in the annotation into the cell representation. The entity relationship extraction and verification module is configured to: perform named entity recognition and relationship extraction based on the document object model using a large language model fine-tuned with knowledge of the power field, and perform constraint verification based on a pre-built physical constraint knowledge base of the power field after extraction, to obtain a set of entity and relationship triples that have passed physical consistency verification. The rule structuring and storage module is configured to: assemble the entity and relation triple set into a structured quality rule record according to a predefined rule classification system and data model, and store the quality rule record in a graph database; The human-machine collaborative optimization module is configured to: perform human-machine collaborative quality checks on the quality rule records, and iteratively optimize the multimodal large language model and the large language model fine-tuned with power domain knowledge using human feedback data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for automatic extraction and knowledge construction of power equipment data quality rules as described in any one of claims 1-7.

Images