Method for constructing power grid engineering review knowledge base based on modal perception

Abstract

This invention discloses a method for constructing a power grid engineering review knowledge base based on modality perception. The method includes the following steps: acquiring and preprocessing full-modal review data; converting images and numerical modalities into standardized text through heterogeneous mapping; classifying and recognizing the text using rules and machine learning, and adaptively dividing and outputting standard text blocks; extracting modality-specific features and performing differentiated embedding, generating modality fusion vectors through spatial alignment and attention fusion; constructing a three-layer architecture and embedding a regional characteristic adaptation layer; performing alignment and relationship completion based on the fusion vectors, binding regionally specific attributes and priority review rules to related entities, and performing conflict resolution to construct the fusion knowledge base. This invention achieves deep semantic alignment of multimodal review information, adapts to special regional engineering environments, and improves the quality of intelligent compliance verification in power grid reviews.

Description

Technical Field

[0001] This invention belongs to the field of intelligent power grid engineering review technology, and specifically relates to a method for constructing a power grid engineering review knowledge base based on modal perception. Background Technology

[0002] Currently, in the field of intelligent power grid engineering review, mainstream technologies mainly focus on three areas: single text knowledge base, simple multimodal feature splicing, and general power intelligent question answering. An integrated technical system that adapts to all review scenarios, covers all modalities, and takes into account regional characteristics has not yet been formed.

[0003] Some existing solutions focus on constructing single-text knowledge graphs. For example, Chinese invention patent application CN111930784A discloses a method and system for constructing a power grid knowledge graph. This solution trains an NLP model by constructing a dictionary of professional terms and a multi-source corpus, combines Spark to process structured data with a natural language processing model to extract unstructured text knowledge, and introduces a dual verification mechanism to ensure accuracy. However, this solution only focuses on text data processing and relies on natural language extraction tools. It cannot perform deep heterogeneous semantic mapping on the massive amounts of non-textual modal information such as engineering images and monitoring values ​​in power grid reviews, resulting in blind spots in the multimodal information linkage retrieval in review scenarios.

[0004] Based on the aforementioned shortcomings, some technologies attempt to fuse multi-source heterogeneous data. For example, Chinese invention patent application CN114239683A discloses a data fusion processing method, system, and device for power systems. It collects multi-source heterogeneous data from power grid operation, transactions, and marketing, performs two-stage preprocessing (cleaning, denoising, and structuring), extracts four modal features (image, video, speech, and point cloud), uses graph convolution and NLP semantic recognition to determine modal relationships, and then inputs the associated features into a generative adversarial network for modal-level and feature-level fusion, ultimately deciding on the fusion output to improve fusion accuracy. However, this solution generally adopts a direct, general feature extraction and splicing method, without designing a dedicated discriminative embedding strategy for the essential differences between structured, semi-structured, and unstructured text in power grid review data. This results in low semantic accuracy and poor alignment of text features with different degrees of structure and non-text features after fusion.

[0005] In summary, existing power grid engineering review knowledge base technologies are all general-purpose designs and have not been specifically optimized for all review scenarios. They suffer from the following core defects: First, incomplete modal coverage and semantic separation between non-text and textual data; second, the general text vectorization method does not perform differentiated embedding based on the text's structuring level, resulting in the loss of domain-specific rule semantics; and third, the lack of regionally specific adaptation and conflict resolution mechanisms means that regional rules such as typhoon resistance and geological avoidance cannot be triggered preferentially, making it difficult for existing knowledge bases to meet the core needs of professional, refined, and regionalized power grid engineering reviews. Summary of the Invention

[0006] This invention provides a method for constructing a knowledge base for power grid engineering review based on modality awareness. It achieves this by heterogeneous semantic mapping of multimodal data, embedding modality-aware differentiated features for different structured texts, and deeply configuring and dynamically triggering regional characteristic adaptation logic such as typhoon, geological, and ecological red lines in the underlying graph and rule base. Ultimately, it completes cross-modal entity alignment and hierarchical resolution of knowledge conflicts. This aims to solve the problems of semantic fragmentation caused by incomplete multimodal data coverage in existing power grid review systems, loss of professional domain semantics due to general text feature extraction, and insufficient accuracy of compliance verification in actual engineering reviews due to the lack of specific adaptation and conflict resolution mechanisms for specific regional scenarios.

[0007] To address the aforementioned technical problems, this invention proposes a method for constructing a power grid engineering review knowledge base based on modal perception, comprising the following steps: Acquire and preprocess structured, semi-structured, and unstructured data, engineering images, and monitoring values ​​for power grid engineering review; The preprocessed engineering images and monitoring values ​​are mapped into standardized text through heterogeneous semantic mapping. A method combining rule matching and machine learning is used to identify text and output a standard set of text blocks. Modality-specific features of the standard text block set are extracted and differentially embedded. After linear spatial alignment and cross-modal attention fusion, a modality fusion vector with a unified dimension is generated. Build a basic architecture that includes a data layer, a knowledge layer, and an application layer, and embed a regional characteristic adaptation layer to generate a complete architecture system; By binding region-specific attributes and region-specific review rules to related entities in the complete architecture system, and performing knowledge conflict resolution, a fused knowledge base is constructed.

[0008] Preferably, after constructing the fused knowledge base, the method further includes the following steps: A five-dimensional quality inspection index system is constructed based on accuracy, completeness, consistency, relevance, and practicality, and quality verification is performed on the fused knowledge base. For knowledge data that fails verification, the error is traced back to the corresponding stage, and the model parameters and review rules are iteratively optimized. When a trigger instruction is received for a new review project, a revision of a standard, or an update of regional data, incremental collection and processing of multimodal data is performed, and the incremental knowledge is associated and integrated into the fused knowledge base for dynamic incremental updates and verification.

[0009] Preferably, the method for generating the standard text is as follows: The image parsing model is called to extract the core information of the engineering image and generate a structured text description. Based on the power monitoring threshold matching rule library, the monitoring values ​​are converted into structured text to obtain a uniformly formatted mapping text. Based on a pre-built power engineering professional dictionary and a context clustering algorithm, semantic distinctions are made between easily confused terms, and entities in the mapped text are associated with pre-trained knowledge graph nodes to complete the deep injection of domain knowledge. The text after domain knowledge injection is divided into different categories according to function. The domain knowledge and the mapped text are semantically weighted and fused according to preset weights. After removing redundant semantics and completing the logic, the standard text is generated.

[0010] Preferably, the method for generating the standard text block set is as follows: Based on rule matching and machine learning classification models, the modality type of text modal data is identified, and the text modal data is divided into structured text, semi-structured text and unstructured text, and corresponding exclusive modality type tags are embedded. The structured text is divided using an attribute group clustering strategy, the semi-structured text is divided using a tagging and logic dual-driven strategy, and the unstructured text is divided using a clue and event dual-guided strategy, generating initial text blocks. The initial text block is subjected to semantic integrity verification of core information and granularity appropriateness verification of word count range. Based on the semantic similarity of adjacent text blocks, boundary merging and splitting optimization is performed, and the standard text block set bound with modal attribute information and verification results is output.

[0011] Preferably, generating a uniform-dimensional modality fusion vector includes the following steps: For structured, semi-structured, and unstructured text, we extract attribute association and parameter logic features, marker symbols and hierarchical logic features, spatiotemporal clues and event causal features, and use them as modality-specific features to concatenate with text word embedding vectors to generate an initial feature vector; The initial feature vector is input into the embedding model built based on the multi-task training mechanism, and the output is a differentiated dimensionality reduction text block vector, wherein the multi-task training mechanism includes modality classification, entity recognition, relationship prediction and semantic matching auxiliary tasks; Construct transformation matrices specific to each text modality, multiply the text block vectors by the corresponding transformation matrices respectively, and map them to a unified semantic space to achieve linear space alignment; The engineering image feature vector and the monitoring data feature vector are introduced through a cross-modal attention mechanism, and then weighted and summed with the aligned text block vector according to a preset fusion weight to generate the unified-dimensional modal fusion vector.

[0012] Preferably, the method for constructing the complete architecture is as follows: A hybrid storage architecture is constructed in the data layer, using a distributed file system to store the original images and unstructured documents, a relational database to store the standard text block set and annotation data, and a vector database to store the modality fusion vectors. In the knowledge layer, a knowledge graph containing devices, parameters, specifications, defects, and regional entities and their relationships is constructed through a graph database, and it is linked with a pre-built condition rule base and case base by foreign keys; The application layer deploys a retrieval interface that supports entity and vector retrieval, a reasoning interface that supports causal judgment, and a multimodal linkage interface that supports the synchronous display of text, images, and data. By embedding the regional characteristic adaptation layer, the relevant equipment and engineering entities are extended with regional-specific attributes including typhoon intensity, geological disaster level and ecological red line area, corresponding environmental impact and avoidance correlation are added, and regional-specific rules with higher priority than general review rules are added to complete the binding and adaptation of regional characteristics.

[0013] Preferably, the process of binding region-specific attributes and region-specific review rules to related entities in the complete architecture system, and performing knowledge conflict resolution to construct a fused knowledge base includes the following steps: Cross-modal entity alignment is performed based on similarity calculation and consistency rule verification of entity fusion vectors, and multi-source semantic relationship completion is completed by using a graph attention network model combined with preset rules. During multimodal knowledge fusion, the knowledge conflict resolution is performed based on a preset confidence weighting mechanism and a normative priority determination mechanism. The system binds typhoon, geological, and ecological red line attributes to geographically related entities and increases the weight of geographically related knowledge during retrieval and reasoning. When compliance verification is triggered, the system controls the geographically reviewed rules to prioritize covering general rules.

[0014] A second aspect of the present invention also proposes a modality-aware power grid engineering review knowledge base construction system, the system being used to implement the knowledge base construction method as described in the first aspect of the present invention, comprising: The data acquisition and preprocessing module is used to acquire and preprocess structured data, semi-structured data, unstructured data, engineering images, and monitoring values ​​for power grid engineering review. The heterogeneous semantic mapping module is used to convert preprocessed engineering images and monitoring values ​​into standardized text through heterogeneous semantic mapping. The modality segmentation module is used to identify text using a fusion of rule matching and machine learning methods, and output a standard set of text blocks. The differential embedding and fusion module is used to extract modality-specific features from the standard text block set and perform differential embedding. After linear spatial alignment and cross-modal attention fusion, a modality fusion vector with a unified dimension is generated. The architecture building module is used to build a basic architecture that includes a data layer, a knowledge layer, and an application layer, and embeds a regional characteristic adaptation layer to generate a complete architecture system. The knowledge resolution and construction module is used to bind region-specific attributes and region review rules to related entities in the complete architecture system, and to perform knowledge conflict resolution to build a fused knowledge base.

[0015] A third aspect of the present invention also provides an electronic device comprising: One or more processors; Memory, used to store one or more computer programs; One or more computer programs stored in the memory are executed by the one or more processors, causing the one or more processors to implement the knowledge base construction method as described in the first aspect of the invention.

[0016] In a fourth aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the knowledge base construction method as described in the first aspect of the present invention.

[0017] Compared with the prior art, the present invention has the following technical effects: 1. The knowledge base construction method proposed in this invention achieves deep integration and unified governance of all modal information (structured / semi-structured / unstructured text, engineering images, and monitoring values) for power grid engineering review by constructing a full-link automated processing architecture. It also deeply integrates engineering constraints of specific physical environments into the underlying architecture system, outputting a fusion knowledge base with extremely high business fit, which significantly improves the accuracy of multimodal information retrieval, the precision of intelligent reasoning, and the quality of compliance verification in power grid engineering review scenarios.

[0018] 2. The knowledge base construction method proposed in this invention converts non-textual modalities into standardized text by calling a parsing model and rule matching, and introduces a domain dictionary and knowledge graph for deep knowledge injection. Simultaneously, it implements adaptive segmentation strategies for texts with different structures. This breaks down the underlying structural barriers of cross-modal data, achieves standardized normalization of multi-source review information, and preserves the professional semantics and core logical relationships of the power review field to the greatest extent possible.

[0019] 3. The knowledge base construction method proposed in this invention extracts specific features for different text types and outputs differentiated dimensionality reduction vectors under a multi-task training mechanism. Finally, it maps these vectors to a unified semantic space through a specific transformation matrix and a cross-modal attention mechanism. This accurately captures the essential attribute differences of texts with different levels of structure, achieves seamless linear alignment of high-dimensional features, and significantly enhances the association depth and feature fusion quality of cross-modal data.

[0020] 4. The knowledge base construction method proposed in this invention embeds a regionally specific adaptation layer on top of the basic architecture of data, knowledge, and applications. This dynamically binds entities with regionally specific attributes and rules, including those related to typhoons, geology, and ecological red lines, and configures a rigorous confidence level and priority conflict resolution funnel. This enables the knowledge base to possess a high degree of awareness of specific natural geographical environments, ensuring that the most appropriate regional review red line standards can be accurately triggered and executed during compliance verification.

[0021] 5. The knowledge base construction method proposed in this invention establishes a multi-dimensional quantitative detection index system and parameter iteration backtracking mechanism, and supports dynamic incremental fusion triggered by multiple sources. This endows the knowledge base with self-correction capabilities and lifecycle evolution capabilities, ensuring the accuracy, consistency, and high timeliness of the review knowledge base in long-term, high-frequency applications. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the knowledge base construction method described in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0024] Example 1 This embodiment describes a method for constructing a knowledge base for power grid engineering review based on modal perception. Figure 1 As shown, it includes the following steps one through six: Step 1: Obtain and preprocess structured, semi-structured, and unstructured data, engineering images, and monitoring values ​​for power grid engineering review.

[0025] To ensure that the multimodal data required for power grid engineering review is comprehensive, accurate, and formatted in a consistent manner, this step can be implemented through three sub-stages: data source configuration, multimodal data acquisition, and standardized preprocessing.

[0026] In the data sourcing phase, before data collection, the system first connects to a multi-source, heterogeneous underlying system. This includes connecting to a distributed relational database (e.g., PostgreSQL) storing the full data of the reviewed projects, a RESTful interface-based connection to an equipment operation and maintenance management platform, and an interface to a government platform containing natural resources, ecological environment, and meteorological data. To avoid index conflicts during subsequent data fusion, all access ports are configured with mandatory constraints on associated parameters. This requires that any piece of raw data collected subsequently must carry four core parameters as a unique composite index: a unique project code, a modality type code, a precise timestamp, and a regional identifier code.

[0027] In the multimodal data acquisition sub-stage, differentiated acquisition strategies are executed based on the differences in data modality characteristics.

[0028] For structured data, equipment parameter tables, test result tables, and equipment ledger data are synchronized from the project ledger and equipment operation and maintenance system periodically (e.g., daily at 00:00). To ensure consistency, the data format is strictly limited to Excel, CSV, or relational database table format, with missing fields filled with NULL values ​​and encoded in UTF-8 format. Specific data fields include fixed items such as equipment model, rated voltage, rated current, insulation resistance, dielectric loss value, test time, pass threshold, and test conclusion. During implementation, the maximum number of data entries collected at one time can be limited to 1000 to avoid network congestion.

[0029] For semi-structured data, data containing engineering drawing annotations, power engineering technical specifications, and review comments are retrieved in batches from archives and industry standard libraries. The transmission protocol uses an HTTPS interface and carries a unique project identity token. A maximum limit is set for each batch of exported data, and duplicate annotations with excessive spatial overlap are automatically filtered after export. As an example, engineering drawings are in PDF / A-1 format, power engineering technical specifications are in PDF or Office document format, and review comments are in XML 1.0 format. Specific data fields are fixed and include annotation number, equipment name, technical requirements, specification number, clause level, problem description, rectification requirements, and responsible unit. In this embodiment, when collecting semi-structured data via HTTPS, the request header is strictly limited to include a unique project identity token, the response data format is forced to UTF-8 encoding, the maximum number of data records exported in a single batch is strictly set to 800, and duplicate annotations with a spatial overlap area ≥ 1 square meter are automatically filtered after export.

[0030] For unstructured data, construction logs, site inspection reports, and defect records are collected from on-site construction terminals and inspection platforms using FTP or other file transfer and file sharing protocols. To avoid network congestion, the size of individual files is strictly limited. As an example, typical construction logs are in TXT format, site inspection reports in Office document format, and defect records in PDF format; specific data fields are fixed and include items such as record time, construction location, construction procedure, inspection time, equipment status, defect location, defect type, and rectification measures. For the collected unstructured data, the size of a single file is limited to 50MB, and the text encoding of the collected data is uniformly converted to UTF-8 format.

[0031] For engineering image data, construction drawings, on-site defect photos, and inspection images in image formats (such as JPG and PNG) are directly acquired from on-site high-definition shooting terminals and drone equipment. The image data must retain its original lossless attributes and original resolution, and metadata including latitude and longitude coordinates and shooting time must be extracted. As an example, the engineering image data collected in this step is limited to: image resolution no less than 1920×1080, single image size strictly controlled to no more than 20MB, and image metadata fixed to include four items: shooting time, shooting location, equipment number, and image annotation information. The data source is limited to on-site high-definition shooting equipment and multi-rotor drone inspection terminals, and the image format must retain its original lossless compression attributes; secondary compression is prohibited.

[0032] For monitoring numerical data, real-time streaming data in JSON format is obtained from the substation online monitoring system, and terminal data is obtained from testing instruments. The real-time streaming data acquisition frequency is set to high frequency, and TCP / IP encrypted communication is used during transmission. To ensure high availability, a retry mechanism with fixed time intervals is introduced. If the maximum number of retries is exceeded, an exception log is generated for future reference. As an example, the acquisition frequency is fixed at 15 minutes / time, using TCP / IP 4.0 encrypted communication protocol for transmission, and the data encoding is forced to use AES-256 encryption format. The interface response timeout threshold is strictly set at 60 seconds. After a single acquisition failure, a retry operation is automatically initiated at a fixed time interval of 3 seconds, with a maximum number of retries limited to 3. After a retry failure, an exception acquisition log containing timestamp, device ID, error code, and acquisition scenario is automatically generated and retained in the database for no less than 30 days.

[0033] In the standardization preprocessing sub-stage, data governance is performed on the collected multimodal raw data to finally output a well-organized and effective dataset. The specific preprocessing process includes deduplication and anomaly isolation, data interpolation and completion, semantic parsing and normalization, format unification and multidimensional annotation.

[0034] Deduplication and Anomaly Isolation: A hash algorithm is used to calculate the hash value of the data. This hash value, combined with the project code and timestamp, is used to perform deduplication, retaining only the latest valid data. For numerical data, the 3σ (Laida) principle is used to calculate the mean and standard deviation to identify outliers. Outlier data is isolated and stored. If the outlier ratio exceeds a certain limit, a re-collection process for that data source is automatically triggered. If the project's unique code and precise timestamp are used as the primary key for deduplication, the duplication criteria are strictly set to require identical hash values, matching primary keys, and a time difference of ≤3 seconds. After deduplication, only the valid data corresponding to the latest timestamp is retained, and all duplicate copies are automatically deleted. When the proportion of outlier data exceeds 5%, a re-collection process for the corresponding data source is automatically triggered.

[0035] Data interpolation and completion: Data is classified and filled according to its characteristics. For missing time-series monitoring values, a linear interpolation algorithm is used, with the interpolation step size matching the aforementioned real-time streaming data acquisition frequency (e.g., 15 minutes). For missing spatial data (e.g., latitude and longitude or geological coordinates), a common Kriging interpolation algorithm is used, with the interpolation semi-variogram function fixed at a spherical model. Data with an interpolation error greater than a set value (e.g., 8%) is marked as data to be reviewed. For missing key text fields, a pre-configured power industry dictionary is used for matching and completion. If the missing rate is less than a set value (e.g., 5%), automatic filling is performed; if it is greater than the set value, the corresponding data source is re-acquired.

[0036] Semantic parsing and normalization: A natural language processing model (such as the BERT-base model) fine-tuned with corpus data from the power industry review domain is invoked to perform preliminary semantic parsing of unstructured text under specific hyperparameters (such as learning rate 4e-5, batch size=12, hidden layer dimension 768, confidence threshold ≥0.7, and 25 iterations). Simultaneously, numerical data is normalized using the Z-Score algorithm, mapping the values ​​to a fixed standard interval (such as [-3,3]). Text data that fails parsing is automatically returned to the human expert annotation process.

[0037] Standardized formatting and multidimensional annotation: All time-series data were standardized by converting them to a standard time format (e.g., YYYY-MM-DD HH:MM:SS), spatial data was normalized to a unified coordinate system (e.g., WGS-84), and all text was converted to UTF-8 encoding. Finally, four-dimensional labels were added to the cleaned data: project code, modality type, region identifier, and project type, forming a standardized dataset with normalized format and no invalid data, providing a high-quality input source for subsequent heterogeneous semantic mapping.

[0038] In addition, this embodiment also performs quality verification and invalid data filtering operations on the collected data: Integrity verification: Verify the missing rate of core fields. The missing rate of core fields must be ≤4%. Data that does not meet the standard should be returned to the corresponding collection stage for re-collection. Accuracy verification: Perform cross-source data comparison to verify the data matching degree. The data matching degree must be ≥97%. Data that does not meet the standard is marked as data to be reviewed and manually checked by power engineering review engineers. Timeliness verification: Verify data synchronization latency, strictly requiring data synchronization latency ≤20 minutes, and automatically trigger incremental data collection process for data that exceeds the time limit; Invalid data filtering: Remove invalid data with abnormal format, excessive values, encoding errors, or data collection failure marks. Valid data after filtering is stored in the database. The size of a single structured data storage is strictly controlled to not exceed 1.8KB. The database index uses a four-dimensional composite index (project code + modality type + region code + timestamp) to avoid index conflicts.

[0039] Step 2: Convert the preprocessed engineering images and monitoring values ​​into standardized text through heterogeneous semantic mapping.

[0040] In this embodiment, the method for generating standard text includes the following steps S21 to S23: S21: Call the image parsing model to extract the core information of the engineering image and generate a structured text description. Based on the power monitoring threshold matching rule library, convert the monitoring values ​​into structured text to obtain a uniformly formatted mapping text.

[0041] Specifically, for engineering image data in non-text modalities, the CLIP (such as the ViT-B / 32 architecture) image parsing model is invoked to extract three core types of information: image annotation text, equipment contour features, and defect morphology features. Based on this, structured text descriptions are automatically generated. This parsing process strictly requires an annotation recognition accuracy of no less than 95%, and the output text format is forcibly standardized to JSON 1.0. The generated text fields consistently include four items: modality source, image content, associated entities, and confidence level. The system sets a confidence threshold of 0.85; mapping results below this threshold are automatically discarded to ensure mapping quality.

[0042] For monitoring numerical data in non-text modalities, a power monitoring threshold matching rule library containing a large number of rules (e.g., ≥1000 rules) is constructed. The monitoring indicators, real-time values, units, operating status and qualified thresholds are deeply correlated, thereby automatically generating structured text with the output format forced to JSON 1.0. The rule matching accuracy is required to reach over 98%.

[0043] Finally, the two types of mapping text were standardized by converting them to UTF-8 encoded JSON 1.0 format, forcing a fixed field order that could not be adjusted, and uniformly filling missing fields with specific markers (such as NULL) to ensure absolute consistency in the mapping text format.

[0044] S22: Based on a pre-built electrical engineering professional dictionary and a context clustering algorithm, semantic distinctions are made between easily confused terms, and entities in the mapped text are associated with pre-trained knowledge graph nodes to complete the deep injection of domain knowledge.

[0045] After standardizing the format, to avoid semantic discrepancies in power engineering review scenarios using generic terms, this embodiment constructs a dedicated power engineering dictionary, including over 8,200 terms across four categories: equipment, parameters, specifications, and defects. Each term is assigned a unique 5-digit code. To further improve semantic accuracy, the system employs a contextual clustering algorithm to identify homonymous terms, specifically distinguishing easily confused terms such as grounding, insulation, and sag.

[0046] Based on this, a pre-built knowledge graph for power grid engineering is imported (the graph has ≥100,000 nodes and ≥80,000 relationships). The graph contains four types of entities: equipment, parameters, specifications, and defects, and four types of relationships: belonging, associating, conforming, and exceeding standards. By fixing the knowledge embedding dimension to 256 dimensions, the system automatically associates and matches entities in the mapped text with graph nodes, requiring a matching rate of over 98%, thereby achieving deep injection of professional knowledge in the power field.

[0047] S23: The text after domain knowledge injection is divided into different categories according to function. The domain knowledge and the mapped text are semantically weighted and fused according to preset weights. After removing redundant semantics and completing the logic, the standard text is generated.

[0048] After knowledge injection is completed, the mapping texts are first divided into three categories according to their function in the review: rule-based, descriptive, and conclusion-based. Rule-based texts must retain the complete condition-requirement logical chain, descriptive texts must retain on-site details and causal relationships, and conclusion-based texts must highlight the core judgment results.

[0049] The semantic fusion stage then proceeds, employing a weighted fusion approach. The weight of domain knowledge is strictly set to 0.6, while the weight of the mapped text is set to 0.4. This weighted fusion automatically removes redundant semantics and completes missing logic. The final semantic integrity of the fused text is required to be no less than 90%.

[0050] Finally, the generated representation results are subjected to rigorous validation, with validation metrics including semantic retention rate (≥90%) and terminology accuracy (≥95%). For text that fails to meet the standards, the system automatically returns it to the image / numerical parsing stage to re-execute heterogeneous semantic mapping until a fully compliant standard text is output.

[0051] Step 3: Use a method that combines rule matching and machine learning to identify all text in the standard text and output a set of standard text blocks.

[0052] In this embodiment, the method for generating the standard text block set includes the following steps S31 to S33: S31: Based on rule matching and machine learning classification models, identify the modality type of text modality data, divide the text modality data into structured text, semi-structured text and unstructured text, and embed the corresponding exclusive modality type tags.

[0053] Specifically, a rule-based matching process is first used for initial screening. Text containing tables and fixed fields is accurately identified and classified as structured text; text containing hierarchical numbers and special markers (such as ①, ②, clauses, etc.) is accurately identified and classified as semi-structured text; all other text is classified as unstructured text. Subsequently, machine learning classification is used, employing a classification model (such as the BERT-base model) fine-tuned from the power grid review corpus. A 768-dimensional text feature vector is input, and the model outputs three modalities: structured, semi-structured, and unstructured. The model's recognition accuracy is strictly required to reach over 96.8%. Finally, unique markers are added to the recognition results: structured text is marked as [STRUCT], semi-structured text as [SEMI], and unstructured text as [UNSTRUCT]. These unique markers are embedded at the beginning of the text and are set to be undeletable and unmodifiable to ensure the accuracy of subsequent data flow.

[0054] S32: The structured text is divided using an attribute group clustering strategy, the semi-structured text is divided using a tagging and logic dual-drive strategy, and the unstructured text is divided using a clue and event dual-guide strategy, generating initial text blocks.

[0055] During adaptive partitioning, to ensure consistency of the partitioning results, all parameters of the partitioning strategy are locked and cannot be modified. The specific modal partitioning logic is as follows: For structured text, an attribute group clustering strategy is adopted, with "entity-parameter-criteria" as the core of clustering. The clustering similarity threshold is strictly set at 0.8, the number of clustering iterations is fixed at 100, the convergence threshold is set at 0.01, and the number of characters in a single text block is strictly controlled within the range of 50-200 characters.

[0056] For semi-structured text, a dual-driven strategy of tagging and logic is adopted. For example, engineering drawing annotation text is divided by annotation number, specification text is divided by three-level numbering, and review comment text is divided by question keywords. Tag matching accuracy is required to be ≥90%, and the number of characters in a single text block is strictly controlled between 80 and 250.

[0057] For unstructured text, a dual-guidance strategy of clues and events is adopted. For example, chronological text is divided by natural days, spatial text is divided by independent regions, and defective text is divided by causal chains. The accuracy of clue recognition is required to be ≥92%, and the number of characters in a single text block is strictly controlled within the range of 150-300 characters.

[0058] S33: Perform semantic integrity verification of core information and granularity appropriateness verification of word count range on the initial text block, and perform boundary merging and splitting optimization based on the semantic similarity of adjacent text blocks, and output the standard text block set bound with modal attribute information and verification results.

[0059] To ensure that the generated text blocks can be accurately vectorized, a semantic integrity check is first performed to check the completeness of the core information (entity-relationship-attribute) of the text block, with a strict requirement that the completeness rate be ≥96.5%. At the same time, a granularity suitability check is performed, which mandates that the number of characters in the text block be between 50 and 500. For overly bold text exceeding the word limit, automatic splitting is performed, and for overly fine text with insufficient word limit, automatic merging is performed.

[0060] In the boundary optimization stage, the system calculates the semantic similarity of adjacent text blocks. When the similarity is ≥0.6, the adjacent text blocks are automatically merged, and redundant text is split or removed. The optimized F1 score needs to reach more than 95%.

[0061] Finally, a unique 8-bit numeric ID is assigned to each optimized text block, which is then bound to the corresponding modality type, partitioning strategy, and verification result. The text is then uniformly formatted as UTF-8 encoded text, thus outputting a standard text block set.

[0062] Step 4: Extract the modality-specific features of the standard text block set and perform differential embedding. After linear spatial alignment and cross-modal attention fusion, generate a modality fusion vector with a unified dimension.

[0063] In this embodiment, generating a modality fusion vector of uniform dimension specifically includes the following steps S41 to S44: S41: Extract attribute association and parameter logic features, marker symbols and hierarchical logic features, spatiotemporal clues and event causal features from structured, semi-structured and unstructured text, and concatenate them with text word embedding vectors to generate an initial feature vector as modality-specific features.

[0064] Specifically, this embodiment accurately extracts three types of features from structured text: attribute association, numerical distribution, and parameter logic; accurately extracts three types of features from semi-structured text: marker symbols, hierarchical logic, and clause association; and accurately extracts three types of features from unstructured text: time clues, spatial clues, and event causality. To standardize input specifications, the vector dimension of each of the above modality-specific features is fixed at 64 dimensions. Feature concatenation is then performed, directly concatenating the 64-dimensional modality feature vector with the 768-dimensional text word embedding vector to generate an initial feature vector of 832 dimensions. This concatenation process strictly ensures no information loss and no dimensionality compression.

[0065] S42: Input the initial feature vector into the embedding model constructed based on the multi-task training mechanism, and output the differentiated dimensionality reduction text block vector, wherein the multi-task training mechanism includes modality classification, entity recognition, relationship prediction and semantic matching auxiliary tasks.

[0066] This embodiment constructs a 6-layer Transformer embedding model for differentiated processing. The model includes 8 attention heads, a hidden layer dimension of 512, a feedforward network dimension of 2048, and uses ReLU as the activation function. During multi-task training, the system sets four auxiliary tasks: modality classification, entity recognition, relationship prediction, and semantic matching, with corresponding loss weights strictly set to 0.5, 0.2, 0.2, and 0.1, respectively. The initial configuration of the model training parameters is as follows: learning rate 5e-5, batch size 32, number of iterations 20, dropout 0.3, and random seed fixed at 42. After processing by this model, different modalities of text achieve differentiated output representations: structured text features emphasize attribute associations, semi-structured text features emphasize logical relationships, and unstructured text features emphasize event causality. The model ultimately outputs a 512-dimensional text block vector after dimensionality reduction. In this stage, the training accuracy of the embedding model is strictly required to be ≥94%.

[0067] S43: Construct a transformation matrix specific to each text modality, multiply the text block vectors by the corresponding transformation matrices respectively, and map them to a unified semantic space to complete linear space alignment.

[0068] To eliminate the underlying representation differences between different modalities, this embodiment constructs dedicated transformation matrices for structured, semi-structured, and unstructured text, with a fixed matrix dimension of 512×512. During the matrix training phase, the system inputs cross-modal paired samples with a similarity ≥0.85 as the training set, uses Euclidean distance as the loss function, and iterates for a fixed 50 rounds to optimize the matrix parameters. After training, spatial mapping is performed, multiplying the three types of 512-dimensional text block vectors output by S42 with their corresponding dedicated transformation matrices, successfully mapping them to a unified 512-dimensional semantic space. Subsequently, the system performs strict alignment verification, requiring that the similarity of the aligned cross-modal vectors improve by ≥20% compared to before alignment. For mapping results that do not meet the standard, a mechanism to retrain the transformation matrix is ​​triggered.

[0069] S44: Introduce engineering image feature vectors and monitoring data feature vectors through a cross-modal attention mechanism, and perform weighted summation with the aligned text block vectors according to preset fusion weights to generate the unified-dimensional modal fusion vector.

[0070] Building upon linear spatial alignment, this embodiment further introduces a cross-modal attention mechanism, setting the number of attention heads to 8 and the computational dimension to 512, and performing normalization processing on the generated attention weights. Through this mechanism, the system introduces 512-dimensional engineering image feature vectors and 512-dimensional monitoring data feature vectors for multimodal fusion. During the fusion process, the fusion weights for each modality are strictly set as follows: text weight 0.6, image weight 0.2, and data weight 0.2. Weighted summation is performed according to these weights to finally generate a 512-dimensional modal fusion vector. Finally, a multimodal fusion verification is performed, strictly requiring a multimodal retrieval accuracy of ≥90% based on this fusion vector. If this target is not met, the preset fusion weights are automatically adjusted until the verification is passed.

[0071] Step 5: Build a basic architecture that includes a data layer, a knowledge layer, and an application layer, and embed a regional characteristic adaptation layer to generate a complete architecture system.

[0072] Specifically, the method for constructing a complete architecture includes the following steps S51 to S54: S51: Construct a hybrid storage architecture in the data layer, using a distributed file system to store the original images and unstructured documents, a relational database to store the standard text block set and labeled data, and a vector database to store the modality fusion vector.

[0073] The data layer employs heterogeneous storage configurations for data with different characteristics: For large volumes of raw images and unstructured documents, HDFS 3.3.6 distributed file system is used for storage, with files organized in a project-modality-time hierarchical directory structure, and the size of a single file is limited to no more than 500MB. For standard text block sets and labeled data with structured characteristics, DM 8.0 relational database is used for storage, with table fields fixed to include a unique ID, modality type, content, association ID, and verification result, and the size of a single data entry is strictly controlled within 1.8KB. For high-dimensional feature vectors output by S4, Milvus 2.2.8 vector database is used to specifically store 512-dimensional modality fusion vectors, with the index type set to IVF_FLAT, the number of clusters set to 1024, and the retrieval response time required to be no more than 50ms. To ensure the stability of the architecture, key parameters such as the system's storage path, index type, and retrieval threshold are all locked to fixed configurations and cannot be modified arbitrarily.

[0074] S52: In the knowledge layer, a knowledge graph containing devices, parameters, specifications, defects, and regional entities and their relationships is constructed through a graph database, and it is linked with a pre-built condition rule base and case base by foreign keys.

[0075] In the knowledge layer organization and construction phase, this embodiment uses the Nebula 3.0 graph database to build the underlying knowledge graph. The nodes of this graph include five core entity types: equipment, parameters, specifications, defects, and regions. The edges between nodes include five types of relationships: belonging, associating, conforming, exceeding standards, and regional association. Simultaneously, a conditional rule base is built using the DM 8.0 database, storing three types of rules—technology, review, and region—in IF-THEN format, and setting fixed priority control logic for the rules. Furthermore, a power grid review case library is constructed, containing fields such as project information, multimodal data, review conclusions, and regional characteristics. Finally, the knowledge graph establishes a deep mapping with the conditional rule base and case library through entity ID foreign keys, requiring an accuracy rate of ≥98% for knowledge association between the libraries.

[0076] S53: Deploy a retrieval interface that supports entity and vector retrieval, a reasoning interface that supports causal judgment, and a multimodal linkage interface that supports the synchronous display of text, images, and data at the application layer.

[0077] The application layer, serving as the system's external output port, uniformly adopts the RESTful 1.2 protocol and JSON 1.0 data exchange format. Its three core deployments are: a retrieval interface, an inference interface, and a multimodal linkage interface. The retrieval interface supports entity retrieval based on knowledge graphs, vector retrieval based on Milvus, and multi-condition combination retrieval, with a system response time strictly limited to within 80ms. The inference interface supports forward derivation, reverse tracing, and causal inference based on the knowledge graph topology and rule base, achieving an accuracy rate of over 90%. The multimodal linkage interface supports the synchronous linkage and aligned display of text descriptions, on-site images, and monitoring data on the power grid review front-end interface, with a matching rate of ≥92% between different modalities.

[0078] S54: Embed the aforementioned regional characteristic adaptation layer to extend the relevant equipment and engineering entities with regional exclusive attributes including typhoon intensity, geological disaster level and ecological red line area, add corresponding environmental impact and avoidance correlation, and supplement regional exclusive rules with higher priority than general review rules to complete the binding and adaptation of regional characteristics.

[0079] To overcome the limitations of general-purpose power grid knowledge graphs in adapting to specific natural environments, this embodiment statically extends the underlying model by embedding a regionally specific adaptation layer. First, three exclusive attribute slots—typhoon intensity, geological hazard level, and ecological red line zone—are added to the graph ontology structure for equipment and engineering entities, ensuring the uniqueness of attribute codes. Second, three types of relationship edges—affected by typhoons, geologically related, and avoiding ecological red lines—are added to the relationship template, with the initial weight of these regional relationships fixed at 0.15. Subsequently, three types of regionally specific rules are added to the conditional rule base (e.g., coastal tower wind resistance ≥35m / s, ecological red line zone avoidance distance ≥500m, and mountainous geological avoidance rules), and the priority of these regionally specific rules is set to be absolutely higher than other general review rules in the system execution logic. Finally, 300 specialized regional review cases are imported as basic corpora and integrated with regional entities, thereby completing the deep configuration and adaptation of regional features at the system infrastructure level.

[0080] Step Six: Bind region-specific attributes and region review rules to related entities in the complete architecture system, and perform knowledge conflict resolution to build a fused knowledge base.

[0081] Specifically, the steps S61 to S63 are included: S61: Similarity calculation and consistency rule verification based on entity fusion vectors are used to perform cross-modal entity alignment, and a graph attention network model combined with preset rules is used to complete multi-source semantic relationship completion.

[0082] In the cross-modal entity alignment stage, the cosine similarity of the fused vectors of various entities is first calculated, and the judgment threshold is fixed at 0.85 to initially screen candidate alignment entities. Then, a strict consistency rule check is performed, filtering invalid alignment results by comparing whether the device number, parameter value, and project ID are completely consistent. To ensure the rigor and security of power grid knowledge entry into the database, high-risk entities with similarity scores between 0.8 and 0.85 are blocked and transferred to review experts with senior engineer titles for manual confirmation. After completing the above process, a unified ID is assigned to successfully aligned entities.

[0083] In the multi-source semantic relation completion stage, the system constructs and invokes a Graph Attention Network (GAT) model, setting its hyperparameters to: 256-dimensional hidden layers, 4 attention heads, 50 iterations, and a dropout rate of 0.3. The model outputs prediction confidence scores to filter out potential semantic relations with a confidence score ≥ 0.7. Based on this inference result, and combined with pre-defined power rules, specific edge relationships are further added for specific regions and regulatory requirements, achieving an F1 score of over 92.8% for relation completion. Finally, automated verification of the results is performed to remove redundant relationships and correct logical errors, ensuring the rigor of the underlying graph relationships.

[0084] S62: During multimodal knowledge fusion, the knowledge conflict resolution is performed according to the preset confidence weighting mechanism and the normative priority determination mechanism.

[0085] Since the fusion of multi-source and multi-modal data inevitably involves conflicts between knowledge and rules, this embodiment establishes a hierarchical conflict resolution mechanism. First, a confidence-weighted approach is implemented, assigning fixed quantitative confidence levels to different dimensions of knowledge sources: technical specifications have the highest weight of 0.9, expert experience is next at 0.7, and historical cases have a weight of 0.6. Second, a specification priority determination mechanism is introduced. When conflicting review specifications at different levels occur, the system strictly adheres to the legally mandated priority order of "national specifications over industry specifications, industry specifications over local specifications" and "local specifications over general specifications" for coverage determination. For high-confidence conflicts that cannot be resolved after the first two rounds of mechanisms, or conflicts involving core engineering safety, they are intercepted and transferred to review experts for manual adjudication. Through this closed-loop hierarchical conflict resolution process, the final conflict resolution rate of the knowledge base reaches over 96.5%.

[0086] S63: Bind typhoon, geological and ecological red line attributes to regionally related entities, and increase the weight of regionally related knowledge during retrieval and reasoning. When compliance verification is triggered, control the regional review rules to prioritize covering general rules.

[0087] To ensure that the regional characteristics preset at the architecture layer are truly implemented in the data fusion stage, this embodiment dynamically performs deep fusion of regional characteristic knowledge at runtime. For engineering projects that trigger regional determination, the corresponding regionally associated equipment or engineering entities are dynamically bound with three specific physical environmental attributes: typhoon, geological, and ecological red line. During multimodal retrieval and intelligent reasoning, the node weight coefficients are dynamically adjusted, forcibly increasing the weight of knowledge with regionally relevant tags by 20% to ensure that it is prioritized for retrieval and reasoning output.

[0088] More importantly, during defect identification and drawing compliance verification, once a relevant entity has a regional label, the system's control logic will prioritize triggering region-specific rules (such as requiring higher wind resistance standards or greater avoidance distances), directly overriding the original nationally applicable review rules. Finally, the fusion results undergo rigorous quantitative acceptance testing. After ensuring a regional knowledge correlation rate of ≥94% and a regional adaptation accuracy rate of ≥95%, multimodal knowledge layered fusion is completed, outputting a final usable multimodal fusion knowledge base with deep regional adaptability.

[0089] In this embodiment, after completing step six and using the constructed fusion knowledge base as the final knowledge base, step seven is also included: constructing a five-dimensional quality inspection index system, performing knowledge base quality verification, parameter iterative optimization, and dynamic incremental updates to complete the quality closed loop and dynamic maintenance, and dynamically updating the fusion knowledge base obtained in step six.

[0090] Step seven specifically includes the following steps S71 to S73: S71: Construct a five-dimensional quality inspection index system based on accuracy, completeness, consistency, relevance, and usability, and perform quality verification on the fused knowledge base.

[0091] To ensure the high availability and reliability of the multimodal fusion knowledge base in power grid engineering reviews, this embodiment constructs a rigorous five-dimensional quality inspection system to quantitatively accept the output knowledge base. The threshold requirements for specific inspection indicators are strictly limited as follows: Accuracy verification: The entity accuracy rate of the map should be ≥97.8%, the relation accuracy rate ≥96.5%, and the rule accuracy rate ≥98.2%. Completeness verification: The knowledge graph must have an entity coverage rate of ≥93.2% in business scenarios and a review scenario coverage rate of ≥91.8%. Consistency verification: Requires logical consistency of entities ≥99.1% and consistency of compliance judgment rules ≥98.7%; Relevance verification: The cross-modal correlation rate between different data sources should be ≥94.3%, and the multimodal joint retrieval accuracy rate should be ≥95.7%. Usability verification: The accuracy of automated review and reasoning based on this knowledge base should be ≥92.3%, and the system retrieval response time should not exceed 80ms.

[0092] S72: For knowledge data that fails verification, perform error backtracking to the corresponding stage and iteratively optimize the model parameters and review rules.

[0093] During the quality verification process, if knowledge data that fails to meet the above threshold indicators is found, it will not be discarded directly. Instead, an error backtracking mechanism will be automatically triggered to accurately backtrack it to the corresponding steps (steps two to six) of the preceding steps, such as multimodal preprocessing, heterogeneous mapping, or differentiated embedding, for reprocessing.

[0094] Simultaneously, addressing the root causes of errors, this embodiment also initiates fine-tuning of model parameters and iterative optimization of rules: at the model network level, the system fine-tunes the learning rate by ±1e-5, adjusts the number of iterations by ±5 rounds, and corrects dropout by ±0.05; at the knowledge rule level, logical conflicts are corrected and missing review rules are supplemented. To ensure the rational use of system computing resources and avoid getting stuck in an infinite loop, the iterative optimization process sets a mandatory termination condition, namely, the upper limit of the number of iterations is set to ≤3 rounds, and the iteration terminates immediately if the reconstructed knowledge meets the target within this period.

[0095] S73: Upon receiving a trigger instruction for adding a review project, revising a standard, or updating regional data, perform incremental collection and processing of multimodal data, and integrate the incremental knowledge into the fused knowledge base for dynamic incremental updates and verification.

[0096] To ensure the timeliness of power grid engineering review rules and on-site environmental data, this embodiment has automated dynamic incremental update capabilities. When the system detects new review projects being initiated, revisions to national / industry standards, or updates to regional data (such as typhoon paths and geological disaster sites) issued by government platforms such as the Department of Natural Resources, the incremental update process will be automatically triggered.

[0097] Upon triggering, incremental data collection is first executed, synchronously pulling new data and reusing the standardized preprocessing and differentiated vectorization processes from steps one to four to generate new modal fusion vectors. Subsequently, incremental fusion is performed, deeply integrating the newly generated entity nodes and standardized logic into the current knowledge base and establishing associations with existing knowledge nodes. Finally, a five-dimensional quality check is performed on the new parts again. Only when the accuracy of the incremental knowledge reaches the red line standard of ≥95% can the dynamic incremental update take effect and be put into production use, thereby achieving automatic evolution and closed-loop maintenance of the entire lifecycle of the review knowledge base.

[0098] Example 2 This embodiment is a knowledge base construction system for power grid engineering review based on modal perception. The system is used to implement the knowledge base construction method as described in Embodiment 1, including: The data acquisition and preprocessing module is used to acquire and preprocess structured data, semi-structured data, unstructured data, engineering images, and monitoring values ​​for power grid engineering review. The heterogeneous semantic mapping module is used to convert preprocessed engineering images and monitoring values ​​into standardized text through heterogeneous semantic mapping. The modality segmentation module is used to identify text using a fusion of rule matching and machine learning methods, and output a standard set of text blocks. The differential embedding and fusion module is used to extract modality-specific features from the standard text block set and perform differential embedding. After linear spatial alignment and cross-modal attention fusion, a modality fusion vector with a unified dimension is generated. The architecture building module is used to build a basic architecture that includes a data layer, a knowledge layer, and an application layer, and embeds a regional characteristic adaptation layer to generate a complete architecture system. The knowledge resolution and construction module is used to bind region-specific attributes and region review rules to related entities in the complete architecture system, and to perform knowledge conflict resolution to build a fused knowledge base.

[0099] Example 3 This embodiment is an electronic device, including: One or more processors; Memory, used to store one or more computer programs; One or more computer programs stored in the memory are executed by the one or more processors, causing the one or more processors to implement the knowledge base construction method as described in Embodiment 1.

[0100] Example 4 This embodiment is a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the knowledge base construction method as described in Embodiment 1.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for constructing a knowledge base for grid engineering review based on modal perception, characterized in that, Includes the following steps: Acquire and preprocess structured, semi-structured, and unstructured data, engineering images, and monitoring values ​​for power grid engineering review; The preprocessed engineering images and monitoring values ​​are mapped into standardized text through heterogeneous semantic mapping. A method combining rule matching and machine learning is used to identify text and output a standard set of text blocks. Modality-specific features of the standard text block set are extracted and differentially embedded. After linear spatial alignment and cross-modal attention fusion, a modality fusion vector with a unified dimension is generated. Build a basic architecture that includes a data layer, a knowledge layer, and an application layer, and embed a regional characteristic adaptation layer to generate a complete architecture system; By binding region-specific attributes and region-specific review rules to related entities in the complete architecture system, and performing knowledge conflict resolution, a fused knowledge base is constructed.

2. The method of claim 1, wherein, After constructing the fusion knowledge base, the method also includes the following steps: A five-dimensional quality inspection index system is constructed based on accuracy, completeness, consistency, relevance, and practicality, and quality verification is performed on the fused knowledge base. For knowledge data that fails verification, the error is traced back to the corresponding stage, and the model parameters and review rules are iteratively optimized. When a trigger instruction is received for a new review project, a revision of a standard, or an update of regional data, incremental collection and processing of multimodal data is performed, and the incremental knowledge is associated and integrated into the fused knowledge base for dynamic incremental updates and verification.

3. The method of claim 1, wherein, The method for generating the standard text is as follows: The image parsing model is called to extract the core information of the engineering image and generate a structured text description. Based on the power monitoring threshold matching rule library, the monitoring values ​​are converted into structured text to obtain a uniformly formatted mapping text. Based on a pre-built power engineering professional dictionary and a context clustering algorithm, semantic distinctions are made between easily confused terms, and entities in the mapped text are associated with pre-trained knowledge graph nodes to complete the deep injection of domain knowledge. The text after domain knowledge injection is divided into different categories according to function. The domain knowledge and the mapped text are semantically weighted and fused according to preset weights. After removing redundant semantics and completing the logic, the standard text is generated.

4. The method of claim 1, wherein, The method for generating the standard text block set is as follows: Based on rule matching and machine learning classification models, the modality type of text modal data is identified, and the text modal data is divided into structured text, semi-structured text and unstructured text, and corresponding exclusive modality type tags are embedded. The structured text is divided using an attribute group clustering strategy, the semi-structured text is divided using a tagging and logic dual-drive strategy, and the unstructured text is divided using a clue and event dual-guide strategy, generating initial text blocks. The initial text block is subjected to semantic integrity verification of core information and granularity appropriateness verification of word count range. Based on the semantic similarity of adjacent text blocks, boundary merging and splitting optimization is performed, and the standard text block set bound with modal attribute information and verification results is output.

5. The method of claim 1, wherein, The generation of a uniform-dimensional modality fusion vector includes the following steps: For structured, semi-structured, and unstructured text, we extract attribute association and parameter logic features, marker symbols and hierarchical logic features, spatiotemporal clues and event causal features, and use them as modality-specific features to concatenate with text word embedding vectors to generate an initial feature vector; The initial feature vector is input into the embedding model built based on the multi-task training mechanism, and the output is a differentiated dimensionality reduction text block vector, wherein the multi-task training mechanism includes modality classification, entity recognition, relationship prediction and semantic matching auxiliary tasks; Construct transformation matrices specific to each text modality, multiply the text block vectors by the corresponding transformation matrices respectively, and map them to a unified semantic space to achieve linear space alignment; The engineering image feature vector and the monitoring data feature vector are introduced through a cross-modal attention mechanism, and then weighted and summed with the aligned text block vector according to a preset fusion weight to generate the unified-dimensional modal fusion vector.

6. The method of claim 1, wherein, The method for constructing the complete architecture is as follows: A hybrid storage architecture is constructed in the data layer, using a distributed file system to store the original images and unstructured documents, a relational database to store the standard text block set and annotation data, and a vector database to store the modality fusion vectors. In the knowledge layer, a knowledge graph containing devices, parameters, specifications, defects, and regional entities and their relationships is constructed through a graph database, and it is linked with a pre-built condition rule base and case base by foreign keys; The application layer deploys a retrieval interface that supports entity and vector retrieval, a reasoning interface that supports causal judgment, and a multimodal linkage interface that supports the synchronous display of text, images, and data. By embedding the regional characteristic adaptation layer, the relevant equipment and engineering entities are extended with regional-specific attributes including typhoon intensity, geological disaster level and ecological red line area, corresponding environmental impact and avoidance correlation are added, and regional-specific rules with higher priority than general review rules are added to complete the binding and adaptation of regional characteristics.

7. The method of claim 1, wherein, The process involves binding region-specific attributes and region-specific review rules to related entities within the complete architecture, resolving knowledge conflicts, and constructing a fused knowledge base. This includes the following steps: Cross-modal entity alignment is performed based on similarity calculation and consistency rule verification of entity fusion vectors, and multi-source semantic relationship completion is completed by using a graph attention network model combined with preset rules. During multimodal knowledge fusion, the knowledge conflict resolution is performed based on a preset confidence weighting mechanism and a normative priority determination mechanism. The system binds typhoon, geological, and ecological red line attributes to geographically related entities and increases the weight of geographically related knowledge during retrieval and reasoning. When compliance verification is triggered, the system controls the geographically reviewed rules to prioritize covering general rules.

8. A modal perception based power grid engineering review knowledge base building system characterized in that, The system is used to implement the knowledge base construction method as described in any one of claims 1-7, including: The data acquisition and preprocessing module is used to acquire and preprocess structured data, semi-structured data, unstructured data, engineering images, and monitoring values ​​for power grid engineering review. The heterogeneous semantic mapping module is used to convert preprocessed engineering images and monitoring values ​​into standardized text through heterogeneous semantic mapping. A modal division module is configured to recognize text by using a rule matching and machine learning fusion method, and output a standard text block set; A differentiated embedding and fusion module is configured to extract modal-specific features of the standard text block set and perform differentiated embedding, linear space alignment, and cross-modal attention fusion to generate a modal fusion vector in a unified dimension; An architecture building module is configured to build a basic architecture including a data layer, a knowledge layer, and an application layer, and embed a regional characteristic adaptation layer to generate a complete architecture system; A knowledge resolution and construction module is configured to bind regional-specific attributes and regional review rules to associated entities in the complete architecture system, perform knowledge conflict resolution, and construct a fusion knowledge base. 9.An electronic device, comprising: one or more processors; a memory for storing one or more computer programs; characterized in that the one or more computer programs stored in the memory are executed by the one or more processors, so that the one or more processors implement the knowledge base construction method according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the knowledge base construction method according to any one of claims 1-7.

Images