Method and system for intelligent generation of power project material inventory based on multi-modal document parsing and knowledge graph semantic mapping

Abstract

This invention discloses an intelligent generation method and system for power project material lists based on multimodal document parsing and knowledge graph semantic mapping, belonging to the fields of artificial intelligence and power engineering. It includes: a first stage of unified parsing and semantic encoding of text documents, engineering drawings, and tabular data to generate a semi-structured raw information set; a second stage of named entity recognition and type labeling of the raw information set based on a bidirectional long short-term memory network and a conditional random field hybrid model to generate a structured entity set; on this basis, a knowledge graph for the power material domain is constructed, a standard key point rule base is established, and semantic mapping and information completion are performed through a multi-level matching mechanism and knowledge reasoning; after quality control, a standardized material list is generated. This invention achieves intelligent and automated generation of power project material lists, improving the accuracy and efficiency of material list generation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and power engineering technology, specifically to a method and system for intelligent generation of power project material inventory based on multimodal document parsing and knowledge graph semantic mapping. Background Technology

[0002] The inventory of materials for power projects is a crucial foundational document for project construction and management, accurately and completely recording information on all equipment, materials, and supplies required for the project. Traditional methods of compiling these inventory lists primarily rely on manual processing, which presents the following significant problems:

[0003] First, the processing efficiency is low. Power project design documents come in various formats, including feasibility study reports, preliminary design documents, construction drawings, engineering drawings, equipment lists, etc. These formats range from text documents like Word and PDF to CAD drawings and Excel spreadsheets. Designers need to manually extract information from a large number of unstructured and semi-structured documents, resulting in a huge workload.

[0004] Secondly, accuracy is difficult to guarantee. Omissions and errors are prone to occur during manual processing. Inconsistencies may exist between different documents, making comprehensive manual verification difficult. The specifications and models of materials are complex, with numerous technical parameters, making transcription prone to errors and severely impacting subsequent procurement and management.

[0005] Third, there is insufficient standardization. Different design units and designers use different standards and specifications, resulting in inconsistent standards for compiling material inventories. The lack of unified standards for material naming, coding, and classification creates difficulties for material management.

[0006] Fourth, insufficient knowledge reuse. Existing methods struggle to fully utilize historical project data and domain knowledge. A large amount of historical project information and expert experience has not been effectively integrated and utilized; each time a materials inventory is compiled, the process must begin afresh, hindering the accumulation and transfer of knowledge.

[0007] Fifth, quality control is difficult. There is a lack of effective consistency verification and integrity checking mechanisms. Information conflicts between different documents are difficult to detect, material matching relationships are easily overlooked, and the matching of technical parameters is difficult to verify.

[0008] In recent years, with the rapid development of artificial intelligence technology, deep learning, natural language processing, knowledge graphs, and other technologies have made significant progress in document understanding, information extraction, and knowledge reasoning. However, existing technologies still have the following shortcomings when applied to the generation of material lists for power projects: insufficient multimodal processing capabilities, making it difficult to uniformly process documents of various modalities such as text, drawings, and tables; insufficient utilization of domain knowledge, with general information extraction models lacking power industry expertise; insufficient semantic understanding depth, with simple keyword matching unable to understand the semantic content of documents; unclear information processing flow, with ambiguous logical relationships between document parsing and entity recognition; and imperfect quality control mechanisms, making it difficult to guarantee the accuracy and completeness of the generated results. Summary of the Invention

[0009] To address the shortcomings mentioned in the background technology, the present invention aims to provide an intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping. Through a two-stage processing flow, firstly, unified parsing and semantic encoding of multimodal documents are performed, followed by precise entity recognition and type labeling to establish clear information flow relationships. Furthermore, by combining knowledge graph technology to accumulate and utilize domain knowledge, intelligent association between document content and material requirements is established through semantic mapping. A quality control mechanism ensures the accuracy and completeness of the generated results, thereby significantly improving the efficiency and quality of material inventory compilation.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] A method for intelligently generating material lists for power projects based on multimodal document parsing and knowledge graph semantic mapping is disclosed. The power project design documents include unstructured text data, engineering drawing data, and semi-structured tabular data. The unstructured text data consists of Word and PDF files such as feasibility study reports, preliminary design specifications, and technical specifications generated during the project design process. The engineering drawing data consists of CAD drawings such as electrical main wiring diagrams and floor plans. The semi-structured tabular data consists of Excel files such as equipment lists and material lists. The power project design documents utilize natural language processing and computer vision technologies for information extraction, semantic mapping, and knowledge reasoning. The intelligent material list generation method specifically includes the following steps:

[0012] Step 10) Obtain the power project design documents, perform format recognition and preprocessing on the documents, and establish a document index set. ,in As a unique identifier for the document, For document type, Document path This is the document's metadata.

[0013] Step 20) Perform unified processing on documents of different formats, converting heterogeneous data into a unified intermediate representation format to generate the original information set. The original information set For a semi-structured intermediate representation, each information entry It includes the following elements: content (text content, which is a text fragment for text, detection results and OCR text for drawings, and cell content for tables); semantic_feature (semantic feature vector, which is a BERT encoded vector for text, and feature description for drawings and tables); source_info (source information, including document ID, chapter, page number, location, etc.); and modality_type (modality type identifier, with values ​​of TEXT, DIAGRAM, or TABLE).

[0014] Output of step 20) It is a semi-structured information collection in a unified format, where the text content has been semantically encoded but has not yet undergone fine-grained entity boundary recognition and type labeling. These tasks will be completed in step 30). The processing in step 20) includes:

[0015] Extract text content from Word and PDF documents while preserving paragraph structure. For long texts, use a sliding window to segment them into several text fragments. (Window length...) Tokens, overlap length One token; text semantic encoding is performed using the BERT pre-trained language model. The BERT model employs a 12-layer Transformer encoder structure, with hidden layer dimensions... Multi-head attention mechanism head count For each text segment, a contextual semantic representation is obtained using the BERT model. ,in The number of tokens. The text document processing unit outputs text fragments and their corresponding BERT semantic vectors as a unified text representation.

[0016] The BERT model is used in a large-scale power domain corpus. Continue pre-training, including: building a corpus in the power sector. We collected power industry technical documents, academic papers, standards and specifications, equipment manuals, etc., from power industry websites, digital libraries, etc., with a corpus size of ≥1000 documents. We used a masked language modeling (MLM) task, randomly masking 15% of the tokens, where 80% were replaced with [MASK], 10% were replaced with random tokens, and 10% remained unchanged. The pre-trained objective function was... ,in Indicates except All tokens outside the specified location; construct sentence pairs using the next sentence prediction NSP task. 50% are consecutive sentence pairs, 50% are random combinations, and the pre-training objective is... Combination loss function Using the AdamW optimizer, the learning rate , ,train , After further pre-training, the model's F1 score for recognizing electrical engineering terminology significantly improved.

[0017] The CAD drawings are rasterized and converted into PNG images with a resolution of R=300 DPI. For the converted PNG images, an improved Faster R-CNN object detection model is used to identify device symbols, obtaining their category and bounding box location information (bounding box coordinates). Simultaneously, OCR technology is used to extract the labeled text from the images, employing the PaddleOCR model, which includes a text detection network (DB) and a text recognition network (CRNN). The drawing recognition processing unit outputs the detected device symbol categories, bounding box locations, and the OCR-recognized text content.

[0018] The improved Faster R-CNN object detection model includes: using ResNet-50 as the backbone network and adding a Feature Pyramid Network (FPN) to enhance multi-scale feature extraction capabilities; and outputting feature maps at the conv2, conv3, conv4, and conv5 layers of ResNet-50. , , , Constructing pyramid features from top to bottom , , , ,in For 2x upsampling, 1×1 convolution; at each pyramid level The independent Region Candidate Network (RPN) is used, with anchor box scales = {32², 64², 128², 256², 512²} and aspect ratios = {1:2, 1:1, 2:1}; in the power equipment symbol annotation dataset. Training is performed; FocalLoss is used for classification loss to address class imbalance. ,in To predict probabilities, As a balance factor, To focus parameters; use ROIAlign instead of ROIPooling to avoid quantization errors; multi-task loss function. ,in For classifying losses, For bounding box regression loss, This is the balance coefficient.

[0019] Use the openpyxl library to read all worksheets from an Excel file; then use the Hough transform to detect table lines, the mathematical expression of which is: ,in The distance from the origin to the line is denoted as . For the normal vector of the line and The angle between axes; determining the table structure based on detected table lines, identifying the header and data areas; parsing the table structure and cell content, identifying cell data types; establishing a preliminary mapping relationship between table columns and material attributes. ,in Name the columns in the table. The material attribute name is matched using a preset column name-attribute mapping rule dictionary. The table analysis and processing unit outputs the structured cell content, cell location information, and column attribute mapping relationships of the table. .

[0020] Through the above processing, a set of original information in a unified format is obtained. . The key features are: uniformity of format, information from different modalities is converted into a uniform data structure; semantic richness, text information contains BERT semantic encoding; information originality, original content is preserved, without fine-grained entity segmentation and type labeling; traceability, each information entry records detailed source information; and semi-structured, an intermediate representation between unstructured original documents and fully structured entity sets.

[0021] Step 30) Based on the bidirectional LSTM+CRF machine learning model, process the original information set generated in step 20). Named entity recognition is performed on the data to determine entity boundaries and types, and an entity set is generated. ,in The total number of entities identified. For the first The text content of each entity, For the first Type label for each entity, For the first The text position range of an entity, The index of information entries from which this entity originates. ,in The specific process is as follows:

[0022] Step 31) Prepare input data

[0023] First, from the original information set The text content (content) and semantic feature vector (semantic_feature) are extracted from the data. Specifically, for the text modality, the content field is extracted directly; for the drawing modality, the text recognized by OCR is extracted; and for the table modality, the cell content is extracted.

[0024] Next, the extracted text sequence Convert to character-level input, where For the sequence number The character at position T, where T is the total length of the sequence. .

[0025] Then, each character is processed through a feature encoding layer. Mapping to word vectors Among them, word vector dimension In the LSTM model, for each position of the sequence... Process sequentially. To fully utilize the BERT semantic information obtained in step 20), the obtained BERT encoding... Additional features are concatenated and fused to form the fused feature representation. Or use attention fusion , ,in , , For learnable parameters, This represents the attention weight.

[0026] Step 32) Represent the fused features obtained in step 31). The input is encoded using a Bidirectional Long Short-Term Memory (BiLSTM) model. BiLSTM consists of a forward LSTM and a backward LSTM: the forward LSTM processes the sequence from left to right, capturing the preceding (left-side) context information; the backward LSTM processes the sequence from right to left, capturing the following (right-side) context information. At each time step t, the calculation formula is as follows:

[0027] Input gate

[0028] Forgotten Gate

[0029] Output gate

[0030] Candidate Memory

[0031] memory unit

[0032] Hidden state

[0033] in, , and These represent the input gate, forget gate, and output gate at the current moment, respectively. and These represent the memory unit at the current moment and the candidate memory, respectively. The state was hidden in the previous moment; This is the current input feature vector; This is element-wise multiplication; for Activation function, defined as ; , , , This is the corresponding weight matrix; , , , This is the corresponding bias vector.

[0034] Hidden state of the forward LSTM Hidden state of backward LSTM By splicing them together, we can obtain Hidden layer dimensions .

[0035] Step 33) Decode using the CRF model

[0036] Define label sequence ,in For position The tag, For a set of tags, This represents the number of label categories. During the training phase, Known manual annotations; during the decoding phase, The variable to be predicted is the hidden state output by the BiLSTM. Mapping to the label space via linear transformation: ,in Indicates the location For tags The emission score. Define the state transition matrix. ,in The number of label categories, where Indicates from the label Transfer to label The transition score. For the label sequence Calculate sequence score Then, the conditional probability is calculated using softmax. ,in Given the set of all possible label sequences. Maximize the log-likelihood during training. During decoding, the Viterbi algorithm is used to find the optimal tag sequence. Based on the predicted label sequence y*, entity labels are performed according to the BIOES labeling system. The label format is "BIO tag-entity type", where BIO tag: B-start, I-inner, O-outer, E-end, S-word; entity type... MAT is the material name, SPEC is the specification, PARAM is the technical parameter, QUANT is the quantity, UNIT is the unit, and MANUF is the manufacturer. Based on the predicted label sequence... Combine consecutive BI*-E tags or single S tags into a complete entity and extract the entity text. ,type ,Location and source Construct entity set .

[0037] The relationship between step 30) and step 20) is as follows: step 20) completes coarse-grained format unification and semantic encoding, outputting semi-structured data. Step 30) Complete fine-grained entity boundary identification and type labeling, and output structured data. The BERT semantic features from step 20) serve as important input features for the BiLSTM-CRF model in step 30), thereby improving recognition accuracy.

[0038] Step 40) Construct a knowledge graph for the power materials domain based on the ontology model. ,in For a set of entity nodes, Let be the set of relation edges. It is a collection of attributes.

[0039] The ontology model design uses OWL to describe the ontology, defining a class hierarchy: the top-level class is ElectricalMaterial, with first-level subclasses including PrimaryEquipment, SecondaryEquipment, CableMaterial, and AuxiliaryMaterial. Second-level and more detailed categories, such as PrimaryEquipment, include Transformer, CircuitBreaker, and Isolator. Object properties are defined, including hasComponent, compatibleWith, replacedBy, and configuredWith relationships. Data properties are defined, including materialCode, materialName, specification, voltageLevel, and capacity.

[0040] Knowledge extraction includes: entity extraction, based on the entity set identified in step 30). Entity nodes are constructed; relations are extracted using an attention-based relation classification model, with input entity pairs. and its context BERT encoding is used to obtain contextual representations and compute relational representations of entity pairs. ,in Entities Vector representation, entity Vector representation and context The vector representation of the matrix is ​​used to predict the relationship type through a softmax classifier. Attribute extraction: Extracting entity attribute values ​​from text using sequence labeling methods.

[0041] Knowledge fusion includes entity alignment, using entity matching algorithms based on edit distance and semantic similarity. ,in To normalize the edit distance, For cosine similarity based on word vectors, For weighting coefficients; entity disambiguation, for entities with the same name but different meanings, the disambiguation score is calculated based on contextual information. ,in It is a feature set, including co-existing entities, attribute features, etc. For feature weights, The feature function is used for knowledge deduplication, which involves merging duplicate knowledge by calculating the similarity of knowledge triples.

[0042] The knowledge graph is stored using the Neo4j graph database. Entity nodes are created using Cypher: `CREATE(n:label{attribute list})`, and relationship edges are created: `CREATE(node1)-[:relation type{relation attribute}]->(node2)`. An index is built to speed up queries: `CREATEINDEXFOR(n:label)ON(n.attribute)`. The completed knowledge graph contains |V| nodes and |E| relationship edges.

[0043] Step 50) Based on the dependency parsing model, process the entity set generated in step 30). Relationship and attribute extraction is performed, and the dependency parser MSTarser based on the maximum spanning tree (MST) theory is used to analyze the dominance and subordination relationships between words in the sentence to reveal its syntactic structure. That is, by analyzing the grammatical components such as subject, verb, object, attributive, adverbial, and complement contained in the sentence, and analyzing the relationships between the components.

[0044] The mathematical model for MSTParser is as follows:

[0045]

[0046] in, Indicates the score. It is a sentence A dependency tree (a set of directed edges). It is a word governing words Dependency edges, It is a high-dimensional binary feature function vector. Represents a node and The dependency relationship between them takes the value 0 or 1. It is a feature The weight vector is obtained through sample training after the feature set is determined. This invention uses the Eisner algorithm for decoding to obtain the dependency tree with the highest score. ,in For sentences The set of all possible dependency trees.

[0047] Based on dependency syntax trees, syntactic relations such as subject-verb-object (SVO), attributive-head relation (ATT), and verb-object relation (VOB) are extracted. Attribute modification relations are identified, numerical attributes are extracted, attribute values ​​are normalized, and a set of entity relation triples is constructed. ,in For the head entity, For the relationship, It is a tail entity.

[0048] Step 60) Establish a mapping relationship between the content of power project documents and standard material requirements, and construct a standard key point rule base. The size is |R| = k rules, each rule Includes the following elements: triggering conditions Define the preconditions for rule activation and the matching mode. Define how to identify relevant information and the conversion rules. Define how to convert to standard material requirements, confidence level Indicates the credibility of the rule and .

[0049] A multi-level matching mechanism is adopted: the first level is exact matching, which uses a hash table to store keywords, with a time complexity of O(1); the second level is fuzzy matching, which calculates the edit distance. and similarity ,when A successful match is considered to occur at this time, by default. The third layer of semantic matching calculates cosine similarity based on word vectors. ,in and They are respectively and The word vector representation, when For a successful semantic match, the default is... .

[0050] Document content is established through a multi-level matching mechanism. With the demand for supplies Mapping relationship between The final confidence score is calculated by comprehensively considering the matching level and contextual information. ,in = For rules Self-confidence level; To determine the confidence level, a value of 1.0 is used for exact matching and 0.0 for fuzzy matching. Or semantic matching ; Contextual confidence.

[0051] Step 70) Use knowledge graphs to reason and automatically complete the implicit material requirements.

[0052] Rule-based reasoning defines a set of reasoning rules. The rule format is as follows: Reasoning can be performed using either forward chain reasoning or backward chain reasoning.

[0053] Graph-based reasoning discovers implicit relationships through paths in the knowledge graph, where the paths are represented as follows: Calculate path credibility Subgraph matching identifies similar resource configuration patterns using a graph isomorphism algorithm; community detection uses the Louvain algorithm to identify closely related resource groups.

[0054] Embedded reasoning uses the TransE model to learn vector representations of entities and relations. loss function ,in For positive sample sets, For negative sample set, Distance function Norm, For boundary parameters, Predicting missing relationships through vector operations, for entity pairs. ,calculate Then, the nearest neighbor is found in the physical space. Through knowledge reasoning, the complete set of material requirements is obtained. .

[0055] Step 80) Perform quality control on the generated material list.

[0056] Cross-document consistency verification involves checking the consistency of information about the same material appearing in different documents and defining a consistency metric. ,in For supplies In the document and The difference measure in the data is that when the consistency is below a threshold, it is marked as a conflict and requires manual confirmation.

[0057] Integrity checks include verifying required fields such as material code, name, specifications, quantity, and unit, as well as checking the completeness of associated material verification equipment. This integrity verification is based on a knowledge graph. ,in For supplies The actual set of attributes included. The required set of attributes defined for a knowledge graph. When Integrity check passed (default) Otherwise, mark it as awaiting manual review and identification of missing attributes. The confidence level is adjusted based on the integrity score.

[0058] Logical consistency verification, quantity relationship verification, checking whether the quantity ratio of main equipment and supporting equipment is reasonable, technical parameter matching verification, checking whether the technical parameters of related equipment match, such as voltage level, capacity, etc., and constraint condition verification, verifying whether they meet the technical specifications and standards requirements.

[0059] Accuracy verification involves calculating the overall confidence level for each material item. ,in Confidence level for entity identification For semantic mapping confidence, For the confidence level of knowledge reasoning, , , Let be the weighting coefficient, satisfying + + =1; Entries below the confidence threshold are marked as requiring manual review, with the default confidence threshold being 1. .

[0060] Step 90) Generate a power project materials list that conforms to the standard format.

[0061] Data structuring: Organize material information into a standardized data structure Material, which includes materialID (unique identifier), materialCode (material code), materialName (material name), category (material category), specification (specification model), technicalParameters (technical parameter dictionary), quantity (quantity), unit (unit), manufacturer (manufacturer), remarks (remarks), sourceDocuments (list of source documents), and confidence (confidence level).

[0062] Classification and Numbering: Classify materials according to the material classification system and generate material codes. The coding rule is: category code - subcategory code - serial number.

[0063] Formatted output: Supports multiple output formats including Excel, XML, JSON, etc. Excel output includes multiple worksheets, namely, primary equipment list, secondary equipment list, cable material list, auxiliary material list, and summary table.

[0064] Report Generation: Generates a materials inventory compilation guide, including data sources, processing methods, confidence level statistics, and explanations of anomalies; generates a quality assessment report, including accuracy, recall, and completeness indicators. The final output is a standardized materials inventory for the power project. And related reports and documents.

[0065] A smart system for generating power project material inventory lists based on multimodal document parsing and knowledge graph semantic mapping includes:

[0066] The document acquisition module is used to receive and manage power project design documents, and supports batch uploading, format checking and preprocessing.

[0067] The multimodal document parsing module, including a text processing unit, a drawing recognition unit, and a table analysis unit, performs unified parsing of documents of different formats and outputs a set of raw information in a unified format. ;

[0068] The entity recognition module uses the BiLSTM-CRF model, based on the original information set. Perform accurate named entity recognition and attribute extraction, and output a structured entity set. ;

[0069] The knowledge graph construction module starts from entity sets. Extract entities and relationships from it, construct and maintain a knowledge graph in the field of power materials, including ontology design, knowledge extraction, knowledge fusion and knowledge storage sub-modules;

[0070] The relation extraction module extracts entity sets based on a dependency parsing model. Perform entity relationship extraction;

[0071] The semantic mapping module establishes a mapping relationship between document content and material requirements based on a standard key point rule base, including rule matching and confidence calculation sub-modules;

[0072] The knowledge reasoning module uses knowledge graphs for reasoning and information completion to obtain a complete set of material requirements, including rule-based reasoning, graph structure reasoning, and embedded reasoning sub-modules.

[0073] The quality control module performs consistency verification, integrity checks, logical verification, and accuracy assessment on the generated material requirements.

[0074] The inventory generation module generates standardized inventory lists for power projects, supporting multiple output formats;

[0075] The modules communicate through message queues and RESTful APIs, and are deployed using a microservice architecture.

[0076] The beneficial effects of this invention are as follows: This invention enables the mining of unstructured text data, engineering drawing data, and semi-structured tabular data, allowing for the full analysis and application of complex engineering data. Through a two-stage processing flow, firstly, unified parsing and semantic encoding of multimodal documents are performed, followed by precise entity recognition and type labeling. This establishes a clear information flow relationship, with clearly defined responsibilities for each step, facilitating system maintenance and performance optimization. The power project design document knowledge graph can meet the automatic retrieval needs of different users and, as a knowledge-sharing platform, breaks down barriers to the exchange of design data between departments, effectively supporting the development of upper-level data applications and significantly improving the efficiency and quality of material inventory compilation. Attached Figure Description

[0077] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0078] Figure 2 This is a schematic diagram of the multimodal document parsing process in this invention.

[0079] Figure 3 This is a schematic diagram of the named entity recognition process in this invention.

[0080] Figure 4 This is a schematic diagram of the process of constructing a knowledge graph of power materials in this invention.

[0081] Figure 5 This is a schematic diagram of dependency parsing and relation extraction in this invention.

[0082] Figure 6 This is a schematic diagram of the semantic mapping and rule matching process in this invention.

[0083] Figure 7 This is a schematic diagram of the system deployment architecture of the present invention.

[0084] Steps 10) to 90) in the figure are executed sequentially. The multimodal documents include text documents, engineering drawings and tabular data. Named entity recognition adopts the BIOES annotation system, and dependency relationship analysis adopts the maximum spanning tree method. Detailed Implementation

[0085] The technical solution of this application will be described in detail below with reference to specific embodiments. Those skilled in the art should understand that the following embodiments are only used to illustrate this application and should not be regarded as limiting the scope of this application.

[0086] like Figure 1 , Figure 2As shown, this embodiment takes a 110kV substation project as an example to illustrate the specific implementation process of the intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping. The project includes one feasibility study report, one preliminary design specification, 15 construction drawing design documents, and electrical main wiring diagrams. Figure 3 Zhang, floor plan Figure 5 There are 8 equipment list forms, totaling 33 documents.

[0087] Step 10) Obtain the power project design documents. Receive the aforementioned 33 documents through the document management system and identify the format of each document. The feasibility study report is in PDF format, the preliminary design specification is in Word format, the construction drawing design document is in Word format, the main electrical wiring diagram is in CAD DWG format, the floor plan is in CAD DXF format, and the equipment list table is in Excel format. Preprocess the documents, extracting metadata including document title, creation time, author, version number, etc. Establish a document index set. Assign a unique identifier to each document. Document type The document format is identified as pdf, docx, dwg, dxf, or xlsx, and the document path is... Points to file storage location, document metadata Record document attribute information.

[0088] Step 20) Perform unified parsing and processing on documents of different formats. For text documents, use the Apache POI library to read Word documents and the PDFBox library to read PDF documents, extracting text content including chapter titles, paragraph text, table data, etc. For the 1850 characters of Chapter 5, Electrical Design, of the Preliminary Design Specification, segment it into 4 segments using a sliding window method, with a window length of... Tokens, overlap length Each token is used for semantic encoding. A BERT model, pre-trained on a power sector corpus, is employed. This corpus contains 520 feasibility study reports, 280 technical specification documents, 150 standard specification documents, and 100 equipment manuals. The BERT model uses a 12-layer Transformer encoder with a hidden layer dimension of [missing information]. For each text segment, a contextual semantic representation is obtained using the BERT model. .

[0089] For engineering drawings, CAD drawings are converted into PNG images with a resolution of 300 DPI. An improved Faster R-CNN model is used to identify equipment symbols in the PNG images. This model is trained on a dataset of power equipment symbols containing 31 annotated drawings. The dataset includes 16 equipment types such as transformers, circuit breakers, disconnectors, current transformers, and voltage transformers, with a total of 520 annotated targets. The model backbone network uses ResNet-50, and a Feature Pyramid Network (FPN) is added. Feature maps are output at conv2, conv3, conv4, and conv5 layers, constructing a pyramid feature structure from top to bottom. The anchor boxes of the Region Candidate Network (RPN) have scales of {32², 64², 128², 256², 512²} and aspect ratios of {1:2, 1:1, 2:1}. Equipment symbols for 3 main transformers, 6 110kV circuit breakers, 12 110kV disconnectors, 8 35kV circuit breakers, and 20 10kV circuit breakers are detected in the main electrical wiring diagram. PaddleOCR was used to extract the annotation text on the drawings and recognize text information such as equipment number, model, and technical parameters.

[0090] For tabular data, the openpyxl library is used to read the Excel file, iterating through eight worksheets, including lists of main transformers, circuit breakers, disconnectors, instrument transformers, cables, switchgear, protection devices, and secondary equipment. Hough transform is used to detect table lines; the mathematical expression is... First, determine the table structure. Then, parse the cell content and identify the data type. Next, establish a mapping relationship between table columns and material attributes, such as mapping the "Equipment Name" column to the "materialName" attribute, the "Specification" column to the "specification" attribute, the "Quantity" column to the "quantity" attribute, and the "Unit" column to the "unit" attribute.

[0091] The above processing generates the original information set. It contains 328 text information entries, 156 drawing information entries, and 1,245 table information entries, totaling 1,729 information entries. Each information entry includes the content (text content), the semantic feature vector (semantic feature vector), the source information (source information), and the modality type identifier (modality type identifier).

[0092] Step 30) Process the original information set Perform named entity recognition. For example... Figure 3 As shown, from Extract the text content. For text modality, directly extract the content field. For drawing modality, extract the text recognized by OCR. For table modality, extract the cell content. Convert the text sequence into character-level input. For example, for the text "This project requires 3 main transformers of 110 kV, with the specification model being SZ11-50000 / 110", it is converted into the character sequence "该" "工" "程" "需" "要" "1" "1" "0" "k" "V" "主" "变" "压" "器" "3" "台" "," "规" "格" "型" "号" "为" "S" "Z" "1" "1" "-" "5" "0" "0" "0" "0" " / " "1" "1" "0".

[0093] Perform feature encoding, map each character to a 300-dimensional word vector, and the word vectors are pre-trained by Word2Vec on 172 power documents. Integrate BERT semantic features, and for the BERT encoding in perform token-level to character-level alignment, and adopt the splicing integration method .

[0094] Use bidirectional LSTM for encoding. Among them, the forward LSTM processes the sequence from left to right, and the backward LSTM processes the sequence from right to left. The dimension of the LSTM hidden layer . For the above text, the forward LSTM captures the left context information such as "This project requires 110 kV", and the backward LSTM captures the right context information such as "台,规格型号为". Concatenate the forward and backward hidden states to obtain .

[0095] Use the CRF layer for decoding, define the state transition matrix , where 13 is the number of label categories, including labels such as B-MAT, I-MAT, E-MAT, S-MAT, B-SPEC, I-SPEC, E-SPEC, S-SPEC, B-PARAM, I-PARAM, E-PARAM, S-PARAM, O, etc. Among them, B - start, I - inside, O - outside, E - end, S - single word, and the entity types include MAT material name, SPEC specification model, PARAM technical parameter, QUANT quantity, UNIT unit, MANUF manufacturer. Calculate the sequence score . Use the Viterbi algorithm to find the optimal label sequence.

[0096] Perform entity recognition on all the text content in all the original information sets to generate the entity set A total of 318 entities were identified, including 12 entities by material name, 35 entities by specification / model, 26 entities by technical parameter, 112 entities by quantity, 108 entities by unit, and 25 entities by manufacturer. The entity identification precision rate was 93.2%, the recall rate was 91.5%, and the F1 score was 92.3%.

[0097] Step 40) Construct a knowledge graph for the power materials sector. For example... Figure 4 As shown, the ontology model is described using OWL, defining a class hierarchy. The top-level class is ElectricalMaterial, and its first-level subclasses include PrimaryEquipment, SecondaryEquipment, CableMaterial, and AuxiliaryMaterial. PrimaryEquipment contains second-level subclasses such as Transformer, CircuitBreaker, Isolator, CurrentTransformer, and VoltageTransformer. Transformer contains third-level subclasses such as PowerTransformer and DistributionTransformer.

[0098] Define object properties: `hasComponent` indicates a composition relationship, such as `transformer hasComponent high-voltage bushing`, `transformer hasComponent gas relay`, `transformer hasComponent pressure relief valve`, etc. `compatibleWith` indicates a compatibility relationship, such as `transformer compatibleWith high-voltage bushing of a specific specification`. `replacedBy` indicates a substitution relationship, such as `transformer of a certain model replacedBy upgraded model transformer`. `configuredWith` indicates a matching relationship, such as `main transformer configuredWith main transformer protection device`.

[0099] Based on the entity set identified in step 30) Construct entity nodes. For the 12 identified material name entities, create 12 entity nodes. For example, for the entity "main transformer", create a node.

[0100] .

[0101] Relation extraction is performed using an attention-based relation classification model. For entity pairs ("main transformer", "high-voltage bushing"), the context "the main transformer requires 6 high-voltage bushings" is extracted, and a context representation is obtained through BERT encoding to calculate the relation representation. The softmax classifier predicts the relation type as hasComponent. A total of 56 relations were extracted, including 75 hasComponent relations, 32 compatibleWith relations, 16 replacedBy relations, and 53 configuredWith relations.

[0102] To perform entity alignment, calculate the edit distance for the two entities "main transformer" and "main transformer". Maximum length Similarity Calculate semantic similarity, specifically the cosine similarity of the word vectors of two entities. Overall similarity If they are determined to be the same entity, they are merged and aligned.

[0103] Entity disambiguation was performed. The entity "transformer" appears 32 times in the document, with 28 referring to the main transformer, 3 to the distribution transformer, and 1 to the instrumentation transformer. Based on the context information, the disambiguation score was calculated for "3 transformers configured on the 110kV side". The characteristics include the co-occurring "110kV", "side", and "3 units", indicating it is a main transformer. The statement "15 transformers configured in the 10kV distribution room" indicates it is a distribution transformer.

[0104] The knowledge graph is stored using the Neo4j graph database. A node is created using the syntax: `CREATE (n:Transformer {id:'T001',name:'Main Transformer',voltageLevel:'110kV',capacity:'50MVA'})`. A relation is created using `CREATE(t:Transformer)-[:HAS_COMPONENT {quantity:6}]->(b:Bushing)`, representing the transformer's `hasComponent` high-voltage bushing, with a quantity of 6. An index is created using `CREATE INDEX ON :Transformer(id)` to speed up queries. The final knowledge graph contains 326 nodes and 847 relations.

[0105] Step 50) Extract relations and attributes. For example... Figure 5As shown, dependency parsing is performed on the sentence "The 220kV switchyard main transformer expansion project designed by Shenyang Electric Power Survey and Design Institute Co., Ltd. is located north of Beisijiazi Village, Shenbei New District". The MST Piser analyzer is used for the sentence... Construct a dependency tree Characteristic function Including part-of-speech features Distance features Lexical features Calculate the dependency tree score. .

[0106] Dependency analysis was performed using the Eisner algorithm to obtain the dependency tree with the highest score. The results show that "located in" is the core predicate, "engineering" is connected to "located in" via SBV subject-predicate relation, and "Shenbei New District Beisijiazi Village" is connected to "located in" via VOB verb-object relation. "Main transformer expansion" modifies "engineering" via ATT nominative-head relation, and "220kV switchyard" modifies "main transformer expansion" via ATT nominative-head relation. "Design" modifies "engineering" via ATT nominative-head relation, and "company" is connected to "design" via SBV subject-predicate relation.

[0107] Extract entity relation triples based on dependency relationships. For the dependency path (“Project”) - [ATT] -> “Design”) - [SBV] -> (“Company”), extract the relation triple (Project, Designer, Company). For the dependency path (“Project”) - [SBV] -> (“Located at”) - [VOB] -> (“Village”), extract the relation triple (Project, Located at, Village). Identify attribute modification relationships; “220kV Switching Station” modifies “Main Transformer Expansion”, extract the attributes (Main Transformer, Voltage Level, 220kV). Extract numerical attributes; “240MVA” is the capacity attribute, and “4 units” is the quantity attribute.

[0108] Extract relations and attributes from all entities to construct a set of entity-relation triples. A total of 245 ternary transformer sets were sampled. Examples include ternary transformer sets such as (main transformer, capacity, 50MVA), (main transformer, quantity, 4 units), (main transformer, supporting equipment, high-voltage bushing), and (main transformer, manufacturer, Shenyang Transformer Co., Ltd.).

[0109] Step 60) Establish semantic mapping relationships. For example... Figure 6 As shown, a standard key point rule base is constructed. It contains 235 rules. An example rule is shown below: Rule ID is RULE_TR_220KV, name is 220kV main transformer configuration rule, trigger condition Cond is that the document contains "220kV" and "main transformer", matching pattern includes capacity matching regular expression "number + MVA", cooling method matching enumeration value ONAN, ONAF, or ODAF, winding type matching enumeration value dual winding or triple winding, conversion rule Trans includes material type "Transformer", material code "TR-240-capacity", quantity extracted as a number, supporting equipment includes 6 high-voltage bushings, 1 gas relay, 1 pressure relief valve, and confidence level Conf is 0.95.

[0110] The document content is matched using a multi-level matching process. The first level is exact matching: for the phrase "220kV main transformer" in the document, a hash table search is used in the rule base to find a match for the rule RULE_TR_110KV. The second level is fuzzy matching: for the phrase "220kV main transformer" in the document, the edit distance between it and "220kV main transformer" is calculated. Maximum length Similarity The fuzzy matching threshold is not met. The third layer of semantic matching calculates the cosine similarity of word vectors. The semantic match was determined to be successful.

[0111] The final confidence score is calculated by comprehensively considering the matching level and contextual information. Confidence level of the rule itself Matching confidence (Semantic matching), contextual confidence (The context includes key information such as capacity and quantity), final confidence level .

[0112] Semantic mapping was performed on all document content to establish a mapping relationship Φ between document content D and material requirements M. A total of 127 material requirement items were mapped, of which 92 had a confidence level ≥ 0.90, 18 had a confidence level between 0.80 and 0.90, and 17 had a confidence level between 0.70 and 0.80.

[0113] Step 70) Perform knowledge reasoning and information completion. Based on rule-based reasoning, define the reasoning rule: IF(project,requires,Transformer)AND(Transformer,hasVoltageLevel,"220kV")THEN(project,requires,TransformerOil)AND(project,requires,GasRelay). According to this rule, from the known (220kV switchyard main transformer expansion project, requires, main transformer) and (main transformer, voltage level, 220kV), deduce (220kV substation project, requires, transformer oil) and (220kV substation project, requires, gas relay).

[0114] Graph-based reasoning discovers implicit relationships through paths in the knowledge graph, where the paths are represented as follows: Calculate path credibility Example: For the main transformer node, searching the `hasComponent` relationship reveals connections to nodes such as high-voltage bushings, gas relays, pressure relief valves, thermometers, and oil level gauges. For the path (main transformer) - [hasComponent] -> (high-voltage bushing), the path confidence is conf = 0.95. For the path (main transformer) - [hasComponent] -> (gas relay) - [compatibleWith] -> (specific model gas relay), the path confidence is conf = 0.95 × 0.88 = 0.84.

[0115] Embedded reasoning uses the TransE model to learn vector representations of entities and relations. For the entity "Master Transformer", the vector representation is... For the relation "hasComponent", the vector representation is... For the entity "high-voltage bushing", vector representation Training the model enables To find out "What additional equipment is needed for the main transformer", calculate... In the physical space, find the nearest neighbor and locate entities such as gas relays, pressure relief valves, and thermometers.

[0116] Through knowledge-based reasoning, the implicit material requirements were completed. The original semantic mapping yielded 127 material requirements; after completion through knowledge-based reasoning, 139 material requirements were obtained, an increase of 12, resulting in a completion rate of 9.4%. The completed materials mainly include supporting equipment, auxiliary materials, and spare parts.

[0117] Step 80) Perform quality control. Cross-document consistency verification: For the main transformer, the preliminary design specification records 4 units, the equipment list table records 4 units, and 4 units are detected in the main electrical wiring diagram; the three sets of information are completely consistent. For the 220kV circuit breaker, the preliminary design specification records 6 units, the equipment list table records 6 units, but 7 units are detected in the main electrical wiring diagram, indicating an inconsistency. Calculate the consistency metric. ,in , .for and , , Reduced. Marked as a conflict and submitted for manual confirmation. Manual review confirmed that one extra standby circuit breaker was tested in the drawings.

[0118] Completeness check: Required fields were checked. For 139 material requirements, the material code, name, specifications, quantity, and unit fields were checked. 15 records were found to be missing specifications and 8 records were missing quantity information; these were marked as incomplete, and the confidence level was reduced by 0.15. Related materials were checked: For 4 main transformers, 18 high-voltage bushings, 3 gas relays, and 3 pressure relief valves should be provided. Only 2 gas relays were found, 1 was missing, and this was marked as incomplete. Knowledge graph-based calculations were performed. .

[0119] Logical consistency check: The quantity ratio of main equipment to auxiliary equipment is checked. There are 4 main transformers and 18 high-voltage bushings, a ratio of 1:6, which meets the specifications. There are 4 main transformers and 2 pressure relief valves, a ratio of 3:2, which does not meet the specifications; it should be 1:1. This is marked as a logical error. Technical parameter matching is checked: the main transformer voltage level is 220kV, and the rated voltage of the high-voltage bushings is 126kV > 110kV, meeting the requirements. The main transformer capacity is 240MVA, and the current transformer ratio is 1000 / 5A. Calculate the rated current on the primary side. The requirements are met.

[0120] Accuracy verification involves calculating the overall confidence level for each material item. , in Qualify the confidence level for the entity (step 30). Set the confidence level for semantic mapping (step 60). Calculate the confidence level of the knowledge reasoning (step 70). The weighting coefficients and For low confidence items ( ,default Marked as awaiting manual review. Example: For the main transformer entry, entity recognition confidence level. Semantic mapping confidence Confidence of knowledge reasoning Overall confidence level Those marked as high confidence do not require manual review. Of the 139 material requirements, 118 have a comprehensive confidence level ≥ 0.90, 13 have a confidence level between 0.80 and 0.90, and 8 have a confidence level < 0.80, meaning that 5.7% require manual review.

[0121] Step 90) Generate a standardized materials list. Organize the materials information into a standardized data structure. For example, the main transformer entry includes materialID as MAT_220_TR_001, materialCode as TR-220-240, materialName as main transformer, category as primary equipment - transformer, specification as SZ20-240000 / 220, technicalParameters as rated capacity 240MVA, high-voltage side voltage 220kV, low-voltage side voltage 35kV and 10kV, cooling method ONAN, winding type three-winding, quantity as 4, unit as unit, manufacturer as Shenyang Transformer, remarks as supporting equipment including high-voltage bushings, gas relays, pressure relief valves, etc., sourceDocuments as Chapter 5 of the preliminary design specification, equipment list table, main transformer list, electrical main wiring diagram, and confidence as 0.924.

[0122] According to the material classification system, primary equipment includes 4 transformers, 34 circuit breakers, 48 ​​disconnect switches, 72 current transformers, 36 voltage transformers, and 54 surge arresters. Secondary equipment includes 15 sets of protection devices, 28 sets of monitoring and control devices, and 12 sets of communication equipment. Cable materials include 1200 meters of 110kV cable, 2800 meters of 35kV cable, 8500 meters of 10kV cable, and 15000 meters of control cable. Auxiliary materials include grounding materials, hardware, and identification plates. Material codes are generated: main transformer codes are TR-220-240-001 to TR-220-240-004, and 220kV circuit breaker codes are CB-220-001 to CB-220-006.

[0123] Supports multiple output formats. Excel output contains 5 worksheets: a primary equipment list with 158 records, a secondary equipment list with 85 records, a cable material list with 127 records, an auxiliary material list with 256 records, and a summary table containing statistical information for all 689 records. XML output conforms to the standard format for power industry material lists. JSON output facilitates data exchange between systems.

[0124] The process included generating a materials inventory compilation guide, with data sourced from 33 design documents. The processing method employed was multimodal document parsing and knowledge graph semantic mapping, achieving an average confidence level of 0.887. An anomaly was noted: inconsistencies in the number of 220kV circuit breakers between documents required manual verification. A quality assessment report was generated, achieving an accuracy rate of 96.5%, a recall rate of 94.2%, a completeness rate of 92.8%, and an automation rate of 92.6%.

[0125] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any limitation on the present invention. Those skilled in the art should fully understand that modifications to the technical solutions described in the foregoing embodiments or equivalent substitutions for any part or all of the technical features are entirely feasible. Such modifications or substitutions, as long as they do not depart from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. A method for intelligent generation of power project material inventory based on multimodal document parsing and knowledge graph semantic mapping, characterized in that, Includes the following steps: Step 10) Obtain the power project design documents, perform format recognition and preprocessing on the documents, and establish a document index set. ,in As a unique identifier for the document, For document type, Document path For document metadata; Step 20) Perform unified processing on documents of different formats, converting heterogeneous data into a unified intermediate representation format to generate the original information set. Each information entry It includes text content (content), semantic feature vector (semantic_feature), source information (source_info), and modality type identifier (modality_type); among which , This represents the total number of information entries; Step 30) Based on the BiLSTM-CRF machine learning model, process the original information set generated in step 20). Named entity recognition is performed on the data to generate an entity set. ,in The total number of entities identified. For the first The text content of each entity, For the first Type label for each entity, For the first The text position range of an entity, The index of information entries from which this entity originates. ; Step 40) From the entity set Extract entities and relationships to construct a knowledge graph in the field of power materials; Step 50) Based on the dependency parsing model, analyze the entity set. Extract relationships and attributes to construct a set of entity relationship triples; Step 60) Establish a standard key point rule base, and establish a mapping relationship between document content and material requirements through a multi-level matching mechanism; specifically as follows: Building a standard key rule base Size |R| rules, each rule Includes triggering conditions Define the preconditions for rule activation and the matching mode. Define how to identify relevant information and conversion rules. Define how to convert into standard material requirements and confidence levels. Indicates the credibility of the rule and ;in , This represents the total number of rules. A multi-level matching mechanism is adopted. The first level is exact matching, which uses a hash table to store keywords. The second layer is fuzzy matching, which calculates the edit distance. and similarity ,when A successful match is considered to occur at this time, by default. The third layer is semantic matching, which calculates cosine similarity based on word vectors. ,in and They are respectively and The word vector representation, when For semantic matching to be successful, ; Document content is established through a multi-level matching mechanism. With the demand for supplies Mapping relationship between ; The final confidence score is calculated by comprehensively considering the matching level and contextual information. ,in = For rules Self-confidence level; To determine the confidence level, a value of 1.0 is used for exact matching and 0.0 for fuzzy matching. Or semantic matching ; For contextual confidence; Step 70) Based on the knowledge graph of the power materials domain, reasoning and information completion are performed. Rule reasoning, graph structure reasoning and embedding reasoning methods are used to obtain the completed set of material requirements. Step 80) Perform quality control on the generated material requirements, including cross-document consistency verification, integrity check, logical consistency verification and accuracy assessment, and calculate the overall confidence level for each material item; Step 90) Organize the material information into a standardized data structure, classify and number it according to the material classification system, and generate a standardized power project material list.

2. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, Step 20) specifically includes the following steps for uniformly processing documents of different formats: Extract text content from Word and PDF documents, and use a sliding window approach to segment the text into several fragments. The window length is... Tokens, overlap length Each text segment is represented by a token; the BERT model is used to semantically encode each text segment to obtain a contextual semantic representation. ,in For the number of tokens, The hidden layer dimension of the BERT model; The CAD drawings are rasterized and converted into PNG images with a resolution of R=300 DPI. An improved Faster R-CNN object detection model is used to identify device symbols in the PNG images, obtaining the category and bounding box location information of the device symbols. Simultaneously, OCR technology is used to extract the annotation text from the images. The improved Faster R-CNN object detection model includes: Introducing the Feature Pyramid Network (FPN), feature maps are output at the conv2, conv3, conv4, and conv5 layers of ResNet-50. , , , Constructing pyramid features from top to bottom , , , ,in For 2x upsampling, It is a 1×1 convolution; At each pyramid level Independently run regional candidate network (RPN), anchor box design scales={32 2 ,64 2 128 2 256 2 512 2 The aspect ratios are set to {1:2, 1:1, 2:1}. Focal Loss is used to address class imbalance during classification. ,in To predict probabilities, As a balance factor, For focusing parameters; Use ROI Align instead of ROI Pooling for sizes of The candidate bounding boxes are divided into a 7×7 grid. In each grid cell, bilinear interpolation is used to calculate feature values ​​at 4 sampling points, and the ROI features are output as 7×7×256. Multi-task loss function ,in For classifying losses, For bounding box regression loss, This is the balance coefficient; Use the openpyxl library to read an Excel file and detect table lines using the Hough transform. The mathematical expression is: ,in The distance from the origin to the line is denoted as . For the normal vector of the line and The angle between axes; determining the table structure based on detected table lines, identifying the header and data areas; parsing the table structure and cell content, identifying cell data types; establishing a mapping relationship between table columns and material attributes. ,in Name the columns in the table. This refers to the name of the material attribute.

3. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 2, characterized in that, In step 30), the entity set The construction methods include: Step 31) Prepare input data First, from the original information set Extract the text content (content) and semantic feature vector (semantic_feature); Next, the extracted text sequence Convert to character-level input, where For the sequence number The character at position 1 , T This is the total length of the sequence; Then, each character is processed through a feature encoding layer. Mapping to word vectors Among them, word vector dimension The BERT semantic information obtained in step 20) Additional features are concatenated and fused to form the fused feature representation. Or use attention fusion , ,in , , For learnable parameters, Attention weights; Step 32) Represent the fused features obtained in step 31). Input is encoded using BiLSTM; at each time step t, the calculation formula is: in, , and These represent the input gate, forget gate, and output gate at the current moment, respectively. and These represent the memory unit at the current moment and the candidate memory, respectively. The state was hidden in the previous moment; This is the current input feature vector; This is element-wise multiplication; for Activation function, defined as ; , , , This is the corresponding weight matrix; , , , This is the corresponding bias vector; Hidden state of the forward LSTM Hidden state of backward LSTM By splicing them together, we can obtain Hidden layer dimensions ; Step 33) Decode using the CRF model Define label sequence ,in For position The tag, For a set of tags, The number of label categories; the hidden states output by BiLSTM. Mapping to the label space via linear transformation: ,in Indicates the location For tags The launch fraction; Define the state transition matrix ,in The number of label categories, where Indicates from the label Transfer to label The transition score; for the label sequence Calculate sequence score Then, the conditional probability is calculated using softmax. ,in Given the set of all possible tag sequences; the Viterbi algorithm is used to find the optimal tag sequence during decoding. Based on the predicted label sequence y*, entity labels are performed according to the BIOES labeling system. The label format is "BIO tag-entity type", and the entity type is... MAT is the material name, SPEC is the specification, PARAM is the technical parameter, QUANT is the quantity, UNIT is the unit, and MANUF is the manufacturer; based on the predicted label sequence. Combine consecutive BI*-E tags or single S tags into a complete entity and extract the entity text. ,type ,Location and source Construct entity set .

4. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, In step 40), the method for constructing a knowledge graph in the field of power materials includes: The ontology is described using OWL, defining a class hierarchy. The top-level class is ElectricalMaterial, and the first-level subclasses include PrimaryEquipment, SecondaryEquipment, CableMaterial, and AuxiliaryMaterial. Object attributes are defined, including composition relationship (hasComponent), compatibility relationship (compatibleWith), replacement relationship (replacedBy), and matching relationship (configuredWith). Data attributes are defined, including materialCode, materialName, specification, voltageLevel, and capacity. Based on the entity set identified in step 30) Entity nodes are constructed; a relation classification model based on an attention mechanism is used for relation extraction, and relation representations of entity pairs are calculated. ,in Entities Vector representation, entity Vector representation and context The vector representation of the matrix is ​​used to predict the relationship type through a softmax classifier. Extracting entity attribute values ​​from text using sequence labeling methods; Entity alignment uses an entity matching algorithm based on edit distance and semantic similarity. ,in To normalize the edit distance, For cosine similarity based on word vectors, These are the weighting coefficients; Entity disambiguation calculates the disambiguation score based on contextual information. ,in For feature set, For feature weights, It is the characteristic function; The knowledge graph is stored using the Neo4j graph database, containing |V| nodes and |E| relation edges. For a set of entity nodes, Let it be the set of relation edges.

5. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, In step 50), the method for constructing the entity relation triplet set is as follows: The mathematical model for MSTParser is as follows: in, Indicates the score. It is a sentence A dependent tree, It is a word governing words Dependency edges, It is a high-dimensional binary feature function vector. Represents a node and The dependency relationship between them takes the value 0 or 1. It is a feature The weight vector; The Eisner algorithm is used for decoding to obtain the dependency tree with the highest score. ,in For sentences The set of all possible dependency trees; Based on dependency syntax trees, Subject-Verb-Object (SVO), Attribution-Head (ATT), and Verb-Object (VOB) relations are extracted. Attribute modification relations are identified, numerical attributes are extracted, and a set of entity relation triples is constructed. ,in For the head entity, For the relationship, It is a tail entity.

6. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, In step 70), the reasoning method based on knowledge graphs includes: Rule-based reasoning defines a set of reasoning rules. The rule format is Reasoning can be performed using either forward chain reasoning or backward chain reasoning. Graph-based reasoning discovers implicit relationships through paths in the knowledge graph, where the paths are represented as follows: Calculate path credibility Subgraph matching is performed using graph isomorphism algorithms, and community detection is performed using Louvain's algorithm or label propagation algorithm. Embedded reasoning uses the TransE model to learn vector representations of entities and relations, satisfying... loss function ， in For positive sample set, For negative sample set, Let be the distance function. For boundary parameters, Missing relationships are predicted through vector operations.

7. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, In step 80), the quality control method includes: Cross-document consistency verification checks the consistency of information about the same material appearing in different documents and defines a consistency metric. ,in For supplies In the document and The difference measure in the data indicates that when the consistency is below a threshold, it is marked as a conflict and requires manual confirmation. Integrity checks include verifying the required fields such as material code, name, specifications, quantity, and unit; and checking the completeness of associated material verification equipment. This integrity verification is based on a knowledge graph. ,in This refers to the actual set of attributes included. The required set of attributes defined for a knowledge graph; when Integrity checks are performed by default. Otherwise, mark it as awaiting manual review and identification of missing attributes. Adjust the confidence level based on the integrity score; Logical consistency verification checks whether the ratio of main equipment to auxiliary equipment is reasonable, whether the technical parameters of related equipment match, and whether they meet the technical specifications and standards. Accuracy verification involves calculating the overall confidence level for each material item. ,in Confidence level for entity identification For semantic mapping confidence, For the confidence level of knowledge reasoning, , , The weighting coefficients and + + =1; Entries below the confidence threshold are marked as requiring manual review, with the default confidence threshold being 1. .

8. The intelligent generation method for power project material inventory based on multimodal document parsing and knowledge graph semantic mapping according to claim 1, characterized in that, In step 90), the method for generating the standardized materials list includes: The material information is organized into a standardized data structure called Material, which includes a unique identifier (materialID), material code (materialCode), material name (materialName), material category, specification, technical parameters dictionary, quantity, unit, manufacturer, remarks, source documents list, and confidence. The materials are classified according to the material classification system, and material codes are generated. The coding rule is: category code - subcategory code - serial number. It supports multiple output formats including Excel, XML, and JSON; the Excel output contains multiple worksheets, namely, primary equipment list, secondary equipment list, cable material list, auxiliary material list, and summary table; The instructions for generating the material inventory include data sources, processing methods, confidence level statistics, and explanations of abnormal situations. The generated quality assessment report includes accuracy, recall rate, and completeness indicators.

9. A power project material inventory intelligent generation system based on multimodal document parsing and knowledge graph semantic mapping, used to implement the power project material inventory intelligent generation method based on multimodal document parsing and knowledge graph semantic mapping as described in any one of claims 1 to 8, characterized in that, include: The document acquisition module is used to receive and manage power project design documents, and supports batch uploading, format checking and preprocessing. The multimodal document parsing module includes a text processing unit, a drawing recognition unit, and a table analysis unit. It performs unified parsing of documents of different formats and outputs a set of raw information in a unified format. The entity recognition module uses the BiLSTM-CRF model to perform accurate named entity recognition and attribute extraction based on the original information set, and outputs a structured entity set. The knowledge graph construction module extracts entities and relationships from the entity set to build and maintain a knowledge graph in the field of power materials. The relation extraction module extracts entity relations from the entity set based on the dependency parsing model and constructs a set of entity relation triples. The semantic mapping module establishes a mapping relationship between document content and material requirements based on a standard key point rule base; The knowledge reasoning module uses a knowledge graph in the field of power materials to perform reasoning and information completion, resulting in a complete set of material requirements. The quality control module performs consistency verification, integrity checks, logical verification, and accuracy assessment on the generated material requirements. The inventory generation module generates standardized inventory lists for power project materials, supporting multiple output formats.

Images