Automatic positioning indexing system for standard content of power grid

Through the automated positioning and indexing system of grid standard content, the improved BERT network structure and multi-level dynamic memory unit are used to solve the problem of difficulty and inefficiency in finding grid standard content, efficient and accurate retrieval is achieved, and the grid practitioners are improved. Work efficiency and grid safety.

CN119003684APending Publication Date: 2024-11-22ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD +1

Patent Information

Application Number
CN202410986639.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

When handling the content of power grid standards, the prior art has problems of insufficient accuracy and limited applicability, which leads to inefficiency in the search and retrieval of power grid practitioners, and is prone to omissions and errors, affecting the safe operation of the power grid.

Method used

An automated positioning indexing system for grid standard content was designed. Through the combination of document segmentation, feature extraction, database storage, user problem processing and matching search modules, the improved BERT network structure is used to realize efficient automatic positioning of grid standard content. and accurate search. Specifically, it includes document segmentation module, feature extraction module, database module, user problem processing module and matching search module, and uses multi-level dynamic memory units and inter-layer information interaction module to enhance the feature expression capabilities of the model.

Benefits of technology

It significantly improves the work efficiency of power grid practitioners and the reliability of safe operation of power grids, realizes fast and accurate positioning and retrieval of grid standard content, reduces the time and energy of manual search, improves the search accuracy and user experience, and enhances the power grid. security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003684A_ABST
    Figure CN119003684A_ABST
Patent Text Reader

Abstract

The invention discloses an automatic positioning index system for power grid standard content. The automatic positioning index system comprises a document segmentation module, a feature extraction module, a database module, a user problem processing module, a matching retrieval module and a result display module. The document segmentation module is used for segmenting a power grid standard document into a plurality of document blocks; the feature extraction module extracts document block features by using an improved BERT network structure; the database module is used for storing document names, summary, document blocks and features thereof; the user question processing module converts a question input by a user into a searchable query vector; the matching retrieval module performs matching with the document summary and the document block according to the query vector; and the result display module displays the matched document and the related document blocks to the user. According to the system, efficient automatic positioning and accurate retrieval of the power grid standard content are realized, the working efficiency of power grid employees and the safety of power grid operation are remarkably improved, and the system has a wide application prospect.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of document information extraction, and in particular to an automatic positioning indexing system for power grid standard content. Background Art

[0002] In the field of grid standard content retrieval, with the continuous development of power systems and the advancement of smart grids, the number and complexity of grid standards are increasing, and grid practitioners need to frequently consult and apply these standards in their daily work. However, due to the large number and complex content of grid standard documents, manual search and positioning of required standard content is time-consuming and laborious, and prone to omissions and errors. This not only reduces work efficiency, but also may bring potential risks to the safe operation of the power grid. Therefore, how to achieve automatic positioning and efficient retrieval of grid standard content has become an urgent problem to be solved.

[0003] In the field of text content classification and retrieval, a variety of technologies and methods have been proposed. For example, a text content classification method extracts the name features, introduction features and chapter title features of the text, and performs multi-label classification on the text content based on these features and their associations. This method can efficiently and accurately classify text content, but its processing power and accuracy still need to be improved for complex power grid standard documents. Another technology involves an image content-based system and method, which generates image-based content by parsing text content. This method can enhance the visualization of text content, but its practicality and effectiveness in the retrieval application of power grid standards still need to be further verified. In addition, there is an external memory access method for a graph computing system based on index positioning, which divides and sorts graph data, establishes an index, and dynamically selects the optimal loading method to improve data access efficiency. This method has certain advantages in data processing and retrieval efficiency, but its applicability needs further study due to the complexity of power grid standard content.

[0004] Some inventions extract features of the target text content by obtaining the name, introduction and chapter title, and perform multi-label classification on the target text content based on these features and their associations. Although this method can achieve multi-label classification, its classification accuracy and efficiency may be insufficient for complex and highly professional texts such as power grid standards. Some inventions receive text-based content items, determine the content scene based on the content and the associated login page content, and generate corresponding image content items. In the process of text-to-image content conversion, this method may not be able to fully express the professional details of power grid standards, and there is still room for improvement in retrieval efficiency and accuracy. Some inventions divide the graph data into multiple sub-graphs, sort and index them respectively, and improve data access efficiency by selecting the optimal loading method. This method is mainly designed for graph computing systems, and there may be applicability issues when it is directly applied to the retrieval of power grid standard content, and certain improvements and optimizations are needed. Some inventions save files that appear in each version in the same index, and record the version in which the file appears, so as to facilitate the rapid location and update of files and versions. Although it can effectively reduce storage overhead and improve retrieval efficiency, the fine-grained retrieval capability for complex documents such as power grid standards is limited.

[0005] In summary, the existing text classification and retrieval technologies still have problems such as insufficient accuracy and limited applicability when processing power grid standard content. Therefore, designing an automatic positioning indexing system for power grid standard content has important theoretical significance and practical application value. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes an automatic positioning and indexing system for power grid standard content. The system combines modules such as document segmentation, feature extraction, database storage, user question processing, matching retrieval and result display, and uses an improved BERT network structure to achieve efficient automatic positioning and accurate retrieval of power grid standard content, greatly improving the work efficiency of power grid practitioners and the reliability of safe operation of the power grid. Specifically, it includes the following modules:

[0007] T1: Document segmentation module, used to segment the entire document into multiple document blocks;

[0008] T2: Feature extraction module, which uses an improved BERT network structure to extract the features of each document block;

[0009] T3: Database module, establishes a database system to store document names, summaries, document blocks and their characteristics;

[0010] T4: User question processing module, which receives questions input by users and converts them into a searchable format;

[0011] T5: matching retrieval module, matching the user's question with the document overview and document blocks in the database;

[0012] T6: Result display module, which displays the matched documents and their related document blocks to the user.

[0013] As a technical preferred solution of the present invention, the document segmentation in the module T1 specifically adopts a segmentation algorithm based on natural language to segment the entire document into multiple document blocks, including the following steps:

[0014] S1: Text preprocessing: preprocess the input text to remove special characters and HTML tags. The calculation formula is as follows:

[0015] cleaned_text=re.sub(\s+,\t+,input_text)

[0016] Among them, re.sub is used to replace multiple spaces with one space;

[0017] S2: Punctuation recognition, identifying sentence endings in the text, including periods, question marks, and exclamation marks. The calculation formula is:

[0018] sentence_endings={'.','! ','? '}

[0019] potential_boundaries={i|cleaned_text[i]∈sentence_endings}

[0020] Among them, entence_endings represents punctuation recognition, potential_boundaries represents potential sentence boundary recognition, and cleaned_text[i] represents the i-th character in the text;

[0021] S3: Result output: divide the text into independent paragraphs according to the confirmed sentence boundaries. The specific calculation formula is as follows:

[0022] sentences=[cleaned_text[0:i1],...,cleaned_text[i k-1 +1:i k ],cleaned_text[i k +1:]]

[0023] Among them, {i1,i2,...,i k} is a set of indices of confirmed sentence boundaries.

[0024] As a technical preferred solution of the present invention, the BERT network structure is improved in the module T1, and a multi-level dynamic memory unit is introduced in each layer to enhance the model's ability to capture long text dependencies and integrate context information, wherein the calculation formula of the multi-level dynamic memory unit is as follows:

[0025]

[0026] in, is the memory state of the lth layer at the tth time step, is the hidden state of the current time step of the lth layer, and is the weight and bias of the memory unit, σ is the sigmoid activation function, ⊙ represents the element-by-element multiplication, where in the multi-level dynamic memory mechanism, by introducing the memory state The model can retain memory of past time steps in each layer, thus better capturing dependencies in long texts;

[0027] Memory Status The hidden state at the current time step and the memory state of the previous time step Joint decision, weight and bias Used to control the update degree of memory status.

[0028] As a technical preferred solution of the present invention, the module T1 improves the BERT network structure and adds an information interaction module between each layer to integrate feature information of different layers and improve the model's feature expression ability and deep information capture ability. The calculation formula for inter-layer interactive information fusion is as follows:

[0029]

[0030] in, is the hidden state of the next time step of the lth layer, is the hidden state of the current time step of the lth layer, is the weighted accumulation of the hidden states of the previous layers, α (l) , β (l) and are learnable weight parameters.

[0031] As a technical preferred solution of the present invention, the improved BERT network structure described in module T1 includes, from its input end to its output end, an input embedding layer, a position encoding layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and an output layer in sequence; wherein the feature extraction layer is composed of a multi-head self-attention mechanism and a feedforward neural network layer, a dynamic memory unit layer, and an inter-layer information interaction module; wherein each layer of the multi-head self-attention mechanism and the feedforward neural network layer includes a multi-head self-attention sublayer and a feedforward neural network sublayer, and each layer of the dynamic memory unit layer introduces multi-level dynamic memory units in each layer to enhance the model's ability to capture long text dependencies, and each layer of the inter-layer information interaction module is used to integrate feature information of different layers.

[0032] As a technical preferred solution of the present invention, in the improved BERT network structure, embedding_size=768 in the input embedding layer; max_position_embeddings=512 in the position encoding layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the first feature extraction layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the second feature extraction layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the third feature extraction layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the output layer.

[0033] As a technical preferred solution of the present invention, the database module storage format in module T3 specifically includes:

[0034] Document name field: stored in the form of a unique identifier, specifically marked with UUID;

[0035] Summary field: uses text summarization technology to generate a brief description of the document and stores it in JSON format for quick retrieval;

[0036] Document block field: Each document block after the entire document is segmented is stored as an independent record in BLOB format. The specific formula is as follows:

[0037] BLOB = pickle.dumps({

[0038] "block_id":Block_ID,

[0039] "content":Block_Content,

[0040] "metadata":Metadata})

[0041] Among them, Block_ID is the unique identifier of the document block; Block_Content is the actual content of the document block, a byte sequence after text encoding; Metadata includes relevant metadata of the document block, including position and length, and the format is: Metadata = {"position":Position,"length":Length}, where Position represents the position and Length represents the length.

[0042] As a technical preferred solution of the present invention, the matching retrieval module in module T4 specifically includes the following operation steps:

[0043] M1: First matching part: Preliminary matching is performed based on the user's question and the document overview in the database to determine relevant documents. The specific process includes:

[0044] M1-1: Receive the user's input question and use natural language processing technology to convert it into a searchable query vector. The calculation formula is:

[0045] Query_Vector=BERT.encode(User_Query)

[0046] Among them, Query_Vector represents the query vector, and User_Query represents the question input by the user;

[0047] M1-2: Search the database for document summaries similar to the query vector and calculate the similarity using cosine similarity, the formula is:

[0048]

[0049] Where Similarity represents similarity, Summary_Vector represents the feature vector of the document summary, · represents the vector dot product, ||Query_Vector|| and ||Summary_Vector|| represent the modulus lengths of the query vector and the document summary vector, respectively;

[0050] M1-3: Sort the documents according to their similarity scores and select the document with the highest similarity.

[0051] M2: Second matching part: In the documents selected in the first matching part, fine-grained matching is performed with the document blocks according to the user's question to determine the relevant paragraphs. The specific process includes:

[0052] M2-1: Extract all document blocks and their feature vectors from the document with the highest similarity selected in step M1-3. The specific calculation formula is:

[0053] Block_Vectors={BERT.extract_features(Block_Content)}

[0054] Among them, Block_Vectors represents the set of document block feature vectors, and Block_Content represents the content of each document block;

[0055] M2-2: Use the query vector and document block feature vector to calculate similarity;

[0056] M2-3: Sort the document blocks according to the similarity scores and select the document block with the highest similarity for display.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] Improve work efficiency: The system can quickly divide the entire power grid standard document into multiple document blocks, and use the improved BERT network structure to extract the features of each document block, realizing the rapid parsing and feature extraction of the document content. In this way, users can find the required power grid standard content in a short time, greatly reducing the time and effort of manual search and improving work efficiency.

[0059] Improve retrieval accuracy: The system uses multi-level dynamic memory units and inter-level information interaction modules to enhance the model's ability to capture long text dependencies and integrate contextual information. By performing natural language processing on user input questions and matching them with document summaries and document blocks in the database, highly accurate retrieval results can be achieved, effectively avoiding possible omissions and errors in the retrieval process.

[0060] Enhanced user experience: The system converts user input questions into a searchable format through the user question processing module, and combines it with the matching retrieval module to match the user's questions with the document overview and document blocks in the database, and finally displays the matched documents and their related document blocks to the user. In this way, users can intuitively view the relevant document content, which significantly improves user experience and satisfaction.

[0061] Improve grid security: Since the system can quickly and accurately locate and retrieve grid standard content, it helps grid practitioners better understand and apply standards, reduces operational risks caused by untimely or incorrect standard search, and thus improves the safe operation level of the grid.

[0062] Efficient data management: The database module in the system uses UUID to uniquely identify the document name, and stores the document summary in JSON format and the document block in BLOB format, effectively organizing and managing the data of a large number of power grid standard documents. This not only improves the efficiency of data storage and retrieval, but also facilitates data maintenance and updating. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A structural diagram of the automatic positioning indexing system for the standard content of the power grid provided by the present invention;

[0064] Figure 2 An improved BERT network structure of an embodiment provided by the present invention;

[0065] Figure 3 A workflow diagram of document content retrieval according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention is further described below in conjunction with the accompanying drawings and examples. However, the present invention can be implemented in many different ways and should not be construed as being limited to the embodiments shown; on the contrary, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0067] Example 1: Figure 1 As shown, this embodiment introduces the implementation and application of the automatic positioning indexing system of power grid standard content. The overall structure of the system includes a document segmentation module, a feature extraction module, a database module, a user question processing module, a matching retrieval module and a result display module.

[0068] First, the document segmentation module uses a segmentation algorithm based on natural language to segment the entire power grid standard document into multiple document blocks. Specifically, special characters and HTML tags are removed through text preprocessing, and multiple spaces are replaced with one space; then sentence terminators in the text, such as periods, question marks, and exclamation marks, are identified to determine potential sentence boundaries; finally, based on the confirmed sentence boundaries, the text is segmented into independent paragraphs and document blocks.

[0069] Next, the feature extraction module uses the improved BERT network structure to extract the features of each document block. The improved BERT network structure includes an input embedding layer, a position encoding layer, multiple feature extraction layers, and an output layer. The feature extraction layer consists of a multi-head self-attention mechanism, a feedforward neural network, a multi-level dynamic memory unit, and an inter-layer information interaction module. By introducing a multi-level dynamic memory unit and an inter-layer information interaction module, the structure enhances the model's ability to capture long text dependencies and integrate contextual information.

[0070] The database module establishes a database system to store the name, summary, document block and its characteristics of the power grid standard document. The document name field uses UUID for unique identification, the summary field uses text summary technology to generate a brief description of the document and stores it in JSON format, and the document block field stores each document block after segmentation as an independent record in BLOB format.

[0071] The user question processing module receives the questions input by the user and converts them into searchable query vectors using natural language processing technology. The matching retrieval module matches the user's questions with the document summaries and document blocks in the database. The initial matching is performed by cosine similarity calculation, matching the user's questions with the document summaries, and selecting the document with the highest similarity. Subsequently, fine-grained matching is further performed in the documents selected by the initial matching, and the document block with the highest similarity is selected by calculating the similarity between the query vector and the document block feature vector.

[0072] Finally, the result display module displays the matched documents and their related document blocks to the user. The system displays the sorted search results through the user interface, and the user can intuitively view the matched document blocks and their contents.

[0073] Through this embodiment, the automatic positioning indexing system of power grid standard content realizes efficient automatic positioning and accurate retrieval, which significantly improves the work efficiency of power grid practitioners and the safety of power grid operation.

[0074] Example 2: Figure 2 As shown, this embodiment introduces in detail the improved BERT network structure used in the automatic positioning indexing system of power grid standard content and its application in feature extraction. The improved BERT network structure introduces multi-level dynamic memory units and inter-layer information interaction modules in each layer to enhance the model's ability to capture long text dependencies and integrate context information.

[0075] Table 1: Improved BERT network structure

[0076] Network structure parameter Input Embedding Layer embedding_size=768 Position Encoding Layer max_position_embeddings=512 Feature extraction layer Multi-head self-attention mechanism, feedforward neural network, multi-level dynamic memory unit and inter-layer information interaction module

[0077] As shown in Table 1, the specific network structure includes the following parts: Input embedding layer: convert the text content of each document block into a vector representation, and the embedding_size is 768. Position encoding layer: add position information to each word, and the max_position_embeddings is 512. Feature extraction layer: includes a multi-head self-attention mechanism, a feedforward neural network, a multi-level dynamic memory unit, and an inter-layer information interaction module.

[0078] Table 2: Feature extraction network structure

[0079] Network structure parameter The first feature extraction layer hidden_size=768, num_attention_heads=12, intermediate_size=3072 Second feature extraction layer hidden_size=768, num_attention_heads=12, intermediate_size=3072 The third feature extraction layer hidden_size=768, num_attention_heads=12, intermediate_size=3072 Output Layer hidden_size=768

[0080] As shown in Table 2, the specific parameters are as follows: First feature extraction layer: hidden_size = 768, num_attention_heads = 12, intermediate_size = 3072. Second feature extraction layer: hidden_size = 768, num_attention_heads = 12, intermediate_size = 3072. Third feature extraction layer: hidden_size = 768, num_attention_heads = 12, intermediate_size = 3072. Output layer: Generate feature vectors for each document block, hidden_size is 768.

[0081] The improved BERT network structure introduces a multi-level dynamic memory unit in each layer to enhance the model's ability to capture long text dependencies and integrate context information. The calculation formula of the multi-level dynamic memory unit is as follows:

[0082]

[0083] in, is the memory state of the lth layer at the tth time step, is the hidden state of the current time step of the lth layer, and is the weight and bias of the memory unit, σ is the sigmoid activation function, ⊙ represents the element-by-element multiplication, where in the multi-level dynamic memory mechanism, by introducing the memory state The model can retain memory of past time steps in each layer, thus better capturing dependencies in long texts;

[0084] Memory Status The hidden state at the current time step and the memory state of the previous time step Joint decision, weight and bias Used to control the update degree of memory status.

[0085] The improved BERT network structure adds an information interaction module between each layer to integrate the feature information of different layers, improve the model's feature expression ability and deep information capture ability, where the calculation formula for inter-layer interactive information fusion is as follows:

[0086]

[0087] in, is the hidden state of the next time step of the lth layer, is the hidden state of the current time step of the lth layer, is the weighted accumulation of the hidden states of the previous layers, α (l) , β (l) and are learnable weight parameters.

[0088] Example 3: Figure 3 As shown, in order to verify the effectiveness of the improved BERT network structure in power grid standard content retrieval, this embodiment conducts a detailed simulation experiment. The experiment selects a large number of power grid standard documents, uses the improved BERT network structure to extract features, and processes and matches the retrieval questions entered by the user to evaluate the performance of the system. The following are the detailed settings and steps of the experiment.

[0089] The dataset used in the experiment contains 500 grid standard documents, which cover standards and regulations in multiple aspects such as grid operation, maintenance, and safety. The average length of each document is 2,000 words, and the total number of words in the entire dataset exceeds 1 million. These documents are divided into multiple document blocks for feature extraction and retrieval.

[0090] In order to comprehensively evaluate the performance of the system, the experiment used retrieval precision (Precision), recall rate (Recall) and F1 value (F1-Score) as evaluation indicators. These indicators can measure the performance of the system in accurate retrieval, comprehensive retrieval and comprehensive performance respectively. Retrieval precision (Precision): refers to the proportion of actually relevant documents among all retrieved documents. Recall rate (Recall): refers to the proportion of correctly retrieved documents among all actually relevant documents. F1 value (F1-Score): The harmonic mean of precision and recall rate, which is an important indicator for comprehensively evaluating system performance.

[0091] The database module establishes a database system to store the name, summary, document block and characteristics of power grid standard documents. The database system uses MongoDB, and the specific storage format is as follows:

[0092] Document name field: Each document name is uniquely identified using UUID.

[0093] Summary field: Generates a brief description of the document using text summarization technology and stores it in JSON format.

[0094] Document block field: Each document block after segmentation is stored as an independent record in BLOB format, which contains the unique identifier of the document block, the actual content and related metadata.

[0095] Assume that the present invention has a power grid standard document, the content of which is as follows: "Document name: Power grid safety operation procedures Document content: 1. Overview of power grid safety operation procedures 1.1 Importance of power grid safety 1.2 Relevant laws and regulations 2. Operator requirements 2.1 Qualification requirements 2.2 Training requirements 3. Operation process 3.1 Daily operation 3.2 Emergency operation"

[0096] The document segmentation module segments the document into multiple document blocks, an example of which is shown in Table 3 below:

[0097] Table 3: Segmentation results

[0098] content: Overview of Power Grid Safety Operation Procedures content: Training requirements content: Importance of Grid Security content: Operation process content: Relevant laws and regulations content: Daily Operation content: Operator requirements content: Emergency Operation content: Qualifications

[0099] The improved BERT network structure is used to extract features from each document block and generate feature vectors. The segmented document blocks and their features are stored in the database. The storage format of each document block is JSON format. The specific example is as follows:

[0100] {"document_id":"123e4567-e89b-12d3-a456-426614174000",

[0101] "document_name":"Grid Safety Operation Procedures","summary":"This procedure introduces the importance of grid safety operations, relevant laws and regulations, operator requirements and operation procedures.","blocks":[{"block_id":"block-001","content":"Grid Safety Operation Procedures Overview","features":[0.123,0.234,0.345,...,0.456],"metadata":{"position":1,"length":10}},{"block_id":"block-002","content":"The Importance of Grid Safety","features":[0.223,0.334,0.445,...,0.556],"metadata":{"position":2,"length":8}},...]}

[0102] When a user enters a query, the system converts the query into a feature vector and matches it with the document block features in the database. Assuming the user queries "grid operator qualification requirements", the system processes it as follows:

[0103] Receive user input: User query: Grid operator qualification requirements

[0104] Generate query vector: Query vector: [0.213, 0.324, 0.435, ..., 0.546]

[0105] Preliminary matching: The system first performs a preliminary match in the document summary to find the most relevant document. Assume that the matching result is: Matching document: Power grid safety operation procedures. Matching degree: 0.92.

[0106] Fine-grained matching: In the initially matched documents, further matching document blocks is performed. Assume that the matching result is: Matching document block: block-005. Matching degree: 0.87. Content: Qualification requirements.

[0107] The system displays the matching results to the user. The user interface is as follows:

[0108] "Query results:

[0109] Document name: Power Grid Safety Operation Procedures

[0110] Matches document block content: Qualification requirements"

[0111] Through the above experimental steps, this experiment achieved the following results: retrieval accuracy (Precision): 92.5%; recall rate (Recall): 89.7%; F1 value (F1-Score): 91.1%.

[0112] It can be seen from the experimental results that the improved BERT network structure shows high precision and recall rate in power grid standard content retrieval, which can effectively improve the accuracy and comprehensiveness of the retrieval results. The specific analysis is as follows: Retrieval accuracy: The system performs well in high-precision retrieval and can accurately locate the power grid standard content required by users. This is mainly due to the powerful ability of the improved BERT network structure in feature extraction. Through multi-level dynamic memory units and inter-layer information interaction modules, the model enhances the capture of long text dependencies and the fusion of context information. Recall rate: The system also performs well in recall rate, which can cover most of the relevant documents and reduce the possibility of omissions. This is mainly attributed to the system's dual guarantee in preliminary matching and fine-grained matching, which ensures the comprehensive retrieval of relevant documents and document blocks. Comprehensive performance: The F1 value, as the harmonic mean of precision and recall rate, reflects the superiority of the system in comprehensive performance. The experimental results show that the system can provide efficient and accurate power grid standard content retrieval services in practical applications, greatly improving the work efficiency of power grid practitioners and the safety of power grid operation.

[0113] In summary, the simulation experiment of this embodiment verifies the effectiveness and advantages of the improved BERT network structure in the retrieval of power grid standard content. The system not only solves the problems of difficulty and low efficiency in finding power grid standard content in the prior art, but also significantly improves the work efficiency of power grid practitioners and the safety of power grid operation, and has important practical value and broad application prospects.

[0114] Through the above specific cases, we can see how the automatic positioning indexing system of power grid standard content of the present invention can achieve efficient and accurate retrieval of power grid standard documents through document segmentation, feature extraction and feature matching. The system uses an improved BERT network structure and a sophisticated database storage solution to significantly improve the retrieval efficiency and accuracy of power grid standard content, and has a wide range of practical value.

[0115] The above embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. Automated positioning indexing system for power grid standard content, Features: The system includes the following modules: T1: Document segmentation module, used to segment the entire document into multiple document blocks; T2: Feature extraction module, which uses an improved BERT network structure to extract the features of each document block; T3: Database module, establishes a database system to store document names, summaries, document blocks and their characteristics; T4: User question processing module, which receives questions input by users and converts them into a searchable format; T5: matching retrieval module, matching the user's question with the document overview and document blocks in the database; T6: Result display module, which displays the matched documents and their related document blocks to the user.

2. The automatic positioning indexing system for power grid standard content according to claim 1, Features: The document segmentation in module T1 specifically adopts a segmentation algorithm based on natural language to segment the entire document into multiple document blocks, including the following steps: S1: Text preprocessing: preprocess the input text to remove special characters and HTML tags. The calculation formula is as follows: cleaned_text=re.sub(\s+,\t+,input_text) Among them, re.sub is used to replace multiple spaces with one space; S2: Punctuation recognition, identifying sentence endings in the text, including periods, question marks, and exclamation marks. The calculation formula is: sentence_endings={'.','! ','? '} potential_boundaries={i|cleaned_text[i]∈sentence_endings} Among them, entence_endings represents punctuation recognition, potential_boundaries represents potential sentence boundary recognition, and cleaned_text[i] represents the i-th character in the text; S3: Result output: divide the text into independent paragraphs according to the confirmed sentence boundaries. The specific calculation formula is as follows: sentences=[cleaned_text[0:i 1 ],…,cleaned_text[i k-1 +1:i k ],cleaned_text[i k +1:]] Among them, {i 1 ,i 2 ,...,i k } is a set of indices of confirmed sentence boundaries.

3. The automatic positioning indexing system for power grid standard content according to claim 1, Features: In module T1, the BERT network structure is improved, and a multi-level dynamic memory unit is introduced in each layer to enhance the model's ability to capture long text dependencies and integrate contextual information. The calculation formula of the multi-level dynamic memory unit is as follows: in, is the memory state of the lth layer at the tth time step, is the hidden state of the current time step of the lth layer, and is the weight and bias of the memory unit, σ is the sigmoid activation function, ⊙ represents the element-by-element multiplication, where in the multi-level dynamic memory mechanism, by introducing the memory state The model can retain memory of past time steps in each layer, thus better capturing dependencies in long texts; Memory Status The hidden state at the current time step and the memory state of the previous time step Joint decision, weight and bias Used to control the update degree of memory status.

4. The automatic positioning indexing system for power grid standard content according to claim 1, Features: In module T1, the BERT network structure is improved, and an information interaction module is added between each layer to integrate the feature information of different layers, improve the model's feature expression ability and deep information capture ability. The calculation formula for inter-layer interactive information fusion is as follows: in, is the hidden state of the next time step of the lth layer, is the hidden state of the current time step of the lth layer, is the weighted accumulation of the hidden states of the previous layers, α (l) , β (l) and are learnable weight parameters.

5. The automatic positioning indexing system for power grid standard content according to claim 1, Features: The improved BERT network structure in module T1 includes an input embedding layer, a position encoding layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and an output layer from its input end to its output end; the feature extraction layer is composed of a multi-head self-attention mechanism and a feedforward neural network layer, a dynamic memory unit layer, and an inter-layer information interaction module; each layer of the multi-head self-attention mechanism and the feedforward neural network layer includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer, and each layer of the dynamic memory unit layer introduces multi-level dynamic memory units in each layer to enhance the model's ability to capture long text dependencies, and each inter-layer information interaction module is used to integrate feature information of different layers.

6. The automatic positioning indexing system for power grid standard content according to claim 5, Features: In the improved BERT network structure, embedding_size=768 in the input embedding layer; max_position_embeddings=512 in the position encoding layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the first feature extraction layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the second feature extraction layer; hidden_size=768, num_attention_heads=12, intermediate_size=3072 in the third feature extraction layer; hidden_size=768 in the output layer.

7. The automatic positioning indexing system for power grid standard content according to claim 1, Features: The database module storage format in module T3 specifically includes: Document name field: stored in the form of a unique identifier, specifically marked with UUID; Summary field: uses text summarization technology to generate a brief description of the document and stores it in JSON format for quick retrieval; Document block field: Each document block after the entire document is segmented is stored as an independent record in BLOB format. The specific formula is as follows: BLOB = pickle.dumps({ "block_id":Block_ID, "content":Block_Content, "metadata":Metadata}) Among them, Block_ID is the unique identifier of the document block; Block_Content is the actual content of the document block, a byte sequence after text encoding; Metadata includes relevant metadata of the document block, including position and length, and the format is: Metadata = {"position":Position,"length":Length}, where Position represents the position and Length represents the length.

8. The automatic positioning indexing system for power grid standard content according to claim 1, Features: The matching retrieval module in module T4 specifically includes the following operation steps: M1: First matching part: Preliminary matching is performed based on the user's question and the document overview in the database to determine relevant documents. The specific process includes: M1-1: Receive the user's input question and use natural language processing technology to convert it into a searchable query vector. The calculation formula is: Query_Vector=BERT.encode(User_Query) Among them, Query_Vector represents the query vector, and User_Query represents the question input by the user; M1-2: Search the database for document summaries similar to the query vector and calculate the similarity using cosine similarity, the formula is: Where Similarity represents similarity, Summary_Vector represents the feature vector of the document summary, · represents the vector dot product, ||Query_Vector|| and ||Summary_Vector|| represent the modulus lengths of the query vector and the document summary vector, respectively; M1-3: Sort the documents according to their similarity scores and select the document with the highest similarity; M2: Second matching part: In the documents selected in the first matching part, fine-grained matching is performed with the document blocks according to the user's question to determine the relevant paragraphs. The specific process includes: M2-1: Extract all document blocks and their feature vectors from the document with the highest similarity selected in step M1-3. The specific calculation formula is: Block_Vectors={BERT.extract_features(Block_Content)} Among them, Block_Vectors represents the set of document block feature vectors, and Block_Content represents the content of each document block; M2-2: Use the query vector and document block feature vector to calculate similarity; M2-3: Sort the document blocks according to the similarity scores and select the document block with the highest similarity for display.

Citation Information

Patent Citations

  • Long document retrieval method based on memory network semantic fusion

    CN113962228A

  • Intelligent question-answering system and method suitable for vertical field and application of intelligent question-answering system and method

    CN116805001A

  • Power file question and answer type intelligent retrieval method and system

    CN117171333A

Cited By

  • Bridge steel structure technical file intelligent retrieval method and device based on embedded model

    CN119759848A

  • Apparatus, method, and program for constructing a database.

    JP7927135B1