Automatic extraction and query system for power grid standard knowledge

Abstract

The invention discloses an automatic extraction and query system for power grid standard knowledge, and the system comprises the following modules: a knowledge segmentation module which is used for segmenting the power grid standard knowledge into a plurality of knowledge blocks; the entity extraction module is used for extracting entities contained in the knowledge blocks and entity information of the entities; the feature extraction module is used for extracting features of the document blocks and the entity information by using an LLM-based Encoder network structure; the database module is used for storing knowledge contents and features thereof and providing retrieval capability based on vector similarity; the user question processing module is used for preprocessing questions of the user so as to carry out matching retrieval; the matching retrieval module is used for carrying out matching according to the query content and the knowledge content, and a plurality of matching paths exist; and the reordering module is used for reordering the multiple paths of matching results to obtain a final ranking, and the ranking is a query result. According to the invention, efficient extraction and retrieval of power grid standard knowledge are realized.

Description

Technical Field

[0001] This invention relates to the field of document information extraction technology, and in particular to an automatic extraction and query system for power grid standard knowledge. Background Technology

[0002] With the continuous development of the power industry, the management and utilization of power grid standard knowledge has become increasingly important. Power grid standard knowledge includes, but is not limited to, power system design specifications, operating procedures, and maintenance standards. This knowledge is scattered across various documents and data sources, such as standard manuals, technical reports, academic papers, and databases. Therefore, how to effectively manage and retrieve this scattered knowledge has become a pressing issue for the power industry. In power grid operation and management, the timely acquisition and correct application of standard knowledge are crucial for ensuring the safe and stable operation of the power grid. However, traditional manual processing methods are not only inefficient but also prone to omissions and errors. With the development of artificial intelligence and big data technologies, automated and intelligent power grid standard knowledge management systems have emerged and are gradually becoming a hot topic in industry research and application.

[0003] Currently, several technical solutions exist in the fields of information retrieval and knowledge management. Some solutions utilize pre-trained word vector models to convert query text into vector representations and employ attention mechanisms to weight these vectors, extracting important feature vectors. Then, they match these vectors with attribute path vectors in a pre-built knowledge graph to output the most relevant search suggestions. This method excels in improving matching accuracy but suffers from high computational complexity when handling large-scale data. Other systems use a pre-defined enterprise information classification system to collect and preprocess structured and unstructured data from both internal and external sources, forming a standardized enterprise knowledge graph. Utilizing low-dimensional embeddings of concepts and relationships, they calculate the semantic similarity between optimized concept and relationship vectors to achieve efficient information management and retrieval. However, this method has limitations in handling the specificities of power grid standard knowledge. Still other systems combine patient medical records with knowledge graphs of specific diseases, using text retrieval and self-attention models to retrieve relevant cases and treatment plans from medical databases. These systems can interpret the relevance of search results to patient medical records, providing doctors with auxiliary diagnostic support. This method is widely used in the medical field, but when applied directly to power grid standard knowledge management, adjustments and optimizations to the knowledge graph and retrieval model may be necessary.

[0004] Existing power grid standard knowledge management systems suffer from the following main drawbacks: 1. Scattered information and low retrieval efficiency: Power grid standard knowledge is distributed across different documents and databases, making it difficult for existing systems to achieve efficient and unified retrieval. 2. High manual processing costs: Manual extraction and organization of power grid standard knowledge requires significant manpower and time, resulting in low automation. 3. Insufficient similarity calculation: Existing technologies need to improve the accuracy and efficiency of calculating the similarity between queries and knowledge content. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies, this invention proposes an automatic extraction and retrieval system for power grid standard knowledge. Utilizing natural language processing and machine learning techniques, it achieves efficient extraction and retrieval of power grid standard knowledge. The system of this invention can significantly improve the management efficiency and query accuracy of power grid standard knowledge and has broad application prospects.

[0006] The technical solution adopted by this invention to solve its technical problem is: an automatic extraction and query system for power grid standard knowledge, comprising the following modules: An automatic extraction and query system for power grid standard knowledge includes the following modules: The knowledge segmentation module is used to segment power grid standard knowledge into multiple knowledge blocks; The entity extraction module is used to extract entities and their information contained in knowledge blocks. The feature extraction module uses an LLM-based encoder network structure to extract features from document blocks and entity information; The database module is used to store knowledge content and its features, and provides retrieval capabilities based on vector similarity. The user issue processing module is used to preprocess user issues for matching and retrieval. The matching and retrieval module is used to match query content with knowledge content, and there are multiple matching paths; The reordering module is used to reorder the matching results from multiple paths to obtain the final ranking, which is the query result. The knowledge segmentation module, entity extraction module, feature extraction module, and database module are connected in sequence; the entity extraction module and feature extraction module are respectively connected to the user question processing module; the user question processing module and database module are respectively connected to the matching retrieval module; and the matching retrieval module is connected to the reordering module.

[0007] Furthermore, the knowledge segmentation module divides the power grid standard knowledge into multiple knowledge blocks using a natural language-based segmentation algorithm, which includes the following steps: S1: Knowledge cleaning, preprocessing the input power grid standard knowledge text to remove special characters; S2: Recursive segmentation algorithm, the specific operation is as follows: S2-1: Set the maximum slice length Overlap length with sliding window ; S2-2: Regarding the length of knowledge Make a judgment if Then the entire knowledge is a slice; if Then proceed to step S2-3; S2-3: Use newline characters as paragraph separators to divide knowledge into paragraphs, specifically, the length of each paragraph... Make a judgment if Then the entire paragraph is a slice; if Then proceed to step S2-4; S2-4: Use Chinese and English periods, exclamation marks, and question marks as sentence separators. Divide knowledge according to sentence structure and specify the length of each sentence. Make a judgment if Then the entire statement is a slice; if Then proceed to step S2-5; S2-5: Based on the statement Perform fixed-length segmentation, with each segment having a gap between it. The overlap; S3: Semantic merging, which calculates the similarity of adjacent fragments and merges similar fragments into one fragment. Feature extraction of fragment content is accomplished using a vector model.

[0008] Furthermore, in step S3, the vector model is trained through contrastive learning using the Loss function, which is designed as follows: ; in, Represents the query vector; Indicates a positive sample; Represents a logarithmic function; Represents an exponential function; Represents the query vector and positive sample vector dot product between; It is a temperature parameter used to control the smoothness of the distribution; Represents the query vector With all negative sample vectors The sum of the dot product exponents between them.

[0009] Furthermore, the entity extraction module specifically employs a bidirectional long short-term memory network and a conditional random field model. The loss function of this conditional random field model consists of the score of the true path and the total score of all paths, as shown in the following formula: ; ; ; ; in, This represents the sum of the scores of the actual label sequence; Represents the logarithm of the sum of fractions of all possible label sequences; The emission score representing the true path is determined by the elements in the emission matrix. Calculated; This represents the feature at position i and the actual label. The corresponding score; The transition score, representing the true path, is determined by the elements in the transition matrix. Calculated; This represents the transition score of the label between the i-th position and the (i+1)-th position; Represents the total score of all possible paths, where Let be the score of the i-th possible path, and e be the exponent.

[0010] Furthermore, the input to the feature extraction module is the knowledge block text output by the knowledge segmentation module and the entity information output by the entity extraction module, and the output is a feature vector for vector retrieval. The operation method is as follows: the knowledge block text and entity information are concatenated and input into a pre-trained Transformer-based LLM Encoder. The contextual semantic features are extracted through a multi-layer self-attention mechanism, pooled using pooling or a specified token representation, and then reduced and normalized by a fully connected network to generate a fixed-length semantic feature vector for storage by the database module and use by the matching retrieval module.

[0011] Furthermore, the database module is a database module with vector storage and retrieval capabilities. When the number of vectors exceeds a threshold, a vector index is constructed, specifically including the following steps: P1: Calculate the cosine similarity among all vectors using the following formula: ; in, and They are two vectors, Represents the dot product. and These represent the magnitudes of the vectors, respectively. P2: Use a clustering algorithm to cluster all vectors. Take all vectors as input and use the K-means clustering algorithm. Initially, set the number of cluster centers k. Calculate the cosine similarity between each vector and the cluster center, and assign the vector to the cluster center with the highest similarity. Update the position of each cluster center so that it is equal to the average of all vectors in the cluster. Repeat the above process until the position of the cluster center no longer changes. P3: Multiple cluster centers are obtained through clustering algorithms, and each cluster center represents a cluster; P4: Build an index within each cluster and sort the vectors within each cluster; use inverted index technology to index and store the vectors within the cluster; establish a fast retrieval mechanism for vectors within the cluster.

[0012] Furthermore, the user problem handling module preprocesses the user's problem, specifically including the following steps: D1: Decompose user questions into multiple sub-questions that may contain multiple semantic information; D2: Extract entities from the decomposed subproblems using the same entity extraction method as in step S3, and identify and extract all entities in the problem. D3: Extract features from the user problem and its sub-problems. Using the same vector model as in step S3, extract features from the user problem, all sub-problems, and extracted entity information to obtain multiple feature vectors.

[0013] Furthermore, the matching and retrieval module matches the query content with the knowledge content by using multiple feature vectors and the basic retrieval capabilities provided by the database module. Specifically, it includes the following steps: obtaining the feature vector of the user's question; using the index in the vector library to match and retrieve the feature vector; and finding the top k results that are most similar to the feature vector of the user's question.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention improves retrieval efficiency: By employing natural language processing and machine learning techniques, and utilizing pre-trained word vector models and attention mechanisms, this invention automatically extracts and retrieves power grid standard knowledge, greatly improving retrieval speed and efficiency. Compared to manual processing and traditional information retrieval methods, this invention can provide more accurate retrieval results in a shorter time.

[0015] 2. This invention reduces manual processing costs: By automating knowledge extraction and retrieval, this invention reduces reliance on manual labor and lowers the cost of manual processing. The system can automatically extract power grid standard knowledge from multiple data sources and perform standardization processing, reducing the need for manual intervention.

[0016] 3. This invention improves retrieval accuracy: This invention utilizes an attention mechanism to weight the query text vector and performs similarity matching through path vectors in a knowledge graph, thereby improving the accuracy of retrieval results. Compared to traditional keyword matching methods, this invention performs better when handling complex queries. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0018] Figure 1 This is a flowchart of an automatic extraction and query system for power grid standard knowledge according to an embodiment of the present invention.

[0019] Figure 2 A flowchart illustrating the segmentation process of an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1: As Figure 1As shown, this embodiment introduces a simulation of an automatic extraction and query system for power grid standard knowledge. To verify the effectiveness of the automatic extraction and query system for power grid standard knowledge of the present invention, the system was simulated and tested. The test data included power grid standard knowledge texts from multiple data sources, covering content such as power system design specifications, operating procedures, and maintenance standards.

[0024] Power grid standard knowledge was extracted from multiple publicly available power system documents and standards, including: power system design manuals; power operation procedures; and power maintenance standards. 5000 power grid standard knowledge entries were selected as a test sample, with data from the following sources: Document A: Power Design Code (2000 entries); Document B: Power Operation Procedures (1500 entries); Document C: Power Maintenance Standards (1500 entries).

[0025] The automatic extraction and query system for power grid standard knowledge includes the following modules: T1: Knowledge segmentation module, used to segment power grid standard knowledge into multiple knowledge blocks; T2: Entity extraction module, used to extract entities and their information contained in knowledge blocks; T3: Feature extraction module, which uses an LLM-based encoder network structure to extract features of document blocks and entity information; T4: Database module, used to store knowledge content and its features, and provides retrieval capabilities based on vector similarity; T5: User Question Processing Module, used to perform entity extraction, feature extraction and other operations on user questions for matching and retrieval; T6: Matching and retrieval module, used to match query content with knowledge content, with multiple matching paths; T7: Reordering module, used to reorder the matching results from multiple paths to obtain the final ranking, which is the query result.

[0026] The knowledge segmentation module, entity extraction module, feature extraction module, and database module are connected in sequence; the entity extraction module and feature extraction module are respectively connected to the user question processing module; the user question processing module and database module are respectively connected to the matching retrieval module; and the matching retrieval module is connected to the re-ranking module.

[0027] like Figure 2 As shown, the specific implementation includes the following steps: T1: Knowledge segmentation module, used to segment power grid standard knowledge into multiple knowledge blocks.

[0028] Knowledge cleaning: Preprocessing the input knowledge text to remove special characters, HTML tags, and irrelevant information.

[0029] Example text: Standard number: DL / T 520-2018 Substation Design Specification... Processed text: Standard number: DL / T 520-2018 Substation Design Specification.

[0030] Recursive segmentation algorithm: Parameter settings: Maximum slice length (max_length): 500 characters; Sliding window overlap length (overlap): 50 characters.

[0031] The cutting steps are as follows: Judge the length of knowledge: If len(text) <= max_length: then the entire text is a slice; else: Continue splitting.

[0032] Divide into paragraphs: Paragraph length check: if len(paragraph) <= max_length: then the entire paragraph is considered a slice.

[0033] Further subdivide the sentences within the paragraph, using periods, exclamation marks, and question marks as units: Statement length check: if len(sentence) <= max_length: then the entire statement is a slice; Fixed-length splitting: The statement is split into fixed lengths according to max_length, and the overlapping parts are preserved.

[0034] Semantic merging: Similarity is calculated for adjacent content segments, and similar segments are merged. The similarity calculation formula is as follows: ; in, and They are two vectors, Represents the dot product. and These represent the magnitudes of the vectors.

[0035] T2: Entity Extraction Module, used to extract entities and their information contained in knowledge blocks.

[0036] Named Entity Recognition (NER): Entity recognition based on the BiLSTM+CRF model.

[0037] Model Training: The model was trained using a publicly available named entity recognition dataset in the power industry. The model's loss function consists of the score of the true path and the total score of all paths, as shown in the following formula: ; ; ; ; in, This represents the sum of the scores of the actual label sequence; Represents the logarithm of the sum of fractions of all possible label sequences; The emission score representing the true path is determined by the elements in the emission matrix. Calculated; This represents the feature at position i and the actual label. The corresponding score; The transition score, representing the true path, is determined by the elements in the transition matrix. Calculated; This represents the transition score of the label between the i-th position and the (i+1)-th position; Represents the total score of all possible paths; where, Let be the score of the i-th possible path, and e be the exponent.

[0038] Example: Standard number: DL / T 520-2018 -> Entity: Standard number.

[0039] T3: Feature extraction module, which uses an LLM-based encoder network structure to extract features from document blocks and entity information.

[0040] Feature Extraction: Features of fragmented content and entity information are extracted using a pre-trained LLM Encoder network. Vector Model: Feature extraction is performed using the same vector model as the knowledge segmentation module, converting text and entity information into vector representations.

[0041] T4: Database module, used to store knowledge content and its features, and provides retrieval capabilities based on vector similarity.

[0042] Vector storage and retrieval: Establish a vector library for storing and retrieving knowledge vectors.

[0043] Clustering Algorithm: When the number of vectors exceeds 10,000, the K-means clustering algorithm is used to cluster the vectors: calculate the similarity between vectors to form initial cluster centers; iteratively update the cluster centers so that each center represents a cluster.

[0044] T5: User Question Processing Module, used to perform entity extraction, feature extraction and other operations on user questions for matching and retrieval.

[0045] Problem preprocessing: Decompose and extract the entities from the user's problem.

[0046] Decomposition steps: Decompose the user's problem into multiple sub-problems, perform entity recognition on each sub-problem, and extract key entity information.

[0047] Feature extraction: Using the same vector model as the segmented content, the user question and sub-questions are converted into vector representations.

[0048] T6: Matching and retrieval module, used to match query content with knowledge content, with multiple matching paths.

[0049] Feature vector matching: Retrieves the top k results that are most similar to the feature vector of the user's question from the vector library.

[0050] Similarity calculation: The cosine similarity is used to calculate the similarity between feature vectors.

[0051] T7: Reordering module, used to reorder the matching results from multiple paths to obtain the final ranking, which is the query result.

[0052] Result reordering: The search results are reordered using the Reverse Flow Ranking (RRF) algorithm.

[0053] ; Calculate the RRF score for each search result. Sort the search results according to the RRF scores to obtain the final list of search results.

[0054] Through the above steps, the system of the present invention performs as follows on 5000 test samples: Search accuracy: 89%.

[0055] Search speed: The average time per query is 0.8 seconds.

[0056] User satisfaction: 92% (obtained through user feedback questionnaire survey).

[0057] The automatic extraction and query system for power grid standard knowledge of this invention achieves automated management and efficient retrieval of power grid standard knowledge through efficient knowledge segmentation, entity extraction, and feature matching. In practical applications, the system demonstrates high accuracy and efficiency, significantly improving the utilization rate and management level of power grid standard knowledge.

[0058] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An automatic extraction and query system for power grid standard knowledge, characterized in that, Includes the following modules: The knowledge segmentation module is used to segment power grid standard knowledge into multiple knowledge blocks; The entity extraction module is used to extract entities and their information contained in knowledge blocks. The feature extraction module uses an LLM-based encoder network structure to extract features from document blocks and entity information; The database module is used to store knowledge content and its features, and provides retrieval capabilities based on vector similarity. The user issue processing module is used to preprocess user issues for matching and retrieval. The matching and retrieval module is used to match query content with knowledge content, and there are multiple matching paths; The reordering module is used to reorder the matching results from multiple paths to obtain the final ranking, which is the query result. The knowledge segmentation module, entity extraction module, feature extraction module, and database module are connected in sequence; the entity extraction module and feature extraction module are respectively connected to the user question processing module; the user question processing module and database module are respectively connected to the matching retrieval module; and the matching retrieval module is connected to the reordering module.

2. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that, The knowledge segmentation module divides power grid standard knowledge into multiple knowledge blocks, specifically employing a natural language-based segmentation algorithm. Includes the following steps: S1: Knowledge cleaning, preprocessing the input power grid standard knowledge text to remove special characters; S2: Recursive segmentation algorithm, the specific operation is as follows: S2-1: Set the maximum slice length Overlap length with sliding window ; S2-2: Regarding the length of knowledge Make a judgment if Then the entire knowledge is a slice; if Then proceed to step S2-3; S2-3: Use newline characters as paragraph separators to divide knowledge into paragraphs, specifically, the length of each paragraph... Make a judgment if Then the entire paragraph is a slice; if Then proceed to step S2-4; S2-4: Use Chinese and English periods, exclamation marks, and question marks as sentence separators. Divide knowledge according to sentence structure and specify the length of each sentence. Make a judgment if Then the entire statement is a slice; if Then proceed to step S2-5; S2-5: Based on the statement Perform fixed-length segmentation, with each segment having a gap between it. The overlap; S3: Semantic merging, which calculates the similarity of adjacent fragments and merges similar fragments into one fragment. Feature extraction of fragment content is accomplished using a vector model.

3. The automatic extraction and query system for power grid standard knowledge according to claim 2, characterized in that, In step S3, the vector model is trained by contrastive learning using Loss. The Loss function used for Loss training is designed as follows: ； in, Represents the query vector; Indicates a positive sample; Represents a logarithmic function; Represents an exponential function; Represents the query vector and positive sample vector dot product between; It is a temperature parameter used to control the smoothness of the distribution; Represents the query vector With all negative sample vectors The sum of the dot product exponents between them.

4. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that, The entity extraction module specifically employs a bidirectional long short-term memory network and a conditional random field model. The loss function of this conditional random field model consists of the score of the true path and the total score of all paths, as shown in the following formula: ； ； ； ； in, This represents the sum of the scores of the actual label sequence; Represents the logarithm of the sum of fractions of all possible label sequences; The emission score representing the true path is determined by the elements in the emission matrix. Calculated; This represents the feature at position i and the actual label. The corresponding score; The transition score, representing the true path, is determined by the elements in the transition matrix. Calculated; This represents the transition score of the label between the i-th position and the (i+1)-th position; Represents the total score of all possible paths, where Let be the score of the i-th possible path, and e be the exponent.

5. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that: The input to the feature extraction module is the knowledge block text output by the knowledge segmentation module and the entity information output by the entity extraction module. The output is a feature vector for vector retrieval. The operation method is as follows: the knowledge block text and entity information are concatenated and input into a pre-trained Transformer-based LLM Encoder. The contextual semantic features are extracted through a multi-layer self-attention mechanism, pooled using pooling or a specified token representation, and then reduced and normalized by a fully connected network to generate a fixed-length semantic feature vector for storage by the database module and use by the matching retrieval module.

6. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that: The database module is a database module with vector storage and retrieval capabilities. When the number of vectors exceeds a threshold, a vector index is constructed, specifically including the following steps: P1: Calculate the cosine similarity among all vectors using the following formula: ； in, and They are two vectors, Represents the dot product. and These represent the magnitudes of the vectors, respectively. P2: Use a clustering algorithm to cluster all vectors. Take all vectors as input and use the K-means clustering algorithm. Initially, set the number of cluster centers k. Calculate the cosine similarity between each vector and the cluster center, and assign the vector to the cluster center with the highest similarity. Update the position of each cluster center so that it is equal to the average of all vectors in the cluster. Repeat the above process until the position of the cluster center no longer changes. P3: Multiple cluster centers are obtained through clustering algorithms, and each cluster center represents a cluster; P4: Build an index within each cluster and sort the vectors within each cluster; use inverted index technology to index and store the vectors within the cluster; establish a fast retrieval mechanism for vectors within the cluster.

7. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that: The user problem handling module preprocesses user problems, specifically including the following steps: D1: Decompose user questions into multiple sub-questions that may contain multiple semantic information; D2: Extract entities from the decomposed subproblems using the same entity extraction method as in step S3, and identify and extract all entities in the problem. D3: Extract features from the user problem and its sub-problems. Using the same vector model as in step S3, extract features from the user problem, all sub-problems, and extracted entity information to obtain multiple feature vectors.

8. The automatic extraction and query system for power grid standard knowledge according to claim 1, characterized in that, The matching and retrieval module matches the query content with the knowledge content. It uses multiple feature vectors and the basic retrieval capabilities provided by the database module to perform the retrieval. Specifically, it includes the following steps: obtaining the feature vector of the user's question; using the index in the vector library to match and retrieve the feature vector; and finding the top k results that are most similar to the feature vector of the user's question.

Images