Knowledge extraction algorithm suitable for knowledge graph in intelligent operation and maintenance field

Through the relationship extraction method of entity recognition model and manual rules based on BERT-BiGRU-CRF, combined with the Protégé ontology construction tool, the problem of accurate extraction and data fusion of knowledge graphs in intelligent operation and maintenance is solved, efficient and accurate knowledge graph construction is achieved, and the efficiency and accuracy of intelligent operation and maintenance are improved.

CN120542534AInactive Publication Date: 2025-08-26SUZHOU JILIANKE IOT TECH CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202510526854.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-25
Publication Date
2025-08-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing knowledge graph construction methods are difficult to accurately extract key entities in the field of intelligent operation and maintenance, and the relationship extraction is inaccurate, there are fuzzy and duplicate data, which cannot meet the requirements of efficient real-timeness, resulting in missing information or incorrect correlation, making it difficult to integrate multi-source heterogeneous data.

Method used

The entity recognition model based on BERT-BiGRU-CRF is used to extract key entities and synonyms, and triple information is constructed based on the relationship extraction method of manual rules, and the ontology model is built using the Protégé ontology construction tool to perform fusion and disambiguation of fuzzy and duplicate data.

Benefits of technology

It improves the accuracy of key entity extraction, enhances the logic and integrity of the knowledge graph, effectively handles fuzzy and duplicate data, improves the efficiency and accuracy of intelligent operation and maintenance, and provides solid knowledge support for fault diagnosis and predictive maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120542534A_ABST
    Figure CN120542534A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of knowledge graph application, and discloses a knowledge extraction algorithm suitable for a knowledge graph in the intelligent operation and maintenance field, and the algorithm comprises the steps: obtaining historical record data in an operation and maintenance process as original data; carrying out key entity extraction on the original data by adopting an entity recognition model based on BERT-BiGRU-CRF, and carrying out synonym processing to carry out data enhancement to obtain a target entity; connecting the key entities to form triple information by adopting a relationship extraction method based on an artificial rule; establishing an operation and maintenance knowledge graph ontology model by utilizing a Protege ontology construction tool; and carrying out fusion disambiguation on fuzzy data and repeated data in the knowledge graph ontology model to obtain a target knowledge graph model. According to the method, key entities are extracted, synonym processing is carried out, data are enhanced, an ontology model is built, fuzzy and repeated data are fused and disambiguated, solid knowledge support is provided for fault diagnosis and predictive maintenance in intelligent operation and maintenance, and the efficiency and accuracy of intelligent operation and maintenance are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph application technology, and in particular to a knowledge extraction algorithm suitable for knowledge graphs in the field of intelligent operation and maintenance. Background Art

[0002] In the field of intelligent operations and maintenance, with the rapid development of information technology, the amount of data generated by various systems and devices has exploded. Traditional operations and maintenance methods are no longer able to meet the needs of efficient and accurate operations and maintenance. Knowledge graphs, as a technology that effectively organizes and represents knowledge, can structure the entities and relationships in operations and maintenance data, providing powerful knowledge support for intelligent operations and maintenance, and facilitating key tasks such as fault diagnosis and predictive maintenance.

[0003] However, existing knowledge graph construction methods face numerous limitations when applied to intelligent operations and maintenance (IOM). They struggle to accurately extract key entities and lack effective data augmentation, leading to information omissions. For relationship extraction, existing technologies rely on complex models with low accuracy, crude rule-making that fails to fully capture inter-entity relationships, and the resulting ontology models often contain ambiguous and duplicated data, making it difficult to construct high-quality knowledge graphs. This ultimately results in insufficient integration capabilities for complex and ever-changing O&M data, making it difficult to integrate heterogeneous data from multiple sources. The constructed knowledge graphs often suffer from missing information or incorrect associations. Furthermore, the construction process is inefficient, failing to meet the real-time requirements of intelligent O&M. Summary of the Invention

[0004] The present invention provides a knowledge extraction algorithm suitable for knowledge graphs in the field of intelligent operation and maintenance. It adopts an entity recognition model based on BERT-BiGRU-CRF to extract key entities, perform synonym processing and enhance data, constructs triple information through a relationship extraction method based on artificial rules, builds an ontology model using the Protégé ontology construction tool, and then fuses and disambiguates fuzzy and duplicate data. It provides solid knowledge support for key tasks in intelligent operation and maintenance, such as fault diagnosis and predictive maintenance, and greatly improves the efficiency and accuracy of intelligent operation and maintenance.

[0005] The present invention provides a knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance, including:

[0006] Obtain historical record data during the operation and maintenance process as raw data; wherein the historical record data includes fault phenomenon, fault cause, treatment measures, fault time, and maintenance records;

[0007] An entity recognition model based on BERT-BiGRU-CRF is used to extract key entities from the original data, and synonym processing is performed on the key entities to perform data enhancement to obtain target entities;

[0008] Adopting the relationship extraction method based on artificial rules, key entities are connected to form triple information by defining multiple relationships;

[0009] Use the Protégé ontology building tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model;

[0010] The fuzzy data and duplicate data in the knowledge graph ontology model are fused and disambiguated to obtain a target knowledge graph model; wherein the fusion disambiguation includes entity disambiguation and coreference resolution.

[0011] Furthermore, the step of extracting key entities from the original data using a BERT-BiGRU-CRF-based entity recognition model and performing synonym processing on the key entities for data enhancement to obtain target entities includes:

[0012] Extracting keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features;

[0013] The keywords are dynamically transformed into word vectors using a BERT model, and the feature sequences of the original data are extracted using a BiGRU-CRF model to obtain key entities;

[0014] Synonym processing is performed on the key entities to perform data enhancement.

[0015] Furthermore, the step of extracting keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features includes:

[0016] For the original data text t i Perform sentence segmentation, i.e. t i ={U1,U2,…,U n}, using the Chinese word segmentation tool Jieba to analyze the sentence U i ∈t i Perform word segmentation and remove stop words by loading the stop word list, retaining the words that meet the conditions after segmentation, that is, U i =u m1 ,u m2 ,....u mi , where u ji ∈U i is the candidate keyword;

[0017] Construct a candidate keyword graph G = (V, E), where V is the set of vertices in the graph, E is the set of edges in the graph, E∈V×V, for each u j,i ∈U i ,u j,i+1 ∈U i ,have <uj,i ,u j,i+1 >∈E, using the co-occurrence relationship between words to build a keyword graph;

[0018] Perform text semantic representation and similarity calculation, and integrate word and text semantic information and word order information into the TextRank algorithm to construct the transition probability matrix and restart probability matrix between words. The formula for calculating the similarity between candidate words and text is:

[0019]

[0020] Among them, v j The vector of the jth word is represented by t, the vector of the text, and n is the dimension of the vector space, n = 512. The cosine similarity between the word and the text is calculated using the formula. The difference between the candidate word and the text is measured by the size of the cosine value of the vector space. The closer the cosine value is to 1, the higher the similarity between the two vectors, that is, the closer the semantic similarity between the candidate word and the text is.

[0021] Score the candidate keywords according to the transition probability matrix, restart probability matrix and PageRank score recursive formula, and select the top N words with the highest scores as keywords according to the scores; specifically, the process includes:

[0022] According to the TextRank iterative calculation process scoring method, the transition probability and restart probability of the TextRank algorithm are modified, and then the PageRank random walk technology is used to score each word node. The calculation formula is as follows:

[0023]

[0024] Among them, TR(u i ) represents node u i That is, the score of word i, the value of which depends on the restart probability r(u i ), the transition probability matrix value Damping coefficient β and TR(u j ), where the damping coefficient β adjusts the problem of node out-degree forming a closed loop during the iteration process, thereby allocating the proportion between the transfer probability and the restart probability, β∈[0,1], usually β=0.85; u j refers to all word nodes u in the word graph i All word nodes connected; e(u j ,u i ) is the node u j and u i The edge weight between them, out(u j ) are all connected nodes u iThe sum of the edge weights of is the element of the transfer probability matrix. The larger the value is, the higher the u i The higher the score, the better; r(u i ) is the restart probability value, and the calculation formula is:

[0025] u j The out-degree out(u j ) is calculated as follows:

[0026] out(u j )=uf:uj→uf e(u j ,u f )

[0027] r(u i ) is calculated as follows:

[0028]

[0029] By calculating the initial weights as above, a new restart probability value is obtained;

[0030] Calculate the edge weight e(u j ,u i ), first calculate the similarity between words and documents, and then calculate the edge weight:

[0031] e(u j ,u i )=s i m(u j ,t)+s i m(u i ,t)

[0032] In the above formula, s i m(u j ,t) is the node u j The similarity between s and text t i m(u i ,t) is the node u i and the similarity between text t, according to e(u j ,u i ) is node u j and u i The edge weights between them, while considering the word node u j Its connected node u i and the similarity relationship between the text, so when the node u j and u i The higher the sum of similarities with document t, that is, the higher the semantic relationship between the word and the document, the greater the proportion of the word in the document;

[0033] The initial value and edge weight are calculated, and then it is iterated to calculate the score of each candidate word. Finally, the words with the highest scores are selected as candidate keywords, that is, the words with the highest semantic similarity to the text.

[0034] Furthermore, the step of converting the keywords into dynamic word vectors based on the BERT model and extracting the feature sequence of the original data based on the BiGRU-CRF model to obtain key entities includes:

[0035] The BERT model uses a bidirectional Transformer encoding structure. It first calculates the correlation between each word and all other words, and then uses these correlations to adjust the importance of each word to generate new semantic feature expressions. The Transformer model uses a multi-head attention mode to increase the representation subspace of the attention mechanism. It also uses a positional embedding method, and its formula is:

[0036]

[0037] Among them, pos is the position index of the word, i is the i-th word, d model The input model has a variable length, PE(pos,2i) is the position vector of the word with even bit index, and PE(pos,2i+1) is the position vector of the word with odd bit index;

[0038] The BiGRU model is used to consider the previous and next text information at the same time to extract bidirectional semantic features, and the CRF model is used to fully consider the connection relationship between entities to obtain the global optimal sequence and obtain key entities.

[0039] Furthermore, the step of performing synonym processing on the key entities to perform data enhancement includes:

[0040] Create a similar entity set list list1 based on the key entity, which covers the synonyms of the key entity. After random processing, form a list of abbreviations and expanded words list2. Randomly extract matching processed words from list1 and list2 to perform cosine similarity calculation:

[0041]

[0042] Among them, S cos is the similarity value between entities; S1 is the synonym entity of the original text; S2 is the abbreviation or expanded word description that matches it; when S cos When the value tends to be between the set threshold and 1, it means that the two texts belong to the same entity. They meet the common entity description in the operation and maintenance history data text and can be included in the similar entity set for generating new text. cosIf the value is not between the set threshold and 1, the key entity is deleted.

[0043] Furthermore, the step of fusing and disambiguating the fuzzy data and duplicate data in the knowledge graph ontology model to obtain a target knowledge graph model includes:

[0044] Entity disambiguation is used to distinguish entities with the same name but different meanings. Coreference resolution is used to identify situations where different expressions refer to the same entity. Cosine similarity and Jaccard similarity are used for fusion disambiguation. The calculation formula is:

[0045]

[0046] Where s1 and s2 are two different sentences or texts, representing the key entities whose similarity is to be calculated. A(s1) and A(s2) are the results of converting the text into sets, representing the word segmentation or splitting of each sentence into word sets.

[0047] The cosine similarity threshold is set to be greater than 0.8, and the Jaccard correlation coefficient threshold is set to be less than 0.7 to judge the similarities and differences between different expressions. When the nodes have high semantic similarity and low word overlap, such nodes are identified as independent nodes. For nodes with high semantic similarity and high word overlap, they are merged into the same node to reduce redundancy in the graph.

[0048] Furthermore, after the steps of fusing and disambiguating the fuzzy data and duplicate data in the knowledge graph ontology model, including entity disambiguation and coreference resolution, the following steps are included:

[0049] Based on the natural language questions input by the user, the corresponding answers are matched from the target knowledge graph ontology model to achieve intelligent question answering; specifically including:

[0050] After the user enters a question, the sentence is segmented using the TF-IDF model, all words are extracted and their word frequencies are calculated to form a word frequency vector. The word frequency vector is used to calculate the cosine similarity between words, and the corresponding query statement is generated based on the extracted keywords and the query results are fed back to the user.

[0051] The present invention also provides a knowledge extraction device suitable for knowledge graphs in the field of intelligent operation and maintenance, comprising:

[0052] An acquisition module is used to acquire historical record data during the operation and maintenance process as raw data; wherein the historical record data includes fault phenomenon, fault cause, treatment measures, fault time, and maintenance records;

[0053] An extraction module is used to extract key entities from the original data using a BERT-BiGRU-CRF-based entity recognition model, and perform synonym processing on the key entities to perform data enhancement to obtain target entities;

[0054] The connection module is used to adopt a relationship extraction method based on artificial rules to connect key entities to form triple information by defining multiple relationships;

[0055] A construction module, used to use the Protégé ontology construction tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model;

[0056] A fusion disambiguation module is used to fuse and disambiguate the fuzzy data and duplicate data in the knowledge graph ontology model to obtain a target knowledge graph model; wherein, the fusion disambiguation includes entity disambiguation and coreference resolution.

[0057] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0058] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0059] The beneficial effects of the present invention are:

[0060] The present invention uses operation and maintenance history record data as the original data source, providing a comprehensive and real data foundation for subsequent analysis. The entity recognition model based on BERT-BiGRU-CRF is adopted to greatly improve the accuracy of key entity extraction. At the same time, data enhancement is performed through synonym processing, which enriches the entity information of the knowledge graph and effectively avoids information omissions. The relationship extraction method based on artificial rules, with the help of a variety of precisely defined relationships, accurately connects key entities to form triple information, thereby enhancing the logic and integrity of the knowledge graph relationship network. The Protégé ontology building tool is used to build an ontology model, making the relationship construction between target entities more standardized and efficient. Finally, the knowledge graph ontology model is fused and disambiguated to effectively process fuzzy data and duplicate data, greatly improving the quality of the target knowledge graph model. Overall, the present invention realizes the efficient conversion from raw operation and maintenance data to high-quality knowledge graphs, provides solid knowledge support for key tasks such as fault diagnosis and predictive maintenance in intelligent operation and maintenance, and greatly improves the efficiency and accuracy of intelligent operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of a method flow according to an embodiment of the present invention.

[0062] Figure 2 FIG. 1 is a schematic diagram of the device structure according to an embodiment of the present invention.

[0063] Figure 3 Schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0064] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0065] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] like Figure 1 As shown, the present invention provides a knowledge extraction algorithm suitable for the knowledge graph in the field of intelligent operation and maintenance, including:

[0067] S1. Obtain historical data from the operation and maintenance process as raw data. The historical data includes, but is not limited to, a list with field headers representing fault symptoms, fault causes, treatment measures, fault time, and maintenance records. The historical data record may be a text file or a table.

[0068] S2. Use the BERT-BiGRU-CRF-based entity recognition model to extract key entities from the original data, and perform synonym processing on the key entities to perform data enhancement to obtain target entities.

[0069] Step S2 specifically includes:

[0070] S201 : extracting keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features.

[0071] (1) Text preprocessing: When historical data is recorded as a table, the text in the table is extracted to form a text, and the text is processed to remove stop words to screen out candidate keywords that meet the conditions.

[0072] For the original data text t i Perform sentence segmentation, i.e. t i ={U1,U2,…,U n}, using the Chinese word segmentation tool Jieba to analyze the sentence U i ∈t i Perform word segmentation and remove stop words by loading the stop word list, retaining the words that meet the conditions after segmentation, that is, U i =u m1 ,u m2 ,…u mi , where u ji ∈U iis the candidate keyword;

[0073] (2) Construct the candidate keyword graph G = (V, E), where V is the set of vertices in the graph, E is the set of edges in the graph, E∈V×V, for each u j,i ∈U i ,u j,i+1 ∈U i ,have j,i ,u j,i+1 >∈E, using the co-occurrence relationship between words to build a keyword graph.

[0074] (3) Perform text semantic representation and similarity calculation, and integrate word and text semantic information and word order information into the TextRank algorithm to construct the transition probability matrix and restart probability matrix between words.

[0075] Text semantic representation is the process of using deep learning technology to represent text data into vectors that can express the semantics of the text. The characteristics of words can be effectively expressed through feature vectors, and the semantic information and word order information between texts and words can be reflected using feature vectors. By representing the document and each word in the document as a word vector, and combining the word order information obtained from the word vector, the semantic information of the text can be fully mined.

[0076] In order to vectorize text and word data, combined with word embedding technology, two public language model tools, Doc2vec model and Word2vec model, are used to vectorize text and candidate words respectively. Among them, the Word2vec model adopts the Skip-gram model during word vector training. First, the words that meet the conditions in the preprocessing stage are mapped to a low-dimensional vector space through the Word2vec model. Combined with experimental experience, the number of iterations is 1000 times, and the dimension of the word vector is set to 512 dimensions. Secondly, the Doc2vec model is used to map the text into a semantic vector representation. The DM model is used to ensure that the dimensions of the document vector and the word vector are the same, which is conducive to similarity calculation. The vector space dimension of the document is also set to 512. Through the above processing, words and documents are represented as vectors with the same dimension. Finally, the cosine similarity between the candidate word and the text is calculated:

[0077]

[0078] Among them, v j It is represented as the vector of the jth word, t represents the vector of the text, and n represents the dimension of the vector space, n = 512. The formula is used to calculate the cosine similarity between the word and the text. The difference between the candidate word and the text is measured by the size of the cosine value of the vector space. The closer the cosine value is to 1, the higher the similarity between the two vectors, that is, the closer the semantic similarity between the candidate word and the text is. ​

[0079] (4) Score the candidate keywords according to the transition probability matrix, restart probability matrix and PageRank score recursive formula, and select the top N words with the highest scores as keywords according to the scores; specifically including:

[0080] To achieve higher keyword extraction accuracy, based on the idea that words that are closer to the document's semantics are more likely to be the document's core keywords, we integrated text semantic information and word order information into the TextRank algorithm framework. Based on the TextRank iterative calculation process scoring method, we modified the transition probability and restart probability of the TextRank algorithm, and then used the PageRank random walk technology to score each word node. The calculation formula is as follows:

[0081]

[0082] Among them, TR(u i ) represents node u i That is, the score of word i, the value of which depends on the restart probability r(u i ), the transition probability matrix value Damping coefficient β and TR(u j ), where the damping coefficient β adjusts the problem of node out-degree forming a closed loop during the iteration process, thereby allocating the proportion between the transfer probability and the restart probability, β∈[0,1], usually β=0.85; u j refers to all word nodes u in the word graph i All word nodes connected; e(u j ,u i ) is the node u j and u i The edge weight between them, out(u j ) are all connected nodes u i The sum of the edge weights is calculated as follows; is the element of the transfer probability matrix. The larger the value is, the higher the u i The higher the score, the better; r(u i ) is the restart probability value, and the calculation formula is as follows:

[0083] u j The out-degree out(u j ) is calculated as follows:

[0084] out(u j )=uf:u j →ufe(u j ,u f )

[0085] r(u i) is calculated as follows:

[0086]

[0087] By calculating the initial weights as above, a new restart probability value is obtained;

[0088] Calculate the edge weight e(u j ,u i ), first use the similarity calculation formula to calculate the similarity between words and documents, and then use the following formula to calculate the edge weight:

[0089] e(u j ,u i )=s i m(u j ,t)+s i m(u i ,t)

[0090] In the above formula, s i m(u j ,t) is the node u j The similarity between s and text t i m(u i ,t) is the node u i and the similarity between text t, according to e(u j ,u i ) is node u j and u i The edge weights between them, while considering the word node u j Its connected node u i and the similarity relationship between the text, so when the node u j and u i The higher the sum of similarities with document t, that is, the higher the semantic relationship between the word and the document, the greater the proportion of the word in the document;

[0091] According to r(u i ) and e(u j ,u i ) formula, calculate the initial value and edge weight, and then iterate, using TR(u i ) formula is used to calculate the score of each candidate word, and finally the words with the highest scores are sorted according to the scores, and the top N words with the highest scores are selected as candidate keywords, that is, the words with the highest semantic similarity to the text.

[0092] S202: Dynamically transform the keywords into word vectors using a BERT model, and extract feature sequences from the original data using a BiGRU-CRF model to obtain key entities.

[0093] BERT is a deep bidirectional language representation model based on the Transformer model. It can be pre-trained based on context, thereby improving language comprehension capabilities. The BERT model's pre-training process consists of two tasks: a masked language model and next sentence prediction. The masked language model is a commonly used pre-training model. It randomly masks some words in the text (typically 15% by default) and then uses the Transformer model to predict these masked words. This approach helps the BERT model incorporate context, leading to a better understanding of the text. In the task of entity recognition in text, the BERT model needs to consider not only the contextual information between words but also the relationship between sentences. Therefore, a next sentence prediction task is constructed. This task is a binary classification task. Specifically, it randomly selects corresponding faulty sentences a and b from the text, with 50% of the faulty sentences being correct and the other 50% being randomly selected faulty sentences. By extracting semantic features from the two sentences, the BERT model can predict whether they are adjacent sentences.

[0094] The BERT model uses a bidirectional Transformer encoding structure composed of multiple encoding units. Each encoding unit includes a self-attention mechanism that is used to find and weight relevant information in the input sequence. The self-attention mechanism helps the model capture long-range dependencies in the input sequence, thereby improving the model's language comprehension capabilities. In the BERT model, each encoding unit uses a multi-head attention mechanism to simultaneously consider different semantic features, further improving model performance. In addition to using Transformer encoding units, the BERT model also employs pre-training and fine-tuning to optimize model performance.

[0095] First, the correlation between each word and all other words is calculated, and then these correlations are used to adjust the importance of each word to generate new semantic feature expressions; the Transformer model adopts a multi-head attention mode to increase the representation subspace of the attention mechanism;

[0096] When processing fault text, words at different positions in the sentence have different meanings. In order to introduce the position information of words in the sentence into the Transformer model, the position embedding method is used. The formula is:

[0097]

[0098] Among them, pos is the position index of the word, i is the i-th word, d model The input model has a variable length, PE(pos,2i) is the position vector of the word with even bit index, and PE(pos,2i+1) is the position vector of the word with odd bit index;

[0099] The BiGRU model is used to consider the previous and next text information at the same time to extract bidirectional semantic features, and the CRF model is used to fully consider the connection relationship between entities to obtain the global optimal sequence and obtain key entities.

[0100] The Bidirectional Gated Recurrent Unit (BiGRU) network consists of a forward and backward GRU network, which can simultaneously consider the preceding and following text information to extract bidirectional semantic features. The BiGRU output is the result of the combined action of the two GRU networks. Because the BiGRU cannot fully consider the dependencies between fault entity labels, it is combined with a CRF to fully consider the connectivity between fault entities, thereby obtaining a globally optimal sequence and improving entity recognition accuracy.

[0101] CRF ensures the logical consistency of entities. CRF can take into account the global information of the sequence. In the named entity recognition task, independently predicting each entity may lead to inconsistent output. For example, certain tag orders (such as the start and end positions of entity tags) have logical dependencies, and independent predictions may violate these order constraints. To address this problem, conditional random fields ensure the logical consistency of predictions by building a global dependency model between tag sequences.

[0102] S203: Perform synonym processing on the key entities to perform data enhancement.

[0103] According to the actual application specifications of the industry, we collect different descriptions of various faults to build a similar entity set, which is used to expand the original text. At the same time, we generate abbreviations and expansion words by random deletion and random insertion. Then, we use cosine similarity to measure the vector space value between texts and select entity words that meet the set threshold for data expansion. Specifically:

[0104] Create a similar entity set list list1 based on the key entity, which covers the synonyms of the key entity. After random processing, form a list of abbreviations and expanded words list2. Randomly extract matching processed words from list1 and list2 to perform cosine similarity calculation:

[0105]

[0106] Among them, S cos is the similarity value between entities; S1 is the synonym entity of the original text; S2 is the abbreviation or expanded word description that matches it; when S cos When the value tends to be between the set threshold and 1, it means that the two texts belong to the same entity. They meet the common entity description in the operation and maintenance history data text and can be included in the similar entity set for generating new text. cosIf the value is not between the set threshold and 1, the key entity is deleted.

[0107] S3. Adopt the relationship extraction method based on artificial rules to connect key entities to form triple information by defining multiple relationships.

[0108] By observing the structural characteristics of the data and combining it with the extracted entity information, we employ a rule-based relationship extraction method, primarily through the definition of relationship sets. Three types of relationships are manually defined, connecting entities to form triples of information. For example, the relationships defined are: fault information generated by the fault location, the cause of the fault information, and the corresponding solution to the fault cause.

[0109] S4. Use the Protégé ontology building tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model.

[0110] S5. Fusion disambiguation is performed on the fuzzy data and duplicate data in the knowledge graph ontology model to obtain a target knowledge graph model; wherein, the fusion disambiguation includes entity disambiguation and coreference resolution.

[0111] To effectively improve the quality of the knowledge graph, knowledge fusion technology is used to fuse and disambiguate fuzzy and duplicate data in the text. This mainly includes entity disambiguation and coreference resolution. Entity disambiguation is used to distinguish entities with the same name but different meanings, while coreference resolution is used to identify different expressions that refer to the same entity.

[0112] Each knowledge node has been converted into a vector representation through the BERT model. Then, the similarity between two nodes is calculated using a combination of cosine similarity and Jaccard correlation coefficient.

[0113] Cosine similarity is used to measure the semantic similarity of nodes and is suitable for longer sentences or descriptions. The calculation formula is:

[0114]

[0115] Jaccard similarity is based on word overlap and is more suitable for identifying situations where there is a lot of word overlap but the semantics may be different (such as "motor output is high" and "motor output is low"), thereby effectively distinguishing faults with different meanings but similar expressions. The calculation method is as follows:

[0116]

[0117] Where s1 and s2 are two different sentences or texts, representing the key entities whose similarity is to be calculated. A(s1) and A(s2) are the results of converting the text into sets, representing the word segmentation or splitting of each sentence into word sets.

[0118] Cosine similarity and Jaccard correlation coefficient can be used to classify and filter nodes. Specifically, the cosine similarity threshold is set to greater than 0.8, and the Jaccard correlation coefficient threshold is set to less than 0.7 to determine the similarities and differences between different expressions. When nodes have high semantic similarity and low word overlap, such nodes are considered independent nodes. Nodes with high semantic similarity and high word overlap are merged into the same node to reduce redundancy in the graph.

[0119] During coreference resolution, for cases where different expressions refer to the same fault cause or phenomenon (for example, "insufficient motor output" and "low output power"), these nodes can be merged into a single, canonical knowledge node. For similar nodes, the most commonly used canonical names are retained, while other expressions are saved as synonyms or labels. This allows these different expressions to point to the same node during graph queries, ensuring the consistency and queryability of the knowledge graph.

[0120] S6. Matching the corresponding answer from the target knowledge graph ontology model based on the natural language question input by the user to achieve intelligent question answering; specifically including:

[0121] In question-answer matching, the TF-IDF model is used to extract keywords, increase the weight of words with important meanings, and use the cosine similarity algorithm to perform fuzzy matching of questions. The cosine similarity calculation formula is as follows:

[0122]

[0123] Among them, A and B are two n-dimensional vectors, A i 、B i are the components in the vector, and their value range is [-1,1], where -1 means completely dissimilar and 1 means completely similar.

[0124] After the user enters a question, the sentence is segmented using the TF-IDF model, all words are extracted and their word frequencies are calculated to form a word frequency vector. The word frequency vector is used to calculate the cosine similarity between words, and the corresponding query statement is generated based on the extracted keywords and the query results are fed back to the user.

[0125] The present invention uses operation and maintenance history record data as the original data source, providing a comprehensive and real data foundation for subsequent analysis. The entity recognition model based on BERT-BiGRU-CRF is adopted to greatly improve the accuracy of key entity extraction. At the same time, data enhancement is performed through synonym processing, which enriches the entity information of the knowledge graph and effectively avoids information omissions. The relationship extraction method based on artificial rules, with the help of a variety of precisely defined relationships, accurately connects key entities to form triple information, thereby enhancing the logic and integrity of the knowledge graph relationship network. The Protégé ontology building tool is used to build an ontology model, making the relationship construction between target entities more standardized and efficient. Finally, the knowledge graph ontology model is fused and disambiguated to effectively process fuzzy data and duplicate data, greatly improving the quality of the target knowledge graph model. Overall, the present invention realizes the efficient conversion from raw operation and maintenance data to high-quality knowledge graphs, provides solid knowledge support for key tasks such as fault diagnosis and predictive maintenance in intelligent operation and maintenance, and greatly improves the efficiency and accuracy of intelligent operation and maintenance.

[0126] like Figure 2 As shown, the present invention provides a knowledge extraction device suitable for knowledge graphs in the field of intelligent operation and maintenance, comprising:

[0127] Acquisition module 1 is used to acquire historical record data during the operation and maintenance process as raw data; wherein the historical record data includes fault phenomenon, fault cause, treatment measures, fault time, and maintenance records;

[0128] Extraction module 2, configured to extract key entities from the raw data using a BERT-BiGRU-CRF-based entity recognition model, and perform synonym processing on the key entities for data enhancement to obtain target entities;

[0129] Connection module 3, used to adopt the relationship extraction method based on artificial rules, connect key entities by defining multiple relationships to form triple information;

[0130] Construction module 4 is used to use the Protégé ontology construction tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model;

[0131] The fusion disambiguation module 5 is used to perform fusion disambiguation on the fuzzy data and duplicate data in the knowledge graph ontology model to obtain the target knowledge graph model; wherein, the fusion disambiguation includes entity disambiguation and coreference resolution.

[0132] In one embodiment, the extraction module 2 includes:

[0133] An extraction unit, configured to extract keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features;

[0134] A conversion unit, configured to perform dynamic word vector conversion on the keywords using a BERT model, and extract feature sequences from the original data using a BiGRU-CRF model to obtain key entities;

[0135] The data enhancement unit is used to perform synonym processing on the key entities to perform data enhancement.

[0136] In one embodiment, the extraction unit comprises:

[0137] For the original data text t i Perform sentence segmentation, i.e. t i ={U1,U2,…,U n}, using the Chinese word segmentation tool Jieba to analyze the sentence U i ∈t i Perform word segmentation and remove stop words by loading the stop word list, retaining the words that meet the conditions after segmentation, that is, U i =u m1 ,u m2 ,....u mi , where u ji ∈U i is the candidate keyword;

[0138] Construct a candidate keyword graph G = (V, E), where V is the set of vertices in the graph, E is the set of edges in the graph, E∈V×V, for each u j,i ∈U i ,u j,i+1 ∈U i ,have j,i ,u j, i +1 >∈E, using the co-occurrence relationship between words to build a keyword graph;

[0139] Perform text semantic representation and similarity calculation, and integrate word and text semantic information and word order information into the TextRank algorithm to construct the transition probability matrix and restart probability matrix between words. The formula for calculating the similarity between candidate words and text is:

[0140]

[0141] Among them, v j The vector of the jth word is represented by t, the vector of the text, and n is the dimension of the vector space, n = 512. The cosine similarity between the word and the text is calculated using the formula. The difference between the candidate word and the text is measured by the size of the cosine value of the vector space. The closer the cosine value is to 1, the higher the similarity between the two vectors, that is, the closer the semantic similarity between the candidate word and the text is. ​

[0142] Score the candidate keywords according to the transition probability matrix, restart probability matrix and PageRank score recursive formula, and select the top N words with the highest scores as keywords according to the scores; specifically, the process includes:

[0143] According to the TextRank iterative calculation process scoring method, the transition probability and restart probability of the TextRank algorithm are modified, and then the PageRank random walk technology is used to score each word node. The calculation formula is as follows:

[0144]

[0145] Among them, TR(u i ) represents node u i That is, the score of word i, the value of which depends on the restart probability r(u i ), the transition probability matrix value Damping coefficient β and TR(u j ), where the damping coefficient β adjusts the problem of node out-degree forming a closed loop during the iteration process, thereby allocating the proportion between the transfer probability and the restart probability, β∈[0,1], usually β=0.85; u j refers to all word nodes u in the word graph i All word nodes connected; e(u j ,u i ) is the node u j and u i The edge weight between them, out(u j ) are all connected nodes u i The sum of the edge weights of is the element of the transfer probability matrix. The larger the value is, the higher the u i The higher the score, the better; r(u i ) is the restart probability value, and the calculation formula is:

[0146] u j The out-degree out(u j ) is calculated as follows:

[0147] out(u j )=uf:u j →uf e(u j ,u f )

[0148] r(u i ) is calculated as follows:

[0149]

[0150] By calculating the initial weights as above, a new restart probability value is obtained;

[0151] Calculate the edge weight e(u j ,u i ), first calculate the similarity between words and documents, and then calculate the edge weight:

[0152] e(u j ,u i )=s i m(u j ,t)+s i m(u i ,t)

[0153] In the above formula, s i m(u j ,t) is the node u j The similarity between s and text t i m(u i ,t) is the node u i and the similarity between text t, according to e(u j ,u i ) is node u j and u i The edge weights between them, while considering the word node u j Its connected node u i and the similarity relationship between the text, so when the node u j and u i The higher the sum of similarities with document t, that is, the higher the semantic relationship between the word and the document, the greater the proportion of the word in the document;

[0154] The initial value and edge weight are calculated, and then it is iterated to calculate the score of each candidate word. Finally, the words with the highest scores are selected as candidate keywords, that is, the words with the highest semantic similarity to the text.

[0155] In one embodiment, the conversion unit comprises:

[0156] The BERT model uses a bidirectional Transformer encoding structure. It first calculates the correlation between each word and all other words, and then uses these correlations to adjust the importance of each word to generate new semantic feature expressions. The Transformer model uses a multi-head attention mode to increase the representation subspace of the attention mechanism. It also uses a positional embedding method, and its formula is:

[0157]

[0158] Among them, pos is the position index of the word, i is the i-th word, d modelThe input model has a variable length, PE(pos,2i) is the position vector of the word with even bit index, and PE(pos,2i+1) is the position vector of the word with odd bit index;

[0159] The BiGRU model is used to consider the previous and next text information at the same time to extract bidirectional semantic features, and the CRF model is used to fully consider the connection relationship between entities to obtain the global optimal sequence and obtain key entities.

[0160] In one embodiment, the data enhancement unit includes:

[0161] Create a similar entity set list list1 based on the key entity, which covers the synonyms of the key entity. After random processing, form a list of abbreviations and expanded words list2. Randomly extract matching processed words from list1 and list2 to perform cosine similarity calculation:

[0162]

[0163] Among them, S cos is the similarity value between entities; S1 is the synonym entity of the original text; S2 is the abbreviation or expanded word description that matches it; when S cos When the value tends to be between the set threshold and 1, it means that the two texts belong to the same entity. They meet the common entity description in the operation and maintenance history data text and can be included in the similar entity set for generating new text. cos If the value is not between the set threshold and 1, the key entity is deleted.

[0164] In one embodiment, the fusion disambiguation module 5 includes:

[0165] Entity disambiguation is used to distinguish entities with the same name but different meanings. Coreference resolution is used to identify situations where different expressions refer to the same entity. Cosine similarity and Jaccard similarity are used for fusion disambiguation. The calculation formula is:

[0166]

[0167] Where s1 and s2 are two different sentences or texts, representing the key entities whose similarity is to be calculated. A(s1) and A(s2) are the results of converting the text into sets, representing the word segmentation or splitting of each sentence into word sets.

[0168] The cosine similarity threshold is set to be greater than 0.8, and the Jaccard correlation coefficient threshold is set to be less than 0.7 to judge the similarities and differences between different expressions. When the nodes have high semantic similarity and low word overlap, such nodes are identified as independent nodes. For nodes with high semantic similarity and high word overlap, they are merged into the same node to reduce redundancy in the graph.

[0169] In one embodiment, it further includes:

[0170] Based on the natural language questions input by the user, the corresponding answers are matched from the target knowledge graph ontology model to achieve intelligent question answering; specifically including:

[0171] After the user enters a question, the sentence is segmented using the TF-IDF model, all words are extracted and their word frequencies are calculated to form a word frequency vector. The word frequency vector is used to calculate the cosine similarity between words, and the corresponding query statement is generated based on the extracted keywords and the query results are fed back to the user.

[0172] The above modules and units are used to execute the corresponding steps in the knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance. The specific implementation method is described in the above method embodiment and will not be repeated here.

[0173] like Figure 3 As shown, the present invention also provides a computer device, which can be a server, and its internal structure can be as follows Figure 3 As shown. The computer device includes a processor, memory, network interface and database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store all data required for the process of the knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements the knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance.

[0174] Those skilled in the art will understand that Figure 3 The structure shown in is merely a block diagram of a portion of the structure related to the present application solution and does not constitute a limitation on the computer device to which the present application solution is applied.

[0175] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the above-mentioned knowledge extraction algorithms applicable to the knowledge graph in the field of intelligent operation and maintenance.

[0176] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0177] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0178] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A knowledge extraction algorithm suitable for knowledge graphs in the field of intelligent operation and maintenance, characterized by: include: Obtain historical record data during the operation and maintenance process as raw data; wherein the historical record data includes fault phenomenon, fault cause, treatment measures, fault time, and maintenance records; An entity recognition model based on BERT-BiGRU-CRF is used to extract key entities from the original data, and synonym processing is performed on the key entities to perform data enhancement to obtain target entities; Adopting the relationship extraction method based on artificial rules, key entities are connected to form triple information by defining multiple relationships; Use the Protégé ontology building tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model; The fuzzy data and duplicate data in the knowledge graph ontology model are fused and disambiguated to obtain a target knowledge graph model; wherein the fusion disambiguation includes entity disambiguation and coreference resolution.

2. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 1 is characterized in that: The step of extracting key entities from the original data using the BERT-BiGRU-CRF-based entity recognition model and performing synonym processing on the key entities to perform data enhancement to obtain target entities includes: Extracting keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features; The keywords are dynamically transformed into word vectors using a BERT model, and the feature sequences of the original data are extracted using a BiGRU-CRF model to obtain key entities; Synonym processing is performed on the key entities to perform data enhancement.

3. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 2 is characterized in that: The step of extracting keywords from the original data using an unsupervised keyword extraction algorithm that integrates semantic features includes: For the original data text t i Perform sentence segmentation, i.e. t i ={U1,U2,…,U n }, using the Chinese word segmentation tool Jieba to analyze the sentence U i ∈t i Perform word segmentation and remove stop words by loading the stop word list, retaining the words that meet the conditions after segmentation, that is, U i =u m1 ,u m2 ,....u mi , where u ji ∈U i is the candidate keyword; Construct a candidate keyword graph G = (V, E), where V is the set of vertices in the graph, E is the set of edges in the graph, E∈V×V, for each u j,i ∈U i ,u j,i+1 ∈U i ,have j,i ,u j,i+1 >∈E, using the co-occurrence relationship between words to build a keyword graph;​ Perform text semantic representation and similarity calculation, and integrate word and text semantic information and word order information into the TextRank algorithm to construct the transition probability matrix and restart probability matrix between words. The formula for calculating the similarity between candidate words and text is: Among them, v j The vector of the jth word is represented by t, the vector of the text, and n is the dimension of the vector space, n = 512. The cosine similarity between the word and the text is calculated using the formula. The difference between the candidate word and the text is measured by the size of the cosine value of the vector space. The closer the cosine value is to 1, the higher the similarity between the two vectors, that is, the closer the semantic similarity between the candidate word and the text is. Score the candidate keywords according to the transition probability matrix, restart probability matrix and PageRank score recursive formula, and select the top N words with the highest scores as keywords according to the scores; specifically, the process includes: According to the TextRank iterative calculation process scoring method, the transition probability and restart probability of the TextRank algorithm are modified, and then the PageRank random walk technology is used to score each word node. The calculation formula is as follows: Among them, TR(u i ) represents node u i That is, the score of word i, the value of which depends on the restart probability r(u i ), the transition probability matrix value Damping coefficient β and TR(u j ), where the damping coefficient β adjusts the problem of node out-degree forming a closed loop during the iteration process, thereby allocating the proportion between the transfer probability and the restart probability, β∈[0,1], usually β=0.85; u j refers to all word nodes u in the word graph i All word nodes connected; e(u j ),u i ) is the node u j and u i The edge weight between them, out(u j ) are all connected nodes u i The sum of the edge weights of is the element of the transfer probability matrix. The larger the value is, the higher the u i The higher the score, the better; r(u i ) is the restart probability value, and the calculation formula is: u j The out-degree out(u j ) is calculated as follows: out(u j )=uf:uj→uf e(u j ,u f ) r(u i ) is calculated as follows: By calculating the initial weights as above, a new restart probability value is obtained; Calculate the edge weight e(u j ,u i ), first calculate the similarity between words and documents, and then calculate the edge weight: I j ,u i )=yes(u j ,t)+sim(u i ,t) In the above formula, sim(u j ,t) is the node u j The similarity between the text t, sim(u i ,t) is the node u i and the similarity between text t, according to e(u j ,u i ) is node u j and u i The edge weights between them, while considering the word node u j Its connected node u i and the similarity relationship between the text, so when the node u j and u i The higher the sum of similarities with document t, that is, the higher the semantic relationship between the word and the document, the greater the proportion of the word in the document; The initial value and edge weight are calculated, and then it is iterated to calculate the score of each candidate word. Finally, the words with the highest scores are selected as candidate keywords, that is, the words with the highest semantic similarity to the text.

4. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 2 is characterized in that: The steps of converting the keywords into dynamic word vectors based on the BERT model and extracting the feature sequences of the original data based on the BiGRU-CRF model to obtain key entities include: The BERT model uses a bidirectional Transformer encoding structure. It first calculates the correlation between each word and all other words, and then uses these correlations to adjust the importance of each word to generate new semantic feature expressions. The Transformer model uses a multi-head attention mode to increase the representation subspace of the attention mechanism. It also uses a positional embedding method, and its formula is: Among them, pos is the position index of the word, i is the i-th word, d model The input model has a variable length, PE(pos,2i) is the position vector of the word with even bit index, and PE(pos,2i+1) is the position vector of the word with odd bit index; The BiGRU model is used to consider the previous and next text information at the same time to extract bidirectional semantic features, and the CRF model is used to fully consider the connection relationship between entities to obtain the global optimal sequence and obtain key entities.

5. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 4 is characterized in that: The step of performing synonym processing on the key entities to perform data enhancement includes: Create a similar entity set list list1 based on the key entity, which covers the synonyms of the key entity. After random processing, form a list of abbreviations and expanded words list2. Randomly extract matching processed words from list1 and list2 to perform cosine similarity calculation: Among them, S cos is the similarity value between entities; S1 is the synonym entity of the original text; S2 is the abbreviation or expanded word description that matches it; when S cos When the value tends to be between the set threshold and 1, it means that the two texts belong to the same entity. They meet the common entity description in the operation and maintenance history data text and can be included in the similar entity set for generating new text. cos If the value is not between the set threshold and 1, the key entity is deleted.

6. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 1 is characterized in that: The step of fusing and disambiguating the fuzzy data and duplicate data in the knowledge graph ontology model to obtain a target knowledge graph model includes: Entity disambiguation is used to distinguish entities with the same name but different meanings. Coreference resolution is used to identify situations where different expressions refer to the same entity. Cosine similarity and Jaccard similarity are used for fusion disambiguation. The calculation formula is: Where s1 and s2 are two different sentences or texts, representing the key entities whose similarity is to be calculated. A(s1) and A(s2) are the results of converting the text into sets, representing the word segmentation or splitting of each sentence into word sets. The cosine similarity threshold is set to be greater than 0.8, and the Jaccard correlation coefficient threshold is set to be less than 0.7 to judge the similarities and differences between different expressions. When the nodes have high semantic similarity and low word overlap, such nodes are identified as independent nodes. For nodes with high semantic similarity and high word overlap, they are merged into the same node to reduce redundancy in the graph.

7. The knowledge extraction algorithm applicable to the knowledge graph in the field of intelligent operation and maintenance according to claim 1 is characterized in that: After the steps of fusing and disambiguating the fuzzy data and duplicate data in the knowledge graph ontology model, including entity disambiguation and coreference resolution, the method further comprises: Based on the natural language questions input by the user, the corresponding answers are matched from the target knowledge graph ontology model to achieve intelligent question answering; specifically including: After the user enters a question, the sentence is segmented using the TF-IDF model, all words are extracted and their word frequencies are calculated to form a word frequency vector. The word frequency vector is used to calculate the cosine similarity between words, and the corresponding query statement is generated based on the extracted keywords and the query results are fed back to the user.

8. A knowledge extraction device suitable for knowledge graph in the field of intelligent operation and maintenance, characterized in that: include: An acquisition module is used to acquire historical record data during the operation and maintenance process as raw data; wherein the historical record data includes fault phenomenon, fault cause, treatment measures, fault time, and maintenance records; An extraction module is used to extract key entities from the original data using a BERT-BiGRU-CRF-based entity recognition model, and perform synonym processing on the key entities to perform data enhancement to obtain target entities; The connection module is used to adopt a relationship extraction method based on artificial rules to connect key entities to form triple information by defining multiple relationships; A construction module, used to use the Protégé ontology construction tool to complete the construction of the relationship between the target entities to establish an operation and maintenance knowledge graph ontology model; A fusion disambiguation module is used to fuse and disambiguate the fuzzy data and duplicate data in the knowledge graph ontology model to obtain a target knowledge graph model; wherein, the fusion disambiguation includes entity disambiguation and coreference resolution.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Equipment operation and maintenance knowledge graph construction method and system based on AI

    CN120806100A

  • Unstructured text data intelligent analysis method and system based on large model

    CN121328569A

  • Intelligent analysis method and system for unstructured text data based on large model

    CN121328569B