Knowledge base-based power operation and maintenance dispatching automation job entity extraction method and system

Abstract

The application provides a knowledge base-based power operation and maintenance dispatching automation operation entity extraction method and system, which is used for power dispatching automation operation entity extraction, mining of potential rules of multi-modal data in the field of dispatching automation, and comprises a multi-modal data extraction module, a text extraction module based on similarity calculation and an entity annotation module based on an LDA algorithm; the text extraction module comprises the multi-modal data extraction module, which uses different tool packages to preliminarily extract text information according to multi-modal data characteristics, and then obtains high-quality text data by comprehensively considering similarity calculation methods such as edit distance and cosine similarity; the entity annotation module uses the LDA algorithm to extract key entities in the text and annotate and construct a power dispatching automation operation text knowledge base; the application can better analyze the potential rules of multi-modal power dispatching automation data, deeply mine potential entity characteristics in the power dispatching automation operation text and annotate, and thus improve the entity extraction accuracy in the field of power automation.

Description

Technical Field

[0001] This invention relates to the field of power grid dispatch automation, and in particular to a method and system for extracting operation entities for power operation and maintenance dispatch automation based on a knowledge base. Background Technology

[0002] Over the long-term process of automated dispatching, power networks have accumulated a wealth of structured and unstructured data. Traditional manual data analysis methods are insufficient to meet the demands of fault diagnosis and data analysis in today's large-scale automated power dispatching operations under the big data paradigm. Knowledge graph technology can effectively address the challenge of knowledge representation in the power automation field. Entity extraction is a crucial step in building a knowledge graph. Given the poor accuracy of entity extraction in power automation, designing a knowledge base to assist in entity extraction is particularly important, as it can significantly improve the accuracy of entity extraction. Summary of the Invention

[0003] This invention proposes a knowledge-based method and system for extracting entities from automated power operation and maintenance dispatching operations. This method can better analyze the potential patterns in multimodal automated power dispatching data, deeply mine and annotate the potential entity features in automated power dispatching operation texts, and thus improve the accuracy of entity extraction in the field of power automation.

[0004] The present invention adopts the following technical solution.

[0005] This knowledge-based power dispatch automation entity extraction system is used for extracting entities from power dispatch automation operations and mining the potential patterns in multimodal data generated in the field of power dispatch automation. It includes a multimodal data extraction module, a text extraction module based on similarity calculation, and an entity annotation module based on the LDA algorithm. The text extraction module uses different toolkits to initially extract text information based on the characteristics of multimodal data, and then comprehensively considers similarity calculation methods such as edit distance and cosine similarity to obtain high-quality text data. The entity annotation module uses the LDA algorithm to extract key entities from the text and annotates them to build a knowledge base for power dispatch automation operation text.

[0006] The knowledge-based power dispatch automation entity extraction method employs a knowledge-based power dispatch automation entity extraction system, including the following steps:

[0007] Step S1: Extract text features from multimodal power dispatch automation data;

[0008] Step S2: Reduce text redundancy by calculating text similarity to obtain high-quality text data;

[0009] Step S3: Design an entity extraction module based on LDA to build a knowledge base by accurately labeling entities;

[0010] Step S4: Based on the knowledge base, use the BERT-BiLSTM-CRF model to extract entities from power dispatch automation data and obtain entities from the text.

[0011] In step S1, text feature extraction is performed on the multimodal power dispatch automation data, specifically including the following steps:

[0012] Step S1A1: For text data such as word and txt, use the python-docx library to extract and classify text according to the structure and format of the text;

[0013] Step S1A2: For PDF data such as ledger information and operation manuals, use OCR technology to extract text information from the PDF files;

[0014] Step S1A3: For work tickets and safety regulations in Excel spreadsheet format, use the pandas library to extract key text.

[0015] Step S1A4: Use a text segmentation method with thresholds to reconstruct different types of text, segment texts with a length greater than the upper threshold, and remove texts with a length less than the lower threshold.

[0016] Step S1A5, following steps S1A1, S1A2, S1A3, and S1A4, yields the preliminarily extracted text data for power dispatch automation.

[0017] In step S2, text redundancy is reduced by calculating text similarity to obtain high-quality text data. Specific steps include:

[0018] Step S2A1: Set a set of weight parameters λ = (λ1, λ2, λ3), which represent the weights of edit distance, cosine similarity, and TF-IDF, respectively; set the threshold Ω, which represents the threshold for text similarity.

[0019] Step S2A2: Manually extract a small number of samples of different types and establish a sample knowledge base D;

[0020] Step S2A3: Calculate the edit distance between the text data and the knowledge base D to obtain the edit distance similarity S1; Step S2A4: Calculate the cosine similarity between the text data and the knowledge base D to obtain the cosine similarity S2; Step S2A5: Calculate the TF-IDF value between the text data and the knowledge base D to obtain the TF-IDF similarity S3; Step S2A6: Obtain the text similarity score S through weighted similarity calculation.

[0021] Step S2A7: Compare the similarity score S with the threshold. If it is greater than the threshold Ω, the text is deleted; if it is less than or equal to the threshold, the text is added to the knowledge base D.

[0022] Step S2A8, repeat steps S2A3-S2A7 to construct the complete entity text knowledge base D.

[0023] In step S2A3, the specific calculations of edit distance and edit distance similarity are as follows:

[0024]

[0025]

[0026] In the formula, It is an indicator function, when a i =b j When lea is active, its value is 0; otherwise, its value is 1. a,b (i,j) represents the edit distance between the first i characters of text a and the first j characters of text b; |a| and |b| represent the lengths of text a and text b, and similarity represents the edit distance similarity.

[0027] In step S2A4, the cosine similarity is calculated as follows:

[0028]

[0029] In the formula, A i B i This represents the values ​​of each dimension of vectors A and B. This represents finding the dot product of vectors A and B. This represents the product of the modulo of A and B;

[0030] In step S2A5, the specific calculation of TF-IDF is as follows:

[0031]

[0032]

[0033] TF-IDF = TF·IDF

[0034] In the formula, n i,j Indicates term t i In document d j The number of times it appears in TF i,j Indicates the term t i In document d j The frequency of occurrence in |D|. |D| represents the total number of documents, |j:t i ∈d j | indicates that the term is included. iThe number of documents; in step S2A6, the weighted similarity calculation is specifically calculated as follows:

[0035] S=λ1S1+λ2S2+λ3S3

[0036] In the formula, λ1, λ2, and λ3 are the weight parameters set in S2A1, and S1, S2, and S3 are the similarities calculated in steps S2A3, S2A4, and S2A5.

[0037] In step S3, an entity extraction module based on LDA is designed to accurately label entities and build a knowledge base. This includes the following steps:

[0038] Step S3A1: Use the LDA algorithm to extract text keywords from knowledge base D.

[0039] Step S3A2: Label the entities extracted using the LDA algorithm according to the actual situation and construct the entity extraction knowledge base D.

[0040] In step S3A1, the specific calculations of the LDA algorithm are as follows:

[0041]

[0042] In the formula, P(W,Z) represents the file probability of a certain combination generated by LDA; P(θ) j ,α) represents the probability distribution of the main body composition of each file; P(Z) represents the probability distribution of each topic; j,t |θ j ) represents the probability multinomial distribution of different topics for a given file; This represents the probability multinomial distribution of words for a given topic.

[0043] In step S4, a BERT-BiLSTM-CRF entity extraction model is constructed, and a knowledge base is used for label correction during entity label prediction. This includes the following steps:

[0044] Step S4A1: Construct a BERT-BiLSTM-CRF model to extract entities from the text and predict labels.

[0045] Step S4A2: Use the knowledge base to correct the erroneous labels of the entities extracted in S4A1.

[0046] The BERT-BiLSTM-CRF model includes the BERT model, the BiLSTM model, and the CRF model.

[0047] In step S4A1, the specific formula for the BERT model is as follows:

[0048]

[0049] V = BERT([CLS], T1, ..., T n ,[SEP])=(e1,...,e n )

[0050] In the formula, E = {E1, E2, ..., E} n} represents the initial text sequence, T = {T1, T2, ..., T} n} represents the text vector obtained after the initial text sequence is encoded by the transformer encoder.

[0051] In step S4A1, the specific formula for the BiLSTM model is as follows:

[0052] i t =σ(W ix X t +W ih h t-1 +b i )

[0053] f t =σ(W fx X t +W fh h t-1 +b f )

[0054] o t =σ(W ox X t +W oh h t-1 +b o )

[0055]

[0056]

[0057] h t =o t Θtanh(C t )

[0058]

[0059] In the formula, f t Indicates the forget gate state at time t, o t This indicates the output gate state at time t. h represents the cell state at time t. t This represents the hidden layer state at time t;

[0060] In step S4A1, the specific formula for the CRF model is as follows:

[0061]

[0062] In the formula, N represents the number of training samples, C represents the regularization parameter, and y (i) Let x represent the true label sequence of the i-th sample. (i) Let p(y) represent the input sequence of the i-th sample. (i) |x (i) ) represents the conditional probability distribution in the CRF model.

[0063] Compared with existing technologies, the present invention has the following advantages: it can better analyze the potential patterns of multimodal power dispatch automation data, deeply mine and annotate the potential entity features in power dispatch automation text, and thus improve the accuracy of entity extraction in the field of power automation. Attached Figure Description

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0065] Appendix Figure 1 This is a schematic diagram of the principle framework of the present invention for entity extraction in power dispatch automation;

[0066] Appendix Figure 2 This is a schematic diagram illustrating the principle framework for constructing a knowledge base in this invention. Detailed Implementation

[0067] As shown in the figure, the knowledge base-based power dispatch automation entity extraction system is used for extracting entities for power dispatch automation operations and mining the potential patterns of multimodal data generated in the field of power dispatch automation. It includes a multimodal data extraction module, a text extraction module based on similarity calculation, and an entity annotation module based on the LDA algorithm. The text extraction module includes a multimodal data extraction module, which uses different toolkits to initially extract text information based on the characteristics of multimodal data, and then comprehensively considers edit distance and cosine similarity calculation methods to obtain high-quality text data. The entity annotation module uses the LDA algorithm to extract key entities in the text and annotates them to build a knowledge base for power dispatch automation operation text.

[0068] A knowledge-based entity extraction method for power dispatch automation is proposed, employing a knowledge-based power dispatch automation entity extraction system. The overall framework of the method is shown in the attached figure. Figure 1As shown, this includes a multimodal data extraction module, a text extraction module, and an LDA-based entity extraction module. To address the characteristics of multimodal data, different toolkits are used for initial text extraction; thresholds are set and multiple similarity calculation methods are used to reduce text redundancy, resulting in high-quality text data; the LDA algorithm is used for text entity extraction, thereby accurately labeling entities and constructing a knowledge base.

[0069] Includes the following steps:

[0070] Step S1: Extract text features from multimodal power dispatch automation data;

[0071] Step S2: Reduce text redundancy by calculating text similarity to obtain high-quality text data;

[0072] Step S3: Design an entity extraction module based on LDA to build a knowledge base by accurately labeling entities;

[0073] Step S4: Based on the knowledge base, use the BERT-BiLSTM-CRF model to extract entities from power dispatch automation data and obtain entities from the text.

[0074] In step S1, text feature extraction is performed on the multimodal power dispatch automation data, specifically including the following steps:

[0075] Step S1A1: For text data such as word and txt, use the python-docx library to extract and classify text according to the structure and format of the text;

[0076] Step S1A2: For PDF data such as ledger information and operation manuals, use OCR technology to extract text information from the PDF files;

[0077] Step S1A3: For work tickets and safety regulations in Excel spreadsheet format, use the pandas library to extract key text.

[0078] Step S1A4: Use a text segmentation method with thresholds to reconstruct different types of text, segment texts with a length greater than the upper threshold, and remove texts with a length less than the lower threshold.

[0079] Step S1A5, following steps S1A1, S1A2, S1A3, and S1A4, yields the preliminarily extracted text data for power dispatch automation.

[0080] In step S2, text redundancy is reduced by calculating text similarity to obtain high-quality text data. Specific steps include:

[0081] Step S2A1: Set a set of weight parameters λ = (λ1, λ2, λ3), which represent the weights of edit distance, cosine similarity, and TF-IDF, respectively; set the threshold Ω, which represents the threshold for text similarity.

[0082] Step S2A2: Manually extract a small number of samples of different types and establish a sample knowledge base D;

[0083] Step S2A3: Calculate the edit distance between the text data and the knowledge base D to obtain the edit distance similarity S1; Step S2A4: Calculate the cosine similarity between the text data and the knowledge base D to obtain the cosine similarity S2; Step S2A5: Calculate the TF-IDF value between the text data and the knowledge base D to obtain the TF-IDF similarity S3; Step S2A6: Obtain the text similarity score S through weighted similarity calculation.

[0084] Step S2A7: Compare the similarity score S with the threshold. If it is greater than the threshold Ω, the text is deleted; if it is less than or equal to the threshold, the text is added to the knowledge base D.

[0085] Step S2A8, repeat steps S2A3-S2A7 to construct the complete entity text knowledge base D.

[0086] In step S2A3, the specific calculations of edit distance and edit distance similarity are as follows:

[0087]

[0088] In the formula, 1 (ai≠bj) It is an indicator function, when a i =b j When lea is active, its value is 0; otherwise, its value is 1. a,b (i,j) represents the edit distance between the first i characters of text a and the first j characters of text b; |a| and |b| represent the lengths of text a and text b, and similarity represents the edit distance similarity.

[0089] In step S2A4, the cosine similarity is calculated as follows:

[0090]

[0091] In the formula, A i B i This represents the values ​​of each dimension of vectors A and B. This represents finding the dot product of vectors A and B. This represents the product of the modulo of A and B;

[0092] In step S2A5, the specific calculation of TF-IDF is as follows:

[0093]

[0094] TF-IDF = TF·IDF

[0095] In the formula, n i,j Indicates the term t i In document d j The number of times it appears in TF i,j Indicates the term t i In document d j The frequency of occurrence in |D|. |D| represents the total number of documents, |j:t i ∈d j | indicates that the term is included. i The number of documents;

[0096] In step S2A6, the weighted similarity calculation is performed as follows:

[0097] S=λ1S1+λ2S2+λ3S3

[0098] In the formula, λ1, λ2, and λ3 are the weight parameters set in S2A1, and S1, S2, and S3 are the similarities calculated in steps S2A3, S2A4, and S2A5.

[0099] In step S3, an entity extraction module based on LDA is designed to accurately label entities and build a knowledge base. This includes the following steps:

[0100] Step S3A1: Use the LDA algorithm to extract text keywords from knowledge base D.

[0101] Step S3A2: Label the entities extracted using the LDA algorithm according to the actual situation and construct the entity extraction knowledge base D.

[0102] In step S3A1, the specific calculations of the LDA algorithm are as follows:

[0103]

[0104] In the formula, P(W,Z) represents the file probability of a certain combination generated by LDA; P(θ) j ,α) represents the probability distribution of the main body composition of each file; P(Z) represents the probability distribution of each topic; j,t |θ j ) represents the probability multinomial distribution of different topics for a given file; This represents the probability multinomial distribution of words for a given topic.

[0105] In step S4, a BERT-BiLSTM-CRF entity extraction model is constructed, and a knowledge base is used for label correction during entity label prediction. This includes the following steps:

[0106] Step S4A1: Construct a BERT-BiLSTM-CRF model to extract entities from the text and predict labels.

[0107] Step S4A2: Use the knowledge base to correct the erroneous labels of the entities extracted in S4A1.

[0108] The BERT-BiLSTM-CRF model includes the BERT model, the BiLSTM model, and the CRF model.

[0109] In step S4A1, the specific formula for the BERT model is as follows:

[0110]

[0111] V = BERT([CLS], T1, ..., T n ,[SEP])=(e1,...,e n )

[0112] In the formula, E = {E1, E2, ..., E} n} represents the initial text sequence, T = {T1, T2, ..., T} n} represents the text vector obtained after the initial text sequence is encoded by the transformer encoder.

[0113] In step S4A1, the specific formula for the BiLSTM model is as follows:

[0114] i t =σ(W ix X t +W ih h t-1 +b i )

[0115] f t =σ(W fx X t +W fh h t-1 +b f )

[0116] o t =σ(W ox X t +W oh h t-1 +b o )

[0117]

[0118] h t =o t Θtanh(C t )

[0119]

[0120] In the formula, f t Indicates the forget gate state at time t, o t This indicates the output gate state at time t. h represents the cell state at time t. t This represents the hidden layer state at time t;

[0121] In step S4A1, the specific formula for the CRF model is as follows:

[0122]

[0123] In the formula, N represents the number of training samples, C represents the regularization parameter, and y (i) Let x represent the true label sequence of the i-th sample. (i) Let p(y) represent the input sequence of the i-th sample. (i) |x (i) ) represents the conditional probability distribution in the CRF model.

[0124] This embodiment includes the following functional modules: a multimodal data extraction module, a text extraction module based on similarity calculation, an entity annotation module based on the LDA algorithm, and an entity extraction module based on a knowledge base and the BERT-BiLSTM-CRF model;

[0125] For multimodal data in power dispatch automation, traditional manual extraction and analysis methods are insufficient to handle the massive amounts of data available today. Using different toolkits based on the characteristics of different modalities to extract text data can reduce manual costs. A text extraction module based on similarity calculation can reduce text redundancy, uncover potential patterns in the text of power dispatch automation data, and obtain high-quality text. An entity extraction method based on the LDA algorithm for power dispatch automation operation text can extract key fields from the text, enabling subsequent use of sequence labeling to build a knowledge base for power dispatch automation operation entity extraction and improving entity extraction accuracy.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A knowledge-based power dispatch automation entity extraction system, used for extracting entities from power dispatch automation operations and mining the potential patterns in multimodal data generated in the field of power dispatch automation, characterized by: It includes a multimodal data extraction module, a text extraction module based on similarity calculation, and an entity annotation module based on the LDA algorithm. The text extraction module includes a multimodal data extraction module, which uses different toolkits to initially extract text information based on the characteristics of multimodal data, and then comprehensively considers edit distance and cosine similarity calculation methods to obtain high-quality text data. The entity annotation module uses the LDA algorithm to extract key entities from the text and annotate them to build a knowledge base for power dispatch automation operation text. The knowledge-based power dispatch automation entity extraction system uses a knowledge-based power dispatch automation entity extraction method, including the following steps: Step S1: Extract text features from multimodal power dispatch automation data; Step S2: Reduce text redundancy by calculating text similarity to obtain high-quality text data; Step S3: Design an entity extraction module based on LDA to build a knowledge base by accurately labeling entities; Step S4: Based on the knowledge base, use the BERT-BiLSTM-CRF model to extract entities from power dispatch automation data and obtain entities from the text; In step S1, text feature extraction is performed on the multimodal power dispatch automation data, specifically including the following steps: Step S1A1: For text data of type word and txt, use the python-docx library to extract and classify text according to the structure and format of the text; Step S1A2: For PDF data such as ledger information and operation manuals, use OCR technology to extract text information from the PDF files; Step S1A3: For work tickets and safety regulations in Excel spreadsheet format, use the pandas library to extract key text. Step S1A4: Use a text segmentation method with thresholds to reconstruct different types of text, segment texts with a length greater than the upper threshold, and remove texts with a length less than the lower threshold. Step S1A5, following steps S1A1, S1A2, S1A3, and S1A4, yields the preliminarily extracted text data for power dispatch automation.

2. The knowledge base-based power dispatch automation entity extraction system according to claim 1, characterized in that: In step S2, text redundancy is reduced by calculating text similarity to obtain high-quality text data. Specific steps include: Step S2A1: Set a set of weight parameters , representing the weights of edit distance, cosine similarity, and TF-IDF, respectively; set the threshold. , representing the threshold for text similarity; Step S2A2: Manually extract a small number of samples of different types and establish a sample knowledge base; Step S2A3: Obtain the edit distance similarity by calculating the edit distance between the text data and the knowledge base. ; Step S2A4: Obtain the cosine similarity by calculating the cosine similarity between the text data and the knowledge base. ; Step S2A5: Obtain the TF-IDF similarity by calculating the TF-IDF value between the text data and the knowledge base. ; Step S2A6: Obtain the text similarity score through weighted similarity calculation. ; Step S2A7: Calculate the similarity score. Compared with the threshold, greater than the threshold If the value is less than or equal to the threshold, the text is deleted; otherwise, it is added to the knowledge base. Step S2A8, repeat steps S2A3-S2A7 to construct a complete entity text knowledge base.

3. The knowledge base-based power dispatch automation entity extraction system according to claim 2, characterized in that: In step S2A3, the specific calculations of edit distance and edit distance similarity are as follows: In the formula, It is an indicator function, when When the time is 0, the value is 1; otherwise, the value is 1. express The former characters and The former Edit distance of one character; , Represented as text With text Length, Indicates edit distance similarity; In step S2A4, the cosine similarity is calculated as follows: In the formula, , Representing vectors , Values ​​of each dimension Representing vectors and Find the inner product. express and Modulus product; In step S2A5, the specific calculation of TF-IDF is as follows: In the formula, Indicates a term In the document The number of times it appears in Indicates a term In the document The frequency of occurrence. This indicates the number of all documents. This indicates that the term is included. The number of documents; In step S2A6, the weighted similarity calculation is performed as follows: In the formula, The weight parameters are set in S2A1, and S1, S2, and S3 are the similarities calculated in steps S2A3, S2A4, and S2A5.

4. The knowledge base-based power dispatch automation entity extraction system according to claim 1, characterized in that: In step S3, an entity extraction module based on LDA is designed to accurately label entities and build a knowledge base. This includes the following steps: Step S3A1: Use the LDA algorithm to extract text keywords from the knowledge base. Step S3A2: Label the entities extracted using the LDA algorithm according to the actual situation and build an entity extraction knowledge base.

5. The knowledge base-based power dispatch automation entity extraction system according to claim 4, characterized in that: In step S3A1, the specific calculations of the LDA algorithm are as follows: In the formula, This represents the probability of a file combination generated by LDA; This represents the probability distribution of the main body composition of each file; This represents the probability distribution of each topic. This represents the probability multinomial distribution of different topics for a given file; This represents the probability multinomial distribution of words for a given topic.

6. The knowledge base-based power dispatch automation entity extraction system according to claim 1, characterized in that: Step S4 also includes the following steps: Step S4A1: Construct a BERT-BiLSTM-CRF model to extract entities from the text and predict labels. Step S4A2: Use the knowledge base to correct the erroneous labels of the entities extracted in S4A1.

Images