Power system supply chain knowledge graph construction method and system

Abstract

The invention provides a power system supply chain knowledge graph construction method and system, and the method comprises the steps: obtaining corresponding initial data through a power system supply chain keyword, and carrying out the selective preprocessing of the initial data; and constructing a knowledge graph of the power system supply chain based on the processed data. According to the technology of introducing the knowledge graph, the supply chain information can be automatically subjected to semantic processing, entities, attributes and relations are represented in a structured mode, and the supply chain information can be intuitively understood and utilized; meanwhile, the invention provides a BERT-based position-aware triple information extraction method, and the extraction performance of the power system supply chain domain relation is improved. Therefore, during the work of analyzing the structure of the supply chain, preventing the risk of the supply chain, optimizing the layout of the supply chain and the like, the current information of the supply chain can be quickly and clearly known, the position of the problem is visually known by means of the definition of the model, and the purposes of quick identification and accurate strategy implementation are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of power market, and in particular to a method and system for constructing a knowledge graph of a power system supply chain. Background Art

[0002] Industrial economy is the lifeblood of the national economy, especially in my country, which is the country with the largest industrial scale and the most comprehensive industrial coverage in the world. However, in recent years, affected by various factors, many industrial chains have been at risk of being "stuck" due to external reasons such as upstream supply interruptions and technological blockades. In order to cope with such risks, it is imperative to discover supply chain bottlenecks and breakpoints, identify bottlenecks, find alternative pathways, and comprehensively optimize the supply chain. The primary task is to build a supply chain network knowledge graph based on complete production factors, and organically integrate information such as the distribution of supply chain elements, development logic, technical routes, and participating entities from a cognitive level, in order to lay a solid foundation and draw a blueprint for supply chain optimization and upgrading. Then, with the help of effective knowledge graphs, supply chain optimization and upgrading and simulation can be carried out, targeted protection of weak nodes in the industrial network, optimization of redundant supply structures, and formulation of protective and preventive policies to ensure the normal operation of the supply chain.

[0003] The complex supply chain structure has brought challenges to industrial analysis, and knowledge graph is a new technology that is very suitable for solving this complex structure. People habitually draw a supply map of enterprises, technologies, parts and other information around a certain type of product, which is called a supply chain or supply chain map. This drawing method is very arbitrary and does not have a unified representation method. In order to realize the semantics of supply chain information, traditional supply chain information descriptions can only be collected manually and only provide text information, lacking semantic expression. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for constructing a knowledge graph of a power system supply chain, aiming to solve the above-mentioned problems in the prior art.

[0005] An embodiment of the present invention provides a method for constructing a knowledge graph of a power system supply chain, including:

[0006] Acquire corresponding initial data through power system supply chain keywords, and selectively preprocess the initial data;

[0007] Construct a knowledge graph of the power system supply chain based on the processed data.

[0008] An embodiment of the present invention provides a power system supply chain knowledge graph construction system, including:

[0009] A data module, used to obtain corresponding initial data through power system supply chain keywords and selectively preprocess the initial data;

[0010] A building block is used to construct a knowledge graph of the power system supply chain based on the processed data.

[0011] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-mentioned method for constructing a knowledge graph of a power system supply chain.

[0012] An embodiment of the present invention also provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the above-mentioned method for constructing a knowledge graph of a power system supply chain are implemented.

[0013] The use of the embodiments of the present invention can include the following beneficial effects: The embodiments of the present invention propose a standardized and unified network relationship construction method based on the knowledge graph, which can quickly and clearly understand the current supply chain information when analyzing the supply chain structure, preventing supply chain risks, optimizing the supply chain layout, etc., and with the help of the definition of the model, the location of the problem can be visualized, thereby achieving the purpose of rapid identification and precise policy implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0015] Figure 1 It is a flow chart of a method for constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention;

[0016] Figure 2 It is a schematic diagram of the process of constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention;

[0017] Figure 3 1 is a schematic diagram of the structure of a BERT-Span entity recognition model according to an embodiment of the present invention;

[0018] Figure 4 is a schematic diagram of a relationship extraction model according to an embodiment of the present invention;

[0019] Figure 5 It is a flow chart of a method for automatically constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention;

[0020] Figure 6It is a schematic diagram of a power system supply chain knowledge graph construction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0022] Method Embodiment

[0023] According to an embodiment of the present invention, a method for constructing a knowledge graph of a power system supply chain is provided. Figure 1 is a flow chart of a method for constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention. Figure 1 As shown, the method for constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention specifically includes:

[0024] Step S101, obtaining corresponding initial data through power system supply chain keywords, and selectively preprocessing the initial data; the initial data includes structured data and unstructured data; specifically includes:

[0025] Using Python crawler technology to input power system supply chain keywords in several resource libraries to obtain corresponding initial data, storing structured data in the initial data in TXT format, writing a power system supply chain crawler program through the Requests and BeautifulSoup libraries in the Python crawler program, crawling the unstructured data in the initial data through the power system supply chain crawler program, and preprocessing the obtained text data and storing it in TXT format;

[0026] Step S102, constructing a knowledge graph of the power system supply chain based on the processed data, specifically including:

[0027] Convert the stored structured data into first knowledge triples;

[0028] The improved pre-trained language model BERT-Span is used to extract entities from the stored unstructured data, and the entities in the sentence are extracted using a location-aware extraction method to obtain the second knowledge triples; specifically, the following steps are included:

[0029] Input the stored unstructured data into the improved pre-trained language model BERT-Span, mark the beginning and end of the sentences in the data to obtain an input sequence, perform feature extraction on the input sequence to obtain a context representation matrix of the sentence, and use Formula 1 and Formula 2 to predict the probability of each word in the sentence being the starting character and ending character of different types of entities according to the context representation matrix, perform argmax operation on the probabilities of the starting character and the ending character through Formula 3 and Formula 4 to obtain the starting position and ending position matrix of the sentence, and match the starting position and ending position matrix according to the proximity principle to obtain the types of all entities in the sentence and their starting and ending positions, and complete the entity recognition of the sentence;

[0030] P start =softmax each row (X.W start ) Formula 1;

[0031] P end =softmax each row [concat(X,P start )·W end ] Formula 2;

[0032] T start = argmax(P start ) Formula 3;

[0033] T end = argmax(P end ) Formula 4;

[0034] Among them, P start and P end Represent the start probability matrix and the end probability matrix respectively, X represents the context representation matrix of the sentence, W start ∈R nx|Y| , W end ∈R (d+|Y|)×|Y| is the weight parameter, Y is the number of predefined entity types including other types, and softmax eachrow () means applying the softmax function to each row of the matrix, T start and T end Represents the matrix of the starting and ending positions of the sentence after operation;

[0035] Mark the head entity and tail entity in the sentence, and convert the marked head and tail entities into feature vectors, convert the relative distance between each word in the sentence and the head and tail entities into position vector A and position vector B through the position embedding matrix, input the context representation matrix of the sentence into the bidirectional long short-term memory network model Bi-LSTM, and obtain the sequence vector representation and embedding representation of the sentence, calculate the feature vector of the sentence by introducing the attention mechanism of the perceptual features of the comprehensive position, the word features and the sentence features according to the position vector A and the position vector B, the sequence vector representation and the embedding representation, concatenate the feature vector of the sentence with the feature vectors of the head and tail entities, and obtain the final sequence representation as shown in Formula 5, input the final sequence representation into the Softmax classifier, and use Formula 6 to perform relationship prediction to obtain the relationship between entities and the corresponding relationship type;

[0036]

[0037]

[0038] Among them, z' represents the final sequence representation of the sentence, z represents the feature vector of the sentence, t1 and t2 represent the feature vectors after the transformation of the head entity and the tail entity of the sentence respectively. represents the real matrix of the corresponding dimension after the splicing transformation, It represents the probability of relationship prediction after Softmax transformation, y represents the relationship between entities, and b E ∈R |R| is a learnable parameter, |R| is the number of categories of the relationship;

[0039] Import the first knowledge triplet and the second knowledge triplet into the Neo4j graph database to generate a final power system supply chain knowledge graph.

[0040] The following is a specific example of the method for constructing a knowledge graph of a power system supply chain according to an embodiment of the present invention. Figure 2 As shown, the above technical solution of the embodiment of the present invention is described in detail.

[0041] An embodiment of the present invention proposes a method for constructing a knowledge graph of an electric power system, which is characterized by comprising using crawler technology to obtain clean documents of a large amount of literature, using a BERT-Span-based method to extract entities from unstructured data, using a location-aware method to extract entity relationships, and finally using a Neo4j graph database to generate a knowledge graph of the electric power system supply chain.

[0042] Specifically, an embodiment of the present invention proposes a method for constructing a knowledge graph of a power system supply chain, which is characterized by using python crawler technology to input power system supply chain keywords on source websites (Tmall, Taobao, etc.), Baidu Encyclopedia, China National Knowledge Infrastructure, etc. to obtain corresponding resource data, and then preprocess the obtained data.

[0043] In data preprocessing, the acquired data includes structured data and unstructured data. For structured data, it is saved in Excel for storage and then converted into knowledge triples. For unstructured data, most academic literature journal websites, such as China National Knowledge Infrastructure and Wanfang, have begun to provide documents in HTML format for academic researchers to read online. Such HTML documents have less noise, and the corresponding text content is embedded in the format framework. A crawler is written using the Requests and BeautifulSoup libraries in the Python crawler to crawl the text content of documents in HTML format. By crawling documents in HTML format through the crawler, concise and clean TXT text content can be directly obtained, reducing the semantic distortion caused by format conversion. Since the knowledge required for knowledge graph extraction is a plain text data set, after extracting the text in the framework through a web crawler, the text needs to be preprocessed again to remove unnecessary text content, such as formulas, quoted characters, etc., and finally obtain documents stored in TXT format.

[0044] The unstructured data text obtained after processing obtains corresponding entities and relationships through knowledge extraction technology, and then is converted into knowledge triples. The knowledge extraction technology used in the embodiment of the present invention includes entity extraction in the power system supply chain field based on BERT-Span and relationship extraction based on location awareness.

[0045] like Figure 3 As shown, input layer: Assume that the sentence Sentence = (w1, w2, ..., w n ), w i Represents the i-th word in the sentence, and n is the length of the sentence. According to the BERT input format, '[CLS]' and '[SEP]' are inserted at the beginning and end of the sentence to construct the input sequence. The input vector of each word in the entity recognition model is obtained by adding character embedding, segment embedding, and position embedding.

[0046] Feature extraction layer: Based on the input layer, the hidden vector of the last layer of the BERT encoding layer is used as the feature representation of the sequence to obtain the context representation matrix X∈R of the sentence S nxd , where R nxd Represents a real matrix of dimension n×d, where d is the vector dimension of the last layer of BERT.

[0047] Decoding output layer: First, BERT outputs the context representation matrix of sentence S, and then the model predicts the probability that each word in sentence S may be the starting character of different types of entities, as shown in equations (1) and (2).

[0048] P start =softmax each row (X.W start ) (1)

[0049] P end =softmax each row [concat(X,P start )·W end ] (2)

[0050] Where: P start and P end Represent the starting probability matrix and the ending probability matrix respectively, W start ∈R nx|Y| , W end ∈R (d +|Y|)×|Y| is the weight parameter, Y is the number of predefined entity types including the "other" type, and softmax eachrow () means applying the softmax function to each row of the matrix separately.

[0051] Through P start and P end Performing argmax operation on each row of T start and T end , which is expressed as:

[0052] T start = argmax(P start ) (3)

[0053] T end = argmax(P end ) (4)

[0054] Where, T start and T end Represents the matrix of the starting and ending positions of the sentence after operation, T start and T end Each non-zero number in represents that the character at the corresponding position in the sentence S is most likely to be the start or end of the entity under the digital label type, and 0 means that the character at this position is not the boundary of the entity, i∈{1,2,...,n}. Finally, for T start and T endThe elements in are matched according to the proximity principle to obtain the types of all entities in sentence S and their starting and ending positions, thereby completing the entity recognition task of the sentence.

[0055] The task of entity relationship extraction is to mine the semantic relationship between entities from text content. The detailed description is given an input sentence S and the head entity t1 and the tail entity t2 in the sentence, so as to identify the relationship between the two entities in the sentence. Figure 4 shown.

[0056] For the sentence "China independently manufactures coal mills", the head entity and tail entity are "China" and "coal mill" respectively. The entity locators "#" and "$" are placed at the beginning and end of the head entity and tail entity respectively for entity marking, so that the model can pay attention to the particularity of these two words. Finally, the model determines the relationship of "place of origin" through the classifier. The following will introduce the various parts of the relationship extraction model in detail.

[0057] Relative position representation of the head and tail entity span: Assume [x a ,x a+1 ,...,x b ] is the span of the head entity, defining the position sequence relative to the span of the head entity as follows:

[0058]

[0059] Where: a and b are the starting and ending positions of the head entity span in the sentence, is the relative position of each word to the head entity, where the superscript s represents the head entity and i represents the i-th word in the head entity. Similarly, the position sequence relative to the tail entity can be obtained o represents the tail entity. The relative distance between each word and the head and tail entities is converted into a vector through a learnable position embedding matrix and Represents a real number matrix with dimension p as the length of sequence p. Based on the sentence representation embedded in Bi-LSTM, the context representation X of sentence S obtained by BERT is input into a bidirectional LSTM. At each moment, the forward hidden state and reverse hidden state of Bi-LSTM are concatenated to obtain Represents a real number matrix with dimension equal to the length of the hidden layer. The final hidden state in each direction of the Bi-LSTM is concatenated to obtain the sequence vector representation of the entire sentence S.

[0060]

[0061]

[0062]

[0063] In the formula, h i is the i-th element in the sentence embedding matrix H, represents the forward LSTM, represents the reverse LSTM, x i Denotes the i-th input of LSTM. Using equations (6) to (8), we get the embedding representation H = [h1; h2; ...; h n ].

[0064] Sentence representation based on attention mechanism: Models with attention mechanism have achieved advanced performance in many NLP tasks. In the relation extraction model, the attention mechanism of comprehensive position perception features, word features and sentence features is also added. The representation z of sentence S is calculated through the attention mechanism:

[0065]

[0066]

[0067]

[0068] Where: v refers to the value vector, u i Refers to the output vector calculated by the attention mechanism, a i Indicates how much attention (i.e., weight) should be given to the i-th element when calculating the context representation of the current element. The larger the weight, the greater the contribution of the element to the current output. n is consistent with the previous description, W q , W s , is a trainable parameter, They represent the real number matrices of different dimensions generated after the attention mechanism. From equations (9) to (11), we can see that this mechanism can evaluate the relative contribution of each word in the sequence, and can also make full use of the global position information of the entity in the entire sequence to convert the types of the head entity and tail entity obtained by the entity recognition model into feature vectors and The real number matrix representing the dimension of vector t is concatenated with the feature vector z of sentence S obtained by formula (11) to obtain the final sequence representation z', which is input into the Softmax classifier for relationship prediction.

[0069]

[0070]

[0071] Where: represents the real matrix of the corresponding dimension after the splicing transformation, represents the probability of relationship prediction after Softmax transformation, b E ∈R |R| is a learnable parameter, and |R| is the number of relation categories. Using formula (13), we can get the entity relation prediction and its relation type, and thus extract the head entity-relationship-tail entity triple.

[0072] Then, the embodiment of the present invention adopts a method based on the Neo4j graph database to store knowledge, imports knowledge triple data by adopting the neo4j-import method, and finally generates a knowledge graph of the power system supply chain.

[0073] In summary, the implementation process of the embodiment of the present invention is summarized as follows (eg Figure 5 shown):

[0074] Step 1, write a crawler using the Requests and BeautifulSoup libraries in Python crawler to obtain documents stored in TXT format.

[0075] Step 2: Use the BERT-Span based entity extraction method to extract entities.

[0076] Step 3: Use location-aware entity relationship extraction method to extract relationships.

[0077] Step 4: Import the generated knowledge triples into Neo4j to generate the power system supply chain knowledge graph.

[0078] The embodiments of the present invention include the following beneficial effects:

[0079] (1) An automatic construction method of a power system supply chain knowledge graph provided by an embodiment of the present invention overcomes the defects of the previous construction method using manual information collection;

[0080] (2) The embodiment of the present invention proposes a BERT-Span-based method to extract entities and a location-aware method to extract entity relationships. This model improves the performance of knowledge extraction.

[0081] System Example

[0082] According to an embodiment of the present invention, a power system supply chain knowledge graph construction system is provided. Figure 6 is a schematic diagram of a power system supply chain knowledge graph construction system according to an embodiment of the present invention. Figure 6As shown, the power system supply chain knowledge graph construction system according to an embodiment of the present invention specifically includes:

[0083] A data module 60, for acquiring corresponding initial data through power system supply chain keywords, and selectively preprocessing the initial data;

[0084] The initial data includes structured data and unstructured data;

[0085] The construction module 62 is used to construct a knowledge graph of the power system supply chain based on the processed data.

[0086] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0087] In summary, the embodiments of the present invention have the following beneficial effects:

[0088] (1) The technology of introducing knowledge graph proposed in the embodiment of the present invention can automatically process the supply chain information semantically, and structure the entities, attributes and relationships, so that the supply chain information can be intuitively understood and used;

[0089] (2) The embodiment of the present invention proposes a location-aware triplet information extraction method based on BERT, which improves the performance of extracting relationships in the power system supply chain field.

[0090] Device Example 1

[0091] An embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps described in the method embodiment when executed by the processor.

[0092] Device Example 2

[0093] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps described in the method embodiment are implemented.

[0094] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a knowledge graph of a power system supply chain, characterized in that include: Acquire corresponding initial data through power system supply chain keywords, and selectively preprocess the initial data; Construct a knowledge graph of the power system supply chain based on the processed data.

2. The method according to claim 1, characterized in that The initial data includes structured data and unstructured data.

3. The method according to claim 2, characterized in that Acquiring corresponding initial data through power system supply chain keywords and selectively preprocessing the initial data specifically include: The Python crawler technology is used to input power system supply chain keywords into several resource libraries to obtain corresponding initial data, and the structured data in the initial data is stored in TXT format. The power system supply chain crawler program is written through the Requests and BeautifulSoup libraries in the Python crawler program, and the unstructured data in the initial data is crawled through the power system supply chain crawler program, and the acquired text data is pre-processed and stored in TXT format.

4. The method according to claim 3, characterized in that The knowledge graph of the power system supply chain based on the processed data includes: Convert the stored structured data into first knowledge triples; The improved pre-trained language model BERT-Span is used to extract entities from the stored unstructured data, and the relationship extraction of entities in the sentence is performed using a location-aware extraction method to obtain the second knowledge triples. Import the first knowledge triplet and the second knowledge triplet into the Neo4j graph database to generate a final power system supply chain knowledge graph.

5. The method according to claim 4, characterized in that The improved pre-trained language model BERT-Span is used to extract entities from stored unstructured data, including: Input the stored unstructured data into the improved pre-trained language model BERT-Span, mark the beginning and end of the sentences in the data to obtain an input sequence, perform feature extraction on the input sequence to obtain a context representation matrix of the sentence, and use Formula 1 and Formula 2 to predict the probability of each word in the sentence being the starting character and ending character of different types of entities according to the context representation matrix, perform argmax operation on the probabilities of the starting character and the ending character through Formula 3 and Formula 4 to obtain the starting position and ending position matrix of the sentence, and match the starting position and ending position matrix according to the proximity principle to obtain the types of all entities in the sentence and their starting and ending positions, and complete the entity recognition of the sentence; P start =softmax eachrow (X.W start ) Formula 1; P end =softmax eachrow [concat(X,P start )·W end ] Formula 2; T start = argmax(P start ) Formula 3; T end = argmax(P end ) Formula 4; Among them, P start and P end Represent the start probability matrix and the end probability matrix respectively, X represents the context representation matrix of the sentence, W start ∈R nx|Y| , W end ∈R (d+|Y|)×|Y| is the weight parameter, Y is the number of predefined entity types including other types, and softmax eachrow () means applying the softmax function to each row of the matrix, T start and T end Represents the start and end position matrix of the sentence after operation.

6. The method according to claim 5, characterized in that The relationship extraction of entities in sentences using the location-aware extraction method specifically includes: Mark the head entity and tail entity in the sentence, and convert the marked head and tail entities into feature vectors, convert the relative distance between each word in the sentence and the head and tail entities into position vector A and position vector B through the position embedding matrix, input the context representation matrix of the sentence into the bidirectional long short-term memory network model Bi-LSTM, and obtain the sequence vector representation and embedding representation of the sentence, calculate the feature vector of the sentence by introducing the attention mechanism of the perceptual features of the comprehensive position, the word features and the sentence features according to the position vector A and the position vector B, the sequence vector representation and the embedding representation, concatenate the feature vector of the sentence with the feature vectors of the head and tail entities, and obtain the final sequence representation as shown in Formula 5, input the final sequence representation into the Softmax classifier, and use Formula 6 to perform relationship prediction to obtain the relationship between entities and the corresponding relationship type; Among them, z' represents the final sequence representation of the sentence, z represents the feature vector of the sentence, t1 and t2 represent the feature vectors after the transformation of the head entity and the tail entity of the sentence respectively. represents the real matrix of the corresponding dimension after the splicing transformation, It represents the probability of relationship prediction after Softmax transformation, y represents the relationship between entities, and b E ∈R |R| is a learnable parameter and |R| is the number of categories of the relation.

7. A knowledge graph construction system for power system supply chain, characterized by include: A data module, used to obtain corresponding initial data through power system supply chain keywords and selectively preprocess the initial data; A building block is used to construct a knowledge graph of the power system supply chain based on the processed data.

8. The system according to claim 7, characterized in that The initial data includes structured data and unstructured data.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for constructing a knowledge graph of a power system supply chain as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by the processor, the steps of the method for constructing a knowledge graph of a power system supply chain as described in any one of claims 1 to 6 are implemented.

Images