A method of recommending categories and groups for power equipment asset management data

Abstract

The application provides a kind of electric power equipment asset management data belonging to class and group recommendation method, comprising: using natural language technology to process the text data of electric power asset management system, obtain effective text data;Effective text data is handled to entity, and knowledge graph is constructed based on entity result;According to the topological structure information of knowledge graph, the influence of knowledge graph node is calculated;Extract the standard entity of user input asset information, and extract the subgraph from knowledge graph that matches the standard entity;According to the theme probability distribution and node influence of matched subgraph, the recommended asset class and asset group are obtained.The application solves the problem that the existing recommendation method for electric power asset data management platform has low recommendation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing systems, in particular to a method for recommending a category and group to which power equipment asset management data belongs. BACKGROUND

[0002] In the constantly improving Internet technology services, various resource platform systems can extract information from the various data provided by users and automatically recommend related content by combining the characteristics of the data, so the content-based recommendation technology has also promoted the progress of the Internet. The recommendation model in the resource platform system usually takes the data provided by the user and the resource content of the database as the basis for judgment, and realizes the recommendation through the calculation of similarity, but in this process, the recommendation result is affected by the richness of the data resource content, and how to train a high-precision recommendation model also faces difficulties. Today, computer technology has entered the rapidly developing information age, and users can obtain various information they need from the vast amount of data resource platform system, so the platform system brings convenience and completeness to the user's information acquisition, for example, the user can quickly obtain the intended purchase items from the platform. Such a resource platform system brings automation and intelligent convenience to users in obtaining information from vast amounts of data content, so the data resource platform system needs to continuously improve and optimize such data recommendation methods. In the vast Internet, information resources continue to grow exponentially, and it is a challenge for users to mine the key content they need from the vast amount of information resources.

[0003] According to the current development situation, various Internet platforms compete and innovate to optimize the automatic recommendation of resource data content. However, it is often difficult to obtain a high-accuracy recommendation result, especially on the power asset data management platform in the power field, and few recommendation methods take the mutual influence degree between text data as the basis for recommendation, and the type definition and relationship definition standards are not clear. Therefore, in view of the problem that the automatic recommendation of the to-be-recommended text data provided by the user is greatly different, continuous technical iteration and exploration are needed. SUMMARY

[0004] In view of the above deficiencies in the prior art, the present application provides a method for recommending a category and group to which power equipment asset management data belongs, which solves the problem of low recommendation accuracy of the existing recommendation method for the power asset data management platform.

[0005] In order to achieve the above-mentioned application purpose, the technical scheme adopted by the present application is as follows: a method for recommending a category and group to which power equipment asset management data belongs, comprising:

[0006] processing the text data of the power asset management system by using natural language technology to obtain effective text data;

[0007] perform entity processing on the effective text data, and construct a knowledge graph based on an entity result;

[0008] calculate influence of nodes of the knowledge graph according to topological structure information of the knowledge graph;

[0009] extract a standard entity of asset information input by a user, and extract a sub-knowledge graph matched with the standard entity from the knowledge graph.

[0010] obtain a recommended asset category and asset group according to a probability distribution of a theme of the matched sub-knowledge graph and the influence of the nodes;

[0011] Further, the natural language processing includes word segmentation, keyword extraction, semantic extraction and stop word removal.

[0012] The entity processing includes entity recognition and entity alignment.

[0013] Further, the calculation of the influence of the nodes of the knowledge graph includes:

[0014] extract relationships between nodes from the knowledge graph to construct a directed graph;

[0015] calculate global influence of the nodes according to the directed graph;

[0016] normalize the global influence to obtain normalized global influence.

[0017] Further, the calculation of the global influence of the nodes adopts a first-order Markov chain model.

[0018] Further, a standard entity of asset information input by a user is extracted by using an entity extraction network model, wherein the entity extraction network model includes an input layer, a word embedding layer, a convolution attention layer, a GRU layer, a global attention layer, a CRF layer and an output layer; an input end of the word embedding layer is connected with the input layer, and an output end thereof is connected with an input end of the convolution attention layer; an input end of the GRU layer is connected with an output end of the convolution attention layer, and an output end thereof is connected with an input end of the global attention layer; an input end of the CRF layer is connected with an output end of the global attention layer, and an output end thereof is connected with the output layer.

[0019] Further, the extraction of the sub-knowledge graph matched with the standard entity from the knowledge graph includes:

[0020] calculate similarity between the standard entity of the user and node information on the knowledge graph;

[0021] construct nodes with a similarity higher than a similarity threshold value as the sub-knowledge graph matched with the standard entity.

[0022] Further, the similarity calculation formula is:

[0023]

[0024] Wherein, sim information (M A ,M B ) is the similarity, M A is the standard entity of the user, M B is the node information on the knowledge graph, |M A | and |M B | are the modulus operations of respective vectors, and · is the dot product.

[0025] Further, the recommended asset category and asset group are obtained according to the probability distribution of the theme of the matched sub-knowledge graph and the node influence, and the recommended asset category and asset group include:

[0026] Extract the knowledge graph node influence in the matched sub-knowledge graph;

[0027] The user input statement is subjected to theme calculation by using an LDA model, and the probability distribution of each theme of the asset category and asset group to which the statement belongs is counted;

[0028] According to the probability distribution of each theme and the node influence, the node recommendation value is calculated;

[0029] According to the node recommendation value, the node corresponding to the maximum node recommendation value is selected as the recommended asset category and asset group.

[0030] The technical scheme of the embodiment of the present application has at least the following advantages and beneficial effects: the present application discloses a power equipment asset management data similarity calculation recommendation method, which solves the problem of inaccurate automatic recommendation result through natural language processing. The introduction of node influence calculation ensures the robustness of power asset data recommendation, and the combination of theme calculation also makes the recommendation process more flexible and targeted, reduces the output of error information. Finally, the method can be applied in various asset management system scenes, provides a solid technical foundation for improving the management level of enterprises, greatly improves the asset management level, realizes information sharing, improves work efficiency, and provides decision basis. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a flowchart of a power equipment asset management data belonging category and group recommendation method;

[0032] Figure 2 It is a detailed flowchart of a power equipment asset management data belonging category and group recommendation method of the present embodiment;

[0033] Figure 3 It is a power asset standard knowledge mode construction graph;

[0034] Figure 4 A knowledge graph part entity relationship structure schematic diagram is shown.

[0035] Figure 5 A structure schematic diagram of an entity extraction network model is shown. DETAILED DESCRIPTION

[0036] To make the technical solutions of the embodiments of the present application clearer and the advantages more apparent, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0037] As shown in Figures 1-2 A method for recommending categories and groups of power equipment asset management data, comprising the following steps:

[0038] S1, processing the text data of the power asset management system by using natural language technology to obtain effective text data;

[0039] In this embodiment, by extracting the semi-structured data information of the power internal asset management system, the text data is sequentially divided into three levels according to the data structure mode of asset category, asset group and asset name. The three levels are the relationship of belonging to each other, and the asset category corresponds to the first level, the asset group corresponds to the second level, and the asset name corresponds to the third level.

[0040] Some techniques of natural language processing are used on the text data, and the data of the three categories are respectively subjected to word segmentation, stop word removal, keyword extraction and semantic extraction operations. The data after these operations form effective power asset management system text data.

[0041] S2, processing the effective power asset management system text data, and constructing a knowledge graph based on the entity result;

[0042] In this embodiment, the effective power asset management system text data is extracted and processed by knowledge extraction and knowledge fusion technology to provide more accurate and comprehensive knowledge representation. Among them, the knowledge extraction adopts a Bi-LSTM-CNN-CRF model to identify the entities of the power asset management system text; the knowledge fusion includes: based on the self-defined entity alignment rule and the graph editing distance, eliminating the repeated information and improving the integrity of the entity information.

[0043] When constructing the knowledge graph, a standard knowledge mode is first constructed:

[0044] According to the data characteristics of the power asset management system, the standard knowledge mode is defined, according to the data content and characteristics of the three types of data of asset name, asset category and asset group, the mutual relationship of the data is analyzed to construct the standard knowledge mode of the power asset, and the entity and the relationship are defined through the relationship mode between the entities, and the knowledge mode is constructed in the form of triple. The specific operation is as follows: according to the data characteristics of the asset category, the entity is defined as the category object and the category state; according to the data characteristics of the asset group, the entity is defined as the group object and the group state; according to the data characteristics of the asset name, the entity is defined as the name primary object, the name auxiliary object and the name state. At the same time, the relationship between the entities is constructed according to the defined entities, and the specific operation is as follows: the name auxiliary object is “name subordinate” to the name primary object, the name primary object is “group subordinate” to the group object, the group object is “category subordinate” to the category object, all state entities are “state subordinate” to the objects in the hierarchy, and are defined as “name state subordinate”, “group state subordinate” and “category state subordinate”, and the basic construction mode is as shown in Figure 3 .

[0045] For example, the asset name, asset category and asset group of a certain asset are “fish breeding station comprehensive building”, “production house” and “other land” respectively, and the entities defined by the above knowledge mode can divide fish, breeding station and comprehensive building into name state, name primary object and name auxiliary object of asset name, and production, house into category state and category object, and finally other and land into group state and group object.

[0046] Constructing a knowledge graph:

[0047] The basic knowledge graph of the text data of the power asset management system is constructed, according to the characteristics and relationships between the data in the power asset management system, through the definition of the standard knowledge mode, the knowledge is rich and the structure is complex, therefore, the knowledge graph is needed to assist understanding and reasoning, the relationship between the asset name, asset category and asset group is abstracted, the ontology model of the data of the power asset management system is constructed in the form of triple, through the defined standard knowledge mode, the construction of the basic knowledge graph is finally completed. The asset name, asset category and asset group are “fish breeding station box type power distribution device”, “power distribution equipment” and “box type transformer” respectively, and the above example is constructed to construct part of the knowledge graph, and the entity relationship structure diagram is as shown in Figure 4 .

[0048] Finally, the entities and relationships are stored in the Neo4j graph database, the graph database has a natural extension of the graph structure, even if the complexity of the entities and their relationships is constantly improved, it will not affect the search efficiency, and there is no need to perform redundant calculation through the complexity of the connection, and finally the storage data entity is presented in the form of the graph structure.

[0049] S3, calculate the influence of the knowledge graph node according to the topological structure information of the knowledge graph;

[0050] The S3 comprises:

[0051] Extract the relationship between nodes from the knowledge graph to construct a directed graph;

[0052] According to the directed graph, calculate the global influence of the node;

[0053] The global influence is normalized to obtain the normalized global influence.

[0054] The calculation of the global influence of the node uses a first-order Markov chain model:

[0055] The transition matrix of the Markov chain model is composed of a linear combination of two parts, one part is the modified data node correlation weighted transition matrix H', and the other part is the arbitrary data node correlation weighted transition matrix A (expanded from the arbitrary data node correlation vector a), and the linear combination coefficient is the damping factor a (0≤a≤1). The random walk Markov chain has a stationary distribution, denoted as π * . * Determined by the following formula:

[0056] π * =H′π * +(1-)a

[0057] The data node characteristic factor vector can be calculated in combination with the data node correlation weighted transition matrix H':

[0058]

[0059] The formula for calculating the global influence of the node is:

[0060]

[0061] S4, extract the standard entity of the user input asset information, and extract the sub-knowledge graph matched with the standard entity from the knowledge graph;

[0062] In this embodiment, as shown in Figure 5 , the entity extraction network model extracts the standard entity of the user input asset information, wherein the entity extraction network model comprises an input layer, a word embedding layer, a convolution attention layer, a GRU layer, a global attention layer, a CRF layer and an output layer. The input end of the word embedding layer is connected with the input layer, and the output end thereof is connected with the input end of the convolution attention layer. The input end of the GRU layer is connected with the output end of the convolution attention layer, and the output end thereof is connected with the input end of the global attention layer. The input end of the CRF layer is connected with the output end of the global attention layer, and the output end thereof is connected with the output layer.

[0063] According to the extracted relevant standard entity, the most matched sub-graph is obtained, in order to obtain the most matched sub-graph of the user-provided data, the power asset management system standard knowledge graph is pre-trained by using a pre-training method, the information of each node of the graph is converted into a low-dimensional vector, and then the low-dimensional vector is mapped into a feature space. The extracted user input entity content and the node of the standard knowledge graph are matched, the semantic similarity is taken as a judgment basis, and the most matched entity node content is calculated. The similarity algorithm used in the similarity calculation of the above user input data and the standard knowledge graph is cosine similarity, so as to match the entity node with the most similar semantics. The similarity calculation formula is as follows, wherein M A is user input information, and M B is node information.

[0064]

[0065] In order to find the most matched sub-graph, whether there is a node similar to the user input entity information in the neighborhood node set of the entity node is judged by using the similarity. Finally, in the process of obtaining the associated sub-graph, the node filtering method is adopted to screen, so as to associate the sub-graph covering the entity node content and the entity and relationship.

[0066] S5, obtaining recommended asset categories and asset groups according to the probability distribution of the theme of the matched sub-graph and the node influence;

[0067] The S5 comprises:

[0068] Extracting the knowledge graph node influence in the matched sub-graph;

[0069] The user input statement is subjected to theme calculation by using an LDA model, and the probability distribution of each theme of the asset category and the asset group to which the statement belongs is counted;

[0070] According to the probability distribution of each theme and the node influence, the node recommendation value is calculated;

[0071] The node recommendation value = node influence * node theme probability (asset category / asset group);

[0072] According to the node recommendation value, the node corresponding to the maximum node recommendation value is selected as the recommended asset category and asset group.

[0073] In this embodiment, S5 is specifically as follows: when the user inputs new data, the asset category and the asset group are recommended. The most matched standard sub-graph obtained according to the new entity content is subjected to theme calculation, and the comprehensive ranking result is output after combining the node influence, so that the final recommendation is completed.

[0074] The entity sequence obtained after the user inputs data is subjected to theme calculation, the user input data is processed by an LDA theme model to obtain a corresponding theme probability distribution, and the corresponding theme probability distribution is applied to the matched subgraph.

[0075] The final presentation is that when the user adds assets, a standard subgraph is obtained according to a series of operations after inputting new data, the comprehensive ranking result of the subgraph is taken as a target recommendation result, and the information of the asset category and the asset group is recommended.

[0076] The asset information added by the user is taken as a data source of the knowledge graph for real-time updating, the new data information input by the user is taken as a new node of the knowledge graph, meanwhile, the asset category and the asset group information to which the new data information belongs are associated, and the knowledge graph is added, modified and deleted to complete real-time updating of the knowledge graph.

[0077] The technical scheme of the embodiment of the present application has at least the following advantages and beneficial effects: the present application discloses a power equipment asset management data similarity calculation recommendation method, which solves the problems of fuzzy asset relationship definition and inaccurate automatic recommendation result in the field of power assets by natural language processing. The recommendation process combines a standard knowledge mode to form a knowledge graph to present data, accurately defines the data category and relationship, can make the recommendation result more accurate and diverse, and has good interpretability. The introduction of node influence calculation ensures the robustness of power asset data recommendation, and the combination of theme calculation can also make the recommendation process more flexible and targeted, reducing the output of error information. Finally, the method can be applied in various fields of asset management system scenes, provides a solid technical foundation for enterprises to improve management level, greatly improves the asset management level, realizes information sharing, improves work efficiency, and provides a basis for decision-making.

[0078] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of recommending a category and group to which power equipment asset management data belongs, characterized by, The utility model relates to an electric power asset management system based on natural language processing, comprising: processing text data of an electric power asset management system by using natural language technology to obtain effective text data; performing entity processing on the effective text data and constructing a knowledge graph based on a standard knowledge mode and entity results; the construction of the standard knowledge mode comprises: analyzing the mutual relationship of three types of data of asset name, asset category and asset group to construct an electric power asset standard knowledge mode; specifically, according to the data characteristics of the asset category, defining the entity as a category object and a category state; according to the data characteristics of the asset group, defining the entity as a group object and a group state; according to the data characteristics of the asset name, defining the entity as a name primary object, a name auxiliary object and a name state; then constructing the relationship between entities according to the defined entities, specifically: the name auxiliary object is subordinate to the name primary object, the name primary object group is subordinate to the group object, the group object category is subordinate to the category object, and all state entities are state subordinate to the objects in the hierarchy; calculating the influence of the nodes of the knowledge graph according to the topological structure information of the knowledge graph; extracting standard entities of user input asset information and extracting a sub-knowledge graph matched with the standard entities from the knowledge graph; obtaining recommended asset categories and asset groups according to the probability distribution of the theme of the matched sub-knowledge graph and the node influence.

2. The method of claim 1, wherein the data belonging to the category and group of the electric power equipment asset is recommended, characterized by, The natural language processing comprises: word segmentation, keyword extraction, semantic extraction and stop word removal; the entity processing comprises: entity recognition and entity alignment.

3. The method of claim 1, wherein the data belonging to the category and group of the electric power equipment asset is recommended. The calculation of the influence of the nodes of the knowledge graph comprises: extracting the relationship between nodes from the knowledge graph to construct a directed graph; calculating the global influence of the nodes according to the directed graph; normalizing the global influence to obtain normalized global influence.

4. The method of claim 3, wherein the data belonging to the category and group of the electric power equipment asset is recommended, characterized by, The calculation of the global influence of the nodes adopts a first-order Markov chain model.

5. The method of claim 1, wherein the data belonging to the category and group of the electric power equipment asset is recommended, characterized by, An entity extraction network model is adopted to extract standard entities of user input asset information, wherein the entity extraction network model comprises: an input layer, a word embedding layer, a convolution attention layer, a GRU layer, a global attention layer, a CRF layer and an output layer; the input end of the word embedding layer is connected with the input layer, and the output end thereof is connected with the input end of the convolution attention layer; the input end of the GRU layer is connected with the output end of the convolution attention layer, and the output end thereof is connected with the input end of the global attention layer; the input end of the CRF layer is connected with the output end of the global attention layer, and the output end thereof is connected with the output layer.

6. The method of claim 1, wherein the data belonging to the category and group of the electric power equipment asset is recommended. The extraction of the sub-knowledge graph matched with the standard entities from the knowledge graph comprises: calculating the similarity between the standard entities of the user and the node information on the knowledge graph; constructing nodes with a similarity higher than a similarity threshold value as a sub-knowledge graph matched with the standard entities.

7. The method of claim 6, wherein the data belonging to the category and group of the electric power equipment asset is recommended, characterized by, The calculation formula of the similarity is: wherein, is a similarity, is a standard entity of a user, is a node information on a knowledge graph, and is a modulo operation of each vector, is a dot product.

8. The method of claim 1, wherein the data is classified into a category and a group of power equipment assets. The calculation of the recommended asset categories and asset groups according to the probability distribution of the theme of the matched sub-knowledge graph and the node influence comprises: extracting the node influence of the knowledge graph in the matched sub-knowledge graph; calculating the probability distribution of each theme of the asset category and the asset group to which the statement belongs by using an LDA model for theme calculation of the user input statement; calculating the node recommendation value according to the probability distribution of each theme and the node influence. According to the node recommendation value, a node corresponding to the maximum node recommendation value is selected as the recommended asset category and asset group.

Images