Data semantic understanding and intelligent cataloguing method based on power industry data

Abstract

The invention belongs to the technical field of power system data processing, and relates to a data semantic understanding and intelligent cataloguing method based on power industry data, which comprises the following steps: acquiring real-time monitoring data, equipment operation data, historical operation and maintenance data and business system data of a power system; carrying out missing value processing, abnormal value elimination, repeated data deletion and normalization processing on the collected data, carrying out integration through a data fusion algorithm to form a unified data set, carrying out semantic analysis on industry text data by utilizing a natural language processing technology, extracting key entities and relationships, and carrying out semantic analysis on the key entities and relationships; mapping the data into a knowledge graph in combination with a power industry ontology model to generate structured and semantic industry knowledge data; finally, through the semantic analysis result and the knowledge graph, the data classification label is automatically generated, intelligent cataloguing and efficient management are achieved, the intelligent level of data management is improved, the data search efficiency and utilization value are remarkably improved, and the complexity of manual intervention is reduced.

Description

Technical Field

[0001] The present application belongs to the technical field of power system data processing, and more specifically, to a data semantic understanding and intelligent cataloging method based on power industry data. Background Art

[0002] The power industry, as a basic industry of the modern economy, undertakes the key tasks of national energy supply, infrastructure operation and maintenance, and intelligent management. In recent years, with the rapid development of smart grids and digital power systems, the power industry is facing a large number of complex real-time data streams, which come from all aspects of the power system, including power generation, transmission, distribution, and power consumption. The amount of data, data types, and diversity of data sources in the power system have greatly increased the difficulty of data management and utilization.

[0003] In the context of power grids, the power industry is in the process of transforming from traditional power monitoring to intelligent, automated and refined management. The construction of smart grids not only relies on real-time monitoring data of power equipment status, but also requires efficient data fusion, intelligent decision-making and precise operation and maintenance management. In order to support this transformation, the power industry urgently needs an effective way to manage and utilize massive heterogeneous data to improve the intelligence level and safety of the power system. Summary of the invention

[0004] The present invention provides a data semantic understanding and intelligent cataloging method based on power industry data, and intends to provide an intelligent cataloging and management method to achieve rapid search and application of data.

[0005] The data semantic understanding and intelligent cataloging method based on power industry data includes the following steps:

[0006] Step 1: Collect real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system;

[0007] Step 2: Process the missing values ​​of the collected data, remove outliers and duplicate data, normalize the processed data, and then integrate the data into a unified data set through a data fusion algorithm;

[0008] Step 3: Use NLP technology to perform semantic analysis on text data in the power industry and extract entities and relationships; use the professional knowledge of the power industry to design an ontology model for the power industry, and use the semantic analysis results and the ontology model to map entities, attributes and their relationships in the data into a knowledge spectrum to generate structured data for the power industry;

[0009] Step 4: Based on the semantic analysis results and knowledge graph, automatically generate classification labels for the data and manage the data through intelligent cataloging.

[0010] The present invention comprehensively covers the data needs of the power industry by collecting real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system; the collected data is processed for missing values, outliers, duplicate data, and normalized, and integrated into a unified data set through a data fusion algorithm to ensure data quality and consistency; natural language processing (NLP) technology is used to perform semantic analysis on industry text data, extract key entities and relationships, and map the data to the knowledge graph in combination with the power industry ontology model to generate structured and semantic industry knowledge data; finally, data classification labels are automatically generated through semantic analysis results and knowledge graphs to achieve intelligent cataloging and efficient management. The beneficial effect of this technical solution is to enhance the intelligence level of data management, significantly improve data search efficiency and utilization value, reduce the complexity of manual intervention, and support the digital and intelligent business development of the power industry.

[0011] Preferably, step 3 comprises the following steps:

[0012] Text data preprocessing: remove stop words in power text data; split the text into separate words; identify the part of speech of each word through part-of-speech tagging, and finally identify entities in the text;

[0013] Entity and relationship extraction: By analyzing the dependency between words in the text, the semantic relationship between entities is extracted. The Transformer-based relationship extraction model is used to take entities as input to predict the relationship category between entities.

[0014] Ontology modeling and industry knowledge integration: Design an ontology model for the power industry based on the professional knowledge of the power industry, map the entities and relationships obtained in entity and relationship extraction to the ontology model, and generate standardized structured data through ontology reasoning;

[0015] Semantic analysis results and knowledge graph mapping: Use the graph database to store the mapped entities and relationships in the form of triples, use the query language of the graph database to reason and query the knowledge graph; and display the knowledge graph through visualization tools.

[0016] Preferably, the text data preprocessing comprises the following steps:

[0017] Remove stop words: Build a stop word library based on the characteristics of the power industry, traverse each word in the text, and compare it with the words in the stop word library. If the word appears in the stop word library, delete it;

[0018] Word segmentation: Use the Jieba word segmentation tool to break the text into words;

[0019] Part-of-speech tagging: Use Jieba word segmentation for part-of-speech tagging, and assign a part-of-speech tag to each word;

[0020] Named Entity Recognition: Identify entities in text based on open source NER. During training, fine-tune the NER model based on data from the power industry. That is, use labeled power industry text data for training to better identify industry-specific entities and define the type of each entity based on the needs of the power industry.

[0021] Preferably, the specific steps of extracting the semantic relationship between entities by analyzing the dependency relationship between words in the text are as follows:

[0022] Use dependency syntax analysis tools to analyze the text and build a dependency syntax tree. The syntax tree describes the dependency relationship between words in the text. Each node represents a word, and the edge represents the grammatical dependency between words.

[0023] Extract the dependency relations between words from the dependency syntax tree.

[0024] Preferably, the relationship extraction model is as follows:

[0025] Positional encoding: The recognized entity is embedded into the input via positional encoding, which combines the positional encoding with the word vector by adding lines:

[0026]

[0027]

[0028] Where: pos represents the position of the word; i represents the dimension in the word vector; d represents the dimension of the word vector;

[0029] Transformer encoder: Using a pre-trained Transformer encoder, the input text is processed by multiple layers of Transformer encoders, and the relationship between each word in the text is captured through the self-attention mechanism:

[0030] The self-attention mechanism assigns different weights to each token of a given input by calculating attention:

[0031]

[0032] Where: Q is the query matrix, which represents the representation of the current token; K is the key matrix, which represents the representation of all tokens; V is the value matrix, which represents the actual content of all tokens; d k Represents the dimension of the key;

[0033] Relation classification: Based on the encoded representation, entities related to the relationship are extracted, and the contextual representation of the two entities is extracted by marking the locations of the entities;

[0034] Relationship prediction: The context representation is input into a fully connected layer and the relationship type between entities is predicted through the softmax function:

[0035] score(R) = softmax(W · [h entity1 ,h entity2 ]+b);

[0036] Where: W represents the weight matrix; b represents the bias term; h entity1 and h entity2 are the context representations of the two entities respectively; R represents the predicted relationship category;

[0037] The softmax function outputs a probability distribution, which indicates the probability of entities belonging to each relationship category, and the category with the highest probability is taken as the relationship category between the entities.

[0038] Preferably, step 4 comprises the following steps:

[0039] Classification label generation: Based on the extracted entities and relationships, machine learning algorithms are used to classify the data;

[0040] Intelligent catalog management: Map the generated classification labels with the entities in the knowledge graph to form an association between the labels and the data, and use the graph database to store the mapping relationship; set different access rights, storage strategies and processing procedures for different categories of data based on the labels.

[0041] Preferably, the machine learning algorithm includes a graph neural network module, a text encoding module, a model fusion layer and a classification layer;

[0042] The graph neural module is used to process graph structure data and capture the relationship between nodes, wherein the input of the graph neural module is a graph G=(V, E), where: V represents a node set, and E represents an edge set, that is, the relationship between devices;

[0043] Based on the graph neural module, convolution operation is performed to update the representation of each node by propagating information from adjacent nodes:

[0044]

[0045] Where: represents the representation of node v at layer l; W represents the set of adjacent nodes of node v; (l) represents the weight matrix of the lth layer; b (l) represents the bias term; σ represents the activation function;

[0046] The graph neural module is operated through multiple layers of convolution, and the final representation of the node Integrate information from other adjacent nodes to better represent device status and fault information;

[0047] Text encoding module: uses BERT as a text encoder to convert text data into context-related vector representations;

[0048] Model fusion layer: weighted fusion of the graph neural module and the text encoding module to form a unified feature vector output:

[0049] X fused =α·X GNN +(1-α)·X text ;

[0050] Where: X GNN represents the node representation vector generated by the graph neural network; X text Represents the text representation vector generated by the BERT model; α∈[0,1] is the weighting coefficient;

[0051] Classification layer: The fused feature vector X fused Pass it into the fully connected layer for the final classification task:

[0052] y=softmax(W·X fused + b);

[0053] Where: W represents the weight matrix; b represents the bias;

[0054] The model output y is a probability distribution, which indicates the predicted probability of each data category; the data category with the highest probability is selected as the data category.

[0055] The beneficial effects of the present invention include:

[0056] The present invention comprehensively covers the data needs of the power industry by collecting real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system; the collected data is processed for missing values, outliers, duplicate data, and normalized, and integrated into a unified data set through a data fusion algorithm to ensure data quality and consistency; natural language processing (NLP) technology is used to perform semantic analysis on industry text data, extract key entities and relationships, and map the data to the knowledge graph in combination with the power industry ontology model to generate structured and semantic industry knowledge data; finally, data classification labels are automatically generated through semantic analysis results and knowledge graphs to achieve intelligent cataloging and efficient management. The beneficial effect of this technical solution is to enhance the intelligence level of data management, significantly improve data search efficiency and utilization value, reduce the complexity of manual intervention, and support the digital and intelligent business development of the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0058] Figure 1 An overall step block diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0060] See also Figure 1 As shown, the best embodiment of the present invention is further described;

[0061] The data semantic understanding and intelligent cataloging method based on power industry data includes the following steps:

[0062] Step 1: Collect real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system;

[0063] Real-time monitoring data: such as sensor data (current, voltage, load, etc.);

[0064] Equipment operation data: such as equipment status, maintenance records, etc.;

[0065] Historical operation and maintenance data: such as equipment failure records, maintenance plans, etc.

[0066] Business system data: such as power dispatching data, power grid configuration data, etc.

[0067] Step 2: Process the missing values ​​of the collected data, remove outliers and duplicate data, normalize the processed data, and then integrate the data into a unified data set through a data fusion algorithm;

[0068] Step 3: Use NLP technology to perform semantic analysis on text data in the power industry and extract entities and relationships; use the professional knowledge of the power industry to design an ontology model for the power industry, and use the semantic analysis results and the ontology model to map entities, attributes and their relationships in the data into a knowledge spectrum to generate structured data for the power industry;

[0069] Step 4: Based on the semantic analysis results and knowledge graph, automatically generate classification labels for the data and manage the data through intelligent cataloging.

[0070] The present invention comprehensively covers the data needs of the power industry by collecting real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system; the collected data is processed for missing values, outliers, duplicate data, and normalized, and integrated into a unified data set through a data fusion algorithm to ensure data quality and consistency; natural language processing (NLP) technology is used to perform semantic analysis on industry text data, extract key entities and relationships, and map the data to the knowledge graph in combination with the power industry ontology model to generate structured and semantic industry knowledge data; finally, data classification labels are automatically generated through semantic analysis results and knowledge graphs to achieve intelligent cataloging and efficient management. The beneficial effect of this technical solution is to enhance the intelligence level of data management, significantly improve data search efficiency and utilization value, reduce the complexity of manual intervention, and support the digital and intelligent business development of the power industry.

[0071] As a possible implementation of this embodiment, step 3 includes the following steps:

[0072] Text data preprocessing: remove stop words in power text data; split the text into separate words; identify the part of speech of each word through part-of-speech tagging, and finally identify entities in the text;

[0073] Entity and relationship extraction: By analyzing the dependency between words in the text, the semantic relationship between entities is extracted. The Transformer-based relationship extraction model is used to take entities as input to predict the relationship category between entities.

[0074] Ontology modeling and industry knowledge integration: Design an ontology model for the power industry based on the professional knowledge of the power industry, map the entities and relationships obtained in entity and relationship extraction to the ontology model, and generate standardized structure data through ontology reasoning; the specific steps are as follows:

[0075] The ontology model of the power industry mainly includes the following aspects of design:

[0076] Concept Definition:

[0077] Equipment: including substations, power equipment, generator sets, transmission lines, etc.

[0078] Event: For example, equipment failure, equipment maintenance, equipment overhaul, etc.

[0079] State: For example, equipment operation status, equipment failure status, equipment shutdown status, etc.

[0080] Fault Type: such as short circuit fault, equipment aging, equipment damage, etc.

[0081] Relationship definition:

[0082] Fault relationship (hasFault): a relationship in which a device fails.

[0083] Operational state relationship (hasState): The device has a certain operational state.

[0084] Equipment containment relationship (contains): A substation contains multiple devices.

[0085] RelatedTo: For example, the relationship between a device and an event. The maintenance of a device may be related to a certain fault.

[0086] Ontology model example:

[0087] Taking equipment and fault type as an example, the RDF triples of the ontology model are expressed as follows:

[0088] (device, hasFault, fault type)

[0089] (device, hasState, running state)

[0090] (Substation, contains, equipment)

[0091] Among them, hasFault, hasState, and contains are relationship types.

[0092] Ontological reasoning:

[0093] Ontology reasoning is to derive new information based on the rules defined in the ontology model through the reasoning engine. Common reasoning methods include:

[0094] Rule-based reasoning: For example, if the operating status of a device is "faulty", it is inferred that the device is faulty.

[0095] Constraint-based reasoning: If a device is in the "downtime" state after the "device failure" event occurs, it is inferred that the device may need maintenance.

[0096] Examples of inference rules are as follows:

[0097] Rule 1: Equipment status == "fault" ==> Equipment needs repair

[0098] Rule 2: Device status == "Normal" ==> Device can operate

[0099] The calculation formula for inference is:

[0100] if State device = "Fault" then requiresRepair device =Ture;

[0101] Where: State device Indicates the device status; requiresRepair device Indicates whether the device requires maintenance;

[0102] Based on the expertise of the power industry, combined with industry standards and practical experience, industry-specific concepts and rules are added to the ontology model. Common integration steps include:

[0103] Integrate equipment fault handling rules: Define the type of equipment fault, detection method and maintenance steps based on the power industry's operating manual or historical data. For example, "short circuit fault" usually leads to "equipment shutdown", while "equipment overload" may trigger the "overload protection" mechanism.

[0104] Integrate equipment operation and maintenance standards: Introduce equipment operation and maintenance standards and maintenance cycles. For example, define the equipment's regular inspection cycle and maintenance duration based on the equipment's model and age.

[0105] Integration of industry-standard classification systems: Standardize data using terminology and classification systems from power industry standards. For example, map all fault data to standard fault categories such as short circuit, equipment aging, etc.

[0106] By integrating these industry knowledge, the ontology model can more accurately reflect the actual needs of the power industry and make the reasoning results more consistent with industry practice.

[0107] Semantic analysis results and knowledge graph mapping: Use the graph database to store the mapped entities and relationships in the form of triples, use the query language of the graph database to reason and query the knowledge graph; and display the knowledge graph through visualization tools.

[0108] As a possible implementation of this embodiment, the text data preprocessing includes the following steps:

[0109] Remove stop words: Build a stop word library based on the characteristics of the power industry, traverse each word in the text, and compare it with the words in the stop word library. If the word appears in the stop word library, delete it;

[0110] Word segmentation: Use the Jieba word segmentation tool to break the text into words;

[0111] Part-of-speech tagging: Use Jieba word segmentation for part-of-speech tagging, and assign a part-of-speech tag to each word;

[0112] Named Entity Recognition: Identify entities in text based on open source NER. During training, fine-tune the NER model based on data from the power industry. That is, use labeled power industry text data for training to better identify industry-specific entities and define the type of each entity based on the needs of the power industry.

[0113] As a possible implementation of this embodiment, the specific steps of extracting the semantic relationship between entities by analyzing the dependency relationship between words in the text are as follows:

[0114] Use dependency syntax analysis tools to analyze the text and build a dependency syntax tree. The syntax tree describes the dependency relationship between words in the text. Each node represents a word, and the edge represents the grammatical dependency between words.

[0115] Extract the dependency relations between words from the dependency syntax tree.

[0116] The specific steps are as follows:

[0117] The goal of dependency parsing is to establish dependency relationships between each word (usually a lexical unit such as a verb, noun, etc.) and other words in the sentence. Each dependency relationship consists of two elements:

[0118] Head: indicates the word that the current word depends on;

[0119] Subword (dependent): represents a word that depends on the current word.

[0120] The output of dependency syntactic analysis is a dependency syntactic tree, where each node represents a word and the edges represent the dependency relationship between words.

[0121] Word segmentation and part-of-speech tagging:

[0122] Use word segmentation tools (such as spaCy, NLTK, etc.) to segment the input text.

[0123] Part-of-speech tagging (such as noun, verb, adjective, etc.) can be performed on each word using a part-of-speech tagging tool (such as the pos_ attribute in spaCy).

[0124] For example, the sentence "power equipment failure" will be segmented into: ["power equipment", "occurrence", "failure"];

[0125] The results of part-of-speech tagging may be: power equipment (noun), occurrence (verb), failure (noun);

[0126] Build dependencies between words: Use a dependency parser (such as spaCy, StanfordNLP, etc.) to perform dependency analysis on the sentence to obtain the dependency relationship between each word and other words in the sentence. The analyzer returns the parent word (head) and dependency type of each word.

[0127] For example, using spaCy's dependency analyzer, we can get the following dependencies:

[0128] "Electrical equipment" <--(nsubj)--> "Occurrence"

[0129] "Occurrence" <--(prep)--> "failure"

[0130] Here, nsubj represents the subject and prep represents the dependency relationship of the prepositional phrase.

[0131] The dependency syntax tree is an undirected graph structure, where each node represents a word and each edge represents the grammatical dependency between words. Each edge is marked with the type of dependency (such as nsubj, prep, etc.).

[0132] Taking the sentence "power equipment fails" as an example, the structure of the syntax tree is as follows:

[0133] occur

[0134] / \

[0135] Power equipment failure

[0136] Among them, "occurrence" is the head word, and other words such as "power equipment" and "failure" are dependent on "occurrence".

[0137] After constructing the dependency syntax tree, the next task is to extract the grammatical relations between entities (e.g., “power equipment” and “fault”) from the tree.

[0138] Extract dependencies:

[0139] The core of extracting dependency relationships from the dependency syntax tree is to extract parent-child word pairs and the dependency tags between them. The following are common dependency relationship types (syntactic tags):

[0140] nsubj(subject): indicates the dependency relationship between the subject and the predicate verb. For example, "power equipment" is the subject of "occurrence".

[0141] dobj (object): indicates the dependency relationship between a verb and its object. For example, "failure" is the object of "occurrence".

[0142] prep (prepositional phrase): indicates the dependency relationship between a preposition and its object. For example, there may be a prepositional phrase relationship between "occurrence" and "failure".

[0143] amod (adjective modification): indicates the relationship of an adjective modifying a noun.

[0144] The steps to extract relations are as follows:

[0145] Traverse each node (vocabulary) of the dependency syntax tree.

[0146] Dependency pairs are extracted based on the parent word and dependency type of each node.

[0147] For example, using spaCy to analyze the sentence "The power equipment failed", we get the following dependency pairs:

[0148] ("Electrical Equipment", "Occurrence", "nsubj")

[0149] ("occurred", "failed", "dobj")

[0150] Inferring semantic relationships between entities using dependency relations

[0151] Through the extracted dependencies, the semantic relationships between entities can be inferred. For example:

[0152] Relation 1: The relationship between "power equipment" and "occurrence" is subject (nsubj), indicating that "power equipment" is the executor of "occurrence" (i.e., the equipment where the fault occurs).

[0153] Relation 2: The relationship between "occurrence" and "failure" is object (dobj), which means that the action of "occurrence" affects "failure".

[0154] These relations help us understand the connections between different entities in a sentence. For example, through the dependency relationship between "power equipment" and "failure", we can infer the event of equipment failure.

[0155] The dependency syntactic tree can be represented by a graph structure. Assuming we have a sentence containing n words, the dependency tree can be represented as a weighted directed graph G = (V, E), where: V represents the vocabulary set, represents the node, V = {w1, w2, ..., w n}; E represents the set of dependency relationships, represents edges, and each edge e i，j Connecting word w i and word w j , and each edge has a label indicating the dependency type relationship (such as nsubj, dobj, etc.), for example:

[0156] For the sentence "The power equipment fails", the dependency graph can be represented as:

[0157] Node V = {power equipment, occurrence, fault}

[0158] Edge E = {(power equipment, occurrence, nsubj), (occurrence, fault, dobj}.

[0159] As a possible implementation of this embodiment, the relationship extraction model is as follows:

[0160] Positional encoding: The recognized entity is embedded into the input via positional encoding, which combines the positional encoding with the word vector by adding lines:

[0161]

[0162]

[0163] Where: pos represents the position of the word; i represents the dimension in the word vector; d represents the dimension of the word vector;

[0164] Transformer encoder: Using a pre-trained Transformer encoder, the input text is processed by multiple layers of Transformer encoders, and the relationship between each word in the text is captured through the self-attention mechanism:

[0165] The self-attention mechanism assigns different weights to each token of a given input by calculating attention:

[0166]

[0167] Where: Q is the query matrix, which represents the representation of the current token; K is the key matrix, which represents the representation of all tokens; V is the value matrix, which represents the actual content of all tokens; d k Represents the dimension of the key;

[0168] Relation classification: Based on the encoded representation, entities related to the relationship are extracted, and the contextual representation of the two entities is extracted by marking the locations of the entities;

[0169] Relationship prediction: The context representation is input into a fully connected layer and the relationship type between entities is predicted through the softmax function:

[0170] score(R) = softmax(W · [h entity1 ,h entity2 ]+b);

[0171] Where: W represents the weight matrix; b represents the bias term; h entity1 and h entity2 are the context representations of the two entities respectively; R represents the predicted relationship category;

[0172] The softmax function outputs a probability distribution, which indicates the probability of entities belonging to each relationship category, and the category with the highest probability is taken as the relationship category between the entities.

[0173] The step 4 comprises the following steps:

[0174] Classification label generation: Based on the extracted entities and relationships, machine learning algorithms are used to classify the data;

[0175] As a possible implementation of this embodiment, the machine learning algorithm includes a graph neural network module, a text encoding module, a model fusion layer, and a classification layer;

[0176] The graph neural module is used to process graph structure data and capture the relationship between nodes, wherein the input of the graph neural module is a graph G=(V, E), where: V represents a node set, and E represents an edge set, that is, the relationship between devices;

[0177] Based on the graph neural module, convolution operation is performed to update the representation of each node by propagating information from adjacent nodes:

[0178]

[0179] Where: represents the representation of node v at layer l; W represents the set of adjacent nodes of node v; (l) represents the weight matrix of the lth layer; b (l) represents the bias term; σ represents the activation function;

[0180] The graph neural module is operated through multiple layers of convolution, and the final representation of the node Integrate information from other adjacent nodes to better represent device status and fault information;

[0181] Text encoding module: uses BERT as a text encoder to convert text data into context-related vector representations;

[0182] Model fusion layer: weighted fusion of the graph neural module and the text encoding module to form a unified feature vector output:

[0183] X fused =α·X GNN +(1-α)·X text ;

[0184] Where: X GNN represents the node representation vector generated by the graph neural network; X text Represents the text representation vector generated by the BERT model; α∈[0,1] is the weighting coefficient;

[0185] Classification layer: The fused feature vector X fused Pass it into the fully connected layer for the final classification task:

[0186] y=softmax(W·X fused + b);

[0187] Where: W represents the weight matrix; b represents the bias;

[0188] The model output y is a probability distribution, which indicates the predicted probability of each data category; the data category with the highest probability is selected as the data category.

[0189] The loss function of the machine learning algorithm adopts the cross entropy loss function to calculate the difference between the predicted label and the true label.

[0190] Intelligent catalog management: Map the generated classification labels with the entities in the knowledge graph to form an association between the labels and the data, and use the graph database to store the mapping relationship; set different access rights, storage strategies and processing procedures for different categories of data based on the labels.

[0191] As a possible implementation of this embodiment, the intelligent catalog management includes the following steps:

[0192] Data classification and label generation:

[0193] In step 4, we have used machine learning algorithms to classify the data and generated multiple category labels (for example, fault type, equipment status, operation and maintenance category, etc.). These labels will be used to label different types of data and subsequently mapped to the knowledge graph.

[0194] Knowledge graph construction:

[0195] The knowledge graph of the power industry is built based on the results of semantic analysis and ontology models, and includes multiple entities such as equipment, faults, and operation and maintenance operations, and their relationships. For example, the relationship between equipment and faults, the relationship between equipment and operation and maintenance operations, etc. The nodes of the knowledge graph include entities (such as generators, substations, and transmission lines), while the edges represent the semantic relationships between them (such as faults, maintenance, monitoring, etc.).

[0196] Mapping process

[0197] Map the generated classification labels to the entities in the knowledge graph to form an association between the data labels and entities. The specific steps are as follows:

[0198] Semantic mapping: By calculating the semantic similarity between labels and entities (such as through TF-IDF or BERT embedding), the relationship strength between labels and entities in the knowledge graph is calculated, and the labels are mapped to the most relevant entities:

[0199]

[0200] Where: L represents the classification label (such as equipment failure); E represents the entity in the knowledge graph (such as generator); f(L) and f(E) are the vector representations of label L and entity E respectively; ||·|| represents the norm of the vector;

[0201] Through the semantic similarity score, the entity most relevant to the label is selected for mapping.

[0202] Graph database selection:

[0203] When storing these mapping relationships, graph databases (such as Neo4j or ArangoDB) can handle data with rich relationships very efficiently and can flexibly store and query the relationships between labels and entities.

[0204] Mapping storage structure design

[0205] The nodes in the graph database represent the classification labels and entities in the knowledge graph, while the edges represent the mapping relationship between them. For example, the nodes "equipment failure" and "generator" can be connected through an "association" relationship.

[0206] The storage structure is designed as follows:

[0207] Node 1: Label node, representing a classification label (such as “generator failure”).

[0208] Node 2: Entity node, representing an entity in the knowledge graph (such as "generator").

[0209] Edge: represents the association relationship between a label and an entity (such as "belongs to").

[0210] The following relational model can be used in graph databases:

[0211] (Label)→(Entity);

[0212] Graph database operations:

[0213] Use the query language of the graph database (such as Cypher) to store and query mapping relationships. For example, you can use the following Cypher query statement to query which entities a label is mapped to:

[0214] MATCH(l:Label)-[:association]->(e:Entity)

[0215] WHERE l.name = "Generator failure"

[0216] RETURN

[0217] This will return all entities associated with the tag "generator failure".

[0218] Tag-based access rights management

[0219] Different categories of data have different security and access rights requirements. For example, equipment operation logs and fault reports may require strict access control, while some non-sensitive operation data can be openly accessed. Based on the classification label, different data access rights can be set:

[0220] Fault data labels: can only be viewed by specific users (such as operation and maintenance personnel, administrators) and cannot be shared externally.

[0221] Monitoring data tags: can be opened to some users (such as data analysts and business personnel).

[0222] Access rights can be managed using role-based access control (RBAC), where each tag corresponds to an access role, which defines which users can access data of which tags. The relationship between roles and permissions is as follows:

[0223] Permission(T)=f(Role(U),Label(T));

[0224] Where: Role(U) represents the role of user U; Label(T) represents label T; f(·) represents the permission mapping function, which determines whether access is allowed based on the role and label;

[0225] Tag-based storage policy

[0226] Set different data storage strategies based on different tag categories:

[0227] Sensitive data (such as failure data): can be stored using encryption and restrict access to the storage location.

[0228] Non-sensitive data (such as normal operation data): Distributed storage can be used to improve data availability and fault tolerance.

[0229] The design of storage strategy can adopt the following functions:

[0230] StoragePolicy(T)=g(Label(T),Sensitive);

[0231] Where: Sensitive indicates whether it is sensitive data, which is determined by the label category; g(·) represents the storage strategy function, which determines the storage method (such as encrypted storage, distributed storage, etc.).

[0232] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. Data semantic understanding and intelligent cataloging method based on power industry data, characterized by: The following steps are involved: Step 1: Collect real-time monitoring data, equipment operation data, historical operation and maintenance data, and business system data of the power system; Step 2: Process the missing values ​​of the collected data, remove outliers and duplicate data, normalize the processed data, and then integrate the data into a unified data set through a data fusion algorithm; Step 3: Use NLP technology to perform semantic analysis on text data in the power industry and extract entities and relationships; use the professional knowledge of the power industry to design an ontology model for the power industry, and use the semantic analysis results and the ontology model to map entities, attributes and their relationships in the data into a knowledge spectrum to generate structured data for the power industry; Step 4: Based on the semantic analysis results and knowledge graph, automatically generate classification labels for the data and manage the data through intelligent cataloging.

2. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 1 is characterized in that: The step 3 comprises the following steps: Text data preprocessing: remove stop words in power text data; split the text into separate words; identify the part of speech of each word through part-of-speech tagging, and finally identify entities in the text; Entity and relationship extraction: By analyzing the dependency between words in the text, the semantic relationship between entities is extracted. The Transformer-based relationship extraction model is used to take entities as input to predict the relationship category between entities. Ontology modeling and industry knowledge integration: Design an ontology model for the power industry based on the professional knowledge of the power industry, map the entities and relationships obtained in entity and relationship extraction to the ontology model, and generate standardized structured data through ontology reasoning; Semantic analysis results and knowledge graph mapping: Use the graph database to store the mapped entities and relationships in the form of triples, use the query language of the graph database to reason and query the knowledge graph; and display the knowledge graph through visualization tools.

3. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 2 is characterized in that: The text data preprocessing comprises the following steps: Remove stop words: Build a stop word library based on the characteristics of the power industry, traverse each word in the text, and compare it with the words in the stop word library. If the word appears in the stop word library, delete it; Word segmentation: Use the Jieba word segmentation tool to break the text into words; Part-of-speech tagging: Use Jieba word segmentation to perform part-of-speech tagging, and assign a part-of-speech tag to each word; Named Entity Recognition: Identify entities in text based on open source NER. During training, fine-tune the NER model based on data from the power industry. That is, use labeled power industry text data for training to better identify industry-specific entities and define the type of each entity based on the needs of the power industry.

4. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 2 is characterized in that: The specific steps of extracting the semantic relationship between entities by analyzing the dependency relationship between words in the text are as follows: Use dependency syntax analysis tools to analyze the text and build a dependency syntax tree. The syntax tree describes the dependency relationship between words in the text. Each node represents a word, and the edge represents the grammatical dependency between words. The dependency relations between words are extracted from the dependency syntax tree.

5. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 2 is characterized in that: The relationship extraction model is as follows: Positional encoding: The recognized entity is embedded into the input via positional encoding, which combines the positional encoding with the word vector by adding lines: Where: pos represents the position of the word; i represents the dimension in the word vector; d represents the dimension of the word vector; Transformer encoder: Using a pre-trained Transformer encoder, the input text is processed by multiple layers of Transformer encoders, and the relationship between each word in the text is captured through the self-attention mechanism: The self-attention mechanism assigns different weights to each token of a given input by calculating attention: Where: Q is the query matrix, which represents the representation of the current token; K is the key matrix, which represents the representation of all tokens; V is the value matrix, which represents the actual content of all tokens; d k Represents the dimension of the key; Relation classification: Based on the encoded representation, entities related to the relationship are extracted, and the contextual representation of the two entities is extracted by marking the locations of the entities; Relationship prediction: The context representation is input into a fully connected layer and the relationship type between entities is predicted through the softmax function: score(R)=softmax(W·[h entity1 ,h entity2 ]+b); Where: W represents the weight matrix; b represents the bias term; h entity1 and h entity2 are the context representations of the two entities respectively; R represents the predicted relationship category; The softmax function outputs a probability distribution, which indicates the probability of entities belonging to each relationship category, and the category with the highest probability is taken as the relationship category between the entities.

6. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 1 is characterized in that: The step 4 comprises the following steps: Classification label generation: Based on the extracted entities and relationships, machine learning algorithms are used to classify the data; Intelligent catalog management: Map the generated classification labels with the entities in the knowledge graph to form an association between the labels and the data, and use the graph database to store the mapping relationship; set different access rights, storage strategies and processing procedures for different categories of data based on the labels.

7. The data semantic understanding and intelligent cataloging method based on power industry data according to claim 6 is characterized in that: The machine learning algorithm includes a graph neural network module, a text encoding module, a model fusion layer, and a classification layer; The graph neural module is used to process graph structure data and capture the relationship between nodes, wherein the input of the graph neural module is a graph G=(V, E), where: V represents a node set, and E represents an edge set, that is, the relationship between devices; Based on the graph neural module, convolution operation is performed to update the representation of each node by propagating information from adjacent nodes: Where: represents the representation of node v at layer l; W represents the set of adjacent nodes of node v; (l) represents the weight matrix of the first layer; b (l) represents the bias term; σ represents the activation function; The graph neural module is operated through multiple layers of convolution, and the final representation of the node Integrate information from other adjacent nodes to better represent device status and fault information; Text encoding module: Use BERT as a text encoder to convert text data into context-related vector representations; Model fusion layer: weighted fusion of the graph neural module and the text encoding module to form a unified feature vector output: X fused =α·X GNN +(1-a)·X text ; Where: X GNN represents the node representation vector generated by the graph neural network; X text Represents the text representation vector generated by the BERT model; α∈[0,1] is the weighting coefficient; Classification layer: The fused feature vector X fused Pass it into the fully connected layer for the final classification task: y=softmax(W·X fused +b) Where: W represents the weight matrix; b represents the bias; The model output y is a probability distribution, which indicates the predicted probability of each data category; the data category with the highest probability is selected as the data category.

Images