Construction method and system of power transformer equipment defect knowledge graph and storage medium

Abstract

The invention discloses a construction method and system of a power transformer equipment defect knowledge graph and a storage medium. The method comprises the following steps: acquiring structured data and unstructured data of transformer equipment operation defects; constructing an ontology layer of the transformer equipment defects, and carrying out structured expression on the relationship among all ontologies in the ontology layer; taking the constructed ontology layer as a reference, extracting the structured data to obtain a first target triple, and extracting the unstructured data to obtain a second target triple; and performing knowledge fusion on the first target triad and the second target triad to obtain transformer equipment defect map data. According to the method, the accuracy of entity recognition and relation recognition can be improved, fusion of multi-dimensional data such as equipment basic information, operation state information and defect information is better realized, so that operation and maintenance personnel are more effectively helped to quickly retrieve, position and process transformer defects, and the workload of the operation and maintenance personnel is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, relates to the field of electric power operation and inspection, and in particular to a method, system and storage medium for constructing a domain knowledge graph of electric power equipment defects. Background Art

[0002] Equipment defect management is based on the "State Grid Corporation Power Grid Equipment Defect Management Regulations" to record and review all kinds of main and auxiliary equipment defects found in inspections, inspections, maintenance, online monitoring, and monitoring, as well as equipment defects automatically identified and pushed, to track the entire process of power grid equipment from defect discovery to defect elimination, and help relevant staff to deal with equipment defects. Among them, transformers, as important equipment in the power system, have more complex functions than before. Their daily defect diagnosis relies too much on the experience accumulation and knowledge reserves of operation and maintenance personnel, and requires continuous memory and query of knowledge. It is difficult to ensure the accuracy and standardization of processing, resulting in low work efficiency.

[0003] Therefore, through the definition of transformer defect knowledge graph, the fusion of multi-dimensional data such as equipment basic information, operating status information, defect information, etc. is realized, which helps the operation and maintenance personnel to quickly retrieve, locate and process transformer defects, and reduce the workload of the operation and maintenance personnel. At the same time, the transformer defect knowledge graph can provide a high-quality data base for subsequent diagnosis, detection, early warning and other applications, which is one of the foundations for further realizing the intelligence of the power field.

[0004] Prior art 1 (CN116993043A) discloses a method and device for tracing the source of power equipment faults. The method includes the steps of extracting information from fault data to obtain triple data; importing triple data into a graph database to construct a knowledge graph of power equipment faults; designing and constructing a power equipment fault tracing model based on a fast graph convolutional network; inputting new power equipment fault data into the constructed power equipment fault tracing model to obtain historical fault cases and disposal measures that are the same or similar to the new power equipment fault data. Among them, this technology uses three serial models, BERT, BiLSTM, and CRF, to jointly extract entities, attributes, and relationships. However, the use of the above-mentioned serial model cannot capture the semantic structure feature information of text words in the power field, resulting in low entity recognition accuracy, which will reduce the quality of knowledge graph construction.

[0005] Prior art 2 (CN114168740A) discloses a transformer concurrent fault diagnosis method based on graph convolutional neural network and knowledge graph. First, a transformer fault knowledge graph is constructed to highlight the correlation of the faulty equipment or components; then a lightweight graph convolutional neural network model is built, combined with the transformer fault knowledge graph, and a corresponding graph convolutional neural network is designed, making full use of its adjacency matrix to characterize the topological information between single fault nodes, effectively mining the correlation between each single fault, and establishing the mapping relationship between each faulty equipment or component in the concurrent fault through a deep graph convolutional architecture, so as to fully model the concurrent fault and determine the equipment or component where the concurrent fault occurs. However, this technology uses BiLSTM+CRF for entity recognition, which can neither further capture global information nor local information for the power field text, so that the entity recognition accuracy is not high. At the same time, this technology uses BiGRU+Attention for relationship recognition, and the accuracy of capturing the relationship within the input text sequence is not high enough, making it difficult to achieve accurate classification of the relationship. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned status quo, the purpose of the present invention is to provide a method, system and storage medium for constructing a knowledge graph of power transformer equipment defects, so as to further improve the accuracy of entity recognition and relationship recognition, thereby helping operation and maintenance personnel to detect and handle equipment defects better and more accurately through the knowledge graph.

[0007] To achieve the above and other purposes, according to a first aspect of the present invention, a method for constructing a knowledge graph of power transformer equipment defects is provided. The method comprises:

[0008] Collect structured and unstructured data of transformer equipment operation defects;

[0009] Construct the ontology layer of transformer equipment defects and express the relationship between each ontology in the ontology layer in a structured manner;

[0010] Taking the constructed ontology layer as a reference, the structured data is extracted to obtain the first target triples, and the unstructured data is extracted to obtain the second target triples;

[0011] The knowledge of the first target triplet and the second target triplet is fused to obtain the transformer equipment defect map data.

[0012] Preferably, the collecting of structured data and unstructured data of transformer equipment operation defects includes:

[0013] Extracting the structured data from the equipment defect record table of the power grid resource business center;

[0014] The unstructured data is collected from a power grid production management system; wherein the unstructured data includes long text, pictures, audio and / or video associated with the transformer.

[0015] Preferably, the step of constructing an ontology layer of transformer equipment defects and expressing the relationship between the ontologies in the ontology layer in a structured manner includes:

[0016] Determine the equipment-component body according to the equipment structure of the transformer;

[0017] Determine the defect information ontology based on the defect standard library;

[0018] The physical relationship between the components is used to determine the relationship between the entities, and the attributes of the entities are determined according to the operating characteristics of the entities; wherein the physical relationship includes a structural association relationship and a functional association relationship.

[0019] Preferably, the extracting of the structured data to obtain the first target triples and the extracting of the unstructured data to obtain the second target triples with reference to the constructed ontology layer includes:

[0020] The D2RQ tool is used to convert the structured data into the first target triples with reference to the ontology layer.

[0021] Preferably, the extracting of the structured data to obtain the first target triples and the extracting of the unstructured data to obtain the second target triples with reference to the constructed ontology layer includes:

[0022] Referring to the ontology layer, an entity recognition model based on BERT and EE-GCN is used to extract entities and target attributes from unstructured data, and target entity data and target attribute data are obtained respectively;

[0023] Referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data;

[0024] The target entity data, the target relationship data and the target attribute data are converted into a second target triple.

[0025] Preferably, an entity recognition model based on BERT and EE-GCN is used to extract entities from unstructured data to obtain target entity data, including:

[0026] The BERT model is used to encode the equipment defect text in the input unstructured data:

[0027] V=V Token +V Segment +V Position

[0028] Where V is the coded representation of the equipment defect text, V Token Represents the vector representation of the current element, V Segment Indicates the sentence where the current element is located in the equipment defect text, V Position Indicates the specific position of the current element in the device defect text;

[0029] Generate semantic dependency analysis tree of equipment defect text;

[0030] Based on the encoding result output by the BERT model, the generated semantic dependency analysis tree is processed by the EE-GCN model, and the label classification probability distribution of each node in the semantic dependency analysis tree is obtained by referring to the ontology layer;

[0031] According to the label classification probability distribution, the optimal label is determined through the CRF layer to obtain the target entity data and target attribute data:

[0032]

[0033] in, represents the total score of each label sequence, represents the set of all possible labels, express The label classification sequence in represents the optimal label.

[0034] Preferably, based on the encoding result output by the BERT model, the generated semantic dependency analysis tree is processed by the EE-GCN model, and the label classification probability distribution of each node in the semantic dependency analysis tree is obtained by referring to the ontology layer, including:

[0035] Using the node update module in the EE-GCN model, the feature vector of the node in the current layer is generated by aggregating semantically connected nodes of specific dependency types in the semantic dependency analysis tree. The specific calculation formula is as follows:

[0036]

[0037] in, It represents the oth channel of E, W z is a learnable parameter matrix; Represents the structure of the semantic dependency analysis tree, which is represented by a three-dimensional matrix. n represents the number of node trees in the analysis tree, and p represents the relationship dimension. Pool() is a function for creating a process pool. ReLU() is an activation function. N is the total number of channels. H represents the hidden state representation of the lth layer of the oth channel; l-1 represents the hidden state representation of the l-1th layer, and the hidden state representation of the first layer is H1 Obtained from the output value of the BERT model; H l represents the hidden state representation of the lth layer, H l is in matrix form, and H l Each row in is the feature vector of the corresponding node.

[0038] By using the relationship update module in the EE-GCN model, the latest feature representation of the relationship in the semantic dependency analysis tree is updated by mining the clues hidden in the dependency edges. The specific formula is as follows:

[0039]

[0040] in, represents the splicing operation, W u represents the learnable weight matrix, is the updated feature vector of node i and node j at layer l, is the relationship feature vector between node i and node j in the current layer, is the relationship feature vector between node i and node j in the previous layer;

[0041] Referring to the ontology layer, the softMax layer is used to calculate the label classification probability distribution of each node in the semantic dependency analysis tree. The specific formula is as follows:

[0042]

[0043] Among them, W t represents the mapping matrix, b t represents the offset, and h is the feature vector of the node.

[0044] Preferably, referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract the relation of the unstructured data to obtain the target relation data, including:

[0045] The BiGRU model is used to process the equipment defect text in the unstructured data to obtain the final state representation of each element in the equipment defect text:

[0046]

[0047] in, and Represent the hidden state of the current layer from left to right and from right to left, respectively. and Represent the hidden state of the previous layer from left to right and from right to left, GRU represents the gated recurrent unit, x t represents the tth element in the device defect text, [:] represents the concatenation operation of the vector, ht It represents the final state of the current layer;

[0048] The final state representation of each element is input into the Self-Attention layer to obtain the feature vector of each node. The specific formula is as follows:

[0049] Q=W q H

[0050] K=W k H

[0051] V=W v H

[0052]

[0053] Among them, W q , W k , W v They represent the query matrix, matching matrix, and extracted information matrix respectively. H represents the final state of each element and represents the concatenated matrix. Q, K, and V represent the query vector, key vector, and value vector respectively. represents the dimension of K; H′ is an eigenvector matrix, each row of which corresponds to the eigenvector of each node;

[0054] The feature vectors of each node are concatenated and processed through the linear layer to obtain the target relationship data. The specific formula is as follows:

[0055] c=[h1′:h2′,...:h n ′]W m

[0056] Among them, c is the relationship recognition result, h1′, h2′, ..., h n ′ are the feature vectors of the nodes in the corresponding rows in H′, W m represents the linear transformation matrix, and m represents the number of relationships.

[0057] Recursive training is performed using a cross entropy loss function so that the difference between the relationship recognition result and the target relationship value is minimized or less than a preset value, and the trained relationship recognition result is used as the target relationship data; wherein the target relationship value is obtained by referring to the relationship in the ontology layer.

[0058] Preferably, the first target triplet and the second target triplet are fused to obtain transformer equipment defect map data, including:

[0059] The target entities and target attributes in the first target triplet and the second target triplet are coreference resolved to unify the data of defective equipment and defect information and form standardized transformer equipment defect map data.

[0060] According to a second aspect of the present invention, a system for constructing a transformer equipment defect knowledge graph is provided. The system comprises:

[0061] A collection module is used to collect structured and unstructured data of transformer equipment operation defects;

[0062] The ontology layer construction module is used to construct the ontology layer of transformer equipment defects and to express the relationship between the various ontologies in the ontology layer in a structured manner;

[0063] An extraction module, used to extract the structured data to obtain the first target triples and to extract the unstructured data to obtain the second target triples with reference to the constructed ontology layer;

[0064] The fusion module is used to fuse the first target triplet and the second target triplet to obtain transformer equipment defect map data.

[0065] Preferably, the ontology layer building module is also used for:

[0066] Determine the equipment-component body according to the equipment structure of the transformer;

[0067] Determine the defect information ontology based on the defect standard library;

[0068] The physical relationship between the components is used to determine the relationship between the entities, and the attributes of the entities are determined according to the operating characteristics of the entities; wherein the physical relationship includes a structural association relationship and a functional association relationship.

[0069] Preferably, the extraction module further comprises a structured knowledge extraction unit for converting the structured data into first target triples by using a D2RQ tool with reference to the ontology layer.

[0070] Preferably, the extraction module further includes an unstructured knowledge extraction unit, which is used to:

[0071] Referring to the ontology layer, an entity recognition model based on BERT and EE-GCN is used to extract entities and target attributes from unstructured data, and target entity data and target attribute data are obtained respectively;

[0072] Referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data;

[0073] The target entity data, the target relationship data and the target attribute data are converted into a second target triple.

[0074] Preferably, the fusion module is also used to perform co-reference resolution on the target entities and target attributes in the first target triplet and the second target triplet, to unify the data of defective equipment and defect information, and to form standardized transformer equipment defect map data.

[0075] Preferably, the system also includes a storage module for storing transformer equipment defect map data in a graph database.

[0076] According to a third aspect of the present invention, there is provided a storage medium of a transformer equipment defect knowledge graph, storing a computer program that can be loaded by a processor and execute the method according to the first aspect of the present invention.

[0077] The beneficial effect of the present invention is that, compared with the prior art,

[0078] 1. By collecting the structured data and unstructured data of transformer equipment operation defects respectively, and extracting knowledge from the structured and unstructured data with reference to the constructed ontology layer, the first and second target triples are obtained, and then the first and second target triples are fused to obtain transformer defect map data, which can improve the integrity of the data and make the obtained information more comprehensive, so as to more comprehensively describe the characteristics and conditions of transformer equipment operation defects.

[0079] 2. Extract the text in the field of power transformers through BERT+EE-GCN. First, capture the global information of the text through BERT, then perform semantic dependency tree analysis on the text of power transformers, and extract the semantic structure information between words on the dependency tree through the EE-GCN model, thereby strengthening the local information extraction of the text of power transformers. In this way, the entity recognition technology of the present invention can not only obtain a relatively high accuracy rate in the field of power transformer texts, but also in other power equipment fields (disconnectors, circuit breakers, etc.), and the model can be migrated.

[0080] 3. By using BiGRU+Self-Attention for relationship recognition, the relationship between each element of the input power field text sequence and all other elements can be calculated, the relationship within the input sequence can be captured, the relationship classification can be better performed, and the accuracy of relationship recognition can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is the overall framework of the method for constructing the transformer defect knowledge graph of the present invention;

[0082] Figure 2 It is the defect ontology layer of the transformer defect knowledge graph constructed by the present invention;

[0083] Figure 3It is the entity recognition model based on BERT and EE-GCN of the present invention;

[0084] Figure 4 is a semantic dependency analysis tree of the transformer equipment defect text of the present invention;

[0085] Figure 5 It is the BiGRU+Self-Attention-based relationship recognition model of the present invention;

[0086] Figure 6 It is a system structure diagram for constructing the transformer defect knowledge graph of the present invention. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0088] The embodiment of the present application provides a method, system and storage medium for constructing a knowledge graph of defects in power transformer equipment. This helps maintenance personnel to quickly retrieve, locate and process transformer defects, thereby reducing their workload. At the same time, the transformer defect knowledge graph can provide a high-quality data base for subsequent diagnosis, detection, early warning and other applications, and is one of the foundations for further realizing the intelligence of the power field.

[0089] Example 1

[0090] Embodiment 1 relates to a method for constructing a knowledge graph of defects of a power transformer device according to the present invention. Figure 1 This is the overall process framework diagram of the method, which specifically includes the following steps:

[0091] S1. Collect structured and unstructured data of transformer equipment operation defects.

[0092] Step S1 specifically includes the following steps:

[0093] S1.1. Extract the structured data from the equipment defect record table of the power grid resource business center;

[0094] S1.2, collecting the unstructured data from the power grid production management system;

[0095] Among them, unstructured data may include long text, pictures, audio / video associated with transformer equipment defects.

[0096] S2. Construct the ontology layer of transformer equipment defects and express the relationship between each ontology in the ontology layer in a structured manner.

[0097] Step S2 specifically includes the following steps:

[0098] S2.1. Determine the equipment-component ontology according to the equipment structure of the transformer, and determine the defect information ontology according to the defect standard library.

[0099] The specific division of transformer equipment defect ontology (Schema) layer is as follows Figure 2 As shown, transformer equipment is subdivided into oil-immersed transformers, dry-type transformers, and SF6 transformers; the subordinate components of the transformer include iron core, oil tank, gas relay, etc.; the defect information categories corresponding to the components include defect description, defect content, defect elimination methods, etc.

[0100] S2.2, after the equipment body is constructed, the physical relationships such as the structural relationship and functional relationship between the components are used to determine the relationship between the bodies, and the attributes of the bodies are determined according to the operating characteristics of the bodies. The specific relationships and attributes are shown in Table 1 and Table 2 respectively:

[0101] Table 1 Ontology relationship types

[0102]

[0103] Table 2 Ontology attributes

[0104]

[0105]

[0106] S3. Taking the constructed ontology layer as a reference, extract the structured data to obtain the first target triples, and extract the unstructured data to obtain the second target triples.

[0107] The step of extracting the structured data to obtain the first target triplet includes: using a D2RQ tool to convert the structured data into the first target triplet with reference to the ontology layer.

[0108] The structured data of equipment defects are mainly reflected in database tables. Therefore, for the extraction of structured data, the data to be extracted from the equipment defect record table is screened, and then the public D2RQ tool is used to convert the relational data into the first target triple data.

[0109] Extracting the unstructured data to obtain the second target triplet specifically includes the following steps:

[0110] S3.1. Referring to the ontology layer, the entity recognition model based on BERT and EE-GCN is used to perform entity extraction and target attribute extraction on unstructured data to obtain target entity data and target attribute data respectively.

[0111] like Figure 3 , step S3.1 specifically includes the following steps:

[0112] S3.1.1. Use the BERT model to encode the equipment defect text in the input unstructured data. The specific formula is as follows:

[0113] V=V Token +V Segment +V Position

[0114] Among them, V represents the vector representation of the current element after encoding, V Token Represents the vector representation of the current element, V Segment The vector representation of which sentence in the text the current element belongs to, V Position A vector representation of the specific position of the current element in the text;

[0115] S3.1.2. Generate a semantic dependency analysis tree for the equipment defect text.

[0116] The generation of semantic dependency analysis can be realized by using the HanLP open source tool. As an example, by using the HanLP open source tool, the semantic dependency analysis tree formed by the text "Manual inspection found that the gas humidity in the C phase gas chamber of the 3rd main transformer of Xiaoxiang Station exceeded the standard" is as follows: Figure 4 shown.

[0117] S3.1.3. Based on the encoding results output by the BERT model, the generated semantic dependency analysis tree is processed by the EE-GCN model, and the label classification probability distribution of each node in the semantic dependency analysis tree is obtained by referring to the ontology layer.

[0118] Among them, a single node in the semantic dependency analysis tree is a single word. The label classification probability distribution needs to refer to the entities and attributes in the ontology layer.

[0119] The S3.1.3 includes the following steps:

[0120] The EE-GCN model is divided into a node update module and a relationship update module. The two modules complement each other and further utilize the semantic structure information and context information of the text. The specific steps are as follows:

[0121] S3.1.3.1. Using the node update module in the EE-GCN model, the node representation of the current layer is generated by aggregating semantically connected nodes of specific dependency types in the semantic dependency analysis tree.

[0122] Among them, the structure of the semantic dependency analysis tree is It is represented by a three-dimensional matrix, where n represents the number of node trees in the analysis tree and p represents the relationship dimension. The specific calculation formula is as follows:

[0123]

[0124] in, It represents the oth channel of E, W z is a learnable parameter matrix; Represents the structure of the semantic dependency analysis tree, which is represented by a three-dimensional matrix. n represents the number of node trees in the analysis tree, and p represents the relationship dimension. Pool() is a function for creating a process pool. ReLU() is an activation function. N is the total number of channels. H represents the hidden state representation of the lth layer of the oth channel; l-1 represents the hidden state representation of the l-1th layer, and the hidden state representation of the first layer is H 1 Obtained from the output value of the BERT model; H l represents the hidden representation of the lth layer, H l is in matrix form, and H l Each row in is the feature vector of the corresponding node.

[0125] S3.1.3.2. Using the relationship update module in the EE-GCN model, we update the latest feature representation of the relationship in the semantic dependency analysis tree by mining the clues hidden in the dependency edges. The specific formula is as follows:

[0126]

[0127] in, and are the relationship feature representations between nodes i and j corresponding to the lth layer and the l-1th layer respectively; Indicates the splicing operation; W u represents the learnable weight matrix; H l The feature vectors of nodes i and j;

[0128] S3.1.3.3 uses the SoftMax layer to calculate the label classification probability distribution of each node in the semantic dependency analysis tree. The specific formula is as follows:

[0129]

[0130] Among them, W t represents the mapping matrix, b trepresents the offset, h represents the feature vector of the node, and SoftMax() is the normalized exponential function.

[0131] S3.1.4. According to the label classification probability distribution, the optimal label is determined through the CRF layer to obtain the target entity data and target attribute data.

[0132] The specific formula is as follows:

[0133]

[0134] in, represents the total score of each label sequence, represents the set of all possible labels, express The label classification sequence in represents the optimal label.

[0135] As an example, when performing entity extraction on the text "Manual inspection found that the gas humidity in the C phase gas chamber of the main transformer of Xiaoxiang Station 3 exceeded the standard", the optimal label classification sequence is (O,O,O,O,O,O,B-sf6,I-sf6,I-sf6,I-sf6,I-sf6,I-sf6,I-sf6,I-sf6,O,O,O,O,B-des,I-des,I-des,I-des,I-des,I-des,I-des,O,O,O), thus obtaining the target entity data (Xiaoxiang Station 3 main transformer C phase, gas humidity exceeded the standard).

[0136] Attributes, as supplementary knowledge of entities, can also be extracted as entity knowledge. Therefore, similarly, the above-mentioned entity extraction steps can be referred to, and the entity recognition model based on Bert and EE-GCN can also be used to extract attributes from equipment defect texts. For example, the equipment defect text "It was found that the voltage level of the C phase of the main transformer of Xiaoxiang Station 3 was AC 1000kv" was obtained through model recognition. The optimal label classification sequence (O, O, O, O, O, B-sf6, I-sf6, I-sf6, I-sf6, I-sf6, I-sf6, I-sf6, I-sf6, O, O, O, O, O, B-vol, I-vol, I-vol, I-vol) was obtained, thereby obtaining the target attribute data (Xiaoxiang Station 3 Main Transformer C Phase, Voltage Level, 1000kv).

[0137] Among them, "O", "B", "I", "sf6", "B-sf6" and so on in the above label classification sequence come from BIO labels (begin, inside, outside), O means irrelevant to the target entity or target attribute, B means the beginning of the target entity or target attribute, and I means the end of the target entity or target attribute. The target entity and target attribute are both referenced to the ontology and attributes corresponding to the entity in the constructed ontology layer. sf6 indicates the type of transformer, and B-sf6 is a label used for model training.

[0138] S3.2, referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data;

[0139] like Figure 5 , step S3.2 specifically includes the following steps:

[0140] S3.2.1. Use the BiGRU model to process the equipment defect text in the unstructured data to obtain the final state representation of each element in the equipment defect text. The specific formula is as follows:

[0141]

[0142] in, and Represent the hidden state of the current layer from left to right and from right to left, respectively. and Represent the hidden state of the previous layer from left to right and from right to left, GRU represents the gated recurrent unit, x t represents the tth element in the defect text, and then concatenates the hidden states in the two directions to obtain the final state representation h of the current layer t ;

[0143]

[0144] Among them, [:] represents the concatenation operation of vectors;

[0145] S3.2.2. Input the final state representation of each element into the Self-Attention layer to obtain the feature vector of each node. The specific formula is as follows:

[0146] Q=W q H

[0147] K=W k H

[0148] V=W v H

[0149] Among them, Wq , W k , W v They represent the query matrix, the matching matrix, and the extracted information matrix respectively, and H represents the final state of each element and the concatenated matrix;

[0150]

[0151] in, represents the dimension of K, H′ represents the final calculation result, which is a two-dimensional eigenvector matrix, and each row of the eigenvector matrix corresponds to the eigenvector of each node.

[0152] S3.2.3 concatenates the feature vectors of each node and processes them through a linear layer to obtain the relationship recognition result.

[0153] The specific formula is as follows:

[0154] c=[h1′:h2′,...:h n ′]W m

[0155] Where c represents the relationship recognition result; W m represents the feature transformation matrix, h1′, h2′, ..., h n ′ are the feature vectors of the nodes in the corresponding rows in H′.

[0156] S3.2.4 Use the cross entropy loss function to perform recursive training so that the difference between the relationship recognition result and the target relationship value is minimized or less than a preset value, and use the trained relationship recognition result as the target relationship data; wherein the target relationship value is obtained by referring to the relationship in the ontology layer.

[0157] Among them, obtaining the target relationship value according to the relationship in the ontology layer can be achieved by data annotation of the relationship in the ontology layer. Data annotation includes defining relationship labels, annotation formats, and annotation tools according to a set of annotation specifications to build a set of all possible relationship labels. For each training instance, its target relationship value can be determined according to the annotation results. This is usually to select one or more labels from the relationship label set to represent the relationship type existing in the instance.

[0158] As an example, for the equipment defect text "Manual inspection found that the gas humidity in the C-phase gas chamber of the Xiaoxiang Station 3 main transformer exceeded the standard", the BiGRU+Self-Attention relation extraction model was used to extract the relationship of the equipment defect text data in the unstructured data, and the relation data "there is a defect" was obtained.

[0159] S3.3, converting the target entity data, the target relationship data and the target attribute data into a second target triple;

[0160] In step S3.5, the data extracted above are converted into the second target triplet, namely the entity relationship triplet (Xiaoxiang Station 3 main transformer C phase, defects, gas humidity exceeds the standard) and the entity attribute triplet (Xiaoxiang Station 3 main transformer C phase, voltage level, 1000kv).

[0161] S4. Perform knowledge fusion on the first target triplet and the second target triplet to obtain transformer equipment defect map data.

[0162] Unify the data of defective equipment and defect information to form standardized map data.

[0163] In S4, since the naming of transformer equipment is relatively standardized and entity ambiguity rarely occurs in the present invention, coreference resolution of transformer equipment defect entities is performed.

[0164] S5. Store the transformer equipment defect map data into the graph database.

[0165] Among them, the graph database neo4j is preferably used.

[0166] In order to better demonstrate the effect of the method of the present invention, a simulation experiment was carried out, and the specific test results are shown in the following table:

[0167] Table 3 Entity recognition experimental results

[0168] Model Accuracy Recall F1 value BERT+EE-GCN 0.9342 0.9148 0.9125 BERT+BiLSTM+CRF 0.9135 0.9341 0.9016 BiLSTM+CRF 0.8976 0.9026 0.8979

[0169] As shown in the table above, the BERT+EE-GCN model of the present invention has achieved better performance in all indicators. The BERT+BiLSTM+CRF model has similar performance to the present technology, but the model does not further extract the structural semantic feature information between words, so it cannot achieve better results.

[0170] Table 4. Experimental results of relation recognition

[0171] Relationship Category Accuracy Recall F1 value belong 0.9021 0.8902 0.8513 Defects 0.9389 0.7652 0.9234 deal with 0.9234 0.9213 0.8521 touch 0.9102 0.9352 0.9053 Adjacent 0.9230 0.9123 0.9528 touch 0.9189 0.9013 0.9136 connect 0.9230 0.8991 0.9048

[0172] The recognition accuracy of the BiGRU+Self-Attention model of the present invention in each category of relations is greater than 0.9000, but the model has a certain degree of difference in the extraction of different relations. The accuracy of "defect occurrence" is significantly higher than that of other relations. This is because the data text in this article is mostly defect information of power transformers, and the seat model has sufficient data samples for learning.

[0173] Example 2

[0174] Embodiment 2 provides a system for constructing a knowledge graph of power transformer equipment defects. Figure 6 As shown, the system includes:

[0175] A collection module is used to collect structured and unstructured data of transformer equipment operation defects;

[0176] The ontology layer construction module is used to construct the ontology layer of transformer equipment defects and to express the relationship between the various ontologies in the ontology layer in a structured manner;

[0177] An extraction module, used to extract the structured data to obtain the first target triples and to extract the unstructured data to obtain the second target triples with reference to the constructed ontology layer;

[0178] The fusion module is used to fuse the first target triplet and the second target triplet to obtain transformer equipment defect map data.

[0179] Preferably, the acquisition module is used to extract the structured data from the equipment defect record table of the power grid resource business center; collect the unstructured data from the power grid production management system; wherein the unstructured data includes long text, pictures, audio and / or video associated with the transformer.

[0180] Preferably, the ontology layer building module is also used for:

[0181] Determine the equipment-component body according to the equipment structure of the transformer;

[0182] Determine the defect information ontology based on the defect standard library;

[0183] The physical relationship between the components is used to determine the relationship between the entities, and the attributes of the entities are determined according to the operating characteristics of the entities; wherein the physical relationship includes a structural association relationship and a functional association relationship.

[0184] Preferably, the extraction module further comprises a structured knowledge extraction unit for converting the structured data into first target triples by using a D2RQ tool with reference to the ontology layer.

[0185] Preferably, the extraction module further includes an unstructured knowledge extraction unit, which is used to:

[0186] Referring to the ontology layer, an entity recognition model based on BERT and EE-GCN is used to extract entities and target attributes from unstructured data, and target entity data and target attribute data are obtained respectively;

[0187] Referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data;

[0188] The target entity data, the target relationship data and the target attribute data are converted into a second target triple.

[0189] Preferably, the fusion module is also used to perform co-reference resolution on the target entities and target attributes in the first target triplet and the second target triplet, to unify the data of defective equipment and defect information, and to form standardized transformer equipment defect map data.

[0190] Preferably, the system also includes a storage module for storing transformer equipment defect map data in a graph database.

[0191] It should be clarified that the implementation scheme of the above-mentioned domain knowledge graph construction system for power equipment defects is similar to the description of the construction method mentioned above, and has the same execution effect as the example method. Therefore, for the details that cannot be described in the implementation of the domain knowledge graph construction system for power equipment defects of the present invention, relevant personnel in related fields are requested to refer to the construction method example.

[0192] Example 3

[0193] Embodiment 3 of the present invention provides a computer-readable storage medium. The computer-readable storage medium stores a program, which, when executed by a processor, implements the steps in the method for constructing a knowledge graph of defects of power transformer equipment as described in Embodiment 1 of the present invention.

[0194] The detailed steps are the same as the method for constructing the power transformer equipment defect knowledge graph provided in Example 1, and will not be repeated here.

[0195] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0196] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0197] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0198] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0199] 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 above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a knowledge graph of power transformer equipment defects, characterized in that: include: Collect structured and unstructured data of transformer equipment operation defects; Construct the ontology layer of transformer equipment defects and express the relationship between each ontology in the ontology layer in a structured manner; Taking the constructed ontology layer as a reference, the structured data is extracted to obtain the first target triples, and the unstructured data is extracted to obtain the second target triples; The knowledge of the first target triplet and the second target triplet is fused to obtain the transformer equipment defect map data.

2. The method for constructing a knowledge graph of power transformer equipment defects according to claim 1, characterized in that: The structured data and unstructured data collected about transformer equipment operation defects include: Extracting the structured data from the equipment defect record table of the power grid resource business center; The unstructured data is collected from a power grid production management system; wherein the unstructured data includes long text, pictures, audio and / or video associated with the transformer.

3. The method for constructing a knowledge graph of power transformer equipment defects according to claim 1, characterized in that: The step of constructing the transformer equipment defect ontology layer and structurally expressing the relationship between the various ontologies in the ontology layer includes: Determine the equipment-component body according to the equipment structure of the transformer; Determine the defect information ontology based on the defect standard library; The physical relationship between the components is used to determine the relationship between the entities, and the attributes of the entities are determined according to the operating characteristics of the entities; wherein the physical relationship includes a structural association relationship and a functional association relationship.

4. The method for constructing a knowledge graph of power transformer equipment defects according to claim 1, characterized in that: The method of extracting the structured data to obtain the first target triples and extracting the unstructured data to obtain the second target triples with reference to the constructed ontology layer includes: The D2RQ tool is used to convert the structured data into the first target triples with reference to the ontology layer.

5. The method for constructing a knowledge graph of power transformer equipment defects according to claim 1, characterized in that: The method of extracting the structured data to obtain the first target triples and extracting the unstructured data to obtain the second target triples with reference to the constructed ontology layer includes: Referring to the ontology layer, an entity recognition model based on BERT and EE-GCN is used to extract entities and target attributes from unstructured data, and target entity data and target attribute data are obtained respectively; Referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data; The target entity data, the target relationship data and the target attribute data are converted into a second target triple.

6. The method for constructing a knowledge graph of power transformer equipment defects according to claim 5, characterized in that: Referring to the ontology layer, an entity recognition model based on BERT and EE-GCN is used to extract entities and target attributes from unstructured data, and target entity data and target attribute data are obtained respectively, including: The BERT model is used to encode the equipment defect text in the input unstructured data: V=V Token +V Segment +V Position Where V is the coded representation of the equipment defect text, V Token Represents the vector representation of the current element, V Segment Indicates the sentence where the current element is located in the equipment defect text, V Position Indicates the specific position of the current element in the device defect text; Generate semantic dependency analysis tree of equipment defect text; Based on the encoding result output by the BERT model, the generated semantic dependency analysis tree is processed by the EE-GCN model, and the label classification probability distribution of each node in the semantic dependency analysis tree is obtained by referring to the ontology layer; According to the label classification probability distribution, the optimal label is determined through the CRF layer to obtain the target entity data and target attribute data: in, represents the total score of each label sequence, represents the set of all possible labels, express The label classification sequence in represents the optimal label.

7. The method for constructing a knowledge graph of power transformer equipment defects according to claim 6, characterized in that: Based on the encoding result output by the BERT model, the generated semantic dependency analysis tree is processed by the EE-GCN model, and the label classification probability distribution of each node in the semantic dependency analysis tree is obtained by referring to the ontology layer, including: Using the node update module in the EE-GCN model, the feature vector of the node in the current layer is generated by aggregating semantically connected nodes of specific dependency types in the semantic dependency analysis tree. The specific calculation formula is as follows: in, It represents the oth channel of E, W z is a learnable parameter matrix; Represents the structure of the semantic dependency analysis tree, which is represented by a three-dimensional matrix. n represents the number of node trees in the analysis tree, and p represents the relationship dimension. Pool() is a function for creating a process pool. ReLU() is an activation function. N is the total number of channels. H represents the hidden state representation of the lth layer of the oth channel; l-1 represents the hidden state representation of the l-1th layer; the hidden state representation of the first layer is H 1 Obtained from the output value of the BERT model; H l represents the hidden representation of the lth layer, H l is in matrix form, and H l Each row in is the feature vector of the corresponding node. By using the relationship update module in the EE-GCN model, the relationship feature vector in the semantic dependency analysis tree is updated by mining the clues hidden in the dependency edges. The specific formula is as follows: in, represents the splicing operation, W u represents the learnable weight matrix, is the updated feature vector of node i and node j at layer l, is the relationship feature vector between node i and node j in layer l, is the relationship feature vector between node i and node j in the l-1th layer; Referring to the ontology layer, the softMax layer is used to calculate the label classification probability distribution of each node in the semantic dependency analysis tree. The specific formula is as follows: Among them, W t represents the mapping matrix, b t represents the offset, and h is the feature vector of the node.

8. The method for constructing a knowledge graph of power transformer equipment defects according to claim 5, characterized in that: Referring to the ontology layer, the BiGRU+Self-Attention relation extraction model is used to extract relations from unstructured data to obtain target relation data, including: The BiGRU model is used to process the equipment defect text in the unstructured data to obtain the final state representation of each element in the equipment defect text: in, and Represent the hidden state of the current layer from left to right and from right to left, respectively. and Represent the hidden state of the previous layer from left to right and from right to left, GRU represents the gated recurrent unit, x t represents the tth element in the device defect text, [:] represents the concatenation operation of the vector, h t It represents the final state of the current layer; The final state representation of each element is input into the Self-Attention layer to obtain the feature vector of each node. The specific formula is as follows: Q=W q H K=W k H V=W v H Among them, W q , W k , W v They represent the query matrix, matching matrix, and extracted information matrix respectively. H represents the final state of each element and represents the concatenated matrix. Q, K, and V represent the query vector, key vector, and value vector respectively. represents the dimension of K; H′ is an eigenvector matrix, each row of which corresponds to the eigenvector of each node; The feature vectors of each node are concatenated and processed through the linear layer to obtain the target relationship data. The specific formula is as follows: c=[h1′:h2′,...:h n ′]W m Among them, c is the relationship recognition result, h1′, h2′, ..., h n ′ are the feature vectors of the nodes in the corresponding rows in H′, W m represents the linear transformation matrix, m represents the number of relationships; Recursive training is performed using a cross entropy loss function so that the difference between the relationship recognition result and the target relationship value is minimized or less than a preset value, and the trained relationship recognition result is used as the target relationship data; wherein the target relationship value is obtained by referring to the relationship in the ontology layer.

9. The method for constructing a knowledge graph of power transformer equipment defects according to claim 1, characterized in that: The first target triplet and the second target triplet are fused to obtain transformer equipment defect map data, including: The target entities and target attributes in the first target triplet and the second target triplet are coreference resolved to unify the data of defective equipment and defect information and form standardized transformer equipment defect map data.

10. A system for constructing a knowledge graph of power transformer equipment defects using the method described in any one of claims 1 to 9, characterized in that: include: A collection module is used to collect structured and unstructured data of transformer equipment operation defects; The ontology layer construction module is used to construct the ontology layer of transformer equipment defects and to express the relationship between the various ontologies in the ontology layer in a structured manner; An extraction module, used to extract the structured data to obtain the first target triples and to extract the unstructured data to obtain the second target triples with reference to the constructed ontology layer; The fusion module is used to fuse the first target triplet and the second target triplet to obtain transformer equipment defect map data.

11. A storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to any one of claims 1 to 9.

Images