A wind power equipment accident diagnosis method and system based on knowledge graph

Abstract

The present invention proposes a wind power equipment accident diagnosis method and system based on knowledge graph, and the method comprises the following steps: obtaining wind power equipment operation and maintenance data and standard specification data, and preprocessing the data to construct a wind power equipment safety data set; performing keyword extraction and text data annotation on text data in the wind power equipment safety data set to construct a keyword corpus, and performing word vector training on the annotated text to obtain a wind power equipment safety data word vector set; constructing an entity recognition model based on a BiLSTM-CRF algorithm combined with a similarity algorithm, and outputting an entity recognition prediction result; and performing relationship extraction and attribute extraction on wind power equipment safety data to construct a wind power equipment safety knowledge graph; by integrating the operation and maintenance data of wind power equipment and standard specification data in the safety field, constructing and optimizing the knowledge graph, accurate diagnosis and cause analysis of wind power equipment faults are achieved, thereby improving the diagnosis efficiency of wind power equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine equipment accident diagnosis, and in particular to a wind power equipment accident diagnosis method and system based on a knowledge graph. Background Art

[0002] In order to meet the demand for energy supply and achieve sustainable energy development, the transformation of fossil energy system to sustainable energy system has become the basic trend of energy development in all countries in the world. As one of the earliest developed and most important renewable energy sources in the comprehensive energy system, wind energy has great potential in greenhouse gas emission reduction due to its abundant storage, renewable, clean and pollution-free characteristics. It has become the mainstream direction of global renewable energy development and an important force in promoting industrial upgrading and driving energy system adjustment.

[0003] Publication number CN109194045A discloses an operation and maintenance system for motor equipment fault diagnosis, including a vibration and noise acquisition analyzer, system software, cloud server and terminal equipment. The vibration and noise acquisition analyzer is connected to the cloud server and the terminal equipment in turn. The vibration and noise acquisition analyzer is deployed around the motor equipment. The vibration and noise acquisition analyzer includes an acquisition sensor, a processor, a data conversion module and a communication module. The terminal equipment deploys the system software to configure monitoring scenarios, basic motor parameters, diagnostic models and operating parameters, etc., and connects to the cloud server to realize health status prediction, fault analysis, operation and maintenance reports and human-computer interaction. It can effectively monitor the status of the motor equipment and perform predictive maintenance on faulty equipment in a timely manner.

[0004] In the current practice of wind power equipment safety management and control, relevant departments have built many systems such as wind power equipment safety monitoring, inspection, and emergency rescue platforms. However, each system is developed by a different software company with different system architectures. Equipment information needs to be repeatedly entered into multiple systems, and the databases are independent of each other. Inspection and supervision business data are not interoperable between inspection agencies and monitoring agencies, and are not related to each other in various functional modules. This leads to complex data processing, difficult fault location, and inaccurate cause analysis in the operation and maintenance of wind power equipment, thereby reducing the diagnostic efficiency of wind power equipment. Summary of the invention

[0005] In view of this, the present invention proposes a wind power equipment accident diagnosis method and system based on knowledge graph. By integrating the operation and maintenance data of wind power equipment and standard specification data in the safety field, the knowledge graph is constructed and optimized, and accurate diagnosis and cause analysis of wind power equipment failures are achieved. The problems of complex data processing, difficult fault location, and inaccurate cause analysis faced in the operation and maintenance of wind power equipment are effectively solved, thereby improving the diagnostic efficiency of wind power equipment.

[0006] The technical solution of the present invention is implemented as follows: In the first aspect, the present invention provides a wind power equipment accident diagnosis method based on a knowledge graph, comprising the following steps:

[0007] S1, obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set;

[0008] S2, extract keywords and annotate text data in the wind power equipment safety data set, build a keyword corpus, and train word vectors on the annotated text to obtain a word vector set of wind power equipment safety data;

[0009] S3, builds an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, inputs the wind power equipment safety data word vector set into the entity recognition model for training, and outputs the entity recognition prediction results;

[0010] S4, extract relations and attributes of wind power equipment safety data, determine knowledge triples based on entity recognition prediction results, and construct a wind power equipment safety knowledge graph;

[0011] S5, based on the TransE algorithm, vectorizes the triple data in the knowledge graph, analyzes the features between entities and relationships, and completes the knowledge graph;

[0012] S6, input the target fault, obtain all the accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph;

[0013] S7, establishes a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnoses the cause of the target fault, and generates an accident analysis report.

[0014] On the basis of the above technical scheme, preferably, in step S1, wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are obtained, and the wind power equipment operation and maintenance data include the operating status of the wind power equipment, historical fault records, maintenance logs and performance parameter information, and the wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are cleaned and segmented, and the pre-processed data are used to construct an equipment safety data set.

[0015] On the basis of the above technical scheme, preferably, in step S2, the PB-TF-IDF algorithm is used to extract keywords from the data in the wind power equipment safety data set, a keyword corpus is constructed based on the extracted keyword data, the BIO tagging method is used to annotate the data in the wind power equipment safety data set, and the skip-word model in Word2vec is used to train the word vector to obtain the wind power equipment safety data word vector set.

[0016] Based on the above technical solution, preferably, step S3 includes the following sub-steps:

[0017] S31, similarity calculation is performed on any two word vectors in the wind power equipment safety data word vector set to obtain a similarity value, a similarity judgment threshold is preset, and the similarity values ​​are judged in turn according to the similarity judgment threshold, and the word vectors corresponding to the word vectors greater than the similarity judgment threshold are extracted, and a corresponding similar word library is constructed, which is expressed as:

[0018]

[0019] In the formula, cosθ is the similarity value of two word vectors, A i and B i is the value of the i-th dimension of the two word vectors, the dimension of each word vector is n, i∈n;

[0020] S32, build a BiLSTM-CRF named entity recognition model, define the entity type of wind power equipment, input the wind power equipment safety data word vector set into the BiLSTM network according to the word vector sequence, and output the score value of each word corresponding to different entity types. The expression is:

[0021] B q =(b q，e , b q，p , b q，a ,b q，t , b q,s , ..., b q，j )

[0022] In the formula, q is the unique index of the current word segment in the keyword corpus, b q，j is the probability value that the current word segment belongs to the jth entity type;

[0023] S34, obtaining the corresponding similar word library according to each word segment, and obtaining the similarity value of each word in the similar word library, and constructing a similarity degree set, the expression is: S = (S1, S2, ..., S m );

[0024] S35, for each word in the text, use the fusion formula to calculate the fusion score of each entity type, select the entity type with the highest fusion score as the entity type of the current word segment, and obtain the entity recognition prediction result. The fusion formula calculation expression is:

[0025]

[0026] In the formula, α and β are weight coefficients, S k is the similarity value between the current word segment and the kth similar word, k∈m, B kis the score value of the k-th similar word in different entity types.

[0027] On the basis of the above technical solution, preferably, step S4 includes the following sub-steps:

[0028] The output entity recognition prediction results are used as nodes in the knowledge graph;

[0029] Define the grammatical rules in the field of wind power equipment safety, and extract the relationship between wind power equipment safety data based on the grammatical rules in the field of wind power equipment safety to obtain the relationship between each entity;

[0030] Use natural language processing technology to extract attributes from wind power equipment safety data and obtain attribute values ​​of various entities;

[0031] According to the output entity recognition prediction results, the relationship between entities and the attribute values ​​of various entities are combined to obtain multiple knowledge triples, and the triple data is stored using the Cypher language in the graph database;

[0032] The py2neo tool is used to store the knowledge triple data one by one into the graph database Neo4j;

[0033] The triple storage operation in the knowledge graph is encapsulated to obtain the wind power equipment safety knowledge graph.

[0034] On the basis of the above technical solution, preferably, step S5 includes the following sub-steps:

[0035] S51, construct a tail entity candidate set, traverse the knowledge graph, obtain all entity pairs connected by the same type of relationship, and store the corresponding tail entity in the tail entity candidate set;

[0036] S52, each knowledge triple is in the form of (h, r, t), where h is the head entity, t is the tail entity, and r is the relationship between the head entity and the tail entity. The TransE algorithm is used to represent the head entity and the relationship as corresponding low-dimensional vectors, respectively, to obtain the head entity vector and the relationship vector;

[0037] S53, add the head entity vector and the relationship vector, and calculate the Euclidean distance between the head entity vector and each tail entity vector in the tail entity candidate set, expressed as:

[0038] d(h+r,t)=||h+rt||2

[0039] Where ||*||2 is the Euclidean norm of the vector;

[0040] S54, arranging all candidate tail entities in the tail entity candidate set in descending order according to the corresponding calculated Euclidean distances, and taking the candidate tail entity with the smallest distance as the result of knowledge graph completion;

[0041] S54, storing the result of knowledge graph completion in the knowledge graph, setting the relationship attribute as the completion item, and setting the corresponding calculated Euclidean distance as the attribute value.

[0042] Based on the above technical solution, preferably, step S7 includes the following sub-steps: establishing a fault diagnosis model based on the knowledge graph sub-graph and the Bayesian network, diagnosing the cause of the target fault, and generating an accident analysis report

[0043] S71, define the Bayesian network structure. The Bayesian network consists of nodes and directed edges. According to the knowledge graph subgraph, the target fault is taken as the root node of the network, and each accident cause is taken as a child node in the network. The root node and the child node are connected by directed edges. The relationship between each directed edge and the target fault and the corresponding accident cause in the knowledge graph subgraph is determined;

[0044] S72, assign a priori probabilities to the sub-nodes corresponding to each accident cause, expressed as:

[0045] P(r)=P(r|p)

[0046] Where P(r) is the prior probability, r is the cause of the accident, p is the accident, and P(r|p) is the probability that the accident is caused by the cause of the accident;

[0047] S73, assigning prior probabilities according to the child nodes and constructing a conditional probability table using the Noisy-or model;

[0048] S74, using the GeNIe tool to import the knowledge graph subgraph into the tool, and input the defined Bayesian network structure, the prior probability of molecular node assignment and the conditional probability table to establish a fault diagnosis model.

[0049] Based on the above technical solution, preferably, the step S73 uses the Noisy-or model to construct the conditional probability table, including defining the association strength between the parent node and the child node, and the expression is:

[0050] P w =P(s=T|g w =Tonly)

[0051] =P(s=T|g1=F, g2=F,...,g w =T, ...g n =F)

[0052] Where P wis the degree of association between nodes, s is the child node, g w is the w-th parent node set, g w =Tonly means only g w The nodes take truth values, T and F represent Boolean values ​​true and false respectively;

[0053] According to the strength of association, the parameter values ​​in the conditional probability table are determined, and the expression is:

[0054]

[0055] Where pa(s) is all parent nodes of s, pa(s) + is the parent node of all true values ​​of s.

[0056] On the basis of the above technical solution, preferably, in step S7, the cause of the target fault is diagnosed and an accident analysis report is generated, wherein the target fault is input into the fault diagnosis model, and based on the Bayesian network diagnostic reasoning, all causes of the fault and the corresponding probabilities are output, the fault cause with the highest probability is obtained, and it is used as the cause propagation chain, and the corresponding cause propagation chain is retrieved in the knowledge graph using the Cypher language to obtain the knowledge graph subgraph of the corresponding processing solution, and an accident analysis report is generated.

[0057] In a second aspect, the present invention further provides a wind power equipment accident diagnosis system based on a knowledge graph, characterized in that: the system comprises:

[0058] The acquisition module is used to obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set;

[0059] A processing module is used to extract keywords and annotate text data in the wind power equipment safety data set, build a keyword corpus, and perform word vector training on the annotated text to obtain a wind power equipment safety data word vector set;

[0060] The entity recognition module is used to build an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, input the wind power equipment safety data word vector set into the entity recognition model for training, and output the entity recognition prediction result;

[0061] The knowledge graph construction module is used to extract relations and attributes from wind power equipment safety data, determine knowledge triples based on entity recognition prediction results, and construct a wind power equipment safety knowledge graph;

[0062] The knowledge graph completion module is used to vectorize the triple data in the knowledge graph based on the TransE algorithm, analyze the features between entities and relationships, and complete the knowledge graph.

[0063] The input module is used to input the target fault, obtain all the accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph;

[0064] The diagnosis module is used to establish a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnose the cause of the target fault, and generate an accident analysis report.

[0065] The wind power equipment accident diagnosis method and system based on knowledge graph of the present invention have the following beneficial effects compared with the prior art:

[0066] (1) By acquiring wind power equipment operation and maintenance data and standard specification data, a wind power equipment safety data set is constructed. The wind power equipment safety data set is processed to construct and optimize the knowledge graph. By combining the knowledge graph subgraph and the Bayesian network, a fault diagnosis model is established to achieve accurate diagnosis and cause analysis of wind power equipment faults, and generate a detailed accident analysis report, thereby improving the diagnostic efficiency of wind power equipment;

[0067] (2) By combining the similarity algorithm with the BiLSTM-CRF algorithm, the model not only improves its entity recognition capability when processing wind power equipment safety data, but also enhances its robustness and generalization capabilities. By constructing a similarity vocabulary and a similarity set, the model can more effectively utilize the semantic relevance between words and improve the accuracy and efficiency of entity recognition.

[0068] (3) The TransE algorithm is used to represent the head entity accident and the cause of the associated relationship as corresponding dimensional vectors, and the Euclidean distance between the sum of these two vectors and the candidate tail entity vector is calculated. Based on the vector space distance method, the potential semantic association between entities and relationships can be captured, thereby achieving accurate prediction of the tail entity. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0070] Figure 1 It is a flow chart of the wind power equipment accident diagnosis method based on knowledge graph of the present invention;

[0071] Figure 2 A framework diagram is constructed for the wind power equipment safety knowledge graph of the wind power equipment accident diagnosis method based on the knowledge graph of the present invention;

[0072] Figure 3 It is a structure diagram of a word-skipping model of a wind power equipment accident diagnosis method based on a knowledge graph of the present invention;

[0073] Figure 4 A schematic diagram of the knowledge graph completion logic of the wind power equipment accident diagnosis method based on the knowledge graph of the present invention;

[0074] Figure 5 This is a theoretical example diagram of the Noisy-or model of the wind power equipment accident diagnosis method based on the knowledge graph of the present invention;

[0075] Figure 6 It is a schematic diagram of the implementation process of the wind power equipment accident diagnosis method based on knowledge graph of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] like Figure 1 and Figure 6 As shown, a wind power equipment accident diagnosis method based on knowledge graph of the present invention comprises the following steps:

[0078] S1, obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set.

[0079] In step S1 of this embodiment, wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are obtained, and the wind power equipment operation and maintenance data include the operating status, historical fault records, maintenance logs and performance parameter information of the wind power equipment. The wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are cleaned and segmented, and the pre-processed data are used to construct an equipment safety data set.

[0080] It should be noted that the operating status data is collected in real time through sensors installed on wind power equipment, including key parameters such as wind speed, wind direction, temperature, humidity, vibration, rotation speed, power output, etc.; these data reflect the real-time operating status and performance of wind power equipment; historical fault records record all faults that occurred in the past operation of wind power equipment and related information, such as fault time, fault phenomenon, fault cause, and treatment measures; maintenance logs record detailed information on regular maintenance, inspection, and repair activities of wind power equipment, including maintenance time, maintenance content, maintenance personnel, maintenance results, etc.; these logs help to understand the maintenance history and current status of the equipment; performance parameter information includes the rated power of wind power equipment Key performance indicators such as rated power, actual output power, power generation efficiency, and availability; these parameters are important bases for evaluating the performance of wind power equipment; standard and specification data in the field of wind power equipment safety mainly come from relevant standards and specifications formulated by the state, industry or enterprises, mainly including equipment safety standards, operating procedures and maintenance specifications. Equipment safety standards stipulate the safety requirements that wind power equipment should meet in various links such as design, manufacturing, installation, operation, and maintenance. The operating procedures describe the operating procedures, precautions, emergency response measures, etc. of wind power equipment to ensure the safe operation of the equipment. The maintenance specifications stipulate the maintenance cycle, content, methods, standards, etc. of wind power equipment to ensure the long-term stable operation of the equipment.

[0081] Among them, data cleaning includes deleting duplicate records in the data set to avoid interference with subsequent analysis, filling or deleting missing data to ensure data integrity, detecting and processing outliers in the data through statistical methods to improve data accuracy, and converting data from different sources and formats into a unified format for subsequent processing and analysis; word segmentation includes word segmentation of text data, splitting sentences into individual words or phrases, and considering professional vocabulary and terminology in the wind power field during word segmentation to ensure the accuracy and professionalism of the word segmentation results. After data cleaning and word segmentation, the pre-processed data is organized according to a certain structure and format to construct an equipment safety data set. This data set should include the operating status, historical fault records, maintenance logs, performance parameter information, and related safety standards and specifications of wind power equipment, providing basic data support for subsequent entity recognition, relationship extraction, and fault diagnosis steps.

[0082] like Figure 2 As shown, S2, keyword extraction and text data annotation are performed on the text data in the wind power equipment safety data set, a keyword corpus is constructed, and word vector training is performed on the annotated text to obtain a word vector set of wind power equipment safety data.

[0083] In step S2 of this embodiment, the PB-TF-IDF algorithm is used to extract keywords from the data in the wind power equipment safety data set, a keyword corpus is constructed based on the extracted keyword data, the BIO tagging method is used to annotate the data in the wind power equipment safety data set, and the skip-word model in Word2vec is used to train the word vector to obtain the wind power equipment safety data word vector set.

[0084] It should be noted that the PB-TF-IDF algorithm is a keyword extraction algorithm that introduces position information based on TF-IDF. PB-TF-IDF enhances the keyword extraction effect by considering the position of words in the text.

[0085] Use the BIO annotation method to annotate the text data in the wind power equipment safety dataset, label the beginning of an entity with B, label the remaining part of an entity with I, label the part that is not an entity with O, and attach the corresponding entity type after the position letter to obtain the node in the knowledge graph;

[0086] like Figure 3 As shown in the figure, since the corpus in the general field cannot fully represent the text features of professional terms in the professional field of wind power equipment, it is necessary to train word vectors specifically in the field of wind power equipment safety, and use the Skip-Gram model in the word2vec method to train word vectors;

[0087] When the skip-gram model is trained, the current word is known, and the left and right corpora of the current word segment are predicted. By setting the output dimension of the word vector and inputting the text segment, an n-dimensional dense vector can be output, in which the numerical size of each element represents the specific features of the current word. The larger the dimension is set, the more features can be considered when training the word vector, and the final output word vector training result can better match the actual situation.

[0088] Specifically, the annotated text data is taken as input, and a word sequence is obtained after word segmentation. The word sequence is trained using a skip-gram model, and the value of the word vector is updated by optimizing the objective function. After the training is completed, the vector representation of each word is obtained, that is, the wind power equipment safety data word vector set.

[0089] S3, builds an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, inputs the wind power equipment safety data word vector set into the entity recognition model for training, and outputs the entity recognition prediction results.

[0090] Wherein, step S3 includes the following sub-steps:

[0091] S31, similarity calculation is performed on any two word vectors in the wind power equipment safety data word vector set to obtain a similarity value, a similarity judgment threshold is preset, and the similarity values ​​are judged in turn according to the similarity judgment threshold, and the word vectors corresponding to the word vectors greater than the similarity judgment threshold are extracted, and a corresponding similar word library is constructed, which is expressed as:

[0092]

[0093] In the formula, cosθ is the similarity value of two word vectors, A i and B i is the value of the i-th dimension of the two word vectors. The dimension of each word vector is n, i∈n, where the closer the cosine value is to 1, the smaller the value of A i and B i The more similar;

[0094] S32, build a BiLSTM-CRF named entity recognition model, define the entity type of wind power equipment, input the wind power equipment safety data word vector set into the BiLSTM network according to the word vector sequence, and output the score value of each word corresponding to different entity types. The expression is:

[0095] B q =(b q,e , b q,p , b q，a ,b q，t , b q,s , ..., b q,j )

[0096] In the formula, q is the unique index of the current word segment in the keyword corpus, b q，j is the probability value that the current word segment belongs to the jth entity type;

[0097] It should be noted that the BiLSTM-CRF model theory is as follows: First, use BiLSTM to extract features. The word vector sequence X = (x1, x1, ..., x n ) Input BiLSTM, where the forward LTSM sequence is trained with the output state sequence h from front to back Lt =[h L0 ,h L1 , ..., h Ln ], and the backward training output is h Rt =[h R0 ,h R1 , ..., h Rn ], merging the features of the two can get h t =[h Rt ,h Lt ],h tIt is the output feature matrix of the hidden layer during the training process. After discarding some features through the dropout layer, the labeling results are displayed by studying the features.

[0098] Use CRF to correct the prediction. Since BiLSTM is aimed at word-level feature analysis and does not focus on the entire sentence, CRF is introduced to comprehensively consider the rationality of each label, define the state transition scoring matrix A and the state observation scoring matrix P, and construct the model's scoring formula for the prediction results based on matrices A and P:

[0099]

[0100] In the formula, S(X,Y) is the score of the input sequence X and the label sequence Y, To transition from state yi to state y i+1 Ratings, For input x j Corresponding label y j Ratings;

[0101] The final output of the BiLSTM model is a vector, in which the elements represent the probability of each entity category of the current word; by adding a CRF layer after the BiLSTM model, the prediction results of the entity label can be corrected so that problems such as the label "B-tools" followed by the label "B-tools" no longer occur, which can improve the accuracy of entity extraction. The conditional probability calculation formula based on the above formula is:

[0102]

[0103] Among them, Y i for all possible tags;

[0104] The optimization goals of model training are:

[0105] maxlog(P(Y|X)).

[0106] S34, obtaining the corresponding similar word library according to each word segment, and obtaining the similarity value of each word in the similar word library, and constructing a similarity degree set, the expression is: S = (S1, S2, ..., S m );

[0107] S35, for each word in the text, use the fusion formula to calculate the fusion score of each entity type, select the entity type with the highest fusion score as the entity type of the current word segment, and obtain the entity recognition prediction result. The fusion formula calculation expression is:

[0108]

[0109] In the formula, α and β are weight coefficients, S k is the similarity value between the current word segment and the kth similar word, k∈m, B k is the score value of the k-th similar word in different entity types.

[0110] It should be noted that combining the similarity algorithm with the BiLSTM-CRF algorithm not only improves the model's entity recognition ability when processing wind power equipment safety data, but also enhances the model's robustness and generalization ability. By building a similarity vocabulary and a similarity set, the model can more effectively utilize the semantic relevance between words and improve the accuracy and efficiency of entity recognition.

[0111] like Figure 4 As shown, S4 performs relationship extraction and attribute extraction on wind power equipment safety data, determines knowledge triples based on entity recognition prediction results, and constructs a wind power equipment safety knowledge graph.

[0112] Step S4 includes the following sub-steps:

[0113] The output entity recognition prediction results are used as nodes in the knowledge graph;

[0114] Define the grammatical rules in the field of wind power equipment safety, and extract the relationship between wind power equipment safety data based on the grammatical rules in the field of wind power equipment safety to obtain the relationship between each entity;

[0115] Use natural language processing technology to extract attributes from wind power equipment safety data and obtain attribute values ​​of various entities;

[0116] According to the output entity recognition prediction results, the relationship between entities and the attribute values ​​of various entities are combined to obtain multiple knowledge triples, and the triple data is stored using the Cypher language in the graph database;

[0117] The py2neo tool is used to store the knowledge triple data one by one into the graph database Neo4j;

[0118] The triple storage operation in the knowledge graph is encapsulated to obtain the wind power equipment safety knowledge graph.

[0119] It should be noted that a text processing method based on grammatical rules is used to extract relationships, and relationships are used as edges connecting nodes in the knowledge graph; based on attribute extraction, relevant attributes and attribute values ​​of various entities in unstructured data are obtained, and based on entity recognition, relationship extraction and attribute extraction, accurate identification of knowledge triples is achieved, and the Cypher language in the graph database is used to store triple data. The py2neo tool is used to process data, and the "node-edge-node" triples are stored one by one in the graph database Neo4j. After the graph database is connected, the visual storage and update of the knowledge triples are realized.

[0120] like Figure 5 As shown, S5, based on the TransE algorithm, vectorizes the triple data in the knowledge graph, analyzes the features between entities and relationships, and completes the knowledge graph.

[0121] Step S5 includes the following sub-steps:

[0122] S51, construct a tail entity candidate set, traverse the knowledge graph, obtain all entity pairs connected by the same type of relationship, and store the corresponding tail entity in the tail entity candidate set;

[0123] S52, each knowledge triple is in the form of (h, r, t), where h is the head entity, t is the tail entity, and r is the relationship between the head entity and the tail entity. The TransE algorithm is used to represent the head entity and the relationship as corresponding low-dimensional vectors, respectively, to obtain the head entity vector and the relationship vector;

[0124] S53, add the head entity vector and the relationship vector, and calculate the Euclidean distance between the head entity vector and each tail entity vector in the tail entity candidate set, expressed as:

[0125] d(h+r,t)=||h+rt||2

[0126] Where ||*||2 is the Euclidean norm of the vector;

[0127] S54, arranging all candidate tail entities in the tail entity candidate set in descending order according to the corresponding calculated Euclidean distances, and taking the candidate tail entity with the smallest distance as the result of knowledge graph completion;

[0128] S54, storing the result of knowledge graph completion in the knowledge graph, setting the relationship attribute as the completion item, and setting the corresponding calculated Euclidean distance as the attribute value.

[0129] It should be noted that the goal of TransE algorithm training is to make (h+r) and t as equal as possible, that is, to achieve h+rt≈0; in addition, for a correct triple, the Euclidean distance d(h+r, t) should be low enough, and for an incorrect triple, it should be high enough, so the objective function is given:

[0130]

[0131] In the formula, S is all the triples in the knowledge graph, S′ is the constructed negative sample triples, γ is an artificially set value, generally set to 1, [x] + Indicates that the minimum value is 0.

[0132] In this embodiment, after the knowledge triples are vectorized, the missing values ​​in the knowledge graph are inferred by analyzing the characteristics between entities and relationships, and finally the knowledge graph is completed to generate a more complete wind power equipment safety knowledge graph. In order to accurately locate the cause of the accident, the main purpose of completing the wind power equipment safety knowledge graph is to predict and complete the tail entity when it is known that the head entity is the accident and the associated relationship is the cause.

[0133] According to the TransE algorithm in this embodiment, the head entity accident and the cause of the associated relationship are represented as corresponding dimensional vectors, and the Euclidean distance between the sum of these two vectors and the candidate tail entity vector is calculated. This method based on vector space distance can capture the potential semantic association between entities and relationships, thereby achieving accurate prediction of tail entities. Taking the candidate tail entity with the smallest distance as the completion result can largely ensure the accuracy of the completion. With the continuous completion of missing values ​​in the knowledge graph, the completeness of the wind power equipment safety knowledge graph has been significantly improved, providing more comprehensive data support for operation and maintenance management, fault diagnosis, etc. in the wind power industry.

[0134] S6, input the target fault, obtain all accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph.

[0135] It should be noted that after completing the missing values ​​of the knowledge graph, the Cypher query language is used to perform retrieval operations on the data in the graph, usually including the MATCH keyword, WHERE keyword, and RETURN keyword. All possible causes of accidents during the operation of wind power equipment can be obtained, and the nodes with edge relationships named causes are retained, and finally a knowledge graph subgraph is obtained.

[0136] S7, establishes a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnoses the cause of the target fault, and generates an accident analysis report.

[0137] Among them, step S7 includes the following sub-steps: establishing a fault diagnosis model based on the knowledge graph sub-graph and the Bayesian network, diagnosing the cause of the target fault, and generating an accident analysis report

[0138] S71, define the Bayesian network structure. The Bayesian network consists of nodes and directed edges. According to the knowledge graph subgraph, the target fault is taken as the root node of the network, and each accident cause is taken as a child node in the network. The root node and the child node are connected by directed edges. The relationship between each directed edge and the target fault and the corresponding accident cause in the knowledge graph subgraph is determined;

[0139] S72, assign a priori probabilities to the sub-nodes corresponding to each accident cause, expressed as:

[0140] P(r)=P(r|p)

[0141] Where P(r) is the prior probability, r is the cause of the accident, p is the accident, and P(r|p) is the probability that the accident is caused by the cause of the accident;

[0142] It should be noted that according to Bayes' theorem, the probability of r-cause leading to accident p is calculated as follows:

[0143]

[0144] Calculate the prior probability of cause r occurring, the calculation formula is:

[0145]

[0146] In industrial practice, it is known that when the cause of an accident r occurs, the accident p has a high probability of occurring. Therefore, in order to increase the safety threshold, let P(P|r)=1, so the calculation formula can be simplified to:

[0147] P(r)=P(r|p)P(p).

[0148] Among them, P(p) represents the probability value of abnormal state;

[0149] Since the cause of the accident is determined only when an accident occurs, it can be seen that |P(p)=1, and the calculation formula can be simplified to:

[0150] P(r)=P(r|p).

[0151] S73, assigning prior probabilities according to the child nodes and constructing a conditional probability table using the Noisy-or model;

[0152] The conditional probability values ​​between nodes are calculated. In order to accurately and efficiently obtain the conditional probability table in the Bayesian network, the Noisy-or model is used to improve the Bayesian network. This model obtains the conditional probability table by coupling the degree of association between each node. Therefore, it is only necessary to determine the degree of mutual association between the parent node and the child node to obtain the conditional probability table of the node through coupling.

[0153] In step S73, the Noisy-or model is used to construct a conditional probability table, including defining the association strength between the parent node and the child node, and the expression is:

[0154] P w =P(s=T|g w =Tonly)

[0155] =P(s=T|g1=F,g2=F,...,g w =T,...g n =F)

[0156] Where P w is the degree of association between nodes, s is the child node, g w is the w-th parent node set, g w =Tonly means only g w The nodes take truth values, T and F represent Boolean values ​​true and false respectively;

[0157] According to the strength of association, the parameter values ​​in the conditional probability table are determined, and the expression is:

[0158]

[0159] Where pa(s) is all parent nodes of s, pa(s) + is the parent node of all true values ​​of s.

[0160] S74, using the GeNIe tool to import the knowledge graph subgraph into the tool, and input the defined Bayesian network structure, the prior probability of molecular node assignment and the conditional probability table to establish a fault diagnosis model.

[0161] In step S7 of this embodiment, the cause of the target fault is diagnosed and an accident analysis report is generated, wherein the target fault is input into the fault diagnosis model, and based on the Bayesian network diagnostic reasoning, all causes of the fault and the corresponding probabilities are output, the fault cause with the highest probability is obtained, and it is used as the cause propagation chain, and the corresponding cause propagation chain is retrieved in the knowledge graph using the Cypher language to obtain the knowledge graph subgraph of the corresponding processing solution, and an accident analysis report is generated.

[0162] In this embodiment, a fault diagnosis model is established based on the knowledge graph subgraph and the Bayesian network, which realizes the accurate diagnosis and efficient analysis of the causes of target fault accidents of wind power equipment. This model not only improves the accuracy and efficiency of fault diagnosis, but also generates a detailed accident analysis report for reference by staff in the process of handling and decision-making.

[0163] The present invention also provides a wind power equipment accident diagnosis system based on knowledge graph, the system comprising:

[0164] The acquisition module is used to obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set;

[0165] A processing module is used to extract keywords and annotate text data in the wind power equipment safety data set, build a keyword corpus, and perform word vector training on the annotated text to obtain a wind power equipment safety data word vector set;

[0166] The entity recognition module is used to build an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, input the wind power equipment safety data word vector set into the entity recognition model for training, and output the entity recognition prediction result;

[0167] The knowledge graph construction module is used to extract relations and attributes from wind power equipment safety data, determine knowledge triples based on entity recognition prediction results, and construct a wind power equipment safety knowledge graph;

[0168] The knowledge graph completion module is used to vectorize the triple data in the knowledge graph based on the TransE algorithm, analyze the features between entities and relationships, and complete the knowledge graph.

[0169] The input module is used to input the target fault, obtain all the accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph;

[0170] The diagnosis module is used to establish a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnose the cause of the target fault, and generate an accident analysis report.

[0171] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind power equipment accident diagnosis method based on knowledge graph, characterized in that: The following steps are involved: S1, obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set; S2, extract keywords and annotate text data in the wind power equipment safety data set, build a keyword corpus, and train word vectors on the annotated text to obtain a word vector set of wind power equipment safety data; S3, builds an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, inputs the wind power equipment safety data word vector set into the entity recognition model for training, and outputs the entity recognition prediction results; S4, extract relations and attributes of wind power equipment safety data, determine knowledge triples based on entity recognition prediction results, and construct a wind power equipment safety knowledge graph; S5, based on the TransE algorithm, vectorizes the triple data in the knowledge graph, analyzes the features between entities and relationships, and completes the knowledge graph; S6, input the target fault, obtain all the accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph; S7, establishes a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnoses the cause of the target fault, and generates an accident analysis report.

2. The wind power equipment accident diagnosis method based on knowledge graph according to claim 1, characterized in that: In step S1, wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are obtained, and the wind power equipment operation and maintenance data include the operating status, historical fault records, maintenance logs and performance parameter information of the wind power equipment. The wind power equipment operation and maintenance data and wind power equipment safety field standard specification data are cleaned and segmented, and the pre-processed data are used to construct an equipment safety data set.

3. The wind power equipment accident diagnosis method based on knowledge graph according to claim 2, characterized in that: In step S2, the PB-TF-IDF algorithm is used to extract keywords from the data in the wind power equipment safety data set, a keyword corpus is constructed based on the extracted keyword data, the BIO tagging method is used to annotate the data in the wind power equipment safety data set, and the skip-word model in Word2vec is used to train the word vector to obtain the wind power equipment safety data word vector set.

4. The wind power equipment accident diagnosis method based on knowledge graph according to claim 3 is characterized in that: Step S3 includes the following sub-steps: S31, similarity calculation is performed on any two word vectors in the wind power equipment safety data word vector set to obtain a similarity value, a similarity judgment threshold is preset, and the similarity values ​​are judged in turn according to the similarity judgment threshold, and the word vectors corresponding to the word vectors greater than the similarity judgment threshold are extracted, and a corresponding similar word library is constructed, which is expressed as: In the formula, cosθ is the similarity value of two word vectors, A i and B i is the value of the i-th dimension of the two word vectors, the dimension of each word vector is n, i∈n; S32, build a BiLSTM-CRF named entity recognition model, define the entity type of wind power equipment, input the wind power equipment safety data word vector set into the BiLSTM network according to the word vector sequence, and output the score value of each word corresponding to different entity types. The expression is: B q =(b q，e ,b q，p ,b q，a ,b q，t ,b q,s ,...,b q，j ) Where q is the unique index of the current word segment in the keyword corpus, bq,j is the probability value of the current word segment belonging to the jth entity type; S34, obtaining the corresponding similar word library according to each word segment, and obtaining the similarity value of each word in the similar word library, and constructing a similarity degree set, the expression is: S = (S1, S2, ..., S m ); S35, for each word in the text, use the fusion formula to calculate the fusion score of each entity type, select the entity type with the highest fusion score as the entity type of the current word segment, and obtain the entity recognition prediction result. The fusion formula calculation expression is: In the formula, α and β are weight coefficients, S k is the similarity value between the current word segment and the kth similar word, k∈m, B k is the score value of the k-th similar word in different entity types.

5. The wind power equipment accident diagnosis method based on knowledge graph according to claim 4, characterized in that: Step S4 includes the following sub-steps: The output entity recognition prediction results are used as nodes in the knowledge graph; Define the grammatical rules in the field of wind power equipment safety, and extract the relationship between wind power equipment safety data based on the grammatical rules in the field of wind power equipment safety to obtain the relationship between each entity; Use natural language processing technology to extract attributes from wind power equipment safety data and obtain attribute values ​​of various entities; According to the output entity recognition prediction results, the relationship between entities and the attribute values ​​of various entities are combined to obtain multiple knowledge triples, and the triple data is stored using the Cypher language in the graph database; The py2neo tool is used to store the knowledge triple data one by one into the graph database Neo4j; The triple storage operation in the knowledge graph is encapsulated to obtain the wind power equipment safety knowledge graph.

6. The wind power equipment accident diagnosis method based on knowledge graph according to claim 5, characterized in that: Step S5 includes the following sub-steps: S51, construct a tail entity candidate set, traverse the knowledge graph, obtain all entity pairs connected by the same type of relationship, and store the corresponding tail entity in the tail entity candidate set; S52, each knowledge triple is in the form of (h, r, t), where h is the head entity, t is the tail entity, and r is the relationship between the head entity and the tail entity. The TransE algorithm is used to represent the head entity and the relationship as corresponding low-dimensional vectors, respectively, to obtain the head entity vector and the relationship vector; S53, add the head entity vector and the relationship vector, and calculate the Euclidean distance between the head entity vector and each tail entity vector in the tail entity candidate set, expressed as: d(h+r,t)=||h+rt||2 Where ||*||2 is the Euclidean norm of the vector; S54, arranging all candidate tail entities in the tail entity candidate set in descending order according to the corresponding calculated Euclidean distances, and taking the candidate tail entity with the smallest distance as the result of knowledge graph completion; S54, storing the result of knowledge graph completion in the knowledge graph, setting the relationship attribute as the completion item, and setting the corresponding calculated Euclidean distance as the attribute value.

7. The wind power equipment accident diagnosis method based on knowledge graph according to claim 6, characterized in that: Step S7 includes the following sub-steps: establishing a fault diagnosis model based on the knowledge graph sub-graph and the Bayesian network, diagnosing the cause of the target fault, and generating an accident analysis report S71, define the Bayesian network structure. The Bayesian network consists of nodes and directed edges. According to the knowledge graph subgraph, the target fault is taken as the root node of the network, and each accident cause is taken as a child node in the network. The root node and the child node are connected by directed edges. The relationship between each directed edge and the target fault and the corresponding accident cause in the knowledge graph subgraph is determined; S72, assign a priori probabilities to the sub-nodes corresponding to each accident cause, expressed as: P(r)=P(r|p) Where P(r) is the prior probability, r is the cause of the accident, p is the accident, and P(r|p) is the probability that the accident is caused by the cause of the accident; S73, assigning prior probabilities according to the child nodes and constructing a conditional probability table using the Noisy-or model; S74, using the GeNIe tool to import the knowledge graph subgraph into the tool, and input the defined Bayesian network structure, the prior probability of molecular node assignment and the conditional probability table to establish a fault diagnosis model.

8. The wind power equipment accident diagnosis method based on knowledge graph according to claim 7, characterized in that: In step S73, the Noisy-or model is used to construct a conditional probability table, including defining the association strength between the parent node and the child node, and the expression is: P w =P(s=T|g w =Tonly) =P(s=T|g1=F,g2=F,...,g w =T,...g m =F) Where P w is the degree of association between nodes, s is the child node, g w is the w-th parent node set, g w =Tonly means only g w The nodes take truth values, T and F represent Boolean values ​​true and false respectively; According to the strength of association, the parameter values ​​in the conditional probability table are determined, and the expression is: Where pa(s) is all parent nodes of s, pa(s) + is the parent node of all true values ​​of s.

9. The wind power equipment accident diagnosis method based on knowledge graph according to claim 8, characterized in that: In the step S7, the cause of the target fault is diagnosed and an accident analysis report is generated, wherein the target fault is input into the fault diagnosis model, and based on the Bayesian network diagnostic reasoning, all causes of the fault and the corresponding probabilities are output, the fault cause with the highest probability is obtained, and it is used as the cause propagation chain, and the corresponding cause propagation chain is retrieved in the knowledge graph using the Cypher language to obtain the knowledge graph subgraph of the corresponding processing solution, and an accident analysis report is generated.

10. A wind power equipment accident diagnosis system based on knowledge graph, characterized by: The system comprises: The acquisition module is used to obtain wind power equipment operation and maintenance data and standard specification data, and pre-process the data to build a wind power equipment safety data set; A processing module is used to extract keywords and annotate text data in the wind power equipment safety data set, build a keyword corpus, and perform word vector training on the annotated text to obtain a wind power equipment safety data word vector set; The entity recognition module is used to build an entity recognition model based on the BiLSTM-CRF algorithm combined with the similarity algorithm, input the wind power equipment safety data word vector set into the entity recognition model for training, and output the entity recognition prediction result; The knowledge graph construction module is used to extract relations and attributes from wind power equipment safety data, determine knowledge triples based on entity recognition prediction results, and construct a wind power equipment safety knowledge graph; The knowledge graph completion module is used to vectorize the triple data in the knowledge graph based on the TransE algorithm, analyze the features between entities and relationships, and complete the knowledge graph. The input module is used to input the target fault, obtain all the accident causes of the entity corresponding to the target fault based on the completed knowledge graph, and construct the knowledge graph subgraph; The diagnosis module is used to establish a fault diagnosis model based on the knowledge graph subgraph and the Bayesian network, diagnose the cause of the target fault, and generate an accident analysis report.

Images