Electromechanical Equipment Fault Prediction Method Based on Dynamic Graph Structure Embedding Model

Abstract

A method for predicting faults of electromechanical equipment based on a dynamic graph structure embedding model belongs to the fields of fault prediction of electromechanical equipment and fault prediction of knowledge graphs. First, the parallel structure information modeling of electromechanical equipment is realized through an ontology construction method, and quadruples are extracted from structured data and unstructured data to construct a time-series knowledge graph of electromechanical equipment. Secondly, the information representation for node classification is learned, and a feature vector is generated for each relevant node. Then, the time evolution information of the operation of electromechanical equipment is captured from the vector sequence. Finally, the hidden state vector of the required time period is output to realize the prediction of the time-series knowledge graph of electromechanical equipment. The method of the present invention specifically constructs a time-series knowledge graph for heterogeneous data of electromechanical equipment for reasoning to achieve fault prediction.

Description

Technical Field

[0001] The present invention belongs to the field of electromechanical equipment fault prediction and knowledge graph fault prediction, and specifically relates to a fault prediction method for electromechanical equipment based on temporal knowledge graph ontology modeling and dynamic graph structure embedding model. Background Art

[0002] Innovations in Industry 4.0 have led to rapid developments in monitoring technology and the generation of vast amounts of data. This has fueled the development of sophisticated artificial intelligence algorithms for big data analysis. This vast amount of heterogeneous data has made fault prediction for electromechanical equipment possible. Currently, processing this vast amount of heterogeneous data remains a pressing challenge. The concept of knowledge graphs was first proposed by Google in 2012. The core idea of knowledge graphs is to represent information as entities and their relationships, thereby forming a structured data network. This concept has since been widely applied in various fields, including but not limited to search engines, question-answering systems, recommendation systems, and fault prediction.

[0003] The application of knowledge graphs to electromechanical equipment can explore potential relationships between heterogeneous data, enabling fault prediction, ensuring both system operation and reliability. A method for constructing domain knowledge graphs for complex mechanical equipment has been proposed, combining them with graph neural networks to achieve fault prediction. A neural network-based named entity recognition model for coal mine equipment maintenance has been proposed, using it to construct a knowledge graph for coal mine equipment fault prediction.

[0004] Neural ODEs (NDOEs) are a method that uses neural network structures to parameterize solutions to ODEs. Since their introduction in 2018, this method has garnered widespread attention in the field of dynamic system modeling. Some have applied NDOEs to supervised learning and time series prediction for image datasets. While these methods utilize NDOEs, they do not consider knowledge graph embedding for prediction. Others have implemented the application of ODEs to knowledge graph embedding by embedding knowledge graph nodes onto ODE trajectories and constructing neural networks.

[0005] Although the above methods adjust the graph embedding structure for specific problems or directly obtain fault knowledge in the target domain, their research has made certain contributions to the field of fault prediction. However, in the field of electromechanical equipment fault prediction, the current challenges include: (1) Most neural network models only process one data type. Since electromechanical equipment data includes text data, sequence data, etc., it is difficult to achieve accurate fault prediction by relying on only a single type of data. (2) When processing electromechanical equipment knowledge graphs containing time information, traditional graph structure embedding models often find it difficult to quickly filter out key information from huge amounts of data. Summary of the Invention

[0006] To overcome the deficiencies of the existing technologies, the present invention proposes a method for predicting faults of electromechanical equipment based on a dynamic graph structure embedding model. First, a parallel structure information model of the electromechanical equipment is established through an ontology construction method, and quadruples are extracted from structured and unstructured data to construct a time-series knowledge graph of the electromechanical equipment. Secondly, information representations for node classification are learned, and feature vectors are generated for each relevant node. Then, the time evolution information of the operation of the electromechanical equipment is captured from the vector sequence; finally, the hidden state vectors for the required time period are output to realize the prediction of the time-series knowledge graph of the electromechanical equipment; this method specifically constructs a time-series knowledge graph for heterogeneous data of electromechanical equipment for reasoning to achieve fault prediction.

[0007] To solve the above technical problems, the present invention proposes a method for predicting faults of electromechanical equipment based on a dynamic graph structure embedding model, and the method includes the following steps:

[0008] The first step is to construct a time-series knowledge graph of the electromechanical equipment, that is, to realize the parallel structure information modeling of the electromechanical equipment through an ontology construction method, and extract quadruples from structured and unstructured data to construct a time-series knowledge graph of the electromechanical equipment;

[0009] The second step is to use a dynamic graph convolutional embedding model to realize the dynamic embedding of the time-series knowledge graph of the operation and maintenance of the electromechanical equipment, and the process is as follows:

[0010] Step (2.1) uses a dynamic graph convolutional model to obtain the tail entity embedding vector of the knowledge graph at the query time;

[0011] Step (2.2) uses the dynamic graph convolutional model in step (2.1) to obtain the head entity embedding vector of the knowledge graph at the query time;

[0012] The third step is to obtain the optimal result of fault prediction through a scoring function, and the process is as follows:

[0013] Step (3.1) uses a scoring function to calculate the scores of all quadruples accompanied by candidate entities;

[0014] Step (3.2) calculates the ranking of the prediction result as the judgment basis for comparing the model effects;

[0015] Step (3.3) selects the object with the highest score as the fault prediction result of the model.

[0016] Furthermore, the process of the first step is as follows:

[0017] Step (1.1) Ontology structure construction. The ontology of the time-series knowledge graph ETKG of the electromechanical equipment is defined as follows,

[0018] ;

[0019] Among them, E represents the set of entities; R represents the set of relationships between these entities; T represents the set of time stamps corresponding to the entities and relationships. Among them, E includes three major categories. C represents the systems and components of the equipment; F represents the faults of the equipment; M represents the maintenance of the equipment.

[0020] Step (1.2) Data collection;

[0021] Step (1.3) Data preprocessing. Use ETL technology to preprocess the data in the previous step, remove invalid data, convert all raw data into a data model, and uniformly store it in a relational database. Among them, ETL technology is used to extract data from one or more sources, and after conversion, it is loaded into the target database.

[0022] Step (1.4) Entity and relationship extraction. Automatically extract entities and relationships from a large amount of structured and unstructured data.

[0023] Preferably, the process of step (1.2) is as follows:

[0024] Step (1.2.1) Collect structured operation data in the control cabinet of the electromechanical equipment, including vibration signals, fault codes, operation data, electrical parameters, temperature, and humidity; use the TCP / IP communication protocol to remotely obtain the real-time operation data of the electromechanical equipment in the sensor; among them, all sensors are connected to the equipment and the gateway in a serial port manner, and data communication is carried out through the Modbus and RS485 protocols.

[0025] Step (1.2.2) Collect unstructured operation and maintenance data in the Internet of Things system of the electromechanical equipment, including maintenance records, repair records, fault causes, and fault phenomena; use web crawler technology to obtain unstructured data on fault information in the field of electromechanical equipment from the Internet of Things system of the electromechanical equipment.

[0026] More preferably, the process of step (1.4) is as follows:

[0027] Step (1.4.1) For structured vibration data Use continuous wavelet transform to achieve entity and relationship extraction;

[0028] Step (1.4.2) For unstructured text data, use Word2vec, Bi-LSTM, and CRF technologies to achieve entity and relationship extraction;

[0029] Step (1.4.3) Classify entities according to the defined ontology. Based on the obtained entities and relationships, draw a time-series knowledge graph of electromechanical equipment in the graph database, and combine the graph database language to effectively query and update the entities and relationships in the graph database.

[0030] More preferably, the process of step (1.4.1) is as follows:

[0031] Step (1.4.1.1) Vibration data After continuous wavelet transform,

[0032] ;

[0033] ;

[0034] Among them, is a wavelet basis function with two parameters a and b, represents the coefficient of wavelet transform, and this coefficient is obtained by using the wavelet basis function to perform integral transform on the vibration data ; represents the conjugate complex root of the wavelet basis function, is the scale parameter, is the translation parameter; ∫ is the integral operation, which is used in mathematics and statistics;

[0035] Step (1.4.1.2) Calculate the energy spectrum,

[0036] ;

[0037] Among them, the energy spectrum describes the energy distribution at different scales and is used to identify the main frequency components; Step (1.4.1.3) Extreme point detection realizes the extraction of fault entities;

[0038] By detecting local extreme points of, determine the sudden events or abnormal vibrations in the signal, and perform entity extraction according to the ontology classification in the first step;

[0039] Step (1.4.1.4) Calculate the co-occurrence matrix to realize relationship extraction,

[0040] ;

[0041] Construct the co-occurrence matrix of wavelet coefficients at different scales and analyze the relationship between different frequency components.

[0042] More preferably, the process of step (1.4.2) is as follows:

[0043] Step (1.4.2.1) Text data embedding:

[0044] The text data needs to be vectorized, mapping the corresponding character sequence to a sequence of word vectors. Through the word embedding method Word2vec, by continuously training and learning to update the word embedding lookup table, finally the input text is matched through the lookup table and converted into the corresponding sequence of word vectors;

[0045] Step (1.4.2.2) The text embedding vector is processed by the Bi-LSTM layer;

[0046] First, randomly initialize multiple initial parameters used in the Bi-LSTM layer. Input the sequence of text embedding vectors output by the embedding layer into the forward LSTM layer and the backward LSTM layer models to learn the context information of the sequence vectors. Finally, concatenate the output vectors in both directions as the output sequence of the Bi-LSTM layer, and introduce the dropout technique to prevent overfitting;

[0047] Step (1.4.2.3) The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0048] The CRF layer considers the relationship between adjacent labels to obtain the globally optimal sequence, models the dependencies between entities based on the output of the Bi-LSTM layer, and selects the sequence with the highest probability from the label space;

[0049] Step (1.4.2.4) Calculate the loss function output by the model to obtain the optimal result;

[0050] Take the output of the CRF layer as the labeled sequence of the final output of the model, and compare it with the actually labeled sequence to calculate the Loss function,

[0051] ;

[0052] Among them, represents the probability of the output sequence Y when the input sequence X is given, represents the scoring function when the input sequence and the label sequence are given. log is the logarithmic operation, and exp is the natural exponential operation;

[0053] By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained.

[0054] More preferably, the process of step (1.4.3) is as follows:

[0055] Step (1.4.3.1) Structured data format conversion;

[0056] Export the structured data processed in step (1.4.1) from the relational database as a csv format. To express the temporal characteristics of the process state data, the traditional triple form is extended to a quadruple with temporal information , where t represents the time information when the triple is established. However, currently the graph database only supports the triple storage form, so the temporal information is converted and embedded into the relationship between entities to establish a temporal knowledge graph of electromechanical equipment based on the graph database;

[0057] Step (1.4.3.2) Unstructured data format conversion;

[0058] Convert the entities and entity relationships extracted in step (1.4.2) into a csv file format. Since the extracted data are all text, uniformly convert the csv file encoding format to UTF-8 encoding;

[0059] Step (1.4.3.3) Import the csv file data into the graph database.

[0060] Furthermore, in the said step (2.1), use the dynamic graph convolutional model to obtain the tail entity embedding vector of the knowledge graph at the query time , and the process is as follows: :

[0061] Step (2.1.1) Embedding of the tail entity based on the traditional graph convolutional model,

[0062] ;

[0063] Among them, is the adjacency matrix, is the hidden state vector of the tail entity at the k+1 layer at time t, is the set of head entities connected to the tail entity through the relationship r in the subgraph at time t, that is , is the normalization constant, represents the hidden state vector of the head entity at the k layer, is the diagonal relationship weight matrix at the k layer, represents the hidden state vector of the initial state embedding, and the activation function uses ReLU, is the trainable parameter matrix;

[0064] Step (2.1.2) Improve the traditional graph convolutional model based on the neural ordinary differential equation;

[0065] In step (2.2), the dynamic graph convolution model in step (2.1) is used to obtain The head entity embedding vector of the knowledge graph .

[0066] Preferably, the process of step (2.1.2) is as follows:

[0067] Step (2.1.2.1) decomposes the propagation depth k of traditional graph convolution into the integration time and step length The combination of , by increasing the variable integration step size dt, the neural ordinary differential equation representation of graph convolution is realized.

[0068] ;

[0069] in, is the hidden state vector of the tail entity at time t, is the hidden state vector embedded in the initial state of the tail entity, For the head entity, the range is 0 to The hidden state vector at time t, is the diagonal relationship weight matrix;

[0070] Step (2.1.2.2) uses Taylor expansion to convert the ordinary differential equation into a differential equation.

[0071] ;

[0072] in, is the hidden state vector of the head entity at time t. Taylor expansion approximates the function with a polynomial near a certain point, which is used to approximate the solution of complex functions in differential equations. is the identity matrix;

[0073] Step (2.1.2.3) , a set of parameters Θ is used to characterize the entire continuous time aggregation process, and the complete tail entity embedding vector is obtained by summing,

[0074] ;

[0075] in, It is a black-box ODE solver for solving neural ordinary differential equations.

[0076] Furthermore, in step (3.1), the score function is used to calculate the scores of all quadruplets accompanying the candidate entity.

[0077] ;

[0078] in, is the query time, represents transpose, is the set of candidate entities, and ;

[0079] In the step (3.2), calculate the ranking of the prediction result as the judgment basis for comparing the model effects,

[0080] ;

[0081] In the step (3.3), select the object with the highest score as the fault prediction result of the model,

[0082] ;

[0083] where arg max means to find the one that makes the largest , is the fault entity of the final prediction.

[0084] The beneficial effects of the present invention are mainly manifested in:

[0085] 1. For the structured data and unstructured data of electromechanical equipment, different methods are respectively adopted for processing, providing data support for the construction of the knowledge graph.

[0086] 2. Define the ontology structure of the electromechanical equipment time series knowledge graph, and extract the entities, relationships, and time of the electromechanical equipment data according to the ontology structure by corresponding methods to form quadruples, realizing the construction of the electromechanical equipment time series knowledge graph.

[0087] 3. Combine the neural ordinary differential equation with the graph convolutional neural network to construct a dynamic graph structure embedding model, continuously process the topological structure of the knowledge graph in time and embed it as a hidden state vector, realizing the in-depth understanding and feature extraction of the electromechanical equipment data. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 is a schematic diagram of the ontology structure of the electromechanical equipment time series knowledge graph defined by the present invention taking an elevator as an example. DETAILED DESCRIPTION OF THE INVENTION

[0089] The present invention will be further described below with reference to the drawings.

[0090] Refer to Figure 1 , a method for predicting faults of electromechanical equipment based on a dynamic graph structure embedding model, the method comprising the following steps:

[0091] The first step is to construct an electromechanical equipment time series knowledge graph, and the process is as follows:

[0092] Step (1.1) Ontology structure construction, as Figure 1 shown, the ontology of the Electromechanical Equipment Time-series Knowledge Graph (ETKG) is defined as follows,

[0093] ;

[0094] Among them, E represents the set of entities; R represents the set of relationships between these entities; T represents the set of time stamps corresponding to entities and relationships. Among them, E includes three major categories. C represents the systems and components of the equipment; F represents the faults of the equipment; M represents the maintenance of the equipment;

[0095] Among them, the ontology structure is a method commonly used in knowledge graph construction, which defines the standard terms related to the operation of electromechanical equipment and their hierarchical structures, and identifies the implicit associations between these terms;

[0096] Step (1.2) Data collection, the process is as follows:

[0097] Step (1.2.1) Collect structured operation data in the control cabinet of electromechanical equipment, including vibration signals, fault codes, operation data, electrical parameters, temperature and humidity; use the TCP / IP communication protocol to remotely obtain the real-time operation data of electromechanical equipment in the sensor. Table 1 gives the corresponding data collection information taking the elevator as an example;

[0098] Table 1

[0099]

[0100] Among them, all sensors are connected to devices and gateways in a serial port manner, and data communication is carried out through Modbus and RS485 protocols;

[0101] Step (1.2.2) Collect unstructured operation and maintenance data in the electromechanical equipment Internet of Things system, including maintenance records, repair records, fault causes, and fault phenomena;

[0102] Use web crawling technology to obtain unstructured data on fault information in the field of electromechanical equipment from the electromechanical equipment Internet of Things system;

[0103] Among them, web crawling technology, also known as web spider or web robot, is an automated program used to browse and collect Internet data;

[0104] Step (1.3) Data preprocessing

[0105] Use ETL (Extract, Transform, Load) technology to preprocess the collected data, remove invalid data, convert all the original data into a data pattern, and uniformly store it in a relational database. Among them, ETL technology is used to extract data from one or more sources, and after transformation, load it into the target database;

[0106] Among them, ETL technology is a common data integration technology, which is used to extract data from one or more sources, and after transformation, load it into the target database or data warehouse;

[0107] Step (1.4) Entity and relationship extraction, automatically extract entities and relationships from a large amount of structured and unstructured data; the process is as follows:

[0108] Step (1.4.1) For structured vibration data Adopt continuous wavelet transform to realize the extraction of entities and relationships, and the process is as follows:

[0109] Step (1.4.1.1) Vibration data After continuous wavelet transform,

[0110] ;

[0111] ;

[0112] Among them, is a wavelet basis function with two parameters a and b, represents the coefficient of wavelet transform, and this coefficient is obtained by using the wavelet basis function to perform integral transform on the vibration data ; represents the conjugate complex root of the wavelet basis function, is the scale parameter, is the translation parameter; ∫ is the integral operation, which is used in mathematics and statistics;

[0113] Step (1.4.1.2) Calculate the energy spectrum,

[0114] ;

[0115] Among them, the energy spectrum describes the energy distribution at different scales and is used to identify the main frequency components; Step (1.4.1.3) Extreme point detection realizes the extraction of fault entities;

[0116] By detecting local extreme points of, determine the sudden events or abnormal vibrations in the signal, and perform entity extraction according to the ontology classification in the first step;

[0117] Step (1.4.1.4) calculates the co-occurrence matrix to realize relationship extraction,

[0118] ;

[0119] Construct the co-occurrence matrix of wavelet coefficients at different scales , analyze the relationships between different frequency components, as Figure 1 shown;

[0120] Step (1.4.2) For unstructured text data, use Word2vec (Word to Vector), Bi-LSTM (Bidirectional Long Short-Term Memory), and CRF (Conditional Random Field) technologies to realize entity and relationship extraction. The process is as follows:

[0121] Step (1.4.2.1) Text data embedding:

[0122] The text data needs to be vectorized, mapping the corresponding character sequence to a sequence of word vectors. In the present invention, through the word embedding method Word2vec, by continuously training and learning to update the word embedding lookup table, finally the input text is matched through the lookup table and converted into the corresponding sequence of word vectors;

[0123] Among them, Word2vec is a commonly used algorithm for natural language processing. By training a neural network model to generate word vectors, it maps vocabulary to a continuous vector space and captures the semantic and syntactic relationships between vocabulary;

[0124] Step (1.4.2.2) The text embedding vector is processed through the Bi-LSTM layer;

[0125] First, randomly initialize multiple initial parameters used in the Bi-LSTM layer, input the sequence of text embedding vectors output by the embedding layer into the forward LSTM layer and the backward LSTM layer models to learn the context information of the sequence vectors, and finally splice the output vectors in the two directions as the output sequence of the Bi-LSTM layer, and introduce the dropout technology to prevent overfitting;

[0126] Among them, dropout is a commonly used regularization technique that prevents overfitting by randomly dropping neurons. LSTM is a commonly used recurrent neural network that controls the storage and reading of information by introducing memory units and gating mechanisms, thereby maintaining and updating information over long time spans. Bi-LSTM is an extended version of LSTM that captures bidirectional context information of sequence data by using two LSTMs in the forward and reverse directions of the time series respectively;

[0127] Step (1.4.2.3): The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0128] The CRF layer obtains the globally optimal sequence by considering the relationships between adjacent labels, models the dependencies between entities based on the output of the Bi-LSTM layer, and selects the sequence with the highest probability from the label space;

[0129] Among them, CRF is a probabilistic graphical model used for annotating and segmenting sequence data, which can consider context information and global features, thereby improving the annotation accuracy;

[0130] Step (1.4.2.4): Calculate the loss function of the model output to obtain the optimal result;

[0131] Take the output of the CRF layer as the labeled sequence of the final output of the model, compare it with the actually annotated sequence, and calculate the Loss function,

[0132] ;

[0133] Among them, represents the probability of the output sequence Y given the input sequence X, represents the scoring function given the input sequence and the label sequence , where log is the logarithmic operation and exp is the natural exponential operation.

[0134] By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained;

[0135] Step (1.4.3): Classify entities according to the ontology defined in Figure 1 . The graph database language can be Cypher, etc., to achieve effective querying and updating of the graph database. The graph database can use Neo4j, which is a high-performance graph database that stores data using a graph structure and is suitable for processing highly connected data. Finally, an electromechanical device time series knowledge graph is drawn in Neo4j; the process is as follows:

[0136] Step (1.4.3.1): Structured data format conversion;

[0137] Export the structured data processed in step (1.4.1) from a relational database (such as Oracle) into csv format. To express the temporal characteristics of process state data, the traditional triple form is extended to a quadruple with temporal information , where t represents the time information when the triple is established. However, currently, graph databases (such as Neo4j) only support the triple storage form. Therefore, the present invention converts the temporal information and embeds it into the relationships between entities to establish a temporal knowledge graph of electromechanical equipment based on the graph database;

[0138] Among them, a triple refers to a set containing three elements in data structures and databases, and is often used to represent entity relationships in relational databases;

[0139] Step (1.4.3.2) Unstructured data format conversion;

[0140] Convert the entities and entity relationships extracted in step (1.4.2) into csv file format. Since the extracted data are all text, the present invention uniformly converts the csv file encoding format to UTF-8 encoding;

[0141] Step (1.4.3.3) Import quadruple data into Neo4j;

[0142] Use a graph database language (such as Cypher) to import the csv file. For example, first store the sorted entity and relationship csv files in the import folder under the Neo4j root path, and use the "LOAD CSV" statement in Cypher to implement the import of the quadruple csv file into Neo4j;

[0143] Second, use a dynamic graph convolutional embedding model to achieve dynamic embedding of the temporal knowledge graph of electromechanical equipment operation and maintenance. The process is as follows:

[0144] Step (2.1) Use a dynamic graph convolutional model to obtain the embedding vector of the tail entity of the knowledge graph at the query time , and the process is as follows:

[0145] Step (2.1.1) Embedding of the tail entity based on the traditional graph convolutional model

[0146] ;

[0147] Among them, is the adjacency matrix, is the hidden state vector of the tail entity at the k + 1 layer at time t, is the subgraph at time t and is related to the tail entity ​The set of head entities connected by the relationship r, i.e., where , is the normalization constant, represents the hidden state vector of the head entity at layer k, and is the diagonal relation weight matrix at layer k, represents the hidden state vector of the initial state embedding, and the activation function is ReLU, is the trainable parameter matrix;

[0148] Among them, traditional graph convolution is a common method for knowledge graph embedding;

[0149] Step (2.1.2) improves the traditional graph convolution model based on neural ordinary differential equations as follows:

[0150] Step (2.1.2.1) decomposes the propagation depth k of traditional graph convolution into the combination of integration time and step size , i.e., , and realizes the neural ordinary differential equation representation of graph convolution by increasing the variable integration step dt,

[0151] ;

[0152] Among them, is the hidden state vector of the tail entity at time t, is the hidden state vector of the initial state embedding of the tail entity, is the hidden state vector of the head entity from 0 to time, is the diagonal relation weight matrix.

[0153] Step (2.1.2.2) uses Taylor expansion to change the ordinary differential equation representation to the differential equation form,

[0154] ;

[0155] Among them, is the hidden state vector of the head entity at time t. Taylor expansion approximates the function near a certain point with a polynomial, which is used to approximate the solution of the complex function in the differential equation, is the identity matrix.

[0156] Step (2.1.2.3) In the present invention, by setting , a set of parameters Θ is used to characterize the entire continuous-time aggregation process, and the complete tail entity embedding vector is obtained by summation,

[0157] ;

[0158] Among them, is a black box ODE solver for solving neural ordinary differential equations;

[0159] Step (2.2) uses the dynamic graph convolution model in step (2.1) to obtain the head entity embedding vector of the knowledge graph at the query time ; ;

[0160] Thirdly, obtain the optimal result of fault prediction through a scoring function, and the process is as follows:

[0161] Step (3.1) uses a scoring function to calculate the scores of all quadruples accompanied by candidate entities,

[0162] ;

[0163] Among them, is the query time, represents transpose, is the set of candidate entities, and ;

[0164] Step (3.2) calculates the ranking of the prediction result as the judgment basis for comparing the model effects,

[0165] ;

[0166] Step (3.3) selects the object with the highest score as the fault prediction result of the model,

[0167] ;

[0168] Among them, arg max means to find the that makes the largest, is the finally predicted fault entity.

[0169] The implementation process of the electromechanical equipment fault prediction method based on the dynamic graph structure embedding model in this embodiment includes the following steps:

[0170] Step 1: Give 10 knowledge graph reasoning models:

[0171] TransE: This method is a static knowledge graph reasoning model based on embedding. It learns the vector representations of entities and relationships by minimizing the difference between the vectors of the head entity and the tail entity of a triple.

[0172] ComplEx: This method is a static knowledge graph embedding model that uses complex number embeddings (real and imaginary parts) to represent entities and relationships.

[0173] DistMult: This method is a simple yet effective static knowledge graph embedding model that captures relationships between entities using symmetric bilinear products.

[0174] TTransE: This method is an extension of the TransE model and is an interpolation method for temporal knowledge graphs. By treating time constraints as part of the triples, it can effectively predict past or future relationships.

[0175] TNTComplEx: This method is an extension of the ComplEx model and is an interpolation method for temporal knowledge graphs. By introducing time factors into the embeddings of entities and relationships, the model can effectively predict time-varying relationships and events in the knowledge graph.

[0176] TA-DistMult: This method is an extension of the DistMult model and is an interpolation method for temporal knowledge graphs.

[0177] By introducing a time factor for each triple, the embeddings can reflect the changes of entities and relationships over time.

[0178] RE-NET: This method is a model designed specifically for extrapolation of temporal knowledge graphs. It captures the evolution of entities and relationships over time by integrating recurrent neural networks.

[0179] RE-GCN: This method combines the advantages of graph convolutional networks and recurrent neural networks. It can capture both the structural information and the time evolution information of the knowledge graph and is a temporal knowledge graph extrapolation model.

[0180] CENET: This method is a temporal knowledge graph event prediction model based on a historical contrast learning framework. It can effectively distinguish potential entities that match the query and is a temporal knowledge graph extrapolation model.

[0181] ETKGFP: The temporal knowledge graph extrapolation method of the present invention. It integrates neural ordinary differential equations and graph convolutional neural networks to dynamically embed the temporal knowledge graph and achieve fault prediction of electromechanical devices.

[0182] Step 2: Experimental dataset

[0183] The experimental dataset is provided by the electromechanical device temporal knowledge graph constructed by the ontology construction method proposed by the present invention, and it contains 12,792 entities and 154 semantic relationships.

[0184] Step 3: Define evaluation metrics

[0185] To verify the prediction performance of each model, the present invention conducts experiments using the following evaluation metrics:

[0186] Hit@n: This metric measures the proportion of correct predictions among all test samples that fall within the top n highest rankings given by the model; smaller values of n emphasize the accuracy of the model in predicting the most relevant options, while larger values of n take more into account the comprehensiveness and diversity of the model.

[0187] ;

[0188] Among them, represents the ranking number of the quadruple, and N is the total number of candidate entities. In this embodiment, Hit@1, Hit@3, and Hit@10 are selected for evaluation.

[0189] MRR: A commonly used international mechanism for evaluating matching algorithms. The calculation principle is that the higher the ranking of the correct result, the higher the score. If there is no correct result, the score is 0. The formula is as follows.

[0190] ;

[0191] Step 4. Analyze the comparison results

[0192] Referring to Table 2, the comparison results of four metrics, namely MRR, Hit@1, Hit@3, and Hit@10, for the fault category entity prediction task are shown.

[0193] Table 2

[0194]

[0195] It can be observed that the three static inference models, TransE, DistMult, and ComplEx, all perform worse than the temporal model in terms of effectiveness because static models cannot consider temporal information and the dependencies between entities at different time snapshots. In the case of temporal inference models, the interpolation models TTtransE, TA-Distmult, and TNTComplEx aim to complete missing relationships and entities rather than predict new events, so they exhibit worse performance than the extrapolation models. Among the extrapolation models, CENET has the best performance in fault entity prediction among the baseline models because CENET, like the model of the present invention, also makes predictions based on the historical interactions between entities, rather than relying highly on the periodicity or recurrence of events like RE-GCN and RE-NET. The contrastive learning method adopted by CENET focuses on the similarities and differences between entities, which makes its dynamic modeling ability completely inferior to ETKGFP that introduces neural ordinary differential equations. The method proposed by the present invention achieves the best results in all four metrics, which are 41.54%, 32.84%, 48.86, and 57.65% respectively.

[0196] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept and is for illustrative purposes only. The protection scope of the present invention should not be regarded as limited to the specific forms stated in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be conceived by those of ordinary skill in the art based on the inventive concept of the present invention.

Claims

1. A method for predicting faults of electromechanical equipment based on a dynamic graph structure embedding model, characterized in that The method includes the following steps: First step: Construct a time-series knowledge graph of electromechanical equipment, that is, realize the parallel structure information modeling of electromechanical equipment through ontology construction methods, and extract quadruples from structured and unstructured data to construct a time-series knowledge graph of electromechanical equipment. The entities, relationships, and time of electromechanical equipment data form quadruples; Second step: Use a dynamic graph convolutional embedding model to realize the dynamic embedding of the time-series knowledge graph of electromechanical equipment operation and maintenance. The process is as follows: Step (2.1): Use a dynamic graph convolutional model to obtain the tail entity embedding vector of the knowledge graph at the query time; Step (2.2): Use the dynamic graph convolutional model in step (2.1) to obtain the head entity embedding vector of the knowledge graph at the query time; Third step: Obtain the optimal result of fault prediction through a scoring function. The process is as follows: Step (3.1): Use a scoring function to calculate the scores of all quadruples accompanied by candidate entities; Step (3.2): Calculate the ranking of the prediction result as the judgment basis for comparing the model effects; Step (3.3): Select the object with the highest score as the fault prediction result of the model; In the step (2.1), a dynamic graph convolutional model is used to obtain the embedding vector of the tail entity of the knowledge graph at the query time and the process is as follows: for the tail entity embedding vector of the knowledge graph Step (2.1.1): Embedding of the tail entity based on the traditional graph convolutional model, ； Among them, is the adjacency matrix, is t the hidden state vector of the tail entity at time in k+ layer 1, is t the set of head entities in the subgraph at time connected to the tail entity r through the relation , that is , is the normalization constant, represents k the hidden state vector of the head entity at layer , is k the diagonal relation weight matrix at layer represents the hidden state vector of the initial state embedding, and the activation function uses ReLU, is the trainable parameter matrix; Step (2.1.2): Improve the traditional graph convolutional model based on neural ordinary differential equations; In the step (2.2), the head entity embedding vector of the knowledge graph at the query time is obtained by using the dynamic graph convolutional model in the step (2.1). ;​ The process of step (2.1.2) is as follows: Step (2.1.2.1) decomposes the propagation depth of the traditional graph convolution k into the integration time and the step size combination, that is , by increasing the variable integration step size d t to achieve the neural ordinary differential equation representation of the graph convolution. ； Among them, is the hidden state vector of the tail entity at time t is the hidden state vector of the initial state embedding of the tail entity, is the hidden state vector of the head entity from 0 to time t, is the diagonal relation weight matrix; Step (2.1.2.2): Use Taylor expansion to change the representation of the ordinary differential equation to the form of a differential equation, ； Among them, is the hidden state vector of the head entity at time t. Taylor expansion approximates a function near a certain point with a polynomial and is used to approximate the solution of a complex function in a differential equation. is the identity matrix; Step (2.1.2.3) Let , and represent the entire continuous-time aggregation process with a set of parameters Θ, and obtain the complete tail entity embedding vector by summation. ； Among them, is a black-box ODE solver for solving neural ordinary differential equations; In step (3.1), use a scoring function to calculate the scores of all quadruples accompanied by candidate entities, ； Among them, is the query time, represents transpose, is the candidate entity set, and ; In the step (3.2), calculate the ranking of the prediction result As the judgment basis for comparing the model effects ； In step (3.3), select the object with the highest score as the fault prediction result of the model, ； Among them, arg max means to find the one that makes the largest , is the finally predicted faulty entity.

2. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 1, wherein The process of the first step is as follows: Step (1.1): Ontology structure construction. The ontology of the time-series knowledge graph ETKG of electromechanical equipment is defined as follows, ； Among them, E represents a set of entities; R represents a set of relationships between these entities; T represents a set of time stamps corresponding to entities and relationships, where E includes three major categories, C represents the systems and components of the device; F represents the faults of the device; M represents the maintenance of the device; Step (1.2): Data collection; Step (1.3): Data preprocessing. Use ETL technology to preprocess the collected data, remove invalid data, convert all raw data into data patterns, and uniformly store them in a relational database. Among them, ETL technology is used to extract data from one or more sources, and after conversion, load it into the target database; Step (1.4): Entity and relationship extraction. Automatically extract entities and relationships from a large amount of structured and unstructured data.

3. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 2, wherein, The process of step (1.2) is: Step (1.2.1): Collect structured operation data in the electromechanical equipment control cabinet, including vibration signals, fault codes, operation data, electrical parameters, temperature, and humidity; use the TCP / IP communication protocol to remotely obtain the real-time operation data of the electromechanical equipment in the sensor; among them, all sensors are connected to the equipment and the gateway in a serial port manner, and data communication is carried out through the Modbus and RS485 protocols; Step (1.2.2): Collect unstructured operation and maintenance data in the electromechanical equipment Internet of Things system, including maintenance records, maintenance records, fault causes, and fault phenomena; use web crawler technology to obtain unstructured data on fault information in the electromechanical equipment field from the electromechanical equipment Internet of Things system.

4. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 2, wherein The process of step (1.4) is: Step (1.4.1) For structured vibration data Use continuous wavelet transform to achieve the extraction of entities and relationships; In step (1.4.2), for unstructured text data, technologies such as Word2vec, Bi-LSTM, and CRF are used to achieve entity and relationship extraction; In step (1.4.3), entities are classified according to the ontology defined in step 1.

1. Based on the obtained entities and relationships, a time-series knowledge graph of electromechanical equipment is drawn in the graph database, and the entities and relationships in the graph database are effectively queried and updated in combination with the graph database language.

5. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 4, wherein The process of step (1.4.1) is as follows: Step (1.4.1.1) Vibration data After continuous wavelet transform, ； ； Among them, is a wavelet basis function with two parameters a and b, represents the coefficient of wavelet transform, which is obtained by using the wavelet basis function to perform integral transformation on the vibration data ; represents the conjugate complex root of the wavelet basis function, is the scale parameter, is the translation parameter; ∫ is the integral operation, which is used in mathematics and statistics; Step (1.4.1.2): Calculate the energy spectrum. ； Among them, the energy spectrum describes the energy distribution at different scales and is used to identify the main frequency components; the extreme point detection in step (1.4.1.3) realizes the extraction of fault entities; By detecting local extreme points of, identify sudden events or abnormal vibrations in the signal, and perform entity extraction according to the ontology classification in the first step; Step (1.4.1.4): Calculate the co-occurrence matrix to achieve relationship extraction. ； Construct the co-occurrence matrix of wavelet coefficients at different scales , and analyze the relationships between different frequency components.

6. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 4, wherein The process of step (1.4.2) is as follows: Step (1.4.2.1): Text data embedding: The text data needs to be vectorized. The corresponding character sequence is mapped to a sequence of word vectors. Through the word embedding method Word2vec, by continuously training and learning to update the word embedding lookup table, finally the input text is matched through the lookup table and converted into the corresponding sequence of word vectors; Step (1.4.2.2): The text embedding vector is processed through the Bi-LSTM layer; First, multiple initial parameters used in the Bi-LSTM layer are randomly initialized. The sequence of text embedding vectors output by the embedding layer is input into the forward LSTM layer and the backward LSTM layer models to learn the context information of the sequence vectors. Finally, the output vectors in the two directions are concatenated as the output sequence of the Bi-LSTM layer, and the dropout technology is introduced to prevent overfitting; Step (1.4.2.3): The output vector of the Bi-LSTM layer is processed through the CRF layer; The CRF layer considers the relationship between adjacent labels to obtain the globally optimal sequence, models the dependence between entities based on the output of the Bi-LSTM layer, and selects the sequence with the highest probability from the label space; Step (1.4.2.4): Calculate the loss function output by the model to obtain the optimal result; Take the output of the CRF layer as the labeled sequence finally output by the model, and compare it with the actually labeled sequence to calculate the Loss function. ； Among them, represents the probability of the output sequence X when the given input sequence is Y ; represents the scoring function when the given input sequence is and the label sequence is , where log is the logarithmic operation and exp is the natural exponential operation; By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained.

7. The electromechanical equipment fault prediction method based on the dynamic graph structure embedding model according to claim 4, wherein, The process of step (1.4.3) is as follows: Step (1.4.3.1): Structured data format conversion; Export the structured data processed in step (1.4.1) from the relational database as csv format. To express the temporal characteristics of the process status data, the traditional triple form is extended to a quadruple with temporal information , where t represents the time information when the triple is established. The graph database supports the triple storage form. The temporal information is converted and embedded into the relationship between entities to realize the establishment of the temporal knowledge graph of electromechanical equipment based on the graph database; Step (1.4.3.2): Unstructured data format conversion; Convert the entities and entity relationships extracted in step (1.4.2) into the csv file format. Since the extracted data are all text, uniformly convert the encoding format of the csv file to UTF-8 encoding; Step (1.4.3.3): Import the csv file data into the graph database.

Images