Predictive maintenance method for electromechanical equipment based on time series knowledge graph

Abstract

A predictive maintenance method for electromechanical equipment integrating a time series knowledge graph belongs to the field of predictive maintenance of electromechanical equipment. The model specifically performs reasoning on the time series knowledge graph of electromechanical equipment operation and maintenance to achieve predictive maintenance. First, the parallel structure information of the electromechanical equipment is obtained to learn the information node representation for node classification, and a feature vector is generated for each related node. Then, the time evolution information of the operation of the electromechanical equipment is captured from the vector sequence, and finally, the hidden state vector is output to participate in the construction of the conditional intensity function, and the prediction of the time series knowledge graph is achieved based on the intensity function to complete the predictive maintenance task of the electromechanical equipment. The present invention effectively realizes predictive maintenance of electromechanical equipment.

Description

Technical Field

[0001] The present invention belongs to the field of predictive maintenance of electromechanical equipment, and specifically relates to a predictive maintenance method for electromechanical equipment that integrates a time series knowledge graph with a neural network ordinary differential equation. Background Art

[0002] Monitoring and diagnosis of target systems and performing maintenance tasks in a timely and resource-efficient manner. Large amounts of heterogeneous data enable predictive maintenance of electromechanical equipment. Predictive maintenance can predict trends, behavior patterns, and correlations through statistical or machine learning models to predict failures that have not occurred in advance, thereby improving the decision-making process of maintenance activities and avoiding downtime. The concept of knowledge graph was first proposed by Google in 2012. The core idea of ​​knowledge graph is to represent information as entities and their relationships, thereby forming a structured data network.

[0003] The application of knowledge graphs on electromechanical equipment can explore the potential relationship between heterogeneous data, but the most important thing for predictive maintenance is time information. Conventional knowledge graphs cannot store triples containing time information, which leads to a temporal knowledge graph composed of quadruple, and predictive maintenance is achieved by reasoning on the temporal knowledge graph. At present, the research on temporal knowledge graph reasoning is very extensive. For example, Xu et al. combined the method of contrastive learning to achieve the prediction of the temporal knowledge graph. Bai et al. completed the temporal knowledge graph based on temporal logic rules. However, relevant research in the field of electromechanical equipment is temporarily missing, so the present invention constructs a temporal knowledge graph based on the characteristics of electromechanical equipment and improves the existing reasoning method to make it more suitable for research in the field of electromechanical equipment.

[0004] Neural ordinary differential equations are a method that uses a neural network structure to parameterize the solution of ordinary differential equations. This method has received widespread attention in the field of dynamic system modeling. Hawkes process is a self-excited point process model used to describe the time series of events. Its characteristic is that the occurrence of one event increases the probability of future events. Since it was proposed by Hawkes in 1971, this method has received widespread attention in fields such as earthquake prediction, social network analysis, and financial market modeling. Zhou et al. applied Hawkes process to information propagation and user behavior modeling in social networks. Mei et al. embedded the nodes of the knowledge graph into the Hawkes process model and constructed a neural network to realize the application of Hawkes process in knowledge graph embedding.

[0005] Although the above methods adjust the graph embedding structure or apply deep learning for predictive maintenance for specific problems, their research has made certain contributions in the field of predictive maintenance. However, the challenges of existing research include: (1) Most existing models are modeled in the discrete time domain, but the operation and maintenance data of electromechanical equipment are mostly irregularly sampled dynamic data, and the changes mainly appear in edge data such as failures and maintenance, and the corresponding main equipment does not change. (2) The length of the input sequence after dynamic graph convolution is too long, and conventional deep learning models cannot fully learn the graph structure. (3) The ultimate goal of predictive maintenance is to prevent the occurrence of downtime failures. For electromechanical equipment, downtime failures are caused by the accumulation of non-stop failures. Existing models do not take advantage of this feature. Summary of the invention

[0006] In order to overcome the shortcomings of the existing technology, the present invention provides a predictive maintenance method for electromechanical equipment that integrates a time series knowledge graph. The model specifically performs reasoning on the time series knowledge graph of electromechanical equipment operation and maintenance to achieve predictive maintenance. First, the parallel structure information of the electromechanical equipment is obtained to learn the information node representation for node classification, and a feature vector is generated for each related node. Then, the time evolution information of the electromechanical equipment operation is captured from the vector sequence, and finally, the hidden state vector is output to participate in the construction of the conditional intensity function, and the prediction of the time series knowledge graph is realized based on the intensity function to complete the predictive maintenance task of the electromechanical equipment.

[0007] In order to solve the above technical problems, the present invention proposes a predictive maintenance method for electromechanical equipment integrating a time series knowledge graph, and the method comprises the following steps:

[0008] The first step is to build a knowledge graph of electromechanical equipment timing. The process is as follows:

[0009] Step (1.1) ontology structure construction, defining the Electromechanical Equipment Temporal Knowledge Graph (ETKG) ontology;

[0010] Step (1.2) data collection;

[0011] Step (1.3) Data preprocessing: Use ETL technology to preprocess the collected data, remove invalid data, convert all raw data into database table structure relationships, and store them in a relational database. ETL technology is used to extract data from one or more sources and load it into the target database after conversion.

[0012] Step (1.4) Entity and relationship extraction, automatically extracting entities and relationships from large amounts of structured and unstructured data;

[0013] The second step is to use the dynamic graph convolution embedding model to realize the dynamic embedding of the electromechanical equipment timing knowledge graph. The process is as follows:

[0014] Step (2.1) uses a dynamic graph convolutional model to obtain the tail entity embedding vector of the knowledge graph at any time;

[0015] 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 query time;

[0016] The third step is to use the dynamic Hawkes Transformer model to achieve the prediction of maintenance entities. The process is as follows:

[0017] Step (3.1) encodes the entity embedding vector obtained in the second step using a dynamic position encoding module;

[0018] Step (3.2) The neural Hawkes decoder module implements maintenance entity prediction;

[0019] Step 4: Treat the entity prediction task as a multi-class classification task and the time prediction task as a regression task for model training. The process is as follows:

[0020] Step (4.1) uses cross entropy loss and mean square error loss as the losses for entity prediction tasks and time prediction tasks;

[0021] For the part of step (4.2) where the Neural Ordinary Differential Equation is used, the present invention adopts the adjoint sensitivity method for back propagation.

[0022] Furthermore, in the step (1.1), the electromechanical equipment timing knowledge graph ETKG ontology is defined as follows:

[0023] ;

[0024] Among them, E represents the set of entities; R represents the set of relationships between these entities; T represents the set of timestamps corresponding to entities and relationships, among which E includes three major categories: C represents the system and components of the equipment; F represents the failure of the equipment; and M represents the maintenance of the equipment.

[0025] Further, the process of step (1.2) is:

[0026] Step (1.2.1) collects structured operating data in the electromechanical equipment control cabinet, including vibration signals, fault codes, operating status data, operation data, electrical parameters, temperature and humidity; uses TCP / IP communication protocol to remotely obtain real-time operating data of electromechanical equipment in sensors; all sensors are connected to devices and gateways via serial ports, and data communication is carried out through Modbus and RS485 protocols;

[0027] Step (1.2.2) collects unstructured operation and maintenance data in the electromechanical equipment IoT system, including repair records, maintenance records, fault causes and fault phenomena; and uses crawler technology to obtain unstructured data on fault information in the electromechanical equipment field from the electromechanical equipment IoT system.

[0028] Furthermore, the process of step (1.4) is as follows:

[0029] Step (1.4.1) for structured vibration data Continuous wavelet transform is used to extract entities and relations;

[0030] Step (1.4.2) uses Word2vec, Bi-LSTM, and CRF technologies to extract entities and relationships for unstructured text data;

[0031] Step (1.4.3) uses the graph database language to import data and draw the electromechanical equipment timing knowledge graph in the graph database.

[0032] Preferably, the process of step (1.4.1) is as follows:

[0033] Step (1.4.1.1) Vibration data After continuous wavelet transform,

[0034] ;

[0035] ;

[0036] in, is a wavelet basis function with two parameters a and b, Represents the coefficients of wavelet transform, which are obtained by using wavelet basis functions Vibration data Perform integral transformation; represents the conjugate complex root of the wavelet basis function, is the scale parameter, is the translation parameter; ∫ is the integral operation, used in mathematics and statistics;

[0037] Step (1.4.1.2) calculates the energy spectrum as follows,

[0038] ;

[0039] Among them, the energy spectrum The energy distribution at different scales is described to identify the main frequency components; Step (1.4.1.3) extreme point detection realizes fault entity extraction;

[0040] Through detection The local extreme points of the signal can be used to determine sudden events or abnormal vibrations in the signal;

[0041] Step (1.4.1.4) calculates the co-occurrence matrix to achieve relationship extraction,

[0042] ;

[0043] Construct the co-occurrence matrix of wavelet coefficients at different scales , analyze the correlation between different frequency components.

[0044] More preferably, the process of step (1.4.2) is as follows:

[0045] Step (1.4.2.1) Text data embedding;

[0046] The text data needs to be vectorized and the corresponding character sequence is mapped into a word vector sequence. The present invention uses the word embedding method Word2vec to continuously train and learn to update the word embedding lookup table, and finally matches the input text through the lookup table to convert it into a corresponding word vector sequence.

[0047] Step (1.4.2.2) The text embedding vector is processed by the Bi-LSTM layer;

[0048] First, the multiple initial parameters used by the Bi-LSTM layer are randomly initialized, and the text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the reverse LSTM layer model to learn the context information of the sequence vector. 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.

[0049] Step (1.4.2.3) The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0050] The CRF layer considers the relationship between adjacent labels to obtain the global 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;

[0051] Step (1.4.2.4) calculates the loss function of the model output to obtain the optimal result;

[0052] The output of the CRF layer is used as the final label sequence of the model output, and compared with the actual labeled sequence to calculate the Loss function.

[0053] ;

[0054] in, It represents the probability of outputting sequence Y when given input sequence X. Represents a given input sequence and tag sequence The score function when , log is the logarithmic operation, exp is the natural exponential operation;

[0055] By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained.

[0056] More preferably, the process of step (1.4.3) is as follows:

[0057] Step (1.4.3.1) structured data format conversion;

[0058] Export the structured data processed in step (1.4.1) from the relational database into csv format. To express the temporal characteristics of process state data, convert the traditional triple form into Expanded to a quadruple with timing information , where t represents the time information when the triple is established. However, the current graph database only supports triple storage, so the present invention converts the timing information and embeds it into the relationship between entities to achieve the establishment of a timing knowledge graph of electromechanical equipment based on the graph database;

[0059] In data structures and databases, a triple refers to a set of three elements, which is often used to represent entity relationships in relational databases;

[0060] Step (1.4.3.2) unstructured data format conversion;

[0061] Convert the entities and entity relationships extracted from the unstructured data in step (1.4.2) into a csv file format. Since the extracted data are all text, the present invention converts the csv file encoding format into UTF-8 encoding;

[0062] Step (1.4.3.3) Data import: Import the csv file data into the graph database.

[0063] Furthermore, in step (2.1), a dynamic graph convolution model is used to obtain the tail entity embedding vector of the knowledge graph at any time t , the process is:

[0064] Step (2.1.1) The tail entity is embedded based on the traditional graph convolution model.

[0065] ;

[0066] in, is the adjacency matrix, is the tail entity at time t In the hidden state of layer k+1, is the entity in the subgraph at time t that is related to the tail entity Head entity connected by relationship r The collection of ; is the normalization constant, Represents the head entity at level k The hidden state of is the diagonal relationship weight matrix when it is k layers, Represents the hidden state embedded in the initial state, and the activation function uses ReLU, is a trainable parameter matrix;

[0067] Step (2.1.2) improves the traditional graph convolution model based on the Neural Ordinary Differential Equation;

[0068] In step (2.2), the dynamic graph convolution model in step (2.1) is used to obtain The embedding vector of the head entity of the knowledge graph .

[0069] Preferably, the process of step (2.1.2) is:

[0070] Step (2.1.2.1) decomposes the propagation depth k of the traditional graph convolution into the integration time and step length The combination of , the Neural ODE representation of graph convolution is realized by adding a variable integration step size dt,

[0071] ;

[0072] in, is the hidden state embedded in the initial state of the tail entity, is the diagonal relationship weight matrix;

[0073] Step (2.1.2.2) uses Taylor expansion to convert the ordinary differential equation into a differential equation form.

[0074] ;

[0075] Among them, 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;

[0076] Step (2.1.2.3) is done by command , 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 up,

[0077] ;

[0078] in, It is a black-box ODE solver used to solve neural ordinary differential equations.

[0079] The process of step (3.1) is:

[0080] Step (3.1.1) trains an ODE module to learn the changes between positions, so that the position encoding is no longer a set of preset static vectors, but a dynamic vector determined by the neural network of ordinary differential equations. The formula of the trained ODETransition Module is as follows:

[0081] ;

[0082] in, is the position code, is the number of the position encoding vector, according to the integration time Divide the position code into piece;

[0083] In step (3.1.2), for the process of processing temporal knowledge graphs, the absolute moments between entities and their relative intervals are very important, so by integrating the original absolute position encoding method with ODE, the temporal position encoding function is:

[0084] ;

[0085] in, and Represent two adjacent points of continuous time position coding, is the frequency parameter associated with each dimension j, is the linear projection of the relation embedding r, s is the sine function, and c is the cosine function;

[0086] Step (3.1.3) is solved using a black box ODE solver to calculate a continuous time encoding vector for any time point t. This new time encoding method needs to be optimized through training data so that the model can adaptively learn dynamic time encoding.

[0087] ;

[0088] in, It is a dynamic position encoding.

[0089] The process of step (3.2) is:

[0090] Step (3.2.1) uses a multi-head attention mechanism to enhance the ability to adaptively identify the temporal information of entities and relationships in the knowledge graph of electromechanical equipment operation and maintenance;

[0091] Step (3.2.2) realizes entity prediction and time prediction by constructing intensity function.

[0092] The process of step (3.2.1) is as follows:

[0093] Step (3.2.1.1) encodes the dynamic position Through adaptive linear mapping, it is converted into i length The vector is used to generate the query vector Q, key vector K and value vector V. The generation formula of Q, K and V is as follows:

[0094] ;

[0095] in, , , It is a learnable weight matrix, which increases the expressiveness of the model, so that the calculated weight matrix has higher generalization ability;

[0096] The hidden state generated by a single attention head in step (3.2.1.2) The calculation formula is as follows,

[0097] ;

[0098] in, is the temporal information bias term in the attention mechanism, represents transpose, is the embedding matrix of the query relation, The function is a normalized exponential function commonly used in multi-class classification problems;

[0099] Step (3.2.1.3) outputs each attention head Do the concatenation operation, and then get the hidden state of the query time output through residual connection and hidden layer regularization ,

[0100] ;

[0101] Among them, the hidden layer regularization Is to normalize the hidden layer in the network to a standard normal distribution. Feedforward Neural Network It is the simplest neural network structure and is widely used in various classification and regression tasks.

[0102] The process of step (3.2.2) is as follows:

[0103] The temporal conditional intensity function constructed in step (3.2.2.1) for all candidate entities is,

[0104] ;

[0105] in, is a projection matrix, At query time The hidden state of With parameters The softplus function is used to ensure the positivity of the intensity function.

[0106] Step (3.2.2.2) predicts the tail entity occurring at query time by maximizing the conditional strength function ,

[0107] ;

[0108] Among them, the arg max function means finding The largest , The maintenance entity is the final predicted one; step (3.2.2.3) predicts the time when the maintenance entity occurs through the conditional time density function

[0109] ;

[0110] ;

[0111] in, To query entities, It is the time parameter used for integration operation.

[0112] The process of step (4.1) is:

[0113] Step (4.1.1) Calculate the cross entropy loss ,

[0114] ;

[0115] in, is the number of sample quadruplets in the training set, is an indicator that is 1 when the i-th sample entity category is M (maintenance), otherwise it is 0. is the probability that the model predicts that the i-th sample entity category is M;

[0116] Step (4.1.2) calculates the mean square error loss ,

[0117] ;

[0118] in, is the time when the maintenance entity is actually generated, It is the time when the maintenance entity is predicted to be generated;

[0119] Step (4.1.3) integrates the losses of the two tasks and improves the prediction accuracy by minimizing the loss function.

[0120] ;

[0121] in, To adjust the hyperparameter of the relative importance between the two loss functions, the value is 0.2;

[0122] In step (4.2), the adjoint sensitivity method is used for back propagation in the part using the Neural Ordinary Differential Equation.

[0123] ;

[0124] in, is the loss in the ordinary differential equation process, is the accompanying quantity, To obtain, is the tail entity embedding vector of the knowledge graph at any time t , The adjoint sensitivity method is the transpose of the adjoint quantity. It directly calculates the gradient of the loss function with respect to the network parameters by solving the augmented ordinary differential equation backward in time, thus eliminating the need for backpropagation through the ODE solver, thereby reducing the memory cost in backpropagation and making the model more efficient.

[0125] The beneficial effects of the present invention are mainly manifested in:

[0126] 1. For the massive heterogeneous data of electromechanical equipment, a knowledge graph ontology structure is constructed, and according to the operation and maintenance data in different formats, corresponding methods are used to extract quadruples to construct graphs, so as to realize the construction of electromechanical equipment operation and maintenance time series knowledge graph (ETKG).

[0127] 2. The Neural Ordinary Differential Equations are combined with the relational graph convolutional neural network to construct a dynamic graph convolution embedding model (DGE). The topological structure of the time series knowledge graph is processed continuously in time and embedded into a hidden state to achieve a deep understanding and feature extraction of electromechanical equipment data.

[0128] 3. Construct a dynamic Hawkes Transformer (DHT) model, improve the position encoding based on the neural ordinary differential equation to enable it to handle large-scale and continuous embedding vectors in the temporal knowledge graph, and add a time information bias term to the multi-head attention mechanism to capture the temporal dependency. Finally, build an intensity function based on the continuous time hidden state to realize entity prediction at any time point or time prediction of expected entities. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] Figure 1 It is a flow chart of the predictive maintenance method for electromechanical equipment integrating the time series knowledge graph. DETAILED DESCRIPTION

[0130] The present invention will be further described below in conjunction with the accompanying drawings.

[0131] Reference Figure 1 , a predictive maintenance method for electromechanical equipment integrating a time series knowledge graph, comprising the following steps:

[0132] The first step is to build a knowledge graph of electromechanical equipment timing. The process is as follows:

[0133] Step (1.1) ontology structure construction, such as Figure 1 As shown, the electromechanical equipment timing knowledge graph (ETKG) ontology is defined as follows:

[0134] ;

[0135] Among them, E represents the set of entities; R represents the set of relationships between these entities; T represents the set of timestamps corresponding to entities and relationships, and E includes three major categories: C represents the system and components of the equipment; F represents the failure of the equipment; M represents the maintenance of the equipment;

[0136] Among them, the ontology structure is to define the standard terms and their hierarchical structures related to the operation of electromechanical equipment and identify the implicit associations between these terms. It is a commonly used method in the construction of knowledge graphs.

[0137] Step (1.2) Data collection, the process is:

[0138] Step (1.2.1) collects structured operating data in the electromechanical equipment control cabinet, including vibration signals, fault codes, operating status data, operating data, electrical parameters, temperature, and humidity.

[0139] Use TCP / IP communication protocol to remotely obtain real-time operating data of electromechanical equipment in sensors;

[0140] Among them, all sensors, devices and gateways are connected by serial port, and data communication is carried out through Modbus and RS485 protocols;

[0141] Step (1.2.2) collects unstructured operation and maintenance data in the electromechanical equipment IoT system, including repair records, maintenance records, fault causes and fault phenomena;

[0142] Use crawler technology to obtain unstructured data on fault information in the field of electromechanical equipment from the electromechanical equipment IoT system;

[0143] Among them, Web Crawling, also known as web spider or web robot, is an automated program used to browse and collect Internet data;

[0144] Step (1.3) data preprocessing;

[0145] Use ETL (Extract, Transform, Load) to pre-process the collected data, remove invalid data, convert all raw data into database table structure relationships, and store them uniformly in a relational database. ETL technology is used to extract data from one or more sources and load it into the target database after transformation;

[0146] Among them, ETL technology is a common data integration technology used to extract data from one or more sources and load it into the target database or data warehouse after transformation;

[0147] Step (1.4) entity and relationship extraction, such as Figure 1 As shown in the figure, entities and relationships are automatically extracted from a large amount of structured and unstructured data. The process is as follows:

[0148] Step (1.4.1) for structured vibration data Continuous wavelet transform is used to extract entities and relationships. The process is as follows:

[0149] Step (1.4.1.1) Vibration data After continuous wavelet transform,

[0150] ;

[0151] ;

[0152] in, is a wavelet basis function with two parameters a and b, Represents the coefficients of wavelet transform, which are obtained by using wavelet basis functions Vibration data Perform integral transformation; represents the conjugate complex root of the wavelet basis function, is the scale parameter, is the translation parameter; ∫ is the integral operation, used in mathematics and statistics;

[0153] Step (1.4.1.2) calculates the energy spectrum as follows,

[0154] ;

[0155] Among them, the energy spectrum The energy distribution at different scales is described to identify the main frequency components; Step (1.4.1.3) extreme point detection realizes fault entity extraction;

[0156] Through detection The local extreme points of the signal can be used to determine sudden events or abnormal vibrations in the signal;

[0157] Step (1.4.1.4) calculates the co-occurrence matrix to achieve relationship extraction,

[0158] ;

[0159] Construct the co-occurrence matrix of wavelet coefficients at different scales , analyze the correlation between different frequency components;

[0160] Step (1.4.2) uses Word2vec (Word to Vector), Bi-LSTM (Bidirectional Long Short-Term Memory), and CRF (Conditional Random Field) technologies to extract entities and relationships for unstructured text data. The process is as follows:

[0161] Step (1.4.2.1) Text data embedding:

[0162] The text data needs to be vectorized and the corresponding character sequence is mapped into a word vector sequence. The present invention uses the word embedding method Word2vec to continuously train and learn to update the word embedding lookup table, and finally matches the input text through the lookup table to convert it into a corresponding word vector sequence.

[0163] Among them, Word2vec is a commonly used algorithm for natural language processing. It generates word vectors by training a neural network model, maps words into a continuous vector space, and captures the semantic and grammatical relationships between words.

[0164] Step (1.4.2.2) The text embedding vector is processed by the Bi-LSTM layer;

[0165] First, the multiple initial parameters used by the Bi-LSTM layer are randomly initialized, and the text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the reverse LSTM layer model to learn the context information of the sequence vector. 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.

[0166] Among them, dropout is a commonly used regularization technique that prevents overfitting by randomly discarding neurons. LSTM is a commonly used recurrent neural network. It controls the storage and reading of information by introducing memory units and gating mechanisms, thereby maintaining and updating information over a long time span. Bi-LSTM is an extended version of LSTM, which captures the bidirectional contextual information of sequence data by using two LSTMs in the forward and reverse directions of the time series;

[0167] Step (1.4.2.3) The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0168] The CRF layer considers the relationship between adjacent labels to obtain the global 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;

[0169] Among them, CRF is a probabilistic graphical model for labeling and segmenting sequence data, which can consider contextual information and global features to improve labeling accuracy;

[0170] Step (1.4.2.4) calculates the loss function of the model output to obtain the optimal result;

[0171] The output of the CRF layer is used as the final label sequence of the model output, and compared with the actual labeled sequence to calculate the Loss function.

[0172] ;

[0173] in, It represents the probability of outputting sequence Y when given input sequence X. Represents a given input sequence and tag sequence The score function when , log is the logarithmic operation, exp is the natural exponential operation;

[0174] By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained;

[0175] Step (1.4.3) uses the graph database language to import data and draw the electromechanical equipment timing knowledge graph in the graph database. The process is as follows:

[0176] Step (1.4.3.1) structured data format conversion;

[0177] Export the structured data processed in step (1.4.1) from the relational database into csv format. To express the temporal characteristics of process state data, convert the traditional triple form into Expanded to a quadruple with timing information , where t represents the time information when the triple is established. However, the current graph database only supports triple storage, so the present invention converts the timing information and embeds it into the relationship between entities to achieve the establishment of a timing knowledge graph of electromechanical equipment based on the graph database;

[0178] In data structures and databases, a triple refers to a set of three elements, which is often used to represent entity relationships in relational databases;

[0179] Step (1.4.3.2) unstructured data format conversion;

[0180] Convert the entities and entity relationships extracted from the unstructured data in step (1.4.2) into a csv file format. Since the extracted data are all text, the present invention converts the csv file encoding format into UTF-8 encoding;

[0181] Step (1.4.3.3) Data import: Import the csv file data into the graph database;

[0182] The second step is to use the dynamic graph convolution embedding model to realize the dynamic embedding of the electromechanical equipment timing knowledge graph. The process is as follows:

[0183] Step (2.1) is as follows Figure 1 As shown, a dynamic graph convolution model is used to obtain the tail entity embedding vector of the knowledge graph at any time t. , the process is:

[0184] Step (2.1.1) The tail entity is embedded based on the traditional graph convolution model.

[0185] ;

[0186] in, is the adjacency matrix, is the tail entity at time t In the hidden state of layer k+1, is the entity in the subgraph at time t that is related to the tail entity Head entity connected by relationship r The collection of , is the normalization constant, Represents the head entity at level k The hidden state of is the diagonal relationship weight matrix when it is k layers, Represents the hidden state embedded in the initial state, and the activation function uses ReLU, is a trainable parameter matrix;

[0187] Among them, traditional graph convolution is a commonly used method for knowledge graph embedding;

[0188] Step (2.1.2) improves the traditional graph convolution model based on the Neural Ordinary Differential Equation. The process is as follows: Step (2.1.2.1) decomposes the propagation depth k of the traditional graph convolution into the integral time and step length The combination of , the Neural ODE representation of graph convolution is realized by adding a variable integration step size dt,

[0189] ;

[0190] in, is the hidden state embedded in the initial state of the tail entity, is the diagonal relationship weight matrix.

[0191] Step (2.1.2.2) uses Taylor expansion to convert the ordinary differential equation into a differential equation form.

[0192] ;

[0193] Among them, 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;

[0194] Step (2.1.2.3) The present invention is carried out by making , 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 up,

[0195] ;

[0196] Step (2.2) uses the dynamic graph convolutional model in step (2.1) to obtain The embedding vector of the head entity of the knowledge graph ;

[0197] in, It is a black-box ODE solver used to solve neural ordinary differential equations.

[0198] The third step is to use the dynamic Hawkes Transformer model to achieve the prediction of maintenance entities. The process is as follows: Step (3.1) uses the dynamic position encoding module to encode the entity embedding vector obtained in the second step. The process is:

[0199] Step (3.1.1) is as follows Figure 1 As shown, the present invention trains an ODE module to learn the changes between positions, so that the position encoding is no longer a set of preset static vectors, but a dynamic vector determined by the neural ordinary differential equation network. The ODE Transition Module formula trained by the present invention is as follows:

[0200] ;

[0201] in, is the position code, is the number of the position encoding vector, according to the integration time Divide the position code into piece.

[0202] Step (3.1.2) is for the process of processing time series knowledge graphs. The absolute time between entities and their relative intervals are very important. Therefore, by integrating the original absolute position encoding method with ODE, the final time position encoding function of the present invention is:

[0203] ;

[0204] in, and Represent two adjacent points of continuous time position coding, is the frequency parameter associated with each dimension j. is the linear projection of the relation embedding r, s is the sine function, and c is the cosine function;

[0205] Step (3.1.3) is solved using a black box ODE solver. The present invention calculates a continuous time coding vector for any time point t. This new time coding method needs to be optimized through training data so that the model can adaptively learn dynamic time coding.

[0206] ;

[0207] in, It is a dynamic position encoding;

[0208] Step (3.2) The neural Hawkes decoder module implements maintenance entity prediction, the process is:

[0209] Step (3.2.1) is as follows Figure 1 As shown, the present invention uses a multi-head attention mechanism to enhance the ability to adaptively identify the time information of entities and relationships in the knowledge graph in the field of electromechanical equipment operation and maintenance. The process is as follows:

[0210] Step (3.2.1.1) encodes the dynamic position Through adaptive linear mapping, it is converted into i length The vector is used to generate the query vector Q, key vector K and value vector V. The generation formula of Q, K and V is as follows:

[0211] ;

[0212] in, , , It is a learnable weight matrix, which increases the expressiveness of the model, so that the calculated weight matrix has higher generalization ability;

[0213] The hidden state generated by a single attention head in step (3.2.1.2) The calculation formula is as follows,

[0214] ;

[0215] in, is the temporal information bias term in the attention mechanism, represents transpose, is the embedding matrix of the query relation, The function is a normalized exponential function commonly used in multi-class classification problems;

[0216] Step (3.2.1.3) outputs each attention head Do the concatenation operation, and then get the hidden state of the query time output through residual connection and hidden layer regularization ,

[0217] ;

[0218] Among them, the hidden layer regularization Is to normalize the hidden layer in the network to a standard normal distribution. Feedforward Neural Network It is the simplest neural network structure and is widely used in various classification and regression tasks;

[0219] Step (3.2.2) is as follows Figure 1 As shown, the present invention realizes entity prediction and time prediction by constructing an intensity function, and the process is as follows:

[0220] The temporal conditional intensity function constructed in step (3.2.2.1) for all candidate entities is,

[0221] ;

[0222] in, is a projection matrix, At query time The hidden state of With parameters The softplus function is used to ensure the positivity of the intensity function;

[0223] Step (3.2.2.2) predicts the tail entity occurring at query time by maximizing the conditional strength function ,

[0224] ;

[0225] Among them, the arg max function means finding The largest , The maintenance entity is the final predicted one; step (3.2.2.3) predicts the time when the maintenance entity occurs through the conditional time density function ,

[0226] ;

[0227] ;

[0228] in, To query entities, It is the time parameter used for integration operation;

[0229] Step 4: Treat the entity prediction task as a multi-class classification task and the time prediction task as a regression task for model training. The process is as follows:

[0230] Step (4.1) The present invention uses cross entropy loss and mean square error loss as the loss of entity prediction task and time prediction task, and the process is:

[0231] Step (4.1.1) Calculate the cross entropy loss ,

[0232] ;

[0233] in, is the number of sample quadruplets in the training set, is an indicator that is 1 when the i-th sample entity category is M (maintenance), otherwise it is 0. is the probability that the model predicts that the i-th sample entity category is M;

[0234] Step (4.1.2) calculates the mean square error loss ,

[0235] ;

[0236] in, is the time when the maintenance entity is actually generated, It is the time when the maintenance entity is predicted to be generated;

[0237] Step (4.1.3) integrates the losses of the two tasks and improves the prediction accuracy by minimizing the loss function.

[0238] ;

[0239] in, To adjust the hyperparameter of the relative importance between the two loss functions, the value is 0.2;

[0240] Step (4.2) For the part using the Neural Ordinary Differential Equation, the present invention adopts the adjoint sensitivity method for back propagation.

[0241] ;

[0242] in, is the loss in the ordinary differential equation process, is the accompanying quantity, To obtain, is the tail entity embedding vector of the knowledge graph at any time t , The adjoint sensitivity method is the transpose of the adjoint quantity. It directly calculates the gradient of the loss function with respect to the network parameters by solving the augmented ordinary differential equation backward in time, thus eliminating the need for backpropagation through the ODE solver, thereby reducing the memory cost in backpropagation and making the model more efficient.

[0243] The implementation process of the electromechanical equipment predictive maintenance method integrating the time series knowledge graph in this embodiment is as follows:

[0244] Step 1: Provide 10 knowledge graph reasoning models:

[0245] TransE: This method is an embedding-based static knowledge graph reasoning model that learns vector representations of entities and relations by minimizing the difference between the vectors of the head and tail entities of a triple.

[0246] ComplEx: This method is a static knowledge graph embedding model that uses complex embeddings (real and imaginary parts) to represent entities and relations.

[0247] DistMult: This method is a simple but effective static knowledge graph embedding model that captures the relationships between entities by using symmetric trilinear products.

[0248] TTransE: This method is an extension of the TransE model and is a temporal knowledge graph inference method. By taking the time constraint as part of the triple, it can effectively predict past or future relations.

[0249] TNTComplEx: This method is an extension of the ComplEx model and is a temporal knowledge graph inference method. By introducing the time factor into the embedding of entities and relationships, the model can effectively predict relationships and events in the knowledge graph that change over time.

[0250] TA-DistMult: This method is an extension of the DistMult model and is a temporal knowledge graph inference method.

[0251] It introduces a time factor for each triple so that the embedding can reflect the changes of entities and relations over time.

[0252] RE-NET: This method is a model specially designed for temporal knowledge graph extrapolation, which captures the evolution of entities and relations over time by integrating recurrent neural networks.

[0253] RE-GCN: This method combines the advantages of graph convolutional networks and recurrent neural networks. It can capture the structural information and time evolution information of the knowledge graph at the same time, and is a temporal knowledge graph extrapolation model.

[0254] CENET: This method is a temporal knowledge graph event prediction model based on a historical contrastive learning framework. It can effectively distinguish potential entities matching the query and is a temporal knowledge graph extrapolation model.

[0255] ETKG-GH: The temporal knowledge graph extrapolation method of the present invention integrates the dynamic graph convolution embedding model and the Hawkes Transformer model to dynamically embed the temporal knowledge graph and realize the predictive maintenance of electromechanical equipment.

[0256] Step 2: Experimental Dataset

[0257] The experimental data set is provided by the electromechanical equipment timing knowledge graph constructed by the ontology construction method proposed in this invention, which contains 12792 entities and 154 semantic relations.

[0258] Step 3: Define evaluation metrics

[0259] In order to verify the prediction performance of each model, the present invention uses the following evaluation indicators for experiments:

[0260] Hit@n: This metric measures the proportion of correct predictions among all test samples that are within the top n highest rankings given by the model. Smaller n values ​​emphasize the accuracy of the model in predicting the most relevant options; larger n values ​​take more into account the comprehensiveness and diversity of the model.

[0261] ;

[0262] in, 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.

[0263] MRR: An internationally accepted mechanism for evaluating matching algorithms. The calculation principle is that the higher the correct result is ranked, the higher the score is. If there is no correct result, the score is 0. The formula is as follows:

[0264] ;

[0265] Step 4: Analyze and compare the results

[0266] Refer to Table 1, which shows the comparison results of four indicators: MRR, Hit@1, Hit@3, and Hit@10 for the maintenance entity prediction task.

[0267]

[0268] Table 1

[0269] The temporal extrapolation model has a significant improvement over the internalization model and the static model. The effect of the extrapolation model RE-NET is significantly increased because the number of repair and maintenance entities for a certain fault entity is much smaller than the number of fault entities for a certain component entity, which leads to a lower demand for historical data, and RE-NET uses a recursive neural encoder to capture time and structure information, which has achieved great success in improving model performance, making it the best baseline model in maintenance entity prediction. However, he only considers the time sequence between events, while the model of the present invention encodes time information into hidden states through ordinary differential equations to capture time dependencies. This enables it to achieve the highest MRR, Hit@1 and Hit@3 in maintenance entity prediction, and ranks third in Hit@10. In particular, in Hit@1, the average gap between ETKGR-GH and the comparison model is the largest. In contrast, the RE-NET and CENET models perform well in Hits@10, but fail to accurately predict the correct entity, making Hits@1 much lower than ETKGR-GH. The method proposed in the present invention has achieved good results in all four indicators, which are 36.3%, 27.84%, 38.86% and 46.65% respectively. Taking all factors into consideration, ETKGR-GH performs best in the cases of MRR, Hit@1 and Hit@3, which means that the model is most likely to include the optimal entity in the candidate set. The method of the present invention fails to have the best performance in Hit@10 because important information will be given greater prominence and weight, resulting in Hit@1 and Hit@3 being higher, and some minor information will be weakened, reducing Hit@10. However, there is no need to pay attention to the top ten prediction results when formulating the maintenance plan, which makes Hit@10 less important than Hit@1 and Hit@3.

[0270] The electromechanical equipment predictive maintenance method integrating the time series knowledge graph in this embodiment is described as follows:

[0271] Step 1: Provide 9 predictive maintenance methods:

[0272] ETKG-GH: Predictive maintenance method of the present invention.

[0273] ETKGR-GH wo DGE: The predictive maintenance method of the present invention removes the dynamic graph convolution embedding module.

[0274] ETKGR-GH w. GCN: The predictive maintenance method of the present invention replaces the dynamic graph convolution embedding module with a graph convolutional network (GCN) module.

[0275] ETKGR-GH w. RGCN: The predictive maintenance method of the present invention replaces the dynamic graph convolution embedding module with a relational graph convolutional network (RGCN) module.

[0276] ETKGR-GH w. Dismult: The predictive maintenance method of the present invention replaces the Hawkes Transformer module with the Dismult module. Dismult is a deep learning framework for solving distributed multi-task learning problems. It can process multiple related tasks simultaneously and improve the learning efficiency and performance of each task by sharing hidden layers.

[0277] ETKGR-GH w. TuckER: The predictive maintenance method of the present invention replaces the Hawkes Transformer module with the TuckER module. TuckER is a knowledge graph embedding model based on tensor decomposition, which uses three-dimensional weight tensors and entity and relationship embedding to predict triple relationships in the knowledge graph.

[0278] ETKGR-GH wo DPE: The predictive maintenance method of the present invention removes the dynamic time sequence position encoding module.

[0279] ETKGR-GH w. APE: The predictive maintenance method of the present invention replaces the dynamic temporal position encoding module with an absolute position encoding module (Absolute Positional Encoding, APE).

[0280] ETKGR-GH w. RPE: The predictive maintenance method of the present invention replaces the dynamic temporal position encoding module with a relative position encoding module (Relative Positional Encoding, RPE).

[0281] Step 2: Experimental Dataset

[0282] The experimental dataset is the same as the implementation scheme.

[0283] Step 3: Define evaluation metrics

[0284] In order to verify the prediction performance of each model, the present invention adds an evaluation index to the evaluation index of the implementation scheme to conduct experiments:

[0285] MAE: This metric measures the mean absolute error between the predicted value and the actual observed value.

[0286] ;

[0287] Where n is the number of predicted values, is the ith observation value, is the i-th predicted value, is the absolute error between the actual observed value and the predicted value.

[0288] Step 4: Analyze and compare the results

[0289] In order to demonstrate the superiority of the ETKGR-GH model in temporal knowledge graph embedding, the present invention compares ETKGR-GH with several variants using different graph structure embedding modules. In order to facilitate the comparison of temporal reasoning capabilities, the present invention normalizes MAE to the range of 0 to 1. Table 2 shows the comparison results of four indicators of predictive maintenance performance of nine models;

[0290]

[0291] Table 2

[0292] The results in Table 2 show that TKG without graph structure embedding learning will lead to a significant decrease in the model's reasoning ability. In addition, DGE has better encoding capabilities than RGCN and GCN, and is more accurate in learning graph structures, because RGCN's convolution can only convolve 3 to 4 layers, while DGE introduces neural ordinary differential equations to achieve multi-layer convolution. By comparing the results of entity and time prediction tasks with RGCN, it is shown that DGE is more suitable for TKG extrapolation reasoning than RGCN. DGE can extract time evolution features from complex structural information to achieve dynamic embedding of TKG.

[0293] In order to reflect the applicability of the dynamic Hawkes Transformer, the present invention compares the decoder part with Distmult and TuckER. As can be seen from Table 2, because time prediction is not involved, the MAE indicator is empty. However, ETKGR-GH has a clear lead in the remaining indicators, because the Hawkes process processes a temporally continuous sequence, and the dynamic Hawkes Transformer of the present invention is implemented precisely for this purpose. In addition, the Hawkes process believes that historical information will have a positive impact on the occurrence of the predicted target, which is a common phenomenon in the field of electromechanical equipment, so the dynamic Hawkes Transformer, as a decoder, is very suitable for predictive maintenance of electromechanical equipment.

[0294] After demonstrating the effect of the dynamic Hawkes Transformer, in order to verify whether the dynamic temporal position encoding proposed by the present invention is effective, the present invention compares ETKGR-GH with several variants using different position encoding methods. In Table 2, it is observed that ETKGR-GH outperforms all other variants in both entity and time prediction tasks. Note that the performance of ETKGR-GH w. APE and ETKGR-GH w. RPE is almost the same as that without using the position encoding function. This indicates that they do not learn position information. In contrast, the continuous time encoding of the present invention can effectively capture temporal information and help the model achieve good performance.

[0295] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A predictive maintenance method for electromechanical equipment integrating a time series knowledge graph, characterized in that: The method comprises the following steps: The first step is to build a knowledge graph of electromechanical equipment timing. The process is as follows: Step (1.1) ontology structure construction, defining the electromechanical equipment timing knowledge graph ontology; Step (1.2) data collection; Step (1.3) data preprocessing; Step (1.4) entity and relationship extraction; The second step is to use the dynamic graph convolution embedding model to realize the dynamic embedding of the electromechanical equipment timing knowledge graph. The process is as follows: Step (2.1) uses a dynamic graph convolutional model to obtain the tail entity embedding vector of the knowledge graph at any time; 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; The third step is to use the dynamic Hawkes Transformer model to achieve the prediction of maintenance entities. The process is as follows: Step (3.1) uses the dynamic position encoding module to encode the entity embedding vector obtained in the second step; Step (3.2) The neural Hawkes decoder module implements maintenance entity prediction and time prediction; Step 4: Treat the entity prediction task as a multi-class classification task and the time prediction task as a regression task for model training. The process is as follows: Step (4.1) uses cross entropy loss and mean square error loss as the losses for entity prediction tasks and time prediction tasks; Step (4.2) uses the adjoint sensitivity method for back propagation for the part using the Neural Ordinary Differential Equation; The process of step (3.2) is: Step (3.2.1) uses a multi-head attention mechanism to enhance the ability to adaptively identify the temporal information of entities and relationships in the knowledge graph of electromechanical equipment operation and maintenance; Step (3.2.2) realizes entity prediction and time prediction by constructing intensity function; The process of step (3.2.1) is as follows: Step (3.2.1.1) encodes the dynamic position P = {p0, p1, ..., p i }Through adaptive linear mapping, it is converted into i pieces of length d K The vector is used to generate the query vector Q, key vector K and value vector V. The generation formula of Q, K and V is as follows: Among them, W Q , W K , W V It is a learnable weight matrix, which increases the expressiveness of the model, so that the calculated weight matrix has higher generalization ability; The hidden state h generated by a single attention head in step (3.2.1.2) i The calculation formula is as follows, Among them, (TP q )(TP k ) is the temporal information bias term in the attention mechanism, ′ represents transposition, is the embedding matrix of the query relation. The Softmax function is a normalized exponential function commonly used in multi-class classification problems. Step (3.2.1.3) converts the h output by each attention head into i Do the concatenation operation, and then get the hidden state of the query time output through residual connection and hidden layer regularization Among them, the hidden layer regularization Add&Norm(·) normalizes the hidden layer in the network to the standard normal distribution; The process of step (3.2.2) is as follows: The temporal conditional intensity function constructed in step (3.2.2.1) for all candidate entities is, Among them, W λ is a projection matrix, is the query time t q The hidden state of g(x) = β· is a softplus function with parameter β, the purpose of which is to ensure the positivity of the intensity function; Step (3.2.2.2) predicts the tail entity e occurring at query time by maximizing the conditional strength function o , Among them, the arg max function means finding The largest e o , e o (t q ) is the maintenance entity of the final prediction; Step (3.2.2.3) predicts the time t of maintenance entity occurrence through conditional time density function * , Among them, e q is the query entity, and τ is the time parameter used for integration operation.

2. The predictive maintenance method for electromechanical equipment integrating a time series knowledge graph as claimed in claim 1 is characterized in that: In the step (1.1), the electromechanical equipment timing knowledge graph ETKG ontology is defined as follows: ETKG=<E,R,T> =<C,F,M,R,T> ; Among them, E represents the set of entities; R represents the set of relationships between these entities; T represents the set of timestamps corresponding to entities and relationships, and E includes three major categories: C represents the system and components of the equipment; F represents the failure of the equipment; M represents the maintenance of the equipment; The process of step (1.2) is: Step (1.2.1) collects structured operating data in the electromechanical equipment control cabinet, including vibration signals, fault codes, operating status data, operating data, electrical parameters, temperature and humidity, and uses the TCP / IP communication protocol to remotely obtain the real-time operating data of the electromechanical equipment in the sensor; wherein all sensors are connected to the equipment and gateway via serial ports, and data communication is carried out via Modbus and RS485 protocols; Step (1.2.2) collects unstructured operation and maintenance data in the electromechanical equipment Internet of Things system, including repair records, maintenance records, fault causes and fault phenomena; and uses crawler technology to obtain unstructured data on fault information in the electromechanical equipment field from the electromechanical equipment Internet of Things system.

3. The predictive maintenance method for electromechanical equipment integrating a time series knowledge graph as claimed in claim 2 is characterized in that: The process of step (1.4) is: Step (1.4.1) uses continuous wavelet transform to extract entities and relationships from structured vibration data F(t); Step (1.4.2) uses Word2vec, Bi-LSTM, and CRF technologies to extract entities and relationships for unstructured text data; Step (1.4.3) uses the graph database language to import data and draw the electromechanical equipment timing knowledge graph in the graph database.

4. The predictive maintenance method for electromechanical equipment integrating a time series knowledge graph as claimed in claim 3 is characterized in that: The process of step (1.4.1) is as follows: Step (1.4.1.1) The vibration data F(t) is transformed by continuous wavelet transform. Among them, ψ a,b is a wavelet basis function with two parameters a and b. W(a,b) represents the coefficient of wavelet transform, which is obtained by using the wavelet basis function ψ a,b Performing integral transformation on the vibration data F(t); It represents the conjugate complex root of the wavelet basis function, a is the scale parameter, b is the translation parameter; ∫ is the integral operation, which is used in mathematics and statistics; Step (1.4.1.2) calculates the energy spectrum as follows, Among them, the energy spectrum E(a) 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 fault entity extraction; By detecting the local extreme points of W(a,b), sudden events or abnormal vibrations in the signal can be determined; Step (1.4.1.4) calculates the co-occurrence matrix to achieve relationship extraction, Construct the co-occurrence matrix C(a1,a2) of wavelet coefficients at different scales and analyze the correlation between different frequency components.

5. The predictive maintenance method for electromechanical equipment integrating a time series knowledge graph as claimed in claim 3 is characterized in that: The process of step (1.4.2) is as follows: Step (1.4.2.1) Text data embedding: Text data needs to be vectorized, and the corresponding character sequence is mapped to a word vector sequence. Through the word embedding method Word2vec, the word embedding lookup table is updated through continuous training and learning, and finally the input text is matched through the lookup table and converted into the corresponding word vector sequence; Step (1.4.2.2) The text embedding vector is processed by the Bi-LSTM layer; First, the multiple initial parameters used by the Bi-LSTM layer are randomly initialized, and the text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the reverse LSTM layer model to learn the context information of the sequence vector. 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 by the CRF layer; The CRF layer considers the relationship between adjacent labels to obtain the global 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; Step (1.4.2.4) calculates the loss function of the model output to obtain the optimal result; The output of the CRF layer is used as the final label sequence of the model output, and compared with the actual labeled sequence to calculate the Loss function. Loss=-logP(Y|X)=-(s(X,Y)-log∑ Y′ exp(s(X,Y′))); Where P(Y|X) represents the probability of output sequence Y given input sequence X, s(X,Y′) represents the score function given input sequence X and label sequence Y′, log represents logarithmic operation, and exp represents natural exponential operation; By continuously adjusting the parameters of each layer to minimize the Loss value, the optimal recognition model is obtained.

6. The electromechanical equipment predictive maintenance method integrating the time series knowledge graph as claimed in claim 3 is characterized in that: 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 into csv format. To express the temporal characteristics of process state data, convert the traditional triple form (e h ,r,e o ) is expanded into a quadruple with timing information (e h ,r,e o ,t), where t represents the time information when the triple is formed. The graph database only supports triple storage. The timing information is converted and embedded into the relationship between entities to realize the establishment of the timing knowledge graph of electromechanical equipment based on the graph database; In data structures and databases, a triple refers to a set of three elements, which is often used to represent entity relationships in relational databases; Step (1.4.3.2) unstructured data format conversion; Convert the entities and entity relationships extracted from the unstructured data in step (1.4.2) into csv file format. Since the extracted data are all text, convert the csv file encoding format to UTF-8 encoding; Step (1.4.3.3) Data import: Import the csv file data into the graph database.

7. The method for predictive maintenance of electromechanical equipment integrating a time series knowledge graph according to any one of claims 1 to 6, characterized in that: In step (2.1), a dynamic graph convolution model is used to obtain the tail entity embedding vector of the knowledge graph at any time t The process is: Step (2.1.1) The tail entity is embedded based on the traditional graph convolution model. Where A is the adjacency matrix, is the tail entity e at time t o In the hidden state of layer k+1, is the entity e in the subgraph at time t o Head entity e connected by relationship r h The collection of is the normalization constant, Represents the head entity e at layer k h The hidden state of is the diagonal relationship weight matrix when it is k layers, H0 represents the hidden state embedded in the initial state, and the activation function uses ReLU, is a trainable parameter matrix; Step (2.1.2) improves the traditional graph convolution model based on the Neural Ordinary Differential Equation. The process is as follows: Step (2.1.2.1) decomposes the propagation depth k of traditional graph convolution into the integration time T * and step length Δt, that is By adding a variable integration step size dt, the Neural ODE representation of graph convolution is realized. in, is the hidden state embedded in the initial state of the tail entity, W r is the diagonal relationship weight matrix; Step (2.1.2.2) uses Taylor expansion to convert the ordinary differential equation into a differential equation form. Among them, Taylor expansion approximates a function with a polynomial near a certain point, which is used to approximate the solution of a complex function in a differential equation, and I is the unit matrix; Step (2.1.2.3) is to let Δt→0, use a set of parameters Θ to characterize the entire continuous time aggregation process, and obtain the complete tail entity embedding vector by summing, Among them, ODESolve is a black-box ODE solver used to solve neural ordinary differential equations; In step (2.2), the dynamic graph convolution model in step (2.1) is used to obtain the query time t q The embedding vector of the head entity of the knowledge graph 8. The method for predictive maintenance of electromechanical equipment integrating a time series knowledge graph according to any one of claims 1 to 6, characterized in that: The process of step (3.1) is: Step (3.1.1) trains an ODE module to learn the changes between positions, so that the position encoding is no longer a set of preset static vectors, but a dynamic vector determined by the neural network of ordinary differential equations. The formula of the trained ODETransition Module is as follows: Among them, p is the position code, i is the number of the position code vector, according to the integration time T * Divide the position code into i equal blocks; In step (3.1.2), for the process of processing time series knowledge graphs, the absolute time between entities and their relative intervals are very important, so by integrating the original absolute position encoding method with ODE, the time position encoding function is: Among them, i and i+1 represent two adjacent points of continuous time position coding, w j is the frequency parameter associated with each dimension j, α rj is the linear projection of the relation embedding r, s is the sine function, and c is the cosine function; Step (3.1.3) is solved using a black box ODE solver to calculate a continuous time encoding vector for any time point t. This new time encoding method needs to be optimized through training data so that the model can adaptively learn dynamic time encoding. Among them, {p0,p1,…,p i } is a dynamic position code.

9. The method for predictive maintenance of electromechanical equipment integrating a time series knowledge graph according to any one of claims 1 to 6, characterized in that: The process of step (4.1) is: Step (4.1.1) calculates the cross entropy loss L G , Among them, |G train | is the number of sample quadruplets in the training set, I M is an indicator that is 1 when the i-th sample entity category is maintained M, otherwise it is 0, p M is the probability that the model predicts that the i-th sample entity category is M; Step (4.1.2) calculates the mean square error loss L τ , Among them, t i is the time when the maintenance entity is actually generated, It is the time when the maintenance entity is predicted to be generated; Step (4.1.3) integrates the losses of the two tasks and improves the prediction accuracy by minimizing the loss function. Loss2=L G +μL τ ; Among them, μ is a hyperparameter that adjusts the relative importance between the two loss functions; In step (4.2), the adjoint sensitivity method is used for back propagation for the part using the Neural Ordinary Differential Equation. Among them, L ODE is the loss in the process of the God ordinary differential equation, α(t) is the adjoint quantity, and is given by Calculate, · is the tail entity embedding vector of the knowledge graph at any time t α(t)′ is the transpose of the adjoint.

Images