An equipment health state evaluation method based on knowledge graph multi-set pooling

Abstract

The application discloses a kind of equipment health state evaluation methods based on knowledge graph multi-set pooling, belongs to equipment health state evaluation technical field.It includes the following steps:1, equipment health state data model construction;2, knowledge graph multi-set pooling network model construction;3, knowledge graph node dimension feature extraction;4, knowledge graph multi-head attention multi-set pooling;5, training and classification.The present application is aimed at the problem that existing health state evaluation method cannot effectively fuse time feature and space feature, proposes a kind of equipment health evaluation method based on knowledge graph multi-set pooling, utilizes time sequence knowledge graph to carry out deep fusion for equipment component and index information into structured graph data model, and extracts time feature and space feature in it by knowledge graph multi-level pooling, improves the accuracy of health evaluation, provides technical support for realizing equipment predictive maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment health status assessment, and specifically relates to an equipment health status assessment method based on multi-set pooling of knowledge graphs. Background Art

[0002] With the rapid development and widespread application of industrial Internet, big data and artificial intelligence technologies, intelligent health status assessment has become a key link in intelligent manufacturing and a potential breakthrough point in the future. By analyzing a large amount of status monitoring data, the current and future working status of equipment can be evaluated, the health status can be accurately grasped in a timely and accurate manner, and early warnings can be issued, so that predictive maintenance measures can be taken to ensure the safety and reliability of equipment operation and effectively reduce maintenance costs.

[0003] Existing equipment health assessment methods are primarily categorized as model-driven, knowledge-driven, and data-driven. Model-driven methods, based on long-term research by experts into the principles of equipment degradation, employ physical models such as the Mahalanobis distance method, fusion weight calculation method, Euclidean distance method, and fuzzy theory method. Song Dengwei (System Simulation Technology, 2017) et al. used a generalized regression neural network to construct an observer for a hydraulic system health assessment based on a health baseline and Mahalanobis distance, obtaining residuals. The Mahalanobis distance from the health baseline was then normalized to the health score. Bai et al. (Microelectronics & Reliability, 2017) constructed a virtual health vector and used the gray correlation method to calculate the correlation between the acquired vector and the virtual vector to assess the health of wind turbine blades. These methods rely on expert experience to determine weights and model parameters, and their idealized modeling assumptions make them difficult to adapt to the complex factors affecting equipment operation in practice.

[0004] Knowledge-driven health assessment uses knowledge acquisition and knowledge expression to evaluate the health status. Peng Yanliang (Computer System Applications, 2017) et al. proposed a health assessment model based on dynamic weights and fuzzy comprehensive evaluation methods for switchgear, but its weights were manually defined and inaccurate. Qian Hong (Journal of Shanghai University of Electric Power, 2019) et al., based on the variable weight fuzzy evaluation method, used entropy weight method, equilibrium function variable weight, etc. to form a comprehensive weighted method, which can more accurately reflect the health of the equipment's operating status. Liu Weiqiang (Computer Integrated Manufacturing Systems, 2021) et al. proposed a health status assessment method based on brittleness and cloud model to address the randomness and fuzziness problems in the evaluation process of the engine cylinder head assembly system, and verified the correctness and effectiveness of the method. This type of method is difficult to apply, and there is little research on health assessment. This is mainly because knowledge and experience are limited, and knowledge expression also faces the problem of knowledge standardization.

[0005] Data-driven is currently the most promising method, which fully utilizes the advantages of machine learning and artificial intelligence. It is also a method that is widely studied in health assessment at home and abroad. Lei Yaguo (Journal of Mechanical Engineering, 2019) proposed a deep migration diagnosis method for mechanical equipment faults, constructed a domain-shared deep residual network, and extracted migration fault features from monitoring data of different mechanical equipment; Pan (KSCE Journal of Civil Engineering, 2018) et al. used deep Bayesian belief network learning to extract structural information and probabilistically determine structural conditions, solving the problems of instability and uncertainty in the health assessment of structural parts; Sun (Mathematical Problems in Engineering, 2018) et al. introduced a hybrid method that combines variance analysis of sequential SCADA variables with RNN, and used long-term and short-term memory networks and component health weighting methods to evaluate the health status of wind turbines; Xu (IJCNN, 2018) et al. proposed a tool wear prediction model based on GRU, which can well handle time series data and avoid the gradient vanishing problem; Huang (IEEE Transactions on Industrial Electronics, 2019) et al. proposed a new prediction method based on bidirectional long-term and short-term memory networks, which integrates multiple sensor data with operating condition data and learns high-level features to generate predicted target outputs; Peng (IEEE Transactions on Industrial Electronics, 2019) and others proposed a Bayesian deep learning-based method that uses variational inference based on a Bayesian neural network. The resulting predictions have improved scalability and generalization capabilities. This type of method primarily focuses on training large amounts of data using machine learning algorithms to automatically understand patterns within the data.

[0006] Existing equipment health assessment methods have not yet achieved the fusion of data from different sources, nor have they integrated the spatial and temporal features in the data, and therefore have significant limitations. Summary of the Invention

[0007] In response to the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide an equipment health assessment method based on multi-set pooling of knowledge graphs, deeply integrating component data and indicator data through a time series knowledge graph, and using a knowledge graph multi-set pooling network model to extract the spatiotemporal features of the time series knowledge graph, and then perform health status assessment based on the spatiotemporal features.

[0008] The present invention provides the following technical solutions:

[0009] A method for equipment health status assessment based on multi-set pooling of knowledge graphs is proposed. This method deeply integrates equipment components and indicator information through a time-series knowledge graph, and uses a graph convolutional neural network and an attention mechanism to perform multi-set pooling representation learning on the knowledge graph to obtain equipment health status classification results. The method includes the following specific steps:

[0010] S1. Engine health status data model construction: Extract the time series data of component entities and corresponding indicators from the engine data, construct a time series knowledge graph based on the relationship between components and indicators, and establish the corresponding adjacency matrix;

[0011] S2. Construct the engine knowledge graph multi-set pooling network model: The knowledge graph feature vector h obtained by the knowledge graph neural network GIN is pooled. G Pooling with the adjacency matrix obtained in step S1 to obtain the engine health status representation learning result;

[0012] S3. Extract node dimension features of the knowledge graph: Use the knowledge graph neural network GIN to perform representation learning on the knowledge graph extracted in step S1 to obtain the full graph feature h of each node feature of the engine knowledge graph G ;

[0013] S4. Graph multi-set pooling based on knowledge graph multi-head attention includes the following steps:

[0014] S4.1. Use the knowledge graph multi-head attention mechanism to learn the clustering scheme of approximate nodes and generate the cluster assignment matrix;

[0015] S4.2. Perform multi-set pooling on the knowledge graph and define the knowledge graph multi-set pooling GMPool. Compress n nodes into k typical representative nodes in an end-to-end pooling manner and input the full-graph feature H obtained by the graph convolution module into GMPool.

[0016] S4.3. Further use the node self-attention mechanism to consider the interactions between n input nodes or k typical representative nodes;

[0017] S5. Graph Multi-set Pooling Network Training and Classification: The graph multi-set pooling network is trained cyclically to continuously reduce the loss, so that the embedding of the engine health status converges and the final engine health status classification result is obtained.

[0018] Furthermore, the formal definition of the temporal knowledge graph is as follows:

[0019] The temporal knowledge graph is a directed label graph G t =(t0,t n,E,R,τ), where E is the vertex set of the knowledge graph, which is used to represent the entity set; R is the edge set of the knowledge graph, which is used to represent the fact relationship set; τ is the function of E×E→R|k, which represents all tuples in the knowledge graph; k represents the time period [t0,t n ] In the knowledge graph triple list sorted by time, there is a k-th relationship R between two entities; there is a pair of nodes with a direct link relationship in the temporal knowledge graph, and the distance between the nodes is 1; let the number of entities be N, and construct the direct adjacency matrix of the relationship graph:

[0020]

[0021] Furthermore, in step S2, the graph multi-head attention pooling network includes an input layer and two hidden layers. The input layer is the initial feature matrix and direct adjacency matrix of the equipment time series knowledge graph. The two hidden layers simultaneously learn feature information of the node dimension and the graph dimension through convolution and self-attention pooling operations. The output of the two hidden layers after training is used as the final vector representation of the health status of the component entity. The specific process is as follows:

[0022] 2.1. First use the graph neural network GNN to learn the feature vector of the node:

[0023] Encoder(X,A)=GNN2(GNN1(X,A),A) (1)

[0024] Among them, X is the input feature matrix, A is the direct adjacency matrix, GNN1 is the first convolution layer, and GNN2 is the second convolution layer;

[0025] 2.2. After obtaining the feature vector of the node, use the graph multiset pooling method to generate the embedding of the entire graph:

[0026] Pooling(H,A)=GMPool1(SelfAtt(GMPool k (H,A)),A′) (2)

[0027] Among them, GMPool is the graph multi-head attention convolution layer, SelfAtt represents the node self-attention mechanism, H is the node feature vector obtained by the GNN layer, A is the direct adjacency matrix, and A′ is the GMPool-derived node feature vector. k After compression, the adjacency matrix is ​​coarsened to obtain the final vector representation of the component entity health status.

[0028] Furthermore, the specific process of step S3 is as follows:

[0029] S3.1. Input the direct adjacency matrix A and feature matrix X of the time series knowledge graph into the convolution module to aggregate the feature vectors and obtain the feature representation of the component and indicator node levels in the time series knowledge graph:

[0030]

[0031] Among them, UPDATE and AGGREGATE are differentiable functions of update and aggregation, N(u) represents a set of adjacent nodes of u, Initialized to the input node feature X u , H (l+1) ∈R n×d It is the node feature obtained after a GNN, which is simplified to:

[0032] H (l+1) =GNN(H (l) ,A (l) ) (4)

[0033] S3.2, after obtaining the node feature representation H (l+1) After that, a readout function generates an embedding vector to obtain the complete full-image representation:

[0034] h G =CONTACT(sum({H v |v∈V})|k=0,1,....K) (5)

[0035] Among them, h G is the full graph feature, CONTACT is the vector concatenation function, and sum is the summation function. Together they constitute the readout function, which sums the nodes of each iteration to obtain the graph features and then concatenates them to obtain the full graph features including the node features.

[0036] Furthermore, the specific process of step S4 is as follows:

[0037] S4.1. In order to enable the pooling method to distinguish the importance of each node in the graph, a graph multi-head attention mechanism is used to extract the clustering scheme;

[0038] S 4.1.1, Assume that there are n node vectors in the graph, and the input of the attention function Att is: query matrix key matrix and value matrix where n q is the number of query vectors, n is the number of input nodes, d k is the dimension of the key matrix, d v is the dimension of the value matrix vector, and the attention function formula is as follows:

[0039] Att(Q,K,A)=ω(QK T )V (6)

[0040] By decomposing the attention function into the dot product of the query and all keys, we can learn the corresponding weighted values ​​and put more weight on the key target nodes:

[0041]

[0042] Among them, ω is the activation function softmax, exp is the exponential of the natural constant e;

[0043] S 4.1.2. Further use multi-head attention instead of single attention, and generate h different projection spaces by linearly projecting the query matrix Q, key matrix K and value matrix V respectively:

[0044] MH(Q,K,V)=[O1,...O h ]W O ;O i =Att(QW i Q ,KW i K ,VW i V ) (8)

[0046] The attention mechanism maps the learned weights to the three weight matrices Q, K, and V by inputting the feature vector of the node, so as to ensure that when encoding any part, it pays attention to the important part as much as possible; W is the adjustment matrix to generate multiple Q, K, V values, which further ensures the effectiveness of encoding learning. h represents the number of projection spaces O, which is composed of the parameter matrix composition, is the output projection matrix, d model is the output dimension of multi-head attention;

[0047] S 4.1.3. In order to generate K and V more accurately in multi-head attention, the graph multi-head attention function is used to learn K and V through GNN. The output of the graph multi-head attention function GMH can be expressed as:

[0048]

[0049] Compared with the linearly embedded KW and VW in the multi-head formula, GNN i The output of contains the adjacency information of the graph. The multi-head attention of the graph can more easily learn the clustering scheme of the approximate nodes and generate the clustering assignment matrix.

[0050] S 4.2, knowledge graph multi-set pooling, the specific process is as follows:

[0051] Define the knowledge graph multi-set pooling GMPool, compress n nodes into k typical representative nodes in an end-to-end pooling manner, and input the full graph feature H obtained by the graph convolution module into GMPool. The formula is as follows:

[0052] GMPool k (H,A)=LN(Z+rFF(Z)); Z=LN(S+GMH(S,H,A)) (10)

[0054] Where H∈R n×d is the node feature derived by GNN, A is the adjacency matrix, rFF is the feedforward layer in any row direction, processing each row feature independently and identically, LN is the layer normalization layer, S∈R k×d It is a parameterized clustering seed matrix used to optimize the pooling operation; the GMH function considers the interaction between the k seed vectors query in S and the n node values ​​in H, and uses the attention similarity between them to compress the n nodes into k typical representative nodes;

[0055] S 4.3. Node self-attention mechanism:

[0056] We further use the node self-attention mechanism to consider the interactions between n input nodes or k typical representative nodes. The formula is as follows:

[0057] SelfAtt(H)=LN(Z+rFF(Z)); Z=LN(S+MH(H,H,H)) (11)

[0058] Compared to GMH in Equation (7) which considers the interactions between k vectors and n nodes, SelfAtt places the node embedding H in the query and key positions of MH in Equation (8) to capture the intrinsic relationships between n nodes.

[0059] Furthermore, the specific process of step S5 is as follows:

[0060] The representation vector H of the final equipment health status obtained above Final Input the final linear classifier to get the final probability in The definition is as follows:

[0061]

[0062] Among them, the sigmoid() function maps the vector to the output probability in the range of [0,1];

[0063] Then according to the above calculated probability Through the loss function and the correct label in the sample Perform loss calculation and minimize the loss function through Adam optimizer. The loss function is defined as:

[0064]

[0065] in, is the number of samples in the sample, is the correct label in the sample;

[0066] The above training process is continuously iterated to converge the embedded representation of the equipment health status and obtain the final equipment health status assessment model; the equipment knowledge graph to be evaluated is input into the above training model to obtain the equipment health status classification result and complete the equipment health status assessment.

[0067] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] In response to the problem that existing health status assessment methods cannot effectively integrate temporal and spatial features, this invention proposes an equipment health assessment method based on multi-set pooling of knowledge graphs. It uses a temporal knowledge graph to deeply integrate equipment components and indicator information into a structured graph data model, and extracts temporal and spatial features through multi-level pooling of the knowledge graph, thereby improving the accuracy of health assessment and providing technical support for the realization of predictive maintenance of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a schematic diagram of the structure of the knowledge graph multi-set pooling network model according to an embodiment of the present invention;

[0070] Figure 2 Schematic diagram of the structure of the engine timing knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0072] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0073] See also Figure 1-2,A method for evaluating equipment health status based on knowledge graph multi-set pooling, including the following steps:

[0074] (1) Construction of engine health status data model

[0075] Extract the time series data of component entities and corresponding indicators from the engine data, and construct a time series knowledge graph ontology based on the relationship between components and indicators, such as Figure 2 As shown in (a); the engine timing knowledge graph is as follows Figure 2 As shown in (c), the equipment knowledge graph at time t is extracted separately. Figure 2 (b) shows the engine status at a certain moment;

[0076] Establish an adjacency matrix corresponding to the knowledge graph. The adjacency matrix reflects whether the nodes are directly connected, for example, Figure 2 In (b), component 2 is connected to components 1, 3, and 4. Then (A) (1,2) =(A) (2,1) =(A) (1,3) =(A) (3,1) =(A) (1,4) =(A s ) (4,1) =1, therefore, Figure 2 The direct adjacency matrix A of the knowledge graph (b) is expressed as:

[0077]

[0078] (2) Constructing a multi-set pooling network model for the engine knowledge graph

[0079] The graph multi-set pooling process includes the following: Figure 1 The two-layer graph multi-set pooling and one self-attention layer shown in the figure convert the knowledge graph feature vector h obtained by GIN into G The adjacency matrix A obtained in step (1) is pooled using formula (2) to obtain the learning result of the engine health status representation.

[0080] (3) Extracting node dimension features of the knowledge graph

[0081] The knowledge graph neural network GIN is used to represent the knowledge graph of the engine at a certain moment extracted in step (1). Its structure includes one layer of input and two layers of convolution, and the vector dimension of each layer is [4, 128, 128]. According to formula (1), the 4-dimensional initial feature vector H and adjacency matrix A obtained in step (1) are first input into the GIN network, and the feature vector H of each node in the previous layer of engine graph is calculated through two layers of convolution using formula (3). (l+1) ; Generate an embedding vector through the full graph angle through the readout function of formula (5) to obtain the representation vector h of each node in the engine knowledge graphG .

[0082] (4) Graph multi-set pooling based on knowledge graph multi-head attention

[0083] (4.1) Graph Multi-Head Attention Module

[0084] The feature vector h learned in step (2) G , adjacency matrix A, randomly initialize the attention matrix Q and input formula (9), learn the corresponding weighted values ​​through the attention function of formula (6) and (7), put more weight on the key components and indicator nodes of the engine, and at the same time learn the clustering scheme of approximate components and indicators through GNN, generate the clustering allocation matrix S to optimize the pooling process;

[0085] (4.2) Graph Multi-set Pooling

[0086] The eigenvector h G Input formula (10) with the adjacency matrix A, learn the interaction between the compressed k seed vectors and the n nodes in the original graph through the multi-head attention part of formula (9) in step (4.1), and learn the clustering distribution matrix S to obtain the representation learning result. Finally, the result of formula (9) is processed by the normalization layer and the feedforward layer to obtain the final representation learning result representing the health status of the engine;

[0087] (4.3) Node Self-Attention Mechanism

[0088] The representation learning result obtained in step (4.2) is input into formula (11), and formulas (6), (7), and (8) are cited. Through node self-attention learning, the interaction between k typical representative nodes in the graph multi-set pooling result is considered, and the multi-head attention coefficient is calculated to preferentially aggregate the feature vectors from each perspective. After activation through the normalization layer and the feedforward layer, the feature vector of the engine health status at the current moment is finally fused.

[0089] (5) Graph Multi-set Pooling Network Training and Classification

[0090] First, the final vector representation of the engine at a certain moment obtained in step (4.3) is input into formula (12) to obtain the final health status classification probability.

[0091] Secondly, the probability calculated in formula (12) is reduced according to the loss function of formula (13) With the correct health status label The loss between them makes the calculated classification results continuously approach the correct labels, making the engine health status representation vector more accurate.

[0092] The training process is continuously cycled to continuously reduce the loss, so that the embedding of the engine health status converges and the final engine health status classification result is obtained.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for equipment health status assessment based on knowledge graph multi-set pooling, characterized by: The equipment components and indicator information are deeply integrated through the time series knowledge graph. The knowledge graph is then represented by multi-set pooling learning using graph convolutional neural networks and attention mechanisms to obtain equipment health status classification results. The specific steps include: S1. Engine health status data model construction: Extract the time series data of component entities and corresponding indicators from the engine data, construct a time series knowledge graph based on the relationship between components and indicators, and establish the corresponding adjacency matrix; S2. Construct the engine knowledge graph multi-set pooling network model: The knowledge graph feature vector h obtained by the knowledge graph neural network GIN is pooled. G Pooling with the adjacency matrix obtained in step S1 to obtain the engine health status representation learning result; S3. Extract node dimension features of the knowledge graph: Use the knowledge graph neural network GIN to perform representation learning on the knowledge graph extracted in step S1 to obtain the full graph feature h of each node feature of the engine knowledge graph G ; S4. Graph multi-set pooling based on knowledge graph multi-head attention includes the following steps: S4.

1. Use the knowledge graph multi-head attention mechanism to learn the clustering scheme of approximate nodes and generate the cluster assignment matrix; S4.

2. Perform multi-set pooling on the knowledge graph and define the knowledge graph multi-set pooling GMPool. Compress n nodes into k typical representative nodes in an end-to-end pooling manner and input the full-graph feature H obtained by the graph convolution module into GMPool. S4.

3. Further use the node self-attention mechanism to consider the interactions between n input nodes or k typical representative nodes; S5. Graph Multi-set Pooling Network Training and Classification: The graph multi-set pooling network is trained cyclically to continuously reduce the loss, so that the embedding of the engine health status converges and the final engine health status classification result is obtained.

2. The equipment health status assessment method based on knowledge graph multi-set pooling according to claim 1 is characterized in that The formal definition of the temporal knowledge graph is as follows: The temporal knowledge graph is a directed label graph G t =(t0,t n ,E,R,τ), where E is the vertex set of the knowledge graph, which is used to represent the entity set; R is the edge set of the knowledge graph, which is used to represent the fact relationship set; τ is the function of E×E→R|k, which represents all tuples in the knowledge graph; k represents the time period [t0,t n ] In the knowledge graph triple list sorted by time, there is a k-th relationship R between two entities; there is a pair of nodes with a direct link relationship in the temporal knowledge graph, and the distance between the nodes is 1; let the number of entities be N, and construct the direct adjacency matrix of the relationship graph:

3. The equipment health status assessment method based on knowledge graph multi-set pooling according to claim 2 is characterized in that In step S2, the graph multi-head attention pooling network includes an input layer and two hidden layers. The input layer is the initial feature matrix and direct adjacency matrix of the equipment time series knowledge graph. The two hidden layers simultaneously learn feature information of the node dimension and the graph dimension through convolution and self-attention pooling operations. The output of the two hidden layers after training is used as the final vector representation of the health status of the component entity. The specific process is as follows: 2.

1. First use the graph neural network GNN to learn the feature vector of the node: E ncoder(X,A)=GNN2(GNN1(X,A),A) (1) Among them, X is the input feature matrix, A is the direct adjacency matrix, GNN1 is the first convolution layer, and GNN2 is the second convolution layer; 2.

2. After obtaining the feature vector of the node, use the graph multiset pooling method to generate the embedding of the entire graph: Pooling(H,A)=GMPool1(SelfAtt(GMPool k (H,A)),A′) (2) Among them, GMPool is the graph multi-head attention convolution layer, SelfAtt represents the node self-attention mechanism, H is the node feature vector obtained by the GNN layer, A is the direct adjacency matrix, and A′ is the GMPool-derived node feature vector. k After compression, the adjacency matrix is ​​coarsened to obtain the final vector representation of the component entity health status.

4. The equipment health status assessment method based on knowledge graph multi-set pooling according to claim 3 is characterized in that The specific process of step S3 is as follows: S3.

1. Input the direct adjacency matrix A and feature matrix X of the time series knowledge graph into the convolution module to aggregate the feature vectors and obtain the feature representation of the component and indicator node levels in the time series knowledge graph: Among them, UPDATE and AGGREGATE are differentiable functions of update and aggregation, N(u) represents a set of adjacent nodes of u, Initialized to the input node feature X u , H (l+1) ∈R n×d It is the node feature obtained after a GNN, which is simplified to: H (l+1) =GNN(H (l) ,A (l) ) (4) S3.2, after obtaining the node feature representation H (l+1) After that, a readout function generates an embedding vector to obtain the complete full-image representation: h G =CONTACT(sum({H v |v∈V})|k=0,1,....K) (5) Among them, h G is the full graph feature, CONTACT is the vector concatenation function, and sum is the summation function. Together they constitute the readout function, which sums the nodes of each iteration to obtain the graph features and then concatenates them to obtain the full graph features including the node features.

5. The equipment health status assessment method based on knowledge graph multi-set pooling according to claim 4 is characterized in that The specific process of step S4 is as follows: S4.

1. In order to enable the pooling method to distinguish the importance of each node in the graph, a graph multi-head attention mechanism is used to extract the clustering scheme; S4.1.

1. Assume that there are n node vectors in the graph, and the input of the attention function Att is: query matrix key matrix and value matrix where n q is the number of query vectors, n is the number of input nodes, d k is the dimension of the key matrix, d v is the dimension of the value matrix vector, and the attention function formula is as follows: That(Q,K,A)=ω(QK T )V (6) By decomposing the attention function into the dot product of the query and all keys, we can learn the corresponding weighted values ​​and put more weight on the key target nodes: Among them, ω is the activation function softmax, exp is the exponential of the natural constant e; S4.1.

2. Further use multi-head attention instead of single attention, and generate h different projection spaces by linearly projecting the query matrix Q, key matrix K and value matrix V respectively: The attention mechanism maps the learned weights to the three weight matrices Q, K, and V by inputting the feature vector of the node, so as to ensure that when encoding any part, it pays attention to the important part as much as possible; W is the adjustment matrix to generate multiple Q, K, V values, which further ensures the effectiveness of encoding learning. h represents the number of projection spaces O, which is composed of the parameter matrix composition, is the output projection matrix, d model is the output dimension of multi-head attention; S4.1.

3. In order to generate K and V more accurately in multi-head attention, the graph multi-head attention function is used to learn K and V through GNN. The output of the graph multi-head attention function GMH can be expressed as: Compared with the linearly embedded KW and VW in the multi-head formula, GNN i The output of contains the adjacency information of the graph. The multi-head attention of the graph can more easily learn the clustering scheme of the approximate nodes and generate the clustering assignment matrix. S4.

2. Knowledge graph multi-set pooling. The specific process is as follows: Define the knowledge graph multi-set pooling GMPool, compress n nodes into k typical representative nodes in an end-to-end pooling manner, and input the full graph feature H obtained by the graph convolution module into GMPool. The formula is as follows: GMPool k (H,A)=LN(Z+rFF(Z));Z=LN(S+GMH(S,H,A)) (10) Where H∈R n×d is the node feature derived by GNN, A is the adjacency matrix, rFF is the feedforward layer in any row direction, processing each row feature independently and identically, LN is the layer normalization layer, S∈R k×d It is a parameterized clustering seed matrix used to optimize the pooling operation; the GMH function considers the interaction between the k seed vectors query in S and the n node values ​​in H, and uses the attention similarity between them to compress the n nodes into k typical representative nodes; S4.

3. Node Self-Attention Mechanism: We further use the node self-attention mechanism to consider the interactions between n input nodes or k typical representative nodes. The formula is as follows: SelfAtt(H)=LN(Z+rFF(Z)); Z=LN(S+MH(H,H,H)) (11) Compared to GMH in Equation (7) which considers the interactions between k vectors and n nodes, SelfAtt places the node embedding H in the query and key positions of MH in Equation (8) to capture the intrinsic relationships between n nodes.

6. The equipment health status assessment method based on knowledge graph multi-set pooling according to claim 5 is characterized in that The specific process of step S5 is as follows: The representation vector H of the final equipment health status obtained above Final Input the final linear classifier to get the final probability in The definition is as follows: Among them, the sigmoid() function maps the vector to the output probability in the range of [0,1]; Then according to the above calculated probability Through the loss function and the correct label in the sample Perform loss calculation and minimize the loss function through Adam optimizer. The loss function is defined as: in, is the number of samples in the sample, is the correct label in the sample; The above training process is continuously iterated to converge the embedded representation of the equipment health status and obtain the final equipment health status assessment model; the equipment knowledge graph to be evaluated is input into the above training model to obtain the equipment health status classification result and complete the equipment health status assessment.

Images