A Method for Aligning Industrial Data Driven by a Hybrid of Knowledge and Data

Abstract

The present invention discloses an industrial data alignment method driven by a mixture of knowledge and data, including collecting industrial data; using the collected industrial data to set up meta-paths to construct a heterogeneous network, constructing multiple views for the constructed heterogeneous network, and generating node embedding representations by multi-view embedding; assigning weights to each meta-path and fusing the node embedding representations; combining the industrial data characterization guided by knowledge, and using contrastive learning to learn the similarity of node sequences to complete alignment. Overcoming the main defects of the above-mentioned mainstream heterogeneous network alignment technologies, the technical solution provided by the present invention reduces the impact brought by noisy data and avoids the error transmission caused by manually setting meta-paths; by combining graph representation learning and contrastive learning, the monitoring data such as temperature generated by sensors used in industry is aligned with the data in the industrial information system, effectively solving the problem of irregular industrial data.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial data processing, and particularly relates to an industrial data alignment method driven by a mixture of knowledge and data. Background Art

[0002] With the development of intelligent manufacturing, technologies such as automation, informatization, and intelligence penetrate into all aspects of the manufacturing production process. From sensors and devices in the industrial field to various information systems in the manufacturing production process (such as manufacturing execution management systems, production monitoring systems, equipment operation and maintenance systems, product quality inspection systems, energy consumption management systems, etc.), a large amount of data of different structural types will be generated. The relationship between data is becoming increasingly close, presenting a situation where data is interrelated and cross-referenced with each other. The growth of data correspondingly generates a large amount of multi-source heterogeneous data. To utilize this data to promote the development of the industry, the research on industrial heterogeneous data alignment technology is extremely urgent. Industrial data includes data generated by various sensors and data scattered in various information systems, including common SCADA systems, CMS systems for vibration status monitoring, ERP systems for assisting operation and maintenance work, etc. Different data sources result in industrial data not being as regular as data in the Internet scenario. Data collected from different information systems such as sensors, controllers, and other external systems all need to be aligned and integrated.

[0003] Currently, the mainstream technologies for processing heterogeneous network data alignment include: using node attribute information as the matching basis, using the relationship structure between nodes as the node alignment feature, and obtaining node representations through network representation learning methods and establishing node matching strategies.

[0004] Network alignment models based on network representation learning can be divided into supervised models and unsupervised models. For supervised models, although they do not rely on the attribute characteristics of nodes, they usually use anchor nodes (i.e., a set of pre-known matching nodes) as clues to establish a machine learning model to obtain node representations. However, the number of pre-known anchor nodes is often very small, or even almost non-existent, and it is only effective for some graphs with anchor nodes; it is not applicable to data sources such as industrial big data, which are neither regular nor tidy. In recent years, some unsupervised methods have been proposed to establish the expression of network nodes without any prior knowledge, mainly including two methods based on machine learning and matrix factorization. Among them, for the method based on matrix factorization, the limitations of its feature setting and the representativeness of the landmark nodes selected by the random sampling strategy affect the algorithm performance to a certain extent. Facing some problems of heterogeneous network alignment, the above-mentioned mainstream heterogeneous network alignment methods have made relatively good improvement effects on a certain specific problem, but there are still certain problems.

[0005] Based on the attribute alignment method, which relies on the attribute features of nodes. Due to considerations such as privacy protection, the alignment attributes of usernames and other important node attribute features are often missing or are of a disguised or false nature, misleading the judgment. Therefore, this method relying on node attribute features has limitations in solving the heterogeneous network alignment problem.

[0006] The premise of alignment based on the relationship structure between nodes is that there must be a certain relationship between nodes in order to find associations, and the applicable scope is relatively narrow. And this method faces a serious challenge, that is, the network structure is very sensitive to noise and structural changes. The industrial field environment is harsh, and the poor data quality is also a special case in the industrial scenario. Due to the harsh environment, the collected data contains a large amount of environmental noise, and there are many data anomaly points, which will cause serious consequences to the network structure. When the network structure changes slightly, the performance of node alignment often decreases.

[0007] The embedding of heterogeneous graphs mainly focuses on the structure information based on meta-paths. Although ESim considers the information of multiple meta-paths, it cannot learn the optimal weight combination when facing specific problems. Meta-path2vec does embedding through random walk and skip-gram algorithms, but it only considers one meta-path; HERec also only considers a single meta-path and filters node sequences through restrictive strategies for embedding. HIN2Vec uses a method of training nodes and meta-paths simultaneously. PME retains the adjacent area of nodes through Euclidean distance. HEER does the embedding of heterogeneous graphs through edge representation. Meta-graph2vec maximally retains the structural and semantic information. There are also embedding models based on meta-graphs that simultaneously consider the hidden relationships of all meta-information of a meta-graph. However, these methods still cannot achieve effective alignment of heterogeneous networks and cannot overcome the impacts and errors generated during the process, especially in the face of the complexity and particularity of industrial sensing data for heterogeneous structures, and cannot solve the problem of irregular industrial data. Summary of the Invention

[0008] To solve the above problems, starting from the defects of the heterogeneous network alignment tasks of network supervised alignment models based on attributes, node structures, and network representation learning, the present invention proposes a knowledge and data hybrid-driven industrial data alignment method, which overcomes the main defects of the above mainstream heterogeneous network alignment technologies. The technical solution provided by the present invention reduces the impact brought by noisy data and avoids the error transmission caused by manually setting meta-paths; aligns the monitoring data such as temperature generated by sensors used in industry with the data in the industrial information system through the combination of graph representation learning and contrast learning, effectively solving the problem of irregular industrial data.

[0009] To achieve the above object, the technical solution adopted by the present invention is: an industrial data alignment method driven by a mixture of knowledge and data, including the steps of:

[0010] S10, collecting industrial data;

[0011] S20, using the collected industrial data to set up meta-paths to construct a heterogeneous network, constructing multiple views for the constructed heterogeneous network, and generating node embedding representations using multi-view embedding;

[0012] S30, assigning weights to each meta-path and fusing the node embedding representations;

[0013] S40, combining the industrial data characterization guided by knowledge, and using contrastive learning to learn the similarity of node sequences to complete alignment.

[0014] Furthermore, the setting of the meta-paths includes sensor-operator-sensor, sensor-production equipment-sensor, and sensor-database-sensor;

[0015] The sensors are operated by control personnel, the production equipment detected by the sensors remains unchanged within a certain period of time, and the data of the sensors are all transmitted to the database by wireless connection.

[0016] Furthermore, the heterogeneous network includes different types of nodes, connections between nodes, and nodes contain different attribute information; the heterogeneous network includes:

[0017] There is an association relationship that the operator controls the sensor between the sensor and the operator;

[0018] There is an association relationship of monitoring and association between the sensor and the production equipment;

[0019] There is a coordination relationship between the sensors;

[0020] There is a control relationship between the operator and the production equipment;

[0021] There is a transmission relationship between the sensor and the information system.

[0022] Furthermore, define multiple views for the constructed heterogeneous network;

[0023] View Gp is composed of meta-path p and heterogeneous graph G = {V, E}, and the view based on the meta-path is composed of a type of proximity or relationship between the nodes represented by a meta-path;

[0024] Capture various aspects of the structural information through the meta-path and dynamically add new nodes.

[0025] Furthermore, using multi-view embedding to generate node embedding representations, generating node embedding representations includes:

[0026] Multiple views are generated from the meta-path. The goal of the multi-view graph embedding layer is to aggregate the attributes and topological information of local neighbors through a learning function and inductively generate node representations;

[0027] For each view based on the meta-path, the node representation is generated by aggregating the features of the meta-path-based neighbors and propagating information across K layers.

[0028] Furthermore, the representation of node v based on meta-path p is as follows:

[0029] First, at the k-th layer, each node aggregates its own representation and the representation of the 1-hop neighborhood N i generated at the (k - 1)-th layer into a single vector as:

[0030]

[0031] where, represents the representation of v j at the (k - 1)-th layer; when k = 0, is defined as the original feature x(v j ); j )

[0032] Then, the weight matrix W (k) p and the bias vector b (k) p are used to transfer information between layers:

[0033]

[0034] Furthermore, to extend the algorithm to the mini-batch setting, first sample the 1-hop neighbor network of the sensors in the batch; the 1-hop neighbor network of node v is defined as all the edges between the 1-hop neighbor nodes of node v and the nodes in the set;

[0035] For each batch, a multi-view subgraph is constructed based on the union of the l-egonnet of all sensor nodes in the batch;

[0036] Then generate the meta-path-based representation of each node in these multi-view subgraphs;

[0037] For more convenient processing, the final representation of v i based on the meta-path p after K layers is denoted as z p (v i ) ≡ z (K) p (v i ).

[0038] Furthermore, weights are assigned to each meta-path to fuse the node embedding representations, including the steps of:

[0039] Step 1: First, introduce a meta-path preference vector a p ∈R |p|*d’ for each meta-path p to guide the semantic attention mechanism; for the meta-path-based representation z (k) p and the meta-path preference vector a p , the more similar they are, the greater the weight assigned to z (k) p ;

[0040] Step 2: Use a non-linear function to convert the d-dimensional meta-path embedding into a d'-dimensional meta-path embedding, and the node representation is z' p (v i );

[0041] z′p(v i ) = σ(W p ·z p (v i ) + b p );

[0042] where Wp ∈ R |P|*d' is the weight parameter, bp ∈ R d' is the transformation bias parameter; z'p(v i ) ∈ R d' is the node representation of the meta-path p based on v i after transformation; the similarity between the transformed representation vector and the preference vector ω p (v i ) is calculated as:

[0043]

[0044] Step 3: Fuse all meta-path-based representation forms into a weighted sum form:

[0045] z(v i ) = ∑ p′∈P ω′ p′ (v i ) * zp'(v i ).

[0046] Furthermore, apply contrastive learning, using the current meta-path as the positive example sample and other irrelevant nodes as the negative example samples for training, and the similarity is calculated using the cosine scoring function;

[0047] Contrastive learning includes the steps of:

[0048] Step 1: Construct negative samples: Negative sampling directly samples non-connected k-hop nodes; select uncorrelated nodes as negative samples; uncorrelated nodes mean that there is no connection at all between the current node and the negative sample node in the heterogeneous network.

[0049] Step 2: Contrastive learning combined with industrial data knowledge traction: Learn similar prior knowledge based on the way of auxiliary learning.

[0050] Step 3: Construct a contrastive learning loss function: The similarity of two samples is calculated by the inner product of vectors. Combine the contrastive learning loss function L1 and the loss function L2 of the auxiliary task to form the final loss function L; continuously train to update the loss function.

[0051] Beneficial effects of adopting this technical solution:

[0052] Sensors on modern industrial manufacturing production lines can detect different data such as temperature, pressure, heat energy, vibration, and noise, and the data structures generated by different types of sensors are different. Workshop production involves structured data such as equipment operation parameters and product processing time from different systems, semi-structured data such as product BOM structure tables and numerical control programs, and unstructured data such as 3D models and inspection images. These data have completely different data structures. Therefore, the data alignment of the present invention not only targets the data of different sensors but also includes the data generated by different systems.

[0053] The present invention utilizes the data heterogeneous graph generated in the industrial production process to align the data collected by sensors with the data in the industrial information system. Modern industrial manufacturing production lines are equipped with thousands of small sensors to detect temperature, pressure, heat energy, vibration, and noise. Since data is collected every few seconds, various forms of analysis can be achieved using these data, including equipment diagnosis, power consumption analysis, energy consumption analysis, quality accident analysis (including violations of production regulations and component failures), etc. First, in terms of production process improvement, using these data during the production process can analyze the entire production process and understand how each link is executed. Once a certain process deviates from the standard process, an alarm signal will be generated, enabling errors or bottlenecks to be discovered more quickly and problems to be solved more easily. Using big data technology, a virtual model of the production process of industrial products can also be established to simulate and optimize the production process. When all processes and performance data can be reconstructed in the system, this transparency will help manufacturers improve their production processes. Another example is in energy consumption analysis. By using sensors to centrally monitor all production processes during equipment production, abnormal or peak energy consumption situations can be discovered, and thus the energy consumption can be optimized during the production process. Analyzing all processes will greatly reduce energy consumption.

[0054] The present invention collects data from external systems, sensors, and controllers, and preprocesses the collected data to avoid the influence caused by environmental noise in industrial data. There are different data structures in the data of different sensors and controllers, such as unstructured data, semi-structured data, and structured data in external systems, etc. Meta-paths are set. The present invention manually sets the meta-paths as: sensor - operator - sensor, sensor - production equipment - sensor, sensor - database - sensor. This avoids the limitation of relying on the attribute features of nodes for network alignment.

[0055] In the present invention, in order not to rely on anchor nodes, a multi-view inductive graph embedding layer is designed to generate node representations for each meta-path. The multi-view graph embedding layer aggregates the attribute and topological information of local neighbors by learning a function to inductively generate node representations. And the present invention designs a semantic attention layer to learn the importance of meta-paths and fuse the meta-path-based representations, and fuses these meta-path-based embeddings into a vector. The present invention introduces multi-view embedding. The views are generated based on the original paths, which maximally preserves the semantic information between nodes, and an attention mechanism is introduced. Each node generates multiple original path representations, and each original path has a different impact on the node. Different weights are assigned to each original path, and finally the original paths of the current node are fused to generate the final representation of the node.

[0056] In the present invention, combined with the industrial data characterization driven by knowledge, contrastive learning is used to learn the similarity of node sequences. Contrastive learning focuses on learning the common features among similar instances and distinguishing the differences among non-similar instances. Compared with generative learning, contrastive learning does not need to pay attention to the cumbersome details on the instances, and only needs to learn to distinguish the data in the feature space at the abstract semantic level. Therefore, the model and its optimization become simpler and have stronger generalization ability. The present invention applies contrastive learning and combines it with knowledge traction to learn the similarity of node sequences. For the negative samples of contrastive learning, non-connected k-hop nodes are directly sampled. The present invention selects uncorrelated nodes as negative samples, which can enhance the accuracy of the present invention model. Uncorrelated nodes mean that there is no connection at all between the current node and the negative sample node in the heterogeneous graph.

[0057] In the present invention, an end-to-end relationship classification model is constructed. The inputs of the model are the heterogeneous graph and the meta-path, and the outputs are the binary classification relationship and the probability value. The entire processing flow of this method is an end-to-end process, which can directly classify industrial data. Description of the Drawings

[0058] Figure 1 It is a schematic flow chart of an industrial data alignment method driven by a mixture of knowledge and data according to the present invention;

[0059] Figure 2Schematic diagram of the principle structure of a method for aligning industrial data driven by a mixture of knowledge and data in an embodiment of the present invention;

[0060] Figure 3 Is a heterogeneous network diagram in an embodiment of the present invention. Detailed implementation manners

[0061] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings.

[0062] In this embodiment, referring to Figure 1 and Figure 2 as shown, the present invention proposes a method for aligning industrial data driven by a mixture of knowledge and data, including the steps of:

[0063] S10. Collect industrial data from sensors, controllers, and other external systems;

[0064] S20. Use the collected industrial data to set meta-paths to construct a heterogeneous network, construct multiple views for the constructed heterogeneous network, and generate node embedding representations using multi-view embedding;

[0065] S30. Assign weights to each meta-path and fuse the node embedding representations;

[0066] S40. Combine the industrial data characterization guided by knowledge and use contrastive learning to learn the similarity of node sequences to complete alignment.

[0067] As an optimized solution of the above embodiment, the setting of the meta-paths includes sensor-operator-sensor, sensor-production equipment-sensor, and sensor-database-sensor;

[0068] The sensors are operated by control personnel, the production equipment detected by the sensors remains unchanged within a certain period of time, and the data of the sensors are all transmitted to the database by wireless connection.

[0069] As Figure 3 shown, the heterogeneous network includes different types of nodes, connections between nodes, and the nodes contain different attribute information; the heterogeneous network includes:

[0070] There is an association relationship that the operator controls the sensor between the sensor and the operator;

[0071] There is an association relationship of monitoring and association between the sensor and the production equipment;

[0072] There is a coordination relationship between the sensors;

[0073] There is a control relationship between the operator and the production equipment;

[0074] The transmission relationship existing between the sensor and the information system.

[0075] Define multi-views for the constructed heterogeneous network;

[0076] View Gp is constructed based on the meta-path p and the heterogeneous graph G = {V, E}. The view based on the meta-path consists of a type of proximity or relationship between the nodes characterized by a meta-path;

[0077] Capture various aspects of the structural information through the meta-path and dynamically add new nodes.

[0078] Generate node embedding representations using multi-view embedding, including:

[0079] Generate multiple views from the meta-path. The goal of the multi-view graph embedding layer is to aggregate the attributes and topological information of local neighbors through a learning function and inductively generate node representations;

[0080] For each view based on the meta-path, the node representation is generated by aggregating the features of the neighbors based on the meta-path and propagating information across K layers.

[0081] As an optimized solution of the above embodiment, the representation of node v based on the meta-path p is as follows:

[0082] First, at the k-th layer, each node aggregates its own representation and the representation of the 1-hop neighborhood N i generated at the (k - 1)-th layer into a single vector as:

[0083]

[0084] In the formula, represents the representation of v j at the (k - 1)-th layer; when k = 0, is defined as the original feature x(v j ) of v; j )

[0085] Then, use the weight matrix W (k) p and the bias vector b (k) p to transfer information between layers:

[0086]

[0087] Preferably, in order to extend the algorithm to the mini-batch setting, first sample the 1-hop neighbor network of the sensors in the batch; the 1-hop neighbor network of node v is defined as all the edges between the 1-hop neighbor nodes of node v and the nodes in the set;

[0088] For each batch, construct a multi-view subgraph based on the union of the l-egonnets of all sensor nodes in that batch;

[0089] Then generate a meta-path-based representation for each node in these multi-view subgraphs;

[0090] For more convenient processing, let v i The final representation based on the meta-path p after K layers is denoted as z p (v i ) ≡ z (K) p (v i ).

[0091] As an optimized solution to the above embodiment, assign weights to each meta-path and fuse the node embedding representations, including the steps of:

[0092] Step 1: First, introduce a meta-path preference vector a p ∈R |p|*d’ to guide the semantic attention mechanism; for the meta-path-based representation z (k) p and the meta-path preference vector a p , the more similar they are, the greater the weight assigned to z (k) p ;

[0093] Step 2: Use a non-linear function to convert the d-dimensional meta-path embedding into a d'-dimensional meta-path embedding, and the node is represented as z' p (v i );

[0094] z′p(v i ) = σ(W p ·z p (v i ) + b p );

[0095] where, Wp ∈ R |P|*d' is the weight parameter, bp ∈ R d' is the bias parameter of the transformation; z'p(v i ) ∈ R d' is the node representation of the meta-path p based on v i after transformation; the similarity between the transformed representation vector and the preference vector ω p (v i ) is calculated as:

[0096]

[0097] Step 3: Fuse all the meta-path-based representations into a weighted sum form:

[0098] z(v i ) = ∑ p′∈P ω′ p′ (v i ) * zp′(v i )。

[0099] As an optimized solution to the above embodiment, contrastive learning is applied. The current meta-path is used as a positive example sample, and other irrelevant nodes are used as negative example samples for training. The similarity is calculated using the cosine scoring function;

[0100] The contrastive learning includes the steps of:

[0101] Step 1: Construct negative samples: Negative sampling directly samples non-connected k-hop nodes. To ensure the quality of negative samples, nodes with k = 5 are selected as non-connected negative examples. Selecting irrelevant nodes as negative samples can enhance the accuracy of the model of the present invention. Irrelevant nodes mean that there is no connection at all between the current node and the negative sample node in the heterogeneous network. For example, there is no connection between the management system and the operator, and one of them can be used as the negative sample of this node.

[0102] Step 2: Contrastive learning combined with industrial data knowledge traction: Learn similar prior knowledge based on the way of auxiliary learning. For example, the production environment of some products has certain limitations, and the data generated by sensors is limited within a certain range. The auxiliary task is to learn this range.

[0103] Step 3: Construct a contrastive learning loss function: The present invention constructs an end-to-end classification model, and the end-to-end deep neural network is a black box. Although it can automatically learn some features with good distinguishability, it often fits to some non-important features. Prior information is introduced into the model, and combining with knowledge can solve the problem that the features that the model is supposed to learn are not learned by the model. The similarity between two samples is calculated using the vector inner product. The contrastive learning loss function L1 and the loss function L2 of the auxiliary task are combined to form the final loss function L; the loss function is updated through continuous training;

[0104] Then the final loss function L is:

[0105] L = L l + L 2 ;

[0106] Wherein,

[0107] x + represents the positive sample of the node, and x - j is the negative sample of the node; exp(f(x) T f(x+ )) represents the dot product of the learned representation f(x) and the positive sample f(x + ), which is actually the score of the positive sample; exp(f(x) T f(x - j )) represents the score of the negative sample.

[0108] The present invention integrates and aligns the data generated by sensors and various systems in industrial data through heterogeneous graph representation learning combined with contrastive learning methods, and models the heterogeneous network of industrial production data. First, a view based on meta-paths is constructed, and the embedding representations of nodes are learned under the constructed multi-view. Combining the semantic attention mechanism, multiple meta-paths with different weights are fused. Finally, the contrastive learning of knowledge fusion is used to learn the similarity of different nodes, and the task of data alignment classification is completed.

[0109] (1) Node embedding representation of multi-view. Previous node embedding methods are either meta-path-based or node-embedding-based. Some only consider the information of one meta-path and perform embedding by filtering node sequences through restrictive strategies. Some can consider the information of multiple meta-paths, but they cannot learn the optimal weight combination when facing specific problems. The multi-view-based node embedding method proposed by the present invention generates views from multiple meta-paths, and maximally preserves the semantic information of the heterogeneous graph.

[0110] (2) Semantic attention layer embedding. Each sensor data obtains multiple meta-path-based representations, and these representations can cooperate with each other. The influence of each meta-path on the sensor node is different. If the attention mechanism is not introduced, the problem of noise data contained in industrial data cannot be avoided, which will cause error transmission and amplification. Therefore, it is assumed that the importance of meta-paths is different, and the attention mechanism of the present invention is used to capture their contributions, so that the weights of meta-paths with greater influence on the current node increase, and the weights of the remaining meta-paths decrease relatively. Finally, the meta-paths based on node representations are fused to obtain the final node feature encoding.

[0111] (3) Combining the industrial data characterization guided by knowledge, and applying contrastive learning to learn the similarity method of node sequences. For a simple contrastive learning method, the accuracy is relatively limited when comparing similarities. Introducing the relevant knowledge of industrial data as prior knowledge can improve the accuracy of the model. The most important thing in contrastive learning is the selection of negative examples. The negative examples of the present invention are directly sampled from unconnected nodes. If a certain node has no direct association with the current node, then it is selected as a negative example. Each type of node with prior knowledge is input into contrastive learning to train the loss function.

[0112] (4) An end-to-end model for data alignment based on heterogeneous graph representation learning combined with contrastive learning. The input of the entire model is a heterogeneous graph and a meta-path, and the output is the classification of the data. To improve the effectiveness of the end-to-end heterogeneous graph representation, a semantic attention layer is constructed, which introduces a weight mechanism for different meta-path representations of each node and integrates it into the node representation.

[0113] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A method for aligning industrial data driven by a mixture of knowledge and data, characterized in that, it includes the steps: S10, Collect industrial data; S20, Use the collected industrial data to set up meta-paths to construct a heterogeneous network, construct multi-views for the constructed heterogeneous network, and generate node embedding representations using multi-view embedding; S30, Assign weights to each meta-path and fuse the node embedding representations, including the steps: Step 1: First, introduce a meta-path preference vector $\mathbf{a}$ for each meta-path $p$ p $\in \mathbb{R}$ |p|*d’ to guide the semantic attention mechanism; for the meta-path based representation $\mathbf{z}$ (k) p and the meta-path preference vector $\mathbf{a}$ p , the more similar they are, the greater the weight assigned to $\mathbf{z}$ (k) p ; Step 2: Use a non-linear function to convert the d-dimensional meta-path embedding into a d'-dimensional meta-path embedding, and the node is represented as z' p (v i ); where, \(W_p\in\mathbb{R}\) |p|*d' is the weight parameter, \(b_p\in\mathbb{R}\) d' is the bias parameter of the transformation; \(z'_p(v i )\in\mathbb{R}\) d' is the node representation of the transformed meta-path \(p\) based on \(v\) i ; the similarity between the transformed representation vector and the preference vector \(\omega p (v i ) is calculated as follows: Step three: Fuse all the meta-path-based representation forms into a weighted sum form: S40, Combine the industrial data characterization guided by knowledge, and use contrastive learning to learn the similarity of node sequences to complete the alignment.

2. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 1, characterized in that, the setting of the meta-paths includes sensor-operator-sensor, sensor-production equipment-sensor, and sensor-database-sensor; The sensors are operated by control personnel, the production equipment detected by the sensors remains unchanged within a certain period of time, and the data of the sensors are all transmitted to the database by wireless connection.

3. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 2, characterized in that, the heterogeneous network includes different types of nodes, connections between nodes, and nodes contain different attribute information; the heterogeneous network includes: There is an association relationship that the operator controls the sensor between the sensor and the operator; There is an association relationship of monitoring and association between the sensor and the production equipment; There is a coordination relationship between the sensors; There is a control relationship between the operator and the production equipment; There is a transmission relationship between the sensor and the information system.

4. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 3, characterized in that, define multi-views for the constructed heterogeneous network; View Gp is composed of meta-path p and heterogeneous graph G = {V, E}, and the view based on the meta-path is composed of a kind of proximity or relationship type between the nodes characterized by a meta-path; Capture various aspects of the structural information through the meta-path and dynamically add new nodes.

5. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 4, characterized in that, generate node embedding representations using multi-view embedding, and generating node embedding representations includes: Generate multiple views from the meta-path, and the goal of the multi-view graph embedding layer is to aggregate the attributes and topological information of local neighbors through a learning function to inductively generate node representations; For each view based on the meta-path, the node representation is generated by aggregating the features of the meta-path-based neighbors and propagating information across K layers.

6. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 5, characterized in that, the representation of node v based on meta-path p is as follows: First, at the k-th layer, each node aggregates its own representation and the representation of the 1-hop neighborhood N i generated at the (k-1)-th layer into a single vector as follows: wherein, represents the representation of v j in the (k - 1)-th layer; when k = 0, is defined as the original feature x(v j ) of j ; Then, use the weight matrix W (k) p and the bias vector b (k) p to transfer information between layers:

7. The method for aligning industrial data driven by a mixture of knowledge and data according to claim 6, characterized in that, First, sample the 1-hop neighbor network of the sensors in the batch; The 1-hop neighbor network of node v is defined as all the edges between the 1-hop neighbor nodes of node v and the nodes in the set; For each batch, construct a multi-view subgraph based on the union of the l-egonnets of all sensor nodes in that batch; Then generate the meta-path-based representations of each node in these multi-view subgraphs; Let v i Denote the final representation of the meta-path p after K layers as z p (v i ) ≡ z (K) p (v i ).

8. A method for aligning industrial data driven by a mixture of knowledge and data according to any one of claims 1-7, characterized in that, Apply contrastive learning, use the current meta-path as a positive example sample, and other irrelevant nodes as negative example samples for training, and calculate the similarity using a cosine scoring function; The contrastive learning includes the steps of: Step 1: Construct negative samples: Negative sampling directly samples non-connected k-hop nodes; Select irrelevant nodes as negative samples; Irrelevant nodes mean that there is no connection at all between the current node and the negative sample node in the heterogeneous network; Step 2: Contrastive learning combined with industrial data knowledge traction: Learn similar prior knowledge based on an auxiliary learning method; Step 3: Construct a contrastive learning loss function: Calculate the similarity of two samples using the vector inner product, and combine the contrastive learning loss function L1 and the loss function L2 of the auxiliary task to form the final loss function L; Continuously train to update the loss function.

Images