A fusion management architecture for multi-source heterogeneous industrial data

Abstract

The application relates to a fusion management architecture for multi-source heterogeneous industrial data, comprising: an edge node, which is used for encapsulating original multi-source heterogeneous industrial data into a Boolean shared representation, and uploading the encrypted data to two independent servers after encryption; a server, which is used for cleaning the received data and sending the data to a multi-source heterogeneous industrial data unified representation platform; the multi-source heterogeneous industrial data unified representation platform, which is used for uniformly representing the cleaned multi-source heterogeneous industrial data, making the data into multi-source homogeneous industrial data, and storing and inquiring the data; and an industry logic mechanism library, which is used for mining the industrial data stored in the multi-source heterogeneous industrial data unified representation platform, and mining the potential relationship between the data. The application maximally utilizes multi-source data and sufficiently mines the semantic relationship between the data based on different business scenes, completes knowledge completion and information fusion, and depicts the connection relationship and structure relationship of full-element interconnection.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to a fusion management architecture for multi-source heterogeneous industrial data. Background Technology

[0002] Industrial production data is scattered across various control systems, with business logic fragmented and exhibiting severe heterogeneous data silos. Harsh production environments mean that collected data may contain significant noise or outliers. Furthermore, industrial data is susceptible to background factors such as parameter settings and operating conditions. Therefore, the industrial sector faces increasingly stringent requirements for the fusion and management of multi-source heterogeneous data.

[0003] In recent years, both domestic and international researchers have conducted extensive research on information models based on production application needs, resulting in new breakthroughs in industrial data management technologies and methods. However, due to the differences in business and application scenarios across industrial sectors, the integrated management solutions for multi-source heterogeneous production data are characterized by their refined nature and narrow applicability. Furthermore, most related research struggles to meet the real-time requirements of production data, lacks a comprehensive top-level design, and still falls short of achieving widespread interconnection. Therefore, constructing and implementing an integrated management framework for multi-source heterogeneous industrial data is of paramount importance. Summary of the Invention

[0004] The purpose of this invention is to provide a fusion management architecture and implementation method for multi-source heterogeneous industrial data, so as to overcome the defects of the above-mentioned industrial big data management system.

[0005] The technical solution adopted by this invention to achieve the above objectives is: a fusion management architecture for multi-source heterogeneous industrial data, comprising:

[0006] Edge nodes are used to encapsulate raw, multi-source, heterogeneous industrial data into a Boolean shared representation, perform encryption operations, and then upload the encrypted data to two independent servers.

[0007] The server is used to clean the received data and send it to the unified characterization platform for multi-source heterogeneous industrial data.

[0008] A unified representation platform for multi-source heterogeneous industrial data is used to uniformly represent cleaned multi-source heterogeneous industrial data, making it multi-source homogeneous industrial data for storage and querying.

[0009] The industry logic mechanism library is used to perform data mining on industrial data stored in a unified representation platform for multi-source heterogeneous industrial data, and to uncover potential relationships between the data.

[0010] The unified representation platform for multi-source heterogeneous industrial data includes a data survey layer, a data integration layer, and a data reporting layer.

[0011] The data survey layer is used to construct entity, process library and data dictionary tables for industrial data, so as to complete the storage format definition of entities and output enterprise business processes;

[0012] Among them, entities are used to store the names and attributes of the currently stored data, so as to classify and define the data and clarify the relationship between business domains and business objects;

[0013] A process library is used to describe the classification, hierarchy, hierarchy boundaries, hierarchy scope, and hierarchy relationships of enterprise business processes.

[0014] The data dictionary table consists of enumerable data and is used to store enumerable data as well as the names of enterprise business process rules.

[0015] The data integration layer includes a logical model that utilizes relationships between entities and describes enterprise business process rules, as well as a physical model for data encapsulation formats used for visualization, in order to achieve a unified representation of multi-source heterogeneous industrial data.

[0016] The data reporting layer enables the creation, deletion, modification, and querying of data, as well as the import and export of data and the aggregation of resources, in order to form a complete industry information architecture.

[0017] The data dictionary is used to statistically analyze and describe data items, data structures, and processing logic, including:

[0018] Dictionary-like dictionary class This is used to store the category to which all data items belong;

[0019] Attribute dictionary dic properity It is used to store the various attributes of all different data items;

[0020] Data type dictionary dic type , a data type used to store all data;

[0021] Unit dictionary dic unit This is used to store all units that may be involved in the data;

[0022] relational dictionary dic relation It is used to store all possible logical relationships between different data;

[0023] rule dictionary dic rule This is used to store various rules for data retrieval in different business scenarios;

[0024] Formula dictionary dic equation It is used to store all possible formulas that may exist between all data.

[0025] The logical model Frame The formal representation is as follows:

[0026] Model Frame ={Name,C,Field,Rel,F,T};

[0027] Field={Fname,F_code,type,standard,key};

[0028] Rel={R_name,R_code,rel,rel_f,rel_child};

[0029] F={F_name,F_code,R_code,sour_chart,sub_chart}

[0030] Name is the model name; C is the model code; Field is the entity field sequence of the model, used to store information about all entities involved in the logical model. Multiple field sequences of multiple entities constitute the field table of the current model. Fname is the entity name, F_code is the entity code, type is the entity type, standard is the entity data standard, and key=1 indicates that the current entity is a key entity; Rel is a single logical relation parameter sequence, used to store the logical relationship between entities. A logical model may contain multiple logical relations, so a set of Rel can form a Rel table, where R_name represents the logical relation name, R_code represents the relation code, rel is the current logical rule, rel_f is the code corresponding to the parent logic, and rel_child is the code corresponding to the child logic. Multiple logics can form a logical relation table; F is the mapping parameter sequence, which establishes the field mapping relationship between the source entity table and the destination entity table, where F_name is the mapping name, F_code is the mapping code, source_chart is the source entity table, and sub_chart is the destination entity table; T is the model information update time;

[0031] The physical model is used to convert the entities, attributes, attribute constraints, and logical relationships defined in the logical model into a table relationship diagram that can be recognized by database software.

[0032] A method for implementing a fusion management architecture for multi-source heterogeneous industrial data includes the following steps:

[0033] Step 1: The edge node encapsulates the raw multi-source heterogeneous industrial data into a Boolean shared representation, performs encryption operations, and then uploads it to two independent servers 1 and 2 in the cloud.

[0034] Step 2: Server 1 and Server 2 respectively calculate the AVF score of the received Boolean shared representation data, and delete the Boolean shared representation data with a score lower than the threshold, thus completing the cleaning operation of the encrypted data;

[0035] Step 3: Upload and store the data obtained after cleaning to the unified characterization platform for multi-source heterogeneous industrial data;

[0036] Step 4: Perform data mining on the industrial data stored in the unified representation platform for multi-source heterogeneous industrial data to uncover potential relationships between the data.

[0037] The edge node encapsulates the original data into a Boolean shared representation and performs encryption operations, including the following steps:

[0038] The raw data x of the j-th attribute collected from the i-th edge node ij Convert to binary mth position Perform a Boolean operation with a randomly selected Boolean number r∈{0,1} at the current edge node, so that the m-th Boolean received by server 1 and server 2 are shared as follows: and Where m = 1, ..., l, l represents a binary number. The total number of digits.

[0039] Step 2, the construction of the unified characterization platform for multi-source heterogeneous industrial data, includes the following steps:

[0040] Step 2.1: The data research layer constructs entity, process library and data dictionary tables for industrial data, which are used to define the storage format of entities and output enterprise business processes;

[0041] Step 2.2: The data integration layer uses the relationships between data entities to describe the logical model of business rules, and constructs a data encapsulation format as a physical model based on the logical model to achieve visualization;

[0042] Step 2.3: The data reporting layer enables data creation, deletion, modification, querying, import, export, and resource aggregation to form a complete industry information architecture.

[0043] The relationship-aware graph neural network and knowledge-aware graph neural network based on the message-passing paradigm are used to mine new information from multi-source heterogeneous data, including the following steps:

[0044] Step 3.1: Based on the data stored in the unified representation platform for multi-source heterogeneous industrial data and the logical relationships between the data, and combined with business scenarios, construct a knowledge graph G with entity data as nodes and logical relationships as directed edges. m =(V m E m), where V m and E m Knowledge graph G m The knowledge graph is composed of a set of nodes and edges, and is divided into a topological graph G that focuses on node information. u =(U,E u The dependency graph G on the relationships between the nodes of interest. v =(V,E) v Two local information graphs, where U and E u The topology diagram G is shown below. u The node set and edge set in V and E v These are the dependency graphs G and G. v The set of nodes and the set of edges;

[0045] Step 3.2: Input the node features and edge features in the knowledge graph into the relationship-aware graph neural network and the knowledge-aware graph neural network based on the message passing paradigm to obtain feature representations of various interaction relationships between different data.

[0046] Step 3.3: According to G u =(U,E u ), G v =(V,E) v The feature representation of different interaction relationships between data nodes can be used to mine new information.

[0047] The relation-aware graph neural network is expressed as follows:

[0048] For a knowledge graph G m =(V m E m The node features and edge features are input into the relational perceptual neural network, that is, from the topological graph G... u =(U,E u ) and dependency graph G v =(V,E) v Two local structures learn node embeddings simultaneously, node u i and nodes The edge between This indicates that the relationship type between the two nodes is k, and the message structure between the two nodes is as follows:

[0049]

[0050] Where, N i Represents node u i The number of neighbors, This indicates a relationship between node u and relation type k. i Connected v j The number of nodes Represents a node The feature embedding, where W1 is the weight matrix corresponding to the node set;

[0051] Includes timing context coding The message propagation pattern is defined as

[0052]

[0053] in, Here, is the LeakyReLU activation function, and l is the number of layers in the current graph neural network. For the node u under timestamp t i and The corresponding context encoding, Represents node u at level (l+1) i Features Represents node u at level l i Features Represents node u at level l i The message being transmitted to oneself Represents node u i The number of neighbors, Represents a node The number of neighbors, For vector addition, This represents the weight matrix when nodes at layer l propagate messages to themselves. Represents the node at level l To node u i The weight matrix used for message propagation Represents the node at level l To node u i Message embedding during message propagation.

[0054] Step 3.3 is as follows:

[0055] Using a knowledge-aware graph neural network to analyze the topological structure graph G of a knowledge graph u =(U,E u ) and dependency graph G v =(V,E) v Two local structures are learned simultaneously, and the embeddings (Z) of the two graphs are learned respectively. u Z v Then, the global embedding is obtained by aggregation using an aggregation function, and the semantic knowledge contained in this knowledge graph is mined, where Z u Represents the topology graph G u Embedded, Z v Representing the dependency graph G v Embedding;

[0056] The update function for graph embedding is as follows:

[0057]

[0058]

[0059] in, Represents graph G u The i-th node u I Embedded, Represents graph G v The J-th node v J The embedding is denoted by η, which represents the normalization function used when performing node aggregation, and ω, which represents the average pooling function for node embedding.

[0060] The present invention has the following beneficial effects and advantages:

[0061] 1. The framework provided by this invention effectively cleans multi-source heterogeneous industrial data while ensuring data privacy. Based on different business scenarios, it maximizes the use of multi-source data, fully explores the semantic relationships between data, completes knowledge completion and information fusion, and depicts the interconnected and structural relationships of all elements.

[0062] 2. This invention, through research on low-quality, multi-source production factors and their integrated management, explores an integration mechanism to meet the needs of the industrial sector, solving the problem of cross-domain, multi-level information interconnection in industry. It consists of three parts: industrial data cleaning based on federated learning, designing a unified representation platform for multi-source heterogeneous industrial data, and constructing an industry logic mechanism library based on knowledge graphs. This addresses the challenge of interconnecting heterogeneous data sources. The integrated management framework for multi-source heterogeneous industrial data provides an integration mechanism that meets the needs of the industrial sector, solving the problem of cross-domain, multi-level information interconnection in industry. Especially in the face of inconsistent protocol standards across industries, dispersed storage of industrial production factors, and the problems of massive heterogeneity and low quality, it meets the needs for accurate representation and integrated management of complex factors. Attached Figure Description

[0063] Figure 1 This is a general framework diagram of the present invention;

[0064] Figure 2 This is a schematic diagram of the federal data cleaning system model of the present invention;

[0065] Figure 3 This is a schematic diagram of the CrowDEC crowdsourcing framework of the present invention;

[0066] Figure 4 This is a schematic diagram of the relation-aware graph neural network based on the message-passing paradigm of the present invention;

[0067] Figure 5 This is a schematic diagram of the knowledge-aware graph neural network of the present invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0069] This invention presents a federated learning-based encrypted data cleaning method that, before storing raw industrial data in a unified representation platform, first encrypts and uploads it to a cloud server for data cleaning. It also utilizes a crowdsourcing strategy to improve data cleaning efficiency, thereby more efficiently checking the consistency of raw industrial data and handling invalid and missing values ​​while protecting data privacy. The method includes the following processes:

[0070] Step 1: Treat different devices in different business scenarios as edge nodes, encapsulate the raw industrial data they collect into a Boolean shared representation, perform encryption operations, and then upload it to two independent cloud servers - Server 1 and Server 2;

[0071] Step 2: Server 1 and Server 2 simultaneously calculate the AVF score of the received data Boolean shared representation and delete the Boolean shared representation with the lower score to complete the cleaning operation of the encrypted data;

[0072] Step 3: Utilize crowdsourcing strategies to improve the efficiency of steps 1 and 2, and upload and store the high-quality data obtained after cleaning to the unified representation platform.

[0073] The framework for the integrated management of multi-source heterogeneous industrial data includes the following steps:

[0074] Step 1: Industrial data cleaning method based on federated learning. The specific process is as follows:

[0075] Step 1.1: Two independent servers are introduced. Before uploading collected data to the servers, users encapsulate the data into a Boolean shared object. The servers then perform Boolean shared operations on the uploaded data. Throughout the entire process, the server does not obtain the original data mapping, thus protecting the data.

[0076] Step 1.2: Performing Boolean operations on each data item as described above would consume a significant amount of memory. Therefore, the incentive-based crowdsourcing framework CrowDEC is used to improve computational efficiency while maintaining quality.

[0077] Step 2: Based on the characteristics of industrial data and incorporating the concept of a data lake, design a third-party multi-source heterogeneous industrial data cloud platform that includes a data research layer, a data integration layer, and a data reporting layer. This platform will unify the definition and representation of the formats of multi-source heterogeneous data. The specific process is as follows:

[0078] Step 2.1: The data research layer comprises a subject library, a process library, and a data dictionary table. The subject library is designed using a three-tier architecture: subject domain grouping, subject domains, and business objects. The process library describes the classification, hierarchy, boundaries, scope, and relationships of enterprise processes. The data dictionary table consists of enumerable data, storing the mapping relationship between enumerable data names and their codes.

[0079] Step 2.2: The data integration layer completes the relationship modeling of industrial production factors. Relationship modeling conforms to 3NF in paradigm theory and includes three levels: conceptual relationship modeling, logical relationship modeling, and physical relationship modeling.

[0080] Step 2.3: The data reporting layer uses the dimensional model as its core to design a fusion management mechanism for multi-source heterogeneous data, realizes functions such as adding, deleting, modifying, querying, importing, exporting, and resource aggregation of data, and forms a complete industry information architecture.

[0081] Step 3: Construct an industry logic mechanism library based on knowledge graphs. The specific process is as follows:

[0082] Step 3.1: Connect various data sources and combine them with prior industry knowledge and logical mechanisms to design an industry-specific knowledge graph G. m =(V m E m ).

[0083] Step 3.2: Design a relation-aware graph neural network based on the message-passing paradigm. Define G. u =(U,E u G represents a topological graph of a knowledge graph, and further defines G as follows: v =(V,E) v ) represents the interdependence graph between nodes v. The key lies in designing the message structure between two nodes in the neural network and the message propagation mode that includes temporal context encoding.

[0084] Step 3.3: Design a knowledge-aware graph neural network. To capture the ability to perceive local and global graph substructures, from G... u =(U,E u ) and G v =(V,E) v Two local structures are learned simultaneously. The embeddings (Zi) of the two graphs are learned separately. u Z v Then, the global embedding is obtained by aggregation through aggregation functions.

[0085] The framework for the fusion and management of multi-source heterogeneous industrial data involved in this invention mainly includes three technologies: First, based on the federated learning concept and combined with crowdsourcing technology, industrial data fusion and cleaning are performed without compromising data privacy, solving the problem of low-quality multi-source heterogeneous data; Second, a third-party data integration cloud platform is built based on the data lake concept to build a bridge for data interconnection between enterprises and to store the cleaned high-quality data; Third, based on the high-quality data stored in the cloud platform, an industry knowledge graph is constructed to manage industry mechanisms, and graph neural network technology is used to mine industry knowledge, build an industry logic mechanism library, and realize the organic combination of industrial mechanisms and big data technology. The overall flowchart of this invention is as follows. Figure 1 As shown in the figure, the construction of the unified representation and fusion management framework for multi-source heterogeneous industrial data in this embodiment is as follows:

[0086] Step 1: Construct an industrial data cleaning method based on federated learning to transform low-quality, multi-source, heterogeneous industrial data into high-quality, multi-source, heterogeneous data. The model structure of the federated data cleaning system is as follows: Figure 2 As shown, the CrowDEC framework is as follows: Figure 3 As shown. The specific process is as follows:

[0087] Step 1.1: As Figure 2 As shown, each different edge node collects raw data. In the data management process, raw data generally refers to data directly from source files (business system databases, offline files, IoT, etc.), or "copy data" directly copied from source files, such as massive amounts of key-value data collected by different sensors like photoelectric, thermal, and force sensors, image and audio data, and interface data in txt, JSON, and XML formats. These source data vary significantly in content, storage format, and parsing software, resulting in a variety of data cleaning methods. For N edge nodes, X... i (i = 1, ..., N) represents the dataset collected by the i-th edge node, where each data point contains M attributes. Then X i This can be represented as an N x M matrix. When data cleaning is required, the edge nodes upload the encrypted data to two independent servers, Server 1 and Server 2, in the cloud. The servers calculate the AVF (Attribute Value Frequency Algorithm, a numerical anomaly detection algorithm; the smaller the value, the more anomalous) score, 'a'. i The function is: Where x ij This represents the value of the j-th attribute collected by the i-th edge node. Lower-scoring data is removed, thus completing the cleaning of a single data point. The core of data encryption lies in how the i-th edge node processes the collected raw data X... i Convert the corresponding Boolean sharing and The data is uploaded to server 1 and server 2 respectively. Specifically, the raw data x of the j-th attribute collected from the i-th edge node is first uploaded. ij Convert to binary Its kth position Performing a Boolean operation with a randomly selected Boolean number r∈{0,1} at the current edge node, the k-th Boolean share received by server 1 and server 2 are respectively... and Subsequently, the server receives the Boolean share. and Perform AVF scoring and sorting, delete the data with low scores, i.e. delete low-quality data, and obtain the remaining high-quality data.

[0088] Step 1.2: Performing Boolean operations on each collected raw data item with a large amount of data would consume a significant amount of memory. Therefore, this framework utilizes the incentive-based crowdsourcing framework CrowDEC to improve computational efficiency while maintaining quality. The framework structure is as follows: Figure 3 As shown, the crowdsourcing framework has two main modules: a rejection module for workers with poor computational quality and an incentive module with different strategies.

[0089] In this embodiment, Figure 2 This ensures data security. Figure 3 While ensuring quality, improve computational efficiency, clean low-quality data into high-quality data, and apply it to subsequent research;

[0090] Step 2: To store the high-quality data obtained from the cleaning process in Step 1, this invention designs a unified representation platform for multi-source heterogeneous data. Addressing the dispersed nature of industrial production factor information and the diverse data storage formats, this invention researches a multi-dimensional unified representation method for multi-source heterogeneous data. The unified representation platform for multi-source heterogeneous data consists of three parts: a data survey layer, a data integration layer, and a data reporting layer. The specific design process is as follows:

[0091] Step 2.1: The data research layer completes the design of entity, process library, and data dictionary tables for industrial data. This is used to define the storage format of data entities and output enterprise business processes, such as the process initiator, business name, executors, and specific order of implementation in the metal component casting business process, enabling data investigation, requirements analysis, and business research. Entities primarily store the names and attributes of currently stored data, such as the name, value, unit, and collection time of data collected by thermal sensors. This facilitates data classification and definition, helps clarify data assets, and defines the relationships between business domains and business objects. The process library is a structured overall framework for processes, describing the classification, hierarchy, boundaries, scope, and relationships of enterprise processes. For example, it categorizes decision-makers, managers, and users in the metal component casting business process, showing the top-down or whole-to-part hierarchy and the specific order of business implementation, reflecting the enterprise's business model and characteristics. The data dictionary table, composed of enumerable data, stores all appearing enumerable data and rule names. This invention defines seven data dictionaries to define and describe data items, data structures, and processing logic, including a class dictionary `dic`. class A dictionary (dic) containing the categories and attributes of all data items. properity A dictionary (dic) that stores the various attributes and data types of all different data items. type Data type and unit dictionary (dic) that stores all data unit A dictionary (dic) containing all units and relationships that may be involved in the data. relation A dictionary (dic) that stores all possible logical relationships and rules between different data. rule A dictionary (dic) containing rules and formulas for data retrieval in different business scenarios. equation It stores all possible formulas that may exist between all the data.

[0092] Step 2.2: Based on the definitions of various industrial data storage and usage frameworks established in Step 2.1, the data integration layer designs a relational model at two levels: a logical model and a physical model. The logical model accurately describes business rules using the relationships between data entities, such as causal relationships and formula derivation relationships. Simultaneously, the logical model provides a strong reference for constructing the physical model, making it a crucial stage in the final successful database design process. This invention's data logical relational model (Model) Frame The formal representation is as follows:

[0093] Model Frame ={Name,C,Field,Rel,F,T};

[0094] Field={Fname,F_code,type,standard,key};

[0095] Rel={R_name,R_code,rel,rel_f,rel_child};

[0096] F={F_name,F_code,R_code,sour_chart,sub_chart}

[0097] Name is the model name; C is the model code; Field is the entity field sequence of the model, used to store information about all entities involved in the relational logic model. Multiple field sequences of multiple entities constitute the field table of the current model. Fname is the entity name, F_code is the entity code, type is the entity type, standard is the entity data standard, and key=1 indicates that the current entity is a key entity; Rel is a single logical relation parameter sequence, used to store the logical relationship between entities. A logical model may contain multiple logical relationships, so a set of Rel can form a Rel table, where R_name represents the logical relation name, R_code represents the relation code, sub_chart represents the destination table, sub_field represents the destination entity, rel is the current logical rule, rel_f is the code corresponding to the parent logic, and rel_child is the code corresponding to the child logic. Multiple logics can form a logical relation table; F is the mapping parameter sequence, which establishes the field mapping relationship between the source entity table and the destination entity table, where F_name is the mapping name, F_code is the mapping code, source_chart is the source entity table, and sub_chart represents the destination entity table; T is the model information update time.

[0098] Based on the logical data model, and combined with various specific technical implementation factors, a physical model is designed, i.e., database architecture design, to realize the storage of data in the database. A physical model refers to a physical description, i.e., a visual data encapsulation format, that transforms the entities, attributes, attribute constraints, logical relationships, and other elements defined in the logical model into a table relationship diagram that database software can recognize, according to certain rules and methods.

[0099] Step 2.3: After addressing the low-level design of industrial data encapsulation and storage functions in the data research and data integration layers, the data reporting layer, centered on the dimensional model, designs a multi-source heterogeneous data fusion management mechanism. This enables functions such as data creation, deletion, modification, querying, import / export, and resource aggregation, forming a complete industry information architecture. Dimensions are perspectives used to observe and analyze business data, supporting data aggregation, drill-down, and slice analysis. Dimensions often have hierarchical structures, such as geographical dimensions (including content at the country, region, province, and city levels), time dimensions (including content at the year, quarter, and month levels), a department containing multiple project teams, or different machines within the same workshop. Creating dimensions involves standardizing the existence and uniqueness of entities (or master data) in the business from the top level. After the dimensional model is established, dimension tables are created to manage dimensions. Dimension modeling is a priori, similar to inclusion relationships, such as chapters: 2, 2.1, 2.1.1; dimensions model these inclusion relationships. For example, business scenarios under a business process are called business dimensions, and dates are called enumeration dimensions.

[0100] This invention designs a third-party unified representation platform for multi-source heterogeneous data, based on industrial data itself and incorporating the concept of data lakes. This platform unifies the definition and representation of the formats of multi-source heterogeneous data, laying a solid foundation for the interconnectedness and structural relationships of all elements in the future Industrial Internet.

[0101] Step 3: Steps 1 and 2 completed the cleaning method for existing multi-source heterogeneous industrial data and the design of a unified representation platform for multi-source heterogeneous data, but did not actively conduct deeper analysis of the data. Therefore, this invention utilizes the industrial data stored in Step 2 to design a relation-aware graph neural network and a knowledge-aware graph neural network based on the message-passing paradigm to mine new knowledge that may exist in the multi-source industrial data. This constructs an optimized industry logic mechanism library, which is added to the unified industrial data representation platform to optimize the industrial data fusion management mechanism. The relation-aware graph neural network based on the message-passing paradigm is as follows: Figure 4 As shown, the knowledge-aware graph neural network is as follows: Figure 5 As shown, the specific process is as follows:

[0102] Step 3.1: Based on the industrial data stored in the industrial data cloud platform in Step 2 and the logical relationships between the data, and combined with the business scenario, construct an industry-specific knowledge graph G with entity data as nodes and directed edges for logical relationships. m =(V m E m This knowledge graph is then broken down into a topological structure graph G. u =(U,E u ) and dependency graph G v =(V,E) v Two partial information diagrams.

[0103] Step 3.2: As Figure 4 As shown, the node features and edge features in the knowledge graph are input into a relation-aware graph neural network based on the message-passing paradigm to obtain feature representations of various interaction relationships between different data. This neural network model consists of two core parts: information construction and information propagation. For a knowledge graph G... m =(V m E m ), node u i and nodes The edge between This indicates that the relationship between the two nodes is of type k (note that the types of the two nodes are also different), and the message structure between the two nodes is defined as follows:

[0104]

[0105] Where, N i Represents node u i The number of neighbors, This indicates a relationship between node u and relation type k. i Connected v j The number of nodes Represents a node The feature embedding, where W1 is the weight matrix corresponding to the node set;

[0106] Includes timing context coding The message propagation pattern is defined as

[0107]

[0108] in, Here, is the LeakyReLU activation function, and l is the number of layers in the current graph neural network. For the node u under timestamp t i and The corresponding context encoding, Represents node u at level (l+1) i Features Represents node u at level l i Features Represents node u at level l i The message being transmitted to oneself Represents node u i The number of neighbors, Represents a node The number of neighbors, representing the number of nodes at level l. To node u i The weight matrix during message propagation represents the weight matrix when a node at layer l propagates a message to itself. This is vector addition.

[0109] Step 3.3: After learning the feature representations of different interaction relationships between various data, it is necessary to further mine potential knowledge from them. Therefore, this invention chooses to utilize a knowledge-aware graph neural network to extract potential knowledge from the topological structure graph G of the knowledge graph. u =(U,E u ) and dependency graph G v =(V,E) v Two local structures are learned simultaneously. The embeddings (Zi) of the two graphs are learned separately. u Z v Then, aggregation is performed using an aggregation function to obtain the global embedding. This allows for the extraction of abstract knowledge contained within the knowledge graph. For example, from a steel smelting knowledge graph generated from the steel smelting process of a certain steel plant, a new relationship between raw material and product yield rates under a specific set of equipment in the steel smelting process can be extracted, and this new relationship can be stored as new knowledge in the industry mechanism database. The graph embedding update function is as follows:

[0110]

[0111]

[0112] During training, node transformations are randomly selected to generate misaligned node relationships as negative samples. Ultimately, new semantic knowledge mined from different knowledge graphs across various business scenarios is used for logical and physical modeling, constructing an industry logic mechanism library and enriching the content of the industrial data cloud platform.

[0113] In this example, knowledge perception learning is performed on the knowledge graph generated from existing industry mechanisms to uncover new semantic knowledge, further enriching the existing industry mechanisms. Knowledge mining is continuously conducted on industry mechanisms under different business scenarios, ultimately constructing an industry mechanism library.

[0114] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. The scope of the present invention is defined only by the appended claims.

Claims

1. A fusion management architecture for multi-source heterogeneous industrial data, characterized in that, Comprise: Edge node, for the original multi-source heterogeneous industrial data encapsulation for Boolean shared representation, after encryption operation, the encrypted data is uploaded to two independent servers respectively; First, the first i The first edge node collected the first j The original data of each attribute Convert to binary its k-th position k=1,…, L Performing a Boolean operation with a randomly selected Boolean number r∈{0,1} at the current edge node, the k-th Boolean share received by the first server and the second server are respectively... and Where k=1,…, L , L Representing binary numbers The total number of digits; Server, for cleaning the received data, and sending to the multi-source heterogeneous industrial data unified representation platform; Multi-source heterogeneous industrial data unified representation platform, for unified representation of the cleaned multi-source heterogeneous industrial data, making it become multi-source homogeneous industrial data, to store and query data; Industry logic mechanism library, for data mining of the industrial data stored in the multi-source heterogeneous industrial data unified representation platform, to mine the potential relationship between data.

2. The fusion management architecture for multi-source heterogeneous industrial data according to claim 1, wherein, The multi-source heterogeneous industrial data unified representation platform comprises a data research layer, a data integration layer and a data report layer; The data research layer is used to build an entity, a process library and a data dictionary table of industrial data, to complete the definition of the storage format of the entity, and output the enterprise business process; The entity is used to store the current storage data name and its attribute, to classify and define the data, and to clarify the association relationship between business domain and business object; The process library is used to describe the classification, level, level boundary, level range and level relationship of the enterprise business process; The data dictionary table is composed of enumerable data, and is used to store the enumeration data and the enterprise business process rule name; The data integration layer comprises a logical model describing the enterprise business process rules by using the relationship between entities, and a physical model of data packaging format for visualization, to realize the unified representation of multi-source heterogeneous industrial data; The data report layer realizes the addition, deletion, modification and query of data, import and export, resource summary, to form a complete industry information architecture.

3. The fusion management architecture for multi-source heterogeneous industrial data according to claim 2, characterized in that: The data dictionary is used to statistically and describe the data item, data structure and processing logic, Comprise: Class dictionary dic class for storing the class to which all data items belong; Attribute dictionary dic properity for storing various attributes of all different data items; Data type dictionary dic type Data type for storing all data; unit dictionary dic unit for storing all units that can be involved in the data relationship dictionary dic relation for storing all possible logical relationships between different data; Rules dictionary dic rule for storing various rules for data call under different service scenarios; Formula dictionary dic equation for storing all formulas that can exist between the data.

4. The fusion management architecture for multi-source heterogeneous industrial data according to claim 2, characterized in that: The logic model Model Frame The formal representation is as follows: Model Frame ={ Name, C, Field, Rel, F, T}; Field ={ Fname, F_code, type, standard, key}; Rel ={ R_name, R_code, rel, rel_f, rel_child}; F = { F_name, F_code, R_code, sour_chart, sub_chart} Name is the model name; C is the model code; Field is the entity field sequence of the model, which stores the information of all entities involved in the logical model, and multiple field sequences of multiple entities constitute the field table of the current model, Fname is the entity name, F_code is the entity code, type is the entity type, standard is the entity data standard, key=1 indicates that the current entity is a key entity; Rel is a single logical relationship parameter sequence, which stores the logical relationship between entities. A logical model may contain multiple logical relationships, so a set of Rel may constitute Rel table, wherein R_name indicates the logical relationship name, R_ code indicates the relationship code, rel is the current logical rule, rel_f is the parent logical corresponding code, rel_child is the child logical corresponding code, and multiple logicals can constitute a logical relationship table; F is a mapping parameter sequence, which indicates the mapping relationship of the fields between the source entity table and the destination entity table, wherein F_name is the mapping name, F_code is the mapping code, sour_chart is the source entity table, sub_chart indicates the destination entity table; T is the model information update time; The physical model is used to convert the entity, attribute, attribute constraint and logical relationship defined in the logical model into a table relationship diagram that can be recognized by database software.

5. A fusion management architecture implementation method for multi-source heterogeneous industrial data, characterized in that, Comprise the following steps: Step 1: the edge node encapsulates the original multi-source heterogeneous industrial data into Boolean shared representation, and uploads it to the first server and the second server on the cloud after encryption operation; Step 2: The first server and the second server respectively perform AVF score calculation on the received Boolean shared representation data, and delete the Boolean shared representation data with scores lower than the threshold, to complete the cleaning operation of the encrypted data; Step 3: Upload and store the data obtained after cleaning to the multi-source heterogeneous industrial data unified representation platform; Step 4: Perform data mining on the industrial data stored in the multi-source heterogeneous industrial data unified representation platform to mine the potential relationship between the data; The encapsulation is a Boolean shared representation, and the encryption operation is specifically as follows: First, the first i The first edge node collected the first j The original data of each attribute Convert to binary its k-th position k=1,…, L Performing a Boolean operation with a randomly selected Boolean number r∈{0,1} at the current edge node, the k-th Boolean share received by the first server and the second server are respectively... and Where k=1,…, L , L Representing binary numbers The total number of digits.

6. The method of claim 5, wherein the method further comprises: In step 2, the construction of the multi-source heterogeneous industrial data unified representation platform includes the following steps: Step 2.1: The data research layer constructs an entity, a process library and a data dictionary table of industrial data, which is used to define the storage format of the entity and output the enterprise business process; Step 2.2: The data integration layer describes the logical model of the business rule by using the relationship between the data entities, and constructs the data encapsulation format as a physical model according to the logical model, to realize visualization; Step 2.3: The data report layer realizes the data adding, deleting, modifying, inquiring, importing and exporting, and resource summarizing, to form a complete industry information architecture.

7. The method of claim 5, wherein the method further comprises: The relationship-aware graph neural network and the knowledge-aware graph neural network based on the message passing paradigm mine new information in the multi-source heterogeneous data, including the following steps: Step 3.1: According to the data stored in the multi-source heterogeneous industrial data unified representation platform and the logical relationship between the data, and combined with the business scene, a knowledge graph is constructed with entity data as nodes and logical relationship as directed edges , wherein and are the node set and edge set of the knowledge graph respectively, and the knowledge graph is split into a topological structure graph focusing on node information and a relationship dependency graph focusing on the relationship between nodes, wherein U and are the node set and edge set in the topological structure graph , and V and are the node set and edge set in the relationship dependency graph . Step 3.2: Input the node features and edge features in the knowledge graph into the relationship-aware graph neural network and the knowledge-aware graph neural network based on the message passing paradigm, to obtain the feature expression of various interaction relationships between different data; Step 3.3: According to , The characteristics of different interaction relationships between node data are expressed, and new information is mined.

8. The method of claim 7, wherein the method is characterized by: The relationship-aware graph neural network is expressed as follows: For a knowledge graph The node features and edge features are input into a relationship-aware graph neural network, i.e., from a topological structure graph And a relationship dependency graph Two local structures learn node embeddings simultaneously, nodes And edges Between nodes Indicate that the relationship type between the two is k The message structure between the two nodes is: ； wherein, denotes the number of neighbors of a node , denotes the number of nodes connected to a node under relation type k, denotes the number of nodes connected to a node under relation type k, denotes the feature embedding of a node is the weight matrix corresponding to the set of nodes; Message propagation pattern definition as comprising temporal context encoding ​ ； wherein, is a LeakyReLU activation function, l is the current layer number of the graph neural network, is t a node at time step t and a corresponding context encoding, denotes l+1 a feature of a node at layer l below, denotes l a feature of a node at layer l above, denotes l a message propagated by a node at layer l below to itself, denotes a number of neighbors of a node, denotes a number of neighbors of a node, is vector addition, denotes l a weight matrix for a node at layer l below when propagating a message to itself, l denotes a weight matrix for a node at layer l above when propagating a message to a node, denotes l an embedding of a message for a node at layer l below when propagating a message to a node, denotes 9. The method of claim 7, wherein the method further comprises: The step 3.3 is specifically as follows: Using knowledge-aware graph neural networks to analyze the topological structure of knowledge graphs Dependency graph Two local structures are learned simultaneously, and the embeddings of the two graphs are learned separately. Then, global embeddings are obtained through aggregation functions, and the semantic knowledge contained in this knowledge graph is extracted. Representing the topology diagram Embedded, Representing a dependency graph Embedding; The update function of the graph embedding is as follows: ； ； wherein, representation graph the embedding of the Ith node in the representation graph the embedding of the Jth node in the representation graph the embedding of the Jth node in the l is the current number of graph neural network layers, denotes a normalization function employed when performing node aggregation, denotes a mean-pooling function of node embeddings.

Images