A method, medium and system for multi-source heterogeneous data integration management based on a digital main line

Abstract

The application provides a multi-source heterogeneous data integration management method, medium and system based on a digital main line, belongs to the technical field of industrial digital main lines, and collects various formats of data and performs standardized processing by using a digital main line platform, realizes cross-domain semantic understanding by using a deep ontology fusion model, improves the semantic understanding efficiency by using adaptive pooling feature dimension reduction and separable attention calculation, optimizes the storage space by using a similarity aggregation deduplication or distributed independent storage strategy according to the data coincidence degree, establishes a digital main line data blood relationship tracing relationship to realize all-life-cycle data link construction, optimizes the semantic understanding parameters by using a fine-tuning or coarse-tuning mode based on a semantic mapping deviation value, constructs a vertical integration mapping from a function model to a performance model and a physical model and a horizontal integration framework of multi-domain comprehensive simulation, and solves the technical problem of insufficient cross-domain model semantic understanding accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial digital main line, and in particular relates to a multi-source heterogeneous data integration management method, medium and system based on a digital main line. BACKGROUND

[0002] In the field of digital manufacturing and product design, enterprises need to integrate heterogeneous data from different systems, including demand models, SYSML models and 3D models and various format data. Traditional data integration methods mainly use ETL tools and data warehouse technology for data aggregation, implement format conversion of different data sources through pre-defined data mapping rules, and use a relational database to store structured data and a file system to manage unstructured data. In the current product lifecycle management system, due to significant semantic differences between data models in different fields, traditional rule-based mapping methods mainly rely on manual definition of mapping rules or simple string matching algorithms, and cannot automatically identify semantic association relationships between cross-domain concepts, which is prone to semantic deviation and mapping errors when processing complex ontology concept mapping. The semantic understanding method in the prior art cannot accurately understand the true meaning of cross-domain concepts when facing different types of data such as demand models, SYSML models and 3D models due to the lack of deep semantic analysis capability, especially when processing large-scale heterogeneous data, which is low in efficiency and low in accuracy, and cannot realize automatic semantic understanding and accurate mapping of cross-domain data, that is, the prior art has the technical problem of insufficient accuracy of cross-domain model semantic understanding. SUMMARY

[0003] Therefore, the present application provides a multi-source heterogeneous data integration management method, medium and system based on a digital main line, which can solve the technical problem of insufficient accuracy of cross-domain model semantic understanding in the prior art.

[0004] The application is implemented in the following manner: a first aspect of the application provides a multi-source heterogeneous data integration management method based on a digital main line to construct a distributed data storage architecture, a NoSQL database and a time series database are combined to establish a data storage layer, and a multi-dimensional index mechanism based on hash fragmentation and time windows is established; demand models, SYSML models and 3D model data are collected through a digital main line platform, intermediate format files are dragged and uploaded and API interfaces are used to obtain data, and the collected raw data is subjected to format standardization processing; different field model data are subjected to semantic analysis by using an ontology mapping algorithm, a field knowledge graph is constructed, and an automatic semantic mapping relationship between cross-domain models is established through a deep ontology fusion model; when the data coincidence degree exceeds a redundancy threshold, a similarity aggregation deduplication strategy is used to optimize the storage space, and if the data coincidence degree is lower than a preset value, a distributed independent storage mode is used to maintain data integrity; after the semantic mapping of the heterogeneous data, the digital main line is constructed, a data blood relationship is established, a data link throughout the product life cycle is formed, and unified management and full-process tracing of multi-source data are realized; different adjustment modes are used to optimize the model mapping parameters according to the semantic mapping deviation value; based on the digital main line architecture, vertical integration and mapping of performance models and physical models by functional models are realized, a horizontal integration framework of multi-domain comprehensive simulation is constructed, and a full-process model collaborative management system is established.

[0005] The time series database refers to a database system for processing time series data, and is used to optimize the storage and query performance of time dimension data.

[0006] The structure of the deep ontology fusion model is that an input layer receives multi-source ontology vectors, a feature dimension reduction optimization algorithm based on adaptive pooling is used to dynamically adjust the pooling window size to realize feature compression, a middle layer uses a calculation efficiency optimization algorithm based on separable attention to decompose multi-head attention into spatial dimension and channel dimension for separate processing, and an output layer generates a unified ontology representation vector after fusion.

[0007] The feature dimension reduction optimization algorithm based on adaptive pooling dynamically adjusts the pooling window size to adapt to different feature distributions, retains key semantic information, realizes feature dimension compression, reduces the computational burden of subsequent processing, and the pooling window size is obtained through a window size function.

[0008] The calculation efficiency optimization algorithm based on separable attention decomposes the traditional multi-head attention mechanism into two independent calculation processes of spatial attention and channel attention, the spatial attention focuses on the positional relationship between concepts, the channel attention focuses on the importance of different feature dimensions, the separation calculation significantly reduces the computational complexity and memory occupation, and the attention weight is obtained through an attention weight function.

[0009] The data coincidence degree refers to a proportion value of same or similar data contents in different data sources.

[0010] The similarity aggregation deduplication strategy refers to calculating the similarity values between data, and combining and storing highly similar data to save storage space.

[0011] The semantic mapping deviation value refers to a difference degree value between concept representation vectors before and after semantic mapping by a distance measurement method in a vector space, when the semantic mapping deviation value is less than a sensitivity threshold value, a fine-tuning mode is used for model mapping parameter optimization, and when the semantic mapping deviation value is greater than a preset value, a coarse adjustment mode is used for rapid mapping adjustment.

[0012] The data bloodline tracing relationship refers to a pedigree relationship of recorded data from generation, transformation to application, and realizes complete tracking of data sources and flow directions.

[0013] The longitudinal integrated mapping refers to establishing a hierarchical mapping relationship from a functional model to a performance model and then to a physical model. The horizontal integrated framework refers to a cross-domain data collaborative processing architecture.

[0014] The second aspect of the application provides a computer readable storage medium, the computer readable storage medium has program instructions stored therein, the program instructions are used for executing the above-mentioned multi-source heterogeneous data integration management method based on a digital main line when running in a computer.

[0015] The third aspect of the application provides a multi-source heterogeneous data integration management system based on a digital main line, comprising the above-mentioned computer readable storage medium, the system is any one of a computer, a server and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0016] The application establishes an adaptive cross-domain concept understanding mechanism by constructing a deep ontology fusion model, realizes automatic semantic understanding and accurate mapping of multi-source heterogeneous data, and significantly improves the accuracy of cross-domain model semantic understanding. The application adopts a feature dimension reduction optimization algorithm based on adaptive pooling and a calculation efficiency optimization algorithm based on separable attention, solves the problems of high computational complexity and insufficient understanding accuracy of traditional semantic understanding methods, and through dynamic adjustment of the pooling window size and separation attention calculation, the efficiency and accuracy of semantic understanding are significantly improved while the integrity of semantic information is maintained. Through the hierarchical processing mechanism of the deep ontology fusion model, the input layer receives multi-source ontology vectors, the middle layer realizes feature compression and attention calculation, and the output layer generates a unified ontology representation vector after fusion, an accurate semantic mapping relationship from the requirement model, the SYSML model to the 3D model is established, and accurate understanding and automatic mapping of cross-domain concepts are realized, thereby effectively solving the technical problem of insufficient accuracy of cross-domain model semantic understanding in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of the method of the application.

[0018] Figure 2 is a trend chart of the mapping relationship accuracy rate changing with the number of iterations in the embodiment. DETAILED DESCRIPTION

[0019] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0020] As shown in Figure 1 , it is a flowchart of a multi-source heterogeneous data integration management method based on a digital main line provided by the first aspect of the application, and the method comprises the following steps:

[0021] S01, a distributed data storage architecture is constructed, a NoSQL database and a time series database combination method is used to establish a data storage layer, and a multi-dimensional index mechanism based on hash sharding and time window is established;

[0022] S02, requirement model, SYSML model and 3D model data are collected through a digital main line platform, data is acquired by using an intermediate format file drag-and-drop upload and an API interface method, and the collected raw data is subjected to format standardization processing;

[0023] S03, the ontology mapping algorithm is used to perform semantic analysis on different domain model data, a domain knowledge graph is constructed, and an automatic semantic mapping relationship between cross-domain models is established through a deep ontology fusion model;

[0024] S04. When the data overlap exceeds the redundancy threshold of 85%, a similarity aggregation deduplication strategy is adopted to optimize storage space. If the data overlap is less than 15%, a distributed independent storage method is adopted to maintain data integrity.

[0025] S05. Construct a digital mainline for the heterogeneous data after semantic mapping, establish a data lineage traceability relationship, form a data link that runs through the entire product life cycle, and realize unified management and full-process traceability of multi-source data.

[0026] S06. If the semantic mapping deviation is less than the sensitivity threshold of 0.05, the fine-tuning mode is used to optimize the model mapping parameters. When the semantic mapping deviation is greater than 0.3, the coarse-tuning mode is used to achieve rapid mapping adjustment.

[0027] S07. Based on the digital mainline architecture, realize the vertical integration mapping of the functional model to the performance model and the physical model, and at the same time build a horizontal integration framework for multi-domain comprehensive simulation and establish a full-process model collaborative management system.

[0028] The distributed data storage architecture refers to a data storage system deployed across multiple nodes, achieving efficient storage and access to massive amounts of heterogeneous data through data sharding and replication mechanisms. The NoSQL database refers to a database system employing a non-relational data model for storing structured and semi-structured heterogeneous data. The time-series database refers to a database system for processing time-series data, optimizing the storage and query performance of time-dimensional data. Hash sharding refers to a data sharding method that calculates hash values ​​based on data characteristics and distributes data evenly across different storage nodes. The time window refers to a mechanism for partitioning data according to the time dimension, facilitating rapid retrieval and management of time-series data. The multi-dimensional indexing mechanism refers to establishing a multi-index structure encompassing time, space, and attribute dimensions to improve data query and retrieval efficiency.

[0029] The digital mainline platform refers to a core management platform for integrating multi-source heterogeneous data and establishing data blood relationship. The requirement model refers to a structured data model describing system functional requirements and performance requirements. The SYSML model refers to a system architecture description model constructed based on system modeling language. The 3D model data refers to digital representation data of three-dimensional geometric shapes and spatial structures. The intermediate format file refers to a standard exchange format file converted from original data of different formats. The API interface mode refers to a communication mode for data transmission and interaction through an application programming interface. The format standardization processing refers to a processing process of converting collected data of various formats into a unified data format. The ontology mapping algorithm refers to an algorithm for establishing semantic correspondence between concepts in different fields based on ontology theory. The semantic analysis refers to a processing process of semantically understanding and structurally analyzing data content. The domain knowledge graph refers to a knowledge representation method representing concepts, entities and their mutual relationships in a domain in the form of a graph structure.

[0030] The structure of the deep ontology fusion model is that the input layer receives multi-source ontology vectors, feature dimension reduction optimization algorithms based on adaptive pooling are used to dynamically adjust the size of the pooling window to realize feature compression, the intermediate layer uses a calculation efficiency optimization algorithm based on separable attention to decompose multi-head attention into spatial dimension and channel dimension for processing respectively, and the output layer generates a unified ontology representation vector after fusion.

[0031] The steps of establishing the training data set of the deep ontology fusion model include collecting standard ontology libraries in different fields as basic data sources, vectorizing ontology concepts to form training samples, establishing artificially annotated ontology mapping relationships as supervision labels, and expanding the scale of training samples through data augmentation techniques.

[0032] The steps of training the deep ontology fusion model include initializing model parameters, updating parameters using a batch gradient descent algorithm, evaluating model performance using a cross-validation method, and optimizing model parameters through loss function minimization until convergence.

[0033] The feature dimension reduction optimization algorithm based on adaptive pooling dynamically adjusts the size of the pooling window to adapt to different feature distributions, while preserving key semantic information, realizes feature dimension compression, and reduces the computational burden of subsequent processing. The size of the pooling window is obtained through a window size function.

[0034] The calculation efficiency optimization algorithm based on separable attention decomposes the traditional multi-head attention mechanism into two independent calculation processes of spatial attention and channel attention. Spatial attention focuses on the positional relationship between concepts, and channel attention focuses on the importance of different feature dimensions. Through separate calculation, the computational complexity and memory occupation are significantly reduced. The attention weight is obtained through an attention weight function.

[0035] The data coincidence degree refers to the proportion value of the same or similar data content in different data sources. The redundancy threshold refers to the critical proportion value for judging whether there is redundancy in data, and the value is 85%. The similarity aggregation deduplication strategy refers to calculating the similarity value between data, and merging and storing highly similar data to save storage space. The distributed independent storage mode refers to storing data with low similarity in different nodes to maintain the independence and integrity of data.

[0036] The semantic mapping refers to the corresponding relationship between concepts in different fields established by the deep ontology fusion model in step S03. The data bloodline tracing relationship refers to the pedigree relationship of recording data from generation, transformation to application, realizing complete tracking of data source and flow. The semantic mapping deviation value refers to the difference value between the concept representation vectors before and after semantic mapping calculated by the distance measurement method in the vector space. The sensitivity threshold refers to the critical value for judging whether the semantic mapping deviation value needs to be finely adjusted, and the value is 0.05.

[0037] The fine-tuning mode refers to an optimization method using small step size parameter adjustment to accurately optimize the model mapping relationship. The coarse adjustment mode refers to an optimization method using large step size parameter adjustment to realize fast convergence of the model mapping relationship. The functional model refers to an abstract model describing the functional behavior and logical relationship of the system. The performance model refers to a mathematical model describing the performance index and constraint condition of the system.

[0038] The physical model refers to an entity model describing the physical structure and geometric characteristics of the system. The longitudinal integrated mapping refers to establishing a hierarchical mapping relationship from the functional model to the performance model and then to the physical model. The multi-domain comprehensive simulation refers to a collaborative simulation analysis method across multiple professional fields. The horizontal integration framework refers to a data collaborative processing architecture across different professional fields.

[0039] The window size function is used to dynamically calculate the optimal size of the pooling window according to the feature distribution characteristics. The input includes the feature vector dimension number, the feature distribution variance value, and the target compression ratio. The output is the pixel value of the side length of the pooling window. The feature vector dimension number is derived from the ontology vectorization representation process in step S03. The feature distribution variance value is derived from the statistical analysis of the semantic analysis process in step S03. The target compression ratio is preset according to the storage capacity requirement. The pixel value of the side length of the pooling window is used in the feature dimension reduction optimization algorithm based on adaptive pooling in step S03. The window size function is expressed as follows: wherein is the normalized window size, is the normalized feature dimension, is the normalized distribution variance, for normalizing the target compression ratio, for normalizing the reference parameter.

[0040] The attention weight function is used to calculate the attention distribution weight of the spatial dimension and the channel dimension, the input includes the spatial position index value and the channel feature intensity value, and the output is the attention weight coefficient, wherein the spatial position index value is derived from the spatial structure information of the domain knowledge graph in step S03, the channel feature intensity value is derived from the feature intensity analysis of the ontology vector in step S03, and the attention weight coefficient is used in the calculation efficiency optimization algorithm based on separable attention in step S03. The attention weight function is expressed as follows: wherein for normalizing the attention weight, for normalizing the spatial position, for normalizing the channel intensity, for normalizing the total number of spaces, for normalizing the maximum intensity value.

[0041] The specific embodiments of the above steps are described in detail below.

[0042] The specific embodiment of step S01 is to build a distributed data storage architecture, and the purpose of this step is to establish an infrastructure that can efficiently store and access massive heterogeneous data. First, MongoDB is selected as a NoSQL database system for storing structured and semi-structured heterogeneous data, which adopts a document storage model to flexibly adapt to the storage needs of different data formats. At the same time, InfluxDB is deployed as a time series database to optimize the storage and query performance of time series data, and its columnar storage structure and time indexing mechanism can significantly improve the processing efficiency of time series data. A hash sharding mechanism is established, which calculates the data feature value through the MD5 hash algorithm, and distributes the data evenly into 3 to 5 storage nodes to ensure load balancing and system scalability. A time window partitioning mechanism is built, which partitions the data according to the day, week, and month three-level time dimension, and the day-level window is used for real-time data access, the week-level window is used for recent data query, and the month-level window is used for historical data archiving. A multi-dimensional indexing mechanism is established, including time dimension index, spatial dimension index, and attribute dimension index, the time dimension index adopts B+ tree structure, the spatial dimension index adopts R tree algorithm, and the attribute dimension index adopts inverted index technology, which comprehensively improves the data query and retrieval efficiency to achieve millisecond-level response.

[0043] The specific implementation of step S02 is to collect multi-source data through the digital mainline platform. The purpose of this step is to obtain and standardize various heterogeneous data sources. A unified data collection interface is established to support two data acquisition methods: drag-and-drop upload and API interface. Drag-and-drop upload supports intermediate format files such as XML, JSON, and STL, with a file size limit of 100MB. The API interface uses RESTful architecture design and supports POST and GET request methods. The requirement model data is collected, including functional requirement description, performance index parameter, and constraint condition definition, which is stored in XML format. The SYSML model data is collected, including system architecture diagram, activity diagram, and sequence diagram model elements, which is represented in UML / SysML standard. The 3D model data is collected, supporting STL, OBJ, and PLY three-dimensional model formats, containing geometric coordinates, topological relationships, and material attribute information. Format standardization processing is implemented to convert the collected multi-format data into a unified JSON format. A data dictionary mapping table is established to define field names, data types, and constraint rules, ensuring data format consistency and interoperability.

[0044] The specific implementation of step S03 is to use ontology mapping algorithm for semantic analysis and knowledge graph construction. The purpose of this step is to realize semantic understanding and association between different domain data. An ontology mapping algorithm based on word vector is used, and the Word2Vec model is used to convert domain concepts into 300-dimensional vector representation. The semantic similarity between concepts is calculated by cosine similarity. A domain knowledge graph is constructed, which represents entities, attributes, and relationships in the form of RDF triples. It is stored and managed using the Neo4j graph database and supports SPARQL query language for semantic retrieval. A deep ontology fusion model is deployed, which uses an encoder-decoder architecture. The encoder part uses a multi-layer perceptron to extract ontology features, and the decoder part uses an attention mechanism to generate a unified ontology representation after fusion. Automatic semantic mapping between cross-domain models is realized by calculating the Euclidean distance between different domain concept vectors. When the distance value is less than 0.3, a semantic mapping relationship is established, and when it is greater than 0.7, it is considered as unrelated concepts. A semantic mapping relationship table is established to record source concepts, target concepts, mapping strength, and confidence, etc., providing a semantic basis for subsequent data integration.

[0045] The specific implementation of step S04 is to implement a data deduplication optimization strategy, and the purpose of this step is to reduce storage redundancy and maintain data integrity. Calculate the data coincidence degree, compare the similarity of data content using the Jaccard similarity algorithm, and calculate the similarity value by the ratio of the intersection and union of sets. Set the redundancy threshold to 85%, and when the data coincidence degree exceeds the threshold, it is considered that there is significant redundancy that needs to be merged. Use a similarity aggregation deduplication strategy, use a hierarchical clustering algorithm to group data with high similarity, and select the most complete data in each group as a representative, and other data establish a reference relationship rather than physical storage. Set the independent storage threshold to 15%, and when the data coincidence degree is lower than the threshold, use a distributed independent storage method to store the data in different nodes, and maintain the independence and integrity of the data. Establish a data version management mechanism, assign a unique identifier and version number to each data entity, record the creation time, modification history and dependency relationship of the data, and support data rollback and incremental update operations.

[0046] The specific implementation of step S05 is to build a digital main line and data bloodline tracing system, and the purpose of this step is to establish a data management link throughout the product life cycle. Build a digital main line architecture, use an event-driven message queue mechanism, use Apache Kafka as a message middleware to realize real-time capture and propagation of data change events. Build data bloodline tracing relationship, record the source of data generation, transformation process and application direction, use directed acyclic graph structure to represent data flow relationship. Establish a data pedigree management system, assign a globally unique identifier to each data object, record the metadata information, processing history and dependency relationship of the data. Realize the unified management of multi-source data, establish a unified data view layer, shield the heterogeneity of the underlying storage, and provide a consistent data access interface for the upper layer application. Build a full-process tracing mechanism, support forward and backward tracing modes, forward tracing from data source to final application, backward tracing from application result to original data source, and tracing depth supports more than 10 levels of data transmission link.

[0047] The specific implementation of step S06 is to implement semantic mapping precision optimization, and the purpose of this step is to improve the accuracy and processing efficiency of semantic mapping. The semantic mapping deviation value is calculated, and the vector space distance measurement method is used to calculate the difference between the concept vectors before and after mapping using Manhattan distance or Euclidean distance. The sensitivity threshold is set to 0.05, and when the semantic mapping deviation value is less than the threshold, a fine-tuning mode is used for optimization. The fine-tuning mode uses a gradient descent algorithm, sets the learning rate to 0.001, and performs small-step parameter adjustment, with each iteration updating the amplitude to within 1% to ensure accurate tuning of the mapping relationship. The coarse adjustment threshold is set to 0.3, and when the semantic mapping deviation value is greater than the threshold, a coarse adjustment mode is used for rapid adjustment. The coarse adjustment mode uses an adaptive learning rate algorithm, with an initial learning rate of 0.1, and dynamically adjusts the learning rate according to the loss function to achieve rapid convergence of the mapping relationship. A mapping quality evaluation mechanism is established, using accuracy, recall rate and F1 score as evaluation indicators, and when the F1 score is less than 0.8, a retraining mechanism is triggered to ensure the reliability of semantic mapping.

[0048] The specific implementation of step S07 is to build a multi-level integrated mapping framework, and the purpose of this step is to realize the collaborative integration and simulation analysis of cross-domain models. A vertical integrated mapping relationship is established to build a hierarchical mapping system from a functional model to a performance model and then to a physical model, and a model conversion rule engine is used to realize automatic mapping conversion. The functional model uses state charts and activity diagrams to represent system behavior logic, the performance model uses mathematical equations and constraints to describe system performance indicators, and the physical model uses CAD geometric models to represent entity structure characteristics. A horizontal integration framework is built to establish a collaborative processing architecture across multiple engineering fields such as mechanics, electronics and software, and federal simulation technology is used to realize multi-domain comprehensive simulation. A distributed simulation engine is deployed to support parallel computing and load balancing, with a single simulation supporting the processing of more than 1 million model elements. A model synchronization mechanism is established to ensure the time consistency of multi-domain models using a time stepping algorithm with a time step precision of microseconds. A full-process model collaborative management system is built to realize model version control, change propagation and consistency maintenance, support multi-user concurrent access and collaborative editing, and support more than 100 users online at the same time.

[0049] It needs to be further explained that the detailed structure of the deep ontology fusion model contains three main levels. The input layer receives ontology vector representations from different domains, each ontology concept is encoded as a 300-dimensional dense vector, and the input layer sets multiple parallel channels to process multi-source ontology data simultaneously. The feature processing layer adopts a feature dimension reduction optimization algorithm based on adaptive pooling, which dynamically calculates the optimal size of the pooling window through a window size function. The input parameters include the feature vector dimension, the feature distribution variance value and the target compression ratio, and the output is the pixel value of the pooling window edge length. The adaptive pooling mechanism dynamically adjusts the window size according to different feature distributions, while preserving key semantic information, and reduces the feature dimension from 300 to 128, reducing the computational burden of subsequent processing. The intermediate layer adopts a computational efficiency optimization algorithm based on separable attention, which decomposes the traditional multi-head attention mechanism into two independent calculation processes: spatial attention and channel attention. Spatial attention focuses on the positional relationship and semantic distance between concepts, and calculates the spatial dimension attention distribution weight through an attention weight function. The input parameter is the spatial position index value, and the output is the normalized spatial attention weight. Channel attention focuses on the importance of different feature dimensions, and the input parameter is the channel feature intensity value, and the output is the normalized channel attention weight. Separated calculation significantly reduces the computational complexity and memory occupancy, and the computational efficiency is improved by about 3 times compared with the traditional attention mechanism. The output layer generates a unified ontology representation vector after fusion, and the vector dimension is fixed at 256. The final fusion representation is generated through a fully connected layer and an activation function.

[0050] The detailed steps of establishing the deep ontology fusion model training dataset include four stages of data collection, preprocessing, labeling and enhancement. In the data collection stage, basic data sources are obtained from standard ontology libraries such as WordNet, ConceptNet, YAGO, covering ontology knowledge in general, engineering and industry fields. In the preprocessing stage, the collected ontology concepts are cleaned and standardized to remove duplicate concepts and noise data and unify the concept naming specification. In the vectorization representation stage, the pre-trained Word2Vec model is used to convert ontology concepts into vector form to form a training sample set. In the labeling stage, artificial labeled ontology mapping relationships are established as supervision labels, and labeling personnel include domain experts and knowledge engineers to ensure labeling quality and consistency. In the data enhancement stage, techniques such as synonym replacement, concept expansion and relationship reasoning are used to expand the training sample size, expanding the original sample number from 100,000 to 500,000. In the model training stage, the model parameters are initialized, and the Xavier initialization method is used to set the weight value. The batch gradient descent algorithm is used for parameter update, with a batch size of 64 and a learning rate of 0.001. The 5-fold cross-validation method is used to evaluate the model performance, and the model parameters are optimized by minimizing the loss function until convergence, with a training round number usually between 100 and 200 rounds.

[0051] It should be noted that the key technical idea of the present application is analyzed as follows. The first key technical idea is the separable attention mechanism of the deep ontology fusion model. The traditional multi-head attention mechanism needs to process the feature relationship of spatial dimension and channel dimension simultaneously, and the computational complexity increases in a square level. The present application decomposes the multi-head attention into two independent processes of spatial attention and channel attention. The spatial attention focuses on the positional relationship and semantic distance between concepts, and the channel attention focuses on the importance weight of different feature dimensions. This separation design significantly reduces the computational complexity, from to , greatly improves the processing efficiency under the premise of maintaining the accuracy of semantic understanding, making large-scale ontology fusion tasks possible.

[0052] The second key technical idea is the full life cycle data bloodline tracing mechanism based on the digital main line. Traditional data management methods often use static data warehouse mode, lacking dynamic tracking ability for data change process. The digital main line architecture built by the present application uses an event-driven message propagation mechanism to capture the generation, transformation and application process of data in real time, and establishes a complete data blood relationship graph. This dynamic tracing capability not only supports the rapid positioning and repair of data quality problems, but also provides a reliable basis for data value analysis and impact assessment, significantly improving the intelligent level of data management.

[0053] The third key technical idea is the adaptive data storage optimization strategy. Existing data storage schemes usually use fixed storage modes and cannot dynamically adjust the storage strategy according to data characteristics. The present application calculates the data coincidence degree, and when the coincidence degree exceeds 85%, it uses a similarity aggregation deduplication strategy to save storage space, and when the coincidence degree is less than 15%, it uses distributed independent storage to maintain data integrity. This adaptive strategy not only avoids information loss caused by excessive deduplication, but also prevents resource waste caused by redundant storage, achieving balanced optimization of storage efficiency and data integrity.

[0054] The synergistic effect of these key technical ideas produces significant comprehensive advantages. The deep ontology fusion model provides efficient and accurate semantic understanding capability, laying the foundation for intelligent integration of heterogeneous data. The digital main line architecture ensures the complete tracing of data flow process, providing support for data quality management and value mining. The adaptive storage strategy optimizes the underlying resource configuration, improving the overall performance of the system. The three work together to form a complete multi-source heterogeneous data integration and management solution. Compared with traditional separate data management methods, it has achieved significant improvement in semantic understanding accuracy, data tracing integrity and storage efficiency, providing strong technical support for digital design and intelligent manufacturing of complex products.

[0055] It should be noted that the present application also solves the following technical problems: firstly, the problem of lack of full life cycle traceability management of multi-source heterogeneous data, in the prior art, when facing data with extensive sources and diversified formats, the traditional data management method lacks an effective data blood relationship tracing mechanism, cannot establish a complete link relationship of data from generation to application, leads to the process of data transfer in different stages is not transparent, the data quality is difficult to control, the data source is not clear, the present application establishes a digital main line architecture to build a data blood relationship tracing relationship, records the complete pedigree of data from generation, transformation to application, realizes the transparent management of the data flow track in the product full life cycle, and establishes a hierarchical tracing link from a functional model to a performance model and then to a physical model through vertical integration mapping, and realizes the collaborative tracing of cross-professional data through a horizontal integration framework, so that the controllability of data quality and the explicitness of data source are ensured.

[0056] Secondly, the problem of heterogeneous data storage efficiency optimization, the traditional relational database storage method has the problems of poor expansibility and low query efficiency when processing massive multi-source heterogeneous data, especially when processing time series data and unstructured data, the performance is seriously reduced, the present application builds a distributed data storage architecture, adopts a combination of NoSQL database and time series database, combines a multi-dimensional index mechanism of hash sharding and time window, realizes efficient storage and fast access of heterogeneous data, and dynamically selects a similarity aggregation deduplication strategy or a distributed independent storage mode according to the data coincidence degree, so that the data integrity is ensured and the storage space utilization rate is optimized.

[0057] In addition, the present application also solves the problem of collaboration of data integration management, the prior art lacks a perfect cross-system data collaboration mechanism, cannot effectively support multi-domain comprehensive simulation and collaborative design requirements, leads to independent data processing in different professional fields, and lacks effective collaborative management, the present application builds vertical integration mapping relationships from a functional model to a performance model and a physical model, and simultaneously establishes a horizontal integration framework of multi-domain comprehensive simulation, realizes unified management and collaborative processing of data in different professional fields, supports the multi-disciplinary collaborative design and simulation analysis requirements of complex products, and ensures the collaboration and consistency of data integration management.

[0058] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, and the program instructions are used for executing the above-mentioned multi-source heterogeneous data integration management method based on a digital main line when running in a computer.

[0059] A third aspect of the present invention provides a multi-source heterogeneous data integration management system based on a digital mainline, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0060] Specifically, the principle of this invention is as follows: This invention establishes semantic understanding relationships between multi-source heterogeneous data through a deep ontology fusion model. This model uses an input layer to receive multi-source ontology vectors, and an intermediate layer dynamically adjusts the pooling window size using an adaptive pooling-based feature dimensionality reduction optimization algorithm to achieve feature compression. Simultaneously, a computational efficiency optimization algorithm based on separable attention decomposes multi-head attention into spatial and channel dimensions for separate processing. The output layer generates a fused unified ontology representation vector. This layered processing mechanism enables efficient cross-domain concept understanding while preserving key semantic information. The adaptive pooling algorithm dynamically calculates the optimal pooling window size based on feature distribution characteristics. The window size function comprehensively considers feature vector dimension, distribution variance, and target compression ratio, ensuring that key semantic information is not lost during feature dimensionality reduction. The separable attention mechanism focuses on the positional relationships between concepts through spatial attention and on the importance of feature dimensions through channel attention, significantly reducing computational complexity while improving semantic understanding accuracy. The domain knowledge graph established by this invention performs semantic parsing on model data from different domains through an ontology mapping algorithm, thereby realizing the establishment of automated semantic mapping relationships between cross-domain models. When the semantic mapping deviation value exceeds the sensitivity threshold, it automatically triggers fine-tuning or coarse-tuning mode to optimize parameters, ensuring the continuous accuracy of cross-domain semantic understanding, thus establishing a complete semantic understanding system for multi-source heterogeneous data.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0062] The specific implementation of step S01 involves constructing a distributed data storage architecture, which includes the calculation process of hash sharding and multidimensional indexes. The hash sharding calculation uses the MD5 algorithm, and its sharding index calculation formula is expressed as follows:

[0063] ;

[0064] In the formula, This is the normalized hash index value, ranging from 0 to 1; Data keywords; This represents the total number of storage nodes, with a default value of 5. This is the MD5 hash function. The formula for calculating the time window partition index is as follows:

[0065] ;

[0066] In the formula, This is the normalized time index value; This is the current timestamp, in seconds. The base timestamp is in seconds. This represents the size of the time window, in seconds; a daily window is 86,400 seconds.

[0067] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0068] The specific implementation of step S03 involves semantic parsing using an ontology mapping algorithm. This step includes vector similarity calculation and an ontology fusion model. The word vector similarity calculation uses the cosine similarity algorithm, and its calculation formula is as follows:

[0069] ;

[0070] In the formula, The normalized cosine similarity value ranges from -1 to 1. and The vector representations of the two concepts are both 300 in dimension. It is the dot product of vectors; and Let be the Euclidean norm of the vector. The formula for calculating the Euclidean distance established by the semantic mapping relationship is as follows:

[0071] ;

[0072] In the formula, For the normalized Euclidean distance value; and For vectors and The One component; The vector dimension is 300 by default. The maximum possible distance value is, empirically, [value]. The window size function is used for feature dimensionality reduction optimization in adaptive pooling, and its calculation formula is expressed as follows:

[0073] ;

[0074] In the formula, To normalize the window size; The normalized feature dimension is derived from the ontology vectorization representation process in step S03; The normalized distribution variance is derived from the statistical analysis of the semantic parsing process in step S03; To normalize the target compression ratio, a preset value of 0.5 is used based on storage capacity requirements. This is the normalized baseline parameter, which defaults to 1. The attention weight function is used to calculate the weight allocation for separable attention, and its calculation formula is as follows:

[0075] ;

[0076] In the formula, To normalize attention weights; The normalized spatial location is derived from the spatial structure information of the domain knowledge graph in step S03; The normalized channel strength is derived from the feature strength analysis of the ontology vector in step S03; This is the normalized total space size, which defaults to 100. The normalized maximum intensity value is 1, based on empirical values.

[0077] The specific implementation of step S04 involves implementing a data deduplication optimization strategy, which includes calculating data overlap. Jaccard similarity calculation is used to quantify the degree of data overlap, and its formula is as follows:

[0078] ;

[0079] In the formula, The value is the Jaccard similarity score, ranging from 0 to 1. and There are two data sets; The number of elements in the intersection of the sets; The number of elements in the union of sets. The data overlap determination function uses a threshold comparison mechanism, and its judgment logic is expressed as follows:

[0080] ;

[0081] In the formula, Select an identifier for the storage strategy: 1 indicates that an aggregated deduplication strategy is used, 0 indicates that an independent storage method is used, and -1 indicates that further analysis is required. The redundancy threshold is set to 0.85. The threshold for independent storage is 0.15.

[0082] The specific implementation method of step S05 is the same as described above, and will not be repeated in detail here.

[0083] The specific implementation of step S06 involves optimizing the semantic mapping accuracy, which includes calculating the semantic mapping deviation value. The semantic mapping deviation value uses a vector space distance metric, and its calculation formula is as follows:

[0084] ;

[0085] wherein, is the normalized semantic mapping deviation value; is the concept vector representation before mapping; is the concept vector representation after mapping; is the Euclidean norm of the vector difference; is the Euclidean norm of the original vector. The optimization strategy selection function adopts a threshold judgment mechanism, which is expressed as follows:

[0086] ;

[0087] wherein, is the optimization mode identifier, 1 represents the fine tuning mode, 2 represents the coarse tuning mode, and 0 represents the standard optimization mode; is the sensitivity threshold, taking a value of 0.05; is the coarse tuning threshold, taking a value of 0.3.

[0088] The specific implementation of step S07 is the same as the foregoing, and will not be described in detail here.

[0089] It needs to be explained that the principle of the hash shard index calculation formula is based on the data uniform distribution theory. The data key is converted into a fixed length hash value by the MD5 hash function, and then the modulo operation is used to realize the uniform distribution of data among multiple storage nodes. The effect of this formula is to eliminate the skewness of data distribution and avoid the problem of hot data concentrated in a single node. Compared with the traditional round robin distribution method, the hash shard can realize more intelligent load balancing according to the data content characteristics, significantly improving the access efficiency and expansion performance of the distributed storage system.

[0090] The time window partition index calculation formula adopts the time modulo operation principle to divide the continuous time axis into fixed size time periods, and each time period corresponds to a storage partition.

[0091] .

[0092] The effect of this formula is to optimize the storage organization method of time series data, so that data of similar time are stored in clusters, greatly improving the performance of time range queries. Compared with the traditional linear time storage method, the time window partition can reduce the query complexity from linear to constant level.

[0093] The principle of the cosine similarity calculation formula is based on vector geometry. The cosine value of the angle between two vectors is calculated to quantify the semantic similarity, avoiding the influence of vector length on similarity judgment.

[0094] .

[0095] The effect of this formula is to provide a standardized semantic similarity measurement method, which can accurately identify the semantic association relationship between different domain concepts. Compared with the traditional string matching method, vector similarity calculation can capture deep semantic information, significantly improving the accuracy of cross-domain semantic mapping.

[0096] The Euclidean distance calculation formula uses the principle of geometric distance, which calculates the straight-line distance between two points in a high-dimensional vector space to evaluate the concept difference.

[0097] .

[0098] The effect of this formula is to provide an accurate semantic difference quantization index, which provides a reliable basis for the establishment of semantic mapping relationship. Compared with the traditional heuristic judgment method, geometric distance measurement has a stronger mathematical theoretical basis and calculation stability.

[0099] The principle of the window size function is based on adaptive optimization theory, which considers multiple factors such as feature dimension, distribution variance, and compression target, and dynamically calculates the optimal pooling window size. The feature dimension term in this function reflects the data complexity, the distribution variance term reflects the data dispersion degree, the target compression ratio term controls the dimensionality reduction strength, and the benchmark parameter term provides a standardized benchmark. The effect of this formula is to realize intelligent control of feature dimensionality reduction, which can automatically adjust the processing strategy according to different data characteristics. Compared with the traditional method of fixed window size, the adaptive window mechanism maximizes the dimensionality reduction efficiency while maintaining key information, significantly improving the processing performance of the ontology fusion model.

[0100] The principle of the attention weight function is based on the attention mechanism theory, which calculates the attention weight by multiplying the spatial position and channel intensity , realizing dynamic attention to important features. The effect of this formula is to optimize the feature weight allocation strategy, which can automatically identify and strengthen key semantic information. Compared with the traditional uniform weight allocation method, the dynamic attention mechanism significantly improves the accuracy and processing efficiency of semantic understanding.

[0101] The principle of the Jaccard similarity calculation formula is based on set theory, which quantifies the overlap between two data sets by the ratio of the intersection and the union of the sets, providing a standardized similarity measurement method.

[0102] .

[0103] The effect of the formula is to accurately identify the data redundancy, and to provide a reliable decision basis for the storage optimization strategy. Compared with the traditional simple repetition detection method, the Jaccard similarity can identify the complex redundancy of partial overlap, and significantly improve the storage space utilization efficiency. The storage strategy selection function adopts a segmented threshold judgment principle, and selects the corresponding storage strategy according to different intervals of the similarity value. The effect of the function is to realize the intelligent selection of the storage strategy, and to achieve the optimal balance between redundancy elimination and data integrity. Compared with the traditional method of single storage mode, the adaptive strategy selection can dynamically optimize the storage scheme according to the data characteristics.

[0104] The principle of the semantic mapping deviation value calculation formula is based on the vector variation measurement theory, and the relative change degree of the vectors before and after mapping is calculated to evaluate the mapping quality. The formula adopts the relative error calculation method, which avoids the problem that the absolute error is affected by the vector magnitude.

[0105] .

[0106] The effect of the formula is to provide an accurate mapping quality evaluation index, and to provide a quantitative basis for the selection of optimization strategy. Compared with the traditional qualitative evaluation method, the numerical deviation calculation has stronger objectivity and comparability. The optimization strategy selection function adopts a multi-threshold segmented judgment principle, and selects different intensity optimization strategies according to the size of the deviation value. The effect of the function is to realize the fine control of the optimization process, and to dynamically adjust the optimization intensity according to the mapping quality. Compared with the traditional method of fixed optimization strategy, the adaptive optimization selection can maximize the processing efficiency while ensuring the mapping precision, and significantly improves the overall performance of the semantic mapping system.

[0107] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application:

[0108] A technical team undertook a full life cycle data management task of a complex industrial equipment, which involves multiple professional fields such as machinery, electronics and software, and generates a large amount of heterogeneous data that needs to be uniformly managed. The technical team decided to adopt a multi-source heterogeneous data integration management method based on digital mainline to solve this technical problem.

[0109] The technical team first built a distributed data storage architecture, deployed a NoSQL database cluster with 12 storage nodes, and configured a time series database to process sensor data generated during device operation. The storage architecture uses a hash sharding mechanism to evenly distribute data to different nodes based on the MD5 value of the data, with each shard's data volume controlled within 2.5GB. The time window mechanism partitions time series data at the hourly level, with each time window containing 3600 seconds of data records. The multi-dimensional indexing mechanism establishes a composite index structure with time dimension, spatial dimension, and attribute dimension, with an index depth of 5 layers to support fast data retrieval.

[0110] During the data collection phase, the technical team collected 8750 demand model data, 3200 SYSML model data, and 1850 3D model files through the digital mainline platform. The demand model data includes 2350 functional requirements, 4200 performance requirements, and 2200 interface requirements. SYSML models are stored in standard XMI format, with an average file size of 125KB. 3D model data is mainly in STEP format, with individual model file sizes ranging from 15 to 45MB. During the collection process, 67% of the data was obtained through intermediate format file drag-and-drop upload, and 33% was automatically collected through RESTful API interfaces. Format standardization processing uniformly converts raw data into a mixed JSON and XML format, with a data conversion accuracy of 98.7%.

[0111] During semantic analysis, the technical team used ontology mapping algorithms to process different domain model data. The mechanical domain established a domain knowledge graph containing 1850 concept nodes, the electronic domain established a knowledge graph containing 2100 concept nodes, and the software domain established a knowledge graph containing 1650 concept nodes. The input layer of the deep ontology fusion model receives a multi-source ontology vector with a dimension of 512, and the feature dimension is compressed to 256 dimensions based on the adaptive pooling feature dimension optimization algorithm, with a compression rate of 50%. The intermediate layer uses a computationally efficient optimization algorithm based on separable attention, with 8 spatial attention heads and 16 channel attention heads, reducing the computational complexity by 42% compared to traditional multi-head attention mechanisms. The output layer generates a 128-dimensional unified ontology representation vector.

[0112] During the training data set establishment process of the deep ontology fusion model, the technical team collected standard ontology libraries in the fields of mechanical engineering, electronic engineering, and software engineering as the basic data source. The ontology concept vectorization adopted the Word2Vec algorithm, and the trained word vector dimension was 300. The artificially annotated ontology mapping relationship totaled 15,600 pairs, of which 8,900 pairs were positive samples and 6,700 pairs were negative samples. The data enhancement technology used synonym replacement, concept hierarchy expansion, and other methods to expand the training sample size to 47,800.

[0113] The model training adopted the batch gradient descent algorithm, with a batch size of 64, an initial learning rate of 0.001, and an exponential decay strategy. The cross-validation used a 5-fold validation method, and the model's accuracy on the validation set reached 94.2%, with a recall rate of 92.8%. The loss function used a weighted combination of cross-entropy loss and mean square error loss, with a weight ratio of 0.7:0.3. After 450 iterations, the model converged.

[0114] In the feature dimension reduction optimization algorithm based on adaptive pooling, the window size function calculated the normalized window size as 0.672 based on the feature vector dimension of 512, the feature distribution variance of 0.156, and the target compression ratio of 0.5. In the computational efficiency optimization algorithm based on separable attention, the attention weight function calculated the weight coefficient based on the spatial position index value and the channel feature intensity value. The average weight of spatial attention was 0.328, and the average weight of channel attention was 0.445.

[0115] In the data storage optimization phase, the technical team analyzed the coincidence degree of the collected heterogeneous data. As shown in Table 1:

[0116] Table 1 Coincidence degree statistics table between different data sources

[0117]

[0118] For data combinations with a coincidence degree exceeding 85%, a similarity aggregation deduplication strategy was adopted. The similarity between data was calculated through cosine similarity, and data with a similarity greater than 0.95 were merged and stored, saving storage space by 38%. For data combinations with a coincidence degree less than 15%, a distributed independent storage method was adopted to ensure data integrity and traceability.

[0119] During the digital mainline construction process, the technical team established data bloodline traceability relationships throughout the product design, manufacturing, testing, and operation life cycle. The data bloodline graph contained a total of 25,600 nodes, 89,400 edges, and a maximum traceability depth of 15 layers. Each data object had a unique bloodline identifier, recording information such as data generation time, source system, processing process, and flow target. The data coverage rate of the full-process traceability reached 96.8%.

[0120] During the semantic mapping quality evaluation process, the technical team calculated the semantic mapping deviation values. As shown in Table 2:

[0121] Table 2 Statistical table of semantic mapping deviation values in different fields

[0122]

[0123] For mapping relationships with a deviation value less than 0.05, a fine-tuning mode is adopted, with a learning rate set to 0.0001 and a parameter adjustment step of 5% of the original value. For mapping relationships with a deviation value greater than 0.3, a coarse adjustment mode is adopted, with a learning rate set to 0.01 and a parameter adjustment step of 25% of the original value, achieving rapid convergence.

[0124] In the longitudinal integrated mapping construction, the technical team established hierarchical mapping relationships from functional models to performance models and then to physical models. The functional model contains 145 functional units, the performance model contains 87 performance indicators, and the physical model contains 234 physical components. The total number of mapping relationships is 1286, including 563 functional-performance mappings and 723 performance-physical mappings. As shown in Table 3, the accuracy of the mapping relationships increases with the number of iterations and stabilizes at 93.6% after the 180th iteration. Figure 2

[0125] The construction of the horizontal integration framework realizes collaborative simulation analysis across mechanical, electronic, and software professional fields. The computational efficiency of multi-domain comprehensive simulation improves with increasing parallelism, reaching optimal performance at a parallelism of 16, with a 78% reduction in computation time compared to the single-thread mode. Collaborative simulation involves various simulation tools such as finite element analysis, circuit simulation, and software testing. Data exchange uses the FMI standard interface, achieving seamless integration between different simulation tools.

[0126] The technical team also established a full-process model collaborative management system, realizing automatic propagation and impact analysis of design changes. The design change propagation network contains 3456 nodes, with an average propagation path length of 4.2 layers. The accuracy of change impact evaluation reaches 91.3%. Model version management uses a distributed version control method based on Git, supporting branching, merging, and rollback operations for models, with version history record integrity reaching 100%.

[0127] ​The present application brings significant technical progress compared with traditional means. The traditional data management method usually adopts a decentralized storage and manual association method, which cannot effectively handle the semantic mapping problem of multi-source heterogeneous data. The present application realizes automatic semantic mapping through a deep ontology fusion model, avoiding the subjectivity and inefficiency of manual processing. Traditional methods are difficult to establish complete data blood relationship, leading to difficulties in data traceability, while the present application constructs a full-process data traceability system through digital mainline technology, realizing complete recording of data sources and flow direction. In terms of storage optimization, the traditional method lacks an intelligent redundancy detection mechanism, and the present application adopts an adaptive deduplication strategy based on similarity, which significantly improves storage efficiency while ensuring data integrity. In terms of computing efficiency, the traditional multi-head attention mechanism has high computational complexity, and the present application separates the spatial dimension and channel dimension through a separable attention algorithm, significantly reducing the computational burden. These technical innovations enable the integration management of multi-source heterogeneous data to shift from traditional passive and decentralized management to active and unified management, providing strong technical support for the digital collaboration of complex products.

[0128] It should be noted that the variables involved in the present application are explained in detail as shown in Table 3.

[0129] Table 3: Variable Explanation Table

[0130]

[0131] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for integrating and managing multi-source heterogeneous data based on a digital thread, characterized in that, A distributed data storage architecture is constructed, using a combination of NoSQL databases and time-series databases to establish a data storage layer, while a multi-dimensional indexing mechanism based on hash sharding and time windows is established; demand model, SYSML model, and 3D model data are collected through the digital mainline platform, and data is acquired by dragging and dropping intermediate format files and API interface, and the collected raw data is standardized in format; Using ontology mapping algorithms, semantic parsing is performed on model data from different domains to construct a domain knowledge graph. A deep ontology fusion model is used to establish automated semantic mapping relationships between cross-domain models. When the data overlap exceeds the redundancy threshold, a similarity aggregation deduplication strategy is used to optimize storage space. If the data overlap is lower than the preset value, a distributed independent storage method is used to maintain data integrity. Construct a digital thread for heterogeneous data after semantic mapping, establish data lineage tracing relationships, and form a data link that runs through the entire product lifecycle; Optimize model mapping parameters using different adjustment modes based on semantic mapping deviation values; Based on the digital mainline architecture, the vertical integration mapping of the functional model to the performance model and physical model is realized. At the same time, a horizontal integration framework for multi-domain comprehensive simulation is constructed, and a full-process model collaborative management system is established. The structure of the deep ontology fusion model is as follows: the input layer receives multi-source ontology vectors, and the feature compression is achieved by dynamically adjusting the pooling window size through an adaptive pooling-based feature dimensionality reduction optimization algorithm; the intermediate layer uses a computational efficiency optimization algorithm based on separable attention to decompose multi-head attention into spatial and channel dimensions for separate processing; and the output layer generates a fused unified ontology representation vector. The semantic mapping deviation value is calculated by using a distance metric in vector space to determine the degree of difference between the concept representation vectors before and after semantic mapping. When the semantic mapping deviation value is less than the sensitivity threshold, a fine-tuning mode is used to optimize the model mapping parameters. When the semantic mapping deviation value is greater than the preset value, a coarse-tuning mode is used to achieve rapid mapping adjustment.

2. The multi-source heterogeneous data integration and management method based on digital mainline according to claim 1, characterized in that, The time-series database refers to a database system used to process time-series data, optimizing the storage and query performance of time-dimensional data.

3. The multi-source heterogeneous data integration and management method based on digital mainline according to claim 2, characterized in that, The adaptive pooling-based feature dimensionality reduction optimization algorithm dynamically adjusts the pooling window size to adapt to different feature distributions, thereby compressing feature dimensions while preserving key semantic information and reducing the computational burden of subsequent processing. The pooling window size is calculated using a window size function.

4. The multi-source heterogeneous data integration and management method based on digital mainline according to claim 3, characterized in that, The computational efficiency optimization algorithm based on separable attention decomposes the traditional multi-head attention mechanism into two independent computational processes: spatial attention and channel attention. Spatial attention focuses on the positional relationship between concepts, while channel attention focuses on the importance of different feature dimensions. By separating the computation, computational complexity and memory usage are significantly reduced. Attention weights are obtained through the attention weight function.

5. The multi-source heterogeneous data integration and management method based on digital mainline according to claim 4, characterized in that, The data overlap refers to the proportion of identical or similar data content in different data sources.

6. The multi-source heterogeneous data integration and management method based on digital mainline according to claim 5, characterized in that, The similarity aggregation deduplication strategy refers to calculating the similarity value between data and merging highly similar data for storage to save storage space.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the multi-source heterogeneous data integration and management method based on a digital mainline as described in any one of claims 1-6.

8. A multi-source heterogeneous data integration management system based on a digital thread, characterized in that, The system includes the computer-readable storage medium of claim 7, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

Images