Performance data intelligent analysis method and device for enterprise management

Abstract

The application relates to the technical field of big data, and discloses a performance data intelligent analysis method and device for enterprise management, wherein the method comprises the following steps: for multi-source heterogeneous performance data, first, identifying heterogeneous characteristics to generate classification labels, establishing a registration center to form a data source registration table; then, constructing a performance term ontology, generating a term mapping relationship table through semantic matching; next, extracting and backstepping index caliber information, resolving caliber conflicts, and establishing a caliber standardization library; then, combining the classification labels, the mapping relationship and the standard library, arranging an execution data fusion pipeline, generating a performance data warehouse and quality data; finally, updating the pipeline based on a fusion effect optimization strategy. The application can adapt to multiple heterogeneous data sources without aligning data one by one manually, unifies data standards and index calibers, greatly reduces manual intervention through automatic processing, reduces errors and processing costs, and improves data processing efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of big data technology, and more specifically, to a method and apparatus for intelligent analysis of performance data in enterprise management. Background Technology

[0002] In current enterprise management scenarios, performance appraisal data is widely distributed across multiple independent business platforms, such as human resource management systems, sales systems, customer relationship management systems, and production execution systems. These systems are developed by different vendors, employing different database structures and data formats. Their data collection dimensions, storage standards, and update mechanisms are incompatible, creating a typical multi-source heterogeneous data environment. Different departments exhibit significant differences in the terminology, statistical definitions, calculation logic, and value ranges for performance indicators. The lack of unified data standards and mapping rules makes direct data integration across systems difficult. Existing technologies primarily rely on manual export, format conversion, field matching, and data validation for data integration. This process is not only time-consuming and labor-intensive, but also prone to human error, leading to decreased data consistency and reliability. This prevents automated fusion and standardized processing of multi-source performance data, resulting in prominent data silos and severely hindering in-depth analysis and efficient utilization of performance data. Summary of the Invention

[0003] The main purpose of this application is to provide an intelligent analysis method for enterprise management performance data, which aims to solve the technical problems of inconsistent standards and definitions of multi-source heterogeneous performance data, making it difficult to automatically integrate and standardize, and relying on manual processing which is inefficient and prone to errors.

[0004] The first aspect of this application proposes one.

[0005] Furthermore, the steps of identifying heterogeneity characteristics and generating data source classification tags for multiple heterogeneous performance data sources, establishing a data source registry center to collect metadata and business semantic information, and generating a data source registry include: Extract structural, syntactic, semantic, and temporal heterogeneity features from multiple heterogeneous performance data sources, calculate and quantify the heterogeneity strength values, and generate a heterogeneous feature vector matrix; Heterogeneous feature maps are constructed based on heterogeneous feature vector matrices. Graph attention networks are used to generate node embedding vectors. Spectral clustering algorithms are used to classify data sources and generate classification labels. The labels and vectors are then merged to generate a data source classification result table. Based on the data source classification result table, a data source registry center is constructed, connector templates are matched to connect various data sources, metadata information is collected, business semantic information is extracted through natural language processing technology, and the two types of information are stored in the data source registry. Deploy a heterogeneous feature map update engine based on the data source registry, monitor metadata changes and reconstruct the heterogeneous feature map, update node embedding vectors and data source classification labels, re-collect metadata and business semantic information, and update the data source registry.

[0006] Further, the step of constructing a performance terminology ontology based on the data source registry, mapping data source terms to standard terms using a semantic matching algorithm, and generating a terminology mapping table includes: Based on the metadata information in the data source registry, a performance terminology ontology containing terminology layer, relationship layer and attribute layer is constructed, terminology node embedding vectors are generated, and the performance terminology ontology and vector information are stored in the terminology ontology library. Based on the term node embedding vectors in the term ontology, a semantic matching algorithm is used to map data source terms to standard terms, determine the mapping relationship between the two, and organize and save the mapping relationship to generate a term mapping relationship table. A quality assessment of term mapping relationships is conducted based on the term mapping relationship table, a term mapping quality report is generated, and low-confidence mapping relationships are marked as mappings to be reviewed. Based on the terminology mapping quality report, low-reliability mapping relationships are reviewed and corrected, and the corrected mapping relationships are updated to the terminology mapping relationship table, and the terminology ontology is updated synchronously.

[0007] Further, the steps of extracting performance indicator caliber information based on the terminology mapping table, reverse-engineering caliber information that is not explicitly recorded, detecting caliber differences and resolving conflicts using preset strategies, and establishing a caliber standardization library include: Based on the terminology mapping table, the performance indicator caliber information is extracted, a caliber causal graph model is constructed, causal relationships in dimensions such as caliber definition and calculation logic are identified, and a caliber causal graph containing causal nodes and directed causal edges is generated. Based on caliber information, a caliber knowledge graph is constructed. Semantic representations of unrecorded calibers are generated through graph reasoning. Similar caliber definitions are retrieved, and caliber information that is not clearly recorded is inferred by combining spatiotemporal evolution laws. Time series modeling is used to capture the dynamic trend of caliber changes, reverse the caliber information of unrecorded time points, identify abrupt changes in caliber definition, generate a report on the spatiotemporal distribution of caliber differences, and quantify the uncertainty of caliber characteristics. Based on various detection and inference results, a preset strategy is adopted to resolve the conflict of caliber differences, and the standardized caliber information is organized to establish a caliber standardization library.

[0008] Furthermore, the steps of using time-series modeling to capture the dynamic change trend of caliber, inferring caliber information at unrecorded time points, identifying abrupt changes in caliber definition, generating a report on the spatiotemporal distribution of caliber differences, and quantifying the uncertainty of caliber characteristics include: A distributed spatiotemporal semantic graph of calibers is constructed, with elements such as caliber definitions and calculation logic defined as graph nodes and semantic relationships defined as graph edges. A word vector model is used to convert caliber terms into vector representations. A federated temporal graph neural network combined with a multi-head attention mechanism is used to model the semantic graph. The model parameters of each node are aggregated through a federated learning algorithm to capture the semantic association strength and temporal evolution features of the caliber and generate a federated caliber embedding vector. Monte Carlo Dropout technique is used to quantify the uncertainty of the federated model, generate relevant feature values ​​of the caliber embedding vector and calculate the uncertainty score, and generate uncertainty-aware federated caliber embedding vector. The embedding vector is used to infer the caliber information for unrecorded time points, retrieve similar caliber definitions among data source nodes, identify caliber definition mutation points and generate a report on the spatiotemporal distribution of caliber differences, store the results in the federated caliber evolution knowledge base and update the model parameters of each node.

[0009] Furthermore, the step of combining the data source classification labels, terminology mapping relationship table, and standardization library to orchestrate and execute the data fusion pipeline to generate a performance data warehouse and quality data includes: Read the data source category tags, identify real-time data sources and batch data sources, build a domain ontology library and associate it with a terminology mapping table and a standardization library, and generate a stream-batch classification knowledge graph. The predictive model is used to input semantic features of the knowledge graph and features of the data source to generate a batch-stream collaborative orchestration scheme and pipeline DAG topology. The pipeline is orchestrated and collaborative nodes are designed to generate a batch-stream collaborative execution plan. Parallel processing pipelines are dynamically orchestrated according to the execution plan to process streaming data and batch data separately to generate corresponding data tables. Terminology mapping and standardization are completed based on knowledge graphs, and a fusion result table is generated by merging collaborative nodes. The merged data is loaded into a Lambda-based data warehouse to generate fact tables and dimension tables. Multi-dimensional quality checks and consistency assessments are then performed on the merged data to generate a performance data warehouse and corresponding quality data.

[0010] Furthermore, the steps of using a predictive model to input semantic features of the knowledge graph and features of the data source to generate a batch-stream collaborative orchestration scheme and pipeline DAG topology, orchestrating the pipeline and designing collaborative nodes, and generating a batch-stream collaborative execution plan include: Read the semantic features of the knowledge graph and the temporal features of the data source, construct a multi-head attention fusion network, calculate the weight matrix by scaling dot product attention, generate fusion feature vectors, extract temporal statistical features and identify important features to form a temporal fusion feature system; A multi-task prediction model is constructed using graph neural networks. The input is fused features, and the output is orchestration mode, pipeline topology, cooperating nodes, estimated execution time and resource consumption. Gradient updates are completed using a meta-learning framework to generate pipeline DAG topology and execution configuration. The system uses a time-series prediction model to input time series sequences, predicts future orchestration requirements, dynamically adjusts the orchestration scheme, generates a dynamic orchestration scheme, generates an interpretable report based on feature importance, performs stream-batch collaborative processing, and records execution logs. Based on the results of multi-task prediction and time-series prediction, the orchestration strategy and execution parameters are integrated to construct a pipeline DAG topology, design flow-batch collaborative nodes, and form a flow-batch collaborative execution plan.

[0011] Furthermore, the steps of constructing a multi-task prediction model using a graph neural network, inputting fused features, and outputting orchestration patterns, pipeline topology, cooperating nodes, estimated execution time, and resource consumption, and using a meta-learning framework to complete gradient updates, generating pipeline DAG topology and execution configuration, include: Read pipeline task data and construct a task graph. Use graph convolutional networks to extract semantic features of the task graph. Integrate knowledge graph features and data source features to construct a multi-task fusion network and generate multi-dimensional fusion features. A multi-task prediction model is constructed using graph convolutional networks. The model is input with fused features and outputs the orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes, generating the initial pipeline topology and configuration. A graph generation network based on fusion features is used to automatically generate pipeline topology graphs, node configuration parameters and edge configuration parameters, and generate corresponding execution configuration files. A graph reinforcement learning environment is constructed and an optimization strategy is trained. The meta-learning framework is used to complete the gradient update and rapid adaptation of model parameters. The topology and configuration are adjusted and optimized, and finally, a pipeline DAG topology and corresponding execution configuration information are generated.

[0012] Furthermore, the step of evaluating the fusion effect based on the performance data warehouse and quality data, optimizing the fusion strategy based on the evaluation results, and updating the data fusion pipeline includes: Extract source domain scenes from the historical optimization scene library, identify new target domain scenes, construct a feature space and use a transfer learning domain adaptive algorithm to complete feature alignment, construct a pipeline optimization causal graph, and generate a transfer training set and causal effect matrix; Meta-training is performed using a meta-learning algorithm to train a transfer reinforcement learning agent, a Bayesian optimizer is constructed and a hyperparameter search space is defined, relevant models and acquisition functions are initialized, causal effect estimation is performed on samples, and counterfactual inference predictions are calculated. Extract scene features from the new scene in the target domain, align them to the feature space of the source domain through a feature mapping function, input them into the transfer reinforcement learning agent to perform gradient updates to generate an initial optimization strategy, fuse Bayesian optimization suggestions and apply causal constraints to generate the optimal pipeline configuration; Deploy the optimal configuration and execute the fusion task, record the execution log, calculate the actual optimization effect, verify the relevant prediction accuracy and optimization efficiency, add new scenario samples to the target domain dataset, and update various models and data fusion pipelines.

[0013] A second aspect of this application also proposes an intelligent performance data analysis device for enterprise management, comprising: The tag generation module is used to identify heterogeneous characteristics and generate data source classification tags for multiple heterogeneous performance data sources, establish a data source registry center to collect metadata and business semantic information, and generate a data source registry. The terminology mapping module is used to construct a performance terminology ontology based on the data source registry, and to map data source terms to standard terms using a semantic matching algorithm to generate a terminology mapping relationship table. The caliber processing module is used to extract performance indicator caliber information based on the terminology mapping table, reverse-engineer caliber information that is not clearly recorded, detect caliber differences and resolve conflicts using preset strategies, and establish a caliber standardization library. The fusion execution module is used to combine the data source classification labels, terminology mapping relationship table and standardization library to orchestrate and execute the data fusion pipeline to generate a performance data warehouse and quality data. The optimization and update module is used to evaluate the fusion effect based on the performance data warehouse and quality data, optimize the fusion strategy and update the data fusion pipeline according to the evaluation results.

[0014] The first aspect of this plan brings the following benefits: This application generates data source classification tags, constructs performance terminology ontology, and performs semantic matching, eliminating the need for manual terminology alignment. It adapts to multiple heterogeneous data source scenarios, reducing manual processing costs. Simultaneously, terminology mapping, caliber inference, and difference resolution can accurately unify performance data standards and indicator calibers, solving the problem of data chaos. Furthermore, the automated data fusion pipeline orchestration and execution, along with optimized fusion effects, further reduce manual intervention. This dual guarantee reduces data processing errors, resulting in a high degree of automation and ease of operation. It can efficiently achieve the standardization and automatic fusion of multi-source performance data, significantly improving data processing efficiency, reducing error rates, eliminating reliance on manual processing, and meeting enterprises' business needs for accurate and efficient performance data analysis. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an intelligent performance data analysis method for enterprise management according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an intelligent performance data analysis device for enterprise management according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0018] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Reference Figure 1 This application provides a method for intelligent analysis of performance data in enterprise management, including: S1: For multiple heterogeneous performance data sources, identify heterogeneous characteristics and generate data source classification tags, establish a data source registry center to collect metadata and business semantic information, and generate a data source registry; S2: Construct a performance terminology ontology based on the data source registry, and use a semantic matching algorithm to map data source terms to standard terms to generate a terminology mapping table; S3: Extract performance indicator caliber information based on the terminology mapping table, reverse-engineer caliber information that is not clearly recorded, detect caliber differences and resolve conflicts using preset strategies, and establish a caliber standardization library; S4: Combining the data source classification labels, terminology mapping relationship table and standardization library, orchestrate and execute the data fusion pipeline to generate a performance data warehouse and quality data; S5: Based on the performance data warehouse and quality data evaluation, optimize the fusion strategy and update the data fusion pipeline according to the evaluation results.

[0020] In step S1, the heterogeneous features of multi-source heterogeneous performance data sources are quantified and classified by using depthwise traversal feature extraction and graph attention network combined with spectral clustering algorithm. At the same time, a data source registry is built based on microservice architecture to collect metadata and business semantic information, and finally a data source registry with classification labels is generated. Taking three heterogeneous data sources—HR system, sales CRM system, and production MES system—as examples, their heterogeneity is manifested in the following ways: the HR system has a simple table structure but low update frequency; the CRM system has inconsistent character sets and chaotic terminology naming; and the MES system has complex data formats and large time-series delays. When an enterprise initiates performance data integration, the system first performs a deep traversal of the three data sources to extract four types of heterogeneous features: structure, syntax, semantics, and time series. This generates a heterogeneous feature vector matrix—each element in the matrix represents the heterogeneity strength value of a certain type of heterogeneous feature (vertical axis) at that position under the corresponding data source (horizontal axis), with a value range from 0 to 1. Subsequently, the traditional clustering algorithm is optimized based on the feature vector matrix: the original algorithm's "single feature dimension clustering" is changed to "multi-feature fusion graph structure clustering," while also employing graph... An attention network aggregates neighbor node features to generate 128-dimensional node embedding vectors, improving classification accuracy and adapting to multi-dimensional heterogeneous feature scenarios. Spectral clustering algorithms are used to calculate these node embedding vectors, matching corresponding categories to each data source from five types of labels, including high structural heterogeneity and high syntactic heterogeneity. The HR system is identified as highly temporally heterogeneous, the CRM system as highly semantically heterogeneous, and the MES system as comprehensively heterogeneous. Then, through a microservice architecture registry center, connector templates are automatically matched to each data source based on the classification labels, collecting metadata such as table structures and field definitions. NLP techniques are used to extract business semantic information such as indicator calculation logic. Finally, within a pre-defined metadata storage specification, classification labels are associated and integrated with metadata and business semantic information to form a data source registry containing complete information for the three data source types. This step provides a categorized data source foundation for subsequent terminology mapping and standardization. All collected and categorized information provides accurate input for subsequent steps, forming a closed loop before data processing. Furthermore, a real-time update engine monitors data source changes to ensure information timeliness, laying the data source layer foundation for fully automated data fusion. This step enables intelligent classification and standardized registration of heterogeneous data sources, replacing manual feature recognition and information collection, significantly improving data source processing efficiency. At the same time, it provides unified and accurate data source information for subsequent steps, solving the data silo problem from the source and ensuring the continuity and accuracy of data processing throughout the entire process.

[0021] In step S2, using graph neural network embedding technology and semantic matching algorithms, combined with the cosine similarity matching criterion, a performance terminology ontology is constructed based on the data source registry generated in S1. This maps heterogeneous data source terms to standard terms, ultimately generating a terminology mapping table. Taking the performance terms from the HR system, CRM system, and MES system in S1 as examples, the HR system refers to employee performance as "performance completion value," the CRM system as "sales achievement rate," and the MES system as "production target achievement value." These are all standard terms for core performance indicators. When the system initiates terminology standardization, it first constructs a performance terminology ontology containing terminology, relationship, and attribute layers based on metadata and business semantic information in the data source registry. A 128-dimensional terminology node embedding vector is generated and stored in the terminology ontology library—the library contains standard terms such as "core performance indicators" and their corresponding embedding vectors. Subsequently, based on the terminology ontology library, the traditional terminology matching method is optimized: the original algorithm's "literal matching" is changed to "semantic embedding vector matching," and a cosine similarity threshold of 0.75 is set. This approach improves the semantic fit of terminology mapping while filtering out low-reliability mapping results. By calculating the cosine similarity between the embedding vectors of "performance completion value," "sales achievement rate," and "production target achievement value" and the embedding vector of the standard terminology "core performance indicators," all three similarities exceed the threshold of 0.75. A terminology mapping quality assessment model then scores these mappings across three dimensions: accuracy, coverage, and consistency, determining them to be high-reliability mappings. Finally, within a predefined terminology mapping storage specification, source terms, target standard terms, similarity scores, and mapping time are integrated to form a terminology mapping relationship table containing terminology mapping information from the three data sources. Simultaneously, this mapping result is fed back to update the terminology ontology, forming a closed loop with the data source classification information in S1, providing a unified terminology foundation for subsequent standardization. This step achieves standardized mapping of performance terms, resolves semantic heterogeneity issues, provides terminology support for standardization, and improves data semantic consistency.

[0022] In step S3, through causal graph modeling and federated temporal graph neural networks, combined with the Monte Carlo Dropout uncertainty quantification method, the performance indicator caliber information is extracted based on the terminology mapping table generated in S2. Unrecorded calibers are deduced and conflicts are resolved, ultimately establishing a caliber standardization library. Taking the standardized terminology "core performance indicators" in S2 as an example, in the HR system, only the calculation logic is recorded as "monthly completion / target amount," without a time range caliber. In the CRM system, the time range is recorded as "natural month," with a default calculation logic. In the MES system, the time range is "production month," and the calculation logic is "daily completion accumulation / monthly target amount." The system first extracts the caliber information of the three systems based on the terminology mapping table, constructing a caliber causal graph—nodes include four types, such as caliber parameters and calculation formulas, and causal edges represent the causal influence direction of each dimension. Subsequently, the traditional caliber completion method is optimized: the original "manual experience completion" is changed to "knowledge graph reasoning + temporal modeling deduction," while a federated temporal graph neural network is used to capture the spatiotemporal nature of the caliber. The evolutionary approach achieves both accurate reverse inference of unrecorded definitions and quantifies the uncertainty of definitional characteristics. By retrieving similar definitions from a definitional knowledge graph, it infers that the HR system's time range is "natural month," and that the CRM system's calculation logic is the same as the HR system. Monte Carlo Dropout sampling is then used to quantify the differences in definitions among the three systems, identifying the abrupt change points between the MES system's "production month" and the standard "natural month." Finally, under a pre-defined definitional conflict resolution strategy, a business scenario adaptation strategy is adopted to unify the MES system's time range to "natural month," and the calculation logic is simultaneously standardized. All unified definitional information is integrated and stored in a definitional standardization library, forming a closed loop with S1 data source classification and S2 terminology mapping, providing a unified basis for subsequent data fusion. This step achieves the standardization and unification of performance indicator definitions, resolves definitional conflicts, and lays the foundation for accurate indicator calculations for data fusion.

[0023] In step S4, through knowledge graph modeling and multi-task prediction model, combined with the Lambda architecture data warehouse construction criteria, and combined with the S1 data source classification labels, S2 terminology mapping relationship table and S3 standardization library, a batch and stream integrated data fusion pipeline is orchestrated and executed to finally generate a performance data warehouse and quality data. Taking the three data sources classified as S1 (HR with high temporal heterogeneity, CRM with high semantic heterogeneity, and MES with comprehensive heterogeneity) as examples, after S2 terminology unification and S3 standardization, the core performance indicator for all three is "actual completed amount / target amount within a natural month". The system first reads the classification labels of the three data sources, identifies HR as a batch processing data source, and CRM and MES as real-time data sources, constructs a domain ontology library and associates it with a terminology mapping table and a standard library, generating a batch / stream classification knowledge graph—the graph contains 1000+ nodes, 5000+ edges, and 128 embedded dimensions. Subsequently, the traditional data fusion and orchestration method is optimized: the original "manual fixed pipeline" is changed to "AI predictive dynamic orchestration", and a multi-task prediction model is constructed using a graph neural network, which can adapt to the different processing needs of batch and stream data, and can also... To improve the efficiency and quality of data fusion, this step generates a stream-batch collaborative DAG topology by using semantic features of the model input graph and data source features. Three stream-batch collaborative nodes are designed, and real-time pipelines for CRM and MES (latency <10 seconds) and batch processing pipelines for HR (24-hour window) are orchestrated. Data is processed in parallel according to plan, and automatic terminology mapping and standardization are completed. A fusion result table is generated by merging the data through collaborative nodes. Finally, under the preset Lambda architecture specification, the fused data is loaded into the real-time layer (1TB, the most recent 7 days) and the batch processing layer (10TB, full historical data), generating performance fact tables and dimension tables. Simultaneously, quality data (overall score ≥0.85) is generated from four dimensions including completeness and accuracy, forming a data processing closed loop with the preceding steps, providing complete data support for optimizing the fusion effect. This step achieves automated stream-batch integrated data fusion, generating a hierarchical performance data warehouse and ensuring the quality and consistency of the fused data.

[0024] In step S5, the fusion effect is evaluated based on the performance data warehouse and quality data generated in S4, using a transfer learning domain adaptive algorithm and a transfer reinforcement learning agent, combined with a Bayesian optimization method. The fusion strategy is then optimized and the data fusion pipeline is updated. Taking the core performance indicator data after the fusion of HR, CRM, and MES systems in S4 as an example, the quality data shows that the MES system data consistency score is 0.78, which does not meet the preset standard of ≥0.85. The system first extracts 50 source domain fusion scenarios from the historical optimization scenario library, identifies new target domain scenarios with low MES data consistency, constructs a 256-dimensional feature space, uses the MMD algorithm to align the source and target domain features, and constructs a pipeline optimization causal graph—containing 30 nodes and 80 edges. Twelve operable configuration variables, such as the number of pipeline nodes and resource allocation, are identified. Subsequently, the traditional strategy optimization method is optimized: the original "manual parameter tuning optimization" is changed to "transfer reinforcement learning + Bayesian fusion optimization," and MAML meta-learning is adopted. The training agent achieves rapid policy adaptation across scenarios and improves optimization accuracy. Through transfer learning, the agent performs a three-step gradient update on target domain features to generate an initial optimization policy. Combined with hyperparameter suggestions from a Bayesian optimizer, an optimal configuration is generated by fusing the transfer weights (0.7) and Bayesian weights (0.3). This increases the number of MES real-time pipeline nodes by two and upgrades CPU resources to eight cores. Finally, under pre-defined pipeline validation specifications, the optimal configuration is deployed and the fusion task is re-executed. The validation shows that the MES data consistency score has improved to 0.92, meeting the quality standard. The new scenario sample is added to the dataset, and the fusion policy and pipeline topology are updated, forming a closed-loop process with the preceding S1-S4 steps, achieving continuous self-optimization of performance data fusion. This step realizes intelligent optimization of the fusion policy and pipeline iteration, continuously improving data fusion quality and achieving a closed-loop self-optimization process.

[0025] In one embodiment, the steps of identifying heterogeneous characteristics and generating data source classification tags for multiple heterogeneous performance data sources, establishing a data source registry center to collect metadata and business semantic information, and generating a data source registry include: S10: Extract structural, syntactic, semantic, and temporal heterogeneity features from multiple heterogeneous performance data sources, calculate and quantify the heterogeneity strength value, and generate a heterogeneous feature vector matrix; S11: Construct a heterogeneous feature map based on the heterogeneous feature vector matrix, use a graph attention network to generate node embedding vectors, use a spectral clustering algorithm to classify the data source to generate classification labels, and merge the labels and vectors to generate a data source classification result table; S12: Construct a data source registry based on the data source classification result table, match connector templates to connect various data sources, collect metadata information, extract business semantic information through natural language processing technology, and store the two types of information in the data source registry. S13: Deploy a heterogeneous feature map update engine based on the data source registry, monitor metadata changes and reconstruct the heterogeneous feature map, update node embedding vectors and data source classification labels, re-collect metadata and business semantic information, and update the data source registry.

[0026] In this embodiment, four types of heterogeneity features—structural, syntactic, semantic, and temporal—are first extracted from multi-source heterogeneous performance data sources. The heterogeneity strength is calculated and quantified to generate a heterogeneity feature vector matrix. Taking three types of heterogeneous performance data sources—enterprise HR system, sales CRM system, and production MES system—as examples, a full-dimensional deep traversal is performed on these three systems. Structural heterogeneity features such as table structure complexity and field type distribution are extracted; the HR system table structure complexity is 0.2, and the MES system reaches 0.8. Syntactic heterogeneity features such as data format consistency and character set consistency are extracted; the CRM system's character set consistency is only 0.3. Semantic heterogeneity features such as terminology naming are extracted. Features such as consistency and business meaning consistency are achieved. The naming consistency of terminology between the MES system and the HR system is 0.4. Features such as time-series heterogeneity extraction update frequency consistency and data latency consistency are achieved. The update frequency consistency of the HR system is 0.2 and that of the CRM system is 0.9. Through a normalization algorithm, various feature values ​​are transformed into heterogeneity intensity values ​​in the range of 0-1. The data is organized according to the dimension of "data source-feature type-feature value-intensity value" to generate a 3-row 4-column heterogeneous feature vector matrix. The horizontal axis of the matrix represents four types of heterogeneous features, and the vertical axis represents three types of data sources. Each cell corresponds to the feature intensity value of a specific data source, which becomes the core data foundation for subsequent data source classification. Then, a heterogeneous feature map is constructed based on the heterogeneous feature vector matrix. A graph attention network is used to generate node embedding vectors. A spectral clustering algorithm is used to classify the data sources and generate classification labels. The labels and vectors are merged to generate a data source classification result table. The three types of data sources are used as nodes in the feature map, with node features being the heterogeneous feature vectors of each data source. The relationship between nodes is constructed using feature similarity as edge weights. A graph attention network with a multi-head attention mechanism aggregates the features of neighboring nodes, generating 128-dimensional node embedding vectors. A spectral clustering algorithm is then used to perform cluster analysis on the embedding vectors. Based on the clustering results, corresponding classification labels are matched to the three types of data sources. The HR system is determined to be highly temporally heterogeneous, the CRM system to be highly semantically heterogeneous, and the MES system to be comprehensively heterogeneous. The classification labels are then associated one-to-one with the 128-dimensional node embedding vectors, generating data containing the data source name, classification label, and embedding vector. The system first generates a source classification result table. Then, based on this table, a data source registry is constructed. Connector templates are matched to link different data sources, metadata is collected, and business semantic information is extracted using natural language processing (NLP). Both types of information are stored to generate a data source registry. The data source registry is built on a microservice architecture. Based on the tags in the classification result table, a time-series connector template is matched for the HR system, a semantic connector template for the CRM system, and a comprehensive connector template for the MES system. Seamless integration with each data source is achieved through these templates. Basic data source information, table structure information, and field definition information from each system are collected. Simultaneously, NLP technology is used to analyze the operation manuals and user manuals of each system, extracting business semantic information such as field business meanings and indicator calculation logic. The metadata and business semantic information are then associated and stored according to a unified standard to generate a complete data source registry. Finally, a heterogeneous feature map update engine is deployed based on the data source registry. This engine monitors metadata changes and reconstructs the heterogeneous feature map, updates node embedding vectors and data source classification labels, re-collects metadata and business semantic information, and updates the data source registry. A real-time monitoring heterogeneous feature map update engine is deployed in the registry center to continuously monitor metadata changes in the data source registry. When a new production indicator field is added to the MES system, causing a change in the table structure, the feature map reconstruction process is automatically triggered. The four types of heterogeneous features of the MES system are re-extracted, and their intensity values ​​are calculated. The heterogeneous feature vector matrix and feature map are updated, and new node embedding vectors are generated again through a graph attention network. The classification labels of the MES system are updated using a spectral clustering algorithm. Subsequently, the connector template is re-matched based on the updated labels, and metadata and business semantic information of the newly added fields are collected. Finally, the entire data source registry update is completed, forming a complete closed-loop processing logic with the preceding feature extraction, classification, and registration steps, ensuring the real-time nature and accuracy of the data source information. This embodiment achieves full-dimensional extraction and accurate classification of heterogeneous performance data source features, establishes a real-time updateable registry center, ensures the uniformity and timeliness of data source information, and lays a solid foundation for subsequent data processing.

[0027] In one embodiment, the step of extracting structural, syntactic, semantic, and temporal heterogeneity features from multiple heterogeneous performance data sources, calculating and quantifying heterogeneity strength values, and generating a heterogeneous feature vector matrix includes: S101: Perform a deep traversal of multiple heterogeneous performance data sources, extract structural, syntactic, semantic, and temporal heterogeneous features respectively, and clarify the specific representation content corresponding to each type of heterogeneous feature; S102: Quantify the extracted heterogeneity features to obtain the heterogeneity intensity value corresponding to each type of feature, and form the feature intensity quantification result; S103: Integrate the names, types, eigenvalues, and corresponding intensity quantification results of various heterogeneous features, and sort out the correlation and correspondence between feature data; S104: Based on the correspondence between the sorted feature data, organize the data according to the corresponding dimensions of the data source and features to generate a heterogeneous feature vector matrix.

[0028] In this embodiment, a deep traversal is first performed on multiple heterogeneous performance data sources to extract structural, syntactic, semantic, and temporal heterogeneity features. The specific representation content corresponding to each type of heterogeneity feature is then clarified. Using three types of heterogeneous performance data sources—enterprise HR systems, sales CRM systems, and production MES systems—as the traversal objects, a deep analysis of each system is conducted through comprehensive data exploration. Structural heterogeneity features focus on table structure complexity, field type distribution, and relational complexity. The HR system contains only 3 performance-related tables with a single field type, while the MES system contains 12 related tables with diverse field types and complex inter-table relationships. Syntactic heterogeneity features revolve around data format consistency, encoding consistency, and character set consistency. The RM system suffers from inconsistent encoding (mixed Chinese and English), with some fields in text format and others in numerical format. Semantic heterogeneity is addressed by ensuring consistency in terminology naming and business meaning: the HR system uses "employee performance completion value," the CRM system uses "sales achievement rate," and the MES system uses "production target value," resulting in different names and varying business interpretations for the same indicator. Temporal heterogeneity focuses on consistency in update frequency and data latency: the HR system updates manually monthly with a latency exceeding 72 hours, the CRM system updates automatically in real-time with a latency of less than 1 minute, and the MES system updates hourly with a latency of approximately 30 minutes. The specific representation dimensions of each of the four types of features are clearly defined to achieve accurate extraction of all-dimensional heterogeneous features. Then, the extracted heterogeneous features are quantified to obtain the heterogeneity strength value corresponding to each type of feature, forming a feature strength quantification result. A quantification scoring standard is set for the specific representation content of each type of feature, with 0 representing no heterogeneity and 1 representing complete heterogeneity. The heterogeneity values ​​of each data source in different feature dimensions are calculated using a feature difference algorithm: HR system structure heterogeneity strength value 0.2, syntax 0.3, semantics 0.5, temporal 0.9; CRM system structure 0.3, syntax 0.8, semantics 0.9, temporal 0.1; MES system structure 0.9, syntax 0.6, semantics 0.7, temporal 0.4. After calibration using a normalization algorithm, all feature dimensions are obtained. The heterogeneity intensity value is used to form a standardized feature intensity quantification result. Secondly, the names, types, feature values ​​and corresponding intensity quantification results of various heterogeneous features are integrated to sort out the correlation and correspondence between feature data. Taking the data source as the core dimension, the structural, syntactic, semantic and temporal feature names, feature type classification, original feature values ​​of each representation content and the calibrated heterogeneity intensity value of the three systems of HR, CRM and MES are matched one by one to establish a five-dimensional correlation relationship of "data source-feature type-feature name-original feature value-heterogeneous intensity value". The precise mapping relationship between each type of data source and four types of features is sorted out to ensure that there is no omission or misalignment of feature data. Finally, based on the identified feature data relationships, the data is organized according to the corresponding dimensions of data sources and features, generating a heterogeneous feature vector matrix. With the data source as the vertical axis and the four types of heterogeneous features as the horizontal axis, the core heterogeneity strength values ​​from the identified five-dimensional relationships are used as the matrix cell values, constructing a 3x4 heterogeneous feature vector matrix. The vertical axis represents HR, CRM, and MES systems, while the horizontal axis represents structural, syntactic, semantic, and temporal features. Each cell corresponds to the heterogeneity strength value of a specific data source in a specific feature dimension. This matrix becomes the core data carrier for subsequent heterogeneous feature map construction and data source classification, forming a closed loop with the preceding feature extraction, quantification, and association analysis stages, providing a standardized and structured data foundation for subsequent intelligent data source classification. This embodiment achieves refined extraction and standardized quantification of heterogeneous performance data source features, constructing a structured feature vector matrix that provides accurate and unified data support for subsequent data source classification.

[0029] In one embodiment, the step of constructing a performance terminology ontology based on the data source registry, mapping data source terms to standard terms using a semantic matching algorithm, and generating a terminology mapping table includes: S20: Based on the metadata information in the data source registry, construct a performance terminology ontology containing a terminology layer, a relationship layer, and an attribute layer, generate terminology node embedding vectors, and store the performance terminology ontology and vector information into the terminology ontology library. S21: Based on the term node embedding vectors in the term ontology, a semantic matching algorithm is used to map the data source terms to standard terms, determine the mapping relationship between the two, and organize and save the mapping relationship to generate a term mapping relationship table. S22: Conduct a quality assessment of term mapping relationships based on the term mapping relationship table, generate a term mapping quality report, and mark low-confidence mapping relationships as mappings to be reviewed; S23: Based on the terminology mapping quality report, review and correct low-reliability mapping relationships, update the corrected mapping relationships to the terminology mapping relationship table, and update the terminology ontology database simultaneously.

[0030] In this embodiment, firstly, based on the metadata information in the data source registry, a performance terminology ontology containing a terminology layer, a relationship layer, and an attribute layer is constructed, and terminology node embedding vectors are generated. The performance terminology ontology and vector information are stored in the terminology ontology library. Based on the data source registry containing HR, CRM, and MES system information generated by S1, all performance-related terminology metadata is extracted. The terminology layer defines a set of general standard terms such as "core performance indicators" and "performance completion rate". The relationship layer defines semantic relationships such as synonymy, hierarchical, and association between terms. For example, "sales achievement rate" and "core performance indicators" are subordinate to each other. The attribute layer defines attribute constraints such as the calculation logic and data source of terms. Graph neural network embedding technology is used to encode the features of each standard term and generate 128-dimensional terminology node embedding vectors. The complete three-layer performance terminology ontology and corresponding embedding vectors are uniformly stored in the terminology ontology library to form a standardized terminology base library. Then, based on the terminology node embedding vectors in the terminology ontology, a semantic matching algorithm is used to map data source terms to standard terms, determine the mapping relationship between the two, and organize and save the mapping relationship to generate a terminology mapping relationship table. Heterogeneous data source terms such as "performance completion value," "sales achievement rate," and "production target achievement value" are extracted from HR, CRM, and MES systems. The cosine similarity between the embedding vectors of each data source term and the embedding vectors of standard terms in the terminology ontology is calculated using a graph neural network matching algorithm. A similarity threshold of 0.75 is set. The similarity between the three data sources and the "core performance indicators" all exceed the threshold, and mapping relationships are established for each. Information such as source term name, target standard term name, similarity score, and mapping time are organized and archived to generate a table containing three types of system terms. The process involves: first, establishing a terminology mapping relationship table; second, conducting a quality assessment of the terminology mapping relationships based on this table, generating a terminology mapping quality report, marking low-reliability mapping relationships as pending review, and constructing a terminology mapping quality assessment model with three dimensions: mapping accuracy, coverage, and consistency. Each record in the mapping relationship table is then quantitatively scored. The mapping scores for "Performance Completion Value" and "Core Performance Indicators" in the HR system (0.92), "Sales Achievement Rate" in the CRM system (0.88), and "Production Target Achievement Value" in the MES system (0.85) are all considered high-reliability mappings. If mapping records exist at the edge of the similarity threshold, they are marked as pending review. All scoring results and marking information are then integrated to generate a complete terminology mapping quality report. Finally, based on the terminology mapping quality report, low-reliability mapping relationships are reviewed and corrected. The corrected mapping relationships are updated to the terminology mapping relationship table, and the terminology ontology is updated synchronously. For mapping records pending review in the quality report, technical personnel conduct manual review and correction in accordance with enterprise business rules. After confirming the rationality of the mapping, the correction results are entered into the system. In this embodiment, there are no low-reliability mappings that do not require correction, and high-reliability mapping relationships are directly confirmed. Simultaneously, the mapping association information between terms from each data source and standard terms is updated to the terminology ontology, improving the semantic relationship definition between terms and enabling data interoperability between the terminology ontology and the mapping relationship table. The entire process is closely connected with the preceding data source registration stage, achieving terminology standardization based on unified metadata, and providing a unified terminology reference for subsequent standardization. This forms a closed-loop processing logic from data source information to terminology mapping, ensuring the uniformity and accuracy of performance terminology. This embodiment constructs a standardized performance terminology ontology, realizing accurate mapping from heterogeneous data source terms to standard terms, ensuring the effectiveness of mapping through quality assessment, and laying a unified terminology foundation for subsequent data processing.

[0031] In one embodiment, the step of mapping data source terms to standard terms using a semantic matching algorithm based on the term node embedding vectors in the terminology ontology, determining the mapping relationship between the two, and organizing and saving the mapping relationship to generate a term mapping relationship table includes: S210: Extract multi-granular semantic features of data source terms, generate multi-dimensional embedding vectors and construct feature representation matrices, fuse features through an attention mechanism to form a multi-granular semantic representation of data source terms; S211: Construct the support set and query set for term mapping, initialize the mapping network parameters and optimize and update them based on the support set, and generate semantic representations of terms within the query set by combining the updated parameters; S212: A Bayesian neural network is used to calculate the probability distribution of term embedding vectors. Relevant feature values ​​are generated through sampling, and term mapping candidates that require manual review are selected based on uncertainty metrics. S213: Construct a terminology knowledge graph and generate graph-enhanced embedding vectors. Combine these vectors to calculate terminology similarity, determine the mapping relationship between data source terms and standard terms, integrate and save all mapping relationship information, and generate a terminology mapping relationship table.

[0032] In this embodiment, firstly, multi-granular semantic features of data source terms are extracted, multi-dimensional embedding vectors are generated, and a feature representation matrix is ​​constructed. Features are then fused using an attention mechanism to form a multi-granular semantic representation of the data source terms. Taking three data source terms—"performance completion value" from the HR system, "sales achievement rate" from the CRM system, and "production target achievement value" from the MES system—as processing objects, three layers of semantic features are extracted: word-level, character-level, and sub-word-level. A BERT pre-trained model is used to generate 768-dimensional embedding vectors for each layer. A multi-granular feature representation matrix is ​​constructed for "performance completion value," containing word-level "performance / completion value," character-level "performance / efficiency / completion / value," and sub-word-level "performance / completion / value." Finally, a 12-head multi-head attention mechanism is used to weight and fuse the three layers of features to generate fusion weights. Vectors are ultimately used to form a 768-dimensional, multi-granular semantic representation containing complete semantic and morphological information, laying the feature foundation for accurate term matching. Then, a support set and query set for term mapping are constructed, the mapping network parameters are initialized and optimized and updated based on the support set, and the semantic representation of terms in the query set is generated by combining the updated parameters. Fifty pairs of labeled enterprise performance terms (such as "performance completion rate - core performance indicators") are selected as the support set, and terms from three data sources are used as the query set. The mapping network parameters are initialized using the MAML meta-learning framework with a learning rate of 0.001. Based on the support set, a three-step gradient update is completed to generate fast adaptation parameters. The terms in the query set are input into the updated network to generate accurate semantic representations of terms under small sample conditions, adapting to scenarios with a small sample size of enterprise performance terms. Secondly, a Bayesian neural network is used to calculate the probability distribution of term embedding vectors. Relevant feature values ​​are generated through sampling. Based on the uncertainty metric, candidate term mappings requiring manual review are selected. The semantic representations of terms from three data sources are input into the Bayesian neural network. The mean and variance of the embedding vectors are generated through 100 Monte Carlo Dropout samplings. An uncertainty metric is calculated with a variance threshold of 0.1 and an entropy threshold of 1.5. The calculated metric values ​​for all three terms are below the thresholds, indicating no high-uncertainty term pairs, thus eliminating the need for manual review. Terms exceeding the thresholds are directly marked as pending review. Finally, a term knowledge graph is constructed, and graph-enhanced embedding vectors are generated. Term similarity is calculated using these vectors to determine the mapping relationship between data source terms and standard terms. All mapping relationship information is integrated, stored, and used to generate term mapping relationships. This paper constructs a performance terminology knowledge graph containing terminology nodes, relation edges, and attribute nodes. "Core performance indicators" are set as core standard nodes. Relationship edges include synonyms and subordinate associations. A two-layer graph attention network (GAT) is used to generate 256-dimensional graph-enhanced embedding vectors. Cosine similarity is calculated by fusing multi-granularity semantic representation and graph-enhanced embedding vectors. The similarity between "performance completion value," "sales achievement rate," "production target achievement value," and "core performance indicators" all exceed the 0.75 threshold, establishing a one-to-one mapping relationship. Subsequently, information such as source terms, target standard terms, similarity scores, and mapping time are integrated and organized according to a unified standard to generate a terminology mapping relationship table. This table, together with the preceding terminology ontology and multi-granularity feature representation, forms a closed loop, providing complete data support for subsequent terminology mapping quality assessment and ensuring the accuracy and traceability of the mapping relationships. This embodiment achieves accurate terminology mapping through multi-granularity feature extraction and multi-technology fusion, filters high-uncertainty candidates to ensure mapping quality, and the generated mapping table provides a unified basis for subsequent standardization.

[0033] In one embodiment, the steps of extracting multi-granular semantic features of data source terms, generating multi-dimensional embedding vectors and constructing feature representation matrices, and fusing features through an attention mechanism to form a multi-granular semantic representation of data source terms include: S2101: Extract word-level, character-level, and sub-word-level semantic features from the data source terms, use a pre-trained language model to generate embedding vectors for each granularity layer, construct a multi-granularity feature representation matrix, and calculate the semantic richness index for each granularity layer. S2102: An autoencoder architecture is used to deconstruct the features of each granularity layer into atomic feature units, generate an atomic feature set, and reconstruct and fuse it. Based on the fusion result, a hierarchical multi-granularity feature vector is generated. S2103: Construct term pairs for positive and negative samples, train a granularity enhancement network, and strengthen feature discriminativeness through contrastive learning to generate contrast-enhanced multi-granularity representation vectors; S2104: A Bayesian neural network is used to calculate the uncertainty score of features at each granularity level. The weights of each granularity level are adjusted through an uncertainty-weighted attention mechanism. After the features are weighted and fused, a multi-granular semantic representation of the data source terms is generated.

[0034] In this embodiment, firstly, word-level, character-level, and sub-word-level semantic features of data source terms are extracted. A pre-trained language model is used to generate embedding vectors for each granularity layer, constructing a multi-granularity feature representation matrix. The semantic richness index for each granularity layer is calculated. Taking "performance completion value" from the HR system, "sales achievement rate" from the CRM system, and "production target achievement value" from the MES system as processing objects, word-level (performance / completion value, sales / achievement rate), character-level (performance / efficiency / completion / value), and sub-word-level (performance / completion / value) semantic features are extracted for each term. A BERT pre-trained model is used to generate 768-dimensional embedding vectors for each layer, constructing a 3-row, 768-column multi-granularity feature representation matrix for each term. Then, the entropy method is used to calculate the semantic distribution entropy of word-level features, the morphological complexity of character-level features, and the structural diversity of sub-word-level features, generating a semantic richness score vector in the 0-1 range. The word-level score for "performance completion value" is 0.91, the character-level score is 0.78, and the sub-word-level score is 0.85, providing a basis for subsequent feature fusion and weighting. Then, an autoencoder architecture is used to deconstruct the features of each granularity layer into atomic feature units, generating atomic feature sets and reconstructing and fusing them. Based on the fusion result, a hierarchical multi-granularity feature vector is generated. An autoencoder architecture with 3 layers of encoder and 3 layers of decoder is built, deconstructing the 768-dimensional embedding vector of each granularity layer into 64-dimensional semantic, morphological, and structural atomic feature units, generating atomic feature sets at each level. Through reconstruction and fusion of the encoder and decoder, the atomic features are mapped to 256-dimensional hidden layer features, finally generating a 768-dimensional hierarchical multi-granularity feature vector, achieving feature dimensionality reduction and effective fusion. Next, term pairs of positive and negative samples are constructed, and the training granularity is increased. A strong network is constructed and feature discrimination is enhanced through contrastive learning. Contrast-enhanced multi-granularity representation vectors are generated, and positive sample pairs (performance completion value - core performance indicators, sales achievement rate - core performance indicators) and negative sample pairs (performance completion value - production plan quantity, sales achievement rate - number of customer visits) are constructed. The hierarchical multi-granularity feature vectors are input into the granularity enhancement network, and the contrast loss temperature parameter is set to 0.07. The features are mapped to a 128-dimensional contrastive learning space through a projection head to maximize the similarity of synonymous term pairs and minimize the similarity of non-synonymous term pairs. After the network training is completed, a contrast-enhanced 768-dimensional multi-granularity representation vector is generated to enhance the recognizability of term features. Finally, a Bayesian neural network is used to calculate the uncertainty score of features at each granularity level. The weights of each granularity level are adjusted through an uncertainty-weighted attention mechanism. After the features are weighted and fused, a multi-granular semantic representation of the data source terms is generated. The contrast enhancement feature vector is input into the Bayesian neural network, and after 50 Monte Carlo Dropout samplings, the mean and variance of the embedding vectors of each granularity level are generated. Uncertainty scores at the word level, character level, and sub-word level are calculated. Features with low uncertainty are given higher weights. For example, for "performance completion value", the uncertainty at the word level is 0.08 and the weight is 0.45, at the character level it is 0.15 and the weight is 0.3, and at the sub-word level it is 0.1 and the weight is 0.25. The three layers of features are weighted and fused through the uncertainty-weighted attention mechanism, and finally a 768-dimensional multi-granular semantic representation of the data source terms is generated. This representation integrates the core features of each granularity level and weakens the uncertainty features. It forms a closed loop with the preceding feature extraction, deconstruction fusion, and contrast enhancement steps, providing an accurate and highly recognizable feature foundation for the semantic matching of subsequent term mapping. This embodiment realizes the deep extraction and fusion of multi-granularity features of data source terms, enhances feature discrimination, and the generated semantic representation provides high-dimensional and high-quality feature support for accurate term mapping.

[0035] In one embodiment, the steps of extracting performance indicator caliber information based on the terminology mapping table, reverse-engineering caliber information that is not explicitly recorded, detecting caliber differences and resolving conflicts using a preset strategy, and establishing a caliber standardization library include: S30: Extract performance indicator caliber information based on the terminology mapping table, construct a caliber causal graph model, identify causal relationships in dimensions such as caliber definition and calculation logic, and generate a caliber causal graph containing causal nodes and directed causal edges. S31: Construct a caliber knowledge graph based on caliber information, generate semantic representations of unrecorded calibers through graph reasoning, retrieve similar caliber definitions, and infer unrecorded caliber information by combining spatiotemporal evolution laws; S32: Use time-series modeling to capture dynamic trends in caliber changes, reverse-engineer caliber information for unrecorded time points, identify abrupt changes in caliber definition, generate a report on the spatiotemporal distribution of caliber differences, and quantify the uncertainty of caliber characteristics. S33: Based on various detection and inference results, a preset strategy is adopted to resolve the conflict of caliber differences, and the standardized caliber information is organized to establish a caliber standardization library.

[0036] In this embodiment, firstly, performance indicator caliber information is extracted based on the terminology mapping table, and a caliber causal graph model is constructed. Causal relationships in dimensions such as caliber definition and calculation logic are identified, and a caliber causal graph containing causal nodes and directed causal edges is generated. Based on the previously generated terminology mapping table, the caliber information corresponding to the unified term "core performance indicators" of the HR, CRM, and MES systems is extracted. The HR system records the calculation logic as "monthly completion amount / target amount" with no time range caliber. The CRM system records the time range as "natural month" with default calculation logic. The MES system records the time range as "production month" with calculation logic as "daily completion amount accumulation / monthly target amount". Based on this information, a caliber causal graph model is constructed, and caliber parameters, calculation formulas, source data tables, and time windows are set as four types of causal nodes. Directed causal edges represent the influence relationship between each node. For example, the "time window" node points to the "calculation formula" node, clarifying the direct causal impact of the time range on the calculation logic. A caliber causal graph containing 12 nodes and 18 directed edges is generated, clearly defining the correlation logic of each caliber dimension. Then, based on the caliber information, a caliber knowledge graph is constructed. Semantic representations of unrecorded calibers are generated through graph reasoning. Similar caliber definitions are retrieved, and the unrecorded caliber information is inferred by combining spatiotemporal evolution patterns. Calibration terms, calculation logic, data sources, time ranges, and business scenarios are set as graph nodes. Four types of edge relationships—definition, dependency, association, and constraint—are defined to construct a caliber knowledge graph for core performance indicators. A two-layer graph attention network is used to reason on the graph, generating semantic representations of the time range missing in the HR system and the calculation logic missing in the CRM system. The most similar caliber definitions are retrieved from the graph. Combining the spatiotemporal evolution patterns of enterprise performance indicator calibers, the time range of the HR system is inferred to be "natural month," and the calculation logic of the CRM system is consistent with that of the HR system as "monthly completion / target amount," thus completing the unrecorded caliber information. Secondly, a time-series... The model captures the dynamic change trend of the caliber, reverse-engineers the caliber information for unrecorded time points, identifies the abrupt change points in the caliber definition, generates a report on the spatiotemporal distribution of caliber differences, and quantifies the uncertainty of caliber characteristics. The temporal graph neural network (TGN) is used to model the spatiotemporal evolution of the caliber, setting a 128-dimensional embedding vector and a time window of 6 time units to capture the dynamic change trend of the caliber of core performance indicators over time, and verify the rationality of the reverse-engineered caliber information. At the same time, the CUSUM algorithm is used to identify the abrupt change points in the caliber definition of the MES system's "production month" and the enterprise's general "natural month", generating a spatiotemporal distribution report containing the location, time of abrupt change, and degree of difference of the caliber of each system. Then, the uncertainty of the caliber characteristics is quantified through 50 Monte Carlo Dropout samplings, resulting in an uncertainty score of 0.23 for the MES system's time range caliber, providing a basis for conflict resolution. Finally, based on various detection and inference results, a pre-set strategy is adopted to resolve discrepancies in statistical definitions. This involves using a business scenario-adapted strategy to resolve these discrepancies. The standardized definition information is then organized to establish a definition standardization library. For conflicts arising from inconsistencies between the MES system's time range and general standards, a business scenario adaptation strategy is employed. Combined with the enterprise's overall performance statistics standards, the MES system's time range is unified to "natural month," and the calculation logic is simultaneously standardized to "monthly completed amount / target amount." The core performance indicator definitions for the three systems are uniformly calibrated. The standardized definition, calculation logic, data source, time range, and applicable scenarios are organized according to a unified standard and stored in the definition standardization library. This library, along with the preceding terminology mapping table and the definition cause-effect graph, forms a closed loop, providing a unified basis for indicator definitions for subsequent data fusion and ensuring consistency in cross-system performance data calculations. This embodiment achieves the completion of performance indicator definitions, discrepancy detection, and conflict resolution, establishing a standardized definition library and laying a unified foundation for indicator calculations in cross-system data fusion.

[0037] In one embodiment, the steps of using time-series modeling to capture the dynamic change trend of caliber, inferring caliber information at unrecorded time points, identifying abrupt changes in caliber definition, generating a report on the spatiotemporal distribution of caliber differences, and quantifying the uncertainty of caliber characteristics include: S320: Construct a distributed spatiotemporal semantic graph of calibers, defining elements such as caliber definitions and calculation logic as graph nodes, and semantic relationships as graph edges. Use a word vector model to convert caliber terms into vector representations. S321: A federated temporal graph neural network combined with a multi-head attention mechanism is used to model the semantic graph. The model parameters of each node are aggregated through a federated learning algorithm to capture the semantic association strength and temporal evolution features of the caliber and generate a federated caliber embedding vector. S322: Monte Carlo Dropout technique is used to quantify the uncertainty of the federated model, generate relevant feature values ​​of the caliber embedding vector and calculate the uncertainty score, and generate uncertainty-aware federated caliber embedding vector. S323: Use the embedded vector to infer the caliber information of unrecorded time points, retrieve similar caliber definitions between data source nodes, identify caliber definition mutation points and generate a report on the spatiotemporal distribution of caliber differences, store the results in the federated caliber evolution knowledge base and update the model parameters of each node.

[0038] In this embodiment, firstly, based on the spatiotemporal feature extension of the caliber knowledge graph, a distributed caliber spatiotemporal semantic graph is constructed. Elements such as caliber definitions and calculation logic are defined as graph nodes, and semantic relationships are defined as graph edges. A word vector model is used to convert caliber terms into vector representations. Using HR, CRM, and MES as the three data source nodes as the distributed deployment entities, each node constructs its own local caliber spatiotemporal semantic graph. The caliber definition, calculation logic, data source, and time range of "core performance indicators" are set as four types of graph nodes. Node types are categorized as terminology, logic, source, and scenario. Synonyms, associations, constraints, and dependencies are set as semantic edges. Each local graph contains 80 nodes and 200 edges. A word vector model is used to convert all caliber terms into 128-dimensional vector representations. Simultaneously, timestamps are added to each node, and differential privacy technology with ε=0.8 is employed. To protect the original local caliber data, a distributed spatiotemporal semantic graph foundation is formed. Then, a federated temporal graph neural network combined with a multi-head attention mechanism is used to model the semantic graph. The model parameters of each node are aggregated through a federated learning algorithm to capture the semantic association strength and temporal evolution features of calibers, generating a federated caliber embedding vector. A 3-layer federated temporal graph neural network (TGN) is deployed on three data source nodes, with an embedding dimension of 384, a 6-head multi-head attention mechanism, and a time window of 8 time units. The semantic association strength between caliber terms at each node is captured through multi-head attention, and the temporal evolution law of caliber semantics is captured by time encoding. The FedAvg federated learning algorithm is used to aggregate the model parameters of each node, with 20 aggregation rounds and a learning rate of 0.005. After multiple rounds of training, a 256-dimensional federated caliber embedding vector is generated, which integrates the spatiotemporal features of calibers at each node. Secondly, Monte Carlo Dropout technology is used to quantify the uncertainty of the federated model, generate relevant feature values ​​of the caliber embedding vector and calculate the uncertainty score, and generate the uncertainty-aware federated caliber embedding vector. Monte Carlo Dropout sampling is performed on the federated TGN model 100 times with a dropout rate of 0.3 to generate the mean and variance of the federated caliber embedding vector. Uncertainty scores at three granularities—word level, character level, and sub-word level—are calculated, with a score range of 0-1. The time range caliber score for the HR system is 0.12, the computational logic score for the CRM system is 0.09, and the time range caliber score for the MES system is 0.25. Through an uncertainty-weighted attention mechanism, higher weights are given to low-uncertainty features, mapping the 256-dimensional federated embedding vector to a 768-dimensional uncertainty-aware federated caliber embedding vector. Finally, the embedding vector is used to infer the caliber information for unrecorded time points, retrieve similar caliber definitions between data source nodes, identify abrupt changes in caliber definitions, and generate a report on the spatiotemporal distribution of caliber differences. The results are stored in the federated caliber evolution knowledge base, and the model parameters of each node are updated. The unrecorded time points of the missing time range in the HR system and the missing calculation logic in the CRM system are used as query times. Through the uncertainty-aware federated caliber embedding vector, the most similar caliber definition is retrieved between data source nodes with a similarity threshold of 0.8, and the time range of the HR system is inferred to be "natural month". The CRM system's calculation logic is "monthly completed amount / target amount." Simultaneously, the CUSUM algorithm is used to identify abrupt changes in the definitions of the MES system's "production month" and the general "natural month," clarifying the timing, dimensions, and extent of these changes. A report on the spatiotemporal distribution of these differences is generated, and the back-calculation results and report information are stored in the federated definition evolution knowledge base. The TGN model parameters of each data source node are updated synchronously, forming a closed loop with the preceding distributed semantic graph construction, federated modeling, and uncertainty quantification stages. This provides accurate spatiotemporal characteristics and difference evidence for subsequent definition conflict resolution. This embodiment achieves spatiotemporal feature modeling and uncertainty quantification of distributed definitions, accurately back-calculates unrecorded definitions, and identifies abrupt changes, providing a spatiotemporal dimension difference analysis basis for definition standardization.

[0039] In one embodiment, the steps of using Monte Carlo Dropout technology to quantify the uncertainty of the federated model, generating relevant feature values ​​of the caliber embedding vector and calculating the uncertainty score to generate the uncertainty-aware federated caliber embedding vector include: S3220: Construct local spatiotemporal caliber maps at each data source node, use spatiotemporal graph neural networks to complete local modeling, generate local spatiotemporal caliber embedding vectors, and use differential privacy technology to protect local original data; S3221: Construct multiple spatiotemporal graph neural network models to form a federated ensemble, use federated variational Dropout ensemble technology to perform distributed spatiotemporal uncertainty quantification, generate local spatiotemporal embedding vector samples and calculate their relevant feature values; S3222: Aggregate the feature values ​​of each node through a secure aggregation protocol to generate global feature data, calculate multi-granularity uncertainty scores based on the global feature data, and integrate them to generate a multi-dimensional uncertainty score vector; S3223: Based on the uncertainty scoring vector, an adaptive weighted voting mechanism is constructed to perform weighted voting on the back-inference results of each model, generating an uncertainty-aware federated caliber embedding vector.

[0040] In this embodiment, a local spatiotemporal caliber graph is first constructed at each data source node. A spatiotemporal graph neural network is used for local modeling, generating local spatiotemporal caliber embedding vectors. Differential privacy technology is employed to protect the original local data. Using HR, CRM, and MES distributed data source nodes as the main components, each node constructs its own local spatiotemporal caliber graph. Each graph contains 60 caliber-related nodes and 150 semantic edges. Timestamps are added to each node to represent temporal features. A three-layer spatiotemporal graph neural network (ST-GCN) is used for modeling, with an embedding dimension of 256. Dropout layers (rate 0.3) are integrated into the hidden and output layers to generate local spatiotemporal caliber embedding vectors for each of the three nodes. Simultaneously, differential privacy technology with ε=0.8 is used to encrypt the original local caliber data to ensure data privacy. Security; then, multiple spatiotemporal graph neural network models are constructed to form a federated ensemble. Federated variational Dropout ensemble technology is used to quantify distributed spatiotemporal uncertainty, generate local spatiotemporal embedding vector samples and calculate their relevant feature values. Eight ST-GCN models with different structures are deployed at each data source node to form a federated ensemble. Joint spatiotemporal sampling is performed on the core performance indicator data at 12 time points. Each model performs 150 forward propagation samplings. Combined with adaptive Dropout distribution parameters μ and σ, each node generates 1200 local spatiotemporal embedding vector samples. The importance scores of four types of features in the samples are calculated: statistical, semantic, graph structure, and temporal. Features are filtered according to a threshold of 0.3 pruning and 0.7 retention, and local mean vector and local local variance vector are generated with feature importance weighting. Secondly, feature values ​​from each node are aggregated using a secure aggregation protocol to generate global feature data. Multi-granularity uncertainty scores are then calculated based on this global feature data, and a multi-dimensional uncertainty score vector is generated. The mean and variance vectors of the three nodes are encrypted and aggregated using the secure aggregation protocol. High-importance features (≥0.7) are aggregated based on feature importance weighting, while low-importance features (≤0.3) undergo standard aggregation, generating a 256-dimensional global mean vector, global variance vector, and spatiotemporal coupling covariance matrix. Uncertainty scores at the word, character, and sub-word levels (time, space, and spatiotemporal coupling dimensions) are calculated based on the global data. Inter-model uncertainty scores are then calculated, and a 10-dimensional uncertainty score vector is generated, with scores ranging from 0 to 1. Finally, an adaptive weighted voting system is constructed based on the uncertainty score vector. The mechanism employs a weighted voting system to calculate the back-calculation results of each model, generating an uncertainty-aware federated caliber embedding vector. An uncertainty-weighted attention mechanism, inversely proportional to the 10-dimensional uncertainty score with a total weight of 1, is constructed. This mechanism is used to calculate the back-calculation results of eight ST-GCN models, with a voting confidence threshold of 0.7. In this embodiment, the confidence level of the back-calculation results related to core performance indicators reaches 0.89, so the voting results are directly adopted. The dimensionality of the voted feature vectors is mapped and fused to generate a 768-dimensional uncertainty-aware federated caliber embedding vector. This vector integrates the core spatiotemporal features of each node and model, while mitigating high-uncertainty features. It forms a closed loop with the preceding local modeling, federated integration quantization, and global feature aggregation stages, providing accurate feature support for subsequent caliber back-calculation and mutation point identification. This embodiment achieves refined uncertainty quantification of the federated model, improves the reliability of the feature vector through weighted voting, and provides high-quality uncertainty-aware features for accurate caliber back-calculation.

[0041] In one embodiment, the steps of constructing multiple spatiotemporal graph neural network models to form a federated ensemble, employing federated variational Dropout ensemble technology for distributed spatiotemporal uncertainty quantification, generating local spatiotemporal embedding vector samples, and calculating their relevant feature values ​​include: S32210: Construct local caliber maps at each data source node, use spatiotemporal graph neural networks to complete local modeling and configure adaptive variational Dropout layers to generate local spatiotemporal feature embedding vectors, and use differential privacy technology to protect the original node data; S32211: Deploy multiple spatiotemporal graph neural network models at each data source node to form a federated integration, perform joint spatiotemporal sampling on the caliber data, and generate local spatiotemporal embedding vector samples by multiple forward propagation sampling and combining adaptive Dropout distribution parameters. S32212: Calculate the importance scores of various features for the sample data, construct feature selection criteria based on the scores, and generate a local mean vector and a local variance vector weighted by feature importance; S32213: Distributed spatiotemporal uncertainty quantification is carried out based on federated variational Dropout integration technology. Statistical information of each node is aggregated through a secure aggregation protocol, and different importance features are classified and aggregated to generate a global mean vector, a global variance vector, and a spatiotemporal coupling covariance matrix.

[0042] In this embodiment, a local caliber graph is first constructed at each data source node. A spatiotemporal graph neural network is used to complete local modeling and an adaptive variational Dropout layer is configured to generate local spatiotemporal feature embedding vectors. Differential privacy technology is used to protect the original data of the nodes. With HR, CRM, and MES as the main processing nodes, each node constructs a local caliber graph around the "core performance indicators" caliber. Each graph contains 80 caliber nodes and 200 semantic edges. The nodes cover four types of features: statistical, semantic, graph structure, and time series. A 3-layer spatiotemporal graph neural network ST-GCN is used for modeling, with the embedding dimension set to 256 dimensions. An adaptive variational Dropout layer is integrated in the hidden layer and the output layer. The Dropout rate is dynamically adjusted from 0.1 to 0.5 according to the complexity of the caliber data. Each node generates a local spatiotemporal feature embedding vector. At the same time, differential privacy technology with ε=0.8 is used to protect the original caliber data of the nodes, taking into account both modeling needs and data security. Then, multiple spatiotemporal graph neural network models are deployed at each data source node to form a federated ensemble, performing joint spatiotemporal sampling on the caliber data. Through multiple forward propagation samplings combined with adaptive Dropout distribution parameters, local spatiotemporal embedding vector samples are generated. Eight structurally differentiated ST-GCN models are deployed at each data source node to form a federated ensemble, conducting joint spatiotemporal sampling on the caliber data for 18 time points of the core performance indicators. Each model performs 120 forward propagation samplings, dynamically adjusting the sampling strategy using adaptive Dropout distribution parameters μ and σ. Each node ultimately generates 960 local spatiotemporal embedding vector samples, fully covering the spatiotemporal feature distribution of the caliber data; secondly... The importance scores of various features in the sample data are calculated. Based on the scores, feature selection criteria are constructed, and local mean vectors and local variance vectors weighted by feature importance are generated. For the local spatiotemporal embedding vector samples of each node, the importance scores of four types of features—statistical, semantic, graph structure, and temporal—are calculated, with a score range of 0-1. Among them, the importance scores of temporal and semantic features are ≥0.7, statistical features are ≥0.5, and graph structure features are ≤0.3. Based on this, feature selection criteria are constructed: features below 0.3 are pruned, and features above 0.7 are retained. High-importance features that are retained are given higher weights. A 256-dimensional local mean vector and local variance vector weighted by feature importance are calculated and generated to weaken the influence of low-importance features. Finally, distributed spatiotemporal uncertainty quantification was carried out based on federated variational Dropout integration technology. Statistical information from each node was aggregated through a secure aggregation protocol, and features of different importance were classified and aggregated to generate a global mean vector, a global variance vector, and a spatiotemporal coupling covariance matrix. Distributed spatiotemporal uncertainty quantification was performed on the data from the three nodes using federated variational Dropout integration technology. The mean and variance vectors of each node were encrypted and aggregated using a secure aggregation protocol. High-importance features (≥0.7) were aggregated according to feature importance weights, low-importance features (≤0.3) were aggregated using standard methods, and medium-importance features (0.3-0.7) were aggregated using mean. Finally, a 256-dimensional global mean vector, a 256-dimensional global variance vector, and a 256×256-dimensional spatiotemporal coupling covariance matrix were generated, which fully characterizes the global spatiotemporal uncertainty features of the core performance indicator. This result forms a closed loop with the preceding local modeling, sample generation, and feature scoring stages, providing an accurate global statistical foundation for subsequent multi-granularity uncertainty scoring calculations. This embodiment realizes the quantification and feature selection of spatiotemporal uncertainty of distributed data, and generates global feature data through classification and aggregation, providing accurate support for uncertainty scoring.

[0043] In one embodiment, the step of combining the data source classification labels, terminology mapping table, and standardization library to orchestrate and execute a data fusion pipeline to generate a performance data warehouse and quality data includes: S40: Read the data source classification labels, identify real-time data sources and batch processing data sources, build a domain ontology library and associate it with a terminology mapping table and a standardization library, and generate a stream-batch classification knowledge graph; S41: Using a predictive model to input semantic features of the knowledge graph and features of the data source, generate a batch-stream collaborative orchestration scheme and pipeline DAG topology, orchestrate the pipeline and design collaborative nodes, and generate a batch-stream collaborative execution plan; S42: Dynamically orchestrate parallel processing pipelines according to the execution plan, process streaming data and batch data respectively to generate corresponding data tables, complete terminology mapping and standardization based on knowledge graph, and generate fusion result tables by merging collaborative nodes; S43: Load the fused data into the Lambda architecture data warehouse, generate fact tables and dimension tables, perform multi-dimensional quality checks and consistency assessments on the fused data, and generate a performance data warehouse and corresponding quality data.

[0044] In this embodiment, the data source classification tags are first read to identify real-time data sources and batch processing data sources. A domain ontology is constructed and associated with a terminology mapping table and a standardization library. A batch-processing classification knowledge graph is generated. The classification tags generated by S1 for HR, CRM, and MES systems are read, and the HR system is determined to be a high-temporal-series heterogeneous batch processing data source, while the CRM and MES systems are real-time data sources. A domain ontology containing 500 concepts and 2000 relationships is constructed around enterprise performance analysis. The terminology mapping table of S2 and the standardization library of S3 are associated with the ontology. The data source classification features, terminology standards, and standardization specifications are integrated to generate a batch-processing classification knowledge graph containing 2000 nodes and 8000 edges. The embedding dimension is set to 256 dimensions. A knowledge support system for stream-batch processing is established. Then, a predictive model is used to input semantic features of the knowledge graph and features of the data source to generate a stream-batch collaborative orchestration scheme and pipeline DAG topology. The pipeline is orchestrated and collaborative nodes are designed to generate a stream-batch collaborative execution plan. The semantic features of the knowledge graph and 50-dimensional features of the data source are extracted and input into the XGBoost predictive model (accuracy ≥ 90%). The model outputs a stream-batch collaborative orchestration scheme, generating a pipeline DAG topology containing a real-time subgraph and a batch processing subgraph. The real-time subgraph has 3 pipelines and the batch processing subgraph has 6 pipelines, totaling 100 nodes and 300 edges. 15 stream-batch collaborative nodes are designed, and a time window strategy of 5 minutes sliding and 30 minutes rolling is defined to form a complete stream-batch collaborative execution plan. Secondly, parallel processing pipelines are dynamically orchestrated according to the execution plan to process streaming and batch data separately to generate corresponding data tables. Terminology mapping and standardization are completed based on knowledge graphs. A fusion result table is generated by merging through collaborative nodes. Six parallel processing pipelines are dynamically orchestrated: two process real-time CRM and MES data (processing 5,000 records per second) to generate 10 real-time data tables, and four process HR batch data (processing 2GB per hour) to generate 30 batch data tables. Cross-system terminology mapping and standardization are automatically completed based on the streaming and batch classification knowledge graph (accuracy ≥95%). Real-time and batch data are merged by time and indicator dimensions through collaborative nodes to generate 20 core performance indicator fusion result tables. An incremental update strategy (5% increment / day) is adopted to ensure data timeliness. Finally, the fused data is loaded into a Lambda-based data warehouse, generating fact tables and dimension tables. Multi-dimensional quality checks and consistency assessments are performed on the fused data, generating a performance data warehouse and corresponding quality data. The fusion result table is loaded into the Lambda-based data warehouse. The real-time layer is configured with 1.5TB of storage for the most recent 7 days of data, and the batch processing layer is configured with 20TB of storage for all historical data, generating 15 performance fact tables and 30 dimension tables. Quality checks are performed on four dimensions: completeness, accuracy, consistency, and timeliness, using a 0-1 scoring system. The overall score is ≥0.88, and a Kappa score ≥0.85 is achieved for both stream and batch consistency. The quality data and performance data are then uniformly stored in the performance data warehouse. This warehouse forms a closed loop with the preceding stream and batch classification, pipeline orchestration, and data fusion stages, providing complete data support for subsequent fusion effect evaluation and strategy optimization. This embodiment achieves automated data fusion integrating stream and batch processing, constructs a standardized performance data warehouse, ensures the quality of fused data, and provides a unified data foundation for performance analysis.

[0045] In one embodiment, the steps of using a predictive model to input semantic features of a knowledge graph and features of a data source to generate a batch-stream collaborative orchestration scheme and a pipeline DAG topology, orchestrating the pipeline and designing collaborative nodes, and generating a batch-stream collaborative execution plan include: S410: Read the semantic features of the knowledge graph and the temporal features of the data source, construct a multi-head attention fusion network, calculate the weight matrix by scaling dot product attention, generate fusion feature vectors, extract temporal statistical features and identify important features to form a temporal fusion feature system; S411: A multi-task prediction model is constructed using a graph neural network. The input is fused features, and the output is orchestration mode, pipeline topology, collaborative nodes, estimated execution time and resource consumption. Gradient updates are completed using a meta-learning framework to generate pipeline DAG topology and execution configuration. S412: The time series prediction model is used to input the time series sequence, predict future orchestration requirements, dynamically adjust the orchestration scheme, generate a dynamic orchestration scheme, generate an interpretable report based on feature importance, perform stream-batch collaborative processing and record execution logs. S413: Based on the results of multi-task prediction and time-series prediction, integrate orchestration strategies and execution parameters, construct a pipeline DAG topology, design flow-batch collaborative nodes, and form a flow-batch collaborative execution plan.

[0046] In this embodiment, firstly, the semantic features of the knowledge graph and the temporal features of the data source are read, and a multi-head attention fusion network is constructed. The weight matrix is ​​calculated by scaling dot product attention to generate a fusion feature vector. Temporal statistical features are extracted and important features are identified to form a temporal fusion feature system. The 256-dimensional semantic features of the batch classification knowledge graph and the 80-dimensional temporal features of the HR, CRM, and MES data sources are read, and an 8-head multi-head attention fusion network is constructed. The weight matrix between features is calculated by scaling dot product attention, and the semantic and temporal features are weighted and fused to generate a 512-dimensional fusion feature vector. Then, 30-dimensional temporal statistical features such as trend, periodicity, and seasonality are extracted. The feature importance algorithm identifies 20 Top-K important features such as data update frequency and data volume scale to form an attention-enhanced temporal fusion feature system. Then, a multi-task prediction model is constructed using a graph neural network. Inputting fused features, it outputs orchestration patterns, pipeline topology, collaborating nodes, estimated execution time, and resource consumption. A meta-learning framework is used to complete gradient updates, generating a pipeline DAG topology and execution configuration. A three-layer graph neural network multi-task prediction model is built, inputting 512-dimensional fused features into the model, and outputting results such as the hybrid orchestration pattern of ELT+CDC, the range of pipeline node / edge counts, and the number of collaborating nodes. Simultaneously, it estimates the CRM / MES real-time pipeline execution latency to <10 seconds, the HR batch processing pipeline window to 24 hours, and CPU resource allocation to 8-32 cores. The MAML meta-learning framework is used to complete a three-step gradient update on a support set of 100 performance data fusion cases, quickly adapting and generating a model containing 80 nodes. The pipeline consists of a 280-edge DAG topology and initial execution configuration. A time-series prediction model is then used to input the time series sequence, predict future orchestration requirements, dynamically adjust the orchestration scheme, generate a dynamic orchestration scheme, generate an interpretable report based on feature importance, perform batch-stream collaborative processing, and record execution logs. A 4-layer LSTM time-series prediction model is used, inputting the data source time series sequences from the most recent 60 time points, to predict orchestration requirements for the next 8 time points. It is predicted that the MES system data volume will increase by 30%, and the orchestration scheme is dynamically adjusted accordingly, adding 2 processing nodes to the MES real-time pipeline. An interpretable report is generated based on feature importance, clarifying that the contribution of data volume features to the orchestration scheme reaches 0.82. Simultaneously, batch-stream collaborative processing is performed, and full logs such as node execution status and resource utilization are recorded. Finally, based on the results of multi-task prediction and time-series prediction, orchestration strategies and execution parameters are integrated to construct a pipeline DAG topology, design batch-stream collaboration nodes, and form a batch-stream collaboration execution plan. The basic orchestration strategy of multi-task prediction and the dynamic adjustment parameters of time-series prediction are integrated to optimize and generate the final pipeline DAG topology. Twenty batch-stream collaboration nodes are designed according to the fusion requirements of real-time / batch processing data, divided into five functional modules: data reception, terminology mapping, calibration, data merging, and quality inspection. A time window strategy of 3 minutes sliding and 40 minutes rolling, and a rule of incremental updates of 5% per day are defined. The resource allocation, execution order, and data interaction methods of each node are clarified, forming a complete batch-stream collaboration execution plan. This plan forms a closed loop with the preceding feature fusion, model prediction, and demand prediction stages, providing accurate and implementable operational basis for the subsequent execution of the data fusion pipeline. This embodiment realizes intelligent orchestration and dynamic adjustment of the fusion pipeline, generates an interpretable execution plan, and ensures the efficiency and adaptability of batch-stream collaboration processing.

[0047] In one embodiment, the steps of constructing a multi-task prediction model using a graph neural network, inputting fused features, and outputting orchestration patterns, pipeline topology, cooperating nodes, estimated execution time, and resource consumption, and using a meta-learning framework to complete gradient updates, generating pipeline DAG topology and execution configuration, include: S4110: Read pipeline task data and construct task graphs, use graph convolutional networks to extract semantic features of task graphs, fuse knowledge graph features and data source features, construct multi-task fusion networks, and generate multi-dimensional fusion features; S4111: A multi-task prediction model is constructed using graph convolutional networks. The model inputs fused features and outputs orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes, generating the initial pipeline topology and configuration. S4112: Using graph generation networks based on fusion features, pipeline topology diagrams, node configuration parameters and edge configuration parameters are automatically generated, and corresponding execution configuration files are generated. S4113: Construct a graph reinforcement learning environment and train and optimize strategies. Combine the meta-learning framework to complete the gradient update and rapid adaptation of model parameters, adjust and optimize the topology and configuration, and finally generate the pipeline DAG topology and corresponding execution configuration information.

[0048] In this embodiment, pipeline task data is first read and a task graph is constructed. A graph convolutional network is used to extract semantic features of the task graph. Based on the multi-head attention fusion features, knowledge graph features and data source features are fused to construct a multi-task fusion network and generate multi-dimensional fusion features. 120 pipeline task data from the fusion of HR, CRM, and MES system performance data are read, covering data collection, terminology mapping, and calibration. A task graph with 120 nodes and 400 edges is constructed, with node features being 80-dimensional task attributes. A 3-layer graph convolutional network is used to extract 256-dimensional semantic features of the task graph. The 256-dimensional semantic features of the batch classification knowledge graph and the 80-dimensional features of the data source are fused to build a 6-head attention multi-task fusion network. Weighted fusion generates 640-dimensional multi-dimensional fusion features, providing comprehensive feature support for subsequent prediction. Then, a multi-task prediction model is constructed using a graph convolutional network. The fused features are input into the model, and the output includes orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes. An initial pipeline topology and configuration are generated. A 4-layer graph convolutional network multi-task prediction model is then built, inputting 640-dimensional fused features. The output is an ELT+stream / batch hybrid orchestration mode, with 80-100 pipeline nodes and 280-320 edges. The CRM / MES real-time pipeline is allocated 16 CPU cores and 32GB of memory, with an execution latency of <10 seconds. The HR batch pipeline is allocated 32 CPU cores and 64GB of memory, with a 24-hour window and 20 collaborative nodes. Based on the prediction results, a multi-task prediction model containing nodes is generated. The initial pipeline topology and basic configuration are defined by node type and edge connection relationship. Then, a graph generation network is used based on fusion features to automatically generate the pipeline topology diagram, node configuration parameters, and edge configuration parameters, generating corresponding execution configuration files. A 4-layer GraphRNN graph generation network is used, with 640-dimensional fusion features as input, to automatically generate a pipeline DAG topology diagram containing 90 nodes and 300 edges. Nodes are categorized into real-time processing, batch processing, and collaborative fusion types. Each node is configured with 30 parameters, including algorithm type and data processing threshold. Each edge is configured with parameters such as data flow type and latency threshold. All configuration information is generated into a JSON format execution configuration file according to specifications, directly adapting to the pipeline execution system. Finally, a graph reinforcement learning environment is constructed and an optimization strategy is trained. Combined with a meta-learning framework, model parameter gradient updates and rapid adaptation are completed. Specifically, a graph reinforcement learning environment is constructed and an optimization strategy aimed at execution efficiency, resource utilization, and data quality is trained. The meta-learning framework is used to complete model parameter gradient updates and rapid adaptation, adjusting and optimizing the topology and configuration. Ultimately, a pipelined DAG topology and corresponding execution configuration information are generated. The graph reinforcement learning environment has a 400-dimensional graph structure and node state features in its state space, and an action space containing 30 actions such as adding nodes and adjusting resources. The reward function balances execution efficiency, resource utilization, and data quality. To optimize source utilization and data quality, a 3-layer Graph-RL training optimization strategy was adopted. After 2000 training rounds, a 3-step gradient update was performed on 80 performance fusion case support sets using the MAML meta-learning framework, with a learning rate of 0.005. This rapidly adapted and optimized the initial topology, removing 5 redundant nodes and adjusting the resource allocation of 10 nodes. The final result generated a pipeline DAG topology with 85 nodes and 290 edges, along with optimized execution configuration information. This result forms a closed loop with the preceding feature fusion, model prediction, and topology generation stages, providing accurate topology and configuration basis for batch and stream collaborative execution plans. This embodiment achieves intelligent generation and optimization of pipeline topology and configuration, combining meta-learning to quickly adapt to business needs and improve orchestration accuracy and execution adaptability.

[0049] In one embodiment, the step of constructing a multi-task prediction model using a graph convolutional network, inputting fused features into the model, and outputting orchestration patterns, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes to generate an initial pipeline topology and configuration includes: S41110: Extract pipeline tasks and construct a hierarchical temporal task requirement graph. Use a hierarchical graph convolutional network to extract hierarchical semantic features, combine a temporal model to extract temporal features, fuse knowledge graph features and data source features, construct an attention fusion network, and generate a hierarchical temporal fusion feature vector. S41111: Construct a hierarchical temporal latent space graph generation multi-task model consisting of an encoder, generator and discriminator. The encoder encodes the hierarchical temporal topology graph into a latent space representation. The generator generates the topology structure based on fused features and latent space noise. The discriminator judges the authenticity of the generation and uses multi-task loss to train the model. S41112: The meta-learning algorithm is used to input the corresponding fusion features, perform gradient updates and fast adaptation on the support set, and output the orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time and prediction results of collaborative nodes. S41113: Based on the prediction results, generate an initial hierarchical time-series pipeline topology and configuration file containing multi-layer topology, node and edge configuration parameters.

[0050] In this embodiment, pipeline tasks are first extracted and a hierarchical temporal task requirement graph is constructed. A hierarchical graph convolutional network is used to extract hierarchical semantic features, and a temporal model is used to extract temporal features. Knowledge graph features and data source features are fused to construct an attention fusion network and generate a hierarchical temporal fusion feature vector. Five types of pipeline tasks for HR batch processing, CRM, and MES real-time performance data fusion—collection, mapping, calibration, fusion, and quality inspection—are extracted. A hierarchical temporal task requirement graph containing a data layer, processing layer, and fusion layer is constructed according to the data processing level. Each layer contains 30 nodes, and timestamps are added to represent temporal features. A 3-layer hierarchical graph convolutional network is used to extract 256-dimensional semantic features for each layer. An LSTM temporal model is used to extract 80-dimensional temporal features. Then, the 256-dimensional features of the batch classification knowledge graph and the 80-dimensional features of the data source are fused to build a 12-head attention fusion network. Weighted fusion is used to generate a 768-dimensional hierarchical temporal fusion feature vector, which fully represents the hierarchical and temporal attributes of the task. Then, a hierarchical temporal latent space graph generation multi-task model is constructed, consisting of an encoder, a generator, and a discriminator. The encoder encodes the hierarchical temporal topology graph into a latent space representation. The generator generates the topology structure based on fusion features and latent space noise. The discriminator judges the authenticity of the generated topology and trains the model using a multi-task loss function. A graph generation multi-task model consisting of a 4-layer encoder, a 5-layer generator, and a 3-layer discriminator is constructed. The encoder encodes the hierarchical temporal task requirement graph into a 512-dimensional latent space representation. The generator generates a pipeline topology structure that fits the performance data fusion requirements based on 768-dimensional fusion features and Gaussian latent space noise. The discriminator judges the authenticity and rationality of the generated topology. The model is trained using a multi-task loss function based on topology matching degree, resource utilization, and execution efficiency, and iterates for 500 rounds. The model achieved a prediction accuracy of 92%. Next, a meta-learning algorithm was used to input the corresponding fusion features, performing gradient updates and rapid adaptation on the support set. The output included orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results for collaborative nodes. The 768-dimensional hierarchical temporal fusion features were input into the MAML meta-learning algorithm. 100 sets of enterprise performance data fusion cases were selected as the support set, with a learning rate of 0.001. Three gradient updates were completed to achieve rapid adaptation. The model output prediction results for an ELT+CDC hybrid orchestration mode, 90 pipeline nodes, 300 edges, a CRM / MES real-time pipeline with 16 CPU cores and an execution latency of <10 seconds, and an HR batch processing pipeline with 32 CPU cores, an execution window of 24 hours, and 20 batch-stream collaborative nodes. Finally, based on the prediction results, an initial hierarchical temporal pipeline topology and configuration file containing multi-layered topology structures and node and edge configuration parameters are generated. Following the hierarchical logic of data layer, processing layer, and fusion layer, the 90 predicted nodes are divided into corresponding layers, clarifying the task type, algorithm configuration, and resource allocation for each node. The data flow direction and interaction rules for 300 edges are defined, and the execution sequence and fusion rules for collaborating nodes at each layer are annotated. Following a unified standard, an initial hierarchical temporal pipeline topology and JSON format configuration file containing multi-layered topology structures and full configuration parameters are generated. This file forms a closed loop with the preceding feature extraction, model training, and meta-learning adaptation stages, providing a foundation for subsequent topology optimization and execution. This embodiment achieves accurate prediction of hierarchical temporal pipeline topology, and combines meta-learning for rapid business adaptation, generating a standardized initial topology and configuration file.

[0051] In one embodiment, the step of evaluating the fusion effect based on the performance data warehouse and quality data, optimizing the fusion strategy based on the evaluation results, and updating the data fusion pipeline includes: S50: Extract source domain scenes from the historical optimization scene library, identify new target domain scenes, construct a feature space and use a transfer learning domain adaptive algorithm to complete feature alignment, construct a pipeline optimization causal graph, and generate a transfer training set and causal effect matrix; S51: Meta-training is performed using a meta-learning algorithm to train a transfer reinforcement learning agent, a Bayesian optimizer is constructed and a hyperparameter search space is defined, relevant models and acquisition functions are initialized, causal effect estimation is performed on samples, and counterfactual inference predictions are calculated. S52: Extract scene features from the new scene in the target domain, align them to the feature space of the source domain through a feature mapping function, input them into the transfer reinforcement learning agent to perform gradient updates to generate an initial optimization strategy, fuse Bayesian optimization suggestions and apply causal constraints to generate the optimal pipeline configuration; S53: Deploy the optimal configuration and execute the fusion task, record the execution log, calculate the actual optimization effect, verify the relevant prediction accuracy and optimization efficiency, add new scenario samples to the target domain dataset, and update various models and data fusion pipelines.

[0052] In this embodiment, source domain scenarios are first extracted from the historical optimization scenario library, new target domain scenarios are identified, a feature space is constructed, and a transfer learning domain adaptive algorithm is used to complete feature alignment. A pipeline optimization causal graph is constructed, and a transfer training set and a causal effect matrix are generated. Fifty source domain fusion optimization scenarios are extracted from the historical optimization scenario library, covering typical problems such as low consistency, high latency, and resource overload. The MES system data in the performance data warehouse output by S4 this time has low consistency (score 0.78, below the threshold of 0.85) and is set as a new target domain scenario. A 256-dimensional feature space including data source features, pipeline features, quality features, and resource features is constructed. The MMD maximum mean difference algorithm is used to align the features of the source domain and the target domain. With fusion quality, execution efficiency, and resource consumption as causal nodes and the relationship between adjustment strategy and effect as directed edges, a pipeline optimization causal graph with 30 nodes and 80 edges is constructed. A transfer training set and a 128×128-dimensional causal effect matrix are generated through causal inference to clarify the influence strength of different configurations on the fusion effect. Then, a meta-learning algorithm is used for meta-training to train the transfer reinforcement learning agent. A Bayesian optimizer is constructed and a hyperparameter search space is defined. Relevant models and acquisition functions are initialized. Causal effect estimation is performed on the samples, and counterfactual inference predictions are calculated. The MAML meta-learning algorithm is used with 50 source domain scenes as support sets for meta-training, with a learning rate of 0.001 and 1000 iterations, to train a transfer reinforcement learning agent that can be quickly adapted. A Bayesian optimizer is constructed, and a hyperparameter search space is defined for the number of nodes, CPU cores, memory, time window, parallelism, etc. A Gaussian model and a PI acquisition function are initialized. Causal effect estimation is performed on the transfer training set samples based on the causal effect matrix. The fusion quality and execution efficiency under different configurations are predicted through counterfactual inference, forming predictive prior knowledge. Secondly, scene features are extracted from the new scene in the target domain, and the features are used to... The mapping function is aligned to the source domain feature space, input to the transfer reinforcement learning agent to perform gradient updates to generate the initial optimization strategy, and Bayesian optimization suggestions are fused and causal constraints are applied. That is, Bayesian optimization suggestions are fused and causal constraints based on the pipeline optimization causal graph and causal effect matrix are applied to generate the optimal pipeline configuration. 256-dimensional scene features of scenarios with low consistency in the MES system are extracted, aligned to the source domain space through the feature mapping function, input to the transfer reinforcement learning agent, and three-step gradient updates are performed to quickly generate the initial optimization strategy. Combined with the hyperparameter suggestions given by the Bayesian optimizer, causal constraints are applied according to the causal effect matrix to eliminate high-risk and low-return configurations. The transfer learning weight is set to 0.7 and the Bayesian weight is set to 0.3 for fusion to generate the optimal configuration: the number of real-time pipeline nodes is increased by 2, the CPU is increased to 8 cores, and the time window is shortened to 5 minutes. Finally, the optimal configuration is deployed and the fusion task is executed. Execution logs are recorded, and the actual optimization effect is calculated, specifically the actual optimization effect in terms of fused data quality and execution efficiency. The relevant prediction accuracy and optimization efficiency are verified. New scenario samples are added to the target domain dataset, and various models and data fusion pipelines are updated. The optimal configuration is deployed again to re-execute performance data fusion, recording execution logs, resource usage, and quality scores in real time. Actual results show that MES data consistency improved to 0.92, meeting the standard, prediction accuracy reached 91%, and optimization efficiency improved by 40%. The new scenario sample and effect data are added to the target domain dataset, expanding the historical optimization scenario library. The transfer reinforcement learning agent, Bayesian optimizer, and data fusion pipeline topology and parameters are updated synchronously, forming a complete closed loop with the preceding S1-S4 processes, achieving continuous self-optimization of performance data fusion. This embodiment achieves intelligent optimization of the fusion strategy through transfer learning and reinforcement learning, continuously improving data quality and efficiency, forming a complete self-optimization closed loop.

[0053] In one embodiment, the steps of employing a meta-learning algorithm for meta-training, training a transfer reinforcement learning agent, constructing a Bayesian optimizer and defining a hyperparameter search space, initializing the relevant model and acquisition function, performing causal effect estimation on samples, and calculating counterfactual inference predictions include: S510: Extract configuration parameters and performance indicators from historical optimization scenarios in the source domain, generate causal graphs and extract causal feature vectors using causal discovery algorithms, collect new scenarios in the target domain and calculate domain differences, learn domain-invariant feature mappings through a domain adaptive network, and generate a causal effect matrix. S511: The reinforcement learning agent is pre-trained using a meta-learning framework. The state space and action space are set, and the causal effect and performance index are integrated to construct the reward function. Meta-training is used to learn source domain optimization meta-knowledge to obtain the transfer reinforcement learning agent. S512: Construct a Bayesian optimizer, define the hyperparameter search space, initialize the Gaussian process model, select the acquisition function, and generate an initial hyperparameter sample set using a sampling method; S513: Perform causal effect estimation and matching operations on hyperparameter samples, calculate counterfactual inference predictions based on the estimation results, complete the training of the Bayesian optimizer, and form an optimization model that combines transfer reinforcement and Bayesian optimization.

[0054] In this embodiment, configuration parameters and performance indicators are first extracted from historical optimization scenarios in the source domain. A causal discovery algorithm is used to generate a causal graph and extract causal feature vectors. New scenarios in the target domain are collected and domain differences are calculated. A domain-adaptive network is used to learn domain-invariant feature mappings to generate a causal effect matrix. Using 50 historical performance data fusion optimization scenarios as the source domain, configuration parameters such as the number of nodes, the number of CPU cores, the time window, and the parallelism are extracted, along with performance indicators such as consistency, latency, and resource utilization. A PC causal discovery algorithm is used to construct a causal graph containing configuration, performance, and results, and a 128-dimensional causal feature vector is extracted. The new scenario with low data consistency in the current MES system is taken as the target domain. The inter-domain distribution differences are calculated, and a domain-adaptive network is used to... The network learns domain-invariant feature mappings, ultimately generating a 128×128 causal effect matrix to quantify the impact of different configurations on fusion quality. Then, a meta-learning framework is used to pre-train a reinforcement learning agent, setting up a state space and action space. The causal effect and performance indicators are fused to construct a reward function. Meta-training learns source domain optimization meta-knowledge to obtain a transfer reinforcement learning agent. Using the MAML meta-learning framework, with the source domain scene as the support set, a 400-dimensional state space and an action space with 36 actions are constructed. The causal effect weights and performance improvement magnitude are fused to construct a reward function. With a learning rate of 0.001 and 1000 iterations, the cross-scene optimization rules are learned to obtain a transfer reinforcement learning agent that can quickly adapt to new scenes. Secondly, a Bayesian optimizer is constructed, defining the hyperparameter search space, initializing the Gaussian process model, selecting a sampling function and generating an initial hyperparameter sample set using sampling methods, building the Bayesian optimizer, defining the range of hyperparameters such as the number of nodes, CPU, memory, time window, and number of collaborative nodes, initializing the Gaussian process model and PI sampling function, and generating 50 initial hyperparameter sample sets through Latin hypercube sampling to cover the global optimum region; finally, causal effect estimation and matching operations are performed on the hyperparameter samples, and counterfactual inference prediction is calculated based on the estimation results to complete the training of the Bayesian optimizer, forming an optimization model that integrates transfer reinforcement learning and Bayesian optimization. Based on the causal effect matrix, the samples are used to estimate the effect, matching similar historical scenarios, and predicting the fusion effect and resource consumption under different configurations through counterfactual inference. The model is iteratively updated, combining the rapid adaptation of transfer reinforcement learning with the global optimization of Bayesian optimization to form an integrated optimization model. This model forms a closed loop with the preceding causal modeling, meta-training, and Bayesian initialization steps, providing accurate and efficient optimization capabilities for subsequent optimal configuration generation. This embodiment achieves rapid adaptation and accurate optimization by training an intelligent optimization model through causal modeling and meta-learning, providing reliable decision support for pipeline optimization.

[0055] refer to Figure 2 A performance data intelligent analysis device for enterprise management, comprising: The tag generation module 100 is used to identify heterogeneous characteristics and generate data source classification tags for multiple heterogeneous performance data sources, establish a data source registration center to collect metadata and business semantic information, and generate a data source registry. The terminology mapping module 200 is used to construct a performance terminology ontology based on the data source registry, and to map data source terms to standard terms using a semantic matching algorithm to generate a terminology mapping relationship table. The caliber processing module 300 is used to extract performance indicator caliber information based on the terminology mapping relationship table, reverse-engineer caliber information that is not clearly recorded, detect caliber differences and resolve conflicts using preset strategies, and establish a caliber standardization library. The fusion execution module 400 is used to combine the data source classification labels, terminology mapping relationship table and standardization library to orchestrate and execute the data fusion pipeline to generate a performance data warehouse and quality data. The optimization and update module 500 is used to evaluate the fusion effect based on the performance data warehouse and quality data, optimize the fusion strategy and update the data fusion pipeline according to the evaluation results.

[0056] Furthermore, the aforementioned label generation module 100 includes: The heterogeneous feature extraction and quantization unit is used to extract structural, syntactic, semantic, and temporal heterogeneous features from multiple heterogeneous performance data sources, calculate and quantify the heterogeneity intensity value, and generate a heterogeneous feature vector matrix. The feature map construction and label generation unit is used to construct heterogeneous feature maps based on heterogeneous feature vector matrices, generate node embedding vectors using graph attention networks, classify data sources using spectral clustering algorithms to generate classification labels, and merge labels and vectors to generate a data source classification result table. The registration center construction and information collection unit is used to construct the data source registration center based on the data source classification result table, match connector templates to connect various data sources, collect metadata information, extract business semantic information through natural language processing technology, and store the two types of information in the data source registry. The dynamic update and closed-loop maintenance unit is used to deploy a heterogeneous feature map update engine based on the data source registry, monitor metadata changes and rebuild the heterogeneous feature map, update node embedding vectors and data source classification labels, re-collect metadata and business semantic information, and update the data source registry.

[0057] Furthermore, the aforementioned heterogeneous feature extraction and quantization unit includes: The multi-dimensional heterogeneous feature extraction unit is used to perform deep traversal of multiple heterogeneous performance data sources, extract structural, syntactic, semantic, and temporal heterogeneous features respectively, and clarify the specific representation content corresponding to each type of heterogeneous feature; The heterogeneous feature quantization calculation unit is used to quantify the extracted heterogeneous features to obtain the heterogeneity intensity value corresponding to each type of feature, thus forming the feature intensity quantization result. The feature data integration and sorting unit is used to integrate the names, types, feature values ​​and corresponding intensity quantification results of various heterogeneous features, and sort out the correlation and correspondence between feature data; The feature vector matrix generation unit is used to organize data according to the corresponding dimensions of the data source and features based on the sorted feature data association relationship, and generate a heterogeneous feature vector matrix.

[0058] Furthermore, the terminology mapping module 200 mentioned above includes: The performance terminology ontology construction unit is used to construct a performance terminology ontology containing a terminology layer, a relationship layer, and an attribute layer based on metadata information in the data source registry, generate terminology node embedding vectors, and store the performance terminology ontology and vector information into the terminology ontology library. The term semantic matching and mapping unit is used to map data source terms to standard terms based on the term node embedding vectors in the term ontology library, using a semantic matching algorithm to determine the mapping relationship between the two, and organize and save the mapping relationship to generate a term mapping relationship table. The terminology mapping quality assessment unit is used to conduct quality assessments on terminology mapping relationships based on the terminology mapping relationship table, generate a terminology mapping quality report, and mark low-confidence mapping relationships as mappings to be reviewed. The mapping relationship review and update unit is used to review and correct low-reliability mapping relationships based on the terminology mapping quality report, update the corrected mapping relationships to the terminology mapping relationship table, and update the terminology ontology database synchronously.

[0059] Furthermore, the aforementioned term semantic matching and mapping unit includes: The multi-granularity semantic feature extraction and fusion unit is used to extract multi-granularity semantic features of data source terms, generate multi-dimensional embedding vectors and construct feature representation matrices, and fuse features through an attention mechanism to form a multi-granularity semantic representation of data source terms. The mapping network initialization and optimization unit is used to construct the support set and query set of the term mapping, initialize the mapping network parameters and optimize and update them based on the support set, and generate the semantic representation of the terms in the query set by combining the updated parameters. The term mapping uncertainty quantification and screening unit is used to calculate the probability distribution of term embedding vectors using a Bayesian neural network, generate relevant feature values ​​through sampling, and screen out term mapping candidates that require manual review based on uncertainty metric values. The term similarity calculation and mapping table generation unit is used to construct a term knowledge graph and generate graph-enhanced embedding vectors. It combines these vectors to calculate term similarity, determine the mapping relationship between data source terms and standard terms, integrate and save all mapping relationship information, and generate a term mapping relationship table.

[0060] Furthermore, the aforementioned multi-granularity semantic feature extraction and fusion unit includes: The multi-granularity semantic feature extraction and matrix construction unit is used to extract word-level, character-level, and sub-word-level semantic features of data source terms. It uses a pre-trained language model to generate embedding vectors for each granularity layer, constructs a multi-granularity feature representation matrix, and calculates the semantic richness index of each granularity layer. The atomic feature deconstruction and reconstruction fusion unit is used to deconstruct the features of each granularity layer into atomic feature units using an autoencoder architecture, generate an atomic feature set and reconstruct and fuse it, and generate a hierarchical multi-granularity feature vector based on the fusion result. The granularity augmentation network training and feature optimization unit is used to construct term pairs of positive and negative samples, train the granularity augmentation network and enhance the feature discrimination through contrastive learning, and generate contrast-enhanced multi-granularity representation vectors. The feature uncertainty quantification and weighted fusion unit is used to calculate the uncertainty score of features at each granularity level using a Bayesian neural network, adjust the weights of each granularity level through an uncertainty weighted attention mechanism, and generate a multi-granularity semantic representation of data source terms after weighted fusion of features.

[0061] Furthermore, the aforementioned caliber processing module 300 includes: The caliber causal graph construction unit is used to extract performance indicator caliber information based on the terminology mapping table, construct a caliber causal graph model, identify causal relationships in dimensions such as caliber definition and calculation logic, and generate a caliber causal graph containing causal nodes and directed causal edges. The Caliber Knowledge Graph Reasoning and Completion Unit is used to construct a caliber knowledge graph based on caliber information, generate semantic representations of unrecorded calibers through graph reasoning, retrieve similar caliber definitions, and infer unrecorded caliber information by combining spatiotemporal evolution laws. The caliber time-series modeling and difference detection unit is used to capture the dynamic change trend of caliber using time-series modeling, reverse the caliber information of unrecorded time points, identify abrupt changes in caliber definition, generate a report on the spatiotemporal distribution of caliber differences, and quantify the uncertainty of caliber characteristics. The unit for resolving caliber conflicts and establishing a standardization library is used to resolve caliber discrepancy conflicts based on various detection and inference results and to organize the standardized caliber information to establish a caliber standardization library.

[0062] Furthermore, the aforementioned time-series modeling and difference detection unit includes: The distributed caliber spatiotemporal semantic graph construction unit is used to construct a distributed caliber spatiotemporal semantic graph. Elements such as caliber definitions and calculation logic are defined as graph nodes, semantic relationships are defined as graph edges, and a word vector model is used to convert caliber terms into vector representations. The federated temporal graph neural network modeling unit is used to model the semantic graph using a federated temporal graph neural network combined with a multi-head attention mechanism. It aggregates the model parameters of each node through a federated learning algorithm, captures the semantic association strength and temporal evolution features of the caliber, and generates a federated caliber embedding vector. The Federation Model Uncertainty Quantization Unit is used to quantify the uncertainty of the federated model using Monte Carlo Dropout technology, generate relevant feature values ​​of the caliber embedding vector and calculate the uncertainty score, and generate the uncertainty-aware federated caliber embedding vector. The caliber back-inference mutation identification and report generation unit is used to back-infer caliber information at unrecorded time points using the embedded vector, retrieve similar caliber definitions between data source nodes, identify caliber definition mutation points and generate a report on the spatiotemporal distribution of caliber differences, store the results in the federated caliber evolution knowledge base and update the model parameters of each node.

[0063] Furthermore, the aforementioned uncertainty quantification unit of the federal model includes: The local spatiotemporal caliber graph construction and modeling unit is used to construct local spatiotemporal caliber graphs at each data source node. It uses a spatiotemporal graph neural network to complete local modeling, generates local spatiotemporal caliber embedding vectors, and uses differential privacy technology to protect the original local data. The federated integration and distributed uncertainty quantification unit is used to construct multiple spatiotemporal graph neural network models to form a federated integration. It uses federated variational Dropout ensemble technology to perform distributed spatiotemporal uncertainty quantification, generate local spatiotemporal embedded vector samples and calculate their relevant feature values. The Feature Security Aggregation and Uncertainty Scoring Unit is used to aggregate the feature values ​​of each node through a security aggregation protocol, generate global feature data, calculate multi-granularity uncertainty scores based on the global feature data, and integrate them to generate a multi-dimensional uncertainty score vector. The adaptive weighted voting and embedding vector generation unit is used to construct an adaptive weighted voting mechanism based on the uncertainty score vector, perform weighted voting on the back-inference results of each model, and generate uncertainty-aware federated caliber embedding vectors.

[0064] Furthermore, the aforementioned federal integration and distributed uncertainty quantification unit includes: The local caliber map construction and adaptive modeling unit is used to construct local caliber maps at each data source node. It uses a spatiotemporal graph neural network to complete local modeling and configures an adaptive variational Dropout layer to generate local spatiotemporal feature embedding vectors. Differential privacy technology is used to protect the original data of the nodes. The federated integration deployment and spatiotemporal sampling unit is used to deploy multiple spatiotemporal graph neural network models on each data source node to form a federated integration, perform joint spatiotemporal sampling on the caliber data, and generate local spatiotemporal embedding vector samples by multiple forward propagation sampling and combining adaptive Dropout distribution parameters. The feature importance assessment and statistical vector generation unit is used to calculate the importance scores of various features in the sample data, construct feature selection criteria based on the scores, and generate a local mean vector and a local difference vector weighted by feature importance. The distributed uncertainty quantification and global aggregation unit is used to carry out distributed spatiotemporal uncertainty quantification based on federated variational Dropout integration technology. It aggregates statistical information from each node through a secure aggregation protocol, classifies and aggregates features of different importance, and generates a global mean vector, a global variance vector, and a spatiotemporal coupled covariance matrix.

[0065] Furthermore, the aforementioned fusion execution module 400 includes: The batch processing classification knowledge graph generation unit is used to read data source classification labels, identify real-time data sources and batch processing data sources, build a domain ontology library and associate it with a terminology mapping table and a standardization library to generate a batch processing classification knowledge graph. The batch-stream collaborative execution plan generation unit is used to take the semantic features of the knowledge graph and the features of the data source as inputs to the prediction model, generate a batch-stream collaborative orchestration scheme and pipeline DAG topology, orchestrate the pipeline and design collaborative nodes, and generate a batch-stream collaborative execution plan. The streaming and batch data processing and fusion unit is used to dynamically orchestrate parallel processing pipelines according to the execution plan, process streaming data and batch data respectively to generate corresponding data tables, complete terminology mapping and standardization based on knowledge graphs, and generate fusion result tables through collaborative node merging. The performance data warehouse construction and quality assessment unit is used to load fused data into a Lambda architecture data warehouse, generate fact tables and dimension tables, perform multi-dimensional quality checks and consistency assessments on the fused data, and generate a performance data warehouse and corresponding quality data.

[0066] Furthermore, the aforementioned batch collaborative execution plan generation unit includes: The temporal fusion feature system construction unit is used to read the semantic features of the knowledge graph and the temporal features of the data source, construct a multi-head attention fusion network, calculate the weight matrix by scaling dot product attention, generate fusion feature vectors, extract temporal statistical features and identify important features to form a temporal fusion feature system; The multi-task prediction and pipeline configuration generation unit is used to build a multi-task prediction model using graph neural networks. It takes fused features as input and outputs orchestration mode, pipeline topology, cooperating nodes, estimated execution time and resource consumption. It uses a meta-learning framework to complete gradient updates and generate pipeline DAG topology and execution configuration. The orchestration scheme dynamic adjustment and logging unit is used to take a time series sequence as input from a time series prediction model, predict future orchestration requirements, dynamically adjust the orchestration scheme, generate a dynamic orchestration scheme, generate an interpretable report based on feature importance, perform stream-batch collaborative processing, and record execution logs. The batch-stream collaborative execution plan integration unit is used to integrate orchestration strategies and execution parameters based on multi-task prediction and time-series prediction results, construct pipeline DAG topology, design batch-stream collaborative nodes, and form a batch-stream collaborative execution plan.

[0067] Furthermore, the aforementioned multi-task prediction and pipeline configuration generation unit includes: The multi-dimensional fusion feature building unit is used to read pipeline task data and build task graphs. It uses graph convolutional networks to extract semantic features of task graphs, fuses knowledge graph features and data source features, builds a multi-task fusion network, and generates multi-dimensional fusion features. The multi-task prediction and initial topology generation unit is used to build a multi-task prediction model using graph convolutional networks. It inputs fused features into the model and outputs the orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes, generating the initial pipeline topology and configuration. The topology automatic generation and configuration file unit is used to automatically generate pipeline topology diagrams, node configuration parameters and edge configuration parameters based on fusion features using graph generation networks, and generate corresponding execution configuration files. The reinforcement learning optimization and adaptation unit is used to build a graph reinforcement learning environment and train optimization strategies. It combines the meta-learning framework to complete the gradient update and rapid adaptation of model parameters, adjust and optimize the topology and configuration, and finally generate the pipeline DAG topology and corresponding execution configuration information.

[0068] Furthermore, the aforementioned multi-task prediction and initial topology generation unit includes: The hierarchical temporal fusion feature vector generation unit is used to extract pipeline tasks and construct a hierarchical temporal task requirement graph. It uses a hierarchical graph convolutional network to extract hierarchical semantic features, combines a temporal model to extract temporal features, fuses knowledge graph features and data source features, constructs an attention fusion network, and generates hierarchical temporal fusion feature vectors. The hierarchical temporal latent space graph generation model building unit is used to construct an encoder, generator and discriminator to form a hierarchical temporal latent space graph generation multi-task model. The encoder encodes the hierarchical temporal topology graph into a latent space representation. The generator generates the topology structure based on fused features and latent space noise. The discriminator judges the authenticity of the generation and uses multi-task loss to train the model. The meta-learning adaptation and prediction result output unit is used to input the corresponding fusion features using the meta-learning algorithm, perform gradient update and fast adaptation on the support set, and output the orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time and collaborative node prediction results. The initial hierarchical temporal topology and configuration file generation unit is used to organize and generate an initial hierarchical temporal pipeline topology and configuration file containing multi-layer topology, node and edge configuration parameters based on the prediction results.

[0069] Furthermore, the aforementioned optimization and update module 500 includes: The source-target domain feature alignment and causal modeling unit is used to extract source domain scenes from the historical optimization scene library, identify new target domain scenes, construct a feature space and use a transfer learning domain adaptive algorithm to complete feature alignment, construct a pipeline optimization causal graph, and generate a transfer training set and causal effect matrix. The meta-learning training and optimization model building unit is used to perform meta-training using meta-learning algorithms, train transfer reinforcement learning agents, build Bayesian optimizers and define hyperparameter search spaces, initialize relevant models and acquisition functions, perform causal effect estimation on samples and calculate counterfactual inference predictions. The optimal pipeline configuration generation unit is used to extract scene features from the new scene in the target domain, align them to the feature space of the source domain through a feature mapping function, input the transfer reinforcement learning agent to perform gradient updates to generate an initial optimization strategy, fuse Bayesian optimization suggestions and apply causal constraints to generate the optimal pipeline configuration; The configuration deployment and full-process update unit is used to deploy the optimal configuration and execute fusion tasks, record execution logs, calculate the actual optimization effect, verify the relevant prediction accuracy and optimization efficiency, add new scenario samples to the target domain dataset, and update various models and data fusion pipelines.

[0070] Furthermore, the aforementioned meta-learning training and optimization model construction unit includes: The causal feature extraction and effect matrix generation unit is used to extract configuration parameters and performance indicators from historical optimization scenarios in the source domain, generate a causal graph and extract causal feature vectors using a causal discovery algorithm, collect new scenarios in the target domain and calculate domain differences, and learn domain-invariant feature mappings through a domain adaptive network to generate a causal effect matrix. The transfer reinforcement learning agent training unit is used to pre-train reinforcement learning agents using a meta-learning framework. It sets up a state space and action space, integrates causal effects and performance indicators to construct a reward function, and learns meta-knowledge from the source domain through meta-training to obtain the transfer reinforcement learning agent. The Bayesian optimizer construction and sample generation unit is used to construct the Bayesian optimizer, define the hyperparameter search space, initialize the Gaussian process model, select the acquisition function, and generate the initial hyperparameter sample set using a sampling method. The counterfactual reasoning and optimization model fusion unit is used to perform causal effect estimation and matching operations on hyperparameter samples, calculate counterfactual reasoning predictions based on the estimation results, complete Bayesian optimizer training, and form an optimization model that combines transfer reinforcement and Bayesian optimization.

[0071] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, this computer device includes a processor, memory, network interface, and database connected via a bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operations, computer programs, and the database. The internal memory provides an environment for the operation of the operations and computer programs stored in the non-volatile storage media. The database stores data such as performance data and intelligent analysis methods for enterprise management. The network interface is used for communication with external terminals via a network connection. When executed by a processor, this computer program implements an intelligent performance data analysis method for enterprise management, including: identifying heterogeneous characteristics and generating data source classification labels for multiple heterogeneous performance data sources; establishing a data source registry to collect metadata and business semantic information; generating a data source registry; constructing a performance terminology ontology based on the data source registry; using a semantic matching algorithm to map data source terms to standard terms; generating a terminology mapping table; extracting performance indicator caliber information based on the terminology mapping table; reverse-engineering unrecorded caliber information; detecting caliber differences and resolving conflicts using preset strategies; establishing a caliber standardization library; combining the data source classification labels, terminology mapping table, and caliber standardization library; orchestrating and executing a data fusion pipeline to generate a performance data warehouse and quality data; evaluating the fusion effect based on the performance data warehouse and quality data; optimizing the fusion strategy and updating the data fusion pipeline based on the evaluation results.

[0072] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements an intelligent performance data analysis method for enterprise management, including: identifying heterogeneous characteristics and generating data source classification labels for multiple heterogeneous performance data sources; establishing a data source registry center to collect metadata and business semantic information; generating a data source registry; constructing a performance terminology ontology based on the data source registry; using a semantic matching algorithm to map data source terms to standard terms; generating a terminology mapping relationship table; extracting performance indicator caliber information based on the terminology mapping relationship table; reverse-engineering unrecorded caliber information; detecting caliber differences and resolving conflicts using a preset strategy; establishing a caliber standardization library; combining the data source classification labels, terminology mapping relationship table, and caliber standardization library; orchestrating and executing a data fusion pipeline to generate a performance data warehouse and quality data; evaluating the fusion effect based on the performance data warehouse and quality data; optimizing the fusion strategy and updating the data fusion pipeline based on the evaluation results.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0074] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for intelligent analysis of performance data in enterprise management, characterized in that, include: For multiple heterogeneous performance data sources, identify heterogeneous characteristics and generate data source classification tags, establish a data source registry center to collect metadata and business semantic information, and generate a data source registry; A performance terminology ontology is constructed based on the data source registry, and a semantic matching algorithm is used to map data source terms to standard terms, generating a terminology mapping table. Based on the terminology mapping table, extract the performance indicator caliber information, reverse-engineer the caliber information that is not clearly recorded, detect caliber differences and use preset strategies to resolve conflicts, and establish a caliber standardization library; By combining the data source classification labels, terminology mapping relationship table and standardization library, a data fusion pipeline is orchestrated and executed to generate a performance data warehouse and quality data. Based on the performance data warehouse and quality data evaluation fusion effect, the fusion strategy is optimized and the data fusion pipeline is updated according to the evaluation results.

2. The intelligent analysis method for enterprise management performance data according to claim 1, characterized in that, The steps of identifying heterogeneous characteristics and generating data source classification tags for multiple heterogeneous performance data sources, establishing a data source registry center to collect metadata and business semantic information, and generating a data source registry include: Extract structural, syntactic, semantic, and temporal heterogeneity features from multiple heterogeneous performance data sources, calculate and quantify the heterogeneity intensity values, and generate a heterogeneous feature vector matrix; Heterogeneous feature maps are constructed based on heterogeneous feature vector matrices. Graph attention networks are used to generate node embedding vectors. Spectral clustering algorithms are used to classify data sources and generate classification labels. The labels and vectors are then merged to generate a data source classification result table. Based on the data source classification result table, a data source registry center is constructed, connector templates are matched to connect various data sources, metadata information is collected, business semantic information is extracted through natural language processing technology, and the two types of information are stored in the data source registry. Deploy a heterogeneous feature map update engine based on the data source registry, monitor metadata changes and reconstruct the heterogeneous feature map, update node embedding vectors and data source classification labels, re-collect metadata and business semantic information, and update the data source registry.

3. The intelligent analysis method for enterprise management performance data according to claim 1, characterized in that, The step of constructing a performance terminology ontology based on the data source registry, mapping data source terms to standard terms using a semantic matching algorithm, and generating a terminology mapping table includes: Based on the metadata information in the data source registry, a performance terminology ontology containing terminology layer, relationship layer and attribute layer is constructed, terminology node embedding vectors are generated, and the performance terminology ontology and vector information are stored in the terminology ontology library. Based on the term node embedding vectors in the term ontology, a semantic matching algorithm is used to map data source terms to standard terms, determine the mapping relationship between the two, and organize and save the mapping relationship to generate a term mapping relationship table. A quality assessment of term mapping relationships is conducted based on the term mapping relationship table, a term mapping quality report is generated, and low-confidence mapping relationships are marked as mappings to be reviewed. Based on the terminology mapping quality report, low-reliability mapping relationships are reviewed and corrected, and the corrected mapping relationships are updated to the terminology mapping relationship table, and the terminology ontology is updated synchronously.

4. The intelligent analysis method for enterprise management performance data according to claim 1, characterized in that, The steps of extracting performance indicator caliber information based on the terminology mapping table, reverse-engineering caliber information that is not explicitly recorded, detecting caliber differences and resolving conflicts using preset strategies, and establishing a caliber standardization library include: Based on the terminology mapping table, the performance indicator caliber information is extracted, a caliber causal graph model is constructed, causal relationships in dimensions such as caliber definition and calculation logic are identified, and a caliber causal graph containing causal nodes and directed causal edges is generated. Based on caliber information, a caliber knowledge graph is constructed. Semantic representations of unrecorded calibers are generated through graph reasoning. Similar caliber definitions are retrieved, and caliber information that is not clearly recorded is inferred by combining spatiotemporal evolution laws. Time series modeling is used to capture the dynamic trend of caliber changes, reverse the caliber information of unrecorded time points, identify abrupt changes in caliber definition, generate a report on the spatiotemporal distribution of caliber differences, and quantify the uncertainty of caliber characteristics. Based on various detection and inference results, a preset strategy is adopted to resolve the conflict of caliber differences, and the standardized caliber information is organized to establish a caliber standardization library.

5. The intelligent analysis method for enterprise management performance data according to claim 4, characterized in that, The steps of using time-series modeling to capture the dynamic change trend of caliber, inferring caliber information at unrecorded time points, identifying abrupt changes in caliber definition, generating a report on the spatiotemporal distribution of caliber differences, and quantifying the uncertainty of caliber characteristics include: A distributed spatiotemporal semantic graph of calibers is constructed, with elements such as caliber definitions and calculation logic defined as graph nodes and semantic relationships defined as graph edges. A word vector model is used to convert caliber terms into vector representations. A federated temporal graph neural network combined with a multi-head attention mechanism is used to model the semantic graph. The model parameters of each node are aggregated through a federated learning algorithm to capture the semantic association strength and temporal evolution features of the caliber and generate a federated caliber embedding vector. Monte Carlo Dropout technique is used to quantify the uncertainty of the federated model, generate relevant feature values ​​of the caliber embedding vector and calculate the uncertainty score, and generate uncertainty-aware federated caliber embedding vector. The embedding vector is used to infer the caliber information for unrecorded time points, retrieve similar caliber definitions among data source nodes, identify caliber definition mutation points and generate a report on the spatiotemporal distribution of caliber differences, store the results in the federated caliber evolution knowledge base and update the model parameters of each node.

6. The intelligent analysis method for enterprise management performance data according to claim 1, characterized in that, The steps of combining the data source classification labels, terminology mapping relationship table, and standardization library to orchestrate and execute a data fusion pipeline to generate a performance data warehouse and quality data include: Read the data source category tags, identify real-time data sources and batch data sources, build a domain ontology library and associate it with a terminology mapping table and a standardization library, and generate a stream-batch classification knowledge graph. The predictive model is used to input semantic features of the knowledge graph and features of the data source to generate a batch-stream collaborative orchestration scheme and pipeline DAG topology. The pipeline is orchestrated and collaborative nodes are designed to generate a batch-stream collaborative execution plan. Parallel processing pipelines are dynamically orchestrated according to the execution plan to process streaming data and batch data separately to generate corresponding data tables. Terminology mapping and standardization are completed based on knowledge graphs, and a fusion result table is generated by merging collaborative nodes. The merged data is loaded into a Lambda-based data warehouse to generate fact tables and dimension tables. Multi-dimensional quality checks and consistency assessments are then performed on the merged data to generate a performance data warehouse and corresponding quality data.

7. The intelligent analysis method for enterprise management performance data according to claim 6, characterized in that, The steps of using a predictive model to input semantic features of a knowledge graph and features of a data source to generate a batch-stream collaborative orchestration scheme and pipeline DAG topology, orchestrating the pipeline and designing collaborative nodes, and generating a batch-stream collaborative execution plan include: Read the semantic features of the knowledge graph and the temporal features of the data source, construct a multi-head attention fusion network, calculate the weight matrix by scaling dot product attention, generate fusion feature vectors, extract temporal statistical features and identify important features to form a temporal fusion feature system; A multi-task prediction model is constructed using graph neural networks. The input is fused features, and the output is orchestration mode, pipeline topology, cooperating nodes, estimated execution time and resource consumption. Gradient updates are completed using a meta-learning framework to generate pipeline DAG topology and execution configuration. The system uses a time-series prediction model to input time series sequences, predicts future orchestration requirements, dynamically adjusts the orchestration scheme, generates a dynamic orchestration scheme, generates an interpretable report based on feature importance, performs stream-batch collaborative processing, and records execution logs. Based on the results of multi-task prediction and time-series prediction, the orchestration strategy and execution parameters are integrated to construct a pipeline DAG topology, design flow-batch collaborative nodes, and form a flow-batch collaborative execution plan.

8. The intelligent analysis method for enterprise management performance data according to claim 7, characterized in that, The steps of constructing a multi-task prediction model using a graph neural network, inputting fused features, and outputting orchestration patterns, pipeline topology, cooperating nodes, estimated execution time, and resource consumption, and using a meta-learning framework to complete gradient updates, generating pipeline DAG topology and execution configuration, include: Read pipeline task data and construct a task graph. Use graph convolutional networks to extract semantic features of the task graph. Integrate knowledge graph features and data source features to construct a multi-task fusion network and generate multi-dimensional fusion features. A multi-task prediction model is constructed using graph convolutional networks. The model is input with fused features and outputs the orchestration mode, number of nodes, number of edges, resource allocation, estimated execution time, and prediction results related to collaborative nodes, generating the initial pipeline topology and configuration. A graph generation network based on fusion features is used to automatically generate pipeline topology graphs, node configuration parameters and edge configuration parameters, and generate corresponding execution configuration files. A graph reinforcement learning environment is constructed and an optimization strategy is trained. The meta-learning framework is used to complete the gradient update and rapid adaptation of model parameters. The topology and configuration are adjusted and optimized, and finally, a pipeline DAG topology and corresponding execution configuration information are generated.

9. The intelligent analysis method for enterprise management performance data according to claim 1, characterized in that, The steps of evaluating the fusion effect based on the performance data warehouse and quality data, optimizing the fusion strategy based on the evaluation results, and updating the data fusion pipeline include: Extract source domain scenes from the historical optimization scene library, identify new target domain scenes, construct a feature space and use a transfer learning domain adaptive algorithm to complete feature alignment, construct a pipeline optimization causal graph, and generate a transfer training set and causal effect matrix; Meta-training is performed using a meta-learning algorithm to train a transfer reinforcement learning agent, a Bayesian optimizer is constructed and a hyperparameter search space is defined, relevant models and acquisition functions are initialized, causal effect estimation is performed on samples, and counterfactual inference predictions are calculated. Extract scene features from the new scene in the target domain, align them to the feature space of the source domain through a feature mapping function, input them into the transfer reinforcement learning agent to perform gradient updates to generate an initial optimization strategy, fuse Bayesian optimization suggestions and apply causal constraints to generate the optimal pipeline configuration; Deploy the optimal configuration and execute the fusion task, record the execution log, calculate the actual optimization effect, verify the relevant prediction accuracy and optimization efficiency, add new scenario samples to the target domain dataset, and update various models and data fusion pipelines.

10. A performance data intelligent analysis device for enterprise management, characterized in that, include: The tag generation module is used to identify heterogeneous characteristics and generate data source classification tags for multiple heterogeneous performance data sources, establish a data source registry center to collect metadata and business semantic information, and generate a data source registry. The terminology mapping module is used to construct a performance terminology ontology based on the data source registry, and to map data source terms to standard terms using a semantic matching algorithm to generate a terminology mapping relationship table. The caliber processing module is used to extract performance indicator caliber information based on the terminology mapping table, reverse-engineer caliber information that is not clearly recorded, detect caliber differences and resolve conflicts using preset strategies, and establish a caliber standardization library. The fusion execution module is used to combine the data source classification labels, terminology mapping relationship table and standardization library to orchestrate and execute the data fusion pipeline to generate a performance data warehouse and quality data. The optimization and update module is used to evaluate the fusion effect based on the performance data warehouse and quality data, optimize the fusion strategy and update the data fusion pipeline according to the evaluation results.

Images