Data copy management intelligent agent training, automation arrangement and data processing method

Abstract

The application discloses the technical field of data copy, and particularly relates to a data copy management intelligent agent training, automatic arrangement and data processing method. First, enterprise self-owned business management specifications and full-amount business data are collected, specification clauses and data attribute attribution mapping is performed, and a correlation relationship from record to specification level is constructed. Then, based on the specification, data correlation mapping data set, multi-dimensional marking dimension standards suitable for training requirements are demarcated, batch hierarchical marking and compliance verification are performed on the data. After that, based on the compliance final version marking training data set, the initialized intelligent agent basic model is subjected to multi-round supervised training and iterative optimization. Based on the above, the application has the advantages of realizing highly accurate autonomous judgment and decision on the sensitive level, operation permission and life cycle strategy of copy data in different business scenarios, and fundamentally eliminating the adaptation gap between the general model and the actual business management requirements of enterprises.

Description

Technical Field

[0001] This invention relates to the field of data replication technology, specifically to methods for training, automating orchestration, and processing data replication management agents. Background Technology

[0002] Currently, enterprises are in the deepening stage of digital transformation, with various business platforms continuously iterating and upgrading. Massive amounts of structured, semi-structured, and unstructured business data are generated across the entire business chain, including production and operation, office collaboration, and business transactions. Data copies, as a fundamental resource supporting core scenarios such as enterprise disaster recovery and backup, business simulation testing, big data analysis, AI model training, and business fault backtracking, directly determine the enterprise's data security level and business operational continuity through their full lifecycle management quality.

[0003] Currently, enterprise data copy management mainly relies on manual offline operations by maintenance personnel. This management model is fragmented, manual, and experience-based, lacking a standardized, intelligent, and integrated end-to-end control system. Although enterprises have generally established their own business platform operation specifications, data security classification and protection specifications, and sensitive information identification and control rules, and have completed the basic classification and archiving of business data, the existing copy data management work has not effectively relied on these standardized specifications. This results in the inability of market-standard data management models to adapt to the private business control scenarios of enterprises, exhibiting core technical defects such as poor agent training adaptability, low decision-making accuracy, and insufficient compliance.

[0004] At its core, the underlying technical pain point lies in the inherent semantic gap between enterprise-specific business norms and the knowledge system of general pre-trained models. This gap manifests in three main technical mismatches: firstly, differences in domain distribution. The training data for general-purpose models is mostly publicly available general-purpose scenario data, which does not cover private domain data such as enterprise-specific business processes, data control rules, and industry compliance details. The data distribution deviates greatly from the actual business scenarios of enterprises, and the model cannot learn the business logic of enterprise private copy management. Secondly, the annotation system is inconsistent. General-purpose models rely on general and universal data annotation standards to build training samples, while the annotation rules for enterprise internal data security classification, sensitive information judgment, copy operation and maintenance management, and permission boundary definition are all highly private and customized standards. The annotation dimensions, judgment thresholds, and classification logic of the two are completely incompatible, and the general annotation system cannot adapt to the personalized compliance management needs of enterprises. Thirdly, the feature space is mismatched. Enterprise business specifications contain a large number of exclusive business terms, scenario-based sensitive features, and customized compliance constraint features, forming a unique private feature space. However, the general pre-trained model only learns general data features and does not internalize the enterprise's private exclusive semantic features and control logic, and cannot accurately identify the hidden sensitive attributes and compliance boundaries of enterprise business data.

[0005] In existing technologies, most enterprises directly reuse general data management models and their training samples without building dedicated training datasets and systems tailored to their own proprietary standards, business data characteristics, and compliance control requirements. This further amplifies the aforementioned technical deficiencies. The trained AI agents are severely disconnected from the enterprise's actual copy management business, failing to accurately adapt to the enterprise's proprietary compliance control scenarios. For example, a financial enterprise's core credit business data contains sensitive information such as customer credit information. Because it failed to build a proprietary annotation system according to the financial industry's specific data security standards and eliminate the semantic gap between models, the AI ​​agents trained directly using general samples cannot identify derived sensitive information in specific business scenarios. This leads to frequent issues of illegally retained sensitive data during subsequent storage, migration, and destruction management operations, failing to meet the strong compliance control requirements of the financial industry and hindering the intelligent, compliant, and precise control of the entire lifecycle of enterprise copy data. Summary of the Invention

[0006] To address the technical problem of poor model adaptability and low accuracy caused by the lack of standardized annotations in enterprise-owned intelligent agent training, this invention provides the following technical solution:

[0007] The training method for the data copy management agent includes the following steps:

[0008] A1, Business Specification Data Collection: Collect the enterprise's own business management specifications and full business data, map the attribution of specification clauses and data attributes, build the association relationship from records to the specification level, and output the specification and data association mapping dataset;

[0009] A2, Business Application Data Annotation: Based on the specifications and data association mapping dataset output by A1, define the multi-dimensional annotation standard to adapt to training requirements, perform batch hierarchical annotation and compliance verification on the data, and output the compliant final version of the annotated training dataset.

[0010] A3, Iterative Training and Optimization of the Agent: Based on the compliant final version labeled training dataset output by A2, the initial agent basic model is subjected to multiple rounds of supervised training and iterative optimization based on contrastive learning with normative constraints or reinforcement learning with compliance rewards, and a mature and usable data copy management agent is output.

[0011] As a preferred embodiment of the training method for the data copy management agent described in this invention, the specific steps of A1 are as follows:

[0012] A11, Comprehensive Collection of Standards System: Comprehensive collection of all currently effective management standards documents within the enterprise. At the same time, structured standard data is sorted out and extracted from the documents to form a raw dataset of enterprise copy management standards, and the raw dataset of enterprise copy management standards is output.

[0013] A12, Comprehensive Business Data Collection: Based on the data collection boundaries determined by the original set of enterprise copy management specifications output by A11, it comprehensively collects production business data, historical copy operation and maintenance data, and various backup, test, and training data generated by the enterprise's daily operations. At the same time, it performs preliminary cleaning on the collected data and outputs the enterprise's full-scale compliant business dataset.

[0014] A13, Data Specification Association Mapping: Based on the full-scale compliant business dataset of the enterprise output by A12, the rule matching engine is used to automatically assign and mark the metadata attributes of the data in the dataset with the corresponding control level, compliance usage requirements and sensitive information categories in the original set of enterprise copy management specifications, establish a mapping relationship from data records to specification category levels, and output specification and data association mapping dataset.

[0015] As a preferred embodiment of the training method for the data copy management agent described in this invention, the specific steps of A2 are as follows:

[0016] A21, precise definition of annotation dimensions: Based on the specifications and data association mapping dataset output by A1, the dimensions for annotating the copy data are determined, and for the determined dimensions, the corresponding annotation judgment thresholds and execution standards are defined, and the intelligent agent data annotation dimension standards are output.

[0017] A22, Batch Data Hierarchical Labeling: Based on the intelligent agent data labeling dimension standard output by A21, call the automated labeling tool to perform batch hierarchical, classification and scenario-specific labeling on all business data in the specification and data association mapping dataset, and output the initial version of the labeled business training dataset;

[0018] A23, Compliance Verification of Annotated Data: Based on the initial version of the annotated business training dataset output in A22, compliance verification is initiated by comparing it with the enterprise's data security specifications and business platform operation specifications, and a compliant final version of the annotated training dataset is output.

[0019] As a preferred embodiment of the training method for the data copy management intelligent agent described in this invention, the specific steps of A3 are as follows:

[0020] A31, Agent Model Initialization: Build a composite deep learning model framework with a Transformer-graph attention network hybrid architecture as the basic model of the dedicated agent for data copy management. At the same time, configure the model's data input interface, multimodal feature extraction module, decision reasoning module based on cross-modal attention mechanism, and multi-task strategy output feedback module to complete the initialization of basic model parameters and output the initialization basic model of the copy management agent.

[0021] A32, Annotated Data Model Training: Based on the initial copy management agent basic model output by A31, the compliance final version annotated training dataset output by A2 is divided into training set and validation set in an 8:2 ratio, and the entire dataset is input into the model. Based on contrastive learning of normative constraints or reinforcement learning of compliance rewards, the initial copy management agent basic model is trained, and the initially trained copy management agent is output.

[0022] A33, Agent Iterative Optimization and Validation: Based on the initially trained replica management agent output by A32, simulated replica management operation and maintenance tests are carried out using an out-of-sample test dataset that was not used in the training. A multi-dimensional evaluation index system is constructed to quantify the model performance. At the same time, based on the deviation between the decision results of the test set and the actual business management needs, the model hyperparameters and network weights are iteratively optimized in reverse, and finally a ready-to-use data replica management agent is output.

[0023] An automated orchestration method for data replica management agents, applied to data replica management agents trained using the above method, includes the following steps:

[0024] B1, Operation and Maintenance Management Requirements Collection: Summarize the replica operation and maintenance requirements of daily and special scenarios, and collect the real-time load status of storage, computing power and bandwidth resources, and output a real-time status table of replica operation and maintenance resources.

[0025] B2, Replica Operation and Maintenance Strategy Generation:

[0026] B21, Intelligent Matching of Operation and Maintenance Window Periods: Input the real-time status table of replica operation and maintenance resources output by B1 into the data replica management intelligent agent, so that the data replica management intelligent agent calls the built-in rule engine, dynamically identifies the peak business periods based on the real-time status table of replica operation and maintenance resources, and combines resource idle prediction to intelligently match exclusive off-peak execution windows for various operation and maintenance tasks, and output exclusive time window solutions for replica operation and maintenance.

[0027] B22, Intelligent Customization of Cyclic Plans: Based on the dedicated time window solution for replica operation and maintenance output by B21, the data replica management intelligent agent customizes the time execution plan, data storage and retention cycle, and periodic verification and maintenance plan of the replica strategy according to the security level, business usage frequency, and compliance retention requirements of different replica data, and outputs the initial strategy set of all dimensions of replica operation and maintenance.

[0028] B23, Policy Compliance Optimization and Calibration: Based on the initial set of replica operation and maintenance policies across all dimensions output by B22, the data replica management intelligent agent once again compares the policies with the enterprise business platform specifications and data security specifications, performs a round of compliance and adaptability optimization and calibration, and outputs a compliant and optimized replica operation and maintenance orchestration policy.

[0029] B3, Orchestration Strategy Implementation: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, it is broken down into execution instructions and sent to the terminal for automatic execution, and optimization suggestions are generated based on the execution log data.

[0030] As a preferred embodiment of the automated orchestration method for the data copy management intelligent agent described in this invention, the specific steps of B1 are as follows:

[0031] B11, Summary of Daily Operation and Maintenance Requirements: Summarize and organize the daily basic operation and maintenance requirements of the replica data, and calculate the execution frequency and basic control requirements of each task, and output a list of daily operation and maintenance requirements for the replica.

[0032] B12, Business Scenario Requirements Collection: Based on the list of daily operation and maintenance requirements for replicas output by B11, additional personalized operation and maintenance requirements for specific replica data service scenarios are collected, and a collection of replica operation and maintenance requirements for all scenarios is output.

[0033] B13, Operation and Maintenance Resource Status Statistics: Based on the full-scenario replica operation and maintenance requirement collection output by B12, a multi-time-scale layered collection and dynamic fusion mechanism is adopted to obtain the real-time load status and historical operation patterns of enterprise data storage resources, computing power scheduling resources and business operation bandwidth resources, respectively, to construct a multi-dimensional resource profile that integrates static constraints and dynamic features, and output a real-time status table of replica operation and maintenance resources.

[0034] As a preferred embodiment of the automated orchestration method for the data copy management intelligent agent described in this invention, the specific steps of B3 are as follows:

[0035] B31, Orchestration Strategy Decomposition and Distribution: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, the data replica management intelligent agent will decompose it into single operation and maintenance operation instructions that can be recognized by the execution terminal according to multiple dimensions, and output a set of replica operation and maintenance sub-items execution instructions;

[0036] B32, Automated Implementation of Policies: Based on the set of replica operation and maintenance instructions output by B31, each operation and maintenance terminal automatically completes the entire process of operation and maintenance according to the time plan and operation standards specified in the instructions, and outputs replica operation and maintenance execution process log data;

[0037] B33, Orchestration Execution Performance Review and Optimization: Based on the replica operation and maintenance execution process log data output by B32, the data replica management agent regularly performs quantitative reviews of the key performance indicators of the orchestration strategy and outputs iterative optimization and adjustment plans for replica orchestration.

[0038] The data processing method for a data copy management agent, applied to a data copy management agent trained using the above method, includes the following steps:

[0039] C1, Requirement Analysis of Replica Services: The application scenario of the replica request is determined by the data replica management intelligent agent, the corresponding exclusive processing specification is retrieved, and the original replica data adapted to the scenario is accurately retrieved according to the specification, and the original replica data adapted to the scenario is output.

[0040] C2, Compliance processing of duplicate data:

[0041] C21, Sensitive Information Desensitization Processing: Based on the scenario adaptation of the original copy data output by C1, perform hierarchical desensitization processing in accordance with the sensitive information control requirements, and output the desensitized and encrypted copy intermediate data;

[0042] C22, Data Compliance Cleaning and Rectification: Based on the anonymized and encrypted intermediate data of the copy output by C21, start the data quality cleaning program and output the cleaned and rectified compliant copy data;

[0043] C23, Copy Data Quality Verification: Based on the cleaned and standardized compliant copy data output from C22, and in accordance with the scenario-specific data processing standards and specifications, automatically perform four quality verifications of data integrity, accuracy, compliance and availability, and output qualified and usable copy data;

[0044] C3, Replica Resource Timeliness Assessment: Based on the verified and usable replica data output by C2, calculate the required resources and time consumption, and integrate them into an integrated service execution solution that includes resource configuration and time planning.

[0045] As a preferred embodiment of the data processing method for the data copy management intelligent agent described in this invention, the specific steps of C1 are as follows:

[0046] C11, Replica Application Scenario Determination: When a business request for replica data service is received, the data replica management intelligent agent analyzes the request intent and context to accurately determine the specific application scenario corresponding to the request and outputs the replica data application scenario determination result.

[0047] C12, Scenario Processing Standard Retrieval: Based on the scenario determination result of the replica data output by C11, the data replica management intelligent agent retrieves the exclusive process standard, data compliance requirements, quality verification specifications and business adaptation rules for replica data processing that are uniquely corresponding to the scenario from the pre-set business specification database, and outputs the scenario-specific data processing standard specification.

[0048] C13, Retrieval of Original Data for Replicas: Based on the scenario-specific data processing standard specifications output by C12, the data replica management intelligent agent generates precise data query instructions, accurately locates and retrieves the original replica basic data corresponding to the business type and time period from the enterprise data storage cluster, and outputs scenario-adapted original replica data.

[0049] As a preferred embodiment of the data processing method for the data copy management intelligent agent described in this invention, the specific steps of C3 are as follows:

[0050] C31, scheduling resource requirement calculation: Based on the verified and available replica data output by C2, the data replica management agent calls the resource calculation model to predict the theoretical demand of various resources required to complete the entire process of this replica data service, and outputs the replica service resource requirement calculation results.

[0051] C32, Service Duration and Feasibility Assessment: Based on the resource requirement calculation results of the replica service output by C31, and combined with the current load status in the enterprise operation and maintenance resource real-time status table, the data replica management intelligent agent assesses the key time nodes and total delivery cycle in the entire service process from data scheduling, transmission, business adaptation to scenario application, and outputs a replica service time cycle assessment report.

[0052] C33, Data Service Process Output and Dynamic Adjustment: Based on the resource requirement calculation results of the replica service output by C31 and the replica service time cycle assessment report output by C32, the data replica management intelligent agent integrates and generates a comprehensive standardized execution process. At the same time, if conflicting changes are detected in the resource status table during the execution of the plan, the execution plan is dynamically adjusted or an early warning is issued, and the integrated replica data service execution plan is output.

[0053] Compared with existing technologies:

[0054] 1. By collecting all currently effective business platform operation specifications, data security classification and protection specifications, and sensitive information control rules within the enterprise, a unique set of original data copy management specifications for the enterprise is constructed. Based on this, a precise mapping relationship is established between business data and specification category levels. Then, multi-dimensional annotation standards are defined to meet the training needs of the intelligent agent. All business data is standardized and verified for compliance. Finally, it is used for iterative training of the intelligent agent. The resulting data copy management intelligent agent deeply internalizes the enterprise's unique business data classification characteristics, security control level requirements, and compliance usage boundaries. It has the ability to make highly accurate autonomous judgments and decisions on the sensitivity level, operation permissions, and lifecycle policies of copy data in different business scenarios. This fundamentally eliminates the adaptation gap between general models and the actual business management needs of enterprises, and significantly improves the decision-making accuracy and compliance rate of the intelligent agent in enterprise private deployment scenarios.

[0055] 2. By inputting real-time resource status and full-scenario operation and maintenance requirements into the intelligent agent, intelligent matching and differentiated strategy customization for off-peak windows are achieved based on dynamically identified peak business periods. This eliminates the arbitrariness and inconsistencies in manual scheduling, realizes the unified, automated generation and dynamic optimization of full-scenario replica operation and maintenance strategies, and significantly improves operation and maintenance efficiency and resource utilization.

[0056] 3. By having the intelligent agent perform compliance processing such as hierarchical desensitization and quality verification before using the copy data, and by completing resource requirement calculation and service time assessment in advance, the system has the beneficial effects of eliminating the risk of sensitive information leakage from the source, identifying and avoiding resource scheduling conflicts and project delivery delays in advance, and ensuring the compliant, orderly and efficient implementation of copy data services. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0058] This invention provides a method for training a data copy management agent, comprising the following steps:

[0059] A1, Business Specification Data Collection: Collect the enterprise's own business management specifications and full business data, map the attribution of specification clauses and data attributes, build the association relationship from records to the specification level, and output the specification and data association mapping dataset;

[0060] The specific steps of A1 are as follows:

[0061] A11, Comprehensive Collection of Standardized Systems: Through pre-set data interfaces and data platforms, comprehensively collect all currently effective management standard documents within the enterprise, focusing on business platform operation management standards, data security classification and protection standards, business data type classification management standards, and sensitive information control details. At the same time, extract structured standard data such as mandatory control clauses, data operation boundary requirements, and sensitive information judgment standards from the documents to form a dedicated original dataset of enterprise copy management standards, and complete the classification, storage, and archiving, outputting the original set of enterprise copy management standards;

[0062] A12, Comprehensive Business Data Collection: Based on the data collection boundaries determined by the original set of enterprise copy management specifications output by A11, it connects with various business branch systems, data storage clusters and business operation backends to comprehensively collect production business data, historical copy operation and maintenance data and various backup, test and training data generated by the enterprise's daily operations. At the same time, it performs preliminary cleaning on the collected data, removes invalid and redundant data, retains compliant and valid core business data, and outputs the enterprise's full compliant business dataset.

[0063] A13, Data Specification Association Mapping: Based on the full-scale compliant business dataset of the enterprise output by A12, the rule matching engine automatically assigns and marks the metadata attributes of the data in the dataset, such as data type, source business system, and included sensitive fields, to the corresponding control level, compliance usage requirements, and sensitive information categories in the original set of enterprise copy management specifications. This establishes a mapping relationship between data records and specification category levels, and outputs a specification and data association mapping dataset.

[0064] A2, Business Application Data Annotation: Based on the specifications and data association mapping dataset output by A1, define the multi-dimensional annotation standard to adapt to training requirements, perform batch hierarchical annotation and compliance verification on the data, and output the compliant final version of the annotated training dataset.

[0065] The specific steps of A2 are as follows:

[0066] A21, Precisely Defined Annotation Dimensions: Based on the specifications and data association mapping dataset output by A1, to enable the agent to learn data security, business attributes, and scenario adaptability, the dimensions for annotating the replica data are determined. These dimensions include at least one or more of the following: data security level annotation, data business attribute annotation, sensitive information concentration annotation, and replica operation and maintenance adaptation scenario annotation. For each determined dimension, the corresponding annotation judgment threshold and execution standard are defined, and the agent's data annotation dimension standard is output. The annotation dimensions are naturally derived from the control level, sensitive category, and other attributes in the specification mapping results, resulting in annotation dimensions such as security level, business attribute, sensitive concentration, and operation and maintenance scenario.

[0067] A22, Batch Data Hierarchical Labeling: Based on the intelligent agent data labeling dimension standard output by A21, the automated labeling tool is called to perform batch hierarchical, classification and scenario-specific labeling on all business data in the specification and data association mapping dataset. For data whose labeling results meet the preset high-risk threshold, the manual review and verification process is initiated to correct possible errors in the automated labeling, ensure that the labeling results meet the preset high accuracy requirements, and finally output the initial version of the labeled business training dataset.

[0068] A23, Compliance Verification of Annotated Data: Based on the initial annotated business training dataset output from A22, compliance verification is initiated against the enterprise's data security specifications and business platform operation specifications: annotation completeness verification, annotation accuracy verification, and annotation compliance verification. For data that fails verification, annotation information completion, erroneous annotation correction, or invalid data removal operations are performed. Finally, through the optimized dataset, a compliant final annotated training dataset is output.

[0069] A3, Iterative Training and Optimization of the Agent: Based on the compliant final version labeled training dataset output by A2, the initial agent basic model is subjected to multiple rounds of supervised training and iterative optimization based on contrastive learning with normative constraints or reinforcement learning with compliance rewards, and a mature and usable data copy management agent is output.

[0070] The specific steps for A3 are as follows:

[0071] A31, Agent Model Initialization: Build a composite deep learning model framework with a Transformer-Graph Attention Network (GAT) hybrid architecture, integrating natural language processing, feature classification and sequence prediction capabilities, as the basic model for a dedicated agent for data copy management. The model framework specifically includes: a data input interface for uniformly receiving structured business data metadata, semi-structured specification clause text, and unstructured data description information; a multimodal feature extraction module, consisting of a BERT-baseTransformer encoder and a 2-layer Graph Attention Network (GAT) submodule, where the Transformer encoder extracts semantic features of the specification text and features of the data description text, and the GAT submodule constructs a ternary association graph of "specification clause - data attribute - control requirement" and extracts topological association features; a decision reasoning module based on a cross-modal attention mechanism, used to fuse text semantic features and graph topological features to generate interpretable decision reasoning paths; and a multi-task strategy output feedback module, including 3 classification output heads and 2 regression output heads. The classification output heads are used to output data security level determination, operation permission allocation, and operation and maintenance scenario matching results, respectively, while the regression output heads are used to output data lifecycle duration and resource scheduling demand prediction results, respectively. The basic model parameters are initialized, with an initial learning rate of 1e-4, batch size of 32, maximum training epochs of 50, and weight decay coefficient of 1e-5, outputting the initialized replica management agent basic model.

[0072] A32, Annotated Data Model Training: Based on the initialization copy management agent basic model output from A31, the compliance final version annotated training dataset output from A2 is divided into training and validation sets in an 8:2 ratio and fully input into the model. Multiple rounds of supervised training and feature learning are conducted based on contrastive learning with specification constraints or reinforcement learning with compliance rewards. The model is iteratively optimized by minimizing the combined loss function. Specifically, when using contrastive learning with specification constraints, the combined loss function consists of three weighted parts: ① Specification alignment contrastive loss (weight 0.4), used to constrain data features corresponding to the same specification clause to cluster in the feature space and to separate data features corresponding to different specification clauses, bridging the semantic gap between the general model and enterprise-specific specifications; ② Multi-class cross-entropy loss (weight 0.4), used to optimize the prediction accuracy of classification tasks such as data security level and operational scenarios; ③ Mean squared error loss (weight 0.2), used to optimize the prediction accuracy of regression tasks such as lifecycle duration and resource requirements. When using compliance-based reinforcement learning, a loss function is constructed based on the Proximal Policy Optimization (PPO) algorithm. The compliance compliance rate of the decision outcome is the core reward term (weight 0.7), while decision latency penalty term (weight 0.2) and resource consumption penalty term (weight 0.1) are introduced. The model's decision strategy is optimized by maximizing the cumulative reward. An early stopping mechanism is adopted during training. Training is terminated when the overall loss on the validation set does not decrease for 5 consecutive rounds, and the initially trained replica management agent is output.

[0073] A33, Agent Iterative Optimization and Validation: Based on the initially trained replica management agent output from A32, a simulated replica management operation and maintenance test covering the entire process of replica creation, storage, migration, destruction, and data service was conducted using an out-of-sample test dataset (10% of the total labeled dataset) that was not used in training. A multi-dimensional evaluation index system was constructed to quantitatively evaluate model performance: Training phase indicators included training loss, validation loss, classification task accuracy, recall, F1 score, regression task mean absolute error, and root mean square error; deployment phase core indicators included decision accuracy ≥95%, compliance compliance rate ≥98%, sensitive information misclassification rate ≤1%, sensitive information missed classification rate ≤0.5%, and single data decision response latency ≤100ms. Based on the deviation between the decision results of the test set and actual business management needs, a method combining grid search and Bayesian optimization was used to iteratively optimize the model hyperparameters and network weights. Through repeated optimization until all core indicators reached and exceeded the enterprise's preset thresholds, a ready-to-use data replica management agent was finally output.

[0074] An automated orchestration method for data replica management agents, applied to data replica management agents trained using the above method, includes the following steps:

[0075] B1, Operation and Maintenance Management Requirements Collection: Summarize the replica operation and maintenance requirements of daily and special scenarios, and collect the real-time load status of storage, computing power and bandwidth resources, and output a real-time status table of replica operation and maintenance resources.

[0076] The specific steps of B1 are as follows:

[0077] B11, Summary of Daily Operation and Maintenance Requirements: Connects with the enterprise data operation and maintenance management backend and the operation and maintenance ports of various business departments, summarizes and organizes the daily basic operation and maintenance requirements of the replica data, focusing on routine operation and maintenance tasks such as regular replica creation, routine data backup, replica compliance checks and cleanup of existing replicas, and calculates the execution frequency and basic control requirements of each task, and outputs a list of daily operation and maintenance requirements for replicas.

[0078] B12, Business Scenario Requirements Collection: Based on the list of daily operation and maintenance requirements for replicas output by B11, additional personalized operation and maintenance requirements from specialized replica data service scenarios such as disaster recovery backup, business simulation testing, big data analysis and AI model training are collected. The specific requirements of each scenario for replica usage timeliness, data specifications and operation and maintenance priority are clarified, and a total set of operation and maintenance requirements covering all scenarios is generated and output as a collection of replica operation and maintenance requirements for all scenarios.

[0079] B13, Operation and Maintenance Resource Status Statistics: Based on the full-scenario replica operation and maintenance requirement collection output by B12, a multi-time-scale layered collection and dynamic fusion mechanism is adopted to obtain the real-time load status and historical operating patterns of enterprise data storage resources, computing power scheduling resources, and business operation bandwidth resources. A multi-dimensional resource profile integrating static constraints and dynamic features is constructed, and a real-time status table of replica operation and maintenance resources is output. Specifically, the following steps are included:

[0080] Multi-timescale data stratified acquisition

[0081] Real-time data acquisition: Through lightweight acquisition agents deployed on each resource node, real-time performance indicators such as CPU utilization, memory usage, disk IOPS, and network bandwidth utilization are collected at a frequency of 1-5 seconds. Sliding window technology is used to smooth the raw data and filter out instantaneous noise interference.

[0082] Historical data statistics: Extract resource operation data from the enterprise's operation and maintenance data warehouse for the past 30-90 days, aggregate and statistically analyze it at three time granularities: hourly, daily, and weekly, calculate characteristic indicators such as average load, peak load, and idle time ratio for each time period, and identify the periodic patterns and abnormal fluctuations in resource usage.

[0083] Resource constraints and feature extraction

[0084] Static constraint extraction: sort out and define the hard constraints on resource usage, including the maximum number of concurrent tasks per node, the threshold for remaining disk space, the upper limit of network bandwidth, and the period when the business system is not operational;

[0085] Dynamic feature extraction: Extract fixed peak periods of business operations, trends in resource load, and resource consumption characteristics of different business types from historical data to establish a resource demand prediction model.

[0086] Multi-dimensional resource profile fusion construction

[0087] A three-layer fusion architecture consisting of a static foundation layer, a dynamic feature layer, and a real-time state layer is used to construct resource profiles.

[0088] Static base layer: Fixed attribute information of storage resources, such as hardware configuration, capacity limit, and constraints;

[0089] Dynamic feature layer: Stores long-term features such as periodic patterns, peak periods, and load trends obtained from historical statistics;

[0090] Real-time state layer: Stores instantaneous state information such as current resource load, available capacity, and task queue length;

[0091] By using timestamp alignment technology to associate data at different time scales, and employing a weighted fusion algorithm to calculate the comprehensive availability score of each resource node, the historical patterns are used to correct and supplement the real-time status.

[0092] Real-time status table dynamically updated

[0093] The replica operation and maintenance resource real-time status table includes key fields such as resource node ID, resource type, current load, available capacity, overall availability, predicted idle period, and constraints.

[0094] Real-time status layer data is updated every 10 seconds, dynamic feature layer data is updated once every day at midnight, and static basic layer data is updated when resource configuration changes.

[0095] When abnormal fluctuations in resource load or significant changes in business requirements are detected, a full data update and resource profile reconstruction are immediately triggered.

[0096] B2, Replica Operation and Maintenance Strategy Generation: Using the data replica management intelligent agent, combined with the real-time status table of replica operation and maintenance resources output by B1, the operation and maintenance window period is matched and the cycle plan is customized. The generated strategy is then optimized for compliance, and the optimized replica operation and maintenance orchestration strategy is output.

[0097] The specific steps for B2 are as follows:

[0098] B21, Intelligent Matching of Operation and Maintenance Window Periods: The real-time status table of replica operation and maintenance resources output by B1 is input into the data replica management intelligent agent, so that the data replica management intelligent agent calls the built-in rule engine, dynamically identifies the peak business periods based on the real-time status table of replica operation and maintenance resources, takes the priority of avoiding the peak period as the core principle, and combines resource idle prediction to intelligently match exclusive off-peak execution windows for various operation and maintenance tasks, and output exclusive time window solutions for replica operation and maintenance.

[0099] B22, Intelligent Customization of Cyclic Plans: Based on the dedicated time window solution for replica operation and maintenance output by B21, the data replica management intelligent agent customizes the time execution plan, data storage and retention cycle, and periodic verification and maintenance plan of the replica strategy according to the security level, business usage frequency, and compliance retention requirements of different replica data. This process strictly distinguishes the differentiated rules between daily operation and maintenance and special scenario operation and maintenance, forming an initial full-dimensional strategy set and outputting the initial full-dimensional strategy set for replica operation and maintenance;

[0100] B23, Policy Compliance Optimization and Calibration: Based on the initial set of full-dimensional replica operation and maintenance policies output in B22, the data replica management intelligent agent once again compares the policies against the enterprise business platform specifications and data security specifications, and performs a round of compliance and adaptability optimization and calibration on all generated operation and maintenance policies. This step will automatically eliminate non-compliant or conflicting policies, adjust suboptimal policies, and fill in missing control links, thereby avoiding resource occupation conflicts and compliance risks from the source, and finally outputting a compliant and optimized replica operation and maintenance orchestration policy.

[0101] B3, Orchestration Strategy Implementation: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, it is broken down into execution instructions and sent to the terminal for automatic execution, and optimization suggestions are generated based on execution log data;

[0102] The specific steps for B3 are as follows:

[0103] B31, Orchestration Strategy Decomposition and Distribution: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, the data replica management intelligent agent will decompose the operation and maintenance into single operation and maintenance instructions that can be recognized by the execution terminal according to multiple dimensions such as operation and maintenance time period, business scenario, and operation and maintenance type. These instructions are accurately distributed to the enterprise data operation and maintenance execution terminal and each business node to complete the full coverage deployment of execution instructions and output the replica operation and maintenance sub-item execution instruction set;

[0104] B32, Automated Implementation of Policies: Based on the replica operation and maintenance sub-execution instruction set output by B31, each operation and maintenance terminal automatically completes the entire process of operation and maintenance operations such as replica creation, storage archiving, periodic control, data verification and expiration cleanup according to the time plan and operation standards specified in the instructions. The entire process does not require manual intervention and automatically generates operation logs with timestamps, outputting replica operation and maintenance execution process log data.

[0105] B33, Orchestration Execution Performance Review and Optimization: Based on the replica operation and maintenance execution process log data output by B32, the data replica management intelligent agent regularly performs quantitative reviews of key performance indicators such as the execution completion rate, business impact, resource utilization, and compliance rate of the orchestration strategy. For execution bottlenecks and policy adaptation shortcomings identified in the review, it generates policy optimization suggestions or candidate correction parameters and pushes them to the operation and maintenance management terminal. After confirmation, it fine-tunes the subsequent orchestration strategy logic in reverse, forming a continuously optimized automated orchestration closed loop and outputting iterative optimization and adjustment schemes for replica orchestration.

[0106] The data processing method for a data copy management agent, applied to a data copy management agent trained using the above method, includes the following steps:

[0107] C1, Requirement Analysis of Replica Services: The application scenario of the replica request is determined by the data replica management intelligent agent, the corresponding exclusive processing specification is retrieved, and the original replica data adapted to the scenario is accurately retrieved according to the specification, and the original replica data adapted to the scenario is output.

[0108] The specific steps of C1 are as follows:

[0109] C11, Replica Application Scenario Determination: When a replica data service request is received from businesses such as disaster recovery, simulation, data analysis, and AI training, the data replica management intelligent agent analyzes the request intent and context to accurately determine the specific application scenario corresponding to the request, clarify the core business objectives of the scenario, and output the replica data application scenario determination result.

[0110] C12, Scenario Processing Standard Retrieval: Based on the scenario determination result of the replica data output by C11, the data replica management intelligent agent retrieves the exclusive process standard, data compliance requirements, quality verification specifications and business adaptation rules for replica data processing that are uniquely corresponding to the scenario from the pre-set business specification database, and outputs the scenario-specific data processing standard specification.

[0111] C13, Retrieval of Original Data for Replicas: Based on the scenario-specific data processing standard specifications output by C12, the data replica management intelligent agent generates precise data query instructions, accurately locates and retrieves the original replica basic data corresponding to the business type and time period from the enterprise data storage cluster. This process aims to lock the data retrieval scope, avoid redundancy and delay caused by full scan, and output scenario-adapted original replica data.

[0112] C2, Compliance Processing of Copy Data: Based on the scenario output of C1, adapt the original copy data, perform hierarchical sensitive information desensitization and quality cleaning and standardization, and perform multi-dimensional quality verification on the processed data, outputting qualified and usable copy data;

[0113] The specific steps of C2 are as follows:

[0114] C21, Sensitive Information De-identification Processing: Based on the scenario adaptation of the original copy data output by C1, and in accordance with the requirements for sensitive information control, hierarchical de-identification processing is performed: According to the preset hierarchical de-identification strategy, different processing operations such as anonymization, de-identification or encryption are performed on different types of sensitive fields to prevent the unauthorized use of exposed sensitive data, and output the de-identified and encrypted copy intermediate data.

[0115] C22, Data Compliance Cleaning and Consolidation: Based on the anonymized and encrypted intermediate data of the copy output by C21, the data quality cleaning program is started. According to the preset rules, missing values, duplicate values ​​and format errors are removed from the data. In accordance with the data format standards and accuracy requirements of this application scenario, the data is consolidated, formatted and field calibrated to improve data quality and business adaptability, and the cleaned and consolidated compliant copy data is output.

[0116] C23, Replica Data Quality Verification: Based on the cleaned and standardized compliant replica data output from C22, the system automatically performs four quality checks on data integrity, accuracy, compliance, and availability, in accordance with the scenario-specific data processing standards and specifications. Data that fails the checks will be marked and returned to the C22 stage for re-cleaning; data that passes the checks will be included in the resource pool to be scheduled, and qualified and usable replica data will be output.

[0117] C3, Replica Resource Timeliness Assessment: Based on the verified and usable replica data output by C2, calculate the required resources and time consumption, and integrate them into an integrated service execution solution that includes resource configuration and time planning;

[0118] The specific steps of C3 are as follows:

[0119] C31, scheduling resource requirement calculation: Based on the verified and usable replica data output by C2, the data replica management agent calls the resource calculation model to predict the theoretical demand for various resources such as computing power, storage, and bandwidth required to complete the entire process of this replica data service. This overcomes the shortcomings of inaccurate manual assessment and outputs the replica service resource requirement calculation results.

[0120] C32, Service Duration and Feasibility Assessment: Based on the resource requirement calculation results of the replica service output by C31, and combined with the current load status in the enterprise operation and maintenance resource real-time status table, the data replica management intelligent agent assesses the key time nodes and total delivery cycle in the entire service process from data scheduling, transmission, business adaptation to scenario application. This assessment can effectively identify time-consuming bottlenecks and potential resource conflict risks, provide a basis for task initiation decisions, and output a replica service time cycle assessment report.

[0121] C33, Data Service Process Output and Dynamic Adjustment: Based on the resource requirement calculation results of the replica service output in C31 and the replica service time cycle assessment report output in C32, the data replica management intelligent agent integrates and generates a comprehensive standardized execution process. The content covers a detailed resource configuration plan, a phased time progress plan, and key points for risk prevention and control. This plan is simultaneously pushed to the business department and the operation and maintenance department to support the safe, orderly, and rapid implementation of replica data services. At the same time, if conflicting changes are detected in the resource status table during the execution of the plan, the execution plan is dynamically adjusted or an early warning is issued, and an integrated replica data service execution plan is output.

[0122] In summary, including but not limited to the following embodiments:

[0123] This embodiment uses a nationwide joint-stock commercial bank's enterprise-level data copy management scenario as an example to fully demonstrate the implementation process of the training, automated orchestration, and data processing methods for the data copy management intelligent agent described in this invention. The bank currently has three core business segments: retail lending, corporate settlement, and wealth management. It generates an average of 1.2TB of structured transaction data and 3.5TB of unstructured customer behavior data daily. The data copies support three core scenarios: disaster recovery backup, business simulation testing, and risk control model training. The original manual management model suffers from high compliance risks, low operational efficiency, and insufficient resource utilization.

[0124] I. Training and Implementation of the Data Copy Management Agent

[0125] Step A1: Business Specification Data Collection

[0126] A11 Standard System Comprehensive Collection

[0127] Through the bank's data platform interface, 12 core management documents currently in effect, including the "Commercial Bank Data Security Management Measures", "Technical Specifications for Personal Financial Information Protection", "Credit Business Data Classification and Grading Standards", and "Detailed Rules for Data Copy Lifecycle Management", were collected. A total of 327 structured and standardized data were extracted, including 189 mandatory control clauses, 4 levels of data security classification standards, and 17 categories of sensitive information, forming the original set of bank copy management specifications and completing the classification and archiving.

[0128] A12 Business Data Fully Collected

[0129] Based on the data collection boundaries defined by the standard original set, and by connecting with core business systems, data warehouses, and backup storage clusters, a total of 128TB of production transaction data, historical copy operation and maintenance logs, risk control model training datasets, and business simulation test data from the past 12 months were collected. After removing 17TB of invalid and duplicate data, a total of 111TB of full-scale compliant business dataset for the enterprise was output.

[0130] A13 Data Specification Association Mapping

[0131] A rule-matching engine based on regular expressions and semantic matching is deployed to automatically map the metadata attributes of business data (data type, source system, sensitive fields, creation time) to the control levels and compliance requirements in the original specification set. For example, "Retail Credit Customer Credit Report" data is mapped to the specification clauses of "Data Security Level 3, Sensitive Category - Personal Financial Information, Retention Period of 5 Years, Access Only Allowed by Risk Control Department", ultimately outputting a specification-data association mapping dataset containing 120 million data records.

[0132] Step A2: Business Application Data Labeling

[0133] A21 annotation dimensions are precisely defined.

[0134] Based on the association mapping dataset, four core annotation dimensions and decision thresholds were determined:

[0135] Data security levels: 1-4 (corresponding to Public, Internal, Secret, and Confidential)

[0136] Sensitive information concentration: Low (<5% of sensitive fields), Medium (5%-30%), High (>30%)

[0137] Business Attributes: Retail lending, corporate settlement, wealth management

[0138] Operation and maintenance adaptation scenarios: disaster recovery backup, simulation testing, model training output of intelligent agent data annotation dimension standards and solidification into annotation tools.

[0139] A22 Batch Data Hierarchical Labeling

[0140] The automated annotation tool was used to perform batch annotation on the entire dataset. Manual review was initiated on 1.2 million data entries labeled "security level 4, high sensitivity concentration". 32,000 annotation errors were corrected, and the annotation accuracy rate reached 99.7%. The initial version of the annotated business training dataset was output.

[0141] A23 Data Compliance Verification

[0142] In accordance with the bank's data security standards, a triple verification of completeness, accuracy, and compliance was conducted. 87,000 missing labeled data entries were filled in, and 13,000 invalid labeled data entries were removed. The final compliant labeled training dataset was output, containing 119 million valid labeled samples.

[0143] Step A3: Iterative training and optimization of the agent

[0144] A31 agent model initialization

[0145] Build a basic model with a Transformer-Graph Attention Network (GAT) hybrid architecture, configure a BERT-based encoder (12 layers, 768-dimensional hidden layers) and a 2-layer GAT submodule, and set up 3 classification output heads (security level, operation permissions, and operation and maintenance scenario) and 2 regression output heads (lifetime duration and resource requirements). Initialize model parameters: learning rate 1e-4, batch size 32, maximum training epochs 50, weight decay coefficient 1e-5, and output the basic model of the initialization replica management agent.

[0146] A32 labeled data model training

[0147] The final compliant annotation dataset was divided into a training set (95.2 million records) and a validation set (23.8 million records) in an 8:2 ratio. Comparative learning with canonical constraints was used for training. The combined loss function weights were set as follows: canonical alignment contrastive loss 0.4, multi-class cross-entropy loss 0.4, and mean squared error loss 0.2. An early stopping mechanism was enabled during training; training was terminated when the overall loss on the validation set did not decrease for five consecutive rounds. A total of 28 rounds of training were completed, and the initially trained copy management agent was output.

[0148] A33 agent iterative optimization verification

[0149] We extracted 10% of the total labeled dataset as out-of-sample test data (11.9 million records) and conducted full-process simulated operation and maintenance tests. We constructed a multi-dimensional evaluation index system, and after three rounds of grid search and Bayesian optimization, all core model indicators met the standards: decision accuracy 96.8%, compliance compliance rate 99.2%, sensitive information false positive rate 0.7%, sensitive information false negative rate 0.3%, and single-data-line decision response latency 72ms. The final output was a ready-to-use intelligent agent for managing bank-specific data copies.

[0150] II. Automated Orchestration and Implementation of Data Copy Management Agents

[0151] Step B1: Gathering Operation and Maintenance Management Requirements

[0152] Summary of B11's daily maintenance requirements

[0153] Connect to the bank's operations and maintenance management backend to summarize daily operations and maintenance requirements: one full backup per day, one incremental backup every 4 hours, one copy compliance check per week, and one expired copy cleanup per month, and output a list of daily operations and maintenance requirements for copies.

[0154] B12 Business Scenario Requirements Collection

[0155] Collect specific scenario requirements: Disaster recovery backup scenario requires RPO ≤ 15 minutes and RTO ≤ 1 hour; Risk control model training scenario requires providing the latest transaction data copy every morning; Business simulation test scenario requires providing a full business snapshot copy every Monday, and output a collection of full-scenario copy operation and maintenance requirements.

[0156] B13 Operation and Maintenance Resource Status Statistics

[0157] A multi-timescale layered data acquisition mechanism is adopted: real-time data is collected every 3 seconds, including CPU, memory, disk IOPS, and bandwidth utilization; historical data is used to analyze resource operation patterns over the past 30 days, identifying the peak business period as 9:00-17:00 on weekdays. A three-dimensional resource profile is constructed, consisting of a "static foundation layer, dynamic feature layer, and real-time status layer." The real-time status table is updated every 10 seconds, outputting a real-time status table for replica operation and maintenance resources.

[0158] Step B2: Generating Replica Operation and Maintenance Strategies

[0159] B21 Maintenance Window Intelligent Matching

[0160] Input the real-time resource status table into the intelligent agent. The intelligent agent automatically identifies peak business periods and prioritizes avoiding them, matching off-peak windows for different operation and maintenance tasks: full backup is matched with 0:00-4:00 daily, incremental backup is matched with 1:00, 5:00, 9:00, 13:00, 17:00, and 21:00 daily (non-peak periods), and compliance checks are matched with 22:00-24:00 every Sunday, outputting a dedicated time window plan for replica operation and maintenance.

[0161] B22 Cycle Plan Intelligent Customization

[0162] The intelligent agent customizes strategies based on the data security level: credit core data of security level 4 is retained for 5 years and verified every 7 days; corporate settlement data of security level 3 is retained for 3 years and verified every 15 days; office data of security level 2 is retained for 1 year and verified every 30 days, and an initial strategy set for full-dimensional copy operation and maintenance is output.

[0163] B23 Strategy Compliance Optimization and Calibration

[0164] The intelligent agent calibrates the initial strategy against the bank's compliance standards, eliminates the non-compliant strategy of "storing Level 4 data copies to public storage nodes", adjusts the suboptimal strategy of "retaining model training copies for more than 7 days", completes the control link of "data erasure before copy destruction", and outputs the compliant and optimized copy operation and maintenance orchestration strategy.

[0165] Step B3: Implementing the orchestration strategy

[0166] B31 orchestration strategy breakdown distributed

[0167] The intelligent agent breaks down the orchestration strategy into 217 executable instructions according to the operation and maintenance period and business scenario, and accurately distributes them to 12 storage nodes and 8 operation and maintenance terminals, outputting a set of execution instructions for replica operation and maintenance.

[0168] B32 Strategy Automated Implementation

[0169] Each terminal automatically completes operations such as replica creation, storage archiving, periodic verification, and expiration destruction according to instructions, without any manual intervention. It automatically generates operation logs with timestamps and outputs replica operation and maintenance execution process log data.

[0170] B33 orchestration execution performance review and optimization

[0171] The intelligent agent conducts a quantitative review of the execution results monthly, and the statistics show that: the operation and maintenance task completion rate is 100%, the business impact is reduced by 85%, the storage resource utilization rate is increased by 42%, and the compliance rate is 99.5%. In response to the problem of "increased backup latency on the end-of-month settlement day", an optimization suggestion of "starting the end-of-month backup at 23:00 in advance" is generated to form a continuous optimization closed loop.

[0172] III. Data Processing Implementation of the Data Copy Management Agent

[0173] This section uses a copy data service requested by a bank's risk control department for training a fraud detection model as an example to demonstrate the data processing flow:

[0174] Step C1: Analysis of Replica Service Requirements

[0175] C11 Copy Application Scenario Determination

[0176] On May 10, 2026, the risk control department submitted a copy service application, requesting to "obtain the full volume of retail credit transaction data for April 2026 for fraud detection model training." The intelligent agent analyzed the application intent and context, accurately determined the application scenario to be "AI model training - risk control fraud detection," and output the scenario determination result.

[0177] C12 scene processing standard call

[0178] The intelligent agent retrieves the scenario-specific standards from the specification database: sensitive information must undergo medium-level desensitization, data format is CSV, precision is retained to 2 decimal places, data integrity requirement is 100%, delivery cycle is ≤4 hours, and outputs scenario-specific data processing standard specifications.

[0179] C13 Copy Original Data Retrieval

[0180] The intelligent agent generates precise data query commands, locates and retrieves 127GB of retail credit transaction data from April 1 to April 30, 2026 from the core storage cluster, avoiding a full scan, and outputs original copy data adapted to the scenario.

[0181] Step C2: Compliance processing of duplicate data

[0182] C21 Sensitive Information De-identification Processing

[0183] Following a medium-level anonymization strategy, the ID card number is de-identified (keeping the first 6 digits and the last 4 digits), the mobile phone number is anonymized (replaced with ***********), and the customer's name is encrypted. The anonymized and encrypted copy of the intermediate data is then output.

[0184] C22 Data Compliance Cleaning and Organization

[0185] The data quality cleaning process was initiated, removing 237 missing values, 112 duplicate values, and 45 data entries with incorrect formatting. The data was then uniformly converted to CSV format, with the precision adjusted to two decimal places. A cleaned and standardized compliant copy of the data was output.

[0186] C23 Copy Data Quality Verification

[0187] The system performs four checks: integrity, accuracy, compliance, and availability. With a 100% pass rate, the data is included in the resource pool to be scheduled, and 126.8GB of valid and usable replica data is output.

[0188] Step C3: Timeliness Assessment of Copy Resources

[0189] C31 scheduling resource demand calculation

[0190] The agent invokes the resource calculation model to predict the resources required for this service: 8-core CPU, 16GB memory, 100Mbps bandwidth, and 200GB temporary storage, and outputs the resource requirement calculation results.

[0191] C32 Service Duration and Feasibility Assessment

[0192] Based on the current real-time resource status (CPU utilization 28%, bandwidth utilization 35%), the assessment results are as follows: data scheduling takes 30 minutes, transmission takes 1.5 hours, and service adaptation takes 1 hour, with a total delivery cycle of 2 hours and 50 minutes, meeting the 4-hour delivery requirement. A service time cycle assessment report is then output.

[0193] C33 Data Service Process Output and Dynamic Adjustment

[0194] The intelligent agent integrates and generates a unified execution plan, clearly defining resource allocation, phased timelines, and key risk control points, which is simultaneously pushed to the risk control and operations departments. During execution, there were no resource anomalies, and data delivery was completed 2 hours and 42 minutes after the application submission, outputting a unified service execution plan for duplicate data.

[0195] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for training a data copy management agent, characterized in that, Includes the following steps: A1, Business Specification Data Collection: Collect the enterprise's own business management specifications and full business data, map the attribution of specification clauses and data attributes, build the association relationship from records to the specification level, and output the specification and data association mapping dataset; A2, Business Application Data Annotation: Based on the specifications and data association mapping dataset output by A1, define the multi-dimensional annotation standard to adapt to training requirements, perform batch hierarchical annotation and compliance verification on the data, and output the compliant final version of the annotated training dataset. A3, Iterative Training and Optimization of the Agent: Based on the compliant final version labeled training dataset output by A2, the initial agent basic model is subjected to multiple rounds of supervised training and iterative optimization based on contrastive learning with normative constraints or reinforcement learning with compliance rewards, and a mature and usable data copy management agent is output.

2. The training method for the data copy management agent according to claim 1, characterized in that, The specific steps of A1 are as follows: A11, Comprehensive Collection of Standards System: Comprehensive collection of all currently effective management standards documents within the enterprise. At the same time, structured standard data is sorted out and extracted from the documents to form a raw dataset of enterprise copy management standards, and the raw dataset of enterprise copy management standards is output. A12, Comprehensive Business Data Collection: Based on the data collection boundaries determined by the original set of enterprise copy management specifications output by A11, it comprehensively collects production business data, historical copy operation and maintenance data, and various backup, test, and training data generated by the enterprise's daily operations. At the same time, it performs preliminary cleaning on the collected data and outputs the enterprise's full-scale compliant business dataset. A13, Data Specification Association Mapping: Based on the full-scale compliant business dataset of the enterprise output by A12, the rule matching engine is used to automatically assign and mark the metadata attributes of the data in the dataset with the corresponding control level, compliance usage requirements and sensitive information categories in the original set of enterprise copy management specifications, establish a mapping relationship from data records to specification category levels, and output specification and data association mapping dataset.

3. The training method for the data copy management agent according to claim 1, characterized in that, The specific steps of A2 are as follows: A21, precise definition of annotation dimensions: Based on the specifications and data association mapping dataset output by A1, the dimensions for annotating the copy data are determined, and for the determined dimensions, the corresponding annotation judgment thresholds and execution standards are defined, and the intelligent agent data annotation dimension standards are output. A22, Batch Data Hierarchical Labeling: Based on the intelligent agent data labeling dimension standard output by A21, call the automated labeling tool to perform batch hierarchical, classification and scenario-specific labeling on all business data in the specification and data association mapping dataset, and output the initial version of the labeled business training dataset; A23, Compliance Verification of Annotated Data: Based on the initial version of the annotated business training dataset output in A22, compliance verification is initiated by comparing it with the enterprise's data security specifications and business platform operation specifications, and a compliant final version of the annotated training dataset is output.

4. The training method for the data copy management agent according to claim 1, characterized in that, The specific steps for A3 are as follows: A31, Agent Model Initialization: Build a composite deep learning model framework with a Transformer-graph attention network hybrid architecture as the basic model of the dedicated agent for data copy management. At the same time, configure the model's data input interface, multimodal feature extraction module, decision reasoning module based on cross-modal attention mechanism, and multi-task strategy output feedback module to complete the initialization of basic model parameters and output the initialization basic model of the copy management agent. A32, Annotated Data Model Training: Based on the initial copy management agent basic model output by A31, the compliance final version annotated training dataset output by A2 is divided into training set and validation set in an 8:2 ratio, and the entire dataset is input into the model. Based on contrastive learning of normative constraints or reinforcement learning of compliance rewards, the initial copy management agent basic model is trained, and the initially trained copy management agent is output. A33, Agent Iterative Optimization and Validation: Based on the initially trained replica management agent output by A32, simulated replica management operation and maintenance tests are carried out using an out-of-sample test dataset that was not used in the training. A multi-dimensional evaluation index system is constructed to quantify the model performance. At the same time, based on the deviation between the decision results of the test set and the actual business management needs, the model hyperparameters and network weights are iteratively optimized in reverse, and finally a ready-to-use data replica management agent is output.

5. An automated orchestration method for data copy management agents, characterized in that, The data copy management agent trained by the method of any one of claims 1-4 includes the following steps: B1, Operation and Maintenance Management Requirements Collection: Summarize the replica operation and maintenance requirements of daily and special scenarios, and collect the real-time load status of storage, computing power and bandwidth resources, and output a real-time status table of replica operation and maintenance resources. B2, Replica Operation and Maintenance Strategy Generation: B21, Intelligent Matching of Operation and Maintenance Window Periods: Input the real-time status table of replica operation and maintenance resources output by B1 into the data replica management intelligent agent, so that the data replica management intelligent agent calls the built-in rule engine, dynamically identifies the peak business periods based on the real-time status table of replica operation and maintenance resources, and combines resource idle prediction to intelligently match exclusive off-peak execution windows for various operation and maintenance tasks, and output exclusive time window solutions for replica operation and maintenance. B22, Intelligent Customization of Cyclic Plans: Based on the dedicated time window solution for replica operation and maintenance output by B21, the data replica management intelligent agent customizes the time execution plan, data storage and retention cycle, and periodic verification and maintenance plan of the replica strategy according to the security level, business usage frequency, and compliance retention requirements of different replica data, and outputs the initial strategy set of all dimensions of replica operation and maintenance. B23, Policy Compliance Optimization and Calibration: Based on the initial set of replica operation and maintenance policies across all dimensions output by B22, the data replica management intelligent agent once again compares the policies with the enterprise business platform specifications and data security specifications, performs a round of compliance and adaptability optimization and calibration, and outputs a compliant and optimized replica operation and maintenance orchestration policy. B3, Orchestration Strategy Implementation: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, it is broken down into execution instructions and sent to the terminal for automatic execution, and optimization suggestions are generated based on the execution log data.

6. The automated orchestration method for data copy management agents according to claim 5, characterized in that, The specific steps of B1 are as follows: B11, Summary of Daily Operation and Maintenance Requirements: Summarize and organize the daily basic operation and maintenance requirements of the replica data, and calculate the execution frequency and basic control requirements of each task, and output a list of daily operation and maintenance requirements for the replica. B12, Business Scenario Requirements Collection: Based on the list of daily operation and maintenance requirements for replicas output by B11, additional personalized operation and maintenance requirements for specific replica data service scenarios are collected, and a collection of replica operation and maintenance requirements for all scenarios is output. B13, Operation and Maintenance Resource Status Statistics: Based on the full-scenario replica operation and maintenance requirement collection output by B12, a multi-time-scale layered collection and dynamic fusion mechanism is adopted to obtain the real-time load status and historical operation patterns of enterprise data storage resources, computing power scheduling resources and business operation bandwidth resources, respectively, to construct a multi-dimensional resource profile that integrates static constraints and dynamic features, and output a real-time status table of replica operation and maintenance resources.

7. The automated orchestration method for data copy management agents according to claim 5, characterized in that, The specific steps for B3 are as follows: B31, Orchestration Strategy Decomposition and Distribution: Based on the compliance-optimized replica operation and maintenance orchestration strategy output by B2, the data replica management intelligent agent will decompose it into single operation and maintenance operation instructions that can be recognized by the execution terminal according to multiple dimensions, and output a set of replica operation and maintenance sub-items execution instructions; B32, Automated Implementation of Policies: Based on the set of replica operation and maintenance instructions output by B31, each operation and maintenance terminal automatically completes the entire process of operation and maintenance according to the time plan and operation standards specified in the instructions, and outputs replica operation and maintenance execution process log data; B33, Orchestration Execution Performance Review and Optimization: Based on the replica operation and maintenance execution process log data output by B32, the data replica management agent regularly performs quantitative reviews of the key performance indicators of the orchestration strategy and outputs iterative optimization and adjustment plans for replica orchestration.

8. A data processing method for a data copy management intelligent agent, characterized in that, The data copy management agent trained by the method of any one of claims 1-4 includes the following steps: C1, Requirement Analysis of Replica Services: The application scenario of the replica request is determined by the data replica management intelligent agent, the corresponding exclusive processing specification is retrieved, and the original replica data adapted to the scenario is accurately retrieved according to the specification, and the original replica data adapted to the scenario is output. C2, Compliance processing of duplicate data: C21, Sensitive Information Desensitization Processing: Based on the scenario adaptation of the original copy data output by C1, perform hierarchical desensitization processing in accordance with the sensitive information control requirements, and output the desensitized and encrypted copy intermediate data; C22, Data Compliance Cleaning and Rectification: Based on the anonymized and encrypted intermediate data of the copy output by C21, start the data quality cleaning program and output the cleaned and rectified compliant copy data; C23, Copy Data Quality Verification: Based on the cleaned and standardized compliant copy data output from C22, and in accordance with the scenario-specific data processing standards and specifications, automatically perform four quality verifications of data integrity, accuracy, compliance and availability, and output qualified and usable copy data; C3, Replica Resource Timeliness Assessment: Based on the verified and usable replica data output by C2, calculate the required resources and time consumption, and integrate them into an integrated service execution solution that includes resource configuration and time planning.

9. The data processing method for a data copy management intelligent agent according to claim 8, characterized in that, The specific steps of C1 are as follows: C11, Replica Application Scenario Determination: When a business request for replica data service is received, the data replica management intelligent agent analyzes the request intent and context to accurately determine the specific application scenario corresponding to the request and outputs the replica data application scenario determination result. C12, Scenario Processing Standard Retrieval: Based on the scenario determination result of the replica data output by C11, the data replica management intelligent agent retrieves the exclusive process standard, data compliance requirements, quality verification specifications and business adaptation rules for replica data processing that are uniquely corresponding to the scenario from the pre-set business specification database, and outputs the scenario-specific data processing standard specification. C13, Retrieval of Original Data for Replicas: Based on the scenario-specific data processing standard specifications output by C12, the data replica management intelligent agent generates precise data query instructions, accurately locates and retrieves the original replica basic data corresponding to the business type and time period from the enterprise data storage cluster, and outputs scenario-adapted original replica data.

10. The data processing method for a data copy management intelligent agent according to claim 8, characterized in that, The specific steps of C3 are as follows: C31, scheduling resource requirement calculation: Based on the verified and available replica data output by C2, the data replica management agent calls the resource calculation model to predict the theoretical demand of various resources required to complete the entire process of this replica data service, and outputs the replica service resource requirement calculation results. C32, Service Duration and Feasibility Assessment: Based on the resource requirement calculation results of the replica service output by C31, and combined with the current load status in the enterprise operation and maintenance resource real-time status table, the data replica management intelligent agent assesses the key time nodes and total delivery cycle in the entire service process from data scheduling, transmission, business adaptation to scenario application, and outputs a replica service time cycle assessment report. C33, Data Service Process Output and Dynamic Adjustment: Based on the resource requirement calculation results of the replica service output by C31 and the replica service time cycle assessment report output by C32, the data replica management intelligent agent integrates and generates a comprehensive standardized execution process. At the same time, if conflicting changes are detected in the resource status table during the execution of the plan, the execution plan is dynamically adjusted or an early warning is issued, and the integrated replica data service execution plan is output.