Intelligent agent learning method, apparatus and device based on multi-source data

Abstract

The application provides an agent learning method, device and equipment based on multi-source data, and relates to the field of big data processing. Multi-source operation data generated in the running process of a target system is acquired, multi-dimensional knowledge assets of the target system are extracted, an intelligent analysis engine is used, the multi-source operation data and the multi-dimensional knowledge assets are dynamically fused and analyzed according to a preset multi-channel strategy, a full-dimensional system representation of the target system is generated, and a structured automated capability module that can be called by an agent is generated based on the full-dimensional system representation. The method can be adapted to a no-document legacy system, a black box system and a heterogeneous system architecture, is suitable for various agent platforms, greatly reduces the cost of manual intervention and customized development, and effectively improves the system access adaptation efficiency and generalization ability.

Description

Technical Field

[0001] This application relates to the field of big data processing, and in particular to an intelligent agent learning method, apparatus and device based on multi-source data. Background Technology

[0002] As agent technology evolves from L2 reasoners to L3 agents, its core capabilities have expanded from "dialogue interaction" to "task closure". Enterprise-level agents are widely used in scenarios such as automating internal system operations, cross-system task collaboration, and digital transformation of legacy systems.

[0003] In existing technologies, methods for intelligent agents to learn new systems are mainly divided into three categories: first, developing capability modules interface by interface, which involves manually analyzing interface documents to write operation tools and workflows; second, reverse engineering of source code, which relies on a complete code library; and third, log-driven methods, which infer interface call chains and data flows by collecting system operation logs.

[0004] However, existing technologies suffer from problems such as inability to handle undocumented systems, poor generalization ability, and low efficiency. Summary of the Invention

[0005] This application provides intelligent agent learning methods, apparatus, and devices based on multi-source data to solve the problems of existing technologies, such as inability to handle undocumented systems, poor generalization ability, and low efficiency.

[0006] In a first aspect, embodiments of this application provide an intelligent agent learning method based on multi-source data, including:

[0007] Acquire multi-source operational data generated by the target system during its operation;

[0008] Extract the multidimensional knowledge assets of the target system;

[0009] Using an intelligent analysis engine, the multi-source operational data and the multi-dimensional knowledge assets are dynamically fused and analyzed according to a preset multi-channel strategy to generate a full-dimensional system representation of the target system.

[0010] Based on the full-dimensional system representation, a structured automation capability module that can be invoked by an intelligent agent is generated.

[0011] As an optional implementation, the method of utilizing an intelligent analysis engine to dynamically fuse and analyze the multi-source operational data with the multi-dimensional knowledge assets according to a preset multi-channel strategy, thereby generating a full-dimensional system representation of the target system, includes:

[0012] A change perception analysis is performed on the structural, behavioral, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain the change correlation.

[0013] Based on the aforementioned changing relationships, adaptive correlation modeling is performed on the multi-source operational data and the multi-dimensional knowledge assets to generate dynamic cognitive results.

[0014] Based on the dynamic cognitive results, the business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system are inferred to obtain the inference results;

[0015] Based on the inference results, the system's cognitive structure is dynamically adjusted to generate a full-dimensional system representation of the target system.

[0016] As an optional implementation, the step of performing change-aware analysis on the structural, behavioral, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain change correlations includes:

[0017] The multi-source operational data is divided according to a preset time window, and the multi-dimensional knowledge assets are divided according to version iteration records to form multiple change sets;

[0018] The current change set is obtained from the multiple change sets, and the behavioral features or business path information associated with the current change set is extracted and updated. The business rule model generated in the previous iteration is dynamically adjusted using a lightweight model to form the incremental cognitive result of this iteration.

[0019] By using dependency analysis techniques, the scope of the impact of the changed elements in the incremental cognitive results on the structured automation capability modules generated in previous iterations is determined.

[0020] Based on the scope of influence, a local update operation is performed on the affected structured automation capability modules generated in previous iterations to generate change relationships.

[0021] As an optional implementation, the step of performing adaptive correlation modeling on the multi-source operational data and the multi-dimensional knowledge assets based on the changing correlation to generate dynamic cognitive results includes:

[0022] Obtain the integrity metrics of the multi-source operational data and the availability metrics of the multi-dimensional knowledge assets;

[0023] Based on the integrity index and the availability index, the weight allocation of the running data analysis channel and the knowledge asset analysis channel is adjusted through a dynamic channel selection algorithm to form a preliminary fusion result;

[0024] Based on the preliminary fusion results, the logical consistency between the multi-source operational data and the multi-dimensional knowledge assets is detected;

[0025] When there is a logical inconsistency between the multi-source operational data and the multi-dimensional knowledge assets, the data sources with inconsistencies are prioritized based on the business context, and the preliminary fusion results are semantically corrected based on the high-priority data sources to obtain semantic verification results.

[0026] The dynamic cognitive result is generated based on the semantically corrected fusion result.

[0027] As an optional implementation, the step of inferring business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system based on the dynamic cognitive results, and obtaining the inference results, includes:

[0028] A dynamic graph of operational data is constructed based on the multi-source operational data, wherein the behavior sequence or component call relationship in the multi-source operational data is mapped to the nodes of the dynamic graph, and the execution sequence or interaction logic between each node is mapped to the edges of the dynamic graph.

[0029] A static graph of knowledge assets is constructed based on the multidimensional knowledge assets, wherein the structural definition information and business logic rules in the multidimensional knowledge assets are mapped to the attribute information of the nodes in the static graph of knowledge assets.

[0030] The dynamic graph of the running data and the static graph of the knowledge assets are input into a pre-trained graph neural network model. The graph neural network model aggregates the node features of the dynamic graph and the attribute information of the static graph to generate a global semantic representation vector of the target system. The implicit business logic relationship is inferred based on the global semantic representation vector to obtain the inference result.

[0031] As an optional implementation, the multi-source operational data includes at least two of the following: real-time operational logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data. The multi-dimensional knowledge assets include at least two of the following: business rules of the target system, architecture design documents, historical operation and maintenance experience, and domain knowledge graphs.

[0032] As an optional implementation, after generating a structured automation capability module that can be invoked by an intelligent agent based on the full-dimensional system representation, the method further includes:

[0033] The two-way verification strategy includes: performing forward runtime verification on the generated structured automation capability module using the multi-source runtime data to obtain runtime verification results, and performing reverse consistency verification on the inferred logical rules using the multi-dimensional knowledge assets to obtain consistency verification results.

[0034] When the runtime verification result and / or the consistency verification result are inconsistent with the expected result, an error report is generated and a re-analysis process is automatically triggered. Based on the error feedback information contained in the error report, the analysis strategy weight of each channel in the multi-channel strategy is dynamically adjusted.

[0035] As an optional implementation, the structured automation capability module includes: operation tool definition and workflow orchestration, and the method further includes:

[0036] Load the operation tool definitions and workflow orchestrations into the agent's task execution engine;

[0037] The task execution engine parses high-level business instructions into specific operation sequences for the target system.

[0038] Secondly, embodiments of this application provide an intelligent agent learning device based on multi-source data, comprising:

[0039] The acquisition module is used to acquire multi-source operational data generated by the target system during its operation.

[0040] The extraction module is used to extract the multidimensional knowledge assets of the target system;

[0041] The generation module is used to utilize an intelligent analysis engine to dynamically fuse and analyze the multi-source operating data and the multi-dimensional knowledge assets according to a preset multi-channel strategy, thereby generating a full-dimensional system representation of the target system.

[0042] The generation module is also used to generate a structured automation capability module that can be invoked by an intelligent agent based on the full-dimensional system representation.

[0043] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: an analysis module and an inference module;

[0044] The analysis module is used to perform change perception analysis on the structural changes, behavioral changes, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain the change correlation.

[0045] The generation module is also used to perform adaptive association modeling on the multi-source operational data and the multi-dimensional knowledge assets based on the changing correlation, and generate dynamic cognitive results;

[0046] The inference module is used to infer the business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system based on the dynamic cognitive results, and obtain the inference results;

[0047] The generation module is used to dynamically adjust the system's cognitive structure based on the inference results and generate a full-dimensional system representation of the target system.

[0048] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a division module, an adjustment module, and a determination module;

[0049] The partitioning module is used to partition the multi-source operational data according to a preset time window and partition the multi-dimensional knowledge assets according to version iteration records to form multiple change sets.

[0050] The adjustment module is used to obtain the current change set from the multiple change sets, extract and update the behavioral features or business path information associated with the current change set, and dynamically adjust the business rule model generated in the previous iteration using a lightweight model to form the incremental cognitive result of this iteration.

[0051] The determining module is used to determine, through dependency analysis technology, the scope of the impact of the changed elements in the incremental cognitive results on the structured automation capability module generated in the previous iteration;

[0052] The generation module is further configured to perform a local update operation on the affected previously generated structured automation capability modules based on the scope of influence, in order to generate change correlations.

[0053] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a detection module;

[0054] The acquisition module is also used to acquire the integrity index of the multi-source operating data and the availability index of the multi-dimensional knowledge asset;

[0055] The generation module is also used to adjust the weight allocation of the running data analysis channel and the knowledge asset analysis channel according to the integrity index and the availability index through a dynamic channel selection algorithm to form a preliminary fusion result;

[0056] The detection module is used to detect the logical consistency between the multi-source operational data and the multi-dimensional knowledge asset based on the preliminary fusion results.

[0057] The generation module is also used to determine the priority of inconsistent data sources based on the business context when there is a logical inconsistency between the multi-source running data and the multi-dimensional knowledge asset, and to perform semantic correction on the preliminary fusion result based on the high-priority data source to obtain a semantic verification result.

[0058] The generation module is also used to generate the dynamic cognitive result based on the semantically corrected fusion result.

[0059] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a construction module;

[0060] The construction module is used to construct a dynamic graph of running data based on the multi-source running data, wherein the behavior sequence or component call relationship in the multi-source running data is mapped to the nodes of the dynamic graph, and the execution sequence or interaction logic between each node is mapped to the edges of the dynamic graph.

[0061] The construction module is further configured to construct a static graph of knowledge assets based on the multidimensional knowledge assets, wherein the structural definition information and business logic rules in the multidimensional knowledge assets are mapped to the attribute information of the nodes in the static graph of knowledge assets.

[0062] The generation module is further configured to input the dynamic graph of the running data and the static graph of the knowledge asset into a pre-trained graph neural network model, aggregate the node features of the dynamic graph and the attribute information of the static graph through the graph neural network model, generate the global semantic representation vector of the target system, and infer the implicit business logic relationship based on the global semantic representation vector to obtain the inference result.

[0063] As an optional implementation, the determining module is further configured to include at least two of the following: the multi-source operational data includes real-time operational logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data; and the multi-dimensional knowledge assets include at least two of the following: business rules of the target system, architecture design documents, historical operation and maintenance experience, and domain knowledge graph.

[0064] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: an execution module and an adjustment module;

[0065] The execution module is used to execute a two-way verification strategy, which includes: using the multi-source runtime data to perform forward runtime verification on the generated structured automation capability module to obtain runtime verification results, and using the multi-dimensional knowledge assets to perform reverse consistency verification on the inferred logical rules to obtain consistency verification results.

[0066] The adjustment module is used to generate an error report and automatically trigger a re-analysis process when the runtime verification result and / or the consistency verification result are inconsistent with the expected result, and dynamically adjust the analysis strategy weight of each channel in the multi-channel strategy based on the error feedback information contained in the error report.

[0067] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a loading module and a parsing module;

[0068] The loading module is used to load the operation tool definition and workflow orchestration into the agent's task execution engine;

[0069] The parsing module is used to parse high-level business instructions into specific operation sequences for the target system through the task execution engine.

[0070] Thirdly, embodiments of this application provide an intelligent agent learning device based on multi-source data, including: a receiver, a transmitter, a memory, and a processor;

[0071] Receiver, used to receive instructions and data;

[0072] A transmitter is used to send commands and data;

[0073] The memory stores computer-executed instructions;

[0074] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0075] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0076] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0077] The intelligent agent learning method based on multi-source data provided in this application collects multi-source data from the target system during its operation phase, mines and extracts multi-dimensional knowledge assets, and relies on an intelligent analysis engine combined with multi-channel preset strategies to achieve dynamic fusion analysis of data and knowledge assets, constructing a full-dimensional system representation, and further encapsulating and generating structured automated capability modules that can be directly called by intelligent agents. On the one hand, it can be compatible with complex architecture scenarios such as undocumented legacy systems, black-box systems, and heterogeneous systems, breaking through the limitations of traditional system adaptation that relies on manual document sorting and customized development and modification, reducing the workload of manual intervention and special customized development investment, and compressing project implementation costs. On the other hand, by leveraging data knowledge fusion and full-dimensional representation modeling, it improves the completeness and accuracy of system analysis and characterization, enhances the universality, flexibility, and generalization ability of cross-type system access adaptation, effectively shortens the system docking cycle, and improves the overall adaptation and implementation efficiency. Attached Figure Description

[0078] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0079] Figure 1 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 1 ;

[0080] Figure 2 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 2 ;

[0081] Figure 3 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 3 ;

[0082] Figure 4 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 4 ;

[0083] Figure 5 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 5 ;

[0084] Figure 6 A schematic diagram of the structure of the intelligent agent learning device based on multi-source data provided in this application;

[0085] Figure 7 A schematic diagram of the structure of the intelligent agent learning device based on multi-source data provided in this application.

[0086] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0087] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0088] As agent technology evolves from L2 inferential agents to L3 agents, its core capabilities have expanded beyond simple "dialogue interaction" to encompass "task closure." Enterprise-level agents are being widely used to automate internal systems and efficiently handle cross-system task collaboration.

[0089] In existing technologies, the methods for intelligent agents to learn and access new systems mainly rely on three traditional paths: first, developing capability modules interface by interface, that is, manually writing operation tools and workflows by deeply analyzing interface documents; second, using source code reverse analysis, but this is highly dependent on obtaining the complete system code library; and third, running log-driven, that is, inferring the interface call chain and internal data flow by collecting and analyzing system running logs.

[0090] However, existing technologies not only struggle to handle legacy or black-box systems lacking documentation, but also exhibit poor generalization capabilities when facing different system architectures, often requiring extensive customization. This model, heavily reliant on manual intervention and specific conditions, results in overall low integration efficiency.

[0091] To address the aforementioned issues, this application provides an intelligent agent learning method based on multi-source data. This method collects multi-source measured data from the target system's operational phase, extracts and consolidates multi-dimensional knowledge assets from this data, and leverages an intelligent analysis engine and a pre-defined multi-channel strategy to achieve dynamic fusion and analysis of operational data and knowledge assets. This results in the construction of a comprehensive system representation that fully characterizes the system's features. Based on this representation, standardized, structured, and automated capability modules are encapsulated and generated for direct invocation by intelligent agents. This solution does not rely on native system documentation, internal source code, or architecture specifications. It is naturally adaptable to undocumented legacy systems, black-box systems, and heterogeneous architecture scenarios, reducing manual analysis and customized development workload, compressing adaptation costs, and simultaneously improving the efficiency of rapid integration and adaptation across different system types, as well as its generalization and deployment capabilities.

[0092] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0093] Figure 1 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 1 The implementing entity in this embodiment is, for example, an intelligent agent learning system based on multi-source data. Figure 1 As shown, the method includes:

[0094] S101: Obtain multi-source operational data generated by the target system during operation.

[0095] Multi-source operational data refers to the collection of various dynamic data generated in real time during the normal operation, abnormal triggering, and user operation of the target system. These data include interface call chains, runtime logs, network requests, system events, user operation behaviors, and abnormal logs. They are dynamic data carriers that reflect the actual operating status, operation path, and behavior pattern of the system.

[0096] Specifically, by deploying an independent data acquisition module, the operating status of the target system is monitored in real time. Log parsing technology (such as regular expression matching) is used to extract key fields from the runtime logs. The complete sequence of interface calls is recorded through network request capture tools. At the same time, the time nodes, operation content and triggered system responses of user operations are collected, and information such as error codes and triggering scenarios of abnormal events are collected synchronously.

[0097] S102: Extract multidimensional knowledge assets from the target system.

[0098] Among them, multidimensional knowledge assets refer to various static knowledge collections formed during the design, development and deployment of the target system, covering source code, interface documents, design documents, database structure, ER diagram (Entity-Relationship Diagram), business rule descriptions, etc., and are static data carriers that reflect the system design intent, interface definition, business logic and data model.

[0099] Specifically, through the knowledge acquisition module, targeted parsing strategies are used to extract different types of knowledge assets. For textual assets such as interface documents and design documents, Natural Language Processing (NLP) technology is used to extract key information such as interface definitions and business rules. For source code, code parsing tools such as Abstract Syntax Tree (AST) analysis are used to extract interface parameters, return fields, and embedded business logic. For database structure, SQL parsing tools are used to extract table structure, field types, field constraints, and inter-table relationships.

[0100] S103: Utilizing an intelligent analysis engine, based on a preset multi-channel strategy, it dynamically integrates and analyzes multi-source operational data with multi-dimensional knowledge assets to generate a full-dimensional system representation of the target system.

[0101] Among them, the intelligent analysis engine is the core module for realizing joint understanding of multi-source data. It has the capabilities of multi-channel analysis, semantic parsing, and logical reasoning. It can parse document content through natural language processing (NLP), model interface call relationships through graph neural networks (GNN), and extract business logic through the rule engine.

[0102] The multi-channel strategy refers to three preset analysis strategies: data-driven + knowledge asset-assisted, knowledge asset-driven + data-driven, and dual-channel collaborative. These strategies can be dynamically adjusted according to data characteristics.

[0103] Full-dimensional system representation refers to the complete description of system functions, business logic, data models, and interface relationships formed by integrating the dynamic operation behavior and static design logic of the system.

[0104] Specifically, the intelligent analysis engine first preprocesses the structured multi-source runtime dataset and knowledge asset repository, cleaning up invalid data and filling in missing information. Then, it dynamically selects the corresponding multi-channel strategy based on data completeness and quality. If the runtime data is of high quality and the interface call chain is complete, but the knowledge assets are missing, the runtime data is used as the main channel, and business rules are inferred by analyzing interface call patterns and operation paths. If the knowledge assets are complete but the runtime data is sparse, the knowledge assets are used as the main channel, and the system operation path is inferred by using static rules. If both exist and there is a conflict, the conflict detection module identifies logical contradictions, prioritizing the actual call logic of the runtime data, and performing semantic verification in conjunction with the knowledge assets. By modeling interface call relationships through GNN and parsing business rules using NLP, the engine comprehensively analyzes the system functions, business logic, and data models from the dynamic perspective of runtime data and the static perspective of knowledge assets. Finally, it generates a full-dimensional system representation that includes interface call graphs, business rule sets, and data dependencies, achieving a leap from "fragmented understanding" to "system-level cognition."

[0105] S104: Based on full-dimensional system representation, generate structured automation capability modules that can be invoked by intelligent agents.

[0106] Among them, the structured automation capability module refers to a standardized module that transforms the full-dimensional system representation and conforms to the intelligent agent calling specification. Its core includes operation tool definition, workflow orchestration, error handling strategy, etc. It has the characteristics of reusability and portability, enabling intelligent agents to directly call the module to realize the operation and business processing of the target system without human intervention. It is a key carrier for intelligent agents to quickly adapt to new systems.

[0107] Specifically, based on information such as interface call relationships, business rules, and data dependencies contained in the full-dimensional system representation, the system cognition is transformed into standardized structured modules through an automated capability generation module. For interface calls, operation tool definitions are generated, clarifying interface call parameters, call order, and return value parsing rules; for business processes, workflow orchestration modules are generated, clarifying the sequence of each operation step, triggering conditions, and exception handling logic; and for potential errors, error handling strategies are generated, clarifying how to handle abnormal scenarios.

[0108] The intelligent agent learning method based on multi-source data provided in this application acquires multi-source operational data generated by the target system during operation, extracts multi-dimensional knowledge assets of the target system, and uses an intelligent analysis engine to dynamically fuse and analyze the multi-source operational data and multi-dimensional knowledge assets according to a preset multi-channel strategy to generate a full-dimensional system representation of the target system. Based on the full-dimensional system representation, a structured automated capability module that can be invoked by the intelligent agent is generated. This method can be adapted to documentless legacy systems, black-box systems, and heterogeneous system architectures, significantly reducing the cost of manual intervention and customized development, and effectively improving the system access adaptation efficiency and generalization ability.

[0109] Figure 2 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, a detailed description of the agent learning method based on multi-source data is provided. This method includes:

[0110] S201: Obtain multi-source operational data generated by the target system during operation.

[0111] Step S201 is similar to step S101, and will not be described again here.

[0112] Optionally, multi-source operational data includes at least two of the following: real-time operational logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data.

[0113] Among them, the real-time operation log is text data such as operation events, error messages, and status changes recorded in real time during system operation.

[0114] Component interaction data refers to the calling relationships, data transmission content, and interaction sequence data between various components (such as interfaces, modules, and services) within the target system.

[0115] User action behavior data refers to the operation records generated by users during the use of the target system, including operation type, operation time, operation path, and triggered system response.

[0116] System performance monitoring data are indicators that reflect the system's operating performance, such as CPU utilization, memory usage, interface response time, and concurrency.

[0117] Specifically, based on the characteristics of the target system and actual analysis needs, at least two of the aforementioned data types should be flexibly selected for collection. For example, in a financial risk control system, real-time operation logs (recording interface call errors and rule trigger logs) and user operation behavior data (recording risk control personnel's review operations and parameter adjustment behaviors) can be collected simultaneously. In a manufacturing PLC (Programmable Logic Controller) control system, component interaction data (interaction data between PLC modules and sensors / actuators) and system performance monitoring data (PLC operating load and response latency) can be collected simultaneously. Within the system, real-time operation logs, component interaction data, and user operation behavior data can be collected simultaneously to ensure that the collected data comprehensively covers the core scenarios of system operation.

[0118] S202: Extract the multidimensional knowledge assets of the target system.

[0119] Step S202 is similar to step S102, and will not be described again here.

[0120] Optionally, multidimensional knowledge assets include at least two of the following: business rules of the target system, architecture design documents, historical operation and maintenance experience, and domain knowledge graphs.

[0121] Among them, business rules are the logical constraints and judgment criteria that the target system must follow when performing specific business operations, such as order approval rules, data verification rules, and risk control threshold rules.

[0122] An architecture design document is a standardized document that records the architecture of a target system, its module division, interface definitions, and component relationships.

[0123] Historical operation and maintenance experience refers to the experiential knowledge accumulated by operation and maintenance personnel in the long-term maintenance of the target system, including problem solutions, fault handling methods, and optimization suggestions.

[0124] Domain knowledge graphs are knowledge networks built on the domain to which the target system belongs, and include entities, relationships and domain specifications within the domain.

[0125] Specifically, depending on the type of the target system and the analysis requirements, at least two types of knowledge assets can be flexibly selected for extraction. For example, in a medical electronic medical record system, business rules (permission verification and data format rules required for medical record storage) and domain knowledge graphs (medical terminology, medical record classification, and treatment process graphs) can be extracted simultaneously. In legacy system transformation scenarios, historical operation and maintenance experience (fault handling records and system optimization logs) and architecture design documents (if available) can be extracted simultaneously to compensate for the lack of documentation in legacy systems. In a financial risk control system, business rules (credit scoring rules and risk control interception rules) and domain knowledge graphs (financial entities and risk relationships) can be extracted simultaneously.

[0126] S203: Perform change perception analysis on structural changes, behavioral changes, or semantic changes in multi-source operational data and multi-dimensional knowledge assets to obtain change correlations.

[0127] Among them, structural changes refer to changes in the architectural hierarchy of system components, interface topology, and data models.

[0128] Behavioral changes refer to dynamic behavioral anomalies in interface calls, business operations, and operational links during system operation.

[0129] Semantic changes refer to changes in the textual semantics and business definitions of business rules, document descriptions, and field interpretations.

[0130] The relationship between changes refers to the correspondence, coupling, and linkage constraints among various change characteristics that have been identified.

[0131] Specifically, in actual execution, the engine first performs layered decomposition of preprocessed multi-source runtime data and multi-dimensional knowledge assets, extracting features from three dimensions: system architecture structure, runtime behavior trajectory, and business semantic text. Then, using a differentiated comparison algorithm, it compares the real-time collected dynamic runtime data with historical baseline data and a standardized knowledge asset library item by item, accurately locating various change points such as structural additions / reductions, behavioral deviations, and semantic changes. Simultaneously, combined with preset business constraints and topology association rules, it uncovers the linkage logic between different changes, such as changes in interface call behavior caused by component structure changes and changes in process execution brought about by business semantic updates. Finally, it sorts out, quantifies, and solidifies the coupling, causal, and temporal relationships of various changes, forming a standardized dataset of change-related relationships.

[0132] S204: Based on changing relationships, adaptive correlation modeling is performed on multi-source operational data and multi-dimensional knowledge assets to generate dynamic cognitive results.

[0133] Among them, dynamic cognitive results are the real-time cognitive state of the system output after modeling is completed. They are intermediate results that integrate static design knowledge of the system with dynamic operation changes, and accurately record the current real operation state and architecture adaptation state of the system.

[0134] Specifically, adaptive modeling and adaptation are achieved by combining multi-channel analysis strategies. For scenarios with complete operational data but missing knowledge assets, a data-driven real-time operational association model is built with dynamic behavioral change association as the core. For scenarios with complete knowledge assets but sparse operational data, a knowledge-driven architectural association model is improved based on static structure and semantic change association. For scenarios where there are conflicting changes between the two types of data, an adaptive conflict adaptation mechanism is activated to prioritize the correction of the knowledge association model to match the change characteristics of the actual operational data. At the same time, the association topology of components, interfaces, and business is reconstructed through GNN graph neural networks, and the semantic association logic is calibrated through NLP to complete the dynamic association modeling of multi-source data and knowledge, ultimately generating dynamic cognitive results that can reflect the changing characteristics of the system in real time.

[0135] S205: Based on the dynamic cognitive results, infer the business logic relationships, component dependencies, or behavioral evolution relationships in the target system to obtain the inference results.

[0136] Among them, business logic relationships refer to the logical framework of system business processes, rule constraints, and function execution.

[0137] Component dependency refers to the calling dependencies, resource dependencies, and topological dependencies between various functional components, modules, and interfaces of a system.

[0138] Behavioral evolution relationship refers to the dynamic evolution of system operation behavior as time, scenario, and business changes occur.

[0139] The inference results are derived from logical reasoning and pattern mining, revealing the underlying operational and architectural logic of the system.

[0140] Specifically, the engine uses the generated dynamic cognitive results as the basis for inference, combining a built-in rule engine and intelligent reasoning algorithms to conduct multi-dimensional relationship inference. For business logic relationships, based on semantic changes and operational behavior changes, it analyzes the execution rules, branch logic, and triggering conditions of business processes, correcting outdated and biased business rules in traditional static knowledge. For component dependency relationships, relying on structural change correlations and topology modeling results, it identifies dependency changes caused by component additions, deletions, and iterations, clarifying strong and weak dependencies between modules. For behavioral evolution relationships, combining the time-series operational data change trajectory, it mines the evolution trend, abnormal evolution characteristics, and scenario adaptation patterns of system operational behavior. Through multi-dimensional cross-validation and logical verification, invalid and erroneous associations are eliminated, ultimately outputting accurate multi-dimensional system relationship inference results.

[0141] S206: Dynamically adjust the system's cognitive structure based on the inference results to generate a full-dimensional system representation of the target system.

[0142] Among them, the system cognitive structure is the overall definition of the original static cognitive framework of the target system, which includes basic cognitive content such as fixed system function definitions, architecture models, business rules, data associations, and interface topologies.

[0143] Specifically, in the final output generation stage, the engine comprehensively iterates and optimizes the original fixed system cognitive structure based on the multi-dimensional relationship inference results. On the one hand, it corrects outdated content in the original cognition that is inconsistent with the actual operation of the system, and updates static cognitive content such as component architecture, interface relationships, and business rules; on the other hand, it incorporates dynamic cognitive content such as system behavior evolution, dynamic dependencies, and real-time change characteristics to make up for the limitations of traditional static cognition. By integrating all dimensions of information such as system functional attributes, business logic rules, data dependency links, interface call models, operational performance characteristics, and behavioral evolution patterns, the engine completes the dynamic reconstruction and upgrade of the system cognitive structure, ultimately forming a full-dimensional system representation that can truly, completely, and in real-time reflect the state of the target system, achieving a comprehensive and accurate understanding of the target system.

[0144] S207: Based on full-dimensional system representation, automatically generate structured automation capability modules that can be invoked by intelligent agents.

[0145] Among them, the structured automation capability module refers to the standardized module that transforms the full-dimensional system representation into a module that conforms to the intelligent agent calling specification. It has the characteristics of reusability, portability, and scalability. Its core is to transform system cognition into operational logic that the intelligent agent can directly execute. It can realize the intelligent agent's operation on the target system without manual coding and is a key carrier for the intelligent agent to quickly adapt to new systems.

[0146] Specifically, based on the core information contained in the full-dimensional system representation, such as system functions, business rules, and interface relationships, an automated capability generation module automatically constructs modules. This module transforms the unstructured and semi-structured information in the full-dimensional system representation into standardized structured code and configuration files, clearly defining the module's calling method, input and output parameters, and execution logic. During the generation process, the module's interfaces and business logic are fully matched with the target system's characteristics, while adhering to the intelligent agent's task execution specifications. This ensures that the generated modules can be directly called by the intelligent agent without additional adaptation development, enabling rapid adaptation of the intelligent agent to the target system.

[0147] Optional, the structured automation capability module includes: operation tool definition and workflow orchestration, and the methods also include:

[0148] Load the operational tool definitions and workflow orchestration into the agent's task execution engine;

[0149] The task execution engine parses high-level business instructions into specific operation sequences for the target system.

[0150] Among them, the definition of operation tools is a core component of the structured automation capability module, which clarifies the specific ways in which the intelligent agent calls the target system interface and components, including interface address, request parameters, return value parsing rules, error handling methods, etc.

[0151] Workflow orchestration is a standardized definition of the business processes of a target system, which clarifies the sequence of each operation step, triggering conditions, branch logic, and exception handling process.

[0152] The task execution engine is the core execution module of the intelligent agent, responsible for loading automation capability modules, parsing high-level business instructions, generating specific operation sequences, and executing them.

[0153] Specifically, firstly, the generated operation tool definitions and workflow orchestration modules are loaded into the agent's task execution engine through a standardized interface, ensuring that the engine can recognize and invoke these modules. When the agent receives a high-level business instruction (such as "complete purchase order creation and inventory synchronization"), the task execution engine invokes the loaded workflow orchestration module to parse the business process corresponding to the instruction. Then, in conjunction with the operation tool definitions, it breaks down the high-level instruction into a series of specific operation sequences (such as "call the order creation interface → call the inventory verification interface → call the inventory synchronization interface"), and clarifies the execution order, parameter configuration, and error handling methods for each operation. Finally, the operation sequence is sent to the target system to complete the automatic execution of the business instruction.

[0154] S208: Execute a two-way verification strategy, which includes: using multi-source runtime data to perform forward runtime verification on the generated structured automation capability modules to obtain runtime verification results, and using multi-dimensional knowledge assets to perform reverse consistency verification on the inferred logical rules to obtain consistency verification results.

[0155] Among them, the two-way verification strategy is the core strategy to ensure the accuracy and reliability of the structured automation capability module, which is divided into two dimensions: forward runtime verification and reverse consistency verification.

[0156] Forward runtime verification refers to simulating the target system's operating environment, calling structured automation capability modules, and using multi-source runtime data to verify the actual operating effect of the modules.

[0157] Reverse consistency verification refers to using multi-dimensional knowledge assets (such as business rules and architecture documents) to verify whether the logical rules inferred in the module are consistent with the system design intent and business specifications.

[0158] Specifically, the process begins with forward runtime verification. This involves building a simulated runtime environment identical to the target system, loading the structured automation capability module, simulating user operations and system operation scenarios, calling the module to execute specific business operations, and simultaneously collecting dynamic data during the simulated runtime. This data is then compared with multi-source runtime data to verify whether the module's operation paths, interface calls, and error handling conform to the actual system operation logic, yielding runtime verification results. Following this, reverse consistency verification is performed. Logical rules from the structured automation capability module are extracted and compared with business rules and architecture design documents in the multi-dimensional knowledge assets to verify whether the inferred logical rules align with the system design intent. Figure 1 To verify whether the module logic conforms to domain specifications, a consistency verification result is obtained to ensure the accuracy of the module logic.

[0159] S209: When the runtime verification results and / or consistency verification results are inconsistent with the expected results, an error report is generated and a re-analysis process is automatically triggered. The analysis strategy weights of each channel in the multi-channel strategy are dynamically adjusted based on the error feedback information contained in the error report.

[0160] Among them, the error report refers to the detailed report generated when the two-way verification result is inconsistent with the expected result, which includes the error type, error scenario, error cause, and error location.

[0161] The reanalysis process refers to the process of re-invoking the intelligent analysis engine and re-performing multi-source data fusion analysis based on error feedback information.

[0162] Analysis strategy weight refers to the analysis priority and weight allocation between the running data channel and the knowledge asset channel in a multi-channel strategy. Adjusting the weight can optimize the accuracy of the fusion analysis.

[0163] Specifically, the system compares runtime verification results, consistency verification results, and preset expected results in real time. When any verification result deviates from the expectation, an error report generation process is automatically triggered, recording detailed error information (such as "interface call order is inconsistent with business rules" or "inferred inventory verification rules do not match document definitions"). After the error report is generated, a re-analysis process is automatically triggered. The intelligent analysis engine reads the feedback information in the error report, locates the root cause of the error (such as unreasonable multi-channel strategy weight allocation or data parsing deviation), and dynamically adjusts the analysis weight of each channel in the multi-channel strategy. For example, if the error originates from a deviation in business rule inference, the analysis weight of the knowledge asset channel can be increased; if the error originates from a deviation in runtime behavior parsing, the analysis weight of the runtime data channel can be increased, and multi-source data fusion analysis and automated capability module generation are performed again, forming a closed loop of "analysis → verification → optimization".

[0164] The intelligent agent learning method based on multi-source data provided in this application collects various types of data from the target system during runtime and mines multi-dimensional knowledge assets from them. It perceives changes in the structure, behavior, and semantics of data and knowledge, identifies correlation patterns, and completes adaptive correlation modeling accordingly, outputting dynamic cognitive results. Based on these results, it infers the evolutionary relationships of system business, components, and behaviors, and iteratively optimizes the system's cognitive structure by combining the inference conclusions, forming a complete, full-dimensional system representation. Based on this representation, it automatically constructs structured, automated capability modules that can be invoked by the intelligent agent and conducts bidirectional verification: on the one hand, it uses runtime data to positively verify the module's operational effect; on the other hand, it uses knowledge assets to reversely verify the consistency of logical rules. If the verification result deviates from expectations, an error report is immediately generated and a re-analysis process is initiated, while the strategy weights of each analysis channel are dynamically adjusted based on the error feedback. This method is adaptable to undocumented legacy systems, black-box systems, and heterogeneous architecture systems, significantly reducing manual operation and customized development costs, and improving system adaptation efficiency and versatility.

[0165] Figure 3 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 3 ,like Figure 3 As shown, in this embodiment... Figure 2 Based on the implementation examples, this paper details the specific implementation method for performing change-aware analysis on structural, behavioral, or semantic changes in multi-source operational data and multi-dimensional knowledge assets to obtain change correlations. The method includes:

[0166] S301: Divide multi-source runtime data according to a preset time window, and divide multi-dimensional knowledge assets according to version iteration records to form multiple change sets.

[0167] The preset time window is a fixed time interval set in advance based on the target system's operating characteristics and data generation frequency. Common time intervals include hourly, daily, and business cycle levels. It is used to standardize and segment continuously generated real-time operating data.

[0168] Version iteration records are complete historical records of multidimensional knowledge assets during long-term revision, update, and iteration. They include core information such as version number, update time, and specific changes, and can accurately trace the iteration trajectory of knowledge assets.

[0169] The change set is the basic analysis unit generated in this step. It is divided into two categories: runtime data change set and knowledge asset change set. Specifically, it refers to the runtime data or knowledge asset dataset within a single time window after the division is completed. It is the core foundation for realizing incremental joint understanding and iterative cognition of the system.

[0170] Specifically, a differentiated approach of classification and segmentation is used to construct change sets. For multi-source operational data, the first step is to match and adapt time windows based on the target system's business scenario and data generation frequency. For example, high-frequency data in a financial risk control system is adapted to a 1-hour time window, while low-frequency business data is adapted to a 1-day time window. Then, through precise matching using data timestamps, multi-source operational data such as real-time logs and component interaction data are categorized into corresponding time windows. The complete operational data within each independent time window constitutes an operational data change set. For multi-dimensional knowledge assets, the version iteration records of assets such as interface documents, business rules, and architecture materials are comprehensively reviewed. According to the version update sequence, different versions of knowledge assets are independently split to form corresponding version knowledge asset change sets. The version information, update time, and core change content of each change set are uniformly labeled, and finally, standardized and iterative multi-type change sets are generated in batches.

[0171] S302: Obtain the current change set from multiple change sets, extract and update the behavioral features or business path information associated with the current change set, and use a lightweight model to dynamically adjust the business rule model generated in the previous iteration to form the incremental cognitive results of this iteration.

[0172] The current change set refers to the target change set to be analyzed incrementally within the current iteration cycle. It can be selected sequentially according to the iteration order of time and version, and serves as the analysis carrier for single incremental cognition.

[0173] Behavioral characteristics are the dynamic operational attributes of the system carried in multi-source operational data, including operational behavior information such as interface call frequency, business operation path, and exception triggering mode.

[0174] Business path information is static rule information such as standardized business processes, interface call specifications, and business execution logic defined in multidimensional knowledge assets.

[0175] Lightweight models are intelligent models that require low computing power and do not require full retraining. They rely on incremental update algorithms and lightweight neural networks to achieve rapid iterative optimization.

[0176] Incremental cognitive results are lightweight cognitive outcomes of a single iteration, only including the newly added and modified cognitive content corresponding to the current change set, without overlaying historical iteration information, thus achieving incremental updates of the system's cognition.

[0177] Specifically, the current change set to be analyzed is selected sequentially according to the iteration order, and differentiated feature extraction and information updates are performed based on the change set type. If the current change set is a runtime data change set, the feature extraction module captures core behavioral features such as new interface call chains, changes in business operation paths, and new anomaly triggering patterns, and iteratively updates the system behavioral feature library and business path records simultaneously. If the current change set is a knowledge asset change set, the changes in the new version of knowledge assets, such as rule additions, interface definition revisions, and process optimizations, are accurately extracted to clarify the adjustment direction and correction logic of the business path. Subsequently, all change information is input into a lightweight model. The model does not need to be fully retrained on the historical complete business rule model; it only fine-tunes the model parameters, updates rule nodes, and corrects business logic for the current changes. Finally, the updated behavioral features, business path information, and optimized business rule model are integrated to generate incremental cognitive results containing only the content of this iteration.

[0178] S303: Using dependency analysis techniques, determine the extent of the impact of changed elements in incremental cognitive results on the structured automation capability modules generated in previous iterations.

[0179] Among them, dependency analysis technology is a specialized technology used to mine the relationships between various core elements of a system. Relying on algorithms such as graph traversal and dependency graph analysis, it accurately identifies the coupling relationships between elements such as business rules, interface definitions, and business paths. Its core purpose is to locate the linkage relationship between changed content and existing automation modules.

[0180] Change elements refer to the system cognitive content that has changed in this incremental cognitive result, including core content such as newly added, modified, and deleted business rules, interface paths, and process logic.

[0181] The scope of impact refers to the existing structured and automated capability modules that the changed elements can influence and affect. Specifically, it covers local content of modules such as operation tool definition and workflow orchestration. It is used to accurately define the target scope of subsequent updates and avoid invalid updates.

[0182] Specifically, based on the incremental cognitive results generated in this iteration, a comprehensive dependency and correlation analysis is conducted. First, all changed elements in the incremental cognitive results are identified and extracted, clarifying the specific system cognitive content that has been added, modified, or deleted. Then, dependency analysis technology is used to construct a dedicated dependency graph between the changed elements and the structured automation capability modules generated in previous iterations, clearly mapping the topological relationships between the changed elements and various operation tools and workflow orchestration modules. Simultaneously, combined with graph traversal algorithms, the transmission paths and linkage logic of the changed elements are traced layer by layer, accurately identifying the local content of automation modules affected by the changes, distinguishing between affected modules and unrelated modules, and finally outputting accurate and detailed results determining the scope of module impact.

[0183] S304: Based on the scope of impact, perform a local update operation on the structured automation capability modules generated in the previous iteration that are affected, in order to generate change relationships.

[0184] Among them, partial update operation is a lightweight iterative optimization method that is different from full update. It only modifies, supplements or deletes the partial content of the affected automation module corresponding to the changed element, and completely preserves the existing modules and cognitive content of the system that are not affected. It has the advantages of low time consumption, low computing power and high efficiency.

[0185] The structured automation capability module is a standardized system capability carrier formed in the previous iterations, including core modules such as operation tool definition and business workflow orchestration.

[0186] Specifically, based on the defined scope of impact, precise and differentiated local updates are executed. For affected operational tool definition modules, details such as interface call parameters, compliance verification rules, and exception handling logic are supplemented or corrected according to the changed elements. For affected workflow orchestration modules, the order of business operation steps is adjusted accordingly, new compliance verification and other business links are added, and invalid process nodes are deleted. During the update process, the principle of local updates is strictly followed, fully preserving the content of all modules unaffected by the changes. At the same time, the consistency between the updated module logic and the incremental cognitive results and the actual system operation logic is verified to ensure that the update is accurate and effective. After completing the local iterative optimization of the modules, all changes and updates are integrated to form a complete system change relationship, realizing the synchronous iteration of the system automation capability modules and system cognition.

[0187] Figure 4 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 4 ,like Figure 4 As shown, in this embodiment... Figure 2 Based on the examples, this paper details how to perform adaptive correlation modeling on multi-source operational data and multi-dimensional knowledge assets based on changing relationships to generate dynamic cognitive results. The method includes:

[0188] S401: Obtain integrity metrics for multi-source operational data and availability metrics for multi-dimensional knowledge assets.

[0189] Among them, the integrity index of multi-source operation data is the core parameter for quantitatively evaluating the quality and coverage of operation data such as real-time logs and component interaction data of the system. It is mainly composed of dimensions such as interface call chain coverage, data field integrity rate, abnormal data ratio, and data temporal continuity. It is used to characterize whether the current operation data can comprehensively, continuously, and reliably support the analysis of system behavior.

[0190] Multidimensional knowledge asset availability metrics are core parameters for quantitatively evaluating the completeness and availability of knowledge assets such as business rules, architecture documents, and interface specifications. These metrics include dimensions such as document completeness rate, rule clarity, update timeliness, and content parsing success rate. They are used to measure whether static knowledge assets can accurately and effectively assist in system logic parsing and architecture understanding.

[0191] Specifically, during execution, the system performs dual-indicator quantitative calculations through a data quality inspection module and a knowledge asset assessment module. For multi-source operational data, it statistically analyzes the coverage ratio of core interface call records, the percentage of missing data fields, the number and percentage of abnormal logs, and the continuity of time-series data, quantifying each item to obtain a completeness indicator value. For multi-dimensional knowledge assets, it verifies the completeness of the architecture documents and business rules, the clarity of rule descriptions, the timeliness of asset updates, and the success rate of automatic machine parsing, comprehensively evaluating the accuracy and usability of the knowledge assets, and finally outputting standardized, comparable usability indicator values.

[0192] S402: Based on integrity and availability indicators, adjust the weight allocation of the runtime data analysis channel and the knowledge asset analysis channel through a dynamic channel selection algorithm to form a preliminary fusion result.

[0193] Among them, the dynamic channel selection algorithm is a core algorithm based on the adaptive allocation of analysis weights using dual-dimensional quantitative indicators. It is used to dynamically adjust the analysis priority and calculation ratio of the running data analysis channel and the knowledge asset analysis channel.

[0194] The runtime data analysis channel focuses on real-time runtime data, analyzing dynamic features such as system dynamic behavior, interface interactions, and operation paths.

[0195] The knowledge asset analysis channel takes static knowledge assets as its core and focuses on the static logic analysis of system architecture design, business rules, interface specifications, etc.

[0196] The preliminary fusion result is an intermediate result of the initial fusion after independent analysis of the two channels according to their weights, achieving the initial aggregation of dynamic and static information, but has not yet undergone logical conflict verification and semantic correction.

[0197] Specifically, the obtained integrity and availability metrics are input into a dynamic channel selection algorithm, which adaptively allocates weights based on a preset weight model. When operational data integrity is high and knowledge asset availability is low, the weight of the operational data analysis channel is increased, adopting an analysis mode dominated by operational data and supplemented by knowledge assets. When knowledge asset availability is high and operational data integrity is poor, the weight of the knowledge asset analysis channel is increased, relying on static rules and architectural knowledge to dominate the analysis. When both metrics are at a high level, a balanced weighting of the two channels is adopted to achieve dynamic and static collaborative analysis. The two channels conduct specialized analyses according to their allocated weights. The operational data channel extracts operation paths and interaction behaviors, while the knowledge asset channel parses business rules and architectural logic. Finally, the analysis results from the two channels are initially spliced ​​and merged to generate a preliminary fusion result that has not yet undergone consistency verification.

[0198] S403: Based on the preliminary fusion results, detect the logical consistency between multi-source operational data and multi-dimensional knowledge assets.

[0199] Specifically, the system calls a conflict detection module to perform item-by-item consistency checks on the preliminary fusion results. First, two types of core logical information are extracted from the preliminary fusion results: one is the actual operational logic derived from multi-source operational data, including the actual interface call sequence, on-site executed business branch rules, and the actual usage status of data fields; the other is the standard design logic derived from multi-dimensional knowledge assets, including the interface call specifications stipulated in the documents, standardized business constraints, and data model structure definitions. Combining a rule engine and semantic analysis technology, the operational logic and design logic are compared and verified point by point, accurately identifying logical points with deviations, contradictions, and mismatches, thus completing the overall logical consistency check.

[0200] S404: When there is a logical inconsistency between multi-source operational data and multi-dimensional knowledge assets, the data sources with inconsistencies are prioritized based on the business context, and the initial fusion results are semantically corrected based on the high-priority data sources to obtain semantic verification results.

[0201] Among them, the business context refers to the process specifications, operating rules, operational requirements and scenario constraints of the target system in the corresponding business scenario, and is the core basis for conflict determination and priority selection.

[0202] The semantic verification result is a standardized fusion result that has undergone conflict correction, logical unification, and is consistent with real business scenarios.

[0203] Specifically, when a logical inconsistency is detected, the first step is to locate the specific business scenario corresponding to the conflict, extract the business context information of the current scenario, and clarify the core execution specifications and business objectives of the scenario. This information is then used to determine the priority of the data source. In typical business scenarios, the actual system operation data accurately reflects the on-site execution status and has a higher priority than document knowledge assets, which may be outdated or biased. If the knowledge asset is the latest iteration of core business rules, and the conflicting operation data is invalid data such as temporary testing or abnormal operations, then the knowledge asset is determined to have a high priority. After determining the priority, the logic of the high-priority data source is used as the standard to specifically correct, replace, and regulate the conflicting semantics, contradictory rules, and biased links in the preliminary fusion results. Finally, a logically unified semantic verification result that aligns with actual business needs is output.

[0204] S405: Generate dynamic cognitive results based on the fusion results after semantic correction.

[0205] Among them, dynamic cognitive results are real-time cognitive results of the system obtained after weight adaptive fusion, logical consistency detection and business semantic correction. Unlike static solidified cognition, dynamic cognitive results can be synchronously compatible with the static design logic of the system and dynamic operation changes. They accurately reflect the current real business execution logic, component interaction relationship and behavior operation rules of the target system and are the core intermediate results for iteratively generating full-dimensional system representation.

[0206] Specifically, the system takes the semantically corrected, logically unified, and scenario-adapted fusion result as its core input to complete the integrated generation of dynamic system cognition. Through the semantic aggregation module, the corrected multi-dimensional analysis information is organized, refined, and structured. This unifies and integrates business rules, interface relationships, data dependencies, operation paths, component interaction logic, and other content obtained from dual-channel parsing, eliminating conflicting and redundant information, retaining valid, accurate, and up-to-date system cognition content, reconstructing the system's current dynamic cognition system, and ultimately generating dynamic cognition results that can accurately depict the system's actual operating state and business logic in real time.

[0207] Figure 5 A flowchart illustrating the agent learning method based on multi-source data provided in this application. Figure 5 ,like Figure 5 As shown, in this embodiment... Figure 2 Based on the examples, this paper details the method of inferring business logic relationships, component dependencies, or behavioral evolution relationships in a target system based on dynamic cognitive results, and obtaining the inference results. The method includes:

[0208] S501: Construct a dynamic graph of runtime data based on multi-source runtime data, where the behavior sequence or component call relationship in the multi-source runtime data is mapped to the nodes of the dynamic graph, and the execution timing or interaction logic between each node is mapped to the edges of the dynamic graph.

[0209] Among them, the dynamic graph of running data is a graph structure model built on multi-source running data. It has the characteristic of real-time updating and is used to intuitively depict the real running status of the system.

[0210] A behavior sequence is a collection of user operations, API calls, and other behaviors arranged chronologically. Component call relationships are the call associations between modules, interfaces, and services within the system.

[0211] Nodes are the basic units of a dynamic graph, carrying the core content of behavior and component calls, while edges are used to express the interaction logic between nodes, such as execution sequence and call triggering.

[0212] Specifically, the multi-source runtime data is first preprocessed to remove invalid and redundant data, and standardized behavior sequences and component call relationships are extracted. A graph is constructed according to preset mapping rules: each behavior and each group of component call relationships is set as an independent node, and node attributes such as behavior type, component name, and call time are labeled; then, the nodes are connected to form edges based on the actual execution order and the interaction relationship between the caller and the called, while the temporal relationship and interaction type are labeled on the edges, ultimately completing the construction of the dynamic graph of the runtime data.

[0213] S502: Construct a static graph of knowledge assets based on multidimensional knowledge assets, where the structural definition information and business logic rules in the multidimensional knowledge assets are mapped to the attribute information of the nodes in the static graph of knowledge assets.

[0214] Among them, the knowledge asset static graph is a graph structure generated based on multi-dimensional knowledge assets. Its state is stable and does not change with real-time running data. It is used to reflect the original design logic of the system.

[0215] The structural definition information includes standardized content such as system architecture, interface specifications, and data models, while business logic rules are the constraints that business execution must follow.

[0216] Node attribute information is the descriptive content attached to the static graph nodes, used to fully represent the relevant elements of the system's static design.

[0217] Specifically, multi-dimensional knowledge assets such as interface documents, architecture materials, and business rules are classified, analyzed, and standardized to extract structural definition information and business logic rules. Core elements such as system components, interfaces, data tables, and business rules are mapped to static graph nodes, and the extracted structural definitions and business rules are uniformly set as attribute information of the corresponding nodes. The nodes are arranged according to the inherent architecture and relationships of the system, and finally, a static graph of knowledge assets is generated.

[0218] S503: Input the dynamic graph of the running data and the static graph of the knowledge assets into the pre-trained graph neural network model. The graph neural network model aggregates the node features of the dynamic graph and the attribute information of the static graph to generate the global semantic representation vector of the target system. Based on the global semantic representation vector, the implicit business logic relationship is inferred to obtain the inference result.

[0219] Among them, pre-trained graph neural networks (GNNs) are models trained on massive amounts of similar graph data, and are good at processing graph structure data, realizing feature aggregation and association mining.

[0220] Node characteristics are dynamic runtime information carried by dynamic graph nodes, such as call frequency and execution duration, while attribute information is static content such as interface definitions and business rules corresponding to static graph nodes.

[0221] Global semantic representation vectors are low-dimensional representations formed by integrating dynamic and static information, while implicit business logic relationships are potential processes and constraints that are not explicitly recorded in data and knowledge but are obtained through semantic reasoning.

[0222] Specifically, the process begins by aligning nodes in the dynamic graph of the running data and the static graph of the knowledge assets using a graph matching algorithm, ensuring semantic correspondence between the two types of graphs. The aligned graphs are then input into a pre-trained GNN model, which performs deep aggregation and fusion of dynamic node features and static node attributes, outputting a global semantic representation vector that uniformly represents the system state. Semantic reasoning is then conducted based on this vector to uncover potential processes, data dependencies, and rule linkages not directly reflected in the data and documents, ultimately outputting a complete business logic inference result.

[0223] Figure 6 A schematic diagram of the structure of the intelligent agent learning device based on multi-source data provided in this application is shown below. Figure 6 As shown, the intelligent agent learning device 600 based on multi-source data provided in this embodiment includes:

[0224] The acquisition module 601 is used to acquire multi-source operational data generated by the target system during operation;

[0225] Extraction module 602 is used to extract multidimensional knowledge assets from the target system;

[0226] The generation module 603 is used to utilize the intelligent analysis engine to dynamically fuse and analyze multi-source operating data and multi-dimensional knowledge assets according to a preset multi-channel strategy, and generate a full-dimensional system representation of the target system.

[0227] The generation module 603 is also used to generate a structured automation capability module that can be invoked by an intelligent agent based on the full-dimensional system representation.

[0228] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: an analysis module 604 and an inference module 605;

[0229] Analysis module 604 is used to perform change perception analysis on structural changes, behavioral changes, or semantic changes in multi-source operational data and multi-dimensional knowledge assets to obtain change correlations;

[0230] The generation module 603 is also used to perform adaptive correlation modeling on multi-source operational data and multi-dimensional knowledge assets based on changing correlation relationships, and generate dynamic cognitive results;

[0231] The inference module 605 is used to infer the business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system based on the dynamic cognitive results, and obtain the inference results;

[0232] The generation module 603 is used to dynamically adjust the system's cognitive structure based on the inference results and generate a full-dimensional system representation of the target system.

[0233] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a division module 606, an adjustment module 607, and a determination module 608;

[0234] The partitioning module 606 is used to partition multi-source runtime data according to a preset time window and partition multi-dimensional knowledge assets according to version iteration records to form multiple change sets;

[0235] The adjustment module 607 is used to obtain the current change set from multiple change sets, extract and update the behavioral features or business path information associated with the current change set, and dynamically adjust the business rule model generated in the previous iteration using a lightweight model to form the incremental cognitive result of this iteration.

[0236] Module 608 is used to determine the scope of the impact of changed elements in the incremental cognitive results on the structured automation capability modules generated in previous iterations, using dependency analysis techniques.

[0237] The generation module 603 is also used to perform local update operations on the affected previously generated structured automated capability modules based on the scope of influence, in order to generate change relationships.

[0238] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a detection module 609;

[0239] The acquisition module 601 is also used to acquire integrity indicators of multi-source operational data and availability indicators of multi-dimensional knowledge assets;

[0240] The generation module 603 is also used to adjust the weight allocation of the running data analysis channel and the knowledge asset analysis channel according to the integrity index and the availability index through a dynamic channel selection algorithm to form a preliminary fusion result;

[0241] The detection module 609 is used to detect the logical consistency between multi-source operational data and multi-dimensional knowledge assets based on the preliminary fusion results.

[0242] The generation module 603 is also used to determine the priority of inconsistent data sources based on the business context when there is a logical inconsistency between multi-source running data and multi-dimensional knowledge assets, and to perform semantic correction on the preliminary fusion result based on the high-priority data source to obtain the semantic verification result.

[0243] The generation module 603 is also used to generate dynamic cognitive results based on the semantically corrected fusion results.

[0244] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a construction module 610;

[0245] Module 610 is used to construct a dynamic graph of runtime data based on multi-source runtime data. In this graph, the behavior sequence or component call relationship in the multi-source runtime data is mapped to the nodes of the dynamic graph, and the execution sequence or interaction logic between the nodes is mapped to the edges of the dynamic graph.

[0246] The construction module 610 is also used to construct a static graph of knowledge assets based on multidimensional knowledge assets, wherein the structural definition information and business logic rules in the multidimensional knowledge assets are mapped to the attribute information of the nodes in the static graph of knowledge assets.

[0247] The generation module 603 is also used to input the dynamic graph of running data and the static graph of knowledge assets into a pre-trained graph neural network model. The graph neural network model aggregates the node features of the dynamic graph and the attribute information of the static graph to generate a global semantic representation vector of the target system. Based on the global semantic representation vector, the implicit business logic relationship is inferred to obtain the inference result.

[0248] As an optional implementation, the determination module 608 is also used for multi-source operational data including at least two of the following: real-time operational logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data; and multi-dimensional knowledge assets including at least two of the following: business rules of the target system, architecture design documents, historical operation and maintenance experience, and domain knowledge graph.

[0249] As an optional implementation, the generation module 603 is also used to execute a two-way verification strategy, which includes: using multi-source runtime data to perform forward runtime verification on the generated structured automation capability module to obtain runtime verification results, and using multi-dimensional knowledge assets to perform reverse consistency verification on the inferred logical rules to obtain consistency verification results.

[0250] The adjustment module 607 is used to generate an error report and automatically trigger a re-analysis process when the runtime verification results and / or consistency verification results are inconsistent with the expected results. Based on the error feedback information contained in the error report, the analysis strategy weights of each channel in the multi-channel strategy are dynamically adjusted.

[0251] As an optional implementation, the intelligent agent learning device based on multi-source data further includes: a loading module 611 and a parsing module 612;

[0252] Loading module 611 is used to load operation tool definitions and workflow orchestrations into the task execution engine of the agent;

[0253] The parsing module 612 is used to parse high-level business instructions into specific operation sequences for the target system through the task execution engine.

[0254] Figure 7 A schematic diagram of the structure of the intelligent agent learning device based on multi-source data provided in this application. Figure 7 As shown, this application provides an intelligent agent learning device based on multi-source data. The intelligent agent learning device 700 based on multi-source data includes: a receiver 701, a transmitter 702, a processor 703, and a memory 704.

[0255] Receiver 701 is used to receive commands and data;

[0256] Transmitter 702 is used to send commands and data;

[0257] Memory 704 is used to store instructions executed by the computer;

[0258] The processor 703 is used to execute computer execution instructions stored in the memory 704 to implement the various steps of the intelligent agent learning method based on multi-source data in the above embodiments. For details, please refer to the relevant descriptions in the foregoing embodiments of the intelligent agent learning method based on multi-source data.

[0259] Optionally, the memory 704 can be either standalone or integrated with the processor 703.

[0260] When the memory 704 is set up independently, the electronic device also includes a bus for connecting the memory 704 and the processor 703.

[0261] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0262] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0263] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0264] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0265] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0266] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0267] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0268] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0269] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0270] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for intelligent agent learning based on multi-source data, characterized in that, include: Acquire multi-source operational data generated by the target system during operation, wherein the multi-source operational data includes at least two of the following: real-time operation logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data; Extract the multidimensional knowledge assets of the target system, which include at least two of the following: business rules, architecture design documents, historical operation and maintenance experience, and domain knowledge graphs of the target system. Using an intelligent analysis engine, the multi-source operational data and the multi-dimensional knowledge assets are dynamically fused and analyzed according to a preset multi-channel strategy to generate a full-dimensional system representation of the target system. The multi-channel strategy refers to three preset analysis strategies: operational data-driven and knowledge asset-assisted, knowledge asset-driven and operational data-assisted, and dual-channel collaborative, which can be dynamically adjusted according to data characteristics. Based on the full-dimensional system representation, a structured automation capability module that can be invoked by an intelligent agent is generated; The process involves using an intelligent analysis engine to dynamically fuse and analyze the multi-source operational data with the multi-dimensional knowledge assets according to a preset multi-channel strategy, generating a full-dimensional system representation of the target system, including: A change perception analysis is performed on the structural, behavioral, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain the change correlation. Based on the aforementioned changing relationships, adaptive correlation modeling is performed on the multi-source operational data and the multi-dimensional knowledge assets to generate dynamic cognitive results. Based on the dynamic cognitive results, the business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system are inferred to obtain the inference results; Based on the inference results, the system's cognitive structure is dynamically adjusted to generate a full-dimensional system representation of the target system.

2. The method according to claim 1, characterized in that, The step of performing change-aware analysis on the structural, behavioral, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain change correlations includes: The multi-source operational data is divided according to a preset time window, and the multi-dimensional knowledge assets are divided according to version iteration records to form multiple change sets; The current change set is obtained from the multiple change sets, and the behavioral features or business path information associated with the current change set is extracted and updated. The business rule model generated in the previous iteration is dynamically adjusted using a lightweight model to form the incremental cognitive result of this iteration. By using dependency analysis techniques, the scope of the impact of the changed elements in the incremental cognitive results on the structured automation capability modules generated in previous iterations is determined. Based on the scope of influence, a local update operation is performed on the affected structured automation capability modules generated in previous iterations to generate change relationships.

3. The method according to claim 1, characterized in that, The process of performing adaptive correlation modeling on the multi-source operational data and the multi-dimensional knowledge assets based on the changing correlations to generate dynamic cognitive results includes: Obtain the integrity metrics of the multi-source operational data and the availability metrics of the multi-dimensional knowledge assets; Based on the integrity index and the availability index, the weight allocation of the running data analysis channel and the knowledge asset analysis channel is adjusted through a dynamic channel selection algorithm to form a preliminary fusion result; Based on the preliminary fusion results, the logical consistency between the multi-source operational data and the multi-dimensional knowledge assets is detected; When there is a logical inconsistency between the multi-source operational data and the multi-dimensional knowledge assets, the data sources with inconsistencies are prioritized based on the business context, and the preliminary fusion results are semantically corrected based on the high-priority data sources to obtain semantic verification results. The dynamic cognitive result is generated based on the semantically corrected fusion result.

4. The method according to claim 1, characterized in that, Based on the dynamic cognitive results, the inference of business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system is performed to obtain the inference results, including: A dynamic graph of operational data is constructed based on the multi-source operational data, wherein the behavior sequence or component call relationship in the multi-source operational data is mapped to the nodes of the dynamic graph, and the execution sequence or interaction logic between each node is mapped to the edges of the dynamic graph. A static graph of knowledge assets is constructed based on the multidimensional knowledge assets, wherein the structural definition information and business logic rules in the multidimensional knowledge assets are mapped to the attribute information of the nodes in the static graph of knowledge assets. The dynamic graph of the running data and the static graph of the knowledge assets are input into a pre-trained graph neural network model. The graph neural network model aggregates the node features of the dynamic graph and the attribute information of the static graph to generate a global semantic representation vector of the target system. The implicit business logic relationship is inferred based on the global semantic representation vector to obtain the inference result.

5. The method according to claim 1, characterized in that, After generating a structured automation capability module that can be invoked by an intelligent agent based on the full-dimensional system representation, the method further includes: The two-way verification strategy includes: performing forward runtime verification on the generated structured automation capability module using the multi-source runtime data to obtain runtime verification results, and performing reverse consistency verification on the inferred logical rules using the multi-dimensional knowledge assets to obtain consistency verification results. When the runtime verification result and / or the consistency verification result are inconsistent with the expected result, an error report is generated and a re-analysis process is automatically triggered. Based on the error feedback information contained in the error report, the analysis strategy weight of each channel in the multi-channel strategy is dynamically adjusted.

6. The method according to claim 1, characterized in that, The structured automation capability module includes: operation tool definition and workflow orchestration; the method further includes: Load the operation tool definitions and workflow orchestrations into the agent's task execution engine; The task execution engine parses high-level business instructions into specific operation sequences for the target system.

7. An intelligent agent learning device based on multi-source data, characterized in that, include: The acquisition module is used to acquire multi-source operational data generated by the target system during operation. The multi-source operational data includes at least two of the following: real-time operation logs of the target system, component interaction data, user operation behavior data, and system performance monitoring data. An extraction module is used to extract multidimensional knowledge assets of the target system, wherein the multidimensional knowledge assets include at least two of the following: business rules of the target system, architecture design documents, historical operation and maintenance experience, and domain knowledge graph. The generation module is used to utilize an intelligent analysis engine to dynamically fuse and analyze the multi-source operational data and the multi-dimensional knowledge assets according to a preset multi-channel strategy, thereby generating a full-dimensional system representation of the target system. The multi-channel strategy refers to three preset analysis strategies: operational data-driven and knowledge asset-assisted, knowledge asset-driven and operational data-assisted, and dual-channel collaborative, which can be dynamically adjusted according to data characteristics. The generation module is also used to generate a structured automation capability module that can be invoked by an intelligent agent based on the full-dimensional system representation. The analysis module is used to perform change perception analysis on the structural changes, behavioral changes, or semantic changes in the multi-source operational data and the multi-dimensional knowledge assets to obtain the change correlation. The generation module is also used to perform adaptive association modeling on the multi-source operational data and the multi-dimensional knowledge assets based on the changing correlation, and generate dynamic cognitive results; The inference module is used to infer the business logic relationships, component dependency relationships, or behavioral evolution relationships in the target system based on the dynamic cognitive results, and obtain the inference results; The generation module is also used to dynamically adjust the system's cognitive structure based on the inference results, and generate a full-dimensional system representation of the target system.

8. An intelligent agent learning device based on multi-source data, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

Images