Method and system for enhancing and correcting expert intervention agents embedded in business processes

Abstract

The application discloses an embedded business process expert intervention intelligent agent enhancement and correction method and system, first, an intelligent agent decision module is deployed at a decision node of a business process, business scene data and process context information are acquired, an initial decision scheme and basis are generated, second, a multi-dimensional decision credibility dynamic evaluation model is constructed, a comprehensive credibility score is calculated and a corresponding risk signal is generated, then a hierarchical expert intervention triggering mechanism is executed to determine whether to start expert intervention, if intervention is triggered, the decision context and uncertainty analysis result are pushed to an expert interactive workbench for review, correction or veto to generate expert feedback results, finally, the feedback results are synchronized to the business process to update the decision output and used as labeled data for online fine-tuning and offline training of the intelligent agent, the scheme realizes quantitative identification and hierarchical control of intelligent agent decision risk, improves decision credibility and reduces unnecessary manual intervention.

Description

Technical Field

[0001] This invention belongs to the field of power system control technology, and in particular relates to a method and system for enhancing and correcting expert-intervention intelligent agents embedded in business processes. Background Technology

[0002] With the continuous development of artificial intelligence technology, intelligent agents are being applied more and more widely in complex business scenarios such as finance, energy, and power regulation. In the field of power system regulation, as the scale of the power grid expands and the complexity of its operation increases, traditional decision-making methods that rely on human experience are no longer able to meet the requirements of real-time performance and efficiency. Therefore, using intelligent agents to analyze operational data and automatically generate regulation decisions is gradually becoming an important technical means to improve the operational efficiency and intelligence level of the power system.

[0003] In existing technologies, business data is typically analyzed using machine learning models or rule-based and data-fusion decision models to generate corresponding decision results. Some systems verify the agent's decision results through confidence assessment, rule validation, or anomaly detection. When anomalies or non-compliance with business rules are detected, corrections are made through manual review or expert intervention. Simultaneously, some systems record historical decisions and manual correction information for subsequent model training or periodic updates to improve the agent's decision-making capabilities.

[0004] However, existing technologies still have significant shortcomings: on the one hand, intelligent agent decision-making systems usually operate as independent modules, with insufficient integration with actual business processes, making it difficult to dynamically adjust decision-making strategies according to business scenarios; on the other hand, existing expert intervention mechanisms are mostly simple manual review or passive triggering, lacking reasonable triggering logic, which can easily lead to excessive intervention or untimely intervention; in addition, expert feedback is usually only used for current decision correction, making it difficult to form a continuous optimization mechanism, thus limiting the long-term improvement of intelligent agent decision-making capabilities. Summary of the Invention

[0005] Purpose of the Invention: The purpose of this invention is to provide a method and system for enhancing and correcting intelligent agents with expert intervention embedded in business processes. By constructing a decision credibility assessment and hierarchical expert intervention mechanism, and forming a closed-loop optimization process driven by expert feedback, the invention aims to achieve deep integration of intelligent agent decision-making and business processes. This will improve decision credibility, reduce unnecessary expert intervention, and promote the continuous optimization of intelligent agent decision-making capabilities.

[0006] Technical solution: The expert-interventional intelligent agent enhancement and correction method embedded in business processes described in this invention includes the following steps:

[0007] S1. Deploy intelligent agent decision-making modules at key decision-making nodes in the business process, obtain business scenario data and process context information of the corresponding nodes, and generate initial decision schemes, uncertainty analysis results and corresponding decision basis based on the business scenario data;

[0008] S2. Construct a multi-dimensional dynamic evaluation model for decision credibility, evaluate the initial decision scheme, calculate the comprehensive credibility score, and generate corresponding risk signals based on the comprehensive credibility score;

[0009] S3. Based on the comprehensive credibility score and risk signal, execute the hierarchical expert intervention triggering mechanism to determine whether to initiate the expert intervention process.

[0010] S4. When the expert intervention process is triggered, the business scenario data, process context information, initial decision plan, uncertainty analysis results and risk signals are pushed to the expert interactive workbench, where experts review, correct or reject them and generate expert feedback results.

[0011] S5. Synchronize the expert feedback results to the business process to update the decision output, and use the expert feedback results as labeled data for online fine-tuning and offline training of the agent model to achieve continuous optimization of the agent's decision-making ability.

[0012] This solution deploys an intelligent agent decision-making module at key nodes of the business process and constructs a multi-dimensional dynamic evaluation model for decision credibility. This enables real-time generation of comprehensive credibility scores and risk signals for initial decision-making schemes. Based on these scores and signals, a tiered expert intervention triggering mechanism is implemented to precisely screen cases requiring human intervention, initiating expert processes only when necessary. Once triggered, a complete context, including decision justification and uncertainty analysis, is pushed to the expert workbench, supporting efficient expert review and correction. Finally, expert feedback is synchronized to the business process to update the decision output and is transformed into high-quality labeled data to drive online fine-tuning and offline training of the intelligent agent model. This constructs a closed-loop system of "intelligent agent pre-decision-credibility assessment-tiered triggering intervention-expert feedback correction-model continuous optimization." By achieving deep integration of intelligent agent decision-making and business processes, this solution significantly improves the credibility and reliability of decisions, effectively reduces unnecessary expert intervention costs, and continuously iterates and optimizes intelligent agent capabilities based on expert feedback, ultimately forming a virtuous cycle of synergistic enhancement between intelligent decision-making and expert experience within the business process.

[0013] Preferably, the agent decision-making module in step S1 adopts a three-layer decoupled architecture of context awareness, model reasoning, and basis generation, wherein:

[0014] The context-aware layer is used to obtain business process context information;

[0015] The model inference layer is used to generate candidate decision schemes based on the business process context information;

[0016] The generation layer is used to generate corresponding structured decision-making criteria through abductive reasoning.

[0017] This preferred solution decouples the agent decision-making module into a three-layer independent architecture: context awareness, model reasoning, and basis generation. This enables accurate acquisition of business process context information, reliable generation of candidate decision solutions, and abductive reasoning of structured decision basis. This layered design not only enhances the professional processing capabilities and flexibility of each stage but also ensures the interpretability and transparency of the decision-making process. As a result, it provides a clear and traceable decision-making basis for subsequent credibility assessment and expert intervention, thereby enhancing the overall robustness and verifiability of agent decision-making.

[0018] Preferably, the multi-dimensional decision credibility dynamic evaluation model described in step S2 includes four evaluation dimensions: introspection confidence, historical performance, rule compliance, and peer consensus, wherein:

[0019] Introspection confidence represents the degree of confidence an agent has in the outcome of its own decisions;

[0020] Historical performance represents the decision-making accuracy and business adaptation pass rate of the intelligent agent in similar business scenarios;

[0021] Rule compliance indicates the degree to which a decision-making solution conforms to business norms or security rules;

[0022] Peer consensus represents the proportion of consistency in the decision-making results of various agents in a multi-agent collaborative environment.

[0023] This preferred approach constructs a multi-dimensional dynamic evaluation model for decision credibility, encompassing four dimensions: introspective confidence, historical performance, rule compliance, and peer consensus. This model achieves a comprehensive and multi-dimensional credibility measurement of agent decision-making schemes. It integrates the agent's intrinsic confidence, past performance, adherence to business norms, and degree of consensus in multi-agent collaboration. This allows for accurate identification of decision-making risks and the generation of more valuable comprehensive credibility scores. This provides a comprehensive and objective basis for subsequent tiered expert intervention mechanisms, significantly improving the accuracy and robustness of the evaluation results.

[0024] Preferably, step S2, which involves generating a corresponding risk signal based on the comprehensive credibility score, includes:

[0025] Configure dynamic weights for each evaluation dimension that are associated with the business scenario type, and calculate the comprehensive credibility score based on the analytic hierarchy process.

[0026] Based on the preset credibility threshold and risk level classification criteria, the comprehensive credibility score is mapped to risk signals of different levels.

[0027] In particular, scenarios without historical case references or where business parameters change abruptly are identified as high-novelty scenarios and the highest level of risk signal is generated directly.

[0028] This preferred solution achieves precise quantification of decision credibility by configuring dynamic weights for each evaluation dimension in relation to business scenarios and using the analytic hierarchy process (AHP) to calculate a comprehensive credibility score. Furthermore, based on preset thresholds and risk level standards, the score is mapped to differentiated risk signals, and scenarios without historical references or with sudden parameter changes are directly marked as having the highest risk, thus ensuring the sensitivity and adaptability of risk warnings. This mechanism not only dynamically responds to the evaluation needs of different business scenarios but also effectively identifies unknown or unexpected situations, providing accurate and reliable risk triggering criteria for subsequent tiered expert intervention.

[0029] Preferably, the hierarchical expert intervention triggering mechanism described in step S3 includes:

[0030] When the overall credibility score reaches the preset high credibility threshold, the risk signal is low risk, and the intelligent agent autonomously completes the decision and records the decision log.

[0031] When the overall credibility score is in the preset middle range, the risk signal is medium risk, and the decision-making plan and evaluation results are pushed to experts for asynchronous review.

[0032] When the overall credibility score is lower than the preset threshold or is identified as a high-novelty scenario, the risk signal is high-risk, the business process is suspended and expert intervention is triggered.

[0033] This preferred solution establishes a tiered expert intervention trigger mechanism, classifying decision results into low, medium, and high risk levels based on a comprehensive credibility score, and adopting differentiated processing strategies of autonomous decision-making, asynchronous review, and synchronous intervention for each level. This mechanism achieves a dynamic balance between efficient agent autonomy and precise expert intervention. While ensuring the smooth operation of business processes in normal scenarios, it promptly blocks suspicious or high-risk decisions and introduces experts for in-depth control. Thus, it effectively optimizes the allocation of expert resources while improving the overall decision-making efficiency and reliability, ensuring the security and stability of business processes.

[0034] Preferably, the process context information in step S4 includes business scenario parameters, data source information, data integrity description, and decision link node information; the uncertainty analysis results include fuzzy areas of agent decision-making, data gaps, and potential business impact assessments.

[0035] This preferred solution provides experts with a comprehensive and transparent panoramic view of decision-making by pushing contextual information such as business scenario parameters, data sources, integrity descriptions, and decision-making link node information, along with the agent uncertainty analysis results including ambiguous areas, data gaps, and potential business impacts, to the expert interactive workbench. This multi-dimensional information integration enables experts to quickly and accurately understand the internal logic and potential risks of agent decisions, significantly improving the efficiency and accuracy of review and correction. This lays a solid foundation for high-quality correction of subsequent decision outputs and the reliability of model optimization annotation data.

[0036] Preferably, step S5, which involves online fine-tuning and offline training of the agent model, includes:

[0037] The decision parameters of the intelligent agent are updated in real time using the labeled data through an online fine-tuning channel;

[0038] By using an offline training channel, labeled data is regularly aggregated to perform batch optimization training on the agent's decision-making logic and the weights of the evaluation model.

[0039] This preferred solution constructs a dual-channel optimization mechanism of online fine-tuning and offline training. It uses labeled data generated by expert feedback to update the agent's decision parameters in real time and periodically summarizes data to optimize the decision logic and evaluation model weights in batches. This mechanism realizes the organic combination of the agent's ability to respond instantly and iterate continuously. While quickly adapting to dynamic changes in business, it can systematically learn and improve from accumulated expert experience, thereby driving the synergistic evolution of the agent's decision-making ability and the accuracy of the evaluation model, forming an efficient and robust closed-loop optimization system.

[0040] Secondly, the expert-interventional intelligent agent enhancement and correction system embedded in business processes described in this invention includes:

[0041] The intelligent agent autonomous decision-making module is used to embed into business processes and generate initial decision schemes and complete decision basis; it includes: a context awareness submodule, used to access and understand the current running state of the embedded business process in real time and extract the state features required for decision-making; a decision generation submodule, used to generate candidate decision schemes that meet preset multi-objective constraints based on the state features; and a decision tracing and reasoning submodule, used to automatically generate structured explanations for the generated candidate decision schemes and form complete decision basis;

[0042] The multi-dimensional credibility assessment module is used to quantitatively evaluate the initial decision-making scheme and output a comprehensive credibility score and risk signals. It includes: a dimensional data acquisition submodule, which is used to collect basic data in real time from four dimensions: introspective confidence, historical performance, rule compliance, and peer consensus; a dynamic weighted calculation submodule, which is used to dynamically adjust the weights of each dimension according to the business scenario type and calculate the comprehensive credibility score through the analytic hierarchy process; and a risk signal generation submodule, which is used to generate risk signals of corresponding levels and label the core risk dimensions based on the comprehensive credibility score and scenario novelty identification results.

[0043] The tiered intervention triggering module is used to execute the judgment and triggering of tiered expert intervention based on the comprehensive credibility score and risk signals. It includes: an intervention rule base management submodule, used to define and maintain the rules, thresholds, and logic for triggering expert intervention in different scenarios; a multi-level trigger judgment engine, used to perform real-time matching and logical reasoning on the input comprehensive credibility score and risk signals based on the rule base, determining whether intervention is needed for this decision and at what level; a process control and state management submodule, used to actually control the suspension, suspension, redirection, or resumption of the business process according to the instructions of the judgment engine, and manage the state machine throughout the intervention process; and an intervention audit submodule, used to record the reasons, process, results, and time consumption of triggering intervention throughout the process, providing data support for analyzing the effectiveness of intervention and optimizing triggering rules.

[0044] The expert interactive intervention module is used to present the decision context to experts and receive their feedback after review and correction. It includes: a context display submodule, which is used to display the decision chain, business parameters, data information and uncertainty analysis report in a structured manner; a solution editing submodule, which allows experts to modify, annotate or reject the initial decision solution and enter the basis and reasons for the operation; and a feedback synchronization submodule, which is used to push the expert feedback results to the business process and model iteration optimization module in real time.

[0045] The model iteration and optimization module is used to transform the feedback results into labeled data and drive the agent's strategy iteration through online fine-tuning and offline training. It includes: a labeled data processing submodule, used to perform structured processing, deduplication, and classification on the expert feedback results to generate a standardized labeled dataset; an online fine-tuning submodule, used to adjust the agent's decision parameters and evaluate the model weights in real time based on the standardized labeled dataset; and an offline training submodule, used to periodically train and evaluate the model and agent's decision logic in batches, generate an optimized model version, and replace it online after testing and verification, while retaining historical versions for backtracking.

[0046] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program capable of being loaded by the processor and executing the expert-intervention intelligent agent enhancement and correction method for the embedded business process.

[0047] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned expert-intervention intelligent agent enhancement and correction method for embedded business processes.

[0048] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. This solution, by constructing a dynamic evaluation model for decision credibility and a hierarchical expert intervention mechanism, achieves quantitative identification and hierarchical control of agent decision-making risks in business processes. It also relies on expert feedback to construct an "evaluation-intervention-optimization" closed loop, deeply integrating agent decision-making into business logic. This significantly reduces redundant manual intervention while improving the credibility of decisions at key nodes, driving the continuous and robust evolution of agent capabilities. 2. This solution adopts a three-layer decoupled architecture of context awareness, model reasoning, and abductive reasoning. This enables the agent to not only generate decision-making solutions but also simultaneously output structured decision-making basis, achieving full-link transparency and explainability of the decision-making process. This effectively supports business personnel in understanding the agent. 3. This solution introduces a multi-dimensional credibility assessment system that includes introspection confidence, historical performance, rule compliance, and peer consensus. Combined with dynamic weights and a novelty scenario recognition mechanism, it achieves precise quantification of decision quality and risk. Based on this, it implements a layered strategy of autonomous decision-making, asynchronous review, or synchronous intervention, maximizing the operational efficiency of the agent while ensuring business security. 4. This solution establishes a dual-channel optimization channel consisting of online fine-tuning and offline training. The results of expert review and correction are automatically converted into labeled data and fed back to the model in real time, forming a data-driven continuous learning mechanism. This enables the agent to quickly adapt to business changes and achieve robust iteration and performance improvement of decision logic in long-term operation. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0050] Figure 2 This is a schematic diagram of the multi-dimensional decision credibility dynamic evaluation model structure of the present invention;

[0051] Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0053] This invention provides a method for enhancing and correcting expert-intervention intelligent agents embedded in business processes, applicable to an expert-intervention intelligent agent enhancement and correction system embedded in business processes, such as... Figure 1 As shown, the method includes the following steps:

[0054] S1. Deploy intelligent agent decision-making modules at key decision-making nodes in the business process, obtain business scenario data and process context information of the corresponding nodes, and generate initial decision schemes, uncertainty analysis results and corresponding decision basis based on the business scenario data;

[0055] S2. Construct a multi-dimensional dynamic evaluation model for decision credibility, evaluate the initial decision scheme, calculate the comprehensive credibility score, and generate corresponding risk signals based on the comprehensive credibility score;

[0056] S3. Based on the comprehensive credibility score and risk signal, execute the hierarchical expert intervention triggering mechanism to determine whether to initiate the expert intervention process.

[0057] S4. When the expert intervention process is triggered, the business scenario data, process context information, initial decision plan, uncertainty analysis results and risk signals are pushed to the expert interactive workbench, where experts review, correct or reject them and generate expert feedback results.

[0058] S5. Synchronize the expert feedback results to the business process to update the decision output, and use the expert feedback results as labeled data for online fine-tuning and offline training of the agent model to achieve continuous optimization of the agent's decision-making ability.

[0059] Furthermore, the intelligent agent decision-making module adopts a three-layer decoupled architecture of context awareness, model reasoning, and basis generation, wherein:

[0060] The context-aware layer is used to obtain business process context information;

[0061] The model inference layer is used to generate candidate decision schemes based on the business process context information;

[0062] The generation layer is used to generate corresponding structured decision-making criteria through abductive reasoning.

[0063] Furthermore, the multi-dimensional decision credibility dynamic assessment model includes four assessment dimensions: introspective confidence, historical performance, rule compliance, and peer consensus.

[0064] Introspection confidence represents the degree of confidence an agent has in the outcome of its own decisions;

[0065] Historical performance represents the decision-making accuracy and business adaptation pass rate of the intelligent agent in similar business scenarios;

[0066] Rule compliance indicates the degree to which a decision-making solution conforms to business norms or security rules;

[0067] Peer consensus represents the proportion of consistency in the decision-making results of various agents in a multi-agent collaborative environment.

[0068] Furthermore, the corresponding risk signals generated based on the comprehensive credibility score include:

[0069] Configure dynamic weights for each evaluation dimension that are associated with the business scenario type, and calculate the comprehensive credibility score based on the analytic hierarchy process.

[0070] Based on the preset credibility threshold and risk level classification criteria, the comprehensive credibility score is mapped to risk signals of different levels.

[0071] In particular, scenarios without historical case references or where business parameters change abruptly are identified as high-novelty scenarios and the highest level of risk signal is generated directly.

[0072] Furthermore, the tiered expert intervention trigger mechanism includes:

[0073] Based on comprehensive credibility scores and risk signals, a tiered expert intervention mechanism is implemented to achieve a dynamic balance between intervention accuracy and business efficiency, avoiding efficiency losses caused by excessive intervention while preventing business risks caused by insufficient intervention. For low-risk signals (≥80 points), the agent autonomously outputs decision results, simultaneously recording the decision plan, evaluation report, inference chain, and risk signals in the decision log. The log data retention period is no less than the minimum period required by business compliance, for subsequent audit traceability. For medium-risk signals (60-79 points), the tiered intervention trigger module automatically pushes the decision plan, comprehensive evaluation report, and uncertainty analysis summary to the interactive workbench of the corresponding business domain expert, initiating an asynchronous expert review process. The business process proceeds normally, and experts must complete the review within a preset time limit. If a decision deviation is found, a correction command can be initiated in real time, and the system updates the decision output immediately upon receiving the command. For high-risk signals (<60 points) or high-novelty scenarios, the system automatically suspends downstream business processes to prevent risk spread and immediately triggers synchronous expert intervention. Experts are reminded via workbench pop-ups and SMS messages, and they must immediately access the review process. The business process resumes only after expert feedback.

[0074] Furthermore, the process context information includes business scenario parameters, data source information, data integrity description, and decision link node information; the uncertainty analysis results include fuzzy areas in the agent's decision-making, data gaps, and potential business impact assessments.

[0075] Furthermore, online fine-tuning and offline training of the agent model include:

[0076] The decision parameters of the intelligent agent are updated in real time using the labeled data through an online fine-tuning channel;

[0077] By using an offline training channel, labeled data is regularly aggregated to perform batch optimization training on the agent's decision-making logic and the weights of the evaluation model.

[0078] The following is an illustration using a specific example.

[0079] First, the intelligent agent's autonomous decision-making module is embedded in the power dispatch instruction generation process, enabling data interoperability with the control cloud platform, D5000 system, and OMS system. Based on real-time load data, equipment operating parameters, dispatch rules, and historical cases, the intelligent agent autonomously generates an initial load adjustment plan for a certain area's transmission lines, specifying the adjustment range, execution timing, and safety verification basis, and simultaneously outputs an introspection confidence level of 82%.

[0080] Secondly, the multi-dimensional credibility assessment module initiates the assessment process: the dimensional data acquisition submodule acquires basic data for each dimension. In the historical performance dimension, the accuracy rate of the agent's similar load adjustment decision is 91%. In the rule compliance dimension, the solution complies with the power grid dispatch safety specifications. In the peer consensus dimension, the consistency rate of similar decisions among the three collaborative agents is 85%. The dynamic weighted calculation submodule, based on the characteristics of the power dispatch scenario, configures the weights as follows: rule compliance 35%, historical performance 30%, peer consensus 20%, and introspection confidence 15%. The overall credibility score is calculated to be 84 points, generating a low-risk signal.

[0081] Then, the layered intervention trigger module determines that the risk is low, the intelligent agent autonomously outputs a load adjustment plan, and simultaneously records the decision log, evaluation data and inference link. The business process proceeds normally, and the plan is pushed to the scheduling execution stage.

[0082] In the above case, if the introspection confidence level of the intelligent system is 70%, the historical performance accuracy rate is 78%, the peer consensus consistency rate is 72%, and there are no violations in rule compliance, the comprehensive score is calculated to be 73 points, generating a medium-risk signal. The layered intervention trigger module initiates asynchronous expert review, and the system pushes the decision plan, evaluation report, load data, and uncertainty analysis (annotating the reasons for the low introspection confidence level) to the scheduling expert workbench. Experts view the complete information through the workbench, discover that there is room for optimization in the load adjustment range of a certain period in the plan, modify the adjustment parameters through the plan editing submodule, supplement the correction basis, and the feedback results are synchronized to the business process in real time. The updated plan is pushed to the execution stage, and the feedback data is simultaneously transmitted to the model iteration and optimization module.

[0083] Finally, the model iteration and optimization module processes the expert feedback: the online fine-tuning submodule adjusts the calculation parameters of the agent's load adjustment range in real time to improve the decision-making accuracy in similar scenarios; the offline training submodule uses this feedback as labeled data, summarizes it with other feedback data daily, and performs batch training on the weight configuration of the credibility assessment model and the agent's decision-making logic. After the optimized model passes the test, it replaces the original version and goes online.

[0084] like Figure 2As shown, the multi-dimensional decision-making credibility dynamic evaluation model structure provided in this embodiment of the invention includes three layers: the target layer, the dimension layer, and the indicator layer. The target layer is a comprehensive credibility score; the dimension layer includes introspection confidence, historical performance, rule compliance, and peer consensus; the introspection confidence indicators include inference entropy value and parameter determinism; the historical performance indicators include accuracy, adaptation pass rate, and fault feedback rate; the rule compliance indicators include the proportion of compliant items and the impact level of violations; and the peer consensus indicators include multi-agent consistency rate and baseline model fit.

[0085] like Figure 3 As shown, this embodiment of the invention provides an expert-interventional intelligent agent enhancement and correction system embedded in business processes, which consists of an intelligent agent autonomous decision-making module, a multi-dimensional credibility assessment module, a hierarchical intervention triggering module, an expert interactive intervention module, and a model iterative optimization module, wherein:

[0086] The intelligent agent autonomous decision-making module is used to embed into business processes and generate initial decision schemes and complete decision basis; it includes: a context awareness submodule, used to access and understand the current running state of the embedded business process in real time and extract the state features required for decision-making; a decision generation submodule, used to generate candidate decision schemes that meet preset multi-objective constraints based on the state features; and a decision tracing and reasoning submodule, used to automatically generate structured explanations for the generated candidate decision schemes and form complete decision basis;

[0087] The multi-dimensional credibility assessment module is used to quantitatively evaluate the initial decision-making scheme and output a comprehensive credibility score and risk signals. It includes: a dimensional data acquisition submodule, which is used to collect basic data in real time from four dimensions: introspective confidence, historical performance, rule compliance, and peer consensus; a dynamic weighted calculation submodule, which is used to dynamically adjust the weights of each dimension according to the business scenario type and calculate the comprehensive credibility score through the analytic hierarchy process; and a risk signal generation submodule, which is used to generate risk signals of corresponding levels and label the core risk dimensions based on the comprehensive credibility score and scenario novelty identification results.

[0088] The tiered intervention triggering module is used to execute the judgment and triggering of tiered expert intervention based on the comprehensive credibility score and risk signals. It includes: an intervention rule base management submodule, used to define and maintain the rules, thresholds, and logic for triggering expert intervention in different scenarios; a multi-level trigger judgment engine, used to perform real-time matching and logical reasoning on the input comprehensive credibility score and risk signals based on the rule base, determining whether intervention is needed for this decision and at what level; a process control and state management submodule, used to actually control the suspension, suspension, redirection, or resumption of the business process according to the instructions of the judgment engine, and manage the state machine throughout the intervention process; and an intervention audit submodule, used to record the reasons, process, results, and time consumption of triggering intervention throughout the process, providing data support for analyzing the effectiveness of intervention and optimizing triggering rules.

[0089] The expert interactive intervention module is used to present the decision context to experts and receive their feedback after review and correction. It includes: a context display submodule, which is used to display the decision chain, business parameters, data information and uncertainty analysis report in a structured manner; a solution editing submodule, which allows experts to modify, annotate or reject the initial decision solution and enter the basis and reasons for the operation; and a feedback synchronization submodule, which is used to push the expert feedback results to the business process and model iteration optimization module in real time.

[0090] The model iteration and optimization module is used to transform the feedback results into labeled data and drive the agent's strategy iteration through online fine-tuning and offline training. It includes: a labeled data processing submodule, used to perform structured processing, deduplication, and classification on the expert feedback results to generate a standardized labeled dataset; an online fine-tuning submodule, used to adjust the agent's decision parameters and evaluate the model weights in real time based on the standardized labeled dataset; and an offline training submodule, used to periodically train and evaluate the model and agent's decision logic in batches, generate an optimized model version, and replace it online after testing and verification, while retaining historical versions for backtracking.

[0091] The present invention also discloses an electronic device.

[0092] Specifically, the electronic device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0093] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor executes various functional applications and data processing by running non-transitory software programs, instructions, and modules stored in memory. Memory may include a program storage area and a data storage area. The program storage area may store the control unit and the application program required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, memory may include high-speed random access memory and non-transitory memory. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0094] The present invention also discloses a computer-readable storage medium.

[0095] Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above method implementation.

[0096] Those skilled in the art will understand that all or part of the processes in the methods described above can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

Claims

1. A method for enhancing and correcting expert-intervened intelligent agents embedded in business processes, characterized in that, Includes the following steps: S1. Deploy intelligent agent decision-making modules at key decision-making nodes in the business process, obtain business scenario data and process context information of the corresponding nodes, and generate initial decision schemes, uncertainty analysis results and corresponding decision basis based on the business scenario data; S2. Construct a multi-dimensional dynamic evaluation model for decision credibility, evaluate the initial decision scheme, calculate the comprehensive credibility score, and generate corresponding risk signals based on the comprehensive credibility score; S3. Based on the comprehensive credibility score and risk signal, execute the hierarchical expert intervention triggering mechanism to determine whether to initiate the expert intervention process. S4. When the expert intervention process is triggered, the business scenario data, process context information, initial decision plan, uncertainty analysis results and risk signals are pushed to the expert interactive workbench, where experts review, correct or reject them and generate expert feedback results. S5. Synchronize the expert feedback results to the business process to update the decision output, and use the expert feedback results as labeled data for online fine-tuning and offline training of the agent model to achieve continuous optimization of the agent's decision-making ability.

2. The method according to claim 1, characterized in that, The agent decision-making module described in step S1 adopts a three-layer decoupled architecture of context awareness, model reasoning, and basis generation, wherein: The context-aware layer is used to obtain business process context information; The model inference layer is used to generate candidate decision schemes and uncertainty analysis results based on the business process context information; The generation layer is used to generate corresponding structured decision-making criteria through abductive reasoning.

3. The method according to claim 1, characterized in that, The multi-dimensional decision credibility dynamic evaluation model described in step S2 includes four evaluation dimensions: introspection confidence, historical performance, rule compliance, and peer consensus. Introspection confidence represents the degree of confidence an agent has in the outcome of its own decisions; Historical performance represents the decision-making accuracy and business adaptation pass rate of the intelligent agent in similar business scenarios; Rule compliance indicates the degree to which a decision-making solution conforms to business norms or security rules; Peer consensus represents the consistency rate of decision-making outcomes among agents in a multi-agent collaborative scenario.

4. The method according to claim 1, characterized in that, Step S2, which involves generating a corresponding risk signal based on the comprehensive credibility score, includes: Configure dynamic weights for each evaluation dimension that are associated with the business scenario type, and calculate the comprehensive credibility score based on the analytic hierarchy process. Based on the preset credibility threshold and risk level classification criteria, the comprehensive credibility score is mapped to risk signals of different levels. In particular, scenarios without historical case references or where business parameters change abruptly are identified as high-novelty scenarios and the highest level of risk signal is generated directly.

5. The method according to claim 1, characterized in that, The hierarchical expert intervention triggering mechanism described in step S3 includes: When the overall credibility score reaches the preset high credibility threshold, the risk signal is low risk, and the intelligent agent autonomously completes the decision and records the decision log. When the overall credibility score is in the preset middle range, the risk signal is medium risk, and the decision-making plan and evaluation results are pushed to experts for asynchronous review. When the overall credibility score is lower than the preset threshold or is identified as a high-novelty scenario, the risk signal is high-risk, the business process is suspended and expert intervention is triggered.

6. The method according to claim 1, characterized in that, The process context information mentioned in step S4 includes business scenario parameters, data source information, data integrity description, and decision link node information; the uncertainty analysis results include fuzzy areas of agent decision-making, data gaps, and potential business impact assessments.

7. The method according to claim 1, characterized in that, Step S5, which involves online fine-tuning and offline training of the agent model, includes: The decision parameters of the intelligent agent are updated in real time using the labeled data through an online fine-tuning channel; By using an offline training channel, labeled data is regularly aggregated to perform batch optimization training on the agent's decision-making logic and the weights of the evaluation model.

8. A method for enhancing and correcting expert-intervened intelligent agents embedded in business processes, characterized in that, include: The intelligent agent autonomous decision-making module is used to embed into business processes and generate initial decision schemes and complete decision basis; it includes: a context awareness submodule, used to access and understand the current running state of the embedded business process in real time and extract the state features required for decision-making; a decision generation submodule, used to generate candidate decision schemes that meet preset multi-objective constraints based on the state features; and a decision tracing and reasoning submodule, used to automatically generate structured explanations for the generated candidate decision schemes and form complete decision basis; The multi-dimensional credibility assessment module is used to quantitatively evaluate the initial decision-making scheme and output a comprehensive credibility score and risk signals. It includes: a dimensional data acquisition submodule, which is used to collect basic data in real time from four dimensions: introspective confidence, historical performance, rule compliance, and peer consensus; a dynamic weighted calculation submodule, which is used to dynamically adjust the weights of each dimension according to the business scenario type and calculate the comprehensive credibility score through the analytic hierarchy process; and a risk signal generation submodule, which is used to generate risk signals of corresponding levels and label the core risk dimensions based on the comprehensive credibility score and scenario novelty identification results. The tiered intervention triggering module is used to execute the judgment and triggering of tiered expert intervention based on the comprehensive credibility score and risk signals. It includes: an intervention rule base management submodule, used to define and maintain the rules, thresholds, and logic for triggering expert intervention in different scenarios; a multi-level trigger judgment engine, used to perform real-time matching and logical reasoning on the input comprehensive credibility score and risk signals based on the rule base, determining whether intervention is needed for this decision and at what level; a process control and state management submodule, used to actually control the suspension, suspension, redirection, or resumption of the business process according to the instructions of the judgment engine, and manage the state machine throughout the intervention process; and an intervention audit submodule, used to record the reasons, process, results, and time consumption of triggering intervention throughout the process, providing data support for analyzing the effectiveness of intervention and optimizing triggering rules. The expert interactive intervention module is used to present the decision context to experts and receive their feedback after review and correction. It includes: a context display submodule, which is used to display the decision chain, business parameters, data information and uncertainty analysis report in a structured manner; a solution editing submodule, which allows experts to modify, annotate or reject the initial decision solution and enter the basis and reasons for the operation; and a feedback synchronization submodule, which is used to push the expert feedback results to the business process and model iteration optimization module in real time. The model iteration and optimization module is used to transform the feedback results into labeled data and drive the agent's strategy iteration through online fine-tuning and offline training. It includes: a labeled data processing submodule, used to perform structured processing, deduplication, and classification on the expert feedback results to generate a standardized labeled dataset; an online fine-tuning submodule, used to adjust the agent's decision parameters and evaluate the model weights in real time based on the standardized labeled dataset; and an offline training submodule, used to periodically train and evaluate the model and agent's decision logic in batches, generate an optimized model version, and replace it online after testing and verification, while retaining historical versions for backtracking.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the expert-intervention agent enhancement and correction method for embedded business processes as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the expert-interventional agent enhancement and correction method for embedded business processes according to any one of claims 1 to 7.

Images