An intelligent situation analysis orchestration method and system

Abstract

The application provides a kind of intelligent situation analysis arrangement method and system, the method includes: obtaining user business analysis demand and extracting core elements, constructs four-layer neural network cooperation framework containing data layer, processing layer, analysis layer and display layer agent;Forward decomposition business target generates initial task execution network, and forms optimized business process by adjusting through reverse cooperation optimization algorithm;Through the double-parameter risk model evaluation process containing model and data parameter vector, identify risk interest points and generate risk situation data;Based on the data, generate editing operation sequence, use continuous time Markov chain model for non-autoregressive adjustment, get the final scheme.This application realizes BI business process intelligent arrangement, improves process quality, efficiency and analysis result reliability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent situational analysis and orchestration method and system, which is applied to data analysis and business intelligence systems. Background Technology

[0002] Business intelligence (BI) systems are key tools for enterprise data analysis and decision support, with core functions including data acquisition, processing, analysis, and visualization. With the surge in enterprise data volume and the diversification of analytical needs, traditional BI systems face significant challenges in orchestrating business processes.

[0003] Currently, common BI orchestration technologies in the industry mainly fall into two categories: rule-based workflow engines and parameterized templates. The former automates data processing through predefined workflow rules, while the latter provides a series of standardized analysis templates, allowing users to configure parameters to meet specific analytical needs. These technologies can provide a certain degree of automation in specific scenarios.

[0004] Advanced BI orchestration technology employs a visual workflow design interface combined with a modular component library, allowing business users to build data analysis workflows through drag-and-drop. This technology encapsulates complex data processing and analysis algorithms into reusable components and provides a data flow mechanism between components, achieving semi-automated construction of analysis workflows. The system also supports simple parameter optimization based on historical execution data to improve the accuracy of the analysis model.

[0005] However, existing technologies still have significant shortcomings: First, they lack intelligent business logic reasoning capabilities, requiring users to manually design analysis processes based on professional knowledge; second, the collaboration mechanisms between system components are rigid, making it difficult to adapt to complex and ever-changing business scenarios; third, they lack the ability to dynamically assess and visualize potential risks during the analysis process, affecting the reliability of decision-making; and finally, the analysis processes generated by the system often have a fixed structure, making it difficult to flexibly adjust them according to data characteristics and business needs. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent situational analysis orchestration method and system, which aims to solve the technical problems in the existing BI business process orchestration, such as lack of intelligent reasoning ability, rigid component collaboration mechanism, insufficient risk assessment, and low flexibility of process adjustment.

[0007] To achieve the above objectives, the present invention provides an intelligent situational analysis and orchestration method, comprising the following steps:

[0008] Obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range, and constraints.

[0009] Based on the core elements of the aforementioned business requirements, a four-layer agent neural network collaboration framework is constructed, comprising data layer agent, processing layer agent, analysis layer agent, and presentation layer agent, and corresponding core capability modules are encapsulated for each layer agent.

[0010] The business objectives are decomposed into forward tasks, and complex business requirements are broken down into sub-tasks layer by layer and assigned to corresponding agents to generate an initial task execution network.

[0011] Based on the execution results of the initial task execution network, a reverse collaboration optimization algorithm is used to adjust the agent collaboration method to form an optimized business process;

[0012] A two-parameter risk model containing model parameter vectors and data parameter vectors is constructed to conduct risk assessment on the optimized business process, identify risk interest points, and generate risk situation data.

[0013] Based on the risk situation data, an editing operation sequence is generated. The optimized business process is then adjusted non-autoregressively using a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme.

[0014] Preferably, the construction of a four-layer agent neural network collaboration framework comprising a data layer agent, a processing layer agent, an analysis layer agent, and a presentation layer agent includes:

[0015] Based on the core elements of the aforementioned business requirements, a hierarchical structure of data layer proxy, processing layer proxy, analysis layer proxy, and presentation layer proxy is defined to form a complete business process collaboration network.

[0016] The data layer proxy encapsulates data connectivity, data extraction, and data caching capabilities; the processing layer proxy encapsulates data cleaning, transformation, and quality control capabilities; the analysis layer proxy encapsulates model training, reasoning, and result interpretation capabilities; and the presentation layer proxy encapsulates data visualization, interaction design, and narrative expression capabilities.

[0017] Based on the core capability modules of the data layer agent, the processing layer agent, the analysis layer agent, and the presentation layer agent, an inter-agent communication protocol is constructed to support horizontal collaboration between agents in the same layer and vertical collaboration between agents across layers, thus obtaining the four-layer agent neural network collaboration framework.

[0018] Preferably, the construction of the inter-agent communication protocol includes:

[0019] Based on the core capability modules of the data layer proxy, the processing layer proxy, the analysis layer proxy, and the presentation layer proxy, a unified message structure is designed that includes sending proxy identifier, receiving proxy identifier, message type, priority, and timestamp.

[0020] The unified message structure can implement multiple communication modes, including request-response mode, publish-subscribe mode, broadcast mode, and point-to-point direct communication mode.

[0021] Based on the aforementioned multiple communication methods, a standard protocol for proxy collaboration is defined to standardize the processing flow of collaboration request initiation, collaboration invitation response, task division negotiation, and execution status synchronization.

[0022] By establishing data flow processing between agents through the aforementioned standard protocol, it supports the block transmission, incremental synchronization and streaming processing of large-scale data, resulting in a communication protocol that supports efficient collaboration between agents.

[0023] Preferably, the step of performing forward task decomposition on the business objective includes:

[0024] The business objectives are semantically parsed to extract core elements including the analysis object, time range, analysis type, and business constraints, and then converted into a structured target representation.

[0025] Based on the structured target representation, a top-down recursive decomposition algorithm is used to gradually decompose the high-level target into multiple interrelated sub-targets.

[0026] The multiple interrelated sub-goals are further decomposed until they reach a granularity that can be directly assigned to a specific agent for execution, thus forming a subtask specification;

[0027] Based on the logical relationships and data flow between the sub-task specifications, a task dependency graph is constructed.

[0028] Based on the task dependency graph and the available proxy capabilities in the system, the most suitable proxy or proxy team is selected for each subtask to generate the initial task execution network.

[0029] Preferably, the process of adjusting the agent collaboration method based on the execution result of the initial task execution network using a reverse collaborative optimization algorithm includes:

[0030] The execution results of the initial task execution network are evaluated from multiple dimensions, including result quality, execution efficiency, collaborative performance, and user satisfaction.

[0031] Based on the multi-dimensional task evaluation results, the reasons for poor task performance are analyzed, and key improvement points are identified, including data quality issues, inappropriate algorithm selection, suboptimal parameter configuration, or poor collaboration.

[0032] For the aforementioned key improvement points, generate specific optimization signals that include problem descriptions, severity, priority, and recommended measures;

[0033] The specific optimization signal is transmitted from the downstream task to the upstream task through the reverse propagation algorithm, forming a complete optimization chain;

[0034] Based on the specific optimization signals, each agent is guided to adjust its behavior and strategy, while the global collaboration mode between agents is optimized to obtain the optimized business process.

[0035] Preferably, the construction of a two-parameter risk model comprising a model parameter vector and a data parameter vector includes:

[0036] Based on the optimized business process, define model parameter vectors that describe the analysis method, algorithm selection, and parameter configuration, as well as data parameter vectors that describe data quality, data distribution, and sample size.

[0037] Construct a risk assessment function, map the model parameter vector and data parameter vector to a multidimensional risk space, calculate basic risk indicators and aggregate them to form a comprehensive risk score;

[0038] Based on the comprehensive risk score, a risk situation calculation engine is developed to quantify the risk of each link in the optimized business process and obtain risk quantification data.

[0039] The risk interest point identification algorithm automatically discovers high-risk points, inflection points, sensitive points and anomalies in the risk quantification data to obtain risk interest points;

[0040] Based on the risk points of interest and the comprehensive risk score, risk situation data containing risk heatmap data and risk level information is generated.

[0041] Preferably, the step of automatically identifying high-risk points, inflection points, sensitive points, and anomalies in the risk quantification data through a risk interest point identification algorithm to obtain risk interest points includes:

[0042] Threshold detection and regional connectivity analysis are used to identify high-risk points whose risk scores consistently exceed the threshold.

[0043] Gradient analysis and curvature calculation are used to identify inflection points where risk surfaces change drastically;

[0044] Local sensitivity analysis is used to calculate the degree of risk response to parameter changes and to identify sensitive points;

[0045] Local anomaly detection and spatial statistical analysis were used to identify anomalous points that were significantly different from the surrounding risk patterns.

[0046] Feature extraction and attribute analysis are performed on the high-risk points, inflection points, sensitive points and abnormal points to obtain feature analysis results including location features, intensity features, morphological features and dynamic features;

[0047] Based on the feature analysis results, the risk impact and business importance are calculated, the risk interest points are prioritized, and risk situation display data containing key risk area information is generated.

[0048] Preferably, the step of generating an editing operation sequence based on the risk situation data includes:

[0049] Based on the aforementioned risk situation data, a set of editing operations applicable to business processes is defined, including basic operations such as component insertion, deletion, replacement, and parameter adjustment.

[0050] Based on the set of editing operations, a continuous-time Markov chain model describing the evolution of the business process is established, where the state represents the process configuration and the transition rate represents the tendency to execute editing operations per unit time.

[0051] By analyzing historical process versions and edit records, the transition rate matrix parameters of the continuous-time Markov chain model are learned using the maximum likelihood estimation method.

[0052] Based on the risk interest points in the risk situation data, a risk-oriented objective function is constructed, with risk scoring as the main optimization objective.

[0053] By employing a method combining Monte Carlo tree search and Markov chain model, an optimized sequence of editing operations for the business process is generated based on the objective function, resulting in the final intelligent business orchestration scheme.

[0054] Preferably, the step of learning the transition rate matrix parameters of the continuous-time Markov chain model using the maximum likelihood estimation method includes:

[0055] The design includes a set of auxiliary variables, including historical state variables, performance indicator variables, risk assessment variables, and business context variables, to expand the business process state space.

[0056] The set of auxiliary variables is integrated into the state representation through a specific encoding method, and together with the state information of the optimized business process, it participates in the training of the continuous-time Markov chain model.

[0057] Collect successful and unsuccessful edit flow case pairs to implement a training framework based on contrastive learning, which learns to distinguish features by comparing edit sequences with effective and ineffective optimizations;

[0058] The design incorporates a comprehensive reward function that includes dimensions of risk reduction, performance improvement, resource efficiency, and business value, and integrates a reinforcement learning framework to enable the continuous-time Markov chain model to improve editing strategies through interaction with the environment.

[0059] The experience replay buffer stores the editing operations and their effects, and an incremental learning process is established to enable the continuous-time Markov chain model to continuously acquire new knowledge from the operation, thus obtaining a trained continuous-time Markov chain model.

[0060] The present invention also provides an intelligent situational analysis and orchestration system, comprising:

[0061] The requirement parsing module is used to obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range, and constraints.

[0062] The agent collaboration module is used to construct a four-layer agent neural network collaboration framework, including data layer agent, processing layer agent, analysis layer agent and presentation layer agent, based on the core elements of the business requirements, and to encapsulate the corresponding core capability modules for each layer agent.

[0063] The task decomposition module is used to perform forward task decomposition on the business objectives, decompose complex business requirements into sub-tasks layer by layer and assign them to corresponding agents, and generate an initial task execution network.

[0064] The process optimization module is used to adjust the agent collaboration method based on the execution result of the initial task execution network using a reverse collaboration optimization algorithm to form an optimized business process;

[0065] The risk assessment module is used to construct a two-parameter risk model containing model parameter vectors and data parameter vectors, to assess the risks of the optimized business process, identify risk interest points, and generate risk situation data.

[0066] The orchestration generation module is used to generate an editing operation sequence based on the risk situation data, and to perform non-autoregressive adjustments on the optimized business process through a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme.

[0067] The beneficial effects of this invention are:

[0068] 1. By constructing a hierarchical collaboration framework for proxy neural networks, intelligent division of labor and collaboration among various stages of the BI business process were achieved, significantly improving the automation level and adaptability of the business process;

[0069] 2. By adopting a two-stage business process optimization mechanism and drawing on the forward and backward propagation principles of neural networks, the automatic generation and iterative optimization of BI business processes were realized, improving process quality and efficiency.

[0070] 3. A risk visualization system was designed to decouple model parameters from data parameters, intuitively displaying the potential risks in each stage of the BI business process, providing a reliable basis for decision-making, and enhancing the credibility of the analysis results;

[0071] 4. A non-autoregressive business process generation mechanism based on edit flow was developed, which realizes flexible adjustment of the process through a continuous-time Markov chain model, breaking through the limitations of traditional process generation;

[0072] 5. By integrating the above innovations, a fully intelligent BI business orchestration system was built, realizing intelligent orchestration of the entire process from data collection to visualization, significantly reducing the threshold for using BI systems and improving enterprises' data analysis capabilities. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 A flowchart of the intelligent situational analysis and orchestration method provided in this embodiment of the invention;

[0075] Figure 2 This is a structural diagram of the hierarchical collaborative framework for proxy neural networks provided in an embodiment of the present invention;

[0076] Figure 3 A flowchart illustrating the two-stage business process optimization mechanism provided in this embodiment of the invention;

[0077] Figure 4 This is a diagram of the decoupled risk visualization system architecture provided in an embodiment of the present invention;

[0078] Figure 5 This is a structural diagram of the intelligent situation analysis and orchestration system provided in an embodiment of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0080] Please refer to Figure 1 , Figure 1 A flowchart illustrating an intelligent situational analysis and orchestration method provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0081] Step S101: Obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range, and constraints.

[0082] In this embodiment, the system first receives a description of the business objective input by the user. This description may be in natural language (e.g., "Analyze the reasons for the sales decline in the past three months and predict the trend for the next quarter") or a structured requirement (e.g., the analysis type and parameters selected through the interface). Using natural language processing techniques, including semantic analysis, intent recognition, and entity extraction, the core elements of the business objective are parsed out, such as the analysis object (sales data), time frame (the past three months), and analysis type (cause diagnosis and predictive analysis). This complex business description is then converted into a structured target representation, serving as the starting point for task decomposition.

[0083] Step S102: Based on the core elements of the business requirements, construct a four-layer agent neural network collaboration framework that includes data layer agent, processing layer agent, analysis layer agent and presentation layer agent, and encapsulate corresponding core capability modules for each layer agent.

[0084] In this embodiment, based on the characteristics of the BI business process, a four-layer agent structure is designed: data layer agent, processing layer agent, analysis layer agent, and presentation layer agent. Each layer of agents is responsible for specific BI sub-tasks, forming a complete BI business process collaboration network. For agents at different levels, corresponding core capability modules are encapsulated, including data connectivity, data transformation, model training, and data visualization capabilities, making each agent an intelligent node with specific functions. A message-passing-based inter-agent communication protocol is designed to support horizontal collaboration between agents at the same layer and vertical collaboration between agents across layers, realizing information flow and task coordination within the agent network. Through a combination of supervised learning and reinforcement learning, an adaptive learning mechanism for agents is implemented, enabling agents to continuously optimize their capabilities and collaboration strategies based on historical task execution performance.

[0085] Step S103: Perform forward task decomposition on the business objectives, decompose complex business requirements into sub-tasks layer by layer and assign them to the corresponding agents to generate an initial task execution network.

[0086] In this embodiment, based on the parsed business objectives, a top-down recursive decomposition algorithm is used to progressively decompose the high-level objectives into multiple interrelated sub-objectives. The decomposition process follows the general workflow of business logic and data analysis; for example, "sales decline reason analysis" is decomposed into sub-objectives such as "sales data trend analysis," "correlation analysis of influencing factors," and "critical node anomaly detection." Each sub-objective can be further decomposed until it reaches a granularity that can be directly assigned to a specific agent for execution. Clear task specifications are defined for each decomposed subtask, including input data requirements, processing logic, output format, quality standards, and time constraints. Based on the logical relationships and data flow between subtasks, a complete task dependency graph (DAG, Directed Acyclic Graph) is constructed. Based on the task specifications and the available agent capabilities in the system, the most suitable agent or agent team is selected for each subtask. An initial task execution plan is generated, including the execution agent for each subtask, the estimated start time, the estimated completion time, and resource allocation.

[0087] Step S104: Based on the execution results of the initial task execution network, the reverse collaboration optimization algorithm is used to adjust the agent collaboration method to form an optimized business process.

[0088] In this embodiment, a comprehensive task evaluation system is first established to assess the execution effectiveness of each sub-task from multiple dimensions. Evaluation dimensions include result quality (accuracy, completeness, consistency, etc.), execution efficiency (time consumption, resource utilization, etc.), collaboration performance (information transmission efficiency, team coordination, etc.), and user satisfaction (result usability, understandability, etc.). Based on the task evaluation results, specific optimization signals are generated to guide business process improvements. The reasons for poor task performance are analyzed to identify key improvement points, such as data quality issues, inappropriate algorithm selection, suboptimal parameter configuration, or poor collaboration. A backpropagation algorithm is designed to transmit optimization signals from downstream tasks to upstream tasks, forming a complete optimization chain. Based on the received optimization signals, each agent is guided to adjust its behavior and strategies. Based on feedback from multiple task executions, the global collaboration mode between agents is continuously optimized.

[0089] Step S105: Construct a two-parameter risk model containing model parameter vectors and data parameter vectors, conduct a risk assessment on the optimized business process, identify risk interest points, and generate risk situation data.

[0090] In this embodiment, a theoretical framework for risk decoupling is established, decomposing potential risks in the BI business process into two orthogonal dimensions: model risk and data risk. The model parameter vector describes factors related to the processing logic, such as analysis methods, algorithm selection, and parameter configuration; the data parameter vector describes factors related to data characteristics, such as data quality, data distribution, and sample size. Based on the model parameter vector and data parameter vector, a risk assessment function is constructed, mapping the parameter vectors to a multi-dimensional risk space. The risk function adopts a hierarchical design, first calculating basic risk indicators (such as prediction bias risk, overfitting risk, and data leakage risk), and then aggregating them to form a comprehensive risk score. A series of dedicated algorithms are designed and implemented to automatically detect various risk interest points from the risk situation data. For high-risk points, threshold detection and regional connectivity analysis are used to identify continuous regions where the risk score consistently exceeds the threshold; for inflection points, gradient analysis and curvature calculation are used to identify locations where the risk surface changes drastically; for sensitive points, local sensitivity analysis is used to calculate the degree of risk response to parameter changes; and for anomaly points, local anomaly detection and spatial statistical analysis are used to identify points significantly different from surrounding risk patterns.

[0091] Step S106: Generate an editing operation sequence based on the risk situation data, and perform non-autoregressive adjustments on the optimized business process using a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme.

[0092] In this embodiment, the design covers basic editing operation types that encompass all possible changes to the BI business process. Operation types include three core categories: insertion operations (adding new components or steps to the process), deletion operations (removing existing components or steps from the process), and replacement operations (replacing existing components with new ones while maintaining connections). A state space describing the BI business process is established. The state representation uses a graph structure, where nodes represent process components and edges represent data flow relationships between components. Based on the set of editing operations and the process state representation, a transition rate matrix describing the dynamics of state transitions is constructed. Based on risk assessment, an objective function, and a continuous-time Markov chain model, an optimized sequence of editing operations for the BI process is generated. The generation process employs a combination of Monte Carlo Tree Search (MCTS) and Markov chain models, maintaining random exploration capabilities while selectively searching in the direction of risk reduction.

[0093] Please refer to Figure 2 , Figure 2 This is a structural diagram of the hierarchical collaborative framework for proxy neural networks provided in an embodiment of the present invention. Figure 2 As shown, the framework consists of four layers: data layer proxy, processing layer proxy, analysis layer proxy, and presentation layer proxy. Each layer of proxy achieves horizontal and vertical collaboration through communication protocols.

[0094] In a preferred embodiment, the construction of a four-layer agent neural network collaborative framework comprising a data layer agent, a processing layer agent, an analysis layer agent, and a presentation layer agent includes:

[0095] Based on the core elements of the aforementioned business requirements, a hierarchical structure of data layer agents, processing layer agents, analysis layer agents, and presentation layer agents is defined to form a complete business process collaboration network. Specifically, the core elements of the business requirements are first analyzed, including the analysis object, time range, analysis type, and business constraints, to determine the various agents that need to be invoked. Then, a hierarchical agent structure is established, with the data layer agent at the bottom responsible for data acquisition, the processing layer agent at the second layer responsible for data transformation, the analysis layer agent at the third layer performing core analysis tasks, and the presentation layer agent at the top layer responsible for result presentation. This hierarchical structure design draws on the hierarchical characteristics of neural networks, enabling each layer of agents to independently perform its own functions while forming a collaborative network to complete complex BI tasks. Furthermore, the number and type of agents at each layer are dynamically adjusted according to the complexity and characteristics of the business requirements to ensure reasonable resource allocation and efficient task execution.

[0096] The data layer proxy encapsulates data connectivity, data extraction, and data caching capabilities; the processing layer proxy encapsulates data cleaning, transformation, and quality control capabilities; the analysis layer proxy encapsulates model training, inference, and result interpretation capabilities; and the presentation layer proxy encapsulates data visualization, interaction design, and narrative expression capabilities. During capability encapsulation, a modular design principle is adopted, encapsulating various functions into independent capability modules for easy reuse and upgrades. For example, the data layer proxy's connectivity module supports multiple data source protocols, including JDBC, ODBC, and REST, and has built-in authentication and permission management mechanisms; the data extraction module implements multiple modes such as incremental extraction, full extraction, and real-time stream processing, and can automatically select the optimal extraction strategy based on data volume and timeliness requirements. The processing layer proxy's data cleaning module integrates anomaly detection algorithms, missing value handling strategies, and deduplication mechanisms, enabling automatic identification and handling of data quality issues; the transformation module supports complex data structure transformations and format standardization, preparing for subsequent analysis. The analysis layer agent's model training module encapsulates various algorithms, from traditional statistical models to deep learning, supporting automatic model selection and parameter optimization; its results interpretation module generates model interpretation reports, improving the interpretability of analysis results. The presentation layer agent's visualization module automatically recommends the most suitable chart type based on data characteristics; its narrative expression module transforms data insights into natural language descriptions, making them easier for non-technical users to understand.

[0097] Based on the core capability modules of the data layer agent, processing layer agent, analysis layer agent, and presentation layer agent, an inter-agent communication protocol is constructed to support horizontal collaboration between agents within the same layer and vertical collaboration between agents across layers, resulting in the four-layer agent neural network collaboration framework. The communication protocol is a crucial component of the entire collaboration framework, featuring a unified message structure and diverse communication modes. The unified message structure includes a message header (sender ID, receiver ID, message type, priority, and other metadata) and a message body (specific business data or instructions), using JSON format for cross-platform compatibility. Regarding communication modes, it supports request-response mode (suitable for task allocation requiring explicit feedback), publish-subscribe mode (suitable for one-to-many information distribution), broadcast mode (suitable for network-wide notifications), and point-to-point direct communication mode (suitable for high-frequency, low-latency data exchange). In terms of horizontal collaboration, agents within the same layer can share resources, divide tasks, and exchange information. For example, multiple data processing agents can collaboratively process large-scale datasets, automatically performing task sharding and result merging. In terms of vertical collaboration, agents at different levels form a data processing pipeline, establishing clear data transfer and control flow relationships between upper and lower layer agents. For example, the analysis layer agent can provide feedback on data quality requirements to the processing layer agent, guiding the optimization direction of data preprocessing. A dynamic routing mechanism is also implemented, which can automatically adjust message transmission paths based on load and execution efficiency, avoiding communication bottlenecks and single points of failure. Through this multi-layered, multi-mode communication protocol, the four-layer agent network forms a highly collaborative intelligence, capable of flexibly responding to various complex BI business scenarios.

[0098] In intelligent BI business orchestration, the first step is to establish a hierarchical agent structure that covers the entire BI process. This structure divides the entire BI business process into four key layers, each with a dedicated intelligent agent responsible for a specific function:

[0099] Data Layer Proxy: As the foundational layer, it is responsible for data source connection, data extraction, and raw data caching. These proxies possess multi-source heterogeneous data processing capabilities, enabling them to connect to various data sources such as relational databases, NoSQL databases, data lakes, and API interfaces. Each data layer proxy focuses on a specific type of data source or data domain, such as transaction data proxy, user behavior data proxy, or external market data proxy.

[0100] Processing layer agents: Responsible for data cleaning, transformation, integration, and preprocessing. This layer includes data quality inspection agents, data transformation agents, data integration agents, and data standardization agents. They receive raw data from the data layer and, through a series of processing operations, transform it into a structured dataset suitable for analysis. Processing layer agents need to have capabilities such as identifying outliers, handling missing data, and performing feature engineering.

[0101] Analysis Layer Agents: As the core layer, this layer is responsible for data modeling, execution of analytical algorithms, and interpretation of results. This layer includes descriptive analysis agents, predictive analysis agents, diagnostic analysis agents, and decision-making recommendation agents. Each type of agent encapsulates a specific category of analytical algorithms, enabling it to select appropriate analytical methods based on business needs and provide preliminary interpretation of the analysis results.

[0102] Presentation Layer Proxy: Responsible for the visualization and interactive exploration of analysis results. This layer includes proxies for chart generation, dashboard construction, report generation, and interaction design. They receive result data from the analysis layer, automatically select the most suitable visualization format based on data characteristics and business context, and support interactive data exploration for users.

[0103] After defining the hierarchical agent structure, it is necessary to encapsulate specific core capability modules for each type of agent, making it an intelligent node with specialized functions. Agent capability encapsulation is a key step in building the agent neural network, as it determines the specific role and performance of each agent in the BI process.

[0104] Data layer proxy capabilities are encapsulated: This includes data connectivity, data extraction, and data caching capabilities. Data connectivity capabilities include connectors supporting multiple protocols (JDBC, ODBC, REST, etc.) and authentication mechanisms; data extraction capabilities include various modes such as incremental extraction, full extraction, and real-time stream processing; and data caching capabilities include in-memory caching and distributed storage interfaces, supporting temporary storage and fast access to raw data. For example, a SQL query optimizer is encapsulated for relational database proxies, enabling automatic generation of efficient query statements based on data volume and query complexity.

[0105] Processing layer proxy capability encapsulation: This encapsulates data cleaning, transformation, and quality control capabilities for the processing layer proxy. Data cleaning capabilities include anomaly detection algorithms, missing value handling strategies, and deduplication mechanisms; transformation capabilities include type conversion, format standardization, and data structure reorganization functions; and quality control capabilities include data integrity checks, consistency verification, and accuracy assessment methods. In particular, feature engineering components are encapsulated for the data transformation proxy, enabling automatic generation, selection, and transformation of features based on subsequent analysis needs.

[0106] The analytics layer agent encapsulates model training, inference, and result interpretation capabilities. Model training capabilities include various machine learning algorithms, deep learning frameworks, and statistical modeling methods; inference capabilities include model inference engines, rule execution engines, and decision tree evaluators; and result interpretation capabilities include feature importance analysis, model transparency tools, and result confidence assessment. For example, an Automated Machine Learning (AutoML) engine can be encapsulated for the predictive analytics agent, enabling it to automatically select the most suitable algorithm and parameters for the current data.

[0107] Presentation Layer Proxy Capability Encapsulation: This encapsulates data visualization, interaction design, and narrative expression capabilities for the presentation layer proxy. Data visualization capabilities include various chart generators, geographic information visualization components, and multidimensional data display tools; interaction design capabilities include filters, drill-down functions, and parameter adjustment controls; and narrative expression capabilities include automatically generated data insight descriptions, key indicator highlighting, and trend interpretation generators. For example, a visualization recommendation engine can be encapsulated for dashboards, automatically recommending the most suitable visualization method based on data characteristics.

[0108] Based on the core capability modules of the data layer proxy, processing layer proxy, analysis layer proxy, and presentation layer proxy, a unified message structure is designed, including a sending proxy identifier, a receiving proxy identifier, message type, priority, and timestamp. This unified message structure is the foundation for efficient communication between proxies. It adopts a structured message design pattern, dividing messages into two main parts: a message header and a message body. The message header contains metadata information, where the sending and receiving proxy identifiers use globally unique UUID formats to ensure proxy identity recognition in a distributed environment. The message type defines various interaction types, including task allocation, data transmission, status query, and result feedback, facilitating quick understanding of the message intent by the receiver. The priority field uses a numerical level of 1-10 to guide the urgency of message processing; high-priority messages are scheduled and processed first. The timestamp accurately records the message creation time for message sorting, timeout detection, and performance analysis. The message body uses a flexible JSON structure to adapt to the content requirements of different message types, while also supporting Base64 encoding or reference passing of binary data, balancing expressiveness and transmission efficiency. In addition, auxiliary fields such as version number, session ID, and checksum are integrated into the message structure, which are used for protocol compatibility management, related message association, and data integrity verification, respectively, to ensure the reliability and maintainability of communication.

[0109] The unified message structure supports multiple communication modes, including request-response, publish-subscribe, broadcast, and point-to-point direct communication. Based on the unified message structure, four complementary communication modes are implemented to adapt to the interaction needs of different scenarios. The request-response mode is implemented synchronously or asynchronously. The sender includes a unique request ID in the message, and the receiver returns a response message containing the same request ID after processing. This is suitable for task allocation and query operations that require explicit result feedback; it supports request timeout mechanisms and retry strategies to ensure communication reliability. The publish-subscribe mode establishes a topic-based message distribution mechanism. Agents can subscribe to topics of interest and automatically receive notifications when messages are published on relevant topics, decoupling the sender and receiver. This is particularly suitable for one-to-many information distribution scenarios, such as broadcasting analysis results to multiple visualization agents. The broadcast mode is a special case of publish-subscribe, where messages are sent to all agents. It is typically used for situations where all agents need to be aware of state changes, global configuration updates, etc., and implements message deduplication and propagation control mechanisms to avoid message storms. Point-to-point direct communication mode establishes a dedicated channel for agents requiring high-frequency, low-latency interaction, bypassing central message routing to directly exchange data. This significantly reduces communication overhead and is suitable for data-intensive processing, such as large-volume raw data transmission from data layer agents to processing layer agents. It can automatically switch between these four modes based on communication needs, network conditions, and load, optimizing communication efficiency.

[0110] Based on the aforementioned communication methods, a standard protocol for agent collaboration is defined to standardize the processing flow of collaboration request initiation, collaboration invitation response, task allocation negotiation, and execution status synchronization. The standard protocol adopts a state machine model design, clearly defining the complete lifecycle of agent collaboration and the processing rules for each stage. In the collaboration request initiation phase, the initiating agent needs to provide information such as task description, resource requirements, and expected completion time, and will select a suitable target agent based on task characteristics. In the collaboration invitation response phase, the target agent assesses its own capabilities and current load, decides to accept, reject, or propose modification suggestions, and forms a preliminary collaboration intention. In the task allocation negotiation phase, the participating agents, through multiple rounds of communication, determine a detailed task decomposition plan, interface specifications, time nodes, and quality standards, reaching a formal collaboration agreement. In the execution status synchronization phase, each agent periodically reports task progress, intermediate results, and resource usage, and the coordinating agent summarizes the status information and makes necessary collaboration adjustments. The standard protocol also defines an exception handling mechanism, including agent failure detection, task reassignment, and conflict resolution strategies, to ensure the robustness of collaboration. Furthermore, the protocol supports dynamic adjustments to the collaboration mode, such as upgrading from loose collaboration to tight collaboration, or switching from synchronous execution to asynchronous execution, to adapt to changing needs during task execution. This complete set of standard protocols enables orderly collaboration between agents, significantly improving the efficiency and success rate of complex task execution.

[0111] The standard protocol establishes data flow processing between agents, supporting large-scale data chunked transmission, incremental synchronization, and streaming processing, resulting in a communication protocol that supports efficient collaboration between agents. Data flow is one of the most critical aspects of agent collaboration. For the large-scale data processing needs common in BI scenarios, several efficient data transmission strategies are designed. The chunked transmission mechanism divides large datasets into fixed-size data blocks, each transmitted independently and carrying metadata (such as block number and checksum). The receiver performs reassembly and integrity verification, supporting parallel transmission and retransmission upon failure, significantly improving the transmission efficiency of large datasets. The incremental synchronization mechanism maintains data versions and change logs, transmitting only the data that has changed since the last synchronization, greatly reducing redundant data transmission, making it particularly suitable for analysis scenarios with periodic updates. The streaming processing mechanism establishes a persistent data stream channel. Data producers continuously push newly generated data, and consumers process the received data stream in real time, without waiting for the complete dataset transmission to begin processing, significantly reducing latency and suitable for real-time analysis needs. It also implements intelligent data routing, automatically selecting the most suitable transmission method based on data volume, network conditions, and processing urgency. For example, direct transmission is used for small, critical data, while pass-by-reference (transmitting only the data location and access credentials) is used for large, non-urgent data. Data compression and encoding optimization are also key features, automatically selecting the optimal compression algorithm for different data types to balance compression ratio and computational overhead. Through these data flow optimization technologies, it can efficiently process datasets ranging from GB to TB, providing a robust data foundation for the execution of complex BI tasks.

[0112] Effective communication between agents is the foundation for collaboration. This step designs a complete communication protocol that supports horizontal collaboration between agents at the same level and vertical collaboration between agents across different levels, enabling the entire agent network to form a coordinated and unified intelligent system.

[0113] Message Structure Design: A unified message structure is designed, comprising a message header and a message body. The message header includes metadata such as the sending agent ID, receiving agent ID, message type, priority, timestamp, and session identifier; the message body contains specific business data, task description, execution results, or control instructions. The message structure is implemented using JSON format for easy cross-platform transmission and processing. For example, a data transmission message sent from the data layer agent to the processing layer agent might include information such as data source description, data samples, and data statistical characteristics.

[0114] Communication Mode Implementation: Supports multiple communication modes, including request-response, publish-subscribe, broadcast, and point-to-point direct communication. Request-response mode is suitable for task assignments requiring explicit feedback; publish-subscribe mode is suitable for one-to-many information distribution; broadcast mode is suitable for network-wide notifications; and point-to-point direct communication is suitable for high-frequency, low-latency data exchange. For example, an analytics layer agent can use the publish-subscribe mode to distribute analysis results to multiple presentation layer agents, enabling different visualization agents to simultaneously receive and process the same analytical data.

[0115] Collaboration Protocol Definition: This defines a standard protocol for collaboration between agents, specifying how agents initiate collaboration requests, respond to collaboration invitations, negotiate task allocation, synchronize execution states, and handle exceptional situations during collaboration. The protocol employs a state machine model, clearly defining the transition conditions and processing logic between each state during collaboration. For example, when a predictive analytics agent requires multiple data processing agents to collaborate on large-scale data, a team collaboration request can be initiated through the collaboration protocol to negotiate data sharding strategies and processing priorities.

[0116] Data flow mechanism: Implements an efficient data flow mechanism between agents, supporting chunked transmission, incremental synchronization, and streaming processing of large-scale data. Different data transmission strategies are provided for different data volumes and real-time requirements, including directly transmitting complete data, transmitting data references or access permissions, and transmitting data summaries or feature descriptions. For example, when processing large amounts of data, the data layer agent may not directly transmit the raw data but instead provide data access interfaces and preprocessing suggestions, which the processing layer agent can then access as needed.

[0117] Please refer to Figure 3 , Figure 3 A flowchart illustrating the two-stage business process optimization mechanism provided in this embodiment of the invention. Figure 3 As shown, the mechanism includes a forward task decomposition phase and a reverse collaboration optimization phase, which realizes the automatic generation and optimization of business processes through forward and reverse propagation.

[0118] In a preferred embodiment, the forward task decomposition of the business objective includes:

[0119] The business objectives are semantically parsed to extract core elements including the analysis object, time range, analysis type, and business constraints, and converted into a structured target representation. Deep learning-based natural language processing technology is used to perform multi-level semantic parsing of the user-input business objectives. First, named entity recognition technology is used to identify key entities in the text, such as "sales data," "last three months," and "regional distribution." Then, dependency parsing is used to determine the relationships between entities, clarifying the analysis subject and modifiers. Next, an intent classification model is used to determine the analysis type, such as trend analysis, cause diagnosis, predictive analysis, or comparative analysis. Finally, business constraints are identified, such as "accuracy rate not less than 90%" and "response time less than 5 seconds." Composite objectives are also processed; for example, "analyze the reasons for sales decline and predict the trend for the next quarter" will be broken down into two related objectives: diagnostic analysis and predictive analysis. The parsing results are converted into a structured representation in JSON format, containing fields such as the main analysis object, time dimension (start and end time, granularity), spatial dimension (region, department, etc.), analysis type, indicator definition, filtering conditions, and business constraints, forming the basis for subsequent task decomposition. It will also proactively ask users questions when key information is unclear through an interactive clarification mechanism, ensuring the completeness and accuracy of the target representation.

[0120] Based on the structured target representation, a top-down recursive decomposition algorithm is employed to progressively break down high-level targets into multiple interconnected sub-targets. The recursive decomposition algorithm, based on BI domain knowledge graphs and analysis pattern libraries, systematically breaks down complex targets into more specific and actionable sub-targets. The algorithm first identifies the analysis type of the target and applies corresponding decomposition templates according to different types. For example, trend analysis is typically decomposed into data collection, time series processing, trend extraction, and visualization; cause diagnosis is decomposed into sub-targets such as baseline establishment, factor identification, correlation analysis, and impact quantification. The decomposition process considers multiple factors, including data dependencies, computational complexity, business importance, and time constraints, and uses dynamic programming to find the optimal decomposition path. Multi-dimensional decomposition is supported, allowing for both horizontal decomposition according to analysis process stages (data preparation, feature engineering, model application, result interpretation) and vertical decomposition according to analysis granularity (overall analysis, dimensional drill-down, anomaly focusing). For complex analysis objectives, a hierarchical objective tree is constructed. Upper-level nodes represent abstract objectives, lower-level nodes represent specific implementation steps, and connections between nodes indicate logical relationships between objectives, such as sequential order, parallel execution, or conditional dependencies. The decomposition process is iterative, dynamically adjusting the decomposition strategy based on preliminary results and available resources to ensure that the generated sub-objectives meet both business requirements and execution capabilities.

[0121] The multiple interrelated sub-goals are further decomposed until a granularity suitable for direct assignment to a specific agent is reached, forming sub-task specifications. In this stage, the sub-goals obtained in the previous step are further refined until an operable task granularity is achieved. An agent capability model library is maintained, recording the functional boundaries and processing capabilities of various agents, providing a reference standard for task granularity judgment. The refinement process employs pattern matching and heuristic rules; for example, "sales trend analysis" may be refined into specific tasks such as "sales data extraction," "time dimension aggregation," "seasonal decomposition," "trend line fitting," and "trend chart generation." For each refined sub-task, a detailed task specification is generated, including input data requirements (data source, fields, format, quality standards), processing logic (algorithm selection, parameter settings, execution conditions), output specifications (result format, accuracy requirements, delivery method), and performance constraints (response time, resource limitations). Standardized templates are used for task specifications to ensure that different agents can accurately understand and execute them. Interface compatibility between sub-tasks is also checked to ensure that the output of upstream tasks meets the input requirements of downstream tasks, and data transformation tasks are inserted when necessary to eliminate incompatibility. For computationally intensive tasks, we assess whether further decomposition into parallelizable subtasks is necessary to improve processing efficiency. For data-intensive tasks, we consider data sharding strategies to reduce the amount of data processed per task. This refined task decomposition ensures that each subtask can be efficiently executed by a specific agent, laying the foundation for subsequent task allocation.

[0122] Based on the logical relationships and data flow between the subtask specifications, a task dependency graph is constructed. All decomposed subtasks are organized into a Directed Acyclic Graph (DAG) to accurately represent the dependencies and execution order between tasks. Nodes in the graph represent subtasks, and edges represent dependencies between tasks. Edge attributes include dependency type (strong or weak dependency), data flow content, and transmission method. By analyzing the input and output definitions in the task specifications, data dependencies between tasks are automatically identified; by analyzing the business semantics of the tasks, logical dependencies are identified, such as certain analyses only being meaningful after other analyses are completed. During the construction process, potential circular dependencies are detected and resolved to ensure the graph's acyclicity and avoid execution deadlocks. After the dependency graph is constructed, critical path analysis is performed to identify critical task chains affecting overall execution time and mark sets of tasks that can be executed in parallel, providing a basis for subsequent scheduling optimization. The in-degree (number of dependent tasks) and out-degree (number of dependent tasks) of each task are also calculated to assess the task's importance and risk level. Tasks with high out-degrees will affect multiple downstream tasks upon failure and are therefore given higher monitoring priority. For complex dependency graphs, multi-level views are supported, allowing users to view the overall dependency structure or focus on specific subgraphs for easier understanding and management. The dependency graph not only guides the execution order of tasks but also provides contextual information for exception handling. When a task fails, it can quickly identify affected downstream tasks and implement appropriate recovery strategies.

[0123] Based on the task dependency graph and the available agent capabilities in the system, the most suitable agent or agent team is selected for each subtask, generating the initial task execution network. Task allocation is a multi-objective optimization problem, requiring the system to balance various factors, including agent expertise, historical performance, current load, resource consumption, and task priority. The system maintains agent capability profiles, recording information such as the task types each agent excels at, supported algorithms, processing capacity boundaries, and historical success rates. The allocation process first generates a candidate agent list for each subtask, evaluating the suitability between agents and tasks using a capability matching algorithm; then, it considers the current system state, including the load level of each agent, available resources, and network connection quality; finally, it applies heuristic algorithms or reinforcement learning methods to generate the overall optimal allocation scheme. For complex tasks, agent teams are assembled, defining roles within the team, such as coordinators, executors, and supervisors, and establishing a collaborative mechanism within the team. The allocation results form the initial task execution network, containing information such as all subtasks, execution agents, estimated start times, estimated completion times, resource allocation plans, and data flow paths. The system supports visual preview and manual intervention for task allocation, allowing experts to adjust automatically generated allocation schemes based on business knowledge. The initial execution network is not static but adaptive, dynamically adjusting based on feedback during execution. For example, when a performance degradation of an agent is detected, its unstarted tasks are automatically reassigned to other suitable agents, ensuring overall execution efficiency and reliability.

[0124] Forward task decomposition is the first stage of the two-stage business process optimization mechanism. Its core is to gradually decompose the user's high-level business goals into an executable network of sub-tasks, similar to the forward propagation process of a neural network.

[0125] Business Objective Semantic Parsing: First, the system receives user input describing the business objective, which may be in natural language (e.g., "Analyze the reasons for the sales decline in the past three months and predict the trend for the next quarter") or structured requirements (e.g., the analysis type and parameters selected through the interface). Using natural language processing techniques, including semantic analysis, intent recognition, and entity extraction, the core elements of the business objective are parsed out, such as the analysis object (sales data), time frame (the past three months), and analysis type (cause diagnosis and predictive analysis). This complex business description is then transformed into a structured objective representation, serving as the starting point for task decomposition.

[0126] Goal-level decomposition algorithm: Based on the parsed business objectives, a top-down recursive decomposition algorithm is used to progressively break down high-level objectives into multiple interrelated sub-objectives. The decomposition process follows the general workflow of business logic and data analysis. For example, "analysis of the reasons for sales decline" is decomposed into sub-objectives such as "sales data trend analysis," "correlation analysis of influencing factors," and "anomaly detection at key nodes." Each sub-objective can be further decomposed until it reaches a granularity that can be directly assigned to a specific agent for execution. The decomposition algorithm uses dynamic programming to find the optimal decomposition path based on task complexity, dependencies, and resource status.

[0127] Subtask specification definition: Each decomposed subtask has a clearly defined specification, including input data requirements, processing logic, output format, quality standards, and time constraints. For example, the specification for the "Sales Data Trend Analysis" subtask might include required sales record fields, time granularity, trend calculation methods, and visualization requirements. The task specifications use a unified description format to ensure that different agents can accurately understand and execute the task requirements. Data flow compatibility between subtasks is automatically checked to ensure that the output of upstream tasks meets the input requirements of downstream tasks.

[0128] Task dependency modeling: Based on the logical relationships and data flow between subtasks, a complete task dependency graph (DAG, Directed Acyclic Graph) is constructed. The dependency graph clearly represents the execution order and parallel possibilities of subtasks, providing a foundation for subsequent scheduling and optimization. The critical path of the dependency graph is analyzed to identify potential execution bottlenecks, and attempts are made to optimize the overall process through task reorganization or parallel execution. For example, "sales data cleaning" is identified as a common dependency of multiple subsequent analyses, and it is prioritized as a critical task.

[0129] Task Assignment and Initial Orchestration: Based on task specifications and available agent capabilities in the system, the most suitable agent or agent team is selected for each subtask. Task assignment considers various factors, including agent expertise, historical performance, current load, and resource requirements. An initial task execution plan is generated, including the execution agent for each subtask, estimated start time, estimated completion time, and resource allocation. This initial plan forms the framework of the BI business process, providing a foundation for subsequent optimizations.

[0130] In a preferred embodiment, the step of adjusting the agent cooperation method based on the execution result of the initial task execution network using a reverse cooperative optimization algorithm includes:

[0131] The execution results of the initial task execution network are evaluated from multiple dimensions, including result quality, execution efficiency, collaboration performance, and user satisfaction. A comprehensive evaluation system is constructed to quantitatively analyze all aspects of task execution. The result quality dimension assesses the accuracy, completeness, consistency, and reliability of the analysis results, using various indicators such as accuracy, precision, recall, and F1 score, with different weights assigned based on the business scenario. For analysis tasks that cannot be directly quantified, relative evaluation is conducted using expert rules and historical benchmarks. The execution efficiency dimension focuses on resource utilization and time consumption, measuring indicators such as end-to-end response time, CPU / memory utilization, I / O operation count, and data transfer volume. Reasonable efficiency benchmarks are set according to task type to identify abnormally time-consuming or excessively resource-intensive stages. The collaboration performance dimension assesses the degree of cooperation between agents, including communication efficiency (message round trip count, communication latency), coordination consistency (state synchronization success rate), and task handover smoothness (interface matching degree, data conversion overhead), etc. These indicators reflect the overall collaboration quality of the agent team. The user satisfaction dimension integrates explicit and implicit user feedback, including user ratings, interaction behaviors (such as drill-down depth and dwell time), and repeat usage rates, with a particular focus on user comprehension of the analysis results and their decision-support value. The evaluation process employs a multi-layered approach, conducting both micro-evaluations of individual subtasks and macro-evaluations of the overall process to create a comprehensive performance profile, providing a data foundation for subsequent optimization.

[0132] Based on the multi-dimensional task evaluation results, the reasons for poor task performance are analyzed, identifying key improvement points including data quality issues, inappropriate algorithm selection, suboptimal parameter configuration, or poor collaboration. Root cause analysis is employed to delve into the underlying causes of performance problems. First, anomaly detection is performed on evaluation metrics to identify performance dimensions that significantly deviate from expectations. Then, through techniques such as decision trees and association rule mining, a correlation model between performance problems and possible causes is established. Next, execution logs, intermediate results, and resource monitoring data are analyzed to find the specific stages and triggering conditions where problems occur. For data quality issues, anomalies in data integrity, consistency, accuracy, and timeliness are detected, such as excessively high missing value ratios, distribution shifts, or outlier clusters. For algorithm selection issues, the matching degree between the current algorithm and data characteristics and task requirements is evaluated, checking for underfitting, overfitting, or model assumption violations. For parameter configuration issues, parameter sensitivity is analyzed to identify the optimal range of key parameters and deviations from current settings. For poor collaboration issues, communication bottlenecks, resource contention, and inconsistent states are detected as collaboration obstacles. During the analysis, the scope, severity, and difficulty of remediation of the problem are comprehensively considered. Multiple identified issues are prioritized to ensure that subsequent optimizations focus on the most critical improvement opportunities. A problem knowledge base is also established to record historical problem patterns and solutions. Through similarity matching, known problem types are quickly identified, improving diagnostic efficiency.

[0133] For the key improvement points, specific optimization signals are generated, including problem descriptions, severity, priority, and recommended actions. These optimization signals are structured representations conveying improvement intentions, designed in a standardized format to ensure accurate understanding and execution by different agents. The problem description uses precise technical language to clearly explain the nature, scope, and specific manifestations of the problem, such as "The accuracy of the predictive model significantly decreases on long-tail data, affecting marketing decisions for niche customer groups." Severity is categorized into five levels (critical, high, medium, low, and warning), based on the degree to which the problem hinders the achievement of business objectives; for example, data errors affecting core decisions are marked as "critical." Prioritization considers factors such as severity, scope, and timing of remediation to guide resource allocation and processing order, using a numerical rating (1-100) to make priorities comparable and rankable. The recommended measures section is the core of signal optimization, providing specific improvement directions and operational suggestions, encompassing multiple levels: Tactical level suggestions offer direct solutions to specific problems, such as "adding outlier handling steps and using IQR methods to identify and handle outliers"; strategy level suggestions provide broader optimization directions, such as "enhancing the model's learning ability for small sample classes, considering the introduction of data augmentation or transfer learning techniques"; and architectural level suggestions involve adjustments to the process structure, such as "adding a feature importance analysis step between data processing and model training to guide feature selection." Each suggestion also provides an expected effect assessment and warnings of potential side effects to help the agent make comprehensive considerations. The optimized signal generation process combines a rule base and a machine learning model, learning the most effective improvement strategies based on historical success cases and continuously optimizing and enriching them through operation.

[0134] The backpropagation algorithm transmits specific optimization signals from downstream tasks to upstream tasks, forming a complete optimization chain. Borrowing from the backpropagation idea of ​​neural networks, the algorithm enables the reverse flow of optimization information, allowing upstream components to adjust their behavior according to downstream needs. The algorithm first constructs a transpose of the task dependency graph, reversing the original data flow direction to form a transmission path for the optimization signal. Then, starting from the node where the problem is discovered, the optimization signal is transmitted along the reverse path. The transmission process is not a simple signal forwarding but includes signal transformation and aggregation: signal transformation translates the optimization requirements of downstream tasks into a form that upstream tasks can understand and execute. For example, the visualization agent's problem of "insufficient color contrast in charts" is transformed into the analysis agent's optimization signal of "increasing the distinguishability of result data." Signal aggregation processes potentially conflicting optimization requirements from multiple downstream tasks, generating a comprehensive optimization scheme that balances the needs of all parties through methods such as priority weighting, constraint satisfaction, or multi-objective optimization. The algorithm also implements a signal attenuation mechanism, where the influence of the optimization signal weakens as the transmission distance increases, avoiding over-optimization over long distances. Simultaneously, a transmission threshold is set, ensuring that only signals reaching a certain importance continue to be transmitted upwards, reducing noise. For complex dependency networks, the algorithm employs topological sorting to ensure the orderly transmission of optimization signals and avoid signal oscillations caused by cyclic dependencies. During the reverse propagation process, the signal transmission path and transformation history are recorded to form a complete optimization chain, facilitating subsequent analysis and verification of the optimization effect.

[0135] Based on the specific optimization signals, each agent adjusts its behavior and strategies, while simultaneously optimizing the global collaboration mode between agents to obtain the optimized business process. Agent behavior adjustment is the execution phase of the optimization process, defining corresponding adjustment mechanisms for different types of agents. Data layer agents can adjust data source selection, sampling strategies, query optimization, and caching mechanisms; processing layer agents can modify cleaning rules, transformation logic, feature engineering methods, and quality control standards; analysis layer agents can change algorithms, adjust model parameters, increase model complexity, or introduce integration methods; and presentation layer agents can change visualization types, adjust interaction design, enhance explanatory text, or optimize layout structure. The adjustment process adopts a gradual strategy, first conducting small-scale tests to evaluate the effects, and then gradually expanding the application scope to ensure the safety and effectiveness of the optimization. Simultaneously, the global collaboration mode between agents is also optimized, including adjusting communication frequency and content (reducing unnecessary message passing), reorganizing agent team structures (forming closely cooperating agents into close-knit teams), optimizing resource allocation strategies (rationally allocating computing resources based on task importance and complexity), and improving collaboration protocols (simplifying negotiation processes and improving decision-making efficiency). Furthermore, an optimization knowledge base will be established to record the effects and applicable conditions of various optimization operations, continuously improving the accuracy of optimization decisions as experience accumulates. The optimized business processes not only improve performance metrics but, more importantly, enhance adaptability and robustness, better responding to changes in data characteristics and the evolution of business needs, providing continuous and reliable analytical support for enterprises.

[0136] Reverse collaborative optimization is the second stage of the two-stage business process optimization mechanism. Similar to the backpropagation process of a neural network, it optimizes the entire business process through feedback of task execution results.

[0137] A multi-dimensional task evaluation system: First, a comprehensive task evaluation system is established to assess the performance of each sub-task from multiple dimensions. Evaluation dimensions include result quality (accuracy, completeness, consistency, etc.), execution efficiency (time consumption, resource utilization, etc.), collaboration performance (information transmission efficiency, team coordination, etc.), and user satisfaction (result usability, understandability, etc.). Different evaluation weights are set for different types of tasks; for example, accuracy is emphasized for prediction tasks, while response speed is emphasized for real-time analysis tasks. Various objective indicators and user feedback are collected to generate a comprehensive task evaluation report.

[0138] Optimize the signal generation mechanism: Based on task evaluation results, generate specific optimization signals to guide business process improvements. Analyze the reasons for poor task performance and identify key improvement points, such as data quality issues, inappropriate algorithm selection, suboptimal parameter configuration, or poor collaboration. For each improvement point, generate clear optimization suggestions, including the adjustment target, optimization direction, and expected results. Optimization signals use a unified format, including elements such as problem description, severity, priority, and recommended measures. For example, for the problem of insufficient prediction accuracy, optimization signals such as "increase the historical data time span" or "try integrating multiple prediction models" might be generated.

[0139] Implementation of the backpropagation algorithm: A backpropagation algorithm is designed to pass optimization signals from downstream tasks to upstream tasks, forming a complete optimization chain. It analyzes the dependencies between tasks, identifying the root causes and impact paths of problems. For example, poor visualization results may stem from improper upstream data processing or an unreasonable structure of the analysis results. The algorithm, based on the dependency graph, propagates optimization signals along the reverse path, transforming and aggregating them during the process, enabling upstream tasks to understand and respond to the needs of downstream tasks. For conflicting requests from multiple downstream tasks, it balances and coordinates them, generating a comprehensive optimization solution.

[0140] Agent Behavior Adjustment Strategy: Based on received optimization signals, this strategy guides each agent to adjust its behavior and strategies. Corresponding adjustment mechanisms are defined for different types of agents; for example, data processing agents can adjust cleaning rules, analysis agents can change algorithms or adjust parameters, and visualization agents can change their display format. Agents determine the scope and intensity of adjustments based on the priority and severity of the optimization signals. Progressive adjustments and A / B testing are supported, allowing agents to safely try new strategies and evaluate their effectiveness. For example, upon receiving a signal that "data noise affects prediction accuracy," the data processing agent might enhance outlier detection and smoothing.

[0141] Global Collaboration Mode Optimization: Based on feedback from multiple task executions, continuously optimize the global collaboration mode among agents. Analyze information flow patterns, resource allocation efficiency, and collaboration bottlenecks within the agent network to identify collaboration links requiring improvement. Adjust the organizational structure, communication mechanisms, and workflows of agent teams to improve overall collaboration efficiency. Achieve automatic evolution of the collaboration mode, enabling automatic adjustment of the optimal collaboration method according to business changes. For example, if it is found that certain data processing and analysis tasks are always closely related, the agent organizational structure may be adjusted to form fixed teams, reducing communication overhead.

[0142] Please refer to Figure 4 , Figure 4 This is a diagram illustrating the decoupling risk visualization architecture provided in an embodiment of the present invention. Figure 4As shown, this method decouples risk into two dimensions: model parameters and data parameters, thereby achieving accurate quantification and visualization of risk.

[0143] In a preferred embodiment, constructing a two-parameter risk model comprising a model parameter vector and a data parameter vector includes:

[0144] Based on the optimized business process, model parameter vectors describing the analysis methods, algorithm selection, and parameter configuration are defined, along with data parameter vectors describing data quality, data distribution, and sample size. First, a comprehensive analysis of the optimized business process is conducted to identify the analysis methods and data characteristics used in each stage, constructing a comprehensive parameter representation. The model parameter vector is a point in a multi-dimensional vector space, with each dimension corresponding to a model characteristic factor. Core dimensions include algorithm type (e.g., decision trees, neural networks, regression models, represented using one-hot encoding), model complexity (e.g., tree depth, number of network layers, number of features, normalized to the 0-1 range), learning parameters (e.g., learning rate, regularization coefficient, number of iterations), ensemble strategies (e.g., bagging, boosting, stacking), and optimization methods (e.g., gradient descent, Newton's method, genetic algorithms). Vector dimensions are dynamically adjusted for different analysis scenarios; for example, time series analysis adds specific dimensions such as time window size and seasonality handling methods, while text analysis includes dimensions such as word vector methods and language model selection. The data parameter vectors also employ a multidimensional representation, with core dimensions including data volume (logarithmic value of sample size), completeness indicators (missing rates of key fields), distribution characteristics (statistics such as skewness, kurtosis, and entropy), noise level (outlier ratio and signal-to-noise ratio estimation), class balance (class frequency variance and Gini coefficient), and timeliness indicators (data freshness and update frequency). These features are automatically extracted using data profiling technology to form standardized data parameter vectors, ensuring the comparability of different datasets. These two types of parameter vectors together constitute the input space for risk assessment, providing a foundation for subsequent risk modeling.

[0145] A risk assessment function is constructed to map the model parameter vector and data parameter vector to a multi-dimensional risk space, calculate basic risk indicators, and aggregate them to form a comprehensive risk score. The risk assessment function is the core of the two-parameter risk model, establishing a mapping relationship from the parameter space to the risk space. A hierarchical design is adopted, first defining several basic risk indicators covering the main risk types in BI analysis: prediction bias risk (deviation between model predictions and actual values), overfitting risk (the model performs well on training data but has poor generalization ability), data drift risk (inconsistency between data distribution in the production environment and training data), sparse data risk (unreliable predictions due to insufficient samples in specific regions), anomaly sensitivity risk (overreaction of the model to abnormal inputs), and interpretability risk (lack of transparency in the model's decision-making process). Each basic risk indicator is calculated by a dedicated sub-function. These sub-functions comprehensively consider the interaction between model parameters and data parameters. For example, the overfitting risk function assesses the match between model complexity and sample size; the risk value increases when a complex model is applied to a small dataset. The risk function design combines theoretical models and empirical data. By analyzing the correlation patterns between parameter configurations and analysis failures in historical cases, a statistical learning model is established to accurately predict the risk level of different parameter combinations. It supports multiple risk aggregation methods, including weighted average, maximum value selection, and hierarchical aggregation. The default method is weighted average, and the weights can be adjusted according to business scenarios and risk preferences. The final comprehensive risk score is standardized to a range of 0-100 and divided into five risk levels (very low, low, medium, high, and very high) for easy understanding and decision-making. A risk decomposition view is also provided, displaying the contribution of each basic risk indicator to help users understand the sources of risk.

[0146] The specific mathematical expression of the risk assessment function is as follows:

[0147] R(m,d) = w1·R1(m) + w2·R2(d) + w3·R3(m,d)

[0148] Where m represents the model parameter vector, containing n-dimensional features such as algorithm type, model complexity, and parameter configuration; d represents the data parameter vector, containing m-dimensional features such as data size, quality, and distribution characteristics; R1(m) represents the model intrinsic risk function, which evaluates the model's own complexity and stability; R2(d) represents the data quality risk function, which evaluates the uncertainty caused by data characteristics; R3(m,d) represents the model-data fit risk function, which evaluates the degree of matching when a specific model is combined with specific data; w1, w2, and w3 are weight coefficients that satisfy w1+w2+w3=1 and can be dynamically adjusted according to the business scenario.

[0149] In the specific implementation, R1(m) adopts a complexity evaluation model based on random forest, R2(d) adopts a data quality evaluation model based on information entropy and statistical features, and R3(m,d) adopts a similarity matching model based on historical cases. The model training data comes from 500 historical analysis cases labeled with risk levels, including complete parameter configurations and actual risk event records.

[0150] Based on the comprehensive risk score, a risk posture calculation engine is developed to quantify the risk of each step in the optimized business process, obtaining risk quantification data. The risk posture calculation engine extends risk assessment from static analysis to dynamic assessment across the entire process, achieving a panoramic risk scan of the entire BI business process. The engine first constructs a computational graph representation of the business process, with nodes corresponding to processing components and edges representing data flow. Then, it extracts the model parameter vector and data parameter vector for each node and applies a risk assessment function to calculate the node's risk score. Next, it analyzes the risk propagation effect between nodes; some risks accumulate downstream along the data flow (such as data quality issues), while others may be mitigated at specific nodes (such as data cleaning steps reducing noise risk). The engine implements multiple risk calculation modes: a global scan mode performs a comprehensive risk assessment of the entire process, generating a complete risk map; a focus analysis mode performs in-depth risk analysis on specific components or sub-processes, providing a more refined risk profile; a difference comparison mode compares the risk changes between two process versions, assessing the risk impact of optimization measures; and a time series mode tracks the evolution trend of risk over time, providing early warnings of rising risk trends. The calculation process supports incremental updates; when process or data characteristics change, only the affected parts are recalculated, improving efficiency. The engine outputs standardized risk quantification data, including node risk scores, risk type distribution, risk propagation paths, and temporal trends. This data can be directly used for decision support and serves as the foundation for subsequent risk interest point identification. It also implements interpretability functionality for risk calculation, generating risk assessment reports that explain the reasons and impacts of high-risk scores.

[0151] The risk interest point (PI) identification algorithm automatically identifies high-risk points, inflection points, sensitive points, and anomalies in the risk quantification data, thus deriving PIs. This algorithm is the intelligent core of risk visualization, extracting the most noteworthy key points from massive amounts of risk data to guide user attention. The algorithm implements multiple PI detection strategies: high-risk point detection uses threshold analysis and spatial clustering to identify areas where risk scores consistently exceed warning lines, calculating the area's size, shape, and boundary features; inflection point detection applies differential geometry and curvature analysis techniques to locate locations in the parameter space where the risk surface changes drastically. These inflection points typically represent critical conditions for qualitative changes in risk characteristics, providing important guidance for risk control; sensitive point detection uses local sensitivity analysis to calculate the derivative of the risk score with respect to small parameter changes, identifying areas where risk is highly sensitive to specific parameters. These areas require precise parameter control to avoid unexpected risk escalation; and anomaly detection utilizes spatial statistics and outlier analysis to find isolated points significantly different from surrounding risk patterns. These anomalies typically indicate unique risk patterns or potential model defects. For each identified point of interest (POI), the algorithm extracts its key features, including location features (precise coordinates in parameter space), intensity features (absolute value of the risk score and its difference from the surrounding area), morphological features (size, shape, and boundary clarity of the region of interest), and dynamic features (risk trends over time). The algorithm also analyzes the business implications of the POI, assesses potential specific problems such as prediction bias, resource waste, or decision-making errors, and provides targeted risk mitigation suggestions. Interactive exploration of POIs is supported, allowing users to select points of interest for in-depth analysis and view detailed risk composition and parameter sensitivity.

[0152] Based on the risk points of interest and comprehensive risk scores, risk situation data including risk heatmap data and risk level information is generated. Risk situation data is a comprehensive representation of the risk analysis results, employing various complementary data structures and visualization formats to fully showcase risk distribution and characteristics. The risk heatmap is the core representation, using a two-dimensional matrix structure with model parameters and data parameters as the two axes. The color intensity of each cell indicates the risk level of the corresponding parameter combination, typically using a gradient color from green (low risk) to red (high risk). The heatmap supports multiple resolutions, generating both a global overview map and high-resolution detailed maps for high-risk areas. Risk contour data is also generated, connecting parameter points with the same risk score to visually display risk boundaries and gradient changes. For risk points of interest, special marker data is created, using icons of different shapes and colors to represent different types of points of interest, with importance weights assigned to control display priority. To support dynamic analysis, risk trend data is generated, recording the time-varying sequence of risk scores for key parameter combinations, which can be used for trend charts and early warning analysis. All this data is organized into standardized risk posture data packages, including metadata (generation time, data version, assessment scope, etc.) and multiple data views (heatmap data, contour data, point of interest data, trend data, etc.). The data packages use an efficient storage format, supporting incremental updates and rapid retrieval. They can be used directly for visualization or as input for further analysis and optimization. Text summaries of the risk posture are also provided, automatically generating natural language descriptions of key risk findings and recommendations to help non-technical users understand the risk situation.

[0153] A two-parameter risk model is the foundation for decoupling risk visualization. By decoupling risks in the BI business process into two dimensions—model parameters and data parameters—it enables precise quantification and assessment of risk.

[0154] Risk Decoupling Theoretical Framework: First, a theoretical framework for risk decoupling is established, decomposing potential risks in the BI business process into two orthogonal dimensions: model risk and data risk. The model parameter vector describes factors related to the processing logic, such as analytical methods, algorithm selection, and parameter configuration; the data parameter vector describes factors related to data characteristics, such as data quality, data distribution, and sample size. This decoupling method draws on the idea of ​​performance-based prediction, enabling the differentiation between risks caused by model defects and risks caused by data problems, providing a foundation for precise risk management. For example, low prediction accuracy may stem from inappropriate model selection (model risk) or insufficient training data (data risk). After decoupling, different risk response strategies can be adopted specifically.

[0155] Model Parameter Vector Definition: This section defines a detailed parameter vector describing the model's characteristics, encompassing multiple key dimensions. Typical model parameters include: algorithm type (e.g., decision tree, neural network, regression model), model complexity (e.g., tree depth, number of network layers), learning rate, regularization parameters, optimization methods, and ensemble strategies. For different stages of the BI business process, such as data processing, analytical modeling, and results visualization, corresponding subsets of model parameters are defined. Each parameter has a clearly defined value range and sensitivity rating, facilitating subsequent risk quantification. For example, for predictive analytics, the model parameter vector might include key factors such as "prediction algorithm type," "feature selection method," "time window size," and "hyperparameter configuration."

[0156] Data Parameter Vector Definition: This section defines a detailed parameter vector describing the characteristics of the data, encompassing multiple key dimensions. Typical data parameters include: data volume, completeness metrics (such as the proportion of missing values), distribution characteristics (such as skewness and kurtosis), noise level, outlier proportion, category balance, and timeliness metrics. For different types of data sources and business scenarios, corresponding subsets of data parameters are defined. Each parameter has a clear evaluation method and risk threshold, facilitating automated risk assessment. For example, in a customer behavior analysis scenario, the data parameter vector might include key factors such as "proportion of active users," "time span of behavioral data," "completeness of user attributes," and "balance of behavioral category distribution."

[0157] Risk Function Construction: Based on model parameter vectors and data parameter vectors, a risk assessment function is constructed, mapping the parameter vectors to a multi-dimensional risk space. The risk function employs a hierarchical design, first calculating basic risk indicators (such as prediction bias risk, overfitting risk, and data leakage risk), and then aggregating them to form a comprehensive risk score. The design of the risk function combines theoretical models and empirical data, accurately reflecting the impact of parameter changes on risk through historical case analysis and expert knowledge modeling. Customization and adjustment of the risk function are supported to adapt to the risk preferences and business characteristics of different enterprises. For example, the risk function can be expressed as: R(m,d) = w1·R1(m) + w2·R2(d) + w3·R3(m,d), where m represents the model parameter vector, d represents the data parameter vector, and R1, R2, and R3 represent the inherent model risk, data quality risk, and model-data fit risk, respectively.

[0158] Risk posture calculation engine: Based on risk calculation results, it generates a comprehensive risk posture representation. The risk posture adopts a multi-dimensional matrix form, with model parameters and data parameters as the two main axes. Each cell represents the risk level under a specific parameter combination. It supports multiple risk posture views, including a global risk map (covering the entire parameter space), a local risk zoom-in map (focusing on high-risk areas), and a time series risk map (showing risk change trends). It implements multi-level zooming and slicing functions for the risk posture, allowing users to analyze risk distribution from different angles and granularities. For example, data scientists can view the risk posture map formed by the intersection of feature selection methods and data noise levels to understand the sensitivity of different feature selection algorithms to noise.

[0159] In a preferred embodiment, the step of automatically identifying high-risk points, inflection points, sensitive points, and anomalies in the risk quantification data using a risk interest point identification algorithm to obtain risk interest points includes:

[0160] Threshold detection and regional connectivity analysis are employed to identify high-risk points whose risk scores consistently exceed thresholds. First, risk score thresholds are set based on business needs and risk tolerance, typically using an adaptive threshold strategy that dynamically adjusts the threshold level according to historical risk distribution and current business objectives. The threshold detection process performs a multi-dimensional scan in the parameter space, marking all points exceeding the threshold to form preliminary high-risk candidate regions. Then, a regional connectivity analysis algorithm is applied to aggregate adjacent high-risk points into continuous risk regions, calculating the size, shape, and boundary characteristics of each region. Connectivity analysis uses a multi-scale approach, capable of identifying both large-scale high-risk regions and small, concentrated risk hotspots. Statistical characteristics of each high-risk region are also calculated, including average risk value, maximum risk value, risk variance, and risk distribution skewness, which help understand the concentration and distribution patterns of risk. For complex risk regions spanning multiple parameter dimensions, principal component analysis is used to determine the main extension direction of the risk region and identify key parameter combinations leading to increased risk. Furthermore, a time-dimensional persistence analysis is implemented to track the evolution of high-risk regions over time, distinguishing between temporary risk fluctuations and persistent structural risks. This multi-faceted analysis of high-risk points provides precise target areas for risk management, enabling risk control measures to focus on the most critical parameter combinations.

[0161] Gradient analysis and curvature calculation are used to identify inflection points where risk surfaces change drastically. Inflection points are critical locations where risk characteristics undergo qualitative changes, and are crucial for understanding and predicting risk situations. A risk response surface is constructed in parameter space, treating the risk score as a function of parameter vectors, and then differential geometry methods are applied to analyze the surface characteristics. Gradient analysis calculates the first-order derivative vector of the risk function at each point, representing the direction and rate of risk change; high-precision gradient calculation is achieved using the central difference method or automatic differentiation techniques, and regions with significant risk changes are screened based on gradient magnitude. Curvature calculation analyzes the second-order characteristics of the risk surface, including principal curvature, Gaussian curvature, and mean curvature, which reflect the degree of surface bending and morphological changes. Particular attention is paid to the locations where the sign of Gaussian curvature changes, which usually indicate a change from convex to concave or vice versa, corresponding to a fundamental shift in risk characteristics. In multidimensional parameter space, the eigenvalue decomposition of the Hessian matrix is ​​calculated to identify curvature characteristics in various directions. To improve computational efficiency, an adaptive grid strategy is adopted, using denser sampling points in regions with large gradients. For each identified inflection point, its precise location, curvature characteristics, and main direction of change are recorded. The differences in risk characteristics on both sides of the inflection point are analyzed to reveal the patterns of risk transformation. This inflection point information provides important decision-making basis for risk control, especially in guiding the determination of parameter adjustment boundaries and early warning thresholds.

[0162] Local sensitivity analysis is employed to calculate the responsiveness of risk to parameter changes and identify sensitive points. Sensitive points are locations where risk is highly sensitive to changes in specific parameters; these points require precise control because even small parameter shifts can lead to a significant increase in risk. A method combining global and local sensitivity analysis is implemented. Global analysis uses variance decomposition techniques (such as the Sobol index method) to assess the contribution of each parameter to the total risk variation and identify the most important parameters overall. Local analysis calculates partial derivatives or sensitivity coefficients at specific points in the parameter space to measure the responsiveness of risk to small changes in each parameter. Multiple sensitivity indices are used, including standardized sensitivity coefficients, elasticity coefficients, and relative sensitivity, which quantify the strength of the parameter-risk relationship from different perspectives. To comprehensively assess sensitivity, parameter perturbation experiments are conducted, varying parameter values ​​within a small range around a baseline point to observe the changing patterns of risk scores and construct local response surfaces. Particular attention is paid to the locations of sensitivity abrupt changes, which typically represent turning points in the risk mechanism. For high-dimensional parameter spaces, dimensionality reduction techniques are used to identify combinations of sensitive parameters and discover potential parameter interaction effects. Time sensitivity analysis is also implemented to assess the changing trends of parameter sensitivity over time and identify the evolution of risk sensitivity in dynamic environments. The results of sensitivity point analysis directly guide the strategy for fine-tuning parameters, recommending stricter parameter control ranges and more frequent monitoring frequencies for highly sensitive areas.

[0163] This system employs local anomaly detection and spatial statistical analysis to identify outliers that significantly differ from surrounding risk patterns. Outliers typically represent unique risk patterns or potential model flaws, requiring special attention. Several complementary anomaly detection methods are implemented to adapt to different types of anomaly patterns. The Local Outlier Factor (LOF) method calculates the density ratio of each point to its neighborhood, identifying locally sparse outliers; distance-based methods (such as k-nearest neighbor distance) measure the average distance between a point and its neighbors, discovering isolated points far from the main distribution; and spectral methods analyze the low-dimensional representation of risk data to identify points deviating from the main pattern in the projected space. Particular attention is paid to risk mutation points, whose risk values ​​contrast sharply with the surrounding environment, often representing unique risk mechanisms. In terms of spatial statistical analysis, local Moran's I index and Getis-Ord Gi* statistics are calculated to identify spatial clustering and discrete patterns of risks, discovering hot and cold areas. Context-sensitive anomaly detection is also implemented, considering business context and parameter semantics to distinguish meaningful anomalies from noise. For each outlier, its degree of anomalousness (quantified by Z-score or anomaly probability), anomaly type (e.g., value anomaly, relational anomaly, or time series anomaly), and possible causes are assessed. The temporal stability of the outliers is also analyzed, distinguishing between persistent structural anomalies and temporary fluctuations. These outlier analysis results provide important clues for improving risk models and handling special cases, helping to uncover risk blind spots that might be overlooked in routine analysis.

[0164] Feature extraction and attribute analysis are performed on the high-risk points, inflection points, sensitive points, and anomaly points to obtain feature analysis results including location features, intensity features, morphological features, and dynamic features. Feature extraction is a key step in the deep understanding of risk points of interest, and a dedicated feature set is designed for each type of point of interest. Location features describe the precise coordinates and relative positions of the points of interest in the parameter space, including Euclidean coordinates, relative center distance, boundary distance, and nearest neighbor distance; the projected positions of the points of interest on the main business dimensions are also calculated to facilitate understanding the risk distribution from a business perspective. Intensity features quantify the risk level and significance of the points of interest, including absolute risk value, relative risk ratio (ratio to average risk), risk gradient magnitude, and statistical significance (such as p-value or Z-score); these indicators assess the importance of the points of interest from different perspectives. Morphological features describe the geometric and topological characteristics of the region where the points of interest are located, including region area, perimeter, compactness, directionality, and boundary clarity; for high-risk regions, their internal structure is also analyzed, such as the uniformity and centralization of risk distribution. The system dynamically tracks the evolution of points of interest over time, including duration, rate of change, periodic patterns, and trend direction. It implements time series analysis methods such as trend decomposition, change point detection, and periodic analysis to reveal the dynamic characteristics of risk. In addition to these four core feature categories, it extracts correlation features to analyze the correlation between points of interest and other risk and business indicators, establishing a risk-business mapping relationship. The feature extraction process employs a multi-scale approach, considering both local details and the global context to create a comprehensive feature profile. It also implements feature importance assessment to identify the key features that best distinguish different types of points of interest, improving the targeting and efficiency of subsequent analysis.

[0165] Based on the aforementioned feature analysis results, risk impact and business importance are calculated to prioritize the risk points of interest, generating risk posture data containing information on key risk areas. Prioritization is a crucial step in transforming analysis results into action guidelines, employing a comprehensive scoring mechanism to balance the technical characteristics and business value of the risks. Risk impact assessment quantifies the potential harm of points of interest from multiple dimensions, including scope of impact (affected business processes and data range), depth of impact (potential error severity), persistence of impact (risk duration and recovery difficulty), and propagation (the possibility of risk spreading to other areas). A mapping model between risk characteristics and actual impact is established by combining historical cases and expert rules. Business importance assessment considers the relevance of points of interest to key business objectives, including financial impact (potential cost increases or revenue losses), customer impact (impact on user experience and satisfaction), compliance impact (risk of violating regulations or internal policies), and strategic impact (hindrance to long-term business goals). A business domain knowledge graph is maintained to achieve correlation analysis between technical risks and business value. Priority calculation uses a weighted scoring model, adjusting the weights of each factor according to the organization's risk preferences and business strategies to generate a standardized priority score from 0 to 100. It supports multiple sorting views, allowing sorting by overall priority or by specific dimensions (such as scope of impact or financial importance) to meet the needs of different decision-making scenarios. The sorting results directly impact the risk posture display, with high-priority points of interest receiving more prominent visual presentation and more detailed information. The risk posture display data adopts a multi-layered structure, including an overview layer (global risk distribution and key point of interest locations), a focus layer (detailed risk characteristics of high-priority areas), and a details layer (complete feature sets and business explanations for individual points of interest). It also generates risk narrative data, transforming digitized risk analysis into natural language descriptions, highlighting key findings and recommendations for easier understanding and action by non-technical decision-makers. This priority-based risk posture display ensures that limited attention resources are focused on the most critical risk areas, improving the efficiency and effectiveness of risk management.

[0166] The risk interest point identification mechanism is a key component of decoupling risk visualization. By automatically discovering and highlighting key areas in the risk landscape, it helps users quickly locate risk points that require focused attention.

[0167] Definition and Classification of Risk Interest Points: First, the concept and classification system of risk interest points are clearly defined. Risk interest points are areas or points of special significance in a risk landscape, including high-risk points (areas where risk scores exceed a critical threshold), inflection points (areas where risk scores change significantly), sensitive points (areas that react strongly to small changes in parameters), anomalies (isolated areas that are significantly different from the surrounding risk landscape), and conflict points (areas where model risk and data risk assessments differ significantly). Each type of interest point has clear identification criteria and business implications, facilitating subsequent analysis and processing. For example, an inflection point typically represents a critical condition where risk characteristics undergo a qualitative change, providing important guidance for risk control.

[0168] Interest Point Detection Algorithms: A series of specialized algorithms were designed and implemented to automatically detect various types of risk interest points from risk situation data. For high-risk points, threshold detection and regional connectivity analysis are used to identify continuous regions where the risk score consistently exceeds the threshold; for inflection points, gradient analysis and curvature calculation are used to identify locations where the risk surface changes drastically; for sensitive points, local sensitivity analysis is used to calculate the degree of risk response to parameter changes; for outliers, local anomaly detection and spatial statistical analysis are used to identify points that are significantly different from the surrounding risk patterns; for conflict points, the differences between model risk and data risk scores are compared to identify areas with inconsistent assessment results. These algorithms combine traditional statistical methods and machine learning techniques, enabling efficient processing of large-scale risk data. For example, local polynomial fitting and curvature calculation may be used to accurately identify the location of inflection points in the risk surface.

[0169] Points of Interest (POIs) Feature Extraction: This involves extracting features and analyzing attributes of detected risky POIs to obtain their key characteristics and business implications. Extracted features include location features (coordinates in parameter space), intensity features (risk score and its local statistical characteristics), morphological features (shape, size, and boundary characteristics of the region of interest), and dynamic features (risk trends over time). By combining business metadata and historical cases, the business implications and potential impacts of POIs are analyzed, such as potential decision-making biases, data quality issues, or model failures. Structured POI descriptions are generated, including quantitative indicators and qualitative explanations, facilitating user understanding and decision-making. For example, for high-risk POIs detected in a customer segmentation model, their parametric features (such as location in areas of high data sparsity and complex models) are extracted, and it is explained that this may lead to overfitting of customer segmentation, resulting in incorrect allocation of marketing resources.

[0170] Prioritization of Points of Interest: An intelligent prioritization mechanism is implemented based on the risk impact and business importance of points of interest. The prioritization considers multiple factors, including risk severity (absolute level of the risk score), scope of impact (affected business processes and decision-making scope), urgency (speed and urgency of risk evolution), and intervention feasibility (difficulty and cost of resolving the risk). A multi-criteria decision analysis method is used to comprehensively consider all factors and calculate the final priority score. Customized prioritization strategies are supported for different business scenarios; for example, compliance-oriented scenarios focus more on compliance risks, while business growth scenarios focus more on opportunity costs. An ordered list of points of interest is generated to guide user attention allocation. For example, data drift risks affecting the accuracy of monthly sales forecasts might be prioritized at the highest level because they directly relate to inventory management decisions and require timely intervention.

[0171] In a preferred embodiment, generating the editing operation sequence based on the risk situation data includes:

[0172] Based on the aforementioned risk profile data, a set of editing operations applicable to business processes is defined, including basic operations such as component insertion, deletion, replacement, and parameter adjustment. First, the structural characteristics and component types of the business processes are analyzed, and a comprehensive and precise editing operation language is designed. The component insertion operation defines rules for adding new processing units at specific locations in the process, including insertion point selection strategies (such as sequential insertion, parallel branch insertion, or conditional branch insertion), component type matching rules, and interface compatibility verification. A component library is maintained, containing various predefined components such as data processing, analysis algorithms, and visualization, each with clear input / output specifications and functional descriptions. The component deletion operation specifies methods for safely removing process nodes, including dependency checks (ensuring deletion does not disrupt data flow), alternative generation (providing functional replacements when necessary), and side effect assessment (predicting the impact of deletion on downstream components). The component replacement operation defines a mechanism for replacing existing components with functionally similar but higher-performing components, including functional equivalence verification, interface adaptation strategies, and performance improvement prediction. Component similarity calculation is implemented, enabling the recommendation of the most suitable replacement component. Parameter adjustment operations optimize the internal configuration of components, including parameter space definition (parameter type, value range, and constraints), adjustment step size control, and parameter dependency handling. Multiple parameter representations are supported, such as discrete enumerated values, continuous numerical ranges, and structured configurations. In addition to these four basic operations, composite operations are defined, such as sub-process refactoring (involving coordinated changes to multiple basic operations) and pattern application (applying predefined best practice patterns to the process). Each operation has clear preconditions, execution rules, and post-validation to ensure the integrity and consistency of the process after execution. Semantic information about the operation, such as its intent, expected effects, and risk impact, is also recorded to provide context for subsequent optimization.

[0173] Based on the set of editing operations, a continuous-time Markov chain model describing the evolution of business processes is established, where states represent process configurations and transition rates represent the tendency to execute editing operations per unit time. The continuous-time Markov chain (CTMC) model is an ideal tool for capturing the dynamic evolution characteristics of processes, and a CTMC framework specifically designed for BI business processes is constructed. In this model, the state space consists of all possible process configurations, and each state is a complete description of a process component and its parameters. Considering the high dimensionality of the state space, eigenvector representation is used to map the process configuration to a fixed-dimensional vector space, preserving key structural and parameter information. The transition rate matrix Q is the core of the model, where the element q_{ij} represents the transition tendency from state i to state j, corresponding to the probability rate of executing a specific editing operation. Transition rate functions are designed for different types of editing operations, considering factors such as operation complexity, risk impact, and historical preferences. A key characteristic of the CTMC model is its ability to handle the continuity of time, capturing the temporal dynamics of operation execution through an exponentially distributed dwell time, which is more consistent with the characteristics of actual process evolution than discrete-time models. It also extends the basic CTMC by introducing a semi-Markov property, allowing transition rates to depend on the dwell time of the current state, thus more accurately modeling the temporal correlation of operation sequences. To address the state space explosion problem, a hierarchical modeling strategy is adopted, decomposing the overall process into multiple sub-modules for independent modeling, and then constructing the complete model through combination rules. Furthermore, context-sensitive transition rates are implemented, ensuring that operational tendencies depend not only on the current state but also on external factors (such as changes in business requirements and resource constraints), improving the model's adaptability and predictive accuracy.

[0174] In its implementation, the continuous-time Markov chain model employs a multi-layer neural network architecture. The input layer corresponds to a state representation vector (containing process configuration and auxiliary variable information), the hidden layer uses a three-layer fully connected network with 256 neurons per layer, using the ReLU activation function, and the output layer generates a transition rate matrix. Training data is sourced from the company's internal BI process history repository, containing over 10,000 process evolution sequences covering different business scenarios and optimization objectives. Data preprocessing includes process standardization, unified encoding of editing operations, and time-series alignment. Model training utilizes the Adam optimizer with an initial learning rate of 0.001, employing a learning rate decay strategy, a batch size of 64, and 200 training epochs. To validate model effectiveness, a 5-fold cross-validation method is used, employing perplexity, prediction accuracy, and mean negative log-likelihood (NLL) as evaluation metrics.

[0175] By analyzing historical process versions and edit records, the transition rate matrix parameters of the continuous-time Markov chain model (CTMC) are learned using the maximum likelihood estimation method. Parameter learning is fundamental to model application, and a specialized learning algorithm is designed to extract statistical patterns of process evolution from historical data. First, a training dataset is constructed, collecting process version history and edit logs from multiple BI projects within the enterprise, recording the operation type, execution time, and context information for each change. To enhance data quality, edit log standardization and version difference extraction algorithms are implemented, converting unstructured historical records into standardized operation sequences. Based on this data, the maximum likelihood estimation (MLE) method is applied to learn the transition rate matrix of the CTMC. The core idea is to find parameter values ​​that maximize the probability of observed edit sequences. Specifically, the Expectation-Maximization (EM) algorithm is used to handle hidden states and incomplete observations in the data, iteratively optimizing to gradually improve the accuracy of parameter estimation. To address data sparsity, prior knowledge and regularization techniques are introduced, such as imposing parameter sharing constraints on similar operation types to prevent overfitting and improve generalization ability. An incremental learning mechanism is also implemented, enabling efficient updating of model parameters as new edit records are generated, maintaining model consistency with the latest practices. To evaluate the learning quality, a cross-validation scheme was designed, using a portion of historical data to train the model and another portion to validate prediction accuracy. The model's fit was quantified using metrics such as perplexity. The learning process also included sensitivity analysis to assess the impact of different data subsets and initial conditions on parameter estimation, ensuring the robustness of the results. The final learned transition rate matrix captures the organization's unique process evolution patterns, providing a statistical basis for subsequent operational sequence generation.

[0176] Based on the risk interest points in the risk situation data, a risk-oriented objective function is constructed, with risk scoring as the primary optimization objective. The objective function serves as the guiding principle for generating operation sequences and incorporates a multi-dimensional, risk-sensitive evaluation mechanism. The core component of the function is the risk scoring item, which directly maps to the comprehensive risk score in the risk situation data and has the highest weight. It pays special attention to risk interest points, providing additional rewards for improving risks in high-risk, inflection, sensitive, and anomaly areas, guiding optimization towards addressing key risk issues. The objective function also includes a business value item, assessing the impact of process changes on analysis quality, response time, and resource efficiency. A predictive model of business indicators and process configurations is established to estimate the business impact of specific changes. To balance short-term optimization and long-term stability, the function introduces a change cost item, considering factors such as implementation complexity, learning curve, and compatibility risks. Cost weights are dynamically adjusted based on the organization's change management capabilities to avoid overly aggressive or conservative optimization strategies. The objective function also includes an innovation exploration item, encouraging attempts at new component combinations and parameter configurations to prevent getting trapped in local optima. A novelty scoring mechanism is implemented, comparing the similarity between candidate solutions and historical configurations to reward innovative attempts. To adapt to different business scenarios, the system supports dynamic configuration of the objective function, allowing adjustments to various weights and the addition or removal of specific objective items. It also implements a multi-objective optimization framework capable of generating a Pareto-optimal candidate solution set, demonstrating the trade-offs between risk reduction and other objectives. The objective function design adopts a modular architecture, with each evaluation component implemented independently, facilitating expansion and customization. Furthermore, it provides a visual explanation of the objective function to help users understand the scoring structure and optimization direction.

[0177] A method combining Monte Carlo Tree Search (MCTS) and Markov Chain Model (MCTS) is employed to generate an optimized sequence of editing operations for the business process based on the objective function, resulting in the final intelligent business orchestration scheme. Operation sequence generation is a complex search problem, and this method combines advanced search algorithms and statistical models to achieve efficient solution exploration. MCTS is the core algorithm, which iteratively constructs a search tree through four stages: the selection stage uses the UCB (Upper Confidence Bound) strategy to balance exploration and utilization, selecting the most promising path from the root node; the expansion stage generates new candidate operations based on the transition probabilities of the CTMC model; the simulation stage uses a fast evaluation strategy to predict the long-term effects of the operation sequence; and the backtracking stage updates node statistics to guide subsequent searches. By tightly integrating the CTMC model with MCTS and using a transition rate matrix to guide the operation selection probability, the generated sequence not only conforms to historical evolution patterns but also effectively reduces risk. To improve search efficiency, several optimization techniques were implemented: domain knowledge pruning to eliminate obviously ineffective operation combinations; parallel MCTS to explore multiple search tree branches simultaneously; progressive widening to control the number of nodes expanded at each level, balancing search breadth and depth; and adaptive simulation depth to adjust the simulation length based on node characteristics, improving evaluation accuracy. An operation sequence optimization mechanism was also designed to post-process the initially generated sequences, merging redundant operations, adjusting execution order, and refining parameter settings to improve the simplicity and executability of the solution. To address uncertainty, multiple alternative solutions were generated, and their robustness under different scenarios was evaluated through Monte Carlo simulation, selecting the most stable solution as the final output. The generated intelligent business orchestration solution includes detailed operation steps, execution order, expected effects, and risk impact analysis, which can directly guide manual implementation and support automated execution; it also provides a visual preview of the solution, showing a comparison of the process before and after changes and key improvements, helping users understand and evaluate the optimization solution.

[0178] The edit operation set is the foundation for implementing a non-autoregressive business process generation mechanism. By defining a complete set of basic operations, it enables the flexible construction and adjustment of BI business processes.

[0179] Basic Editing Operation Type Design: The design covers all possible changes to BI business processes. Operation types include three core categories: Insertion (adding new components or steps to the process), Deletion (removing existing components or steps from the process), and Replacement (replacing existing components with new ones while maintaining connections). Each core operation type is further subdivided into several specific operations. For example, Insertion includes component insertion, branch insertion, and parallel process insertion; Deletion includes single component deletion, subprocess deletion, and conditional branch deletion; Replacement includes similar replacement, upgrade replacement, and downgrade replacement. In addition, auxiliary operation types are defined, such as parameter adjustment (modifying component parameters without changing the process structure), connection redirection (changing the connection relationships between components), and reordering (adjusting the execution order of parallel components). Each operation type has a clear semantic definition and execution effect. For example, the "conditional branch insertion" operation adds a condition judgment node and multiple branch paths at a specified location, enabling the process to execute different processing logic based on conditions.

[0180] Continuous-time Markov chain model: Based on the set of edit operations and the representation of process states, a transition rate matrix is ​​constructed to describe the dynamics of state transitions. Unlike traditional discrete-time Markov chains, the continuous-time model uses transition rates rather than transition probabilities to more accurately describe the dynamic characteristics of process evolution. Each possible edit operation corresponds to a transition rate, representing the tendency to execute that operation per unit of time. The transition rate is influenced by various factors, including operation type, current process state, historical evolution patterns, and changes in business requirements. By analyzing historical process versions and edit records, the parameters of the transition rate matrix are learned using maximum likelihood estimation or Bayesian methods. A dynamic adjustment mechanism for the transition rate is implemented, enabling the model to adapt to constantly changing business environments and user preferences. For example, it may learn that adding data cleaning components will increase the transition rate when data quality deteriorates; while in scenarios where users are more focused on real-time performance, simplifying the processing flow will increase the transition rate.

[0181] Editflow Generation Based on Risk Assessment: This algorithm generates a sequence of edit operations to optimize the BI workflow based on risk assessment, an objective function, and a continuous-time Markov chain model. The generation process combines Monte Carlo Tree Search (MCTS) with a Markov chain model, maintaining random exploration capabilities while selectively searching in directions that reduce risk. The algorithm includes: starting from the current workflow state, simulating multiple possible edit paths; evaluating each path, calculating its risk reduction effect and other performance metrics; selecting the optimal path based on the evaluation results and determining the next edit operation; executing the selected operation and updating the workflow state; repeating the above process until a termination condition is met (e.g., risk reduced to an acceptable level or reaching the maximum number of iterations). Parallel search and distributed evaluation are implemented to improve the optimization efficiency of large-scale workflows. For example, to address performance risks caused by data skew, a series of edit operations may be generated, including "adding data distribution analysis," "inserting data sharding processing," and "adjusting parallelism parameters," forming a coherent optimization scheme.

[0182] In a preferred embodiment, learning the transition rate matrix parameters of the continuous-time Markov chain model using the maximum likelihood estimation method includes:

[0183] The design incorporates a set of auxiliary variables, including historical state variables, performance metric variables, risk assessment variables, and business context variables, to expand the business process state space. Recognizing that a purely structural state representation of the process cannot capture all factors influencing editorial decisions, a multi-dimensional set of auxiliary variables is designed to enrich the state representation. Historical state variables record the evolutionary trajectory information of the process, including previous state sequences, state dwell time distribution, and historical transition frequencies. These variables enable the model to capture time dependencies and sequence patterns. Efficient historical encoding methods are implemented, such as using recurrent neural networks (RNNs) or attention mechanisms to compress variable-length histories into fixed-dimensional vector representations. Performance metric variables quantify the operational effectiveness of the process, including key indicators such as response time, throughput, resource utilization, and accuracy. These indicators are collected through online monitoring and offline analysis, and standardization is applied to make indicators of different dimensions comparable. Risk assessment variables are extracted directly from risk situation data, including comprehensive risk scores, risk type distribution, high-risk area locations, and risk trends. These variables enable the model to perceive risk conditions and guide risk-oriented optimization. Business context variables describe the external environment and business requirements, including user types, business scenarios, time window requirements, and data characteristics. These variables are extracted through a business rule engine and a context-aware framework, enabling the model to adapt to different business situations. Feature engineering techniques are used to process these auxiliary variables, including feature selection (removing redundant and low-information features), feature transformation (such as logarithmic transformation and binning), and feature combination (creating interactive features to capture relationships between variables), optimizing the expressive power and computational efficiency of the feature set.

[0184] The set of auxiliary variables is integrated into the state representation using a specific encoding method, and participates in the training of the continuous-time Markov chain model together with the state information of the optimized business process. The design of the state representation is crucial to the model's performance, implementing multiple encoding strategies to balance expressive power and computational complexity. For process structure information, a graph encoding method is used to represent the process as a directed graph, with nodes corresponding to processing components and edges representing data flow directions. Graph embedding techniques (such as Graph2Vec or GraphSAGE) are used to convert the graph structure into a fixed-dimensional vector representation, preserving topological characteristics and component relationships. For component parameters, a hierarchical encoding scheme is designed, with different types of parameters (such as numerical, categorical, textual, and structured configurations) using specialized encoding methods, and then aggregated through an attention mechanism to form a unified representation. The integration of auxiliary variables adopts multimodal fusion technology. First, various variables are encoded within their respective modalities, such as using recurrent networks to process temporal data and convolutional networks to process spatial data. Then, cross-modal attention mechanisms or multi-head self-attention networks are used to achieve the interaction and fusion of information from different modalities, generating a comprehensive representation. To handle high-dimensional state spaces, an autoencoder architecture was implemented. This architecture discovers low-dimensional representations of states through unsupervised learning, reducing computational complexity while preserving key information. An incremental encoding mechanism was also designed, updating only the relevant parts of the state representation when the process undergoes minor changes, avoiding complete recalculation and improving training efficiency. The state encoding process supports interpretable design, allowing analysis of the structure and key dimensions of the encoded vectors to help understand the basis of model decisions. Through this comprehensive and efficient state representation, the continuous-time Markov chain model can accurately capture various factors influencing editing decisions, improving the accuracy of transition rate prediction.

[0185] This study collects successful and unsuccessful edit flow case pairs to implement a contrastive learning-based training framework. This framework learns to distinguish features by comparing edit sequences with effective and ineffective optimizations. The contrastive learning framework is an innovative method for model training, improving the model's discriminative ability by constructing positive and negative sample pairs. First, a case library is established, collecting edit sequences and their effect evaluations from historical projects. Cases are categorized into successful and unsuccessful categories based on predefined success criteria (such as risk reduction, performance improvement, and user satisfaction). To ensure data quality, a case screening mechanism is implemented to exclude incomplete records and abnormal cases, and data augmentation techniques are applied to expand the samples for scarce categories. Within the contrastive learning framework, three types of sample pairs are constructed: anchor-positive pairs (two successful optimizations in the same process), anchor-negative pairs (a comparison of successful and unsuccessful optimizations), and hard pairs (successful and unsuccessful cases with minimal difference in effect). The training process employs a contrastive loss function, such as InfoNCE or triplet loss, aiming to maximize the similarity of positive pairs while minimizing the similarity of negative pairs. Multiple similarity calculation methods were implemented, including cosine similarity, Mahalanobis distance, and learned similarity functions, automatically selecting the most suitable metric based on data characteristics. To improve training efficiency, a hard example mining strategy was adopted, focusing on case pairs that the model struggles to distinguish and dynamically adjusting sample weights. Contrastive learning not only improved the model's discriminative ability but also helped discover key distinguishing features between successful and unsuccessful editing sequences. These features were explicitly extracted and used to guide subsequent editing decisions. Cross-project transfer learning was also implemented, leveraging successful experiences from different business domains to improve the model's adaptability in new scenarios. Through the contrastive learning framework, the continuous-time Markov chain model can more accurately evaluate the potential effects of different editing operations and generate more effective optimized sequences.

[0186] The design incorporates a comprehensive reward function encompassing dimensions of risk reduction, performance improvement, resource efficiency, and business value. An integrated reinforcement learning framework enables the continuous-time Markov chain model to improve editing strategies through interaction with the environment. Combining the continuous-time Markov chain model with the reinforcement learning framework achieves feedback-based adaptive optimization. The reward function is the core of reinforcement learning, and a multi-dimensional, configurable evaluation mechanism is designed. The risk reduction dimension directly maps changes in risk profile, calculating the difference in risk scores before and after editing operations, and assigning higher weight to improvements in high-risk areas; risk sensitivity analysis is implemented to identify the improvement effects on different types of risks. The performance improvement dimension evaluates changes in metrics such as response time, query latency, and processing throughput, using relative improvement rates rather than absolute values ​​to ensure fairness in the evaluation; it supports scenario-aware performance evaluation, setting different performance targets for different load conditions. The resource efficiency dimension measures the utilization of computing resources, storage space, and network bandwidth, rewarding optimizations that reduce resource consumption while maintaining functionality; it implements resource-function balance evaluation to prevent excessive sacrifice of functionality for resource conservation. The accuracy, completeness, timeliness, and interpretability of the business value dimension analysis results support business decisions. A business impact model transforms technical indicators into business value assessments, such as decision support effectiveness and user satisfaction. The reward function integrates scores from various dimensions using a weighted summation method, with weights dynamically adjusted based on business priorities. A non-linear reward mapping is also implemented, providing superlinear rewards for breakthrough improvements to encourage innovative optimization. The reinforcement learning framework employs a policy gradient method, treating the transition rate parameters of the continuous-time Markov chain model as policy parameters, optimizing the expected cumulative reward through gradient ascent. A baseline function is implemented to reduce the variance of gradient estimation and improve training stability. To address long-term dependency issues, temporal difference learning and experience trajectory replay are used to effectively capture delayed reward signals. A curiosity-driven exploration mechanism is also implemented to encourage trying new editing strategies and avoid getting trapped in local optima.

[0187] An experience replay buffer stores editing operations and their effects. Incremental learning is implemented to enable the continuous-time Markov chain model to continuously acquire new knowledge during operation, resulting in a fully trained continuous-time Markov chain model. A continuous learning mechanism is designed to allow the model to continuously evolve from practical applications. The experience replay buffer is the core component, storing complete context information of editing operations, including the initial state, operation sequence, intermediate states, final state, and effect evaluation. A hierarchical storage architecture is adopted, with hot data stored in memory for immediate access and cold data migrated to persistent storage to ensure long-term availability. Buffer management uses a priority sampling strategy, allocating storage priorities based on the novelty, scarcity, and information content of samples. An automatic expiration mechanism is implemented to gradually eliminate outdated experience data, maintaining the timeliness of knowledge. The incremental learning framework supports multiple update modes: batch updates periodically retrain the model using accumulated new data; online updates immediately adjust model parameters when new experiences are generated; and hybrid updates combine the advantages of both methods to balance learning efficiency and model stability. To prevent catastrophic forgetting (new knowledge overwriting old knowledge), techniques such as elastic weight merging, knowledge distillation, and experience replay are implemented to maintain the model's adaptability to historical scenarios. A knowledge transfer mechanism was also designed to generalize optimization strategies learned in specific scenarios to similar new scenarios, accelerating the learning process. Model evaluation employs a sliding window method, continuously monitoring the model's performance on the latest data. Deep retraining is triggered when performance drops below a threshold. An A / B testing framework was also implemented to safely verify the effectiveness of model updates in a production environment. Through this closed-loop continuous learning mechanism, the continuous-time Markov chain model can adapt to dynamic changes in the business environment, continuously improving the effectiveness of editing strategies and forming truly intelligent business orchestration. The trained model not only includes a parameterized transition rate matrix but also a state representation encoder, reward evaluator, and policy optimizer, forming a complete decision-making process capable of generating the optimal sequence of editing operations based on the current business state and risk profile.

[0188] Extended state-space training is a key technique for improving the performance of edit flow models. By introducing auxiliary variables and advanced learning strategies, it enables more efficient learning and generation of optimal editing strategies.

[0189] Auxiliary Variable Design and Integration: A set of auxiliary variables is designed to expand the BI process state space and enhance the model's expressive power and learning efficiency. Auxiliary variables include historical state variables (recording the process's historical states and editing history), performance indicator variables (tracking the performance of various aspects of the process), risk assessment variables (representing different types of risk levels), and business context variables (describing current business needs and environment). These auxiliary variables do not directly represent the process state but provide important contextual information, helping the model understand the motivation and effects of state transitions. Auxiliary variables are integrated into the state representation through specific encoding methods, participating in model training and inference together with the core state variables. For example, historical state variables may record the most recent N editing operations, helping the model capture editing patterns and trends; risk assessment variables may contain scores for various risk dimensions, guiding the model to search in directions that reduce risk.

[0190] Contrastive Learning Framework: This framework implements a contrastive learning-based training framework to improve the efficiency of model learning from historical optimization cases. Contrastive learning learns to distinguish key features that differentiate between effective and ineffective optimizations by comparing "good" and "bad" edit flow examples. Specific methods include: collecting pairs of successful and unsuccessful edit flow cases; successful cases are edit sequences that significantly reduce risk or improve performance, while unsuccessful cases are ineffective or harmful edit sequences; designing a contrastive loss function to enable the model to increase the probability of generating successful cases and decrease the probability of generating unsuccessful cases; and implementing hard case mining, focusing on easily confused boundary cases to improve the model's discriminative ability. Contrastive learning is particularly suitable for handling sequential decision-making problems like edit flows, effectively capturing the impact of local edit decisions on the overall process quality. For example, the model can learn that adding a data validation step after data cleaning is generally beneficial (positive example), while directly deleting anomaly detection steps after critical processing nodes is generally harmful (negative example).

[0191] Reinforcement Learning and Reward Design: Integrating a reinforcement learning framework enables the model to continuously improve its editing strategy through interaction with the environment. Editing flow generation is modeled as a Markov Decision Process (MDP), where the state is the current process configuration, actions are possible editing operations, and rewards represent changes in process performance and risk after the operation. A comprehensive reward function is designed, covering multiple evaluation dimensions, including risk reduction (improvement of various risk scores), performance enhancement (optimization of metrics such as accuracy and response time), resource efficiency (saving computational and storage resources), and business value (degree of support for business decisions). A long-term reward mechanism is implemented, focusing not only on the effect of a single editing step but also on the cumulative benefits of the entire editing sequence. Algorithms such as policy gradient, Q-learning, or Monte Carlo tree search are used to train the editing policy network. For example, adding an editing operation to check data quality may result in a short-term increase in processing latency (negative immediate reward), but in the long run, it improves the reliability of the analysis results (high cumulative reward), and the reinforcement learning framework can correctly evaluate this long-term value.

[0192] Incremental Learning and Continuous Optimization: An incremental learning mechanism is designed to enable the model to continuously acquire new knowledge from runtime, thereby continuously improving its editflow generation capabilities. An experience replay buffer is implemented to store observed editing operations and their effects during runtime, serving as training data for model updates. A priority sampling strategy is designed to focus on learning novel, unexpected, or high-impact cases, improving learning efficiency. A periodic model update mechanism is implemented to adjust model parameters based on accumulated new experience, adapting to changes in the business environment. A model version management and evaluation framework is established to monitor model performance changes and ensure that updates do not lead to performance degradation. Model rollback and A / B testing are supported to safely verify the effectiveness of new models. For example, when a new data anomaly pattern is observed, relevant cases and handling methods are recorded, and this new knowledge is integrated into the editflow model through incremental learning to improve its ability to handle similar situations.

[0193] The intelligent BI business orchestration system comprises five layers: infrastructure layer, data service layer, agent collaboration layer, business orchestration layer, and user interaction layer, realizing intelligent orchestration of the entire process from data collection to visualization.

[0194] Please refer to Figure 5 , Figure 5 This is a structural diagram of an intelligent situational analysis and orchestration system provided in an embodiment of the present invention. Figure 5 As shown, the system includes:

[0195] The requirement parsing module 701 is used to obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range and constraints.

[0196] The agent collaboration module 702 is used to construct a four-layer agent neural network collaboration framework, including data layer agent, processing layer agent, analysis layer agent and presentation layer agent, based on the core elements of the business requirements, and to encapsulate corresponding core capability modules for each layer agent.

[0197] The task decomposition module 703 is used to perform forward task decomposition on the business objectives, decompose complex business requirements into sub-tasks layer by layer and assign them to corresponding agents, and generate an initial task execution network.

[0198] The process optimization module 704 is used to adjust the agent collaboration method based on the execution result of the initial task execution network using a reverse collaboration optimization algorithm to form an optimized business process;

[0199] The risk assessment module 705 is used to construct a two-parameter risk model containing model parameter vectors and data parameter vectors, to assess the risks of the optimized business process, identify risk interest points, and generate risk situation data.

[0200] The orchestration generation module 706 is used to generate an editing operation sequence based on the risk situation data, and to perform non-autoregressive adjustments on the optimized business process through a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme.

[0201] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described intelligent situational analysis and orchestration method.

[0202] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described intelligent situational analysis and orchestration method.

[0203] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for intelligent situational analysis and orchestration, characterized in that, Includes the following steps: Obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range, and constraints. Based on the core elements of the aforementioned business requirements, a four-layer agent neural network collaboration framework is constructed, comprising data layer agent, processing layer agent, analysis layer agent, and presentation layer agent, and corresponding core capability modules are encapsulated for each layer agent. The business objectives are decomposed into forward tasks, and complex business requirements are broken down into sub-tasks layer by layer and assigned to corresponding agents to generate an initial task execution network. Based on the execution results of the initial task execution network, a reverse collaboration optimization algorithm is used to adjust the agent collaboration method to form an optimized business process; A two-parameter risk model containing model parameter vectors and data parameter vectors is constructed to conduct risk assessment on the optimized business process, identify risk interest points, and generate risk situation data. Based on the risk situation data, an editing operation sequence is generated. The optimized business process is then adjusted non-autoregressively using a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme. The step of generating an editing operation sequence based on the risk situation data includes: Based on the aforementioned risk situation data, a set of editing operations applicable to business processes is defined, including basic operations such as component insertion, deletion, replacement, and parameter adjustment. Based on the set of editing operations, a continuous-time Markov chain model describing the evolution of the business process is established, where the state represents the process configuration and the transition rate represents the tendency to execute editing operations per unit time. By analyzing historical process versions and edit records, the transition rate matrix parameters of the continuous-time Markov chain model are learned using the maximum likelihood estimation method. Based on the risk interest points in the risk situation data, a risk-oriented objective function is constructed, with risk scoring as the main optimization objective. By employing a method combining Monte Carlo tree search and Markov chain model, an optimized sequence of editing operations for the business process is generated based on the objective function, resulting in the final intelligent business orchestration scheme.

2. The method according to claim 1, characterized in that, The construction of a four-layer agent neural network collaboration framework, comprising data layer agent, processing layer agent, analysis layer agent, and presentation layer agent, includes: Based on the core elements of the aforementioned business requirements, a hierarchical structure of data layer proxy, processing layer proxy, analysis layer proxy, and presentation layer proxy is defined to form a complete business process collaboration network. The data layer proxy encapsulates data connectivity, data extraction, and data caching capabilities; the processing layer proxy encapsulates data cleaning, transformation, and quality control capabilities; the analysis layer proxy encapsulates model training, reasoning, and result interpretation capabilities; and the presentation layer proxy encapsulates data visualization, interaction design, and narrative expression capabilities. Based on the core capability modules of the data layer agent, the processing layer agent, the analysis layer agent, and the presentation layer agent, an inter-agent communication protocol is constructed to support horizontal collaboration between agents in the same layer and vertical collaboration between agents across layers, thus obtaining the four-layer agent neural network collaboration framework.

3. The method according to claim 2, characterized in that, The construction of the inter-agent communication protocol includes: Based on the core capability modules of the data layer proxy, the processing layer proxy, the analysis layer proxy, and the presentation layer proxy, a unified message structure is designed that includes sending proxy identifier, receiving proxy identifier, message type, priority, and timestamp. The unified message structure can implement multiple communication modes, including request-response mode, publish-subscribe mode, broadcast mode, and point-to-point direct communication mode. Based on the aforementioned multiple communication methods, a standard protocol for proxy collaboration is defined to standardize the processing flow of collaboration request initiation, collaboration invitation response, task division negotiation, and execution status synchronization. By establishing data flow processing between agents through the aforementioned standard protocol, it supports the block transmission, incremental synchronization and streaming processing of large-scale data, resulting in a communication protocol that supports efficient collaboration between agents.

4. The method according to claim 1, characterized in that, The forward task decomposition of the business objectives includes: The business objectives are semantically parsed to extract core elements including the analysis object, time range, analysis type, and business constraints, and then converted into a structured target representation. Based on the structured target representation, a top-down recursive decomposition algorithm is used to gradually decompose the high-level target into multiple interrelated sub-targets. The multiple interrelated sub-goals are further decomposed until they reach a granularity that can be directly assigned to a specific agent for execution, thus forming a subtask specification; Based on the logical relationships and data flow between the sub-task specifications, a task dependency graph is constructed. Based on the task dependency graph and the available proxy capabilities in the system, the most suitable proxy or proxy team is selected for each subtask to generate the initial task execution network.

5. The method according to claim 4, characterized in that, The process of adjusting the agent collaboration method based on the execution results of the initial task execution network using a reverse collaborative optimization algorithm includes: The execution results of the initial task execution network are evaluated from multiple dimensions, including result quality, execution efficiency, collaborative performance, and user satisfaction. Based on the multi-dimensional task evaluation results, the reasons for poor task performance are analyzed, and key improvement points are identified, including data quality issues, inappropriate algorithm selection, suboptimal parameter configuration, or poor collaboration. For the aforementioned key improvement points, generate specific optimization signals that include problem descriptions, severity, priority, and recommended measures; The specific optimization signal is transmitted from the downstream task to the upstream task through the reverse propagation algorithm, forming a complete optimization chain; Based on the specific optimization signals, each agent is guided to adjust its behavior and strategy, while the global collaboration mode between agents is optimized to obtain the optimized business process.

6. The method according to claim 1, characterized in that, The construction of a two-parameter risk model comprising a model parameter vector and a data parameter vector includes: Based on the optimized business process, define model parameter vectors that describe the analysis method, algorithm selection, and parameter configuration, as well as data parameter vectors that describe data quality, data distribution, and sample size. Construct a risk assessment function, map the model parameter vector and data parameter vector to a multidimensional risk space, calculate basic risk indicators and aggregate them to form a comprehensive risk score; Based on the comprehensive risk score, a risk situation calculation engine is developed to quantify the risk of each link in the optimized business process and obtain risk quantification data. The risk interest point identification algorithm automatically discovers high-risk points, inflection points, sensitive points and anomalies in the risk quantification data to obtain risk interest points; Based on the risk points of interest and the comprehensive risk score, risk situation data containing risk heatmap data and risk level information is generated.

7. The method according to claim 6, characterized in that, The risk interest point identification algorithm automatically identifies high-risk points, inflection points, sensitive points, and anomalies in the risk quantification data to obtain risk interest points, including: Threshold detection and regional connectivity analysis are used to identify high-risk points whose risk scores consistently exceed the threshold. Gradient analysis and curvature calculation are used to identify inflection points where risk surfaces change drastically; Local sensitivity analysis is used to calculate the degree of risk response to parameter changes and to identify sensitive points; Local anomaly detection and spatial statistical analysis were used to identify anomalous points that were significantly different from the surrounding risk patterns. Feature extraction and attribute analysis are performed on the high-risk points, inflection points, sensitive points and abnormal points to obtain feature analysis results including location features, intensity features, morphological features and dynamic features; Based on the feature analysis results, the risk impact and business importance are calculated, the risk interest points are prioritized, and risk situation display data containing key risk area information is generated.

8. The method according to claim 1, characterized in that, The process of learning the transition rate matrix parameters of the continuous-time Markov chain model using the maximum likelihood estimation method includes: The design includes a set of auxiliary variables, including historical state variables, performance indicator variables, risk assessment variables, and business context variables, to expand the business process state space. The set of auxiliary variables is integrated into the state representation through a specific encoding method, and together with the state information of the optimized business process, they participate in the training of the continuous-time Markov chain model. Collect successful and unsuccessful edit flow case pairs to implement a training framework based on contrastive learning, and learn to distinguish features by comparing edit sequences with effective and ineffective optimizations; The design incorporates a comprehensive reward function that includes dimensions of risk reduction, performance improvement, resource efficiency, and business value, and integrates a reinforcement learning framework to enable the continuous-time Markov chain model to improve editing strategies through interaction with the environment. The experience replay buffer stores the editing operations and their effects, and an incremental learning process is established to enable the continuous-time Markov chain model to continuously acquire new knowledge from the operation, thus obtaining a trained continuous-time Markov chain model.

9. A business intelligence orchestration system based on artificial intelligence, characterized in that, include: The requirement parsing module is used to obtain the business analysis requirements input by the user, perform semantic parsing and intent recognition on the business analysis requirements, and extract the core elements of the business requirements, including business objectives, analysis objects, time range, and constraints. The agent collaboration module is used to construct a four-layer agent neural network collaboration framework, including data layer agent, processing layer agent, analysis layer agent and presentation layer agent, based on the core elements of the business requirements, and to encapsulate the corresponding core capability modules for each layer agent. The task decomposition module is used to perform forward task decomposition on the business objectives, decompose complex business requirements into sub-tasks layer by layer and assign them to corresponding agents, and generate an initial task execution network. The process optimization module is used to adjust the agent collaboration method based on the execution result of the initial task execution network using a reverse collaboration optimization algorithm to form an optimized business process; The risk assessment module is used to construct a two-parameter risk model containing model parameter vectors and data parameter vectors, to assess the risks of the optimized business process, identify risk interest points, and generate risk situation data. The orchestration generation module is used to generate an editing operation sequence based on the risk situation data, and to perform non-autoregressive adjustments on the optimized business process through a continuous-time Markov chain model to obtain the final intelligent business orchestration scheme. The step of generating an editing operation sequence based on the risk situation data includes: Based on the aforementioned risk situation data, a set of editing operations applicable to business processes is defined, including basic operations such as component insertion, deletion, replacement, and parameter adjustment. Based on the set of editing operations, a continuous-time Markov chain model describing the evolution of the business process is established, where the state represents the process configuration and the transition rate represents the tendency to execute editing operations per unit time. By analyzing historical process versions and edit records, the transition rate matrix parameters of the continuous-time Markov chain model are learned using the maximum likelihood estimation method. Based on the risk interest points in the risk situation data, a risk-oriented objective function is constructed, with risk scoring as the main optimization objective. By employing a method combining Monte Carlo tree search and Markov chain model, an optimized sequence of editing operations for the business process is generated based on the objective function, resulting in the final intelligent business orchestration scheme.

Images