Complex business process reconfiguration method based on data integration and multi-agent collaboration

Abstract

The application discloses a complex business process reconstruction method based on data integration and multi-agent cooperation, relates to the field of business process management, and comprises the following steps: a multi-source heterogeneous data integration layer is constructed to collect and integrate data of each node of a business process in real time, so that a data graph with consistent semantics and complete events is formed; a multi-agent system based on five types of functional agents and a three-layer cooperative architecture is used to analyze and reconstruct the process by means of an event-driven and mixed cooperation mechanism; relying on the data graph, an agent converts an event sequence into a weighted directed graph to perform dynamic diagnosis, and combines multi-dimensional indexes and root cause analysis to locate a process bottleneck; and an optimization decision agent organizes and cooperates to propose a reconstruction scheme, and realizes closed-loop self-adaptive dynamic optimization through reinforcement learning. The application has the advantages that a standardized business process data graph is formed through a data integration layer, a multi-agent system is matched, full-dimensional accurate diagnosis, cooperative reconstruction optimization and closed-loop self-adaptive optimization of a complex business process are realized.

Description

Technical Field

[0001] This invention relates to the field of business process management, and in particular to a method for refactoring complex business processes based on data integration and multi-agent collaboration. Background Technology

[0002] As digital transformation deepens, enterprises face the challenge of increasingly complex and fragmented business processes. Traditional monolithic systems struggle to adapt to dynamically changing market demands, resulting in data silos, slow response times, and low collaboration efficiency. Against this backdrop, business process reengineering methods based on data integration and multi-agent collaboration have emerged.

[0003] Current mainstream methods for complex business process refactoring have significant shortcomings and limitations. First, they lack data support capabilities. Most methods only collect fragmented business data, lacking a unified multi-source heterogeneous data semantic integration and end-to-end quality control system. They fail to construct complete event sequences and consistent entity relationships in the business process data graph, resulting in data silos, semantic conflicts, and data gaps in the process links, thus failing to provide accurate and complete multi-dimensional data support for process analysis. Second, their collaborative decision-making mechanisms are rigid, often employing centralized, static manual modeling and optimization modes. They lack a distributed, multi-agent, layered, hybrid collaborative architecture, making it difficult to balance real-time event response, cross-node resource negotiation, and global consensus on refactoring goals. Cross-stage collaboration is inefficient and cannot adapt to the dynamic changes in complex business needs. Third, they lack long-term adaptive optimization capabilities. Most are one-off process refactorings, failing to form a closed-loop reinforcement learning mechanism of diagnosis-refactoring-optimization. They cannot dynamically iterate and optimize based on operational data, resulting in severely insufficient scenario adaptability and long-term optimization effectiveness of the refactoring solutions. Summary of the Invention

[0004] To improve existing methods, this paper provides a method for refactoring complex business processes based on data integration and multi-agent collaboration. This method forms a standardized business process data map by constructing a multi-source heterogeneous data integration layer, and is paired with a multi-agent system with a three-layer hybrid collaborative architecture. This enables accurate diagnosis of complex business processes in all dimensions, collaborative refactoring optimization and closed-loop adaptive tuning, and efficiently solves the core pain points of high data barriers, low collaborative efficiency and insufficient dynamic adaptability in traditional process refactoring.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] Methods for refactoring complex business processes based on data integration and multi-agent collaboration include:

[0007] Construct a multi-source heterogeneous data integration layer for complex business processes, and perform real-time data collection, semantic mapping, cleaning and quality assessment on data from each node of the business process to form a business process data map with complete event sequences and consistent entity associations.

[0008] Design a multi-agent system based on a hybrid collaborative mechanism. The system defines and organizes five types of functional agents: process awareness, task decomposition, resource negotiation, execution monitoring, and optimization decision-making. The agents collaborate through a three-layer collaborative architecture of event-driven, contract network protocol, and joint intent. The system analyzes and reconstructs business processes based on private and shared knowledge bases.

[0009] Based on business process data graphs, intelligent agents collaboratively perform data-driven dynamic analysis and diagnosis of business processes, transforming event sequences into weighted directed graph models to identify structural features, extracting and associating multi-dimensional performance indicators of time, resources, and quality, and applying a root cause analysis engine to locate key activities, data interfaces, or resource contention points that lead to process bottlenecks or defects.

[0010] The multi-agent collaborative process refactoring and dynamic optimization is initiated by the optimization decision agent. The organizing agent, through a collaborative process of multiple rounds of proposal feedback, generates and optimizes the refactoring plan for the diagnosed problems in response to the refactoring goal. After reaching a consensus, an executable plan is formed that includes activity adjustment, resource migration, rule update and interface adaptation. During operation, closed-loop adaptive optimization is achieved through reinforcement learning mechanism.

[0011] Preferably, the construction of a multi-source heterogeneous data integration layer for complex business processes specifically includes:

[0012] By deploying adapters at each execution node of the business process, structured business data, unstructured document data, and system log stream data are collected in real time. The adapters use log-based incremental capture technology for databases, an event listening mechanism for file systems, and a long-connection callback mode for application interfaces to achieve low-latency data synchronization.

[0013] Establish a unified business semantic model, perform semantic annotation and mapping on the collected raw data, formally define business entities, events and relationships using domain ontology, map fields and terms in heterogeneous data to concepts in ontology through semantic parsing, and generate a unified metadata description containing timestamps, data sources and semantic types.

[0014] The design incorporates a three-layer cleaning rule: the syntax layer uses regular expressions and format templates to repair data format anomalies; the logic layer uses predefined business rules and constraints to detect data consistency; and the semantic layer uses ontology reasoning to identify and resolve conceptual conflicts and relational contradictions.

[0015] The cleaned data is aggregated by subject domain to form a business process data graph with complete event sequences and consistent entity links.

[0016] Preferably, the establishment of a unified business semantic model specifically includes:

[0017] The domain ontology is constructed using the OWL language, defining the core classes "business activities", "business objects", "participants", and "rules" as well as their object attributes and data attributes;

[0018] Based on the constructed domain ontology, semantic annotation tools are used to map terms from different data sources to standard concepts in the ontology through synonym expansion, context matching, and manual verification. The mapping confidence is recorded to form a business semantic knowledge base stored in the form of an RDF graph.

[0019] Preferably, the multi-agent system based on a hybrid cooperative mechanism specifically includes:

[0020] Five types of functional agents are defined: the process-aware agent identifies the start, activity switching, and end states of process instances by subscribing to event streams in the data integration layer; the task decomposition agent receives process fragments and decomposes them into atomic task sequences based on resource constraints and business rules; the resource negotiation agent manages a virtual resource pool including human, equipment, and computing resources, and allocates appropriate resource instances to tasks; the execution monitoring agent tracks the status of atomic tasks and collects performance indicators and abnormal events; and the optimization decision-making agent evaluates the overall efficiency of the process and generates refactoring suggestions.

[0021] A layered and hybrid collaborative architecture is adopted. The bottom layer is reactive collaboration, where agents publish and subscribe to key state changes through a shared event bus. The middle layer is contract network protocol collaboration, which negotiates resource allocation and task delegation among agents through bidding, tendering, and winning bid mechanisms. The top layer is collaboration based on joint intent, initiated by the optimization decision agent, which organizes agents to make multiple rounds of proposals and commitments around the process refactoring goal to form a consensus solution.

[0022] A two-layer knowledge base is designed for intelligent agents. The private knowledge base stores rules, cases, and experience data in the area of ​​expertise, while the shared knowledge base stores global constraints, historical performance patterns, and collaboration protocol templates extracted from the business process data graph.

[0023] Preferably, the shared event bus adopts a publish-subscribe model, defining a set of standard event types and message formats; agents subscribe to corresponding events according to their responsibilities, and when there is a state update or a key event in the business process data graph, the data integration layer or the sensing agent publishes the event message, triggering the immediate response of the subscriber agents.

[0024] Preferably, the dynamic analysis and diagnosis of business processes driven by data collaboratively executed by intelligent agents specifically includes:

[0025] The process-aware intelligent agent transforms the event sequence in the data graph into a weighted directed graph model, where nodes represent business activities, edges represent the flow relationships between activities, and weights are determined by the historical execution frequency and the strength of data dependencies; it also identifies high-frequency paths, loop structures, and redundant nodes through graph mining algorithms.

[0026] The execution monitoring agent extracts three types of performance metrics from the event sequence: time metrics, including activity duration, workflow interval, and process cycle; resource metrics, including resource utilization, load balancing, and number of conflicts; and quality metrics, including first-pass yield, error rate, and rework rate. These metrics are associated with process nodes in real time.

[0027] The root cause analysis engine for the optimized decision-making agent is used to bind abnormal performance indicators to specific nodes, edges, or resources in the process model. The engine uses pattern matching based on association rules and classification based on decision trees to locate key business activities, data interfaces, or resource contention points that lead to bottlenecks or quality defects, and generates diagnostic reports.

[0028] Preferably, the identification of high-frequency paths, cyclic structures, and redundant nodes using graph mining algorithms specifically includes:

[0029] The frequent subgraph mining algorithm is used to find activity sequences whose execution frequency exceeds a threshold; the loop detection algorithm is used to identify explicit or implicit loops in the flowchart that cause delays; and by analyzing the data dependencies and functional similarities between activities, a set of activities with highly overlapping inputs and outputs and fixed execution order is identified as potential candidate nodes for merging.

[0030] Preferably, the process refactoring and dynamic optimization of multi-agent collaboration specifically includes:

[0031] The optimization decision-making agent initializes reconstruction goals based on the diagnostic report, including shortening the cycle, improving quality, or reducing costs; based on the goals, it retrieves similar historical cases from the shared knowledge base and extracts candidate reconstruction patterns, including activity merging, parallelization transformation, resource reallocation, or rule simplification.

[0032] The optimization decision-making agent broadcasts the reconstruction goal and candidate modes to relevant agents; the task decomposition agent evaluates the impact of the modes on the task sequence and simulates the decomposed atomic task chain; the resource negotiation agent evaluates the constraints of the resource pool and calculates the resource reallocation scheme; the process awareness agent predicts the data flow and event sequence under the new mode; and each agent gradually revises the reconstruction scheme through multiple rounds of proposal feedback loops.

[0033] Once a consensus is reached, an executable refactoring plan is generated, which includes specific activity adjustment steps, resource migration instructions, business rule update scripts, and data interface adaptation strategies; the resource negotiation agent executes resource reservation and allocation according to the plan; the process awareness agent updates the process model and sends model change notifications to relevant business systems.

[0034] During the reconstructed process, the execution monitoring agent initiates a reinforcement learning mechanism to collect real-time performance data of the new process and compare it with the expected target. If the deviation continues to exceed the tolerance range, a new round of collaborative diagnosis and local optimization is triggered, forming a closed-loop adaptive optimization cycle.

[0035] Preferably, the multi-round proposal feedback loop specifically includes:

[0036] The optimization decision-making agent proposes an initial reconstruction plan; the task decomposition and resource negotiation agents evaluate the plan from the perspectives of task logical feasibility and resource availability, and return modification suggestions or constraint conflicts; the optimization decision-making agent integrates feedback and adjusts the plan using constraint satisfaction problem-solving techniques; the loop terminates after a pre-set maximum number of rounds or until all agents accept the plan.

[0037] Preferably, the reinforcement learning mechanism specifically includes:

[0038] Set a target value range for each performance indicator; the execution monitoring agent continuously collects the actual values ​​and calculates the gap with the target range; when the gap continues to exceed the preset period, the current process status and performance gap are used as status inputs to call the optimization decision agent to start a small-scale collaborative optimization; the optimization actions and their effects are recorded and used to update the agent's private experience knowledge base.

[0039] Compared with the prior art, the advantages of the present invention are:

[0040] By integrating multi-source heterogeneous data and constructing a business process data graph, real-time collection, semantic unification, and multi-dimensional quality control of business data across the entire chain are achieved, completely breaking down data silos and providing complete and consistent entity associations and event sequences to support process analysis. A three-layer hybrid collaborative architecture with five types of functional intelligent agents is designed, balancing real-time response, distributed resource negotiation, and global goal consensus. A two-layer knowledge base enables efficient collaboration between professional capabilities and global constraints, overcoming the rigidity of traditional centralized decision-making. Data-driven, multi-dimensional process diagnosis can accurately pinpoint the root causes of bottlenecks. Combined with multi-agent collaborative solution iteration and reinforcement learning closed-loop optimization, dynamic adaptation and continuous iteration of process refactoring are achieved, significantly improving the optimization efficiency and operational resilience of complex business processes. Attached Figure Description

[0041] Figure 1This is a schematic diagram of the complex business process refactoring method based on data integration and multi-agent collaboration proposed in this invention. Detailed Implementation

[0042] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0043] See Figure 1 As shown, the method for refactoring complex business processes based on data integration and multi-agent collaboration includes:

[0044] Step 1: Construct a multi-source heterogeneous data integration layer for complex business processes, and collect, semantically map, clean and assess the quality of data from each node of the business process in real time to form a business process data map with complete event sequences and consistent entity associations;

[0045] Step 2: Design a multi-agent system based on a hybrid collaborative mechanism. The system defines and organizes five types of functional agents: process awareness, task decomposition, resource negotiation, execution monitoring, and optimization decision-making. The agents collaborate through a three-layer collaborative architecture of event-driven, contract network protocol, and joint intent. The system analyzes and reconstructs business processes based on private and shared knowledge bases.

[0046] Step 3: Based on the business process data graph, intelligent agents collaboratively perform data-driven dynamic analysis and diagnosis of business processes, transforming event sequences into weighted directed graph models to identify structural features, extracting and associating multi-dimensional performance indicators of time, resources, and quality, and applying a root cause analysis engine to locate key activities, data interfaces, or resource contention points that lead to process bottlenecks or defects.

[0047] Step 4: Multi-agent collaborative process restructuring and dynamic optimization is initiated by the optimization decision-making agent. The organizing agents, through a collaborative process of multiple rounds of proposal feedback, generate and optimize restructuring solutions for the diagnosed problems in response to the restructuring goals. After reaching a consensus, an executable plan is formed that includes activity adjustments, resource migration, rule updates, and interface adaptation. During operation, closed-loop adaptive optimization is achieved through a reinforcement learning mechanism.

[0048] In step one, for structured data sources such as relational databases, the adapter employs log-based incremental capture technology, such as parsing MySQL's binlog or Oracle's redo log, to monitor changes in transaction logs in real time, capturing and transmitting only data records with added, deleted, or modified operations and their changed content. For unstructured or semi-structured document data stored in the file system, the adapter uses an event listening mechanism to monitor the creation, modification, and deletion events of files in a specified directory, and immediately triggers the reading of file content and extraction of metadata. For services or streaming data sources that provide application programming interfaces (APIs), the adapter uses a long-connection callback mode, such as WebSocket connections or persistent subscriptions to message queues, to maintain a stable session with the server and receive structured event message payloads in real time when business events occur. All raw data collected by the adapter is appended with basic metadata such as source identifiers and collection timestamps, and uniformly transmitted to the buffer queue of the data integration layer.

[0049] Construct a formal business semantic model, using OWL to define a domain ontology closely related to business processes. This ontology explicitly defines core business entities, business events, and the relationships between them. The semantic mapping engine leverages this ontology to automatically semantically annotate raw data fields and business terms from different data sources. For a given data item d, the overall confidence level of its mapping to ontology concept c is calculated. Calculated using the following formula:

[0050]

[0051] in, The textual similarity between the label of data item d and the standard term or synonym of concept c; This indicates the degree of matching between the context information of data item d and the common contexts defined for concept c in the ontology; It is the prior probability of the historical mapping of concept c; , , The weighting coefficients are adjustable, satisfying α+β+γ=1. The mapping results are associated with the original data, forming a unified, machine-understandable semantic metadata description.

[0052] The design employs a three-tiered, progressive data cleaning system: Syntax-level cleaning utilizes regular expressions and data format template validation to correct basic formatting errors, encoding anomalies, and illegal characters. Logic-level cleaning, based on predefined business rules and constraints, detects, marks, or corrects data inconsistencies that violate business logic. Semantic-level cleaning leverages a constructed business semantic model and ontology reasoning capabilities to identify deeper conceptual conflicts and relational contradictions, resolving them through reasoning rules or manually predefined strategies.

[0053] The cleaned data is not stored in isolation, but rather aggregated and linked according to its semantic relationships and subject areas. For example, data such as the creation event, payment event, and tracking number of the same order are linked together through shared business identifiers or semantic relationships, gradually building a structured business process data graph with a complete event sequence and clear entity relationships.

[0054] In step two, five types of intelligent agents with different specialized responsibilities are defined: process-aware intelligent agents, task-decomposition intelligent agents, resource-negotiation intelligent agents, execution-monitoring intelligent agents, and optimization-decision intelligent agents. A layered hybrid collaborative architecture is adopted, and the most suitable interaction mode is used for different collaborative scenarios.

[0055] The underlying reactive collaboration is implemented through a shared event bus. All agents can act as publishers or subscribers of events. The bus defines a set of standard event types and structured message formats. For example, when a process-aware agent detects a process initiation, it publishes a `ProcessInstanceStarted` event; the task decomposition agent and the optimization decision agent subscribe to this event to trigger subsequent processing. Similarly, when a resource negotiation agent releases a resource, it publishes a `ResourceFreed` event; the task decomposition agent can subscribe to this event to wake up the task decomposition process waiting for that resource. This mechanism achieves millisecond-level distributed event response.

[0056] The mid-level contract network protocol is used for resource allocation and complex task delegation. When a task decomposition agent generates an atomic task or an optimization decision agent generates a refactoring subtask that needs evaluation, it can initiate a tender as a tendering party. The tender message includes the task description, required resources and capabilities, and deadline. Other agents capable of handling the task can evaluate and submit bids as bidders. The bid evaluation function... It is the core of decision-making; for resource bidding, it is usually as follows:

[0057]

[0058] in, To estimate the execution time, To estimate costs, The historical reliability score for this resource is determined, with α, β, and γ serving as weighting coefficients. After collecting all bids, the bidding party will... The successful bidder is selected, and a contractual relationship is established.

[0059] Top-level collaborative intent is triggered when the optimization decision-making agent initiates a restructuring of complex business processes. The optimization decision-making agent first proposes initial restructuring goals and preliminary solutions, inviting all relevant agents to participate as team members. Each team member evaluates the feasibility of the solution from their own professional perspective and proposes constraints or modification suggestions. This is a multi-round proposal-feedback-revision process until a consensus solution is reached. The degree of consensus can be evaluated using the following formula:

[0060]

[0061] in, For the sake of consensus, It refers to the number of intelligent agents participating in the collaboration. It is the predicted value of the current scheme in the i-th target dimension that agent i is interested in. It is the agent's expected value for that dimension. and This represents the possible range of values ​​for this dimension. Consensus is reached when the consensus level exceeds a preset threshold and all key agents make commitments.

[0062] Each agent is equipped with a two-layer knowledge base to support its decision-making and collaboration: a private knowledge base stores core knowledge in the agent's own area of ​​expertise. For example, a task decomposition agent stores task decomposition rules and algorithms; a resource negotiation agent stores resource models and pricing strategies; and an optimization decision-making agent stores optimization algorithms and a library of historical reconstruction cases. The shared knowledge base stores public knowledge extracted from the global business process data graph and historical collaborations. This includes: global business constraints, historical process performance patterns, and validated collaboration protocol templates.

[0063] The intelligent agent updates its knowledge base through periodic synchronization and incremental learning. For example, after each reconstruction, the optimization decision-making agent stores the reconstruction goal, solution, and execution result as new cases in its private knowledge base, and may abstract successful patterns into templates to contribute to the shared knowledge base. The agent's success rate and task completion quality in contract bidding are used to dynamically update its credit rating, which affects its probability of being selected in future bids. The credit update formula can be simplified to:

[0064]

[0065] in, It is a historical credit decay factor. It is a measure of recent task success rate. It is a score for the quality of task output. and It is the weighting coefficient.

[0066] In step three, the process-aware agent serves as the starting point for analysis, automatically transforming the time-ordered business event sequence generated by the data integration layer into a dynamically updated weighted directed graph model G=(V,E,W). This model forms the structured foundation for all subsequent analyses. Here, the node set V represents uniquely identified business activities; the directed edge set E represents the sequential flow relationships between activities; and the weight set W quantifies the strength and characteristics of these flow relationships. The edge weight calculation comprehensively considers historical behavior frequency and data dependency strength to more realistically reflect the process logic. For any activity... To the event The weight of an edge is calculated by the following formula:

[0067]

[0068] in, The normalized value representing the historical execution frequency is typically calculated within the observation window. Execute immediately after completion The number of instances accounts for The percentage of the total number of times completed, i.e. , This indicates the strength of data dependencies, calculated by analyzing the overlap of business data objects passed between activities in an event sequence. A higher value indicates a tighter data coupling between the two activities. This is an adjustment factor used to balance the importance of frequency and data dependence in the weighting.

[0069] Based on this weighted graph, the process-aware agent employs a series of graph mining algorithms for in-depth structural analysis: High-frequency path identification utilizes an improved frequent subgraph mining algorithm to search for activity sequences in the process instance set whose frequency exceeds a preset threshold. These high-frequency paths represent the main execution patterns of the process. Loop structure detection employs loop detection algorithms such as Depth-First Search (DFS) and Strongly Connected Component Analysis (SCC) to identify explicit and implicit loops in the graph that cause execution delays or inefficient resource usage. The severity of a loop can be assessed by its average number of iterations and the average time spent per iteration. Redundant node identification involves analyzing the functional similarity and data dependencies between activities to discover potential process simplification points. For a pair of activities, their functional similarity and data overlap are calculated. If they satisfy the condition that their execution order in the process is fixed and both exceed a threshold, they are identified as potential candidate node pairs for merging.

[0070] The execution monitoring agent extracts and calculates three types of core performance indicators in real time from the continuously flowing event sequences and resource status data, and precisely associates them with specific nodes and edges in the process model to form a model-indicator relationship view, including time indicators, resource indicators, and quality indicators.

[0071] The root cause analysis engine, integrated with the optimization decision-making agent, acts as the brain for diagnosing process bottlenecks and quality issues. When the execution monitoring agent detects anomalies in performance metrics, the optimization decision-making agent initiates a diagnostic process, abstracting the anomaly into a feature vector. This vector contains contextual features such as the type of anomaly, its severity, involved process nodes / edges / resources, concurrent events, and historical data from the same period. The engine first uses association rule mining based on an improved Apriori algorithm in a historical diagnostic case library to find other process elements or environmental factors that frequently co-occur with the current anomaly features. For more complex anomalies involving multiple factors, the engine employs a pre-trained decision tree or random forest classification model for root cause inference. This model is trained on historical anomaly events and their ultimately determined root causes. Given the current feature vector, the model outputs the most likely root cause category and its probability.

[0072] For process nodes or resources identified as root causes, the engine further calculates their contribution to the overall performance anomaly. For example, for latency caused by resource contention, the contribution of a specific resource *r* to the total latency. It can be estimated as follows:

[0073]

[0074] Here, the numerator is the sum of the waiting times of all tasks delayed due to waiting for resource r.

[0075] Based on the above analysis, the optimized decision-making agent generates a structured diagnostic report. The report clearly identifies the root cause of the anomaly, pinpoints the specific process activity, data interface, or resource entity, quantifies its impact, and provides preliminary directions for improvement.

[0076] In step four, after receiving the diagnostic report from the root cause analysis engine, the optimization decision-making agent first transforms the specific performance problem into one or more quantified refactoring objectives. Each objective is defined as a triple. ,in For target type, The specific target value or improvement range to be achieved. This assigns a priority weight to the objective. The optimization decision-making agent, acting as a collaborative initiator, retrieves similar historical reconstruction cases from the shared knowledge base's historical case library, based on the current set of reconstruction objectives and the current process model. The retrieval process is based on a case similarity score. calculate:

[0077]

[0078] in, and These are the target set of historical cases and the process model at that time. To optimize the decision-making agent, target similarity can be measured by the overlap of target types and the vector distance of target values. The similarity of process structures can be evaluated using graph kernel functions or key subgraph matching degree; , , where represents the weighting coefficient. After retrieving the N most similar historical cases, the optimization decision agent extracts validated candidate reconstruction patterns from them, which serve as the initial input for subsequent collaborative design.

[0079] The optimization decision-making agent initiates a formal joint intent coordination session, broadcasting the reconstruction goal G and the initially selected candidate reconstruction pattern set to all relevant functional agents. The core of the coordination is a multi-round proposal-feedback loop. In each round:

[0080] The optimization decision-making agent, based on the current consensus, proposes a draft restructuring plan, detailing which restructuring modes will be applied to which process nodes. Each agent evaluates the draft restructuring plan in parallel from its own expertise. The task decomposition agent simulates the execution of changes in the draft restructuring plan, verifying the logical feasibility and deadlock-free nature of the restructured task sequence. It calculates a logical feasibility score, which combines the sequence's completeness and rule consistency. The resource negotiation agent evaluates the impact of the proposed plan on the resource pool. It calculates a resource conflict index. :

[0081]

[0082] in, This is a forecast of the new peak demand for resource r under the draft restructuring plan. It is its available capacity. This means taking the larger value between x and 0. It is the importance weight of resource r. The higher the value, the greater the risk of resource contention arising from the plan.

[0083] The process-aware agent predicts data flow and event sequence under the new pattern and assesses whether data dependencies will be disrupted. The optimization decision agent collects all evaluation feedback and integrates it into a constraint conflict set. Then, it models the solution optimization as a constraint satisfaction problem or constraint optimization problem, aiming to maximize the expected benefit to the reconstruction objective while satisfying all hard constraints.

[0084] After each revision, the optimization decision-making agent calculates the current consensus level. A simple metric is to calculate the average support level for the current solution from all key agents. When the consensus level exceeds a preset threshold or reaches the maximum number of iterations, the collaborative loop terminates, reaching a final consensus solution.

[0085] Based on the final consensus scheme, the optimized decision-making agent generates an executable refactoring plan that is machine-readable and partially human-readable. This plan is a collection of atomic operation instructions, including: instructions for adding, deleting, modifying, and merging active nodes; instructions for updating resource quotas and scheduling policies; scripts for enabling, disabling, or modifying rules in the business rule engine; and adaptation and mapping change strategies for data interfaces.

[0086] The resource negotiation agent, acting as the execution coordinator, first performs the pre-allocation or reservation of resources according to plan, ensuring that resources are in place when the new process starts. The process-aware agent then formally synchronizes the updated process model to the process data graph and, as the producer of process model change events, sends model change notifications to relevant business systems that have subscribed to the event via the event bus, driving the business systems to load and execute the new process definition.

[0087] After the refactoring goes live, the execution monitoring agent enters a reinforcement learning-guided continuous monitoring and fine-tuning phase, forming a closed loop. The execution monitoring agent continuously observes the actual performance metrics of the new process, constituting its state. It compares the actual value of each key performance indicator with the expected target value range during the refactoring. An immediate reward signal is designed to quantify the achievement of the target.

[0088] The system maintains an estimate of the policy value function Q(s,a) for the current process, which represents the long-term cumulative reward expectation of taking the locally optimized action a in state s. The execution monitoring agent periodically evaluates the value of the current state s. If the actual reward consistently falls below expectations for multiple consecutive monitoring periods, the execution monitoring agent determines that the current policy is inadequate and triggers a small-scale locally optimized process coordinated by the optimization decision agent. This process can be viewed as a new, smaller-scale collaborative refactoring, but its decisions are guided by the reinforcement learning value function.

[0089] Each local optimization action, along with its resulting state transition and actual reward, is stored as an experience tuple in an experience replay buffer. The optimization decision-making agent, or a dedicated learning agent, periodically samples these experiences and updates the estimate of the value function Q(s,a) using the Q-learning algorithm. Simultaneously, successful or unsuccessful experiences are extracted and used to update the private knowledge base of the relevant agents, thereby enabling continuous learning and evolution of the entire multi-agent system. Ultimately, this forms a highly adaptive monitoring-diagnosis-negotiation-adjustment optimization loop.

[0090] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0091] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for refactoring complex business processes based on data integration and multi-agent collaboration, characterized in that: include: Construct a multi-source heterogeneous data integration layer for complex business processes, and perform real-time data collection, semantic mapping, cleaning and quality assessment on data from each node of the business process to form a business process data map with complete event sequences and consistent entity associations. Design a multi-agent system based on a hybrid collaborative mechanism. The system defines and organizes five types of functional agents: process awareness, task decomposition, resource negotiation, execution monitoring, and optimization decision-making. The agents collaborate through a three-layer collaborative architecture of event-driven, contract network protocol, and joint intent. The system analyzes and reconstructs business processes based on private and shared knowledge bases. Based on business process data graphs, intelligent agents collaboratively perform data-driven dynamic analysis and diagnosis of business processes, transforming event sequences into weighted directed graph models to identify structural features, extracting and associating multi-dimensional performance indicators of time, resources, and quality, and applying a root cause analysis engine to locate key activities, data interfaces, or resource contention points that lead to process bottlenecks or defects. The multi-agent collaborative process refactoring and dynamic optimization is initiated by the optimization decision agent. The organizing agent, through a collaborative process of multiple rounds of proposal feedback, generates and optimizes the refactoring plan for the diagnosed problems in response to the refactoring goal. After reaching a consensus, an executable plan is formed that includes activity adjustment, resource migration, rule update and interface adaptation. During operation, closed-loop adaptive optimization is achieved through reinforcement learning mechanism.

2. The method for reconstructing complex business processes based on data integration and multi-agent collaboration as described in claim 1, characterized in that, The construction of a multi-source heterogeneous data integration layer for complex business processes specifically includes: By deploying adapters at each execution node of the business process, structured business data, unstructured document data, and system log stream data are collected in real time. The adapters use log-based incremental capture technology for databases, an event listening mechanism for file systems, and a long-connection callback mode for application interfaces to achieve low-latency data synchronization. Establish a unified business semantic model, perform semantic annotation and mapping on the collected raw data, formally define business entities, events and relationships using domain ontology, map fields and terms in heterogeneous data to concepts in ontology through semantic parsing, and generate a unified metadata description containing timestamps, data sources and semantic types. The design incorporates a three-layer cleaning rule: the syntax layer uses regular expressions and format templates to repair data format anomalies; the logic layer uses predefined business rules and constraints to detect data consistency; and the semantic layer uses ontology reasoning to identify and resolve conceptual conflicts and relational contradictions. The cleaned data is aggregated by subject domain to form a business process data graph with complete event sequences and consistent entity links.

3. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 2, characterized in that, The establishment of a unified business semantic model specifically includes: The domain ontology is constructed using the OWL language, defining core classes such as "business activities", "business objects", "participants", and "rules", as well as their object attributes and data attributes; Based on the constructed domain ontology, semantic annotation tools are used to map terms from different data sources to standard concepts in the ontology through synonym expansion, context matching, and manual verification. The mapping confidence is recorded to form a business semantic knowledge base stored in the form of an RDF graph.

4. The method for reconstructing complex business processes based on data integration and multi-agent collaboration as described in claim 1, characterized in that, The multi-agent system based on a hybrid cooperative mechanism specifically includes: Five types of functional agents are defined: the process-aware agent identifies the start, activity switching, and end states of process instances by subscribing to event streams in the data integration layer; the task decomposition agent receives process fragments and decomposes them into atomic task sequences based on resource constraints and business rules; the resource negotiation agent manages a virtual resource pool including human, equipment, and computing resources, and allocates appropriate resource instances to tasks; the execution monitoring agent tracks the status of atomic tasks and collects performance indicators and abnormal events; and the optimization decision-making agent evaluates the overall efficiency of the process and generates refactoring suggestions. A layered and hybrid collaborative architecture is adopted. The bottom layer is reactive collaboration, where agents publish and subscribe to key state changes through a shared event bus. The middle layer is contract network protocol collaboration, which negotiates resource allocation and task delegation among agents through bidding, tendering, and winning bid mechanisms. The top layer is collaboration based on joint intent, initiated by the optimization decision agent, which organizes agents to make multiple rounds of proposals and commitments around the process refactoring goal to form a consensus solution. A two-layer knowledge base is designed for intelligent agents. The private knowledge base stores rules, cases, and experience data in the area of ​​expertise, while the shared knowledge base stores global constraints, historical performance patterns, and collaboration protocol templates extracted from the business process data graph.

5. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 4, characterized in that, The shared event bus adopts a publish-subscribe model, defining a set of standard event types and message formats. Intelligent agents subscribe to corresponding events according to their responsibilities. When there is a state update or a key event in the business process data graph, the data integration layer or the sensing intelligent agent publishes the event message, triggering the immediate response of the subscriber intelligent agents.

6. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 1, characterized in that, The dynamic analysis and diagnosis of business processes driven by data collaboratively executed by intelligent agents specifically includes: The process-aware intelligent agent transforms the event sequence in the data graph into a weighted directed graph model, where nodes represent business activities, edges represent the flow relationships between activities, and weights are determined by the historical execution frequency and the strength of data dependencies; it also identifies high-frequency paths, loop structures, and redundant nodes through graph mining algorithms. The execution monitoring agent extracts three types of performance metrics from the event sequence: time metrics, including activity duration, workflow interval, and process cycle; resource metrics, including resource utilization, load balancing, and number of conflicts; and quality metrics, including first-pass yield, error rate, and rework rate. These metrics are associated with process nodes in real time. The root cause analysis engine for the optimized decision-making agent is used to bind abnormal performance indicators to specific nodes, edges, or resources in the process model. The engine uses pattern matching based on association rules and classification based on decision trees to locate key business activities, data interfaces, or resource contention points that lead to bottlenecks or quality defects, and generates diagnostic reports.

7. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 6, characterized in that, The specific steps of identifying high-frequency paths, loop structures, and redundant nodes using graph mining algorithms include: The frequent subgraph mining algorithm is used to find activity sequences whose execution frequency exceeds a threshold; the loop detection algorithm is used to identify explicit or implicit loops in the flowchart that cause delays; and by analyzing the data dependencies and functional similarities between activities, a set of activities with highly overlapping inputs and outputs and fixed execution order is identified as potential candidate nodes for merging.

8. The method for reconstructing complex business processes based on data integration and multi-agent collaboration as described in claim 1, characterized in that, The process refactoring and dynamic optimization of multi-agent collaboration specifically includes: The optimization decision-making agent initializes reconstruction goals based on the diagnostic report, including shortening the cycle, improving quality, or reducing costs; based on the goals, it retrieves similar historical cases from the shared knowledge base and extracts candidate reconstruction patterns, including activity merging, parallelization transformation, resource reallocation, or rule simplification. The optimization decision-making agent broadcasts the reconstruction goal and candidate modes to relevant agents; the task decomposition agent evaluates the impact of the modes on the task sequence and simulates the decomposed atomic task chain; the resource negotiation agent evaluates the constraints of the resource pool and calculates the resource reallocation scheme; the process awareness agent predicts the data flow and event sequence under the new mode; and each agent gradually revises the reconstruction scheme through multiple rounds of proposal feedback loops. Once a consensus is reached, an executable refactoring plan is generated, which includes specific activity adjustment steps, resource migration instructions, business rule update scripts, and data interface adaptation strategies; the resource negotiation agent executes resource reservation and allocation according to the plan; the process awareness agent updates the process model and sends model change notifications to relevant business systems. During the reconstructed process, the execution monitoring agent initiates a reinforcement learning mechanism to collect real-time performance data of the new process and compare it with the expected target. If the deviation continues to exceed the tolerance range, a new round of collaborative diagnosis and local optimization is triggered, forming a closed-loop adaptive optimization cycle.

9. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 8, characterized in that, The multi-round proposal feedback loop specifically includes: The optimization decision-making agent proposes an initial reconstruction plan; the task decomposition and resource negotiation agents evaluate the plan from the perspectives of task logical feasibility and resource availability, and return modification suggestions or constraint conflicts; the optimization decision-making agent integrates feedback and adjusts the plan using constraint satisfaction problem-solving techniques; the loop terminates after a pre-set maximum number of rounds or until all agents accept the plan.

10. The method for reconstructing complex business processes based on data integration and multi-agent collaboration according to claim 9, characterized in that, The reinforcement learning mechanism specifically includes: Set a target value range for each performance indicator; the execution monitoring agent continuously collects the actual values ​​and calculates the gap with the target range; when the gap continues to exceed the preset period, the current process status and performance gap are used as status inputs to call the optimization decision agent to start a small-scale collaborative optimization; the optimization actions and their effects are recorded and used to update the agent's private experience knowledge base.

Images