Search

CN-121542094-B - Complex task multi-agent collaborative disassembly method and system based on business process

CN121542094BCN 121542094 BCN121542094 BCN 121542094BCN-121542094-B

Abstract

The invention provides a complex task multi-agent collaborative disassembly method and a system based on a business process, which relate to the technical field of artificial intelligence and comprise the steps of constructing a map by analyzing a business process document, matching the complex task with the flow node, identifying a closed-loop dependent path, selecting an agent based on constraint logic, establishing a causal trace-back chain to process abnormal conditions, and carrying out semantic alignment to ensure data transfer among subtasks. The invention improves the automation degree and the execution efficiency of complex tasks processed by the cooperation of multiple agents, and enhances the fault tolerance capability of the system and the reliability of task completion.

Inventors

  • ZHAO LEIZHEN
  • SUN SHUMENG
  • Jia Songrui

Assignees

  • 北京亦庄智能城市研究院集团有限公司

Dates

Publication Date
20260508
Application Date
20260119

Claims (9)

  1. 1. The complex task multi-agent collaborative disassembly method based on the business process is characterized by comprising the following steps of: analyzing a business process document, extracting process nodes and dependency relations, and constructing a business process map; matching the complex task with the process nodes in the business process map to obtain a matched node set; Identifying closed-loop dependent paths in the matched node set, marking a circulating node group, extracting service termination conditions, setting a circulating monitoring mechanism, and converting the circulating node group into sub-tasks based on a termination signal; Expressing the execution constraint of the subtasks as constraint logic, expressing the history execution record of the agent as an inference rule base, and selecting the agent with the highest logic consistency score for binding by matching and calculating the logic consistency score; When an execution abnormality is detected, tracing back to an abnormality source subtask along the causal tracing chain, extracting an abnormality feature vector of the abnormality source subtask, constructing an abnormality propagation diagram based on the abnormality feature vector, extracting node features from a service flow diagram for matching, selecting a flow node with the highest matching degree as a substitute flow node, and generating a substitute subtask, wherein the method comprises the following steps: Deconstructing the abnormal feature vector to obtain a time sequence feature sequence, calculating the change rate and fluctuation amplitude of the time sequence feature sequence based on a sliding time window, and generating an abnormal evolution feature set; constructing an abnormal propagation diagram based on the abnormal evolution feature set, recording the transmission intensity and the diffusion direction among the process nodes at the diffusion layer, recording the performance loss and the resource occupation of the process nodes at the influence layer, and obtaining a candidate abnormal node set and a risk node set through the mapping calculation of the diffusion layer and the influence layer; extracting task types, input and output specifications and resource consumption of flow nodes from the business flow map, constructing execution feature vectors, and calculating similarity among the execution feature vectors to obtain a node connection weight matrix; constructing a node matching network based on the node connection weight matrix, clustering the node matching network by using a community discovery algorithm to obtain a node matching group, calculating the execution success rate and the resource utilization rate of each flow node in the node matching group, and selecting the flow node with the highest comprehensive score of the execution success rate and the resource utilization rate from the node matching group where the candidate abnormal node set is positioned as a substituted flow node; combining the historical execution data of the alternative flow node with the execution success rate threshold value to generate an alternative subtask; And carrying out semantic alignment on the output data of the front sub-task and the input constraint of the follow-up sub-task, injecting the aligned data into the follow-up agent and triggering execution.
  2. 2. The method of claim 1, wherein identifying closed-loop dependent paths in the set of matched nodes and marking the set of circulating nodes, extracting traffic termination conditions to set a circulating monitoring mechanism, converting the set of circulating nodes into sub-tasks based on the termination signals comprises: Traversing the dependency relationship among the flow nodes in the matched node set, and marking the flow nodes on the closed-loop dependency path as a circulating node group when the closed-loop dependency path is detected to be formed among the flow nodes; Extracting service termination conditions corresponding to the circulating node group from the service flow diagram, expressing constraint variables in the service termination conditions as state variables, and converting judgment logic into a Boolean logic expression; Setting an iteration counter and a state tracker at the initial flow node of the loop node group, wherein the iteration counter records the execution times of the loop node group, and the state tracker records the real-time numerical value of the constraint variable, and updates the iteration counter and the state tracker when the loop node group is executed each time; Substituting the real-time numerical value recorded by the state tracker into a Boolean logic expression for calculation, and generating a cycle termination signal when the calculation result is true; converting each flow node in the loop node group into a subtask, taking the current value of the iteration counter as the iteration identification of the subtask, taking the loop termination signal as the termination judgment basis of the subtask, and generating the subtask comprising the iteration identification and the termination judgment basis.
  3. 3. The method of claim 1, wherein expressing the execution constraint of the subtask as a constraint logical formula, expressing the historical execution record of the agent as an inference rule base, calculating the logical consistency score by matching, and selecting the agent with the highest logical consistency score for binding comprises: Extracting a pre-condition variable and a post-result variable from the execution constraint of the subtask, expressing the pre-condition variable as a front predicate of a logic formula, expressing the post-result variable as a post predicate of the logic formula, and constructing a subtask constraint logic formula; extracting an input state and an output state from a history execution record of the intelligent agent, mapping the input state into a precondition of an inference rule, mapping the output state into a conclusion of the inference rule, and generating an inference rule base; decomposing a precursor predicate in a subtask constraint logic formula into an atomic predicate sequence, decomposing a precondition in an inference rule base into an atomic precondition sequence, calculating structural isomorphism between the atomic predicate sequence and the atomic precondition sequence one by one, and generating a precondition matching score based on the structural isomorphism; Decomposing the post predicate in the subtask constraint logic formula into a target state vector, decomposing a conclusion item in the reasoning rule base into an actual state vector, calculating a vector distance between the target state vector and the actual state vector, and generating a conclusion matching score based on the reciprocal of the vector distance; and carrying out weighted summation on the premise matching score and the conclusion matching score to obtain a logic consistency score, and selecting an agent with the highest logic consistency score for binding.
  4. 4. The method of claim 1, wherein creating a causal trace-back chain as the agent performs the subtasks, and tracing back to the source subtask of the anomaly when an execution anomaly is detected along the causal trace-back chain, and wherein extracting the anomaly feature vector of the source subtask of the anomaly comprises: when the intelligent agent executes the subtask, carrying out hash operation on input data to generate an input data fingerprint, and storing the input data fingerprint and the subtask identifier in an associated mode; after the execution of the subtasks is completed, carrying out hash operation on output data to generate output data fingerprints, extracting the data transfer relation of the subtasks to subsequent subtasks, constructing the output data fingerprints, the subtask identifications and the data transfer relation as causal tracing nodes, and connecting the causal tracing nodes in series according to the execution time sequence to form a causal tracing chain; Monitoring state information of an agent executing subtasks, triggering an abnormality detection signal when a predefined abnormality mode appears, extracting an abnormality subtask identifier, positioning a corresponding causal tracing node from a causal tracing chain, acquiring a data transfer relationship, and traversing a leading causal tracing node along a reverse path of the data transfer relationship; In traversal, extracting an input data fingerprint and an output data fingerprint from each leading-cause-effect traceability node, performing consistency check on the output data fingerprint and the input data fingerprint of a subsequent node, marking the leading-cause-effect traceability node as an abnormal source node when the check fails, extracting a corresponding subtask, and determining an abnormal source subtask; And extracting the abnormal characteristics of the input data, the abnormal characteristics of the execution logic and the abnormal characteristics of the output data from the execution records of the subtasks of the abnormal source head, and combining the abnormal characteristics and the abnormal characteristics of the output data into abnormal characteristic vectors.
  5. 5. The method of claim 4, wherein the consistency check of the output data fingerprint with the input data fingerprint of the subsequent node comprises: Extracting a fingerprint conversion mode from the data transfer relation from an output data fingerprint to an input data fingerprint, wherein the fingerprint conversion mode records the change rule of hash values when data are transferred between nodes; Performing fingerprint space mapping on the output data fingerprint by applying a fingerprint transformation mode, generating an expected input data fingerprint, calculating the bitwise exclusive-or result of the expected input data fingerprint and the input data fingerprint of the subsequent node, counting the number of bits with the median of 1 in the exclusive-or result, and determining the fingerprint deviation degree; and when the fingerprint deviation exceeds a preset deviation threshold, judging that the verification fails, and marking the leading cause and effect traceable node as an abnormal source node.
  6. 6. The method of claim 1, wherein semantically aligning the predecessor subtask output data with the successor subtask input constraints, injecting the aligned data into the successor agents and triggering execution comprises: Constructing a multidimensional association graph by data items in the data output by the preamble subtasks, extracting the change rule of the data items in the time dimension, extracting the association strength among the data items in the space dimension, extracting the action level of the data items in the service dimension, and generating a data feature association matrix; Inputting constraint expression of a subsequent subtask as a constraint network, calculating importance degree and mutual exclusion degree of constraint conditions, constructing a constraint priority sequence, mapping a data characteristic association matrix into the constraint priority sequence, carrying out hierarchical conversion and combined optimization on data items according to a mapping result, and generating semantic alignment data with optimal constraint satisfaction degree; Extracting a processing mode sequence from a history execution record of a subsequent intelligent agent, identifying a processing bottleneck point and a fluctuation interval, and constructing a processing performance curve of the intelligent agent; According to the processing performance curve, semantic alignment data are calculated to be a core data stream and an auxiliary data stream, the core data stream is transmitted according to the peak of the processing performance curve, the auxiliary data stream is transmitted according to the trough, the transmission proportion of the core data stream and the auxiliary data stream is calculated based on the processing state feedback information of the agent, and the adjusted data are injected into the following agent and an execution signal is sent.
  7. 7. A complex task multi-agent collaborative disassembly system based on business processes for implementing the method of any of the preceding claims 1-6, comprising: The flow modeling module is used for analyzing the business flow document, extracting the flow nodes and the dependency relationship and constructing a business flow map; The task mapping module is used for matching the complex task with the process nodes in the business process map to obtain a matched node set; the circulating disassembly module is used for identifying closed-loop dependent paths in the matched node set, marking a circulating node group, extracting service termination conditions and setting a circulating monitoring mechanism, and converting the circulating node group into sub-tasks based on termination signals; The agent selection module is used for expressing the execution constraint of the subtasks as constraint logic formula, expressing the history execution record of the agents as an inference rule base, and selecting the agent with the highest logical consistency score for binding by matching and calculating the logical consistency score; The abnormal replacement module is used for establishing a causal trace chain when the intelligent agent executes the subtasks, tracing back to an abnormal source subtask along the causal trace chain when the execution abnormality is detected, extracting an abnormal characteristic vector of the abnormal source subtask, constructing an abnormal propagation diagram based on the abnormal characteristic vector, extracting node characteristics from the service flow diagram for matching, and selecting a flow node with the highest matching degree as a substituted flow node to generate a substituted subtask; the semantic linking module is used for carrying out semantic alignment on the output data of the preamble subtask and the input constraint of the following subtask, injecting the aligned data into the following agent and triggering execution.
  8. 8. An electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1 to 6.
  9. 9. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 6.

Description

Complex task multi-agent collaborative disassembly method and system based on business process Technical Field The invention relates to the technical field of artificial intelligence, in particular to a complex task multi-agent collaborative disassembly method and system based on a business process. Background With the rapid development of artificial intelligence technology, multi-agent collaborative systems exhibit increasingly important values in processing complex business processes. In modern enterprise environments, business processes often involve collaboration of multiple links and roles, which often have complex dependencies, loop structures, and exception handling mechanisms. Traditional business process automation relies primarily on predefined rules and flow diagrams, which are difficult to deal with dynamically changing business requirements. With the development of large language models and intelligent agent technologies, complex tasks are automatically disassembled and distributed to a plurality of specialized intelligent agents to finish in a cooperated manner, and the method can effectively improve the automation level and the processing efficiency of the business process. The prior art still has some defects and shortcomings in the aspect of complex task multi-agent collaborative disassembly based on business processes, most of the prior multi-agent system adopts a static task allocation mechanism, the task disassembly strategy cannot be adaptively adjusted according to the dynamic changes of the business processes, particularly for the business processes comprising complex circulation structures, the task execution efficiency is low or the task execution is trapped in dead circulation, the prior art mainly depends on simple rule matching or keyword matching in the aspect of matching of agents and subtasks, lacks deep analysis and logic consistency assessment on the historical execution capacity of the agents, cannot guarantee to select the most suitable agents to execute specific subtasks, is particularly obvious when the prior multi-agent system generally adopts a simple retry or rollback strategy in the aspect of exception handling, lacks systematic exception traceability and propagation analysis mechanism, cannot accurately position exception sources and generate proper alternatives, and causes poor system robustness and difficult continuity and reliability in complex business scenes. Disclosure of Invention The embodiment of the invention provides a complex task multi-agent collaborative disassembly method and system based on a business process, which can solve the problems in the prior art. In a first aspect of the embodiment of the present invention, a complex task multi-agent collaborative disassembly method based on a business process is provided, including: analyzing a business process document, extracting process nodes and dependency relations, and constructing a business process map; matching the complex task with the process nodes in the business process map to obtain a matched node set; Identifying closed-loop dependent paths in the matched node set, marking a circulating node group, extracting service termination conditions, setting a circulating monitoring mechanism, and converting the circulating node group into sub-tasks based on a termination signal; Expressing the execution constraint of the subtasks as constraint logic, expressing the history execution record of the agent as an inference rule base, and selecting the agent with the highest logic consistency score for binding by matching and calculating the logic consistency score; when an execution abnormality is detected, tracing back to an abnormality source subtask along the causal tracing chain, extracting an abnormality feature vector of the abnormality source subtask, constructing an abnormality propagation diagram based on the abnormality feature vector, extracting node features from a service flow diagram for matching, and selecting a flow node with the highest matching degree as a substitute flow node to generate a substitute subtask; And carrying out semantic alignment on the output data of the front sub-task and the input constraint of the follow-up sub-task, injecting the aligned data into the follow-up agent and triggering execution. In an alternative embodiment, identifying closed-loop dependent paths in the set of matching nodes and marking the set of circulating nodes, extracting service termination conditions and setting up a circulating monitoring mechanism, converting the set of circulating nodes into subtasks based on the termination signals includes: Traversing the dependency relationship among the flow nodes in the matched node set, and marking the flow nodes on the closed-loop dependency path as a circulating node group when the closed-loop dependency path is detected to be formed among the flow nodes; Extracting service termination conditions corresponding to the circulating node group from the se