Multi-agent adaptive collaborative anomaly detection method based on task decomposition and recombination

Abstract

The invention discloses a multi-agent adaptive collaborative anomaly detection method based on task decomposition and recombination. The method comprises the following steps: S1, collecting data of different modes; s2, performing different initialization processing on the data of different modalities; s3, a task decomposition agent constructs a task ontology model based on the meta-information of the multi-source data, disassembles an overall anomaly detection task into a plurality of independently executable single-mode sub-tasks, and constructs a hierarchical dependency graph at the same time; s4, the cooperative control agent dynamically allocates the tasks to corresponding modal agents according to the modal affiliation information of the subtasks; and S5, the global abstract agent receives the anomaly detection reports reported by all the modal agents, semantic aggregation and logic integration are carried out, multiple rounds of debate are organized, and finally a global anomaly detection comprehensive report is output. According to the method, a multi-agent cooperation mechanism is introduced, and modal characteristics and a task scheduling optimization strategy are fused, so that efficient and explainable anomaly detection with self-adaptive capability is realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent information processing technology, specifically relating to a multi-agent adaptive collaborative anomaly detection method based on task decomposition and reorganization. Background Technology

[0002] With the explosive growth of multimodal data in various application scenarios, cross-modal anomaly detection has gradually become a key technology in intelligent sensing systems. In complex environments such as medical diagnosis, civil aviation safety, and intelligent manufacturing, systems need to comprehensively process information from heterogeneous modalities such as images, voice, text, and the Internet of Things in order to accurately identify potential anomalies.

[0003] However, traditional anomaly detection methods face three major challenges in multimodal scenarios: First, significant semantic differences between modalities make feature fusion difficult, and direct splicing often fails to achieve effective information integration; second, mainstream methods mostly rely on black-box models, lack interpretability, and make it difficult to trace the cause of anomalies; third, existing systems are mostly based on static architectures, making it difficult to dynamically adapt to changes such as the addition of new modalities or changes in computing resources. Therefore, new frameworks are urgently needed to solve these major challenges. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention proposes a multi-agent collaborative framework that effectively improves detection performance through task layering and decoupling, modality-specific optimization, and consensus-driven mechanisms.

[0005] Specifically, the Task Decomposition Agent (T-Agent) achieves dynamic task scheduling by constructing a dependency graph, the Modal Agent (M-Agent) integrates knowledge graphs for intra-modal verification to enhance accuracy, and the Global Summary Agent (G-Agent) introduces multi-round semantic debate to aggregate text reports, achieving multi-modal consistency judgment. Simultaneously, an end-to-end optimization strategy based on the GRPO (Graph-based Resource-aware Policy Optimization) algorithm improves resource utilization efficiency while ensuring detection accuracy.

[0006] The technical solution of this invention is as follows:

[0007] Step a. Obtain detection data for at least two heterogeneous modes;

[0008] Step b. Perform modal-related initialization processing on the heterogeneous modal data respectively;

[0009] Step c. The task decomposition agent constructs a task ontology model based on the metadata of the data, decomposes the total anomaly detection task into several independently executable subtasks, and dynamically generates a hierarchical dependency graph based on the dependencies and resource constraints between the subtasks.

[0010] Step d. The collaborative control agent assigns the subtasks to the corresponding modal agents based on their modal affiliation, and monitors the progress of each modal agent in real time during execution to achieve orderly execution;

[0011] Step e. The global summary agent collects the anomaly detection reports generated by each modality agent, aligns them, triggers multi-agent debate, until the preset termination condition is met, and finally outputs a cross-modal anomaly detection comprehensive report;

[0012] Step f. Utilize a graph-based multi-objective reward strategy optimization algorithm, with task completion, resource utilization, detection accuracy, and response latency as reward factors, to perform end-to-end dynamic optimization of the strategy parameters of the task decomposition agent, collaborative control agent, and each modal agent, thereby achieving continuous evolution of system performance.

[0013] The beneficial effects of this invention are:

[0014] 1. Enhanced cross-modal adaptability

[0015] This invention addresses the issues of large differences and semantic inconsistencies in multimodal data by designing a unified feature mapping space and a modality-adaptive anomaly detection mechanism. This enables anomaly data from different modalities to possess semantic consistency and structural regularity, significantly improving the interpretability and operability of cross-modal comparison and collaborative analysis.

[0016] 2. Intelligent and dynamic task scheduling

[0017] By introducing a task decomposition agent (T-Agent) to construct a task ontology and a hierarchical dependency graph, the complex anomaly detection process can be intelligently decomposed into multiple independently executable subtasks, and the execution path can be dynamically optimized based on resources and dependencies, thereby achieving adaptive task scheduling and system flexibility.

[0018] 3. Multi-agent collaboration mechanism enables efficient execution.

[0019] The Collaborative Control Agent (C-Agent) coordinates multiple Modal Agents (M-Agents) to perform parallel operations and monitors progress in real time, effectively solving the problems of multi-task synchronization and resource scheduling, and improving task execution efficiency and system stability.

[0020] 4. Global information fusion and multi-agent consensus mechanism

[0021] The Global Summary Agent (G-Agent) can integrate text reports output by various modal agents. Through multimodal semantic aggregation and multi-round debate supported by AutoGen, it can achieve logically unified, cross-modal coordinated, and traceable global anomaly summaries, meeting the needs of multi-source data decision fusion in complex scenarios.

[0022] 5. System optimization and autonomous evolution capabilities

[0023] By constructing a multi-objective reward mechanism using the GRPO algorithm, this invention can perform end-to-end policy optimization based on indicators such as agent task completion, resource utilization efficiency, and response latency, thereby promoting the system's development towards intelligent directions of autonomous learning, adaptive optimization, and continuous evolution. Attached Figure Description

[0024] Figure 1 This is a framework diagram of a multi-agent adaptive collaborative anomaly detection method based on task decomposition and reorganization according to the present invention.

[0025] Figure 2 This is a flowchart of a multi-agent adaptive collaborative anomaly detection method based on task decomposition and reorganization according to the present invention. Detailed Implementation

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0028] This invention aims to address the problems of feature fusion difficulties, poor interpretability, and weak dynamic adaptability in traditional multimodal anomaly detection. To solve these technical problems, such as... Figure 1 As shown, the present invention includes the following steps:

[0029] Step 1: Collect data from different modalities

[0030] Step 2: Perform different initialization processes on data of different modalities.

[0031] Step 3: The Task Decomposition Agent (T-Agent) integrates metadata from multi-source heterogeneous data (including data source type, collection frequency, context description, etc.) to construct a task ontology model. It decomposes the complex multimodal anomaly detection problem into several independently executable single-modal subtasks. Based on the dependencies and resource constraints between tasks, it dynamically generates a hierarchical dependency graph with priority ranking and execution path optimization to achieve intelligent and adaptive task scheduling.

[0032] Step 4: The collaborative control agent (C-Agent) assigns the decomposed subtasks to the corresponding modal agents (M-Agent) based on the modal attribution information of the subtasks, and monitors the processing progress of each modal agent in real time during the task execution process to achieve the orderly execution of the tasks.

[0033] Step 5: After receiving the anomaly detection reports generated by each modality, the Global Summary Agent (G-Agent) first aligns them and then generates a global summary. Then, the modality agents exchange anomaly detection reports with each other and modify their own reports based on the anomaly detection reports received from other agents. They then send the reports back to the Global Summary Agent to complete the multi-agent debate. Finally, after the debate termination conditions are met, the Global Summary Agent generates the final anomaly detection summary report.

[0034] Step 6: Introduce the GRPO (Generalized Reward Policy Optimization) algorithm based on gradient policy optimization. Design a multi-objective reward function for the performance of three types of agents (T-Agent, C-Agent, and M-Agent) during task execution. Take into account indicators such as task completion, resource utilization, detection accuracy, and system response latency to perform end-to-end policy optimization, realize dynamic adjustment of agent behavior and continuous evolution of overall system performance, and promote the development of autonomous learning and self-evolution capabilities of the multimodal anomaly detection system.

[0035] This invention can be applied to anomaly identification scenarios involving multimodal heterogeneous data, such as intelligent manufacturing quality inspection, industrial equipment operation and maintenance, financial risk control, smart city video security, network security intrusion detection, and medical multi-source monitoring.

[0036] This application provides a multi-agent adaptive cooperative anomaly detection method based on task decomposition and reorganization, such as... Figure 2 As shown, an example for industrial applications includes the following steps:

[0037] Step 1: First, acquire the following three types of data: 1) RGB image; 2) 3D point cloud data; 3) Process parameter data: including but not limited to time-series parameters such as temperature, pressure, speed, and time during the production process.

[0038] Step 2: Extract local block features using a pre-trained ViT model:

[0039]

[0040] in For the extracted local block features, This involves normalizing the three types of data mentioned in step 1.

[0041] Step 3: Task Breakdown:

[0042] T-Agent first defines a complete pool of subtasks based on the modality of the input data and the type of anomaly to be detected. Let the total anomaly detection task be... The set of subtasks obtained after decomposition is defined as follows:

[0043]

[0044] Each subtask It corresponds to a specific data processing unit.

[0045] In this embodiment, the system input includes three types of heterogeneous data (RGB image, 3D point cloud, and process parameters). T-Agent performs preliminary task mapping based on the data modality: 1): RGB image modality: can be decomposed into (Surface texture anomaly detection) (Color anomaly detection) (Subtasks such as shape defect detection) 2): 3D point cloud modality: can be decomposed into (Volume anomaly detection) (Surface geometric deformation detection) (Key dimension measurement) and other sub-tasks. 3): Process parameter modes: can be decomposed into (Trend Anomaly Detection) (Detection of anomalies exceeding the threshold) Subtasks such as (multi-parameter correlation anomaly detection) are also included.

[0046] Furthermore, T-Agent uses a dependency matrix This formally describes the dependencies between these relational subtasks. The elements of the matrix... The definition is as follows:

[0047]

[0048] Among them, the dependency matrix It is a square matrix whose dimension is equal to the number of subtasks. It clearly depicts the global dependency network between subtasks. After the dependencies are established, the T-Agent assigns the subtasks to the most suitable modal agent (M-Agent) for execution.

[0049] Step 4: Collaborative Control

[0050] Collaborative Control Agent (C-Agent) is based on the modality attribution information of subtasks. The decomposed subtasks are then assigned to the corresponding modal proxies. (M-Agent) and monitors the processing progress of each modal agent in real time during task execution to achieve orderly task execution.

[0051] Step 5: Generate a global summary:

[0052] After collecting anomaly detection reports from different modal agents, the global summary agent G-Agent first performs semantic alignment of the reports based on spatial location (the mapping relationship between image pixels and point cloud points) or time (process parameters) to generate a global summary. This summary is generated by inputting a prompt to the G-Agent large model, such as: "Please understand these reports, identify the contradictions between the reports, adjust their format and content to resolve the contradictions, and generate a global summary."

[0053] Then, all modal agents read the aligned reports from other agents, modify their own reports, and send them back to the global summary agent.

[0054] Specifically, each modal agent The multi-agent debate process is completed by updating its arguments based on the content of other modal reports. Anomaly detection report. The generation is based on a debate function. :

[0055] ,

[0056] in Indicates except Reports from other agents This is the collection of reports from all agents in round t-1. For T-1 round agency The report, function Indicates agent The report processing operation, where t represents the iteration number after the current report.

[0057] All the revised reports from the agents make up a new round of report collections:

[0058] ,

[0059] when ), To control the threshold, the default value is 0.1, or The debate will then end, in the final round. This is the final overall anomaly detection report.

[0060] Step 6: GRPO Multi-Objective Reward Optimization:

[0061] Define a multi-objective reward function:

[0062]

[0063] in , , , .

[0064] The overall loss function is as follows:

[0065]

[0066] in For three different proxy strategy parameters, The discount factor (default is 0.95), KL divergence term. Constraint policy update magnitude It refers to the specific strategy adopted by the model.

[0067] Furthermore, the specific gradient updates are as follows:

[0068]

[0069]

[0070]

[0071] All values ​​represent the learning rate, with default values ​​set as follows: , The derivative of the corresponding parameter.

[0072] The present invention has been described in detail with reference to the foregoing embodiments. For those skilled in the art, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions claimed by the present invention.

Claims

1. A multi-agent adaptive cooperative anomaly detection method based on task decomposition and reorganization, characterized in that, include: Step a. Obtain detection data for at least two heterogeneous modes; Step b. Perform modal-related initialization processing on the heterogeneous modal data respectively; Step c. The task decomposition agent constructs a task ontology model based on the metadata of the data, decomposes the total anomaly detection task into several independently executable subtasks, and dynamically generates a hierarchical dependency graph based on the dependencies and resource constraints between the subtasks. Step d. The collaborative control agent assigns the subtasks to the corresponding modal agents based on their modal affiliation, and monitors the progress of each modal agent in real time during execution to achieve orderly execution; Step e. The global summary agent collects the anomaly detection reports generated by each modality agent, aligns them, triggers multi-agent debate, until the preset termination condition is met, and finally outputs a cross-modal anomaly detection comprehensive report; Step f. Utilize a graph-based multi-objective reward strategy optimization algorithm, with task completion, resource utilization, detection accuracy, and response latency as reward factors, to perform end-to-end dynamic optimization of the strategy parameters of the task decomposition agent, collaborative control agent, and each modal agent, thereby achieving continuous evolution of system performance.

2. The method according to claim 1, characterized in that: The heterogeneous modal data mentioned in step a includes at least two of the following: images, point clouds, text, audio, and time-series sensor signals.

3. The method according to claim 1 or 2, characterized in that: The initialization process described in step b is as follows: performing local block feature extraction based on a pre-trained visual Transformer on the normalized heterogeneous modal data.

4. The method according to claim 1, characterized in that: The task ontology model described in step c uses data source type, collection frequency and context description as meta-information input, and uses a directed acyclic graph to express the dependencies between subtasks.

5. The method according to claim 4, characterized in that: The dependency is through Dependency matrix D describes the matrix elements. ∈{0,1}, when =1 indicates a subtask Must be in subtask It can only be started after completion.

6. The method according to claim 1, characterized in that: The collaborative control agent described in step d obtains the progress information of each modal agent in real time through periodic status queries, and dynamically reallocates the remaining subtasks when resource bottlenecks occur.

7. The method according to claim 1, characterized in that: The multi-agent debate process described in step e includes: The e1 global summary agent performs semantic alignment on reports from different modalities; e2 Each modal agent reads the aligned reports from other agents and updates its own report; e3 Repeat e1-e2 until the reported difference is less than the first threshold or the round is greater than the second threshold, then terminate the debate and output the final comprehensive report.

8. The method according to claim 7, characterized in that: The semantic alignment achieves cross-modal information matching based on spatial mapping or timestamp association.

9. The method according to claim 1, characterized in that: The multi-objective reward function in step f rewards task completion, resource utilization, and detection accuracy as positive rewards and punishes response latency as a negative reward. The sum of the weights of each reward item is 1, and the weights are adaptively adjusted according to the system state during training.

10. The method according to claim 9, characterized in that: The policy optimization algorithm is the GRPO algorithm, whose policy update is constrained by KL divergence, and the learning rate ratio of the task decomposition agent, the cooperative control agent and the modal agent is preset to 1.2:1.0:0.8.

Images