CN-121981024-A - Intelligent tool box full life cycle control method based on data cockpit

CN121981024ACN 121981024 ACN121981024 ACN 121981024ACN-121981024-A

Abstract

The invention discloses an intelligent tool box full life cycle control method based on a data cockpit, and relates to the technical field of artificial intelligence and micro-service operation and maintenance. The method comprises the steps of constructing an agent identity and access boundary map based on multi-source interaction data to generate a pre-deployment strategy, collecting natural language output sequences and mapping the natural language output sequences into continuous medium fields in a multi-agent execution task, calculating fluid dynamics parameters comprising entropy coupling semantic Reynolds numbers, inputting the parameters into a data cockpit to render a flow field distribution view, and if the fact that the agent has cognitive drift is judged according to the parameters, injecting system resistance to the intelligent agent based on a preset track attenuation resistance function to control retirement. The method is used for solving the problems that the traditional discrete monitoring is difficult to quantitatively identify the hidden cognitive drift of the intelligent agent and the framework cascade avalanche is easy to be caused by forced fusing of abnormal nodes.

Inventors

YANG XICHENG
SHAO XIANG
WANG CHENG
PAN HUI
GAO BO
WU ZHENZHONG
SUN JUAN
XU LU

Assignees

合肥大多数信息科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The intelligent tool box full life cycle management and control method based on the data cockpit is characterized by comprising the following steps of: constructing an intelligent body identity and access boundary map based on multisource system interaction data; generating a pre-deployment strategy based on the identity of the intelligent agent and the access boundary map, and issuing the strategy to a target computing node so as to control the intelligent agent to execute an automatic task; In the process of executing an automation task by multiple agents, acquiring a natural language output sequence, mapping the natural language output sequence into a continuous medium field, and calculating fluid dynamics parameters of the continuous medium field containing entropy coupling semantic Reynolds numbers; inputting the fluid dynamic parameters into a data cockpit to render a flow field distribution view; If the cognitive drift exists in the target agent according to the flow field distribution view, the system resistance is injected into the target agent, so that the target agent is controlled to execute the retired operation.
2. The method of claim 1, wherein calculating the hydrodynamic parameters of the continuous media field including the entropy coupling semantic reynolds number comprises: Converting the output sequence into a semantic vector; calculating semantic flow speed and semantic fluid density based on the time change rate and the distribution degree of the semantic vector; calculating semantic information entropy increment based on logarithmic probability distribution of the output sequence; mapping the decoding temperature parameter for generating the output sequence into a model temperature compensation coefficient; And calculating the entropy coupling semantic Reynolds number according to the semantic flow speed, the semantic fluid density, the semantic information entropy increment, the model temperature compensation coefficient and the semantic viscosity coefficient.
3. The method according to claim 2, wherein the calculation formula of the entropy coupling semantic reynolds number is: Wherein, the For the entropy coupling of the semantic reynolds numbers, For a semantic fluid density, For the semantic flow rate, Is the characteristic sequence length of the natural language output sequence, Is a preset semantic viscosity coefficient, For the increase in the entropy of the semantic information, Is a model temperature compensation coefficient.
4. The method of claim 1, wherein determining that the target agent has cognitive drift from the flow field distribution view comprises: If the entropy coupling semantic Reynolds number corresponding to the target intelligent agent is not greater than a preset threshold value, judging that the semantic flow of the target intelligent agent is in a laminar state, and judging that the target intelligent agent does not have cognitive drift; And if the entropy coupling semantic Reynolds number is larger than the threshold value, generating a vortex representation image in a region corresponding to the target intelligent agent in the flow field distribution view, and judging that the target intelligent agent has cognitive drift.
5. The method of claim 1, wherein injecting system resistance for the target agent to control the target agent to perform decommissioning operations comprises: Acquiring a service activity index of a target intelligent agent and a dependency weight aiming at a downstream service node; substituting the liveness index and the dependence weight into the track attenuation resistance function, and calculating to obtain a resistance distribution coefficient; injecting time delay into an API call link of the target intelligent agent according to the resistance distribution coefficient, and cutting off non-core read-write permission of the target intelligent agent according to the corresponding proportion; and destroying the authentication token of the target intelligent agent when the concurrent connection number of the target intelligent agent is monitored to be attenuated to zero.
6. The method of claim 1, wherein during execution of the automated task by the multi-agent, further comprising: capturing a full-link execution log of the multi-agent when executing an automation task; mapping the automatic operation flow in the full-link execution log into a lattice growth process in a three-dimensional space; If an abnormal execution result exists in the full-link execution log, generating a corresponding crystallographic defect map in the lattice growth process; Synchronously rendering a three-dimensional crystal strain view including the crystallographic defect map in a data cockpit.
7. The method of claim 6, wherein generating a corresponding crystallographic defect map during lattice growth comprises: if the abnormal execution result is a local logic error report of a single computing node, generating point defect mapping at a corresponding position in a three-dimensional space; if the abnormal execution result is cascade error caused by cross-system data pollution, generating a linear defect dislocation map in a three-dimensional space, and calculating the strain tensor torsion degree corresponding to the linear defect dislocation map.
8. The method of claim 7, wherein the method further comprises: when a drill-down instruction for a specific vortex characterization image in a flow field distribution view is received, linear defect dislocation mapping with the same time sequence label as that of the specific vortex characterization image region in the three-dimensional crystal strain view is highlighted.
9. The method of claim 1, wherein prior to generating the pre-deployment policy based on the agent identity and the access boundary map, further comprising: calculating the basic infection number of each node in the identity and access boundary map when encountering malicious instruction injection; And rendering a right infection thermodynamic diagram in the data cockpit based on each basic infection number so as to predict cascade diffusion probability of safety risks.
10. The method of claim 1, wherein generating a pre-deployment policy based on an agent identity and an access boundary map comprises: acquiring API calling frequency and resource consumption time period of an agent to be deployed so as to construct a multidimensional ecological niche feature; calculating the feature overlapping degree of the multidimensional ecological niche features and the operated intelligent agent based on a preset ecological niche overlapping index model; If the feature overlapping degree is larger than a preset overlapping threshold value, generating an ecological niche competition phase diagram in the data cockpit, and refusing to generate a pre-deployment strategy containing the intelligent agent to be deployed.

Description

Intelligent tool box full life cycle control method based on data cockpit Technical Field The invention relates to the technical field of artificial intelligence and micro-service operation and maintenance, in particular to an intelligent tool box full life cycle management and control method based on a data cockpit. Background Large language model agents have been widely packaged as intelligent toolkits, deeply integrated into enterprise-level microservice architecture and automation business links. In a complex distributed computing environment, the coordination of multiple agents for executing high concurrency tasks becomes normal, which requires that the underlying management and control system not only guarantee the robust scheduling of computing resources, but also precisely govern the access boundaries, execution logic and state evolution in the full life cycle of the agents so as to maintain the high availability and service continuity of the global system. At present, a conventional life cycle management and control scheme for an intelligent agent or a micro-service cluster mainly depends on a role-based access control model to carry out static permission division, and maintains operation by combining a gateway current-limiting fusing mechanism and basic physical resource alarms. At the level of exception capture and node retirement, the industry typically relies on preset regular expressions to extract structured error reporting logs, and once rule thresholds are triggered, a forced termination instruction is directly issued to instantaneously cut off container processes and associated network session connections. However, the above solution has significant technical limitations. Firstly, the traditional monitoring system based on the discrete error code is almost invalid for unstructured natural language output of a large model, and real-time feature extraction and quantitative instability identification of hidden logic illusion and hidden 'cognitive drift' cannot be carried out at all. Secondly, in a mass waterfall flow log, the system is difficult to intuitively distinguish isolated local errors from cascading errors which are enough to cause global avalanche, and the capability of carrying out prior infection wave surface prediction on override detection and malicious injection is also lacking. Finally, for the 'one-cut' forced fusing of abnormal intelligent bodies, large-scale link suspension and distributed transaction rollback are extremely easy to be initiated in a highly-coupled micro-service network, and a soft landing retirement control mechanism with toughness compensation is lacked, so that the existing architecture is extremely fragile when the special unstructured semantic risk of a large model is handled. Disclosure of Invention In order to overcome the defects in the prior art, the embodiment of the invention provides an intelligent tool box full life cycle control method based on a data cockpit, which is used for constructing an identity and access boundary map for front-end control, mapping large model natural language output into a continuous medium field and calculating hydrodynamic parameters to visually and accurately identify cognitive drift, and simultaneously implementing flexible retirement on abnormal intelligent bodies by combining a track attenuation resistance function so as to solve the technical problems that the hidden logic illusion of the intelligent bodies is difficult to quantitatively capture in real time in the traditional discrete log monitoring and the cascade avalanche of a micro-service architecture is extremely easy to be caused by forced fusing of abnormal nodes. In order to achieve the above purpose, the present invention provides the following technical solutions: A full life cycle management and control method of an intelligent tool box based on a data cockpit comprises the following steps of constructing an agent identity and an access boundary map based on multi-source system interaction data, generating a pre-deployment strategy based on the agent identity and the access boundary map, issuing the strategy to a target computing node to control a plurality of agents to execute an automation task, acquiring a natural language output sequence and mapping the output sequence into a continuous medium field in the process of executing the automation task by the plurality of agents, calculating fluid dynamics parameters of the continuous medium field containing entropy coupling semantic Reynolds numbers, inputting the fluid dynamics parameters into the data cockpit to render a flow field distribution view, and if cognitive drift exists in the target agents according to the flow field distribution view, injecting the system resistance into the target agents based on a preset track attenuation resistance function to control the target agents to execute decommissioning operation. In a preferred embodiment, the method for calculating the hydrodynamic paramet