A Constraint Alignment Method for Intelligent Agent Skill Invocation Based on Industrial Knowledge Graph

By using a hard constraint interception layer based on industrial knowledge graphs, the problem of insufficient constraint processing in intelligent agent skill invocation is solved, enabling fast and effective constraint verification and simulation resource optimization, and reducing the cost of retry loops and hard-coded maintenance.

CN123070338APending Publication Date: 2026-11-10LONG SIYUN (WUXI) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611450471.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies lack a unified and evolvable constraint expression when large language model agents invoke skills, which cannot effectively handle equation coupling constraints, resulting in wasted simulation resources and retry loops, and the review process is lagging and cannot prevent errors in advance.

Method used

A hard constraint interception layer is constructed based on an industrial knowledge graph. By mapping the boundaries of physical equations and process specification data, a dynamic constraint solver is generated to perform static solving and constraint propagation, record dependency chains, intercept unsatisfiable instructions bidirectionally, and generate constraint violation feedback prompts to guide the agent to re-reason.

Benefits of technology

It achieves millisecond-level constraint verification, avoids invalid simulations, reduces retry loops, improves simulation resource utilization, continuously evolves constraint benchmarks, reduces hard-coded maintenance lag, and improves design efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070338A_ABST
    Figure CN123070338A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent agent skill call constraint alignment method based on an industrial knowledge graph. This method constructs an industrial knowledge graph with design parameter entities as nodes and constraint relationships as edges, based on physical equation boundaries and process specification data, mapping it to the variable domain and constraint conditions of a dynamic constraint solver. Skill call instructions carrying candidate design parameters are intercepted by the intelligent agent and substituted into the solver for static solution through constraint propagation. If the variable domain converges to an empty set, it is locally intercepted. The minimum conflict constraint set that has been violated is located through dependency chain backtracking, and its flow to the execution and review sides is blocked bidirectionally. Constraint violation feedback prompts containing violation parameters, constraint boundary values, and feasible domain directions are generated to prompt the intelligent agent to re-reason. If the conditions are met, execution is allowed, and the review results are written back to the graph to update the constraint weights. This invention performs pre-verification before the instruction is issued, eliminating unnecessary simulation consumption, reducing the review burden, and avoiding blind retries.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and industrial software technology, and in particular to a method for aligning intelligent agent skill invocation constraints based on industrial knowledge graphs. Background Technology

[0002] The application of large language model intelligent agents in industrial design automation has taken shape: after undertaking a design task, the agent autonomously plans the steps, generates skill invocation instructions carrying candidate design parameters, calls simulation and design services encapsulated as skills to complete parameter writing and performance verification, and iteratively infers based on the execution results. Existing technologies mainly employ three architectures for this application approach.

[0003] The first type is an unverified architecture where the agent directly invokes skills. Its shortcomings stem from the generation mechanism of large language models: the model generates parameter values ​​by sampling word probabilities, and there are no explicit variables of physical equations and process boundaries internally. There is no consistency guarantee between parameters and feasible domains; the skill server only performs type and format verification, and non-compliant parameters are executed without question. The simulation computing power of tens of minutes to several hours is wasted in a single instance, and the erroneous results are also used as context feedback, causing errors to accumulate and amplify along the inference rounds.

[0004] The second approach is a post-processing manual review architecture. Its drawbacks are: review lags behind parameter generation, and non-compliant parameters have already entered the simulation queue before review approval, making it impossible to prevent such errors beforehand; the review benchmark is static process documents, and there is a time lag in updating standards after revisions; review conclusions exist as natural language annotations, which cannot be distilled into machine-executable verification rules, leading to repeated occurrences of similar errors.

[0005] The third type is a pre-coded validation script. Its drawbacks are: constraints are hard-coded with scattered conditional statements, lacking relational expression, and new constraints require modification and redeployment of the code; the script can only perform single-parameter independent threshold judgments and cannot handle joint constraints coupled with equations; after interception, it only returns an error code without carrying boundary values ​​and correction directions, causing the agent to blindly retry, forming a repetitive retry loop.

[0006] What the above architectures all lack is a pre-verification mechanism that, based on a unified and evolvable constraint expression, can handle equation coupling constraints before instructions are issued and provides correction directions after interception. Summary of the Invention

[0007] (a) Technical problems to be solved The technical problem to be solved by the present invention is to provide a method for aligning intelligent agent skill invocation constraints based on industrial knowledge graph.

[0008] (II) Technical Solution A method for aligning intelligent agent skill invocation constraints based on an industrial knowledge graph includes the step of obtaining skill invocation instructions generated by the intelligent agent for an industrial design task, wherein the skill invocation instructions carry candidate design parameters, and further includes: S1. Construct a hard constraint interception layer: Obtain physical equation boundary data and process specification data, construct an industrial knowledge graph, wherein the industrial knowledge graph uses design parameter entities as nodes and the constraint relationships between design parameter entities as edges; map the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver, wherein the nodes are mapped to variables, the value boundaries corresponding to the nodes are mapped to the variable domain of the variables, and the edges are mapped to the constraint conditions applied to the corresponding variables; S2. Constraint Propagation: Intercept the skill call instruction, substitute the candidate design parameters carried by the skill call instruction into the dynamic constraint solver for static solution, perform arc consistency check on each constraint condition, iteratively delete values ​​in the variable domain that are incompatible with the constraint condition, until the variable domain of each variable no longer changes or the variable domain of any variable is empty, and the variable domain of each variable converges through constraint propagation. S3. Two-way interception judgment: During the constraint propagation process, record the constraint conditions and related variables that are deleted each time a value is taken, forming a dependency chain; judge whether the static solution result can be satisfied. If the variable domain of any variable converges to an empty set, it is determined that it cannot be satisfied. Intercept the skill call instruction locally, backtrack along the dependency chain, extract the minimum conflict constraint set that leads to the empty domain, determine the constraint items in the minimum conflict constraint set as the violated constraint items, and block the submission of the skill call instruction to the skill execution side and the flow of the candidate design parameters to the review side. S4. Constraint violation feedback prompt generation: Based on the association path of the violated constraint item in the industrial knowledge graph, a constraint violation feedback prompt is generated. The constraint violation feedback prompt includes the violation parameter, the constraint boundary value corresponding to the violated constraint item, and the feasible region direction. The feasible region direction is determined according to the comparison relationship between the violation value and the constraint boundary value. The correction goal is to cross the constraint boundary while retaining the boundary margin. The constraint violation feedback prompt is sent to the agent so that the agent can re-infer and generate a new skill call instruction based on the constraint violation feedback prompt, and then return to step S2. S5. If the static solution result is satisfactory, then the skill call instruction is allowed to be executed on the skill execution side, the execution result is obtained, the execution result is reviewed afterward, and the review result is written back to the industrial knowledge graph to update the constraint weight of the corresponding constraint relationship.

[0009] Preferably, in step S1, the edge carries a constraint expression, constraint source identifier, constraint weight, and effective time. The types of the edge include lower limit constraint, upper limit constraint, interval constraint, equation coupling constraint, and mutual exclusion constraint. The equation coupling constraint is mapped to a constraint condition that uses physical equations as expressions and acts on multiple variables.

[0010] Preferably, in step S1, mapping the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver includes: listening to change events of the industrial knowledge graph, recompiling the variable domain and constraint conditions involved in the changed subgraph, merging the recompiled result with the unchanged part, and generating a new versioned constraint set carrying version difference records.

[0011] Preferably, in step S4, generating constraint violation feedback prompts includes: retrieving associated design parameters affected by the violated constraint along the associated path, and generating constraint violation feedback prompts that include the violation parameter, the constraint boundary value, the feasible domain direction, and the adjustment priority of each associated design parameter; or, using the violated constraint as the search key, retrieving passing parameter combinations under the same constraint in historical approved cases, and attaching the retrieved passing parameter combinations as reference examples to the constraint violation feedback prompts.

[0012] Preferably, step S4 further includes: counting the number of re-inference attempts of the agent; when the number of re-inference attempts reaches a preset retry limit, terminating the loop that returns to step S2, and transferring the candidate design parameters and the violated constraints to the manual arbitration queue.

[0013] To address the aforementioned technical problems, this invention also provides an intelligent agent skill invocation constraint alignment device based on an industrial knowledge graph, comprising: The graph construction module is configured to acquire physical equation boundary data and process specification data, and construct an industrial knowledge graph, wherein the industrial knowledge graph uses design parameter entities as nodes and the constraint relationships between design parameter entities as edges. The constraint compilation module is configured to map the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver, wherein the nodes are mapped to variables, the value boundaries corresponding to the nodes are mapped to the variable domain of the variables, and the edges are mapped to the constraint conditions applied to the corresponding variables. The interception module is configured to intercept skill invocation instructions generated by the intelligent agent for an industrial design task, the skill invocation instructions carrying candidate design parameters; The constraint solving module is configured to substitute the candidate design parameters into the dynamic constraint solver for static solving, perform arc consistency checks on each constraint condition, iteratively delete values ​​in the variable domain that are incompatible with the constraint condition, until the variable domain of each variable no longer changes or the variable domain of any variable is empty, and record the constraint conditions and related variables on which each value deletion depends during the constraint propagation process, forming a dependency chain; The interception and judgment module is configured to determine whether the static solution result is satisfactory. If the variable domain of any variable converges to an empty set, it is determined that it is not satisfactory. The skill call instruction is intercepted locally, the dependency chain is traced back, the minimum conflict constraint set that leads to the empty domain is extracted, the constraint items in the minimum conflict constraint set are identified as the violated constraint items, and the submission of the skill call instruction to the skill execution side and the flow of the candidate design parameters to the review side are blocked. The prompt word generation module is configured to generate constraint violation feedback prompt words based on the association path of the violated constraint item in the industrial knowledge graph, and send the constraint violation feedback prompt words to the intelligent agent. The constraint violation feedback prompt words include violation parameters, the constraint boundary value corresponding to the violated constraint item, and the feasible domain direction. The execution review module is configured to allow the skill call instruction to be executed on the skill execution side when the static solution result is satisfactory, obtain the execution result, perform post-execution review on the execution result, and write the review result back to the industrial knowledge graph to update the constraint weights of the corresponding constraint relationship.

[0014] Preferably, when the execution audit module updates the constraint weights, if the audit result is passed, the constraint weight of the corresponding constraint relationship is increased; if the audit result is rejected, the corresponding constraint relationship is marked and the constraint boundary review process is triggered. The constraint compilation module is also used to listen for change events of the industrial knowledge graph, recompile the variable domains and constraint conditions involved in the changed subgraphs, and generate a new version of the versioned constraint set.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described intelligent agent skill invocation constraint alignment method based on industrial knowledge graph.

[0016] (III) Beneficial Effects 1. By substituting candidate design parameters into the dynamic constraint solver for millisecond-level constraint propagation static solution before issuing instructions, non-compliant parameters are identified before entering the simulation, thereby eliminating invalid simulation computing power consumption. The time for a single verification is on the order of milliseconds.

[0017] 2. By blocking unsatisfactory instructions from entering the simulation queue through bidirectional interception, simulation computing resources are allocated only to satisfactory instructions, avoiding the invalid occupation of computing resources; the review side only accesses parameters that have passed pre-verification, reducing the overhead of invalid processing of unsatisfactory parameters in the review process.

[0018] 3. By carrying violation parameters, constraint boundary values, and feasible region directions through constraint violation feedback prompts, the re-inference is transformed from blind retries to targeted searches. In the actual test of the implementation, the average number of retry rounds was reduced from 3.4 rounds to 1.3 rounds, thereby avoiding the amplification of overhead from repeated retry loops.

[0019] 4. By uniformly expressing constraints through graphs and updating constraint weights by writing back the audit results, the constraint benchmark can be continuously evolved, thereby eliminating the drift in verification caliber caused by the lag in maintenance of hard-coded scripts.

[0020] 5. By generating versioned constraint sets through incremental compilation, the granularity of solving the model is reduced to the change subgraph, thereby reducing compilation overhead and maintaining baseline timeliness. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the intelligent agent skill invocation constraint alignment method provided in an embodiment of the present invention.

[0022] Figure 2 The timing diagram of interactions between the parties is provided for embodiments of the present invention.

[0023] Figure 3 This is a schematic diagram of the constraint propagation and conflict localization process provided in an embodiment of the present invention.

[0024] Attached Figure Labeling Reference Table: 101, Graph Construction Module, 102, Constraint Compilation Module, 103, Call Interception Module, 104, Constraint Solving Module, 105, Interception Judgment Module, 106, Prompt Word Generation Module, 107, Execution Review Module, 201, Agent, 202, Hard Constraint Interception Layer, 203, Skill Execution Side, 204, Review Side. Detailed Implementation

[0025] The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection. The execution entity in this embodiment is the hard constraint interception layer 202 deployed in the industrial design automation system. The interception layer is located between the intelligent agent 201 and the skill execution side 203 and includes modules 101 to 107.

[0026] Combination Figure 1 The overall process of the method consists of steps S1 to S5, where S2 to S4 constitute a pre-verification closed loop for each round of skill invocation, and S5 is executed after each release.

[0027] Step S1 is executed by the graph construction module 101 and the constraint compilation module 102. The graph construction module 101 extracts design parameter entities as nodes from the physical equation boundary data and process specification data, categorizing them into four types: geometric, material, process, and performance. It also extracts constraint relationships as edges. The constraint compilation module 102 maps nodes to variables, node value boundaries to variable domains, and edges to constraint conditions. After compilation, the static solution problem is mathematically represented as a constraint satisfaction problem, its elements being a set of variables X, a set of variable domains D, and a set of constraint conditions C. The goal is to determine whether there exists a set of assignments that makes all constraint conditions true.

[0028] Step S2 is executed by the interception module 103 and the constraint solving module 104. The interception module 103 intercepts instructions at the instruction exit of the agent 201, parses the skill identifier and candidate design parameters, and assigns values ​​to the corresponding variables. The constraint solving module 104 performs constraint propagation: for the constraint condition cij between variables xi and xj, if the value a in variable domain Di has a value b in variable domain Dj that makes cij true, then the value a is retained; otherwise, the value a is deleted from Di; the iteration continues until each variable domain no longer changes or any variable domain is empty.

[0029] The above arc consistency constraint propagation uses the AC-3 algorithm, whose worst-case time complexity is O(e·d). 3 ), where e is the number of constraints and d is the maximum number of values ​​in the variable domain; in the hardware environment of this embodiment (Intel Xeon Gold6330 processor, 2.0 GHz, 256 GB memory), the actual measured time for a single static solution with 10,000 constraints is about 8.6 milliseconds. The millisecond-level algorithm overhead replaces the invalid simulation of tens of minutes, thereby improving the efficiency of computer computing resources.

[0030] Combination Figure 3 The conflict localization process is as follows: Constraint solving module 104 records the constraint conditions that triggered the deletion and related variables to form a dependency chain each time a value is deleted; after detecting a spatial domain, interception and judgment module 105 backtracks along the dependency chain, starting from the last deletion that directly caused the spatial domain, collecting dependency records level by level, and backtracking from the end of the dependency chain, attempting to remove each constraint condition in the set and re-executing constraint propagation. If the spatial domain no longer reappears after removing a constraint condition, the constraint condition is retained in the set; if the spatial domain still reappears after removal, it is removed. The final set of constraint conditions where the spatial domain no longer reappears after deleting any constraint condition is the minimum conflict constraint set. For equation coupling constraints, constraint solving module 104 uses interval arithmetic, treating the physical equation as an interval function iteratively shrinking the variable interval boundary. An empty interval is considered unsatisfactory. The minimum conflict constraint set consists of the constraint terms and boundary constraint terms involved in the equation.

[0031] Step S3 is executed by the interception and judgment module 105: if the condition cannot be met, the local termination instruction is terminated and bidirectional blocking is implemented; if the condition can be met, the task is transferred to the execution review module 107.

[0032] Step S4 is executed by the prompt word generation module 106. The logic for determining the feasible region direction is as follows: Let the violation value be v, and the constraint boundary value of the violated constraint term be b. When the value is below the lower limit, the direction is the increasing direction, and the correction target is not less than b. When the value is above the upper limit, the opposite is true. The suggested correction step size λ is determined by the following formula, where ε is the boundary margin: λ=|bv|+ε The prompt word generation module 106 organizes the violation parameters and their values, constraint boundary values, and feasible region directions into constraint violation feedback prompt words and sends them to the agent 201. The agent re-infers and generates new instructions, and the process returns to step S2. The interception judgment module 105 counts the number of re-inference attempts. When the preset retry limit is reached, the closed loop is terminated and the system is transferred to the manual arbitration queue. In this embodiment, the retry limit is three times, which is a value set through simulation verification.

[0033] Step S5 is executed by the audit module 107. For example... Figure 2 As shown, satisfyable instructions are forwarded to the skill execution side 203 for execution via the hard constraint interception layer 202. The execution result is reviewed by the review side 204, and the review conclusion is written back to the graph and the constraint weights are updated according to the following formula: w′=w+η·δ Where w is the constraint weight before the update, w′ is the constraint weight after the update, η is the update step size coefficient, and δ takes a positive value when the review is passed and a negative value when it is rejected. In this embodiment, η is 0.05, and δ is +1 or -1, which are the parameter values ​​used in simulation verification. When the review is rejected, the corresponding constraint relationship is marked and the constraint boundary review process is triggered.

[0034] The uses of constraint weights in subsequent stages include: First, when the minimum conflict constraint set contains multiple constraint items, the interception judgment module 105 sorts the constraint items from high to low according to their constraint weights, and presents the violated constraint items in the constraint violation feedback prompts in that order, so that the agent 201 prioritizes adjusting the parameters corresponding to high-weight constraints; Second, the candidate design parameters and violated constraint items transferred to the manual arbitration queue are prioritized for arbitration processing according to their constraint weights from high to low, with high-weight constraint violation cases being allocated review resources first; Third, when the constraint compilation module 102 performs incremental compilation, it determines the verification range based on the constraint weights, performs a full re-verification of the variable domains involved in constraint relationships with constraint weights higher than a preset threshold, and only marks the change impact for constraint relationships with weights lower than the threshold, thereby controlling compilation overhead in scenarios where specifications are frequently revised.

[0035] Example 1: Implementation of the basic scheme for die casting wall thickness design. In a die casting design task, the drawing pre-sets nodes such as wall thickness, material grade, and projected area. A pre-set lower limit constraint for wall thickness is given by the expression "wall thickness not less than 3 mm," derived from the enterprise's casting process specifications. Additionally, a pre-set constraint is established between wall thickness, material yield strength, and design load through a strength equation coupled with the stress equation. The constraint compilation module 102 compiles the entire code to generate a versioned constraint set. The agent 201 infers a wall thickness of 2 mm and generates an instruction to invoke the design skill. The interception module 103 intercepts the instruction. After the constraint solving module 104 assigns a value, arc consistency propagation is performed. If the assigned value is incompatible with the lower limit constraint, the wall thickness variable domain converges to empty, and the constraint is deemed unsatisfactory. The single solution takes milliseconds. The interception judgment module 105 identifies the violated constraint as the lower limit constraint for wall thickness and bidirectionally blocks instruction submission and flow. The prompt word generation module 106 generates a constraint violation feedback prompt word: "The wall thickness of 2 mm is lower than the casting process lower limit of 3 mm. Please re-infer along the increasing direction within the feasible domain of not less than 3 mm and not more than 5 mm." Agent 201 re-infers and outputs a wall thickness of 3.5 mm. Once the new instruction is satisfied, execution is allowed. The review side 204 verifies and approves the instruction, and the constraint weight of the lower wall thickness constraint is increased according to the above formula. In this embodiment, the entire process from capturing the prompt word to generating the constraint violation feedback prompt word takes less than ten milliseconds in actual testing, which is expected to avoid approximately forty minutes of invalid structural simulation.

[0036] Example 2: Optimization Scheme Implementation for Multi-Parameter Coupling Scenarios. In a task to lightweight a motor housing, there are tens of thousands of constraints and the specifications are revised monthly. The constraint compilation module 102 uses incremental compilation, only recompiling the changed subgraphs and retaining version difference records. Agent 201 generates candidate parameters: wall thickness 2.5 mm, material grade A, and load condition B. The constraint solving module 104 uses interval propagation for the coupled constraints of the strength equation and maintains the deletion dependency chain. The propagation result is that the wall thickness interval is empty. The interception judgment module 105 backtracks and extracts the minimum conflict constraint set, which contains only the coupled constraints of the strength equation and the lower limit constraint of the wall thickness. The prompt word generation module 106 retrieves the affected associated parameters as material grade and stiffener arrangement parameters, and generates a multi-parameter joint correction prompt word: the combination of wall thickness 2.5 mm and material grade A does not satisfy the strength equation, the feasible region of wall thickness is not less than 3.2 mm, and the adjustment priority is wall thickness, material grade, and stiffener arrangement in that order; and attaches the combination of wall thickness 3.5 mm and material grade C from the historical successful cases as a reference example. Agent 201 outputs a parameter combination that satisfies all constraints after a single re-inference. Simulation verification shows that the measured time for solving a single constraint in this embodiment is approximately 8.6 milliseconds. Before the introduction of this interception layer, the average number of invalid simulations per month due to parameter violations in this scenario was approximately 37; after its introduction, this number decreased to 0. The average number of retry rounds decreased from 3.4 rounds to 1.3 rounds, and the convergence rate of a single re-inference reached over 70%. In this embodiment, the average size of the minimum conflict constraint set is two to three.

[0037] Device Implementation: The hard constraint interception layer 202 is deployed as a local component within the runtime node of agent 201, with each module interacting within its process. An alternative deployment is an independent cloud-based interception service, where agents interact via a message bus, suitable for multi-agent concurrency and centralized constraint set management. For the alternative solver implementation, a satisfiability module theory solver can be used to compile the graph into first-order logic formulas for solving, or a custom propagation engine can be implemented based on an interval arithmetic library. For the alternative graph construction implementation, a document parsing pipeline can be introduced, using named entity recognition to extract parameter entities, and using relation extraction models to identify constraint semantics and generate graph edges to be confirmed. These edges are then stored in the database after confirmation by a knowledge engineer.

[0038] From the perspective of system operation, the pre-verification closed loop ensures that all instructions received by the skill execution side 203 pass the constraint solution, and no more parameter violation tasks appear in the simulation queue; all parameters encountered by the review side 204 meet the current constraint benchmarks; the graph constraint weights are updated on a rolling basis with the review conclusions, and adding new constraints only requires adding nodes and edges to the graph and taking effect through incremental compilation, without modifying the interception layer code.

[0039] 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 and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for aligning intelligent agent skill invocation constraints based on industrial knowledge graphs, comprising the step of obtaining skill invocation instructions generated by the intelligent agent for an industrial design task, wherein the skill invocation instructions carry candidate design parameters, characterized in that, Also includes: S1. Construct a hard constraint interception layer: Obtain physical equation boundary data and process specification data, construct an industrial knowledge graph, wherein the industrial knowledge graph uses design parameter entities as nodes and the constraint relationships between design parameter entities as edges; map the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver, wherein the nodes are mapped to variables, the value boundaries corresponding to the nodes are mapped to the variable domain of the variables, and the edges are mapped to the constraint conditions applied to the corresponding variables; S2. Constraint Propagation: Intercept the skill call instruction, substitute the candidate design parameters carried by the skill call instruction into the dynamic constraint solver for static solution, perform arc consistency check on each constraint condition, iteratively delete values ​​in the variable domain that are incompatible with the constraint condition, until the variable domain of each variable no longer changes or the variable domain of any variable is empty, and the variable domain of each variable converges through constraint propagation. S3. Two-way interception judgment: During the constraint propagation process, record the constraint conditions and related variables that are deleted each time a value is taken, forming a dependency chain; judge whether the static solution result can be satisfied. If the variable domain of any variable converges to an empty set, it is determined that it cannot be satisfied. Intercept the skill call instruction locally, backtrack along the dependency chain, extract the minimum conflict constraint set that leads to the empty domain, determine the constraint items in the minimum conflict constraint set as the violated constraint items, and block the submission of the skill call instruction to the skill execution side and the flow of the candidate design parameters to the review side. S4. Constraint violation feedback prompt generation: Based on the association path of the violated constraint item in the industrial knowledge graph, a constraint violation feedback prompt is generated. The constraint violation feedback prompt includes the violation parameter, the constraint boundary value corresponding to the violated constraint item, and the feasible region direction. The feasible region direction is determined according to the comparison relationship between the violation value and the constraint boundary value. The correction goal is to cross the constraint boundary while retaining the boundary margin. The constraint violation feedback prompt is sent to the agent so that the agent can re-infer and generate a new skill call instruction based on the constraint violation feedback prompt, and then return to step S2. S5. If the static solution result is satisfactory, then the skill call instruction is allowed to be executed on the skill execution side, the execution result is obtained, the execution result is reviewed afterward, and the review result is written back to the industrial knowledge graph to update the constraint weight of the corresponding constraint relationship.

2. The intelligent agent skill invocation constraint alignment method based on industrial knowledge graph according to claim 1, characterized in that, In step S1, the edge carries a constraint expression, constraint source identifier, constraint weight and effective time. The types of the edge include lower limit constraint, upper limit constraint, interval constraint, equation coupling constraint and mutual exclusion constraint. The equation coupling constraint is mapped to a constraint condition that is expressed by a physical equation and acts on multiple variables.

3. The intelligent agent skill invocation constraint alignment method based on industrial knowledge graph according to claim 1, characterized in that, In step S1, mapping the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver includes: listening to change events of the industrial knowledge graph, recompiling the variable domain and constraint conditions involved in the changed subgraph, merging the recompiled result with the unchanged part, and generating a new versioned constraint set carrying version difference records.

4. The intelligent agent skill invocation constraint alignment method based on industrial knowledge graph according to claim 1, characterized in that, In step S4, generating constraint violation feedback prompts includes: retrieving associated design parameters affected by the violated constraint along the associated path, and generating constraint violation feedback prompts that include the violation parameter, the constraint boundary value, the feasible domain direction, and the adjustment priority of each associated design parameter; or, using the violated constraint as the search key, retrieving passing parameter combinations under the same constraint in historical approved cases, and attaching the retrieved passing parameter combinations as reference examples to the constraint violation feedback prompts.

5. The intelligent agent skill invocation constraint alignment method based on industrial knowledge graph according to claim 1, characterized in that, Step S4 further includes: counting the number of re-inference attempts of the agent; when the number of re-inference attempts reaches a preset retry limit, terminating the loop that returns to step S2, and transferring the candidate design parameters and the violated constraints to the manual arbitration queue.

6. A smart agent skill invocation constraint alignment device based on industrial knowledge graph, characterized in that, include: The graph construction module is configured to acquire physical equation boundary data and process specification data, and construct an industrial knowledge graph, wherein the industrial knowledge graph uses design parameter entities as nodes and the constraint relationships between design parameter entities as edges. The constraint compilation module is configured to map the industrial knowledge graph to the variable domain and constraint conditions of a dynamic constraint solver, wherein the nodes are mapped to variables, the value boundaries corresponding to the nodes are mapped to the variable domain of the variables, and the edges are mapped to the constraint conditions applied to the corresponding variables. The interception module is configured to intercept skill invocation instructions generated by the intelligent agent for an industrial design task, the skill invocation instructions carrying candidate design parameters; The constraint solving module is configured to substitute the candidate design parameters into the dynamic constraint solver for static solving, perform arc consistency checks on each constraint condition, iteratively delete values ​​in the variable domain that are incompatible with the constraint condition, until the variable domain of each variable no longer changes or the variable domain of any variable is empty, and record the constraint conditions and related variables on which each value deletion depends during the constraint propagation process, forming a dependency chain; The interception and judgment module is configured to determine whether the static solution result is satisfactory. If the variable domain of any variable converges to an empty set, it is determined that it is not satisfactory. The skill call instruction is intercepted locally, the dependency chain is traced back, the minimum conflict constraint set that leads to the empty domain is extracted, the constraint items in the minimum conflict constraint set are identified as the violated constraint items, and the submission of the skill call instruction to the skill execution side and the flow of the candidate design parameters to the review side are blocked. The prompt word generation module is configured to generate constraint violation feedback prompt words based on the association path of the violated constraint item in the industrial knowledge graph, and send the constraint violation feedback prompt words to the intelligent agent. The constraint violation feedback prompt words include violation parameters, the constraint boundary value corresponding to the violated constraint item, and the feasible domain direction. The execution review module is configured to allow the skill call instruction to be executed on the skill execution side when the static solution result is satisfactory, obtain the execution result, perform post-execution review on the execution result, and write the review result back to the industrial knowledge graph to update the constraint weights of the corresponding constraint relationship.

7. The intelligent agent skill invocation constraint alignment device based on industrial knowledge graph according to claim 6, characterized in that, When the execution review module updates the constraint weights, if the review result is passed, the constraint weight of the corresponding constraint relationship is increased; if the review result is rejected, the corresponding constraint relationship is marked and the constraint boundary review process is triggered. The constraint compilation module is also used to listen for change events of the industrial knowledge graph, recompile the variable domains and constraint conditions involved in the changed subgraphs, and generate a new version of the versioned constraint set.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent agent skill invocation constraint alignment method based on industrial knowledge graph as described in any one of claims 1 to 5.