Agent-based external operation equipment manufacturing full life cycle collaborative control method

Abstract

The application provides a full life cycle collaborative control method for equipment manufacturing based on an agent external operation, comprising the following steps: S1, receiving a large and complex equipment manufacturing task; S2, identifying an engineering context; S3, generating a runtime task description; S4, generating a budget-aware memory snapshot; S5, generating a skill runtime snippet; S6, generating a protocol invocation constraint; S7, constructing an externalization module collaborative control graph; S8, obtaining a structured action intent; S9, performing a check; and S10, generating a structured execution trajectory and a write-back. The application constructs a landable industrial agent governance mechanism, and realizes a safe conversion from general reasoning to controlled manufacturing execution.

Description

Technical Field

[0001] This invention relates to the fields of industrial artificial intelligence, intelligent manufacturing, and runtime control technology of large language model intelligent agents, and in particular to collaborative control methods, systems, electronic devices, and storage media for the entire life cycle of equipment manufacturing based on intelligent external operation. Background Technology

[0002] The manufacturing of large and complex equipment (such as ships, aircraft, and generators) is characterized by long project cycles, complex professional collaboration, strict version control, and long quality responsibility chains. Currently, industrial systems (such as PLM, MES, ERP, and QMS) operate independently, with inconsistent interface protocols, permission models, and data standards, resulting in low efficiency in cross-system collaboration and difficulties in error tracing.

[0003] In recent years, intelligent agent technology based on large language models has been attempted for enterprise knowledge question answering, tool invocation, and workflow orchestration, but the following problems still exist: Insufficient continuity of engineering status: It is difficult to maintain stable manufacturing status across stages and versions; Experience not being capitalized: Experience in design, process, and exception handling has not been systematically reused; Lack of governance in API calls: Industrial system API calls lack a unified mechanism for permissions, risks, and auditing; Contextual budget conflict: Multiple sources of data can easily exceed the model's processing capacity, and key information is overwhelmed; High-risk actions lack control: such as modifying the BOM and closing quality issue reports, which lack approval and sandbox linkage. Lack of a closed-loop evolution mechanism: The execution results do not generate reusable experience feedback. Summary of the Invention

[0004] This invention provides a collaborative control method, system, electronic device, and storage medium for the entire lifecycle of equipment manufacturing based on intelligent externalized operation, aiming to achieve the following objectives: improve the consistency of cross-system engineering data calls by unifying index versions and manufacturing stages to avoid data mismatch; reduce the risk of state loss in long-cycle tasks by persistently storing engineering states and approval trajectories through externalized memory; improve the execution stability of complex process tasks by reducing reasoning errors through skill asset accumulation and gradual disclosure; enhance the security and auditability of industrial system calls through protocol gateways, permission verification, and approval gate mechanisms; reduce context redundancy and call costs by optimizing information loading through unified budget allocation and collaborative control charts; improve the closed-loop efficiency of anomaly handling by accelerating problem localization and resolution through trajectory write-back and experience reuse; and construct a feasible industrial intelligent agent governance mechanism to achieve a secure transition from general reasoning to controlled manufacturing execution.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation includes: Step S1: Receive the task of manufacturing large and complex equipment; Step S2: Identify the project context associated with the task and generate project context information; Step S3: Generate a runtime task description based on the project context information; Step S4: Based on the runtime task description, retrieve the externalized memory object from the externalized memory bank and generate a budget-aware memory snapshot; Step S5: Based on the runtime task description and the budget-aware memory snapshot, select skill asset objects from the skill asset library and generate skill runtime fragments; Step S6: Based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, determine the industrial protocol endpoint object from the industrial protocol endpoint library and generate protocol call constraints; Step S7: Construct an externalized module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraints, the context budget, the permission constraints, the approval strategy, and the risk level; Step S8: Based on the externalization module collaborative control diagram, generate the runtime context of the large language model agent, and input the runtime context into the large language model agent to obtain the structured action intent output by the large language model agent; Step S9: Perform verification on the structured action intent, the verification including at least one of protocol verification, permission verification, risk verification, approval verification and sandbox verification; Step S10: After the structured action intent passes the verification, the controlled operation is executed and a structured execution trajectory is generated; the structured execution trajectory is written back to the engineering externalization memory, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

[0006] In this specification, when generating the budget-aware memory snapshot in step S4, the method further includes: filtering the retrieval request based on user role, project permissions, and data security level; normalizing the retrieval results from different data sources into a unified format of externalized project memory objects; and performing a comprehensive scoring and sorting of the externalized project memory objects, prioritizing the retention of several memory objects with the highest comprehensive scores under the constraints of the context budget.

[0007] In this specification, when selecting a skill asset object and generating a skill runtime fragment in step S5, the method further includes: verifying whether the input data requirements of the candidate skill asset object have been met by the budget-aware memory snapshot or the runtime task description; determining the disclosure level of the target skill asset object based on the comprehensive score, the risk level, and the context budget, wherein the disclosure level includes at least a summary layer, a constraint layer, and a detailed rule layer; and trimming the operation steps, decision rules, prohibited items, and approval requirements in the target skill asset object into the skill runtime fragment.

[0008] In this specification, when determining the industrial protocol endpoint object and generating protocol call constraints in step S6, the method further includes: the industrial protocol gateway recalling candidate protocol endpoint objects based on system type, capability tags, and input / output modes; mapping internal data objects to the interface call format of the target industrial system; performing a comprehensive score on the candidate protocol endpoint objects; determining the target protocol endpoint object based on the comprehensive score and permission filtering; and generating the protocol call constraints, which at least include the allowed methods, parameter ranges, and whether approval or sandboxing is required.

[0009] In this specification, step S7, constructing the externalized module collaborative control diagram, specifically includes: writing the engineering externalized memory object in the budget-aware memory snapshot into the control diagram as a memory node; writing the skill asset object corresponding to the skill runtime fragment into the control diagram as a skill node; writing the industrial protocol endpoint object corresponding to the protocol call constraint into the control diagram as a protocol endpoint node; establishing dependency edges based on the correspondence between the input data requirements of the skill node and the evidence reference of the memory node; establishing call edges based on the correspondence between the call parameters of the protocol endpoint node and the output format of the skill node; and establishing approval gate edges and sandbox edges based on the risk level and the approval strategy.

[0010] In this specification, when generating the runtime context of the large language model agent in step S8, the method further includes: assembling the content summary in the budget-aware memory snapshot, the operation steps and decision rules in the skill runtime segment, and the input / output patterns in the protocol call constraints into the runtime context according to the allocation ratio of the context budget, based on the externalization module collaborative control diagram. The runtime context also includes prohibited items and approval requirements.

[0011] In this specification, the risk verification in step S9 further includes: calculating a risk score based on the action type risk, object sensitivity, lifecycle stage, user permission level, and historical failure rate; determining the execution mode based on the risk score, wherein the execution mode includes automatic execution for low risk, generating only suggestions or drafts for medium risk, requiring manual approval for execution for high risk, and allowing verification only in a sandbox or digital twin environment for extremely high risk.

[0012] In this specification, the sandbox verification in step S9 further includes: when the structured action intent is determined to require sandbox execution, the structured action intent is pre-executed in a read-only data copy, a digital twin simulation environment, or a container isolation environment; if the sandbox execution is successful and the result meets expectations, execution continues; if the sandbox execution fails or the result is abnormal, execution is rejected and the reason for failure is returned.

[0013] In this specification, when performing the controlled operation in step S10, the processing is based on the type of the structured action intent: when it is a query operation, the data is obtained by calling the read-only interface of the target system through the industrial protocol gateway; when it is an analysis operation, the simulation or computing service is called to generate an analysis report or solution suggestion; when it is an approval flow initiation operation, an approval draft is generated and submitted to the enterprise approval system; when it is a data modification operation, the controlled write is only performed by the industrial protocol gateway after manual approval and sandbox verification.

[0014] In this specification, when writing the structured execution trajectory back to the externalized engineering memory in step S10, the model call records, memory hit records, protocol call results, approval events and manual feedback in the structured execution trajectory are transformed into new externalized engineering memory objects, and the success or failure status of the execution result and the manual confirmation opinions are recorded.

[0015] In this specification, when writing the structured execution trajectory back to the skill asset library in step S10, the historical success rate and comprehensive score of the skill asset object are updated based on the skill call results, human feedback and final conclusion in the structured execution trajectory; when the skill asset object continuously receives low scores or fails in multiple similar tasks, its ranking weight is reduced or it is marked as requiring manual review.

[0016] In this specification, when writing the structured execution trajectory back to the industrial protocol endpoint library in step S10, the historical availability score and failure rate of the industrial protocol endpoint object are updated according to the protocol call results, failure reasons and compensation strategy execution status in the structured execution trajectory; when the number of consecutive failures of a certain protocol endpoint object exceeds the threshold, it is marked as unavailable or downgraded to a backup endpoint.

[0017] This specification also includes a context budget joint allocation step: when generating the runtime task description in step S3, the allocation ratio of the context budget among the externalized engineering memory, skill assets, protocol constraints, and output format is determined based on the task type, risk level, and historical recovery cost; in step S4, the budget-aware memory snapshot is cropped according to the allocation ratio; and in step S5, the disclosure level of the skill asset object is determined according to the allocation ratio.

[0018] This specification also includes lifecycle phase state machine maintenance steps: pre-defining the phase sequence and transformation relationship of the entire lifecycle of large and complex equipment manufacturing; when generating runtime task descriptions in step S3, determining the scope of accessible data sources, the set of callable protocol endpoints, the set of loadable skill assets, and the approval strategy based on the current lifecycle phase; when the structured action intent involves cross-phase operations, triggering phase adaptation verification, if the current phase does not allow the execution of the action, then refusing to execute or requiring manual confirmation.

[0019] In this specification, if any one of the protocol verification, permission verification, risk verification, approval verification, or sandbox verification in step S9 fails, the system refuses to execute the structured action intent and generates a rejection reason explanation; the rejection reason explanation and the structured action intent are then combined... Figure 1 It is recorded in the structured execution trajectory and written back to the externalized memory of the project.

[0020] In this specification, the large and complex equipment manufacturing task is at least one of the following tasks: design review task, process planning task, bill of materials and drawing consistency review task, design change impact analysis task, assembly interference analysis task, quality anomaly 8D analysis task, test fault location task, operation and maintenance fault knowledge feedback task, and production plan impact assessment task.

[0021] In this specification, the large language model agent does not directly access external industrial systems, and its output structured action intent must be verified in step S9 before it can be executed; the external industrial systems include at least one or more of the following: product lifecycle management system, manufacturing execution system, enterprise resource planning system, quality management system, computer-aided design system, and monitoring and data acquisition system.

[0022] A collaborative control system for the entire lifecycle of equipment manufacturing based on intelligent external operation is provided to implement the collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation described in any one of the above-mentioned methods. The collaborative control system for the entire lifecycle of equipment manufacturing based on intelligent external operation includes: The task access module is used to receive manufacturing tasks for large and complex equipment. The task parsing and equipment object identification module is used to identify the engineering context associated with the task and generate engineering context information; The runtime task description generation module is used to generate runtime task descriptions based on the project context information. The project externalization memory module is used to retrieve project externalization memory objects from the project externalization memory library according to the runtime task description, and generate a budget-aware memory snapshot based on the context budget; The skill asset management module is used to select skill asset objects from the skill asset library based on the runtime task description and the budget-aware memory snapshot, and generate skill runtime fragments. The industrial protocol gateway module is used to determine industrial protocol endpoint objects from the industrial protocol endpoint library based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, and to generate protocol call constraints. The externalization module collaborative control module is used to construct an externalization module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraint, context budget, permission constraint, approval strategy, and risk level. The large language model intelligent agent module is used to output structured action intentions based on the runtime context generated by the collaborative control diagram of the externalization module; The verification module is used to perform at least one of the following on the structured action intent: protocol verification, permission verification, risk verification, approval verification, and sandbox verification; The sandbox execution module is used to pre-execute structured action intents that need to be verified in a sandbox environment; The observation and trajectory write-back module is used to collect the execution trajectory after performing a controlled operation and write the execution trajectory back to the engineering externalized memory library, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

[0023] An electronic device includes at least one processor, at least one memory, and at least one communication interface. The memory stores a computer program, and when the processor executes the computer program, it implements the collaborative control method for the entire life cycle of equipment manufacturing based on intelligent external operation as described above.

[0024] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described above.

[0025] In summary, the present invention has at least the following beneficial effects: 1. Improve the consistency of cross-system engineering data retrieval. By jointly indexing the externalized engineering memory using equipment object ID, project number, configuration version, BOM version, process version, quality status, and manufacturing stage, errors caused by inconsistencies in drawing, BOM, process, quality, and test data versions can be reduced.

[0026] 2. Reduce the risk of state loss in long-cycle manufacturing tasks. By storing project state, change records, anomaly handling processes, approval trajectories, and maintenance feedback through externalized engineering memory, the large language model agent no longer needs to rely on the context of a single dialogue to maintain the long-term engineering state.

[0027] 3. Improve the stability of complex process tasks. By accumulating procedural knowledge such as design review, process review, assembly anomaly analysis, 8D quality management, test fault location, and design change impact analysis through skill assets, and by progressively disclosing this knowledge according to task stage, risk level, and contextual budget, the omissions and misjudgments caused by the agent's ad-hoc generation process can be reduced.

[0028] 4. Improve the security and auditability of industrial system calls. Through industrial protocol gateways, protocol verification, permission verification, risk verification, approval gates, and structured call logs, prevent unauthorized modifications to data in critical systems such as PLM, MES, QMS, and ERP by intelligent agents.

[0029] 5. Reduce context redundancy and invocation costs. By using an externalized module collaborative control diagram, the loading granularity of memory snapshots, skill details, protocol schemas, and inference constraints is uniformly determined, prioritizing the retention of information related to the current equipment object, manufacturing stage, and risk control within a limited context budget.

[0030] 6. Improve the efficiency of the manufacturing anomaly closed loop. Write the retrieval, analysis, suggestions, approval, execution results and human feedback in the anomaly handling process back into the externalized engineering memory, and update skill scores and protocol routing strategies so that verified experience can be reused in subsequent similar tasks.

[0031] 7. Establish a feasible governance mechanism for industrial intelligent agents during runtime. Through risk classification, sandbox execution, manual approval, rollback checkpoints, and trajectory auditing, achieve the transformation from general-purpose large language model reasoning capabilities to controlled industrial manufacturing task execution capabilities. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall architecture of a collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation.

[0033] Figure 2 This is a schematic diagram illustrating the steps of a collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation.

[0034] Figure 3 This is a schematic diagram of the collaborative control diagram of the externalized modules involved in this invention.

[0035] Figure 4 This is a schematic diagram illustrating the risk classification, approval gate, and sandbox execution process involved in this invention.

[0036] Figure 5 This is a schematic diagram of the structured execution trajectory write-back and closed-loop update involved in this invention.

[0037] Figure 6 This is a schematic diagram of the lifecycle stage state machine involved in this invention. Detailed Implementation

[0038] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0039] like Figure 1 As shown, this embodiment provides a collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation, including: Step S1: Receive the task of manufacturing large and complex equipment; Step S2: Identify the project context associated with the task and generate project context information; Step S3: Generate a runtime task description based on the project context information; Step S4: Based on the runtime task description, retrieve the externalized memory object from the externalized memory bank and generate a budget-aware memory snapshot; Step S5: Based on the runtime task description and the budget-aware memory snapshot, select skill asset objects from the skill asset library and generate skill runtime fragments; Step S6: Based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, determine the industrial protocol endpoint object from the industrial protocol endpoint library and generate protocol call constraints; Step S7: Construct an externalized module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraints, the context budget, the permission constraints, the approval strategy, and the risk level; Step S8: Based on the externalization module collaborative control diagram, generate the runtime context of the large language model agent, and input the runtime context into the large language model agent to obtain the structured action intent output by the large language model agent; Step S9: Perform verification on the structured action intent, the verification including at least one of protocol verification, permission verification, risk verification, approval verification and sandbox verification; Step S10: After the structured action intent passes the verification, the controlled operation is executed and a structured execution trajectory is generated; the structured execution trajectory is written back to the engineering externalization memory, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

[0040] In some embodiments, when generating a budget-aware memory snapshot in step S4, the method further includes: filtering the retrieval request based on user role, project permissions, and data security level; normalizing the retrieval results from different data sources into a unified format of externalized project memory objects; and performing a comprehensive scoring and ranking of the externalized project memory objects, prioritizing the retention of a number of memory objects with the highest comprehensive scores under the constraints of the context budget.

[0041] In some embodiments, when selecting a skill asset object and generating a skill runtime fragment in step S5, the method further includes: verifying whether the input data requirements of the candidate skill asset object have been met by the budget-aware memory snapshot or the runtime task description; determining the disclosure level of the target skill asset object based on the comprehensive score, the risk level, and the context budget, wherein the disclosure level includes at least a summary layer, a constraint layer, and a detailed rule layer; and trimming the operation steps, decision rules, prohibited items, and approval requirements in the target skill asset object into the skill runtime fragment.

[0042] In some embodiments, when determining the industrial protocol endpoint object and generating protocol call constraints in step S6, the method further includes: recalling candidate protocol endpoint objects by the industrial protocol gateway based on system type, capability label, and input / output mode; mapping internal data objects to the interface call format of the target industrial system; performing a comprehensive score on the candidate protocol endpoint objects; determining the target protocol endpoint object based on the comprehensive score and permission filtering; and generating the protocol call constraints, which at least include the allowed methods, parameter ranges, and whether approval or sandboxing is required.

[0043] In some embodiments, the construction of the externalization module collaborative control diagram in step S7 specifically includes: writing the engineering externalization memory object in the budget-aware memory snapshot into the control diagram as a memory node; writing the skill asset object corresponding to the skill runtime fragment into the control diagram as a skill node; writing the industrial protocol endpoint object corresponding to the protocol call constraint into the control diagram as a protocol endpoint node; establishing dependency edges based on the correspondence between the input data requirements of the skill node and the evidence reference of the memory node; establishing call edges based on the correspondence between the call parameters of the protocol endpoint node and the output format of the skill node; and establishing approval gate edges and sandbox edges based on the risk level and the approval strategy.

[0044] In some embodiments, when generating the runtime context of the large language model agent in step S8, the method further includes: assembling the content summary in the budget-aware memory snapshot, the operation steps and decision rules in the skill runtime fragment, and the input / output patterns in the protocol call constraints into the runtime context according to the allocation ratio of the context budget, based on the externalization module collaborative control diagram. The runtime context also includes prohibited items and approval requirements.

[0045] In some embodiments, the risk verification in step S9 further includes: calculating a risk score based on action type risk, object sensitivity, lifecycle stage, user permission level, and historical failure rate; and determining an execution mode based on the risk score, wherein the execution mode includes low-risk automatic execution, medium-risk only generating suggestions or drafts, high-risk requiring manual approval before execution, and extremely high-risk only allowing verification in a sandbox or digital twin environment.

[0046] In some embodiments, the sandbox verification in step S9 further includes: when the structured action intent is determined to require sandbox execution, pre-executing the structured action intent in a read-only data copy, a digital twin simulation environment, or a container isolation environment; if the sandbox execution is successful and the result meets expectations, then continuing execution; if the sandbox execution fails or the result is abnormal, then rejecting execution and returning the reason for failure.

[0047] In some embodiments, when performing a controlled operation in step S10, the operation is processed according to the type of the structured action intent: when it is a query operation, data is obtained by calling the read-only interface of the target system through the industrial protocol gateway; when it is an analysis operation, simulation or computing services are called to generate analysis reports or solution suggestions; when it is an approval flow initiation operation, an approval draft is generated and submitted to the enterprise approval system; when it is a data modification operation, the controlled write is only performed by the industrial protocol gateway after manual approval and sandbox verification.

[0048] In some embodiments, when writing the structured execution trajectory back to the externalized engineering memory in step S10, the model call records, memory hit records, protocol call results, approval events and manual feedback in the structured execution trajectory are transformed into new externalized engineering memory objects, and the success or failure status of the execution result and the manual confirmation opinions are recorded.

[0049] In some embodiments, when writing the structured execution trajectory back to the skill asset library in step S10, the historical success rate and comprehensive score of the skill asset object are updated based on the skill call results, human feedback and final conclusion in the structured execution trajectory; when the skill asset object continuously receives low scores or fails in multiple similar tasks, its ranking weight is reduced or it is marked as requiring human review.

[0050] In some embodiments, when writing the structured execution trajectory back to the industrial protocol endpoint library in step S10, the historical availability score and failure rate of the industrial protocol endpoint object are updated according to the protocol call results, failure reasons and compensation strategy execution status in the structured execution trajectory; when the number of consecutive failures of a certain protocol endpoint object exceeds the threshold, it is marked as unavailable or downgraded to a candidate endpoint.

[0051] In some embodiments, a context budget joint allocation step is also included: when generating the runtime task description in step S3, the allocation ratio of the context budget among engineering externalized memory, skill assets, protocol constraints, and output format is determined based on the task type, risk level, and historical recovery cost; in step S4, the budget-aware memory snapshot is cropped according to the allocation ratio; and in step S5, the disclosure level of the skill asset object is determined according to the allocation ratio.

[0052] In some embodiments, a lifecycle phase state machine maintenance step is also included: pre-defining the phase sequence and its transformation relationship of the entire lifecycle of large and complex equipment manufacturing; when generating the runtime task description in step S3, determining the scope of accessible data sources, the set of callable protocol endpoints, the set of loadable skill assets and the approval strategy according to the current lifecycle phase; when the structured action intent involves cross-phase operation, triggering phase adaptation verification, if the current phase does not allow the action to be executed, then refusing to execute or requiring manual confirmation.

[0053] In some embodiments, if any one of the protocol verification, permission verification, risk verification, approval verification, or sandbox verification in step S9 fails, the system refuses to execute the structured action intent and generates a rejection reason explanation; the rejection reason explanation and the structured action intent are then combined. Figure 1 It is recorded in the structured execution trajectory and written back to the externalized memory of the project.

[0054] In some embodiments, the large-scale complex equipment manufacturing task is at least one of the following tasks: design review task, process planning task, bill of materials and drawing consistency review task, design change impact analysis task, assembly interference analysis task, quality anomaly 8D analysis task, test fault location task, operation and maintenance fault knowledge feedback task, and production plan impact assessment task.

[0055] In some embodiments, the large language model agent does not directly access external industrial systems, and its output structured action intent must be verified in step S9 before it can be executed; the external industrial system includes at least one or more of the following: product lifecycle management system, manufacturing execution system, enterprise resource planning system, quality management system, computer-aided design system, and monitoring and data acquisition system.

[0056] A collaborative control system for the entire lifecycle of equipment manufacturing based on intelligent external operation is provided to implement the collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation described in any one of the above-mentioned methods. The collaborative control system for the entire lifecycle of equipment manufacturing based on intelligent external operation includes: The task access module is used to receive manufacturing tasks for large and complex equipment. The task parsing and equipment object identification module is used to identify the engineering context associated with the task and generate engineering context information; The runtime task description generation module is used to generate runtime task descriptions based on the project context information. The project externalization memory module is used to retrieve project externalization memory objects from the project externalization memory library according to the runtime task description, and generate a budget-aware memory snapshot based on the context budget; The skill asset management module is used to select skill asset objects from the skill asset library based on the runtime task description and the budget-aware memory snapshot, and generate skill runtime fragments. The industrial protocol gateway module is used to determine industrial protocol endpoint objects from the industrial protocol endpoint library based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, and to generate protocol call constraints. The externalization module collaborative control module is used to construct an externalization module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraint, context budget, permission constraint, approval strategy, and risk level. The large language model intelligent agent module is used to output structured action intentions based on the runtime context generated by the collaborative control diagram of the externalization module; The verification module is used to perform at least one of the following on the structured action intent: protocol verification, permission verification, risk verification, approval verification, and sandbox verification; The sandbox execution module is used to pre-execute structured action intents that need to be verified in a sandbox environment; The observation and trajectory write-back module is used to collect the execution trajectory after performing a controlled operation and write the execution trajectory back to the engineering externalized memory library, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

[0057] An electronic device includes at least one processor, at least one memory, and at least one communication interface. The memory stores a computer program, and when the processor executes the computer program, it implements the collaborative control method for the entire life cycle of equipment manufacturing based on intelligent external operation as described above.

[0058] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described above.

[0059] The technical concept of this invention is as follows: Overall concept: This invention externalizes the long-term states, procedural knowledge, external system interaction contracts, and governance strategies required for the operation of an intelligent agent into engineering externalized memory objects, skill asset objects, industrial protocol endpoint objects, and runtime strategy objects, respectively. The above objects are dynamically arranged before and during each task execution through the externalized module collaborative control diagram.

[0060] The "large and complex equipment manufacturing tasks" referred to in this invention include, but are not limited to: design review, process planning, BOM and drawing consistency review, design change impact analysis, procurement substitution impact analysis, assembly interference analysis, 8D quality anomaly analysis, test fault location, maintenance fault knowledge feedback, production plan impact assessment, and supplier quality problem closure.

[0061] The "externalized runtime" referred to in this invention does not simply mean storing data in an external database, but rather that during the operation of the intelligent agent, the engineering state maintenance, skill selection, protocol invocation, security approval, and feedback revision, which originally relied on the implicit memory of the model or temporary reasoning, are transformed into externally representable, computable, verifiable, and auditable runtime objects, which are then uniformly scheduled by the system-level control logic.

[0062] The "large language model intelligent agent" referred to in this invention refers to an intelligent computing unit that performs task understanding, planning reasoning, and action intent generation under the constraints of a runtime collaborative controller. It may include a large language model reasoning unit, a task planning unit, an action intent generation unit, a tool or protocol invocation decision unit, a result interpretation unit, and a self-checking and correction unit. The large language model reasoning unit can be implemented using a large language model deployed privately by an enterprise, a cloud-based large language model service, an industry model fine-tuned for large and complex equipment manufacturing corpora, a lightweight edge model, or a combination of multiple models. This invention does not limit the specific source, parameter scale, or deployment form of the large language model; the focus is on constraining its input, output, and external actions through an externalized runtime collaborative control mechanism.

[0063] In this application, the large language model agent does not directly access industrial systems such as PLM, MES, ERP, QMS, CAD, CAE, SCADA, or IoT, nor does it directly modify engineering data. The runtime collaborative controller first generates an externalized engineering memory snapshot, skill disclosure content, protocol call constraints, context budget, and permission approval strategy based on the task description, and inputs these as runtime context into the large language model agent. The large language model agent then generates structured action intentions based on these intentions. These action intentions are then processed by protocol verification, permission verification, risk verification, approval gate, and sandbox execution modules before being converted into controlled query, analysis, simulation, report generation, or approval flow initiation operations. The execution results, exception information, and human feedback are written back to the externalized engineering memory, skill asset library, and protocol routing strategy for collaborative control of subsequent tasks.

[0064] System overall architecture: such as Figure 1 As shown, the system of the present invention is divided into a task access layer, an externalized object layer, a collaborative control layer, an agent reasoning layer, a controlled execution layer, and a feedback evolution layer according to the data flow and control flow. Figure 1 The boundaries of each layer are marked with group boxes, and the flow between task description, externalized object, runtime context, action intent, controlled execution result and feedback write-back is indicated by directed arrows.

[0065] Figure 1 In this module, the task parsing and equipment object identification module is used to identify the equipment objects associated with the task, the project number, the manufacturing stage, the configuration version, and the risk level; the engineering externalization memory is used to store the engineering status related to the equipment objects and life cycle stages; the skill asset library is used to store reusable procedural manufacturing knowledge; the industrial protocol endpoint library is used to describe the callable interfaces and call constraints of external industrial systems; the runtime strategy library is used to store permission, approval, sandbox, and budget rules; the externalization module collaborative controller is used to uniformly generate runtime context and action constraints; and the structured observation and trajectory write-back module is used to achieve closed-loop updates.

[0066] Overall method flow: as follows Figure 2 As shown, the method of the present invention includes steps S1 to S10. The specific steps are as follows.

[0067] S1 receives manufacturing tasks for large and complex equipment. Tasks can originate from user natural language input, PLM change orders, MES exception work orders, QMS quality issue reports, testing system alarms, maintenance work orders, or approval system callbacks. Task content can include queries, analysis, diagnosis, solution generation, impact assessment, simulation triggering, or approval workflow initiation.

[0068] S2 identifies the engineering context associated with the task. The system parses the task text and source system metadata to identify the equipment object ID, project number, manufacturing stage, configuration version, BOM version, process version, quality level, risk level, data source scope, and responsible role. For example, the task "There is a conflict between piping and cable trays in the cabin of section A of a cruise ship. Please provide adjustment suggestions" is parsed as follows: the equipment objects are cruise ship sections and cabins, the manufacturing stage is final assembly, the risk level is medium to high risk, and it involves PLM, CAD, MES, and QMS data sources.

[0069] S3, Generate runtime task description. The runtime task description is one of the core intermediate objects of this invention, used to carry the control conditions before the agent executes. Preferably, the runtime task description includes at least the fields in the following table: The runtime task description is not generated out of thin air by the large language model agent, but rather jointly generated by the task parsing and equipment object recognition module, the runtime cooperative controller, enterprise engineering system data, and the runtime policy library. Preferably, the generation process includes the following sub-steps: Task source identification. The system identifies tasks from user natural language input, PLM change orders, MES exception work orders, QMS quality problem orders, test system alarms, maintenance work orders, or approval system callbacks, and records the source system, source document number, trigger time, and initiator role; Task semantic parsing. The task parsing and equipment object recognition module parses natural language text, work order fields, or alarm content to extract task intent, equipment object name, problem type, expected output, and initial risk factors. When the natural language expression is incomplete, the large language model agent can generate candidate task summaries, execution target classifications, and suggestions for completing missing fields. Equipment object and project matching. Based on enterprise master data, equipment object tree, project ledger, BOM structure, workstation or segmented coding, the system maps natural language objects in the task to specific equipment_object_id and project_id; Lifecycle stage and version baseline completion. The runtime collaborative controller calls records from PLM, MES, QMS, ERP or document systems to complete or verify lifecycle_stage, configuration_version, bom_version, process_version and quality_level to avoid version mismatch caused by relying solely on user descriptions; Access control, approval processes, and data scope matching. The system determines data_scope, permission_constraint, and approval_policy based on user roles, project permissions, target system permissions, task risks, and the enterprise policy library. Risk level and context budget generation. The system calculates risk_level based on action type, equipment object sensitivity, lifecycle stage, historical failure rate, whether external system writes are involved, and whether sandbox verification is required, and generates context_budget based on task complexity, available model context window, and externalized object candidate size; Formal description confirmation. For low-risk tasks, the system can automatically generate a formal runtime task description; for high-risk tasks, tasks with conflicting fields, or tasks with missing key fields, the system will submit the candidate runtime task description to the responsible engineer or approver for confirmation before proceeding to the next step.

[0070] In the aforementioned process, the large language model agent is primarily used for task semantic parsing, target induction, candidate field completion, and explanatory text generation. The formal field values ​​of the runtime task description are determined by the runtime collaborative controller after verification based on enterprise master data, engineering system records, and rule policies. Thus, the runtime task description can both absorb the flexibility of natural language input and maintain the version consistency, permission consistency, and approval traceability required in large and complex equipment manufacturing scenarios.

[0071] S4, Retrieve Externalized Engineering Memory. Based on the runtime task description, the system retrieves design data, process data, manufacturing data, quality data, test data, operation and maintenance data, historical anomaly data, approval records, and manual feedback from the externalized engineering memory repository, which are related to equipment objects, projects, stages, versions, and problem types. During the retrieval, not only textual or vector relevance is calculated, but also version consistency, stage adaptability, data timeliness, accessibility, and evidence credibility.

[0072] The retrieval of the externalized engineering memory does not involve the large language model agent directly generating SQL and accessing the database. Instead, the runtime coordinator generates a structured retrieval request based on the runtime task description, which is then converted into a controlled query by the externalized engineering memory module in conjunction with permission policies. Preferably, the retrieval process includes the following sub-steps: 1. Generate a structured search request. The runtime coordinator reads fields such as equipment_object_id, project_id, lifecycle_stage, configuration_version, bom_version, process_version, data_scope, and permission_constraint from the runtime task description, and generates a structured search request containing the search object, search scope, version constraints, permission constraints, and candidate data types; 2. Agent-assisted query rewriting. When user input includes natural language questions, aliases, abbreviations, or colloquial expressions, the large language model agent can generate candidate keywords, synonyms, question type tags, and rewritten search intent results without directly accessing the database, which can then be selected and used by the engineering externalization memory module. 3. Permission and Data Source Filtering. The project externalization memory module filters the data source range and field range in the structured search request based on user role, project permissions, data security level, source system permissions, and task risk level, generating executable search conditions; 4. Controlled Query Execution. Based on the data storage type, the system converts executable search conditions into SQL queries, graph database queries, vector searches, full-text searches, time-series data queries, or industrial system interface calls through preset security templates or search APIs, and records the query template, parameters, time, and caller identifier; 5. Candidate result normalization. The system converts search results from different data sources into engineering external memory objects, supplementing metadata such as equipment object, version, stage, source system, evidence citation, permission level, effective time and credibility; 6. Comprehensive scoring and budget pruning. The system comprehensively scores and ranks the externalized memory objects of candidate projects, and generates a budget-aware memory snapshot under contextual budget constraints.

[0073] The externalized engineering memory objects are persistently stored in the externalized engineering memory repository. This repository is a persistent memory management layer independent of the large language model context and can be deployed in enterprise private clouds, enterprise data centers, industrial edge servers, or hybrid cloud environments. The externalized engineering memory repository may include metadata storage units, semantic indexing units, evidence citation storage units, and version audit log units. Specifically, the metadata storage unit stores structured fields such as equipment object ID, project number, lifecycle stage, configuration version, BOM version, process version, quality status, permission level, effective time, and expiration time; the semantic indexing unit stores memory summaries, keywords, full-text indexes, and vector indexes; the evidence citation storage unit stores reference addresses, file hashes, snapshot identifiers, or summary copies of source system data such as original drawings, BOMs, work orders, inspection reports, test curves, and approval documents; and the version audit log unit stores records of memory object generation, updates, freezes, expirations, conflict resolution, and manual confirmation.

[0074] In a preferred embodiment, external industrial systems such as PLM, MES, QMS, ERP, CAD, CAE, SCADA, and IoT remain authoritative sources of original engineering data. The externalized engineering memory does not require a complete copy of all original data from these systems. Instead, it stores memory objects, indexes, summaries, metadata, and evidence references suitable for retrieval, pruning, access control, and audit traceability during the runtime of large language model intelligent agents. For high-risk tasks, offline environments, or scenarios with high audit requirements, the externalized engineering memory can also store read-only snapshots, anonymized copies, or hash verification information of the original data to support subsequent traceability and consistency verification.

[0075] Preferably, the externalized memory object for engineering projects includes the following fields: In a preferred embodiment, the system calculates a comprehensive score, Score_mem, for each candidate project externalized memory object and sorts them according to the comprehensive score. The comprehensive score can be calculated using the following formula: Score_mem = w1 * Sim_sem + w2 * Match_obj + w3 * Cons_ver + w4 * Fit_stage + w5 * Fresh_time + w6 * Perm_vis + w7 * Trust_evi - w8 * Penalty_conflict; where Score_mem represents the comprehensive score of the candidate project externalized memory object; Sim_sem represents semantic relevance, which can be obtained from the similarity between the task text vector and the memory summary vector, keyword matching degree, or a combination of both; Match_obj represents the equipment object matching degree, taking different matching values ​​when the equipment object ID of the candidate memory object is the same as the equipment object ID in the runtime task description, belongs to its superior system, or belongs to its subordinate component; Cons_ver represents version consistency, based on configuration version, BOM version, and process... The system determines whether the versions of the candidate memory objects are identical, compatible, traceable, or conflicting; Fit_stage represents stage fit, determined by whether the applicable lifecycle stage of the candidate memory object covers the current task stage; Fresh_time represents data timeliness, determined by whether the current time is between the effective and expiration times, and the interval between the data update time and the current time; Perm_vis represents permission visibility, determined by user role, project permissions, data security level, and source system permissions; Trust_evi represents evidence credibility, determined by the type of evidence source, whether it is a formal system record, whether it has been manually verified, and whether there is original evidence cited; Penalty_conflict represents conflict penalty, determined by whether the candidate memory object has version conflicts, factual conflicts, expired status, or has been manually marked as untrustworthy; w1 to w8 are weighting coefficients, which can be preset by the system or configured according to different task types.

[0076] Preferably, each of the above scoring sub-items is normalized to between 0 and 1. For candidate memory objects with invisible permissions, obvious version conflicts, or that are expired and untraceable, the system can filter them directly before comprehensive scoring, or reduce their ranking by using a higher Penalty_conflict. For high-risk tasks, Cons_ver, Perm_vis, and Trust_evi have higher weights than for ordinary query tasks; for fault diagnosis tasks, the weights of Sim_sem, Fresh_time, and historical anomaly evidence can be increased.

[0077] The system generates a budget-aware memory snapshot based on the above fields. The budget-aware memory snapshot is not a simple concatenation of search results, but rather prioritizes retaining the configuration baseline, key evidence, constraints, anomaly history, and recent approval conclusions necessary for the current task under contextual budget constraints, while retaining low-priority data as references that can be retrieved later.

[0078] S5. Select skill assets and determine progressive disclosure granularity. The system selects candidate skills from the skill asset library that match the current task and determines the skill loading granularity based on task complexity, risk level, input data integrity, context budget, historical success rate, and approval strategy.

[0079] The selection of skill assets is not determined arbitrarily by the large language model agent, but rather by the runtime coordinating controller, which generates structured skill retrieval requests based on runtime task descriptions and pre-generated budget-aware memory snapshots. The skill asset management module then performs candidate skill recall, scoring and ranking, and disclosure granularity determination. Preferably, the process includes the following sub-steps: Generate a skill retrieval request. The runtime coordination controller reads `equipment_object_id`, `lifecycle_stage`, `execution_goal`, `risk_level`, `context_budget`, and data source range from the runtime task description, and combines them with the memory snapshot summary generated by S4 to form a skill retrieval request that includes task type, applicable equipment type, applicable stage, input data conditions, and risk constraints; Recalling Candidate Skills. The skill asset management module recalls applicable candidate skills from the skill asset library based on the skill retrieval request. Recall methods may include keyword matching, semantic vector matching, rule matching, equipment type matching, lifecycle stage matching, or a combination of these methods. Verify input integrity. The system compares the `input_requirement` of the candidate skill with the currently obtained externalized engineering memory, runtime task description, and callable protocol endpoints to determine whether the input required to execute the skill has been met. If not, it generates supplementary retrieval or supplementary protocol call requirements. The system calculates a comprehensive skill score. Based on factors such as task semantic matching, equipment type suitability, stage suitability, input data completeness, historical success rate, risk constraint matching, context usage, and approval requirement matching, the system comprehensively scores candidate skills. Determine the granularity of skill loading. Based on the comprehensive score of candidate skills, task risk level, task complexity, context budget, and approval strategy, the system determines the disclosure level of the target skill, including at least one of the following: summary layer, constraint layer, and detailed rule layer; Generate skill runtime fragments. The system will trim the operation steps, decision rules, prohibited items, output format, approval requirements, and failure handling strategies related to the current task from the selected skill into skill runtime fragments, and input them into the externalization module collaborative control diagram.

[0080] Preferably, the skill asset object includes the following fields: In a preferred embodiment, the system calculates a comprehensive score `Score_skill` for each candidate skill asset object and selects the target skill based on the comprehensive score. The comprehensive score can be calculated using the following formula: Score_skill = a1 * Match_task + a2 * Fit_equip + a3 * Fit_stage + a4 * Complete_input + a5 * Success_hist + a6 * Fit_risk + a7 * Match_approval - a8 * Cost_context; where `Score_skill` represents the comprehensive score of the candidate skill asset object; `Match_task` represents the task semantic matching degree, which can be determined by the semantic similarity between the task intent, execution target, and skill name or skill description; `Fit_equip` represents the equipment type fit degree, which is determined based on whether the equipment type applicable to the skill covers the type of the current equipment object; `Fit_stage` represents the stage fit degree, which is determined based on whether the lifecycle stage applicable to the skill covers the current task stage; `Co` represents the stage fit degree, which is determined based on whether the lifecycle stage applicable to the skill covers the current task stage; and `Co` represents the equipment type fit degree. `mplete_input` represents the completeness of input data, determined by whether the required input data for the skill has been satisfied in the runtime task description, engineering externalization memory snapshot, and callable protocol endpoints; `Success_hist` represents the historical success rate, determined by the number of successful executions, failures, and human ratings for the skill in the same or similar tasks; `Fit_risk` represents the risk constraint matching degree, determined by whether the skill's prohibited items, approval requirements, and failure handling strategies cover the current risk level; `Match_approval` represents the approval requirement matching degree, determined by whether the approval role, approval process, and current task approval strategy required by the skill are consistent; `Cost_context` represents the context occupancy penalty, determined by the ratio of the context length occupied by the skill at different disclosure levels to the current remaining context budget; `a1` to `a8` are weighting coefficients, which can be preset by the system or configured according to the task type.

[0081] Preferably, each of the above scoring sub-items is normalized to between 0 and 1. For candidate skills that are clearly not satisfied by the input data and cannot be supplemented for retrieval, the system can directly filter them; for high-risk tasks, the weights of `Fit_risk` and `Match_approval` are higher than those of ordinary query tasks; for complex fault diagnosis or assembly anomaly handling tasks, the weights of `Complete_input` and `Success_hist` can be increased.

[0082] Skilled assets are preferably disclosed using a progressive approach, including a summary layer, a constraint layer, and a detailed rules layer. Low-risk query tasks only include the summary layer; medium-risk analysis tasks include both the summary and constraint layers; high-risk tasks involving changes, quality conclusions, plan adjustments, or equipment-related matters include the summary, constraint, and detailed rules layers, with mandatory additional approval requirements and prohibitions.

[0083] Typical skill assets include: BOM and drawing consistency review skills, design change impact analysis skills, welding process compliance review skills, assembly interference and process conflict analysis skills, 8D quality anomaly analysis skills, test fault location skills, and maintenance fault knowledge feedback skills.

[0084] S6. Identify industrial protocol endpoints and generate protocol call constraints. Based on task objectives, skill requirements, and permission policies, the system determines callable interfaces from the industrial protocol endpoint library. An industrial protocol endpoint object includes at least the system type, endpoint address or logical identifier, call method, input schema, output schema, lifecycle state, permission scope, risk level, audit requirements, retry policy, and failure compensation policy.

[0085] The determination of the industrial protocol endpoint is not achieved by the large language model agent directly selecting the interface address and initiating the call. Instead, the runtime cooperative controller generates a protocol endpoint matching request, and the industrial protocol gateway performs adaptive mapping, scoring selection, and call constraint generation. Preferably, the process includes the following sub-steps: 1. Generate protocol endpoint matching requests. The runtime coordinator generates protocol endpoint matching requests based on the execution_goal, data_scope, permission_constraint, and risk_level in the runtime task description, and the required_protocols, input_requirement, and output_schema in the target skill asset; 2. Recall candidate protocol endpoints. The industrial protocol gateway recalls candidate endpoints from the industrial protocol endpoint library based on system type, capability tag, data object type, input / output schema, lifecycle status, permission scope, and risk level; 3. Perform adaptive mapping. The industrial protocol gateway maps internal action intents or internal data objects to the interface call format of the target industrial system, including field name mapping, enumeration value mapping, unit conversion, version field mapping, pagination parameter mapping, filter condition mapping, and return result normalization; 4. Endpoint Scoring and Ranking. The system comprehensively scores candidate protocol endpoints based on interface necessity, source system authority, schema matching degree, permission availability, risk matching degree, historical availability, call cost, and audit completeness. 5. Call Constraint Generation. For target endpoints filtered by scoring and permissions, the system generates protocol call constraints, including allowed methods, parameter ranges, maximum number of calls, timeout, retry strategy, audit fields, failure compensation strategy, and whether approval or sandboxing is required; 6. No available endpoint processing. When no endpoint meets the task objectives, permissions, and risk constraints, the system does not allow the large language model agent to bypass the protocol gateway and directly call the external system. Instead, it returns a prompt for supplementary authorization, supplementary data source, manual processing, or downgrades to generating only suggestions. This invention does not require all external systems to use the same underlying protocol. The system uses an industrial protocol gateway to uniformly map different protocol inputs to internal protocol endpoint objects, and performs protocol verification, permission verification, risk verification, and audit configuration before invocation.

[0086] In a preferred embodiment, the comprehensive score of the candidate industrial protocol endpoint is calculated using the following formula: Score_endpoint = b1 * Need_api + b2 * Auth_source + b3 * Match_schema + b4 * Perm_api + b5 * Fit_risk + b6 * Availability_hist + b7 * Audit_complete - b8 * Cost_call; where `Score_endpoint` represents the comprehensive score of the candidate industrial protocol endpoint; `Need_api` represents the necessity of the interface, determined based on whether the target skill or action intention must rely on this endpoint; `Auth_source` represents the authority of the source system, for example, the BOM is based on the authoritative source in PLM or ERP, and quality issues are based on the authoritative source in QMS; `Match_schema` represents the input / output scheme. `MA` matches the endpoint's input and output fields, and determines the coverage of these fields with the current task. `Perm_api` represents permission availability, determined by whether the current user, project, role, and system permissions allow the endpoint to be called. `Fit_risk` represents risk matching, determined by the endpoint's supported call types, write capabilities, approval requirements, and task risk level. `Availability_hist` represents historical availability, determined by the endpoint's historical success rate, failure rate, average latency, and recent health status. `Audit_complete` represents audit completeness, determined by whether the endpoint supports call logs, parameter recording, approval association, and result tracking. `Cost_call` represents call cost penalty, determined by API call time, cost, rate limiting, network overhead, or impact on the production system. `b1` to `b8` are weighting coefficients, which can be preset by the system or configured according to the task type.

[0087] The above process constitutes a controlled adaptive mapping invocation. The large language model agent can generate candidate invocation intents based on the runtime context, such as "query the current BOM version" or "initiate a design change impact analysis approval draft," but it cannot directly determine the interface address, authentication information, or underlying invocation parameters. The industrial protocol gateway is responsible for mapping candidate invocation intents to interface calls of specific systems such as PLM, MES, ERP, QMS, CAD / CAE, and SCADA / IoT, and completes schema conversion, parameter mapping, permission verification, risk verification, and audit configuration before the invocation.

[0088] S7, Construct an Externalized Module Collaborative Control Chart. This chart is used to uniformly represent the control relationships between candidate memories, candidate skills, candidate protocol endpoints, context budgets, permission constraints, risk approval, and execution feedback. Unlike ordinary workflow diagrams or multi-agent DAGs, the nodes in this control chart are not only task steps, but also include memory snapshots, skill disclosure levels, protocol endpoints, permission domains, approval gates, sandbox strategies, and write-back rules.

[0089] like Figure 3 As shown, the collaborative control diagram is used to integrate candidate project externalization memory, candidate skill assets, candidate industrial protocol endpoints and their scoring results with context budget, permission approval strategy and risk constraints to generate runtime context, action constraints and write-back rules.

[0090] When constructing the control chart, S7 no longer recalculates the base scores of candidate objects, but instead inherits the scores and constraint results already generated by S4, S5, and S6. Specifically, S7 receives the candidate project externalized memory objects and their Score_mem, budget-aware memory snapshots, and evidence references output by S4; it receives the candidate skill asset objects and their Score_skill, skill disclosure level, and skill runtime fragments output by S5; it receives the candidate industrial protocol endpoint objects and their Score_endpoint, protocol call constraints, and endpoint mapping relationships output by S6; and it receives the context budget, permission policy, approval policy, risk level, and execution objective from the runtime task description.

[0091] In a preferred embodiment, the process of generating the externalized module collaborative control diagram includes the following sub-steps: Establish task root nodes and constraint nodes. The system establishes task root nodes based on runtime task descriptions and establishes constraint nodes for equipment objects, lifecycle stages, configuration versions, BOM versions, process versions, risk levels, permission domains, and approval policies. Write to memory nodes. The system writes the externalized memory objects in S4 that meet the threshold for comprehensive scores or are identified as key evidence into the control chart as memory nodes, and marks the nodes with Score_mem, evidence reference, version status, permission level and whether they have entered the runtime context; Write the skill node. The system writes the selected target skill or candidate skill in S5 into the control chart as a skill node, and marks the node with Score_skill, disclosure level, input data integrity, prohibited items, approval requirements and failure handling strategy; Write to protocol endpoint nodes. The system writes the target endpoints or alternative endpoints filtered by permissions and risks in S6 into the control graph as protocol endpoint nodes, and marks the Score_endpoint, target system, schema mapping relationship, calling method, parameter range, audit requirements and failure compensation strategy on the nodes; Establish dependency edges and constraint edges. Based on skill input requirements, source of memory evidence, protocol endpoint input / output schema, permission policy, and risk level, the system establishes dependency edges and constraint edges between memory nodes, skill nodes, protocol endpoint nodes, budget nodes, approval nodes, and sandbox nodes. Perform context budget fusion. Based on the context budget allocation results, the system determines which memory objects directly enter the context, which skill content is loaded into the summary layer or the detail layer, which protocol schemas are directly exposed or delayed in exposure, and which evidence is only retained for reference. The system generates runtime context and action constraints based on the control chart. Simultaneously, it generates action constraints, including rules for automatic execution, suggestion generation only, manual approval, sandbox verification, prohibition of execution, and failure write-back. The feedback write-back rules are retained. The system records the target objects that should be written back in the execution trajectory in the control chart, including externalized engineering memory, skill scores, protocol routing, and runtime policies, so that S10 can perform closed-loop updates.

[0092] Through the above fusion generation process, the externalized module collaborative control diagram merges the memory retrieval results of S4, the skill selection results of S5, the protocol endpoint selection results of S6, as well as the context budget, permission approval, and risk constraints into a single runtime control object, thereby avoiding the isolated execution of memory retrieval, skill loading, and interface calls.

[0093] S8 generates the runtime context of the large language model agent and obtains the action intent. The system assembles the budget-aware memory snapshot, skill disclosure content, protocol call constraints, permission constraints, and output format requirements into a runtime context and inputs it into the large language model agent. The large language model agent outputs the action intent, which includes the suggested operation type, target object, call interface, input parameters, expected output, risk statement, and required approval.

[0094] Action intentions should preferably be presented in a structured format, for example: { "action_type":"generate_change_impact_analysis", "target_object":"AIRCRAFT-PART-Frame-Station-XX", "required_protocols":["PLM.query_bom","MES.query_process_route","QMS.query_inspection_plan"], "parameters":{ "configuration_version":"CFG-2026-05-A", "bom_version":"BOM-R17", "process_version":"PROC-R09" }, "risk_level":"L3", "approval_required":true, "expected_output":"impact_matrix_and_approval_draft" }

[0095] S9 performs protocol, permission, risk, approval, and sandbox validation on action intents. The system does not directly execute action intents output by the large language model; instead, the runtime governance module performs multiple validations. Validation includes at least: whether the protocol schema matches, whether the parameters are within the allowed range, whether the user role has the necessary permissions, whether the target object version is consistent, whether the action risk level exceeds the current autonomy level, whether manual approval is required, and whether pre-execution in a digital twin or sandbox environment is necessary.

[0096] The correspondence between risk levels and execution strategies can be set as follows: like Figure 4 As shown, high-risk actions require approval through a combination of approval processes and sandbox strategies.

[0097] S10 executes controlled operations and writes back the trajectory, results, and human feedback. After verification, the system executes queries, analyses, simulations, report generation, approval flow initiation, or other controlled operations, and collects model calls, memory retrievals, skill calls, protocol calls, approval events, sandbox results, anomaly information, human feedback, and final conclusions to form a structured execution trajectory.

[0098] like Figure 5 As shown, the execution trajectory is written back to the project externalization memory, skill asset library, protocol endpoint library, and runtime strategy library.

[0099] Through the aforementioned write-back mechanism, when the intelligent agent subsequently encounters similar equipment objects, similar manufacturing stages, or similar anomalies, it can prioritize recalling verified cases, improve skill ranking, avoid failed protocol paths, and revise control strategies based on human feedback.

[0100] Lifecycle Stage State Machine: The manufacturing tasks of large and complex equipment are strongly correlated with the lifecycle stages. The permissibility, risk level, and approval requirements of the same action differ at different stages. For example, BOM adjustments before design freeze may be normal design activities, while BOM adjustments after manufacturing commencement may affect procurement, inventory, processes, quality, and delivery.

[0101] Preferably, system maintenance such as Figure 6 The lifecycle stage state machine is shown.

[0102] When generating runtime task descriptions, the system determines accessible data sources, invoked protocol endpoints, loadable skills, executable actions, and approval policies based on the stage state machine. For example: 1. During the detailed design phase, the agent can generate design review comments and draft changes, but cannot directly release the design baseline; 2. During the process design phase, the intelligent agent can generate process route suggestions and simulation tasks, but the process route must be approved by the process manager before it can be implemented. 3. During the final assembly and integration phase, the intelligent agent can perform interference analysis and suggest rework solutions, but any changes to on-site work instructions must be approved. 4. During the testing and inspection phase, the intelligent agent can assist in fault location and provide retest suggestions, but it cannot automatically modify the test conclusions; 5. During the service and maintenance phase, the intelligent agent can transform field faults into design or process improvement suggestions and feed them back to the maintenance feedback and improvement phase.

[0103] Distinguishing Features of Externalized Module Collaborative Control Charts: S7 has already described the process of merging and generating externalized module collaborative control charts. This section further explains the distinguishing features of this control chart compared to ordinary workflow diagrams, ordinary multi-agent DAGs, ordinary RAG retrieval chains, and ordinary tool call chains.

[0104] Joint allocation of context budget: Large and complex equipment manufacturing tasks typically involve a large amount of data. To avoid irrelevant data cluttering the model context, this invention sets up a joint allocation of context budget to uniformly allocate budgets for engineering memory, skill content, protocol schema, and output constraints.

[0105] Preferably, the context-based budget joint allocator determines the budget proportions for different modules based on the task phase, risk level, historical recovery costs, and data necessity. For example: When the context budget is insufficient, the system prioritizes retaining evidence, skill constraint layers, and high-risk protocol constraints directly related to the current equipment object and version baseline, while using less relevant historical cases, complete standard texts, and low-risk protocol schemas as lazy-load references. This reduces context redundancy and increases the effective engineering information density of the model input.

[0106] Risk grading, approval gates, and sandbox strategies are linked: In the manufacturing of large and complex equipment, changes to physical manufacturing processes and engineering data often cannot be easily rolled back. This invention establishes a risk grading approval gate and sandbox strategy linkage mechanism.

[0107] The risk score is preferably generated by weighting the following factors: 1. Risks associated with action types, such as querying, generating suggestions, submitting for approval, modifying data, and issuing control instructions; 2. Object sensitivity, such as configuration baseline, BOM, process route, quality conclusions, and equipment control parameters; 3. Lifecycle stages, such as design stage, production stage, final assembly stage, testing stage, or operation and maintenance stage; 4. User permission levels and project roles; 5. Historical failure rate and similar anomaly records; 6. Does it involve supplier, airworthiness, classification society, regulatory or customer delivery constraints? 7. Is there a sandbox or digital twin environment available for pre-execution?

[0108] The system determines the execution mode based on risk scores. Low-risk queries can be executed automatically; medium-risk analyses can automatically generate reports but are not effective; high-risk actions must be manually approved; and extremely high-risk actions can only be entered into a sandbox, digital twin, or manually reviewed, and are not allowed to be executed directly by the intelligent agent.

[0109] Sandbox strategies include, but are not limited to: read-only data copies, digital twin simulation, network isolation, file system whitelisting, command whitelisting, timeout limits, cost limits, rollback checkpoints, and output auditing. For actions involving writes to external industrial systems, the system generates a rollback point or difference summary for manual review before submission for approval.

[0110] System, electronic devices, and storage media implementation: The system of this invention can be deployed in enterprise private clouds, industrial internet platforms, edge servers, or hybrid cloud environments. The system includes at least: The task access module is used to receive natural language tasks, work orders, change orders, quality problem orders, test alarms, or approval callbacks. The task parsing and equipment object identification module is used to identify equipment objects, projects, stages, versions, risks, and data source scope. The runtime task description generation module is used to generate runtime task descriptions based on the project context information. The project externalization memory module is used to build, retrieve, trim, and update project externalization memory objects; The Skills Asset Management module is used for registering, matching, progressively disclosing, scoring, and revoking skills assets. Industrial protocol gateway modules are used to encapsulate interfaces for PLM, MES, ERP, QMS, CAD, CAE, SCADA, IoT, document systems, and approval systems. The externalized module is a collaborative control module used to construct collaborative control graphs, allocate context budgets, and generate runtime contexts. The large language model intelligent agent module is used to generate action intentions based on the runtime context; The verification module is used to perform pre-execution processing on the intent of an action; The sandbox execution module is used to validate high-risk actions in read-only copies, digital twins, or isolated environments; The observation and trajectory write-back module is used to collect execution trajectories and update memories, skills, protocols, and strategies.

[0111] The electronic device includes at least one processor, a memory, a communication interface, and an industrial system access interface. The memory stores a program executable by the processor, which, when executed, implements the method of the present invention. A computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of the present invention.

[0112] The improvements of this invention compared to the prior art are as follows: Therefore, this invention can achieve higher data consistency, execution stability, security controllability, context efficiency, and experience reuse capabilities in large-scale and complex equipment manufacturing scenarios.

[0113] Example 1: Interference handling during cruise ship segment outfitting.

[0114] In this embodiment, the user inputs the task: "In the third deck of section A of a cruise ship project P, there is a spatial conflict between the pipelines and cable trays during on-site installation. Please analyze the cause and provide adjustment suggestions."

[0115] The system executes the following process: (1) Runtime task description generation. The task parsing module identifies the equipment object as a cabin on the third deck of section A of the cruise ship, the lifecycle stage as final assembly integration, and the problem type as assembly interference. The runtime task description may include: task_id=T-CRUISE-20260510-001, equipment_object_id=CRUISE-P-BLOCK-A-DK3-CAB032, project_id=P-CRUISE-2026, lifecycle_stage=final assembly integration, configuration_version=CFG-PA-R12, bom_version=BOM-PA-R08, process_version=PROC-OUTFITTING-R05, risk_level=L3, context_budget=16000 tokens, data_scope=PLM / CAD / MES / QMS / document system, approval_policy=signed by the outfitting manager and the electrical manager; (2) Externalized Engineering Memory Retrieval. The externalized engineering memory module generates a structured retrieval request based on the runtime task description, retrieves and scores candidate memory objects. For example, the snapshot of the 3D model of segment A has a score of 0.94, the piping layout diagram has a score of 0.91, the cable tray path diagram has a score of 0.90, the current construction procedure of MES has a score of 0.85, the rework case of similar compartments has a score of 0.77, and the summary of the classification society layout rules has a score of 0.68. The system directly writes the first three categories into the runtime context, and uses the complete classification society rules as evidence references with delayed loading. (3) Skill Asset Selection. The skill asset management module recalls the "Assembly Interference and Process Conflict Analysis Skill" and calculates Score_skill=0.92, and recalls the "Design Change Impact Analysis Skill" and calculates Score_skill=0.74. Since the task risk level is L3, the system loads the summary layer, constraint layer and detailed rule layer for the former skill, and only loads the constraint layer for the latter skill as an auxiliary judgment; (4) Protocol Endpoint Selection. The industrial protocol gateway recalls and scores the following endpoints: PLM drawing query endpoint (Score_endpoint=0.93), CAD model interference query endpoint (Score_endpoint=0.95), MES process status endpoint (Score_endpoint=0.88), and QMS rework record endpoint (Score_endpoint=0.82). All of these endpoints are configured for read-only queries or draft approval initiation, and direct modification of work orders is not allowed. (5) Collaborative control chart generation. The external module collaborative controller writes high-resolution memory nodes, assembly interference skill nodes, CAD / PLM / MES / QMS endpoint nodes, L3 risk nodes, approval and countersigning nodes, and context budget nodes into the collaborative control chart and generates action constraints: allows automatic querying and generation of adjustment suggestions, allows generation of approval drafts, and prohibits direct modification of PLM drawings, MES work orders, or on-site construction status; (6) Agent Action Intent and Controlled Execution. The large language model agent generates action intent in the runtime context, including calling CAD intervention results, querying the current installation status of MES, generating conflict causes, listing affected pipelines and cable trays, and generating two sets of adjustment plans and approval drafts. After verification by the controlled execution layer, only read-only queries and draft generation are executed; (7) Output and Write-back. The system outputs the cause of the conflict, a list of affected objects, adjustment plans, the impact on subsequent processes, the professional departments requiring review, and the draft approval workflow. The approval conclusions after manual confirmation, the final adjustment plans, rework hours, and on-site photos are written back to the externalized engineering memory, and the successful sample of this "Assembly Interference and Process Conflict Analysis Skill" is included in the skill score.

[0116] Through this embodiment, the system can avoid generating erroneous rework suggestions based solely on partial text data, and integrate multi-source data from CAD, PLM, MES, and QMS into controlled analysis.

[0117] Example 2: Impact analysis of design changes to aircraft structural components.

[0118] In this embodiment, the user inputs the task: "After the design of a certain aircraft structural component is changed, please determine the impact on the process, procurement, assembly and inspection, and generate review materials."

[0119] The system executes the following process: (1) Runtime task description generation. The task parsing module identifies the task source as a PLM design change order or a user natural language request, and the equipment object as an aircraft structural component. The runtime task description may include: task_id=T-AIR-ECO-20260510-014, equipment_object_id=AIR-FRAME-STA420-BRACKET-17, project_id=AIR-PROJ-X, lifecycle_stage=transition from detailed design to process design, configuration_version=CFG-AIR-X-BL23, bom_version=BOM-AIR-X-R17, process_version=PROC-ASSY-R09, risk_level=L3, context_budget=20000 tokens, approval_policy=signed by the design manager, process manager, quality manager and airworthiness compliance manager; (2) Externalized Engineering Memory Retrieval. The externalized engineering memory module retrieves differences in drawings before and after changes, BOM differences, process routes, supplier delivery status, inspection plans, and historical change cases. For example, PLM change difference summary Score_mem=0.95, BOM impact list Score_mem=0.93, MES process route Score_mem=0.88, QMS inspection plan Score_mem=0.81, ERP supplier delivery constraints Score_mem=0.74, and similar ECO historical cases Score_mem=0.78. The system directly writes change differences, BOM impact, and process routes into the context, and uses supplier delivery details as deferred references. (3) Skill Asset Selection. In the Skill Asset Management module, select “Design Change Impact Analysis Skill”, Score_skill=0.94, and load the summary layer, constraint layer, and detail layer; at the same time, select “BOM and Drawing Consistency Review Skill”, Score_skill=0.89, and load the summary layer and constraint layer to verify whether the differences between the drawings and the BOM are consistent; (4) Protocol endpoint selection. The industrial protocol gateway selects the following endpoints: PLM change order query endpoint (Score_endpoint=0.96), ERP procurement status endpoint (Score_endpoint=0.82), MES process route endpoint (Score_endpoint=0.86), and QMS inspection plan endpoint (Score_endpoint=0.84). Among them, PLM is the authoritative source of design changes, QMS is the authoritative source of inspection requirements, and ERP data is only used as supplementary evidence of procurement impact. (5) Collaborative Control Chart Generation. The collaborative control chart includes change difference memory nodes, BOM impact nodes, process route nodes, design change impact analysis skill nodes, BOM consistency review skill nodes, PLM / ERP / MES / QMS endpoint nodes, L3 risk nodes, and countersigning approval nodes. Action constraints are set as follows: allow generation of impact matrices and approval drafts, allow initiation of approval flow drafts, prohibit automatic release of design changes, and prohibit automatic modification of the formal BOM or process route; (6) Agent Action Intent and Controlled Execution. The large language model agent generates structured action intents, requesting the formation of six types of influence matrices: design, process, procurement, assembly, quality, and airworthiness. After verifying protocols, permissions, and risks, the controlled execution layer only allows calls to read-only query interfaces and the generation of review materials; (7) Output and Write-back. The system outputs the impact matrix, items requiring additional verification, a list of affected materials and processes, supplier delivery risks, inspection plan adjustment suggestions, and a draft approval flow. Review committee comments, approval conclusions, final change status, and adopted impact analysis results are written back to the engineering externalization memory, and the "Design Change Impact Analysis Skills" and the historical success rate of relevant protocol endpoints are updated.

[0120] Through this embodiment, the present invention can extend the design change from a single drawing comparison to a full-link impact analysis across PLM, ERP, MES, and QMS, and prevent intelligent agents from releasing design changes without authorization through an approval gate.

[0121] Example 3: Diagnosis of abnormal vibration or temperature rise in large generator tests.

[0122] In this embodiment, the test system generates an alarm: "Abnormal vibration and local temperature rise occurred during the factory test of a large generator. Please assist in locating the cause and provide handling suggestions." The system executes the following process: (8) Runtime task description generation. The task access module receives alarms from SCADA or the test system and identifies the equipment object as a large generator set and its rotor, stator, windings and cooling system. The runtime task description may include: task_id=T-GEN-TEST-20260510-003, equipment_object_id=GEN-G500-UNIT02, project_id=GEN-G500-P02, lifecycle_stage=inspection test, configuration_version=CFG-GEN-G500-R06, bom_version=BOM-GEN-R11, process_version=TEST-PROC-R04, risk_level=L4, context_budget=24000 tokens, approval_policy=reviewed by the test manager, quality manager and chief engineer; (9) Engineering Externalized Memory Retrieval. The engineering externalized memory module retrieves and scores the following data: SCADA vibration and temperature rise curve summary (Score_mem=0.96), rotor assembly record (Score_mem=0.90), dynamic balance record (Score_mem=0.88), stator winding insulation process (Score_mem=0.82), test procedure limit (Score_mem=0.87), and historical similar fault cases (Score_mem=0.79). The system directly writes real-time curves, key assembly parameters, test procedure limits, and dynamic balance records into the context, and uses complete historical fault reports as evidence. (10) Skill Asset Selection. In the Skill Asset Management module, select "Test Fault Location Skill", Score_skill=0.95, and load the summary layer, constraint layer, and detail layer; select "Quality Anomaly 8D Analysis Skill", Score_skill=0.86, and load the summary layer and constraint layer for subsequent quality closed-loop analysis; (11) Protocol Endpoint Selection. The industrial protocol gateway selects the following endpoints: SCADA test data read-only endpoint Score_endpoint=0.94, QMS quality record endpoint Score_endpoint=0.89, MES assembly record endpoint Score_endpoint=0.84, and document system test procedure endpoint Score_endpoint=0.82. Since the task risk level is L4, all endpoints are configured for read-only or sandbox verification, and writing device control commands is not permitted. (12) Collaborative Control Chart Generation. The collaborative control chart is written into real-time test curve nodes, assembly record nodes, test procedure limit nodes, test fault location skill nodes, quality 8D skill nodes, SCADA / QMS / MES / document system endpoint nodes, L4 risk nodes, sandbox verification nodes, and manual review nodes. Action constraints are set as follows: allow the generation of diagnostic suggestions, cause ranking, and retesting plans; allow verification and analysis paths in read-only data copies or digital twin environments; prohibit modification of test conclusions, prohibit closing quality problem forms, and prohibit issuing equipment control commands; (13) Agent action intent and controlled execution. The large language model agent outputs a fault cause ranking, additional measurement items, retesting suggestions, and risk mitigation plans. The controlled execution layer identifies actions that involve test safety and quality conclusions, performs only read-only queries and sandbox verification, and submits the results to the test engineer, quality manager, and chief engineer for review; (14) Output and Write-back. The system outputs a ranking of possible causes, such as rotor dynamic balance deviation, abnormal cooling channels, local insulation defects in windings, or sensor malfunctions; it also outputs retest recommendations, additional measurement points to be collected, risk mitigation plans, and a manual review list. The final fault cause, mitigation measures, retest results, manual confirmation opinions, and quality issue number are written back to the engineering external memory and used as priority recall cases for similar generator test anomalies in the future.

[0123] Through this embodiment, the system can utilize multi-source engineering data to assist in fault location during testing and generate reusable fault diagnosis experience without exceeding its authority to control equipment or automatically modifying quality conclusions.

[0124] This invention does not limit the specific type of large language model and can employ privately deployed models, cloud-based models, industry-tuned models, or combinations of multiple models. The engineering externalization memory can utilize relational databases, graph databases, vector databases, object storage, time-series databases, or combinations thereof. The skill asset repository can take the form of files, databases, knowledge graphs, or package management. The industrial protocol gateway can interface with REST, GraphQL, OPC UA, MQTT, database views, message queues, vendor SDKs, or enterprise service buses.

[0125] In another implementation, the collaborative control diagram can be generated with the assistance of a rule engine, state machine, graph computation engine, or large language model, but it must undergo system-side verification of protocols, permissions, risks, and approvals before final execution. In yet another implementation, sandbox execution can be achieved through a digital twin platform, simulation software, read-only data copies, process engine pre-visualization, or a container isolation environment.

[0126] The above alternatives do not affect the core of this invention: namely, for the entire life cycle of large and complex equipment manufacturing, jointly scheduling engineering externalized memory, skill assets, industrial protocol endpoints, context budget, risk approval, sandbox execution, and execution feedback in the same runtime control mechanism.

[0127] The following describes possible variations of the embodiments. It should be understood that the following variations can be used individually or in combination without departing from the core concept of the invention. The scope of protection of the present invention should not be limited to the specific module names, system types, process sequences, or data fields listed.

[0128] Example of changes to the externalized memory module in engineering: In the above embodiments, the externalized engineering memory simultaneously stores design data, process data, quality data, manufacturing data, test data, operation and maintenance data, and historical anomaly data. In other embodiments, the externalized engineering memory module can be modified as follows: Storage component replacement. The externalized engineering memory can be replaced by a relational database, graph database, object storage, time-series database, document database, knowledge graph, or a combination of multiple databases. For example, design drawing summaries can be stored in a document database, equipment object relationships can be stored in a graph database, and sensor curves can be stored in a time-series database.

[0129] Index fields can be replaced or expanded. Equipment object ID, project number, configuration version, BOM version, and process version can be replaced with the company's existing model number, batch number, segment number, serial number, equipment station account, engineering change order number, or quality issue order number. Supplier number, airworthiness clause number, classification society specification number, customer contract batch number, or site workstation number can also be added.

[0130] The method of memory generation has changed. Externalized engineering memory objects can be automatically generated synchronously from systems such as PLM, MES, and QMS, or they can be written after manual review. Alternatively, they can be generated by a large language model agent based on the execution trajectory, and then written after being approved and confirmed by a rule engine or by humans.

[0131] The memory update strategy has changed. In low-risk scenarios, execution traces can be automatically written into the externalized engineering memory; in high-risk scenarios, execution traces can first enter the review area, and then enter the formal memory bank after confirmation by the responsible engineer. For invalid configurations or expired processes, the system can choose to delete, freeze, demote, archive, or only retain evidence references.

[0132] The memory retrieval order changes. In the best embodiment, a runtime task description is generated first, followed by memory retrieval. In another embodiment, the system can first perform a coarse search based on user input, then complete the equipment object and version information based on the coarse search results, and finally perform a fine search.

[0133] Example of changes to the Skills Assets module: In the above embodiments, skill assets adopt a three-tiered progressive disclosure approach: summary layer, constraint layer, and rules layer. In other embodiments, the skill asset module can be modified as follows: Disclosure hierarchy changes. Skill assets can be set to two, three, four, or more layers. For example, two layers may include a summary layer and an execution layer; four layers may include a summary layer, an input validation layer, a constraint rule layer, and a detailed operation layer.

[0134] Skill organization methods can be replaced. Skill assets can exist as individual skill files, skill packages, plugins, database records, knowledge graph nodes, process templates, rule sets, or executable scripts. For highly standardized process tasks, skills can be represented as fixed process templates; for analytical tasks, skills can be represented as step descriptions, checklists, and output schemas.

[0135] Skill selection strategies have changed. Skills can be selected by a rule engine, a retrieval model, a large language model, a classification model, a graph matching algorithm, or a combination of these methods. Skill ranking can be based on historical success rate, task similarity, expert ratings, most recent update time, applicable equipment type, or failure rate.

[0136] Skill omission scenarios. For simple data retrieval, low-risk summary generation, or fixed report export tasks, the skill asset selection step can be omitted, and the task can be completed using only engineering externalization memory and protocol endpoints.

[0137] Skills enhancement scenarios. During industry expansion, skills assets can be added related to supply chain risk assessment, carbon emission accounting, cost impact analysis, maintainability analysis, spare parts recommendation, and field service knowledge feedback.

[0138] Skill combinations can vary. Multiple skills can be combined sequentially, in parallel, or by a primary skill calling a sub-skill. For example, in a design change impact analysis, the BOM and drawing consistency review skill can be called first, followed by the procurement impact analysis skill, and finally the quality inspection plan update skill.

[0139] Examples of changes in industrial protocol endpoints and protocol gateways: In the above embodiments, the industrial protocol endpoint library covers PLM, MES, ERP, QMS, CAD, CAE, SCADA, IoT, document systems, and approval systems. In other embodiments, the protocol endpoints and protocol gateways can be changed as follows: Endpoint types can be added or removed. Depending on the enterprise's system construction, CAD, CAE, SCADA, or IoT endpoints can be omitted, while endpoints such as WMS, APS, EAM, LIMS, SRM, CRM, digital twin platform, master data platform, data lake, knowledge base, or low-code process platform can be added.

[0140] Protocol implementation alternatives. Industrial protocol gateways can access external systems via REST API, GraphQL, OPC UA, MQTT, WebSocket, message queues, database views, file exchange, vendor SDKs, enterprise service bus, or robotic process automation.

[0141] The direction of invocation changes. Protocol endpoints can support read-only queries, or they can support draft writing, approval flow initiation, status write-back, or formal data writing. For high-risk systems, only read-only endpoints can be enabled; for low-risk systems, automatic write endpoints can be enabled.

[0142] The protocol verification method has changed. Protocol verification can be performed at the gateway layer or within each external system adapter; it can employ schema verification, rule verification, state machine verification, authorization token verification, digital signature verification, or a combination of these verification methods.

[0143] Endpoint discovery methods have changed. Protocol endpoints can be pre-registered by administrators, synchronized by the system from the enterprise service catalog, or discovered by agents within a controlled scope and then manually confirmed.

[0144] Example of changes in the collaborative control diagram of externalized modules: In the above embodiments, the system constructs an externalized module collaborative control diagram to uniformly represent candidate memories, candidate skills, candidate protocol endpoints, context budgets, permission constraints, risk approval, and execution feedback. In other embodiments, this control diagram can be modified as follows: Graph structure replacement. The externalized module collaborative control diagram can be replaced by a state machine, rule table, decision tree, BPMN process, Petri net, directed acyclic graph, blackboard system, task queue or policy matrix, as long as it can express the dependencies, constraints and execution order between externalized objects.

[0145] Node type changes. Control charts can omit some nodes, such as omitting the approval gate node in read-only analysis scenarios; they can also add cost control nodes, model selection nodes, data anonymization nodes, compliance audit nodes, human collaboration nodes, or digital twin verification nodes.

[0146] Changes in edge relationships. Candidate memories, candidate skills, and candidate protocol endpoints can be directly dependent on each other, or they can be indirectly related through intermediate task stage nodes, risk level nodes, budget nodes, or permission domain nodes.

[0147] The generation method has changed. Control charts can be generated by a rule engine, or candidate charts generated by a large language model and then verified by the system, or template charts retrieved from historical successful cases and then partially modified.

[0148] The execution method has changed. The control chart can be generated and executed all at once, or it can be updated incrementally based on observations at each execution stage. For example, when a protocol call fails, the system can temporarily replace the protocol endpoint; when the search results are insufficient, the system can append more search results; when the risk increases, the system can insert an approval node.

[0149] Example of changes in the order of method steps: The above embodiments provide a preferred order for steps S1 to S10. In other embodiments, the method steps can be adjusted while maintaining the technical effect: The order of memory retrieval and skill selection can be reversed. The system can first retrieve the externalized engineering memory based on the task description and then select the skill; or it can first identify the task type and select the skill, and then retrieve the externalized engineering memory according to the skill input requirements.

[0150] Protocol endpoints are determined in advance. For tasks with a clear source, such as quality issue tickets from QMS or test alarms from SCADA, the system can first determine the source system and protocol endpoints before generating a complete runtime task description.

[0151] Risk assessment is conducted in advance. For known high-risk actions, the system can perform an initial risk assessment before memory retrieval and skill selection, thereby limiting subsequent access to data sources and callable interfaces.

[0152] Context budget allocation is executed multiple times. Context budget allocation can be executed once at the start of a task, or multiple times after memory retrieval, skill selection, protocol endpoint determination, and failure recovery.

[0153] The sandbox execution location has changed. Sandbox execution can occur after the large language model generates the action intent, or it can be performed before generating the final solution by calling the digital twin or read-only copy to obtain the verification results, and then inputting the verification results into the model to generate the solution.

[0154] The write-back process can be broken down into steps. Execution trajectory write-back can be done uniformly after the task is completed, or it can be done in stages according to model calls, memory retrieval, protocol calls, approval events, and manual feedback.

[0155] The timing of manual approval varies. Manual approval can occur before action execution, after sandbox execution, before report release, before data writing, or before task closure. Different approval timings can be configured according to risk levels and company policies.

[0156] Example of changes in system connectivity: The system described in the above embodiments adopts a layered structure consisting of a task access layer, an externalized object layer, a collaborative control layer, a controlled execution layer, and a feedback evolution layer. In other embodiments, the connection relationships of the system components can be changed as follows: The centralized deployment has shifted to a distributed deployment. The engineering externalization memory, skill asset library, protocol gateway, and collaborative controller can be deployed on the same server, or they can be deployed separately in enterprise private clouds, edge servers, workshop gateways, and data centers, communicating through message buses or service interfaces.

[0157] The connection location of the protocol gateway has changed. The protocol gateway can be directly connected to the co-controller, or it can be connected to the enterprise service bus first, and then the enterprise service bus accesses external systems such as PLM, MES, and ERP.

[0158] The connection method of large language models has changed. Large language models can be deployed as internal system modules or called as external model services; a single model can be used, or multiple models can be used to handle task parsing, summary generation, action intent generation, verification interpretation, and report generation respectively.

[0159] The connection relationships of the approval module have changed. The approval module can be part of the runtime governance module or connected independently to the enterprise OA, BPM, or PLM approval system. For read-only scenarios, the approval module can only record alarms; for high-risk write scenarios, the approval module can block the execution chain.

[0160] The sandbox module connectivity changes. Sandbox modules can connect to digital twin platforms, simulation systems, read-only database copies, process engine pre-simulation environments, or container isolation environments. Different sandboxes can be dynamically selected by the collaboration controller based on task type.

[0161] Changes in the observation and write-back connection relationships are observed. The observation and trajectory write-back module can be directly written to the externalized engineering memory, or it can be first written to the log platform, data lake, or audit platform, and then new memory objects, skill scores, and protocol routing strategies can be generated by offline tasks.

[0162] Examples of changes in governance strategies: In the above embodiments, the risk levels are divided into L1 to L4. In other embodiments, the implementation of the governance strategy can be varied as follows: The number of risk levels varies. Risk levels can be divided into three, four, five, or more levels, or a continuous risk score can be used. Different companies can configure level thresholds according to their quality systems, information security policies, and production safety policies.

[0163] Changes in approval strategies. Approvals can consist of single-person approval, multi-person countersigning, approval by a professional manager, approval by a quality manager, approval by an airworthiness or classification society role, approval by a customer representative, or approval by automated rules.

[0164] Sandbox strategies can vary. Different risk levels can correspond to different sandbox strategies, such as read-only query sandbox, data replica sandbox, digital twin sandbox, process rehearsal sandbox, or complete isolation sandbox.

[0165] Rollback strategy changes. For rollbackable software configurations or process drafts, rollback points can be saved; for non-rollbackable physical manufacturing actions, a difference summary can be saved, automatic execution can be disabled, and manual confirmation is required before execution by an external system.

[0166] Changes in audit granularity. The system can record brief call logs, or it can record complete model inputs and outputs, retrieval evidence, skill versions, protocol parameters, approval comments, sandbox results, and manual modification traces.

[0167] Examples of changes in application scenarios: The above examples illustrate three scenarios: cruise ships, airplanes, and large generators. In other embodiments, the invention can also be applied to rail transit vehicles, heavy-duty gas turbines, nuclear power equipment, construction machinery, wind power equipment, marine engineering platforms, complex production lines, semiconductor manufacturing equipment, and large-scale medical equipment.

[0168] In different scenarios, equipment objects, lifecycle stages, engineering externalization memory fields, skill asset types, and industrial protocol endpoints can be replaced accordingly. For example: In the context of rail transit vehicles, the equipment may include the car body, bogie, traction system, braking system, and signaling system. In the context of wind power equipment, the equipment objects can include blades, hubs, gearboxes, generators, converters, and towers. In the context of nuclear power equipment, the equipment may include pressure vessels, steam generators, main pumps, piping systems, and safety-grade instrumentation and control systems. In the context of semiconductor manufacturing equipment, the equipment objects can include cavities, vacuum systems, motion platforms, optical systems, and process modules.

[0169] The aforementioned changes in application scenarios do not alter the fundamental technical concept of this invention, which unifies the scheduling of engineering memory, skill assets, industrial protocol endpoints, context budgets, risk approval, sandbox execution, and execution feedback through an externalized module collaborative control diagram.

Claims

1. A collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation, characterized in that, include: Step S1: Receive the task of manufacturing large and complex equipment; Step S2: Identify the project context associated with the task and generate project context information; Step S3: Generate a runtime task description based on the project context information; Step S4: Based on the runtime task description, retrieve the externalized memory object from the externalized memory bank and generate a budget-aware memory snapshot; Step S5: Based on the runtime task description and the budget-aware memory snapshot, select skill asset objects from the skill asset library and generate skill runtime fragments; Verify whether the input data requirements of the candidate skill asset object have been satisfied by the budget-aware memory snapshot or the runtime task description; Based on the comprehensive score, risk level, and the context budget, the disclosure level of the target skill asset is determined, and the disclosure level includes at least a summary level, a constraint level, and a detailed rule level; the operational steps, decision-making rules, prohibited items, and approval requirements in the target skill asset are trimmed into skill runtime fragments; Step S6: Based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, determine the industrial protocol endpoint object from the industrial protocol endpoint library and generate protocol call constraints; Step S7: Construct an externalized module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraints, the context budget, the permission constraints, the approval strategy, and the risk level; Step S8: Based on the externalization module collaborative control diagram, generate the runtime context of the large language model agent, and input the runtime context into the large language model agent to obtain the structured action intent output by the large language model agent; Step S9: Perform verification on the structured action intent, the verification including at least one of protocol verification, permission verification, risk verification, approval verification and sandbox verification; Step S10: After the structured action intent passes the verification, the controlled operation is executed and a structured execution trajectory is generated; The structured execution trajectory is written back to the engineering externalization memory, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

2. The equipment manufacturing full life cycle collaborative control method based on intelligent external operation according to claim 1, characterized in that, When generating the budget-aware memory snapshot in step S4, the method further includes: filtering the retrieval request based on user role, project permissions, and data security level; normalizing the retrieval results from different data sources into a unified format of externalized project memory objects; and performing a comprehensive scoring and sorting of the externalized project memory objects, prioritizing the retention of several memory objects with the highest comprehensive scores under the constraints of the context budget.

3. The equipment manufacturing full life cycle collaborative control method based on intelligent external operation according to claim 1, characterized in that, When determining the industrial protocol endpoint object and generating protocol call constraints in step S6, the process further includes: the industrial protocol gateway recalling candidate protocol endpoint objects based on system type, capability tags, and input / output modes; mapping internal data objects to the interface call format of the target industrial system; performing a comprehensive score on the candidate protocol endpoint objects; determining the target protocol endpoint object based on the comprehensive score and permission filtering; and generating the protocol call constraints, which at least include the allowed methods, parameter ranges, and whether approval or sandboxing is required.

4. The equipment manufacturing full life cycle collaborative control method based on intelligent external operation according to claim 1, characterized in that, Step S7, which involves constructing a collaborative control diagram for externalized modules, specifically includes: writing the engineering externalized memory object from the budget-aware memory snapshot into the control diagram as a memory node; writing the skill asset object corresponding to the skill runtime fragment into the control diagram as a skill node; writing the industrial protocol endpoint object corresponding to the protocol call constraint into the control diagram as a protocol endpoint node; establishing dependency edges based on the correspondence between the input data requirements of the skill node and the evidence reference of the memory node; establishing call edges based on the correspondence between the call parameters of the protocol endpoint node and the output format of the skill node; and establishing approval gate edges and sandbox edges based on the risk level and the approval strategy.

5. The equipment manufacturing full life cycle collaborative control method based on intelligent external operation according to claim 1, characterized in that, When generating the runtime context of the large language model agent in step S8, the method further includes: assembling the content summary in the budget-aware memory snapshot, the operation steps and decision rules in the skill runtime segment, and the input / output patterns in the protocol call constraints into the runtime context according to the allocation ratio of the context budget, based on the externalization module collaborative control diagram. The runtime context also includes prohibited items and approval requirements.

6. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described in claim 1, characterized in that, The risk verification in step S9 also includes: calculating a risk score based on the action type risk, object sensitivity, lifecycle stage, user permission level, and historical failure rate; and determining the execution mode based on the risk score. The execution mode includes automatic execution for low risk, generating only suggestions or drafts for medium risk, requiring manual approval for execution for high risk, and allowing verification only in a sandbox or digital twin environment for extremely high risk.

7. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described in claim 1, characterized in that, The sandbox verification in step S9 further includes: when the structured action intent is determined to require sandbox execution, pre-executing the structured action intent in a read-only data copy, a digital twin simulation environment, or a container isolation environment; if the sandbox execution is successful and the result meets expectations, then execution continues; if the sandbox execution fails or the result is abnormal, then execution is rejected and the reason for failure is returned.

8. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described in claim 1, characterized in that, When performing the controlled operation in step S10, the structured action intent is processed according to its type: when it is a query operation, the data is obtained by calling the read-only interface of the target system through the industrial protocol gateway; when it is an analysis operation, the simulation or computing service is called to generate an analysis report or solution suggestion. When an operation is initiated for an approval flow, an approval draft is generated and submitted to the enterprise approval system; when a data modification operation is initiated, the controlled write is performed by the industrial protocol gateway only after manual approval and sandbox verification.

9. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, In step S10, when writing the structured execution trajectory back to the externalized engineering memory, the model call records, memory hit records, protocol call results, approval events, and manual feedback in the structured execution trajectory are transformed into new externalized engineering memory objects, and the success or failure status of the execution result and the manual confirmation opinions are recorded.

10. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, In step S10, when writing the structured execution trajectory back to the skill asset library, the historical success rate and comprehensive score of the skill asset object are updated based on the skill call results, human feedback and final conclusion in the structured execution trajectory; when the skill asset object continuously receives low scores or fails in multiple similar tasks, its ranking weight is reduced or it is marked as requiring manual review.

11. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, In step S10, when writing the structured execution trajectory back to the industrial protocol endpoint library, the historical availability score and failure rate of the industrial protocol endpoint object are updated according to the protocol call results, failure reasons and compensation strategy execution status in the structured execution trajectory; when the number of consecutive failures of a certain protocol endpoint object exceeds the threshold, it is marked as unavailable or downgraded to a backup endpoint.

12. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described in claim 1, characterized in that, It also includes a context budget joint allocation step: when generating the runtime task description in step S3, the allocation ratio of the context budget among the engineering externalized memory, skill assets, protocol constraints and output format is determined according to the task type, risk level and historical recovery cost; In step S4, the budget-aware memory snapshot is cropped according to the allocation ratio; In step S5, the disclosure level of the skill asset object is determined according to the allocation ratio.

13. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, It also includes lifecycle phase state machine maintenance steps: pre-defining the phase sequence and transformation relationship of the entire lifecycle of large and complex equipment manufacturing; when generating runtime task descriptions in step S3, determining the scope of accessible data sources, the set of callable protocol endpoints, the set of loadable skill assets, and the approval strategy based on the current lifecycle phase; when the structured action intent involves cross-phase operation, triggering phase adaptation verification, if the current phase does not allow the action to be executed, then refusing to execute or requiring manual confirmation.

14. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation as described in claim 1, characterized in that, If any of the protocol verification, permission verification, risk verification, approval verification, or sandbox verification in step S9 fails, the system refuses to execute the structured action intent and generates a rejection reason explanation; the rejection reason explanation and the structured action intent are recorded together in the structured execution trajectory and written back to the project externalization memory.

15. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, The large-scale complex equipment manufacturing task includes at least one of the following tasks: design review task, process planning task, bill of materials and drawing consistency review task, design change impact analysis task, assembly interference analysis task, quality anomaly 8D analysis task, test fault location task, operation and maintenance fault knowledge feedback task, and production plan impact assessment task.

16. The collaborative control method for the entire lifecycle of equipment manufacturing based on intelligent external operation according to claim 1, characterized in that, The large language model agent does not directly access external industrial systems, and its output structured action intent must be verified in step S9 before it can be executed; the external industrial systems include at least one or more of the following: product lifecycle management system, manufacturing execution system, enterprise resource planning system, quality management system, computer-aided design system, and monitoring and data acquisition system.

17. A collaborative control system for the entire lifecycle of equipment manufacturing based on intelligent external operation, characterized in that: The equipment manufacturing lifecycle collaborative control method based on intelligent external operation, as described in any one of claims 1 to 16, comprises: The task access module is used to receive manufacturing tasks for large and complex equipment. The task parsing and equipment object identification module is used to identify the engineering context associated with the task and generate engineering context information; The runtime task description generation module is used to generate runtime task descriptions based on the project context information. The project externalization memory module is used to retrieve project externalization memory objects from the project externalization memory library according to the runtime task description, and generate a budget-aware memory snapshot based on the context budget; The skill asset management module is used to select skill asset objects from the skill asset library based on the runtime task description and the budget-aware memory snapshot, and generate skill runtime fragments. The industrial protocol gateway module is used to determine industrial protocol endpoint objects from the industrial protocol endpoint library based on the runtime task description, the budget-aware memory snapshot, and the skill runtime fragment, and to generate protocol call constraints. The externalization module collaborative control module is used to construct an externalization module collaborative control diagram based on the runtime task description, the budget-aware memory snapshot, the skill runtime fragment, the protocol call constraint, context budget, permission constraint, approval strategy, and risk level. The large language model intelligent agent module is used to output structured action intentions based on the runtime context generated by the collaborative control diagram of the externalization module; The verification module is used to perform at least one of the following on the structured action intent: protocol verification, permission verification, risk verification, approval verification, and sandbox verification; The sandbox execution module is used to pre-execute structured action intents that need to be verified in a sandbox environment; The observation and trajectory write-back module is used to collect the execution trajectory after performing a controlled operation and write the execution trajectory back to the engineering externalized memory library, the skill asset library, the industrial protocol endpoint library, and the runtime strategy library.

18. An electronic device comprising at least one processor, at least one memory, and at least one communication interface, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the collaborative control method for the entire life cycle of equipment manufacturing based on intelligent external operation as described in any one of claims 1 to 16.

19. 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 collaborative control method for the entire life cycle of equipment manufacturing based on intelligent external operation as described in any one of claims 1 to 16.

Images