Large model decision control method and device, edge security execution method and system

Abstract

The application discloses a large model decision control method and device, an edge security execution method and system. The method comprises the following steps: constructing a multi-dimensional semantic constraint space fusing a physical topology and time sequence logic, and synchronizing state data of a bottom device; in response to a natural language business instruction, anchoring to the multi-dimensional semantic constraint space, generating a structured constraint prompt input, and calling a large language model to generate an initial decision strategy; importing a digital twin deduction environment for simulation verification, when a preset constraint is violated, extracting conflict feature data and converting the conflict feature data into a semantic modification feedback that can be parsed by the large language model; driving the large language model to perform iterative modification until a verified target decision strategy is generated; and compiling into bottom control instructions, which are executed after being verified by an edge gateway. The application improves the accuracy and safety of large model decision, reduces decision risk, and is suitable for complex industrial production scenarios.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing, and in particular to a large-scale model decision control method and device, and an edge security execution method and system. Background Technology

[0002] With the increasing complexity of industrial production processes, higher demands are placed on the real-time performance, accuracy, and security of decision control. Traditional industrial decision control methods largely rely on pre-set rules and models, making it difficult to cope with dynamically changing production scenarios and complex business needs. Large Language Models (LLMs), with their powerful natural language understanding and generation capabilities, bring new opportunities to industrial decision control. However, the application of LLMs in industrial scenarios faces challenges such as semantic understanding bias, difficulty in ensuring decision security, and complex interaction with underlying devices. Summary of the Invention

[0003] This application provides a large-scale model decision control method and apparatus, and an edge security execution method and system.

[0004] According to a first aspect of this application, this application provides a large-scale model decision control method, comprising: A multidimensional semantic constraint space integrating physical topology and temporal logic is constructed. The status data of the underlying devices is synchronized to the multidimensional semantic constraint space in real time through a data alignment gateway to establish a mapping relationship between the underlying device entities and the multidimensional semantic constraint space. In response to a natural language business instruction, the natural language business instruction is anchored to the multidimensional semantic constraint space. Based on the anchoring result, a structured constraint prompt input is generated and a large language model is invoked to generate an initial decision strategy. The initial decision strategy is imported into a digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, the corresponding conflict feature data is extracted. Based on the topological relationship of the multidimensional semantic constraint space, the conflict feature data is converted into semantic correction feedback that can be parsed by a large language model. Based on the semantic correction feedback, the large language model is driven to iteratively correct the initial decision strategy and repeatedly perform simulation verification until a verified target decision strategy is generated. The target decision strategy is compiled into low-level control instructions, which are then sent to the corresponding device control unit for execution after pre-execution security verification by the edge gateway. The pre-execution security verification includes at least checking the target device's operating status and security interlock signals.

[0005] In some embodiments, the multidimensional semantic constraint space includes an industrial semantic constraint graph, which includes multiple nodes and associated constraint edges connecting the nodes; The nodes include work order nodes, process nodes, equipment nodes, tooling nodes, material nodes, personnel nodes, and safety constraint nodes; The associated constraint edges include: process timing edges representing timing constraints; equipment capability edges representing physical constraints; resource occupancy edges representing resource constraints; and mutual exclusion constraint edges representing security constraints.

[0006] In some embodiments, the step of converting the conflict feature data into semantic correction feedback that can be parsed by a large language model based on the topological relationship of the multidimensional semantic constraint space includes: Based on the node correlation of the industrial semantic constraint graph, the affected entity nodes are located from the conflict feature data; Tracing back along the associated constraint edges, we can determine the upstream cause of the conflict and the affected downstream objects. Based on the upstream cause and the downstream object, a natural language or structured description containing conflict description, affected entities, taboo operations, and adjustment suggestions is generated as the semantic correction feedback.

[0007] In some embodiments, the step of responding to a natural language business instruction, anchoring the natural language business instruction to the multidimensional semantic constraint space, generating structured constrained prompt input, and invoking a large language model to generate an initial decision strategy includes: Identify the business entities, scheduling targets, and constraints from the natural language business instructions; Match the business entity with the corresponding node in the industrial semantic constraint graph; The scheduling objective is converted into a specific optimization function or priority rule; Based on the matched nodes and the optimization function or rule, the constraint context required to generate the decision strategy is retrieved from the graph.

[0008] In some embodiments, the step of driving the large language model to iteratively revise the initial decision strategy based on the semantic correction feedback, and repeatedly performing simulation verification until a verified target decision strategy is generated, includes: The target decision-making strategy is verified and confirmed by the simulation of the digital twin simulation environment, and the target decision-making strategy can be successfully executed under the predetermined physical constraints, timing constraints, resource constraints and security constraints.

[0009] In some embodiments, the pre-execution security verification via the edge gateway includes: After receiving the underlying control command, the edge gateway checks the real-time online status, operating status, network security status, and interlock security signals of the target device. When a preset prohibition condition is detected to be triggered, hardware interlock is executed to prevent the issuance of instructions, and the reason for the failure is reported upwards.

[0010] According to a second aspect of this application, this application provides an edge-secure execution method for large model decision instructions, applied to an edge gateway, comprising: Collect real-time status data from underlying devices, convert it into standardized status attributes, and synchronize it to the multi-dimensional semantic constraint space of the upper-level computing system; Receive verified decision control commands issued by the host computing system; The decision control instruction undergoes a cross-level pre-execution status verification; if the verification fails, a hardware interlock is activated and the instruction is refused to be executed. If the verification passes, the instruction is converted into a set of action instruction streams that can directly drive physical devices, sent to the production terminal, and the execution result is monitored in a closed loop.

[0011] According to a third aspect of this application, this application provides a large-scale model decision control device, comprising: The constraint construction module is used to construct a multi-dimensional semantic constraint space that integrates physical topology and temporal logic. The status data of the underlying devices is synchronized to the multi-dimensional semantic constraint space in real time through the data alignment gateway to establish a mapping relationship between the underlying device entities and the multi-dimensional semantic constraint space. The strategy generation module is used to respond to natural language business instructions, anchor the natural language business instructions to the multidimensional semantic constraint space, generate structured constraint prompts based on the anchoring results, and call the large language model to generate an initial decision strategy. The strategy inference module is used to import the initial decision strategy into the digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, it extracts the corresponding conflict feature data and converts the conflict feature data into semantic correction feedback that can be parsed by the large language model based on the topological relationship of the multidimensional semantic constraint space. The strategy optimization module is used to drive the large language model to iteratively revise the initial decision strategy based on the semantic correction feedback, and repeatedly perform simulation verification until a verified target decision strategy is generated. The instruction issuing module is used to compile the target decision strategy into low-level control instructions, and after pre-execution security verification, issue them to the corresponding device control unit for execution. The security verification includes at least checking the target device's operating status and security interlock signals.

[0012] According to a fourth aspect of this application, this application provides a large-scale model decision control system, comprising: The upper-level computing node is used to implement the large-scale model decision control method described above; A data alignment gateway is used to collect and transform data from underlying devices and synchronize it to the multidimensional semantic constraint space of the upper-level computing node. Edge gateways are used to implement edge-secure execution methods for large-scale model decision instructions as described above. A cluster of production equipment is used to execute the action command stream issued by the edge gateway; The management terminal is used to display scheduling strategies and execution results, and to receive user commands.

[0013] According to a fifth aspect of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the method described above.

[0014] Due to the adoption of the above technical solutions, the beneficial effects of this application are as follows: The large-scale model decision control method and apparatus, and edge security execution method and system provided in this application include: constructing a multi-dimensional semantic constraint space that integrates physical topology and temporal logic, and synchronizing the status data of the underlying devices; responding to natural language business instructions, anchoring to the multi-dimensional semantic constraint space, generating structured constraint prompt input, and calling the large language model to generate an initial decision strategy; importing the large language model into a digital twin simulation environment for simulation verification, and when the preset constraints are violated, extracting conflict feature data and converting it into semantic correction feedback that can be parsed by the large language model; driving the large language model to perform iterative correction until a verified target decision strategy is generated; compiling the large language model into underlying control instructions, and issuing them for execution after security verification by the edge gateway. This application constructs a multi-dimensional semantic constraint space to associate natural language business instructions with entities and constraints in industrial production scenarios, improving the accuracy of the large language model's understanding of business instructions. By using a digital twin inference environment to simulate and iteratively correct the initial decision-making strategy, the feasibility and security of the decision-making strategy are ensured. The security verification before execution by the edge gateway further improves the security of decision execution and avoids production accidents caused by decision-making errors. This application improves the accuracy and security of large model decision-making, reduces decision-making risks, and is applicable to complex industrial production scenarios.

[0015] ‌ Attached Figure Description

[0016] Figure 1 A flowchart of one implementation of the method provided in this application embodiment; Figure 2 A flowchart of a sub-step in one implementation of the method provided in this application embodiment; Figure 3A flowchart of a sub-step in another implementation of the method provided in this application embodiment; Figure 4 A flowchart of another implementation of the method provided in the embodiments of this application; Figure 5 A schematic diagram of the program modules of the apparatus provided in one embodiment of this application; Figure 6 This is a schematic diagram of the modules of the system provided in one embodiment of the present application. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0018] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0019] The serial numbers assigned to components in this article, such as "first" and "second", are used only to distinguish the objects being described and have no sequential or technical meaning.

[0020] Furthermore, the technical features and solutions described herein can be combined in any suitable manner in one or more embodiments. It will be readily understood by those skilled in the art that the steps or order of operations related to the embodiments provided herein can also be changed. Therefore, any order in the drawings and embodiments is for illustrative purposes only and does not imply a requirement to follow a particular order, unless explicitly stated otherwise.

[0021] like Figure 1As shown, the large-scale model decision control method provided in this application embodiment is particularly applicable to, but not limited to, production scheduling scenarios. One implementation includes the following steps: Step 110: Construct a multi-dimensional semantic constraint space that integrates physical topology and temporal logic. Use a data alignment gateway to synchronize the status data of the underlying devices to the multi-dimensional semantic constraint space in real time, so as to establish a mapping relationship between the underlying device entities and the multi-dimensional semantic constraint space.

[0022] In this system, physical topology expresses the connections between nodes and their physical locations, while temporal logic expresses the order of actions, including what to do first, what to do next, and which actions cannot be performed in parallel. This application first establishes a digital knowledge model of the industrial system. This model is not a simple database, but a structured semantic network that integrates equipment physical topology and temporal logic. Through a data alignment gateway (supporting multiple industrial protocols such as OPC UA, Modbus, and PLC interfaces), the system collects real-time operational status data from underlying devices (such as robots, machine tools, sensors, and PLC controllers), parses and converts this data into standardized attribute tensors (e.g., [device ID, coordinates (x, y, z), status, temperature...]), and synchronously maps it to the corresponding nodes in the constraint graph, ensuring that the graph is a dynamic and accurate reflection of the physical world. By synchronizing the status data of the underlying devices to the multidimensional semantic constraint space in real time through the data alignment gateway, a mapping relationship can be established between the underlying device entities and the nodes in the multidimensional semantic constraint space, thereby achieving real-time perception and management of the underlying device status.

[0023] In some embodiments, the multidimensional semantic constraint space includes an industrial semantic constraint graph, which is constructed with nodes as basic units and constraint edges as links. Nodes are the core elements of the graph, and constraint edges are the logical links between nodes, that is, they include multiple nodes and the associated constraint edges connecting the nodes. Nodes can include work order nodes, process nodes, equipment nodes, tooling nodes, material nodes, personnel nodes, and safety constraint nodes; Associative constraint edges can include: Process timing edges, which represent timing constraints, are used to define the sequential execution logic between process nodes; Equipment capability edges that characterize physical constraints are used to establish adaptive associations between equipment nodes and work order / process nodes; Resource occupancy edges, representing resource constraints, are used to mark the real-time binding relationship between equipment / tooling nodes and work order nodes; Mutually exclusive constraint edges, representing security constraints, are used to limit the rules that cannot be executed simultaneously between specific groups of nodes.

[0024] Step 120: In response to natural language business instructions, anchor the natural language business instructions to a multi-dimensional semantic constraint space, generate structured constraint prompts based on the anchoring results, and call the large language model to generate an initial decision strategy.

[0025] In some embodiments, constrained input prompts may include one or more of the following: business entity field, objective function field, device status field, resource availability field, security constraint field, and output format constraint field. Initial decision-making strategies may include one or more of the following: work order sorting, device allocation, process scheduling start and end times, and exception handling actions.

[0026] Upon receiving a natural language business instruction, the system first identifies the business entity, scheduling objective, and constraints from the instruction. Then, it matches the business entity with corresponding nodes in the industrial semantic constraint graph, converting the scheduling objective into a specific optimization function or priority rule. Next, based on the matched nodes and optimization functions or rules, it retrieves the constraint context required to generate the decision strategy from the graph, generating a structured, constrained input prompt. Finally, this input prompt is used to call the large language model to generate the initial decision strategy.

[0027] Users issue business instructions in natural language, such as, "Prioritize urgent orders delayed due to equipment failure to available equipment of the same type, and minimize the number of mold changes." The system anchors this instruction to a multi-dimensional semantic constraint space; in one implementation, it can be anchored to an industrial semantic constraint graph. Specifically, this includes: Intent analysis: Identify the business objects (such as "urgent orders", "faulty equipment", "available equipment"), scheduling objectives ("prioritize production", "reduce mold changes") and implicit constraints in the instructions.

[0028] Context generation: Based on anchored graph nodes and associated constraint edges, the system constructs a constrained prompt input, which is a structured context containing a list of pending work orders, real-time equipment status, available resources, material availability, safety restrictions, and clear output format constraints (such as requiring LLM to output a structured JSON scheduling scheme). This greatly guides and limits the output scope of LLM, enabling it to generate context within the rule framework.

[0029] Subsequently, the LLM is invoked to generate an initial decision strategy based on this strongly constrained context. This initial decision strategy is structured and machine-readable, for example, a scheduling Gantt chart data structure with one-to-one correspondence between work orders, processes, equipment, and planned timestamps.

[0030] like Figure 2 As shown, in some embodiments, step 120 may specifically include the following steps: Step 1201: Identify business entities, scheduling objectives, and constraints from natural language business instructions; Step 1202: Match the business entities with the corresponding nodes in the industrial semantic constraint graph; Step 1203: Convert the scheduling objective into a specific optimization function or priority rule; Step 1204: Based on the matched nodes and optimization functions or rules, retrieve the constraint context required to generate the decision strategy from the graph.

[0031] Step 130: Import the initial decision strategy into the digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, extract the corresponding conflict feature data. Based on the topological relationship of the multidimensional semantic constraint space, convert the conflict feature data into semantic correction feedback that can be parsed by the large language model.

[0032] Existing technologies typically only output simulation results as failure indicators or error logs, failing to enable large language models to understand the underlying industrial semantics of conflicts. This application utilizes the topological relationships of a multidimensional semantic constraint space to transform conflict feature data into semantic correction feedback containing conflict descriptions, affected entities, taboo operations, and adjustment suggestions. This allows large language models to perform targeted corrections based on the causes of simulation failures, thereby forming a verifiable closed-loop control strategy.

[0033] In some embodiments, conflict feature data may include one or more of the following: conflict type, set of conflicting entities, conflict time interval, conflict location, identifier of violated constraint rule, and conflict severity. An example of conflict feature data is shown below: {Conflict Type: Resource Mutual Exclusion, Conflicting Entities: [M01, A, B], Time: 10:00-11:00, Violated Rule: R-001}. Specifically, the initial decision strategy is imported into a digital twin simulation environment for simulation verification, simulating the execution of the strategy in a real production scenario. When the simulation environment detects that the strategy execution violates preset constraints, the corresponding conflict feature data is extracted. Then, based on the topological relationship of the multidimensional semantic constraint space, the affected entity nodes are located from the conflict feature data, and the upstream causes and affected downstream objects are determined by tracing along the associated constraint edges. Finally, based on the upstream causes and downstream objects, a natural language or structured description containing a conflict description, affected entities, taboo operations, and adjustment suggestions is generated as semantic correction feedback. This application's embodiments utilize an industrial semantic constraint graph to provide clear physical rule boundaries for LLM decision-making, precisely anchoring ambiguous natural language instructions to specific, computable industrial entities and relationships. Digital twin simulation can pre-test and eliminate infeasible solutions before they have real-world consequences.

[0034] In some embodiments, step 130, importing the initial decision strategy into a digital twin simulation environment for simulation verification, may specifically include: The digital twin simulation environment is equipped with a discrete event simulation engine, which is used to perform virtual simulation and dynamic evaluation of the execution process of the initial decision-making strategy based on discrete event sequences, and accurately simulate resource consumption, waiting queues and path conflicts.

[0035] like Figure 3 As shown, step 130, based on the topological relationships of the multidimensional semantic constraint space, converts the conflict feature data into semantic correction feedback that can be parsed by a large language model. Specifically, this may include the following steps: Step 1301: Based on the node correlation of the industrial semantic constraint graph, locate the affected entity nodes from the conflict feature data; Step 1302: Trace along the associated constraint edges to determine the upstream cause of the conflict and the affected downstream objects; Step 1303: Based on the upstream cause and the downstream object, generate a natural language or structured description containing conflict description, affected entities, taboo operations and adjustment suggestions as semantic correction feedback.

[0036] Step 140: Based on semantic correction feedback, drive the large language model to iteratively correct the initial decision strategy and repeat the simulation verification until a verified target decision strategy is generated.

[0037] Specifically, semantic correction feedback is input into the large language model, driving iterative correction of the initial decision strategy. The corrected decision strategy is then imported back into the digital twin inference environment for simulation verification, and this process is repeated until a verified target decision strategy is generated. The target decision strategy is verified by simulation in the digital twin inference environment and can be executed smoothly under predetermined physical constraints, temporal constraints, resource constraints, and security constraints. In other words, the target decision strategy is verified by the digital twin inference environment to have no preset physical constraint conflicts, temporal constraint conflicts, resource occupation conflicts, or security interlock conflicts.

[0038] The generated initial strategy needs to be pre-executed in the digital world, such as through simulation verification, to identify potential physical conflicts. Simulation verification inputs the initial decision strategy into a digital twin simulation environment (typically a discrete event simulation engine). This environment simulates the entire process of equipment occupancy, material flow, and process execution based on the definition of the constraint graph. When the simulation detects a violation of preset constraints (such as equipment conflict, incomplete material supply, or security zone intrusion), it generates a low-level feature tuple (e.g., conflict entity set: [Equipment M01, Work Order A, Work Order B], conflict time: 10:00-11:00, conflict type: resource mutual exclusion). The key to this application is that instead of directly feeding these machine codes back to the large model, it reverse-translates them into semantic correction feedback that the LLM can understand based on the relationships in the semantic constraint graph. For example: "Work order A and work order B are scheduled to occupy equipment M01 simultaneously between 10:00 and 11:00, violating the resource mutual exclusion rule. Work order A is an urgent order, so it is recommended to retain its occupation and relocate work order B to equipment M02 (idle) or delay its execution until 11:00." This human-readable "feedback" is input into the large language model, driving it to revise the initial strategy based on specific problems and improvement suggestions. The revised strategy then re-enters the simulation verification stage. This closed-loop process of "Generation (LLM) -> Verification (Digital Twin) -> Feedback (Semantic Transformation) -> Re-revision (LLM)" continues until a target decision strategy that has passed all constraint checks is obtained.

[0039] By performing cross-modal inverse transformation of conflict features, the barrier between machine language (simulation error code) and the semantic language of large models is broken down, enabling large models to understand "why they are wrong" and "how to correct them," thereby enabling them to perform self-correction and optimization under given constraints.

[0040] In some embodiments, based on semantic correction feedback, the large language model is driven to iteratively correct the initial decision strategy and repeatedly perform simulation verification until a verified target decision strategy is generated. Specifically, this may include: The target decision-making strategy has been verified and confirmed by simulation in the digital twin simulation environment, and the target decision-making strategy can be successfully executed under the predetermined physical constraints, timing constraints, resource constraints and security constraints.

[0041] In some embodiments, when the discrete event simulation engine detects a violation of preset constraints in strategy execution during simulation, it extracts corresponding conflict feature data, which may specifically include: When the discrete event simulation engine detects a virtual resource conflict or deadlock event, it records the specific coordinates and timestamps of the conflict event on the simulation timeline to form the conflict feature data.

[0042] Step 150: Compile the target decision strategy into low-level control instructions, perform pre-execution security verification through the edge gateway, and then send them to the corresponding device control unit for execution. The pre-execution security verification includes at least checking the target device's operating status and security interlock signals.

[0043] The introduction of edge-side pre-execution security checks and hardware interlocks constitutes a third physical security defense line, in addition to semantic security (graph constraints) and logical security (simulation verification). Even if the upper-layer intelligent decision-making logic passes all verifications, if an anomaly is detected at the moment before execution (such as sudden equipment failure or emergency stop activation), the system can immediately block dangerous instructions, achieving intrinsic security.

[0044] In some embodiments, the underlying conflict feature data discovered by digital twin inference is converted into semantic correction feedback that can be parsed by a large language model through the topological relationship of a multidimensional semantic constraint space. This enables the large model to understand the cause of the error and to implement a targeted correction strategy, which is then executed after edge security verification.

[0045] Specifically, the upper-level computing node compiles the target decision strategy into control instructions that the underlying devices can recognize and execute. These instructions are then sent to the edge gateway. Upon receiving the instructions, the edge gateway checks the target device's real-time online status, operational status, network security status, and interlock security signals. When a preset prohibition condition is detected, hardware interlocking is executed to prevent the instruction from being sent, and the reason for the failure is reported upwards. If the verification passes, the instruction is sent to the corresponding device control unit for execution. The verified target strategy is thus transformed from an abstract decision into an executable action.

[0046] The target strategy is compiled into instructions that the underlying devices can understand (such as PLC control code, device motion instructions, message queue instructions, etc.). These instructions are not issued directly, but are first passed to the edge gateway, where cross-layer security checks are performed before execution. Status pre-verification: Check whether the target device is online, whether the network is unobstructed, and whether it is currently in a state that can accept commands (non-emergency stop, safety door closed, etc.).

[0047] Interlock verification: Verify the physical interlock signals to ensure there are no safety risks (such as personnel accidentally entering the work area).

[0048] Only after all security checks are passed will the edge gateway securely send the command to the device control unit for execution and monitor the execution results, thus forming a closed business loop.

[0049] In this embodiment, from instruction parsing, strategy generation, simulation conflict to final instruction issuance, the decision context, cause, constraints and correction path of each step are clearly recorded, forming a complete decision feature fingerprint and audit log, which facilitates accident attribution, model evaluation and system optimization, improves the interpretability and traceability of the system, and ensures the absolute security of the execution end.

[0050] In some embodiments, compiling the target decision strategy into low-level control instructions, performing pre-execution security verification via an edge gateway, and then sending them to the corresponding device control unit for execution may further include: After the underlying control commands are issued and executed, the entire chain of decision data from the generation and deduction of the strategy to its execution is collected; Aggregate and extract features from the entire decision-making data to generate a decision feature fingerprint that can uniquely identify the decision-making process. The decision feature fingerprint and its associated metadata are written into the blockchain ledger or solidified into a read-only storage medium to form an immutable decision evidence record.

[0051] In some embodiments, the large model decision control method provided in this application may further include: recording natural language business instructions, constrained prompt input, initial decision strategy, conflict feature data, semantic feedback signals, target decision strategy, state pre-verification results and execution feedback results to form a traceable decision link record.

[0052] In some embodiments, performing pre-execution security verification via an edge gateway may include the following steps: After receiving control commands from the lower-level layer, the edge gateway checks the target device's real-time online status, operational status, network security status, and interlock security signals. When a preset prohibition condition is detected, it executes a hardware interlock and prevents commands from being sent, while simultaneously reporting the reason for the failure upwards. In one implementation, the edge gateway's prohibition conditions include emergency stop triggering, security door opening, device offline, target device status mismatch, abnormal interlock signal, and abnormal network heartbeat.

[0053] The methods and systems provided in this application are applicable to a variety of fields with extremely high requirements for automation and safety, such as discrete manufacturing (e.g., machining, automobile assembly), process manufacturing (e.g., chemical engineering, metallurgy), logistics warehousing and scheduling, and energy allocation.

[0054] The following are specific application examples of intelligent production scheduling in automobile manufacturing plants: In a car manufacturing plant, the large-scale model decision control method of this invention is applied to achieve intelligent production scheduling. The system first constructs an industrial semantic constraint graph containing nodes such as work orders, processes, equipment, and materials, and then synchronizes the operating status of equipment such as welding robots, painting lines, and final assembly conveyors in real time through a data alignment gateway.

[0055] When the production supervisor issues a natural language instruction to "complete the welding of the B model's emergency order within 2 hours," the system automatically identifies the business entity and scheduling target, anchors them to a multi-dimensional semantic constraint space, generates structured prompts, calls the large language model, and outputs an initial decision strategy—schedule the No. 3 welding robot cluster to prioritize handling this task.

[0056] The strategy was imported into a digital twin simulation environment for verification. The system detected that robot No. 3 would enter the scheduled maintenance window in 10 minutes, resulting in an execution conflict. Subsequently, the conflict feature data was extracted, the affected nodes were located based on the graph topology, and semantic correction feedback was generated: "Adjust task allocation, activate the backup robot No. 5, and reschedule the originally postponed tasks."

[0057] The large language model iteratively optimizes the strategy based on the correction instructions, and generates the target decision strategy after multiple simulations. This strategy is compiled into low-level control instructions, which are then sent to the execution unit after the edge gateway verifies that the device status and safety interlock signals are normal. Ultimately, the welding task of vehicle model B is completed on time while ensuring equipment safety, without any manual intervention throughout the entire process.

[0058] like Figure 4 As shown in the figure, the edge security execution method for large model decision instructions provided in this application embodiment is applied to an edge gateway. One implementation includes the following steps: Step 410: Collect real-time status data from the underlying devices, convert it into standardized status attributes, and synchronize it to the multi-dimensional semantic constraint space of the upper-level computing system; Step 420: Receive the verified decision control command issued by the host computer system; Step 430: Perform cross-level pre-execution status verification on the decision control instruction; if the verification fails, activate the hardware interlock and refuse to execute the instruction; Step 440: If the verification passes, the instruction is converted into a set of action instruction streams that can directly drive physical devices, sent to the production terminal, and the execution result is monitored in a closed loop.

[0059] like Figure 5 As shown in the embodiment of this application, the large model decision control device includes, in one implementation, a constraint construction module 510, a strategy generation module 520, a strategy deduction module 530, a strategy optimization module 540, and an instruction issuance module 550.

[0060] The constraint construction module 510 is used to construct a multi-dimensional semantic constraint space that integrates physical topology and temporal logic. It uses a data alignment gateway to synchronize the status data of the underlying devices to the multi-dimensional semantic constraint space in real time, so as to establish a mapping relationship between the underlying device entities and the multi-dimensional semantic constraint space. The strategy generation module 520 is used to respond to natural language business instructions, anchor the natural language business instructions to the multidimensional semantic constraint space, generate structured constraint prompts based on the anchoring results, and call the large language model to generate an initial decision strategy. The strategy inference module 530 is used to import the initial decision strategy into the digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, it extracts the corresponding conflict feature data and converts the conflict feature data into semantic correction feedback that can be parsed by the large language model based on the topological relationship of the multidimensional semantic constraint space. The strategy optimization module 540 is used to drive the large language model to iteratively revise the initial decision strategy based on semantic correction feedback, and repeatedly perform simulation verification until a verified target decision strategy is generated. The instruction issuing module 550 is used to compile the target decision strategy into low-level control instructions, and after pre-execution security verification, issue them to the corresponding device control unit for execution. The security verification includes at least checking the target device's operating status and safety interlock signals.

[0061] like Figure 6 As shown in the embodiment of this application, one implementation of the large model decision control system includes a host computing node 610, a data alignment gateway 620, an edge gateway 630, a production equipment cluster 640, and a management terminal 650.

[0062] The upper-level computing node 610 is used to implement the large-model decision control method described above; The data alignment gateway 620 is used to collect and transform data from underlying devices and synchronize it to the multidimensional semantic constraint space of the upper computing node 610. Edge gateway 630 is used to implement the edge-secure execution method of large model decision instructions as described above; Production equipment cluster 640 is used to execute the action command stream issued by edge gateway 630; The management terminal 650 is used to display scheduling strategies and execution results, and to receive user commands.

[0063] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the method described above. Those skilled in the art will understand that all or part of the steps of the various methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk or optical disk, etc.

[0064] The above description, in conjunction with specific embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. Those skilled in the art to which this application pertains can make several simple deductions or substitutions without departing from the concept of this application.

Claims

1. A large-scale model decision control method, characterized in that, include: A multidimensional semantic constraint space integrating physical topology and temporal logic is constructed. The status data of the underlying devices is synchronized to the multidimensional semantic constraint space in real time through a data alignment gateway to establish a mapping relationship between the underlying device entities and the multidimensional semantic constraint space. In response to a natural language business instruction, the natural language business instruction is anchored to the multidimensional semantic constraint space. Based on the anchoring result, a structured constraint prompt input is generated and a large language model is invoked to generate an initial decision strategy. The initial decision strategy is imported into a digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, the corresponding conflict feature data is extracted. Based on the topological relationship of the multidimensional semantic constraint space, the conflict feature data is converted into semantic correction feedback that can be parsed by a large language model. Based on the semantic correction feedback, the large language model is driven to iteratively correct the initial decision strategy and repeatedly perform simulation verification until a verified target decision strategy is generated. The target decision strategy is compiled into low-level control instructions, which are then sent to the corresponding device control unit for execution after pre-execution security verification by the edge gateway. The pre-execution security verification includes at least checking the target device's operating status and security interlock signals.

2. The method as described in claim 1, characterized in that, The multidimensional semantic constraint space includes an industrial semantic constraint graph, which includes multiple nodes and associated constraint edges connecting the nodes. The nodes include work order nodes, process nodes, equipment nodes, tooling nodes, material nodes, personnel nodes, and safety constraint nodes; The associated constraint edges include: process timing edges representing timing constraints; equipment capability edges representing physical constraints; and resource occupancy edges representing resource constraints. Mutually exclusive constraint edges that characterize safety constraints.

3. The method as described in claim 2, characterized in that, The topological relationship based on the multidimensional semantic constraint space, which converts the conflict feature data into semantic correction feedback that can be parsed by a large language model, includes: Based on the node correlation of the industrial semantic constraint graph, the affected entity nodes are located from the conflict feature data; Tracing back along the associated constraint edges, we can determine the upstream cause of the conflict and the affected downstream objects. Based on the upstream cause and the downstream object, a natural language or structured description containing conflict description, affected entities, taboo operations, and adjustment suggestions is generated as the semantic correction feedback.

4. The method as described in claim 2, characterized in that, The response to the natural language business instruction involves anchoring the natural language business instruction to the multidimensional semantic constraint space, generating structured constrained prompt input based on the anchoring result, and invoking the large language model to generate an initial decision strategy, including: Identify the business entities, scheduling targets, and constraints from the natural language business instructions; Match the business entity with the corresponding node in the industrial semantic constraint graph; The scheduling objective is converted into a specific optimization function or priority rule; Based on the matched nodes and the optimization function or rule, the constraint context required to generate the decision strategy is retrieved from the graph.

5. The method as described in claim 1, characterized in that, The step of driving the large language model to iteratively revise the initial decision strategy based on the semantic correction feedback, and repeatedly performing simulation verification until a verified target decision strategy is generated, includes: The target decision-making strategy is verified and confirmed by the simulation of the digital twin simulation environment, and the target decision-making strategy can be successfully executed under the predetermined physical constraints, timing constraints, resource constraints and security constraints.

6. The method as described in claim 1, characterized in that, The pre-execution security verification via the edge gateway includes: After receiving the underlying control command, the edge gateway checks the real-time online status, operating status, network security status, and interlock security signals of the target device. When a preset prohibition condition is detected to be triggered, hardware interlock is executed to prevent the issuance of instructions, and the reason for the failure is reported upwards.

7. A method for securely executing large model decision instructions at the edge, applied to an edge gateway, characterized in that, include: Collect real-time status data from underlying devices, convert it into standardized status attributes, and synchronize it to the multi-dimensional semantic constraint space of the upper-level computing system; Receive verified decision control commands issued by the host computing system; Perform cross-level pre-execution status verification on the decision control instructions; If the verification fails, a hardware interlock is activated and the instruction is refused to be executed; If the verification passes, the instruction is converted into a set of action instruction streams that can directly drive physical devices, sent to the production terminal, and the execution result is monitored in a closed loop.

8. A large-scale model decision control device, characterized in that, include: The constraint construction module is used to construct a multi-dimensional semantic constraint space that integrates physical topology and temporal logic. The status data of the underlying devices is synchronized to the multi-dimensional semantic constraint space in real time through the data alignment gateway to establish a mapping relationship between the underlying device entities and the multi-dimensional semantic constraint space. The strategy generation module is used to respond to natural language business instructions, anchor the natural language business instructions to the multidimensional semantic constraint space, generate structured constraint prompts based on the anchoring results, and call the large language model to generate an initial decision strategy. The strategy inference module is used to import the initial decision strategy into the digital twin inference environment for simulation verification. When the inference environment detects that the strategy execution violates the preset constraints, it extracts the corresponding conflict feature data and converts the conflict feature data into semantic correction feedback that can be parsed by the large language model based on the topological relationship of the multidimensional semantic constraint space. The strategy optimization module is used to drive the large language model to iteratively revise the initial decision strategy based on the semantic correction feedback, and repeatedly perform simulation verification until a verified target decision strategy is generated. The instruction issuing module is used to compile the target decision strategy into low-level control instructions, and after pre-execution security verification, issue them to the corresponding device control unit for execution. The security verification includes at least checking the target device's operating status and security interlock signals.

9. A large-scale model decision control system, characterized in that, include: The upper-level computing node is used to implement the large model decision control method as described in any one of claims 1-6; A data alignment gateway is used to collect and transform data from underlying devices and synchronize it to the multidimensional semantic constraint space of the upper-level computing node. An edge gateway for implementing the edge-secure execution method of large model decision instructions as described in claim 7; A cluster of production equipment is used to execute the action command stream issued by the edge gateway; The management terminal is used to display scheduling strategies and execution results, and to receive user commands.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 7.

Images