An industrial AI application design platform and method

Abstract

The application discloses an industrial AI application design platform and method, a cloud service forms a machine-readable agent constraint description through a multi-agent registration center, a natural language requirement entry generates a candidate multi-agent arrangement graph based on industrial application requirements, agent constraint descriptions, device topologies, current canvas states and tenant permission information, and converts the candidate multi-agent arrangement graph into a structured action plan containing action types, port references, parameters, runtimes, dependency relationships and rollback information. A Web visual editing interface performs arrangement according to incremental actions, a consistency checking engine checks ports, parameters, permissions, resource budgets and runtime compatibility before and after action execution; when blocking-level errors occur, only the failed actions and their dependent subgraphs are regenerated. A model-driven generation module generates target code, firmware or running configuration for a network that passes the verification, forms deployment consistency records, and deploys the network to a device end runtime after verification by an edge gateway.

Description

Technical Field

[0001] This invention belongs to the fields of industrial control, artificial intelligence, multi-agent systems, cloud-edge-device collaboration, industrial software design tools, and edge intelligence deployment technology, and specifically relates to an industrial AI application design platform and method. Background Technology

[0002] Existing industrial IDEs primarily target PLC programming, function block editing, fieldbus configuration, or cloud collaboration. They typically treat function blocks as passive software execution units, making it difficult to form an autonomous closed loop integrating perception, decision-making, execution, and collaboration. While general-purpose AI workflow platforms can orchestrate model call flows, their outputs are mostly cloud-based model links, lacking industrial control primitives, edge device deployment capabilities, firmware compilation pipelines, field communication constraints, and industrial safety governance mechanisms.

[0003] With the development of large language models and multi-agent technologies, engineers can describe industrial application requirements using natural language, and the generative model can automatically generate application processes or agent orchestration schemes. However, in industrial field scenarios, directly implementing natural language generation results carries high risks. For example, the generative model may generate non-existent agent identifiers, incorrectly connect event ports to data ports, generate configurations that do not conform to the parameter value range, or ignore tenant permissions, gateway resource budgets, communication latency budgets, or target runtime compatibility, resulting in undeployable generated results, abnormal operation after deployment, or even the generation of insecure control logic.

[0004] Furthermore, existing solutions for digital twins, MQTT communication, cross-compilation, edge gateways, model governance, and multi-tenant access control mostly exist as independent tools or modules, making it difficult to integrate them across the entire process of industrial demand input, agent design, compilation and simulation, firmware deployment, on-site monitoring, anomaly rollback, model training, and continuous optimization. Even if existing solutions can perform process verification, this verification usually occurs after the complete process has been generated. When verification fails, it often requires regenerating the entire process or manual modification, lacking pre-verification, post-verification, blocking-level error feedback, and local regeneration mechanisms for individual incremental actions.

[0005] Therefore, there is a need for a platform and method that can use the Web as a unified entry point and multiple intelligent agents as basic units to transform natural language industrial needs into verifiable, rollbackable, partially repairable, deployable, monitorable, and auditable industrial AI applications. Summary of the Invention

[0006] The purpose of this invention is to provide an industrial AI application design platform and method to solve the problems of the fragmentation between existing industrial IDEs, general AI workflow platforms, edge gateways, digital twin systems and model governance tools, as well as the difficulty in reliably deploying, verifying and repairing the results of natural language automatic orchestration and the difficulty in forming a consistent design and deployment record.

[0007] To achieve the above objectives, this invention provides an industrial AI application design platform, including a cloud service, an edge gateway, and a device runtime. The cloud service serves as a unified design entry point, receiving industrial application requirements, maintaining multi-agent registration information, and performing natural language orchestration, visual incremental editing, consistency verification, partial regeneration, model-driven generation, and deployment management. The edge gateway communicates with the cloud service and the device runtime, undertaking cloud deployment tasks and performing local data acquisition, local rule judgment, local AI inference, on-site protocol bridging, network outage autonomy, and recovery synchronization. The device runtime hosts multi-agent instances generated and deployed by the platform.

[0008] The cloud service includes a natural language requirement entry point, a multi-agent registry center, a web-based visual editing interface, a consistency verification engine, a model-driven generation module, and a deployment management module. The multi-agent registry center maintains the capability declarations, port definitions, parameter constraints, target runtime support information, resource consumption information, version information, and tenant permission information of multi-agents, and forms machine-readable agent constraint descriptions based on the above information. These agent constraint descriptions are used to constrain the generation of candidate multi-agent orchestration graphs by the natural language requirement entry point and are used by the consistency verification engine to perform action-level verification.

[0009] The Natural Language Requirements Entry (NLE) generates at least one candidate multi-agent orchestration diagram based on industrial application requirements, agent constraint descriptions, device topology, current canvas state, and tenant permission information. The candidate multi-agent orchestration diagram includes a set of agents to be instantiated, port connection topology, parameter configurations, runtime goals, and deployment constraints. The NLE further converts the candidate multi-agent orchestration diagram into a structured action plan. The structured action plan includes multiple incremental actions ordered by dependencies. Each incremental action includes at least one or more of the following: action type, target agent identifier, port reference, parameter key-value pair, target runtime, dependent action identifiers, and failure rollback information.

[0010] The web-based visual editing interface executes incremental actions one by one according to the structured action plan, instantiating multi-agent networks, binding resources, setting target runtimes, configuring parameters, establishing port connections, and configuring the event state machine. Unlike generating the final canvas all at once, this invention breaks down the natural language generation results into verifiable, executable, and rollbackable incremental actions, enabling the system to detect errors in a timely manner during action execution and preventing errors from spreading to the complete multi-agent network.

[0011] Before and / or after each incremental action, the consistency verification engine verifies the agent identifier, port direction, port type, data type, parameter range, permissions, resource budget, communication latency budget, and target runtime compatibility based on the agent constraint description, device topology, current canvas state, tenant permission information, and target runtime constraints. When the consistency verification engine generates a blocking-level verification result, it converts the blocking-level verification result into a local regeneration constraint and feeds it back to the natural language requirement entry point. Local regeneration constraints include one or more of the following: prohibited connection relationships, parameter boundaries, alternative agent ranges, runtime filtering conditions, resource limits, or dependent subgraph ranges.

[0012] The natural language requirement entry point is based on local regeneration constraints. It regenerates local structured action plans only for failed incremental actions and their dependent subgraphs corresponding to blocking-level verification results, and then merges the regenerated local structured action plans into the original structured action plans. Through this method, the present invention avoids regenerating the entire canvas for each verification failure, reducing generation and verification costs, minimizing disturbances to already verified parts, and improving the deployability and maintainability of natural language orchestration results in industrial scenarios.

[0013] The model-driven generation module converts the multi-agent network, which undergoes consistency verification, into target code, firmware, or runtime configuration, and generates a deployment consistency record. This record includes the multi-agent network model, target runtime, resource budget, version information, and the hash value of the generated result. The deployment management module deploys the target code, firmware, or runtime configuration, along with the deployment consistency record, to the device runtime via an edge gateway. Before deployment, the edge gateway verifies the deployment consistency record to confirm the consistency between the cloud-based design results, the generated results, and the on-site deployment content.

[0014] This invention also provides an industrial AI application design method, comprising: receiving industrial application requirements and obtaining agent constraint descriptions, device topology, current canvas state, and tenant permission information from a multi-agent registry center; generating at least one candidate multi-agent orchestration diagram based on the above information; determining a structured action plan based on the candidate multi-agent orchestration diagram; executing incremental actions one by one in a Web visual editing interface according to the structured action plan; performing consistency verification before and / or after each incremental action; when a blocking-level verification result occurs, converting the blocking-level verification result into a local regeneration constraint, and regenerating a local structured action plan only for the failed incremental action and its dependent subgraph based on the local regeneration constraint; converting the multi-agent network that has passed the consistency verification into target code, firmware, or runtime configuration, and generating a deployment consistency record; verifying the deployment consistency record through an edge gateway and deploying it to the device for runtime; during runtime, the edge gateway performs local data collection, local rule judgment, local AI inference, and network outage autonomy, and synchronizes the local trigger record to the cloud service after the network is restored.

[0015] The beneficial effects of this invention are as follows:

[0016] (1) By describing the generation of candidate multi-agent orchestration graphs through agent constraints, the natural language generation process can perceive ports, parameters, permissions, resources and runtime constraints, reducing the probability of generating unusable agents, incorrectly connected ports or out-of-bounds parameters.

[0017] (2) By using structured action plans, the candidate multi-agent orchestration graph is split into incremental actions with dependencies, target objects, port references, parameter configurations, runtime targets and rollback information, so as to make the natural language orchestration results executable, verifiable and rollbackable.

[0018] (3) By performing pre- and post-verification checks before and after incremental action execution, errors can be intercepted in time before entering the complete multi-agent network, thereby improving the security and deployment success rate of automatic orchestration of industrial AI applications.

[0019] (4) By converting the blocking-level verification results into local regeneration constraints, and only regenerating the local structured action plan for the failed incremental action and its dependent subgraph, the computational overhead caused by the overall regeneration and the risk of the correctly constructed canvas structure being destroyed are reduced.

[0020] (5) By deploying consistency records, the multi-agent network model, target runtime, resource budget, version information and generated result hash value are bound together, thereby improving the consistency, auditability and traceability between cloud design, model-driven generation, edge deployment and device operation.

[0021] (6) Improve the continuous operation capability of industrial sites under network anomalies by using distributed Watch, local AI inference and network outage self-governance mechanism, and ensure that trigger records after network recovery can be synchronized and audited. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the industrial AI application design platform of the present invention;

[0023] Figure 2 This is a schematic diagram of the standardized multi-agent structure of the present invention;

[0024] Figure 3 This is a schematic diagram of the natural language consensus orchestration and incremental verification process of this invention;

[0025] Figure 4 This is a schematic diagram of the model-driven generation and deployment consistency verification process of the present invention;

[0026] Figure 5 This is a schematic diagram of the distributed Watch network outage self-governance process of the present invention;

[0027] Figure 6 This is a schematic diagram of the edge gateway cluster collaboration process of the present invention;

[0028] Figure 7 This is a flowchart of the industrial AI application design method of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0030] like Figure 1 As shown, the industrial AI application design platform of this invention includes cloud services, edge gateways, and device runtime. The cloud services can be deployed in a public cloud, private cloud, or enterprise server to provide a unified entry point for industrial AI application design. The edge gateway can be a physical edge gateway or a cross-platform virtual gateway running on a Windows platform, Linux platform, or other operating system platform. The device runtime can run on a PLC, industrial controller, embedded device, industrial PC, or edge computing device.

[0031] The cloud services include a natural language request entry point, a multi-agent registry, a web-based visual editing interface, a consistency verification engine, a model-driven generation module, and a deployment management module. The multi-agent registry stores information about the multi-agents that the platform can call, including capability declarations, port definitions, parameter constraints, target runtime support information, resource usage information, version information, and tenant permission information. Based on this information, the multi-agent registry generates machine-readable agent constraint descriptions. These descriptions can be represented in JSON, YAML, XML, Protocol Buffers, or database record format.

[0032] In one implementation, the agent constraint description includes agent type, agent version, capability label, input port list, output port list, port direction, port category, data type, event type, parameter name, parameter type, parameter value range, default parameter value, target runtime supported list, CPU usage estimate, memory usage estimate, communication bandwidth estimate, permission domain, and tenant access scope. Through the agent constraint description, the platform can determine, when generating orchestration schemes, whether a particular agent can meet the requirements of the industrial task, whether it can be deployed to the target runtime, whether it exceeds the resource budget, and whether it complies with tenant permissions.

[0033] like Figure 2 As shown, the multi-agent system includes a perception interface, a decision engine, an execution interface, a goal and policy interface, and a collaboration interface. The perception interface is used to receive device data, message data, or model output data; the decision engine is used to generate control, diagnostic, or scheduling results based on the goal and policy; the execution interface is used to output control commands, alarm information, or data write commands; and the collaboration interface is used to interact with other multi-agent systems through event ports, data ports, shared states, or message topics.

[0034] In one implementation, the multi-agent registry categorizes multi-agents into task-oriented multi-agents, resource-oriented multi-agents, algorithm-oriented multi-agents, collaborative multi-agents, and diagnostic multi-agents. Task-oriented multi-agents describe industrial control or business tasks, such as temperature control, valve regulation, alarm interlocking, or data archiving. Resource-oriented multi-agents encapsulate equipment, databases, gateways, or communication resources. Algorithm-oriented multi-agents encapsulate AI models, data preprocessing algorithms, or control algorithms. Collaborative multi-agents perform task decomposition, collaborative scheduling, or conflict resolution. Diagnostic multi-agents perform anomaly identification, root cause analysis, or risk assessment.

[0035] like Figure 3As shown, the natural language requirement entry point receives industrial application requirements input by engineers. This entry point can encapsulate the industrial application requirements, agent constraint descriptions, device topology, current canvas state, and tenant permission information into industrial domain prompts, and invoke at least two generative models to output candidate multi-agent orchestration diagrams. The candidate multi-agent orchestration diagrams can include the set of agents to be instantiated, port connection topology, parameter configurations, runtime objectives, and deployment constraints.

[0036] The natural language processing (NLP) requirement entry point can include a multi-model consensus unit. The multi-model consensus unit determines the consensus orchestration graph based on the similarity of agent sets, port connection topology, parameter configuration, and target runtime consistency among candidate multi-agent orchestration graphs. Agent set similarity is determined by the degree of overlap in agent types, capability tags, and resource binding objects in the candidate graphs; port connection topology similarity is determined by the matching degree of the source agent, source port, target agent, and target port of the connection edges; parameter configuration similarity is determined by the matching degree of parameter names, parameter values, and parameter ranges; and target runtime consistency is determined by whether the target runtimes of each agent in the candidate graphs are consistent or compatible. When the consensus orchestration graph meets a preset consensus threshold, the multi-model consensus unit outputs the consensus orchestration graph; when the preset consensus threshold is not met, regeneration or manual confirmation can be triggered.

[0037] The Natural Language Processing (NLP) entry point transforms the consensus orchestration diagram into a structured action plan. Instead of generating the final canvas all at once, the structured action plan breaks down the consensus orchestration diagram into multiple incremental actions ordered by dependencies. Incremental actions include creating agent instances, deleting agent instances, binding resources, setting the target runtime, configuring parameters, establishing port connections, configuring the event state machine, and configuring rollback actions. Each incremental action can include an action type, target agent identifier, port reference, parameter key-value pairs, target runtime, dependent action identifiers, and failure rollback information.

[0038] For example, an incremental action to create an agent instance may include the action type `create_agent`, the target agent identifier `temp_collector_01`, the agent type `temperature-collecting agent`, the target runtime `edge_runtime_A`, and the failure rollback information `delete_agent`. An incremental action to establish a port connection may include the action type `connect_port`, the source agent identifier `temp_collector_01`, the source port `temperature_out`, the target agent identifier `threshold_checker_01`, the target port `temperature_in`, the data type `float`, and the dependent action identifier. An incremental action to configure a parameter may include the action type `set_param`, the target agent identifier `threshold_checker_01`, the parameter name `threshold`, the parameter value `85`, the parameter unit `degrees Celsius`, and the parameter range constraint.

[0039] The web-based visual editing interface includes an agent workflow canvas, port connection area, parameter configuration area, and verification prompt area. The interface executes incremental actions step-by-step according to the structured action plan, thereby achieving automatic orchestration of natural language requirements into a multi-agent network. Because incremental actions have dependent action identifiers and failure rollback information, when an action fails to execute or verify, the platform can locate the affected action and its dependent subgraphs, and undo or partially repair the already executed actions.

[0040] The consistency verification engine consists of a pre-verification unit and a post-verification unit. The pre-verification unit verifies the uniqueness of agent identifiers, port direction matching, event port and data port matching, data type compatibility, parameter value range, tenant permissions, resource budget pre-occupancy, and target runtime support before incremental actions are executed. The post-verification unit verifies the feedback loop depth, resource conflicts, gateway memory budget, communication latency budget, deployment constraints, multi-agent network integrity, and the rollbackability of executed incremental actions after incremental actions are executed.

[0041] Consistency verification results can include pass, warning-level error, and blocking-level error. When the verification result is pass, the web-based visual editing interface continues to execute subsequent incremental actions. When the verification result is a warning-level error, the platform may allow continued execution and display a warning message in the web-based visual editing interface. When the verification result is a blocking-level error, the platform refuses to execute the corresponding incremental action, or reverts the executed incremental action that failed the subsequent verification, and converts the blocking-level verification result into a local regeneration constraint.

[0042] Local regeneration constraints can include prohibited connections, parameter boundaries, alternative agent ranges, runtime filtering conditions, resource limits, or dependent subgraph ranges. For example, when the blocking-level error is "data port connected to event port," local regeneration constraints can include prohibiting the connection of the source data port to the target event port and requiring a reselection of the data input port or the addition of an event transformation agent. When the blocking-level error is "parameter out of range," local regeneration constraints can include parameter upper limits, parameter lower limits, and recommended default values. When the blocking-level error is "target runtime does not support this agent," local regeneration constraints can include runtime filtering conditions or alternative agent ranges. When the blocking-level error is "gateway memory budget insufficient," local regeneration constraints can include resource limits, lightweight algorithm agent ranges, or deployment splitting suggestions.

[0043] The natural language requirement entry point is based on local regeneration constraints, regenerating local structured action plans only for failed incremental actions and their dependent subgraphs. The dependent subgraph can be determined based on the dependent action identifiers, port connections, and resource binding relationships in the structured action plan. The regenerated local structured action plan replaces or inserts into the original structured action plan, maintaining consistency with the incremental actions that have passed verification and are unaffected. Therefore, this invention can correct natural language generation errors while maintaining the stability of the original canvas, avoiding the uncontrollable changes caused by full reconstruction.

[0044] like Figure 4 As shown, the model-driven generation module includes a metamodel parsing unit, a template generation unit, a compilation and simulation unit, and a firmware verification unit. The metamodel parsing unit parses the multi-agent network model that has passed consistency verification and extracts agent instances, port connections, parameter configurations, runtime objectives, and deployment constraints. The template generation unit generates target code, firmware, or runtime configuration according to the target runtime. The target code can be in C++, Python, Lua, or other languages ​​adapted to the target runtime. The compilation and simulation unit performs incremental compilation, simulation testing, and compilation log analysis. The firmware verification unit generates a hash value of the generated result and binds this hash value with the multi-agent network model, target runtime, resource budget, and version information to form a deployment consistency record.

[0045] Deployment consistency records can include multi-agent network model identifiers, model versions, agent instance lists, port connection summaries, parameter configuration summaries, target runtime identifiers, resource budgets, generation result hash values, generation time, deployment batch numbers, and tenant identifiers. The deployment management module deploys the target code, firmware or runtime configuration, and deployment consistency records to the device runtime via the edge gateway. Before deployment, the edge gateway calculates the hash value of the received target code, firmware, or runtime configuration and compares it with the generation result hash value in the deployment consistency record. If the comparison passes, deployment continues; if the comparison fails, deployment is rejected, and a deployment consistency error is reported to the cloud service.

[0046] During deployment, version comparison, firmware hash verification, deployment permission verification, canary release, failure rollback, and audit log generation can be performed. Once the device receives the deployment content, it loads the corresponding multi-agent instance and performs data collection, control output, model inference, alarm reporting, or collaborative tasks.

[0047] like Figure 5 As shown, the edge gateway includes a distributed Watch engine. The distributed Watch engine receives and caches Watch rules issued by the cloud service. Watch rules can include sampling objects, trigger conditions, duration, alarm actions, write actions, recording policies, and rule versions. During network interruptions, the edge gateway continues to perform data sampling, condition judgment, alarm output, control writes, and trigger record storage based on the locally cached Watch rules. After network recovery, the edge gateway batches the trigger records to the cloud service according to rule version and timestamp, thereby ensuring the continuity of on-site rule execution and the integrity of cloud audit records.

[0048] like Figure 6 As shown, an edge gateway can include physical edge gateways and cross-platform virtual gateways. Cross-platform virtual gateways have the same device discovery, field protocol bridging, local rule enforcement, local AI inference, MQTT communication, and cluster election capabilities as physical edge gateways. Multiple edge gateways determine a master gateway through a master-slave gateway election mechanism. The master gateway performs deployment task coordination, downloads generated results, verifies deployment consistency records, and distributes generated results within the local area network. When the master gateway loses connection, the remaining edge gateways re-elect a new master gateway, thereby improving the reliability of deployment tasks and field collaboration tasks.

[0049] In one implementation, cloud services, edge gateways, and device runtimes communicate via an MQTT structured communication bus. MQTT topics can employ multi-segment namespaces that include gateway identifiers, function namespaces, device identifiers, processing types, and target ports. Access isolation between different tenants, factories, or devices can be achieved through tenant prefixes, namespace permissions, and gateway binding relationships. This communication method can also be replaced with gRPC, HTTP / REST, ZeroMQ, or other communication methods that support cloud-edge-device message interaction.

[0050] like Figure 7 As shown, the industrial AI application design method of the present invention includes the following steps.

[0051] Step S1: Receive industrial application requirements and obtain agent constraint descriptions, device topology, current canvas status, and tenant permission information from the multi-agent registry center.

[0052] Step S2: Generate at least one candidate multi-agent orchestration diagram based on industrial application requirements, agent constraint description, device topology, current canvas state, and tenant permission information.

[0053] Step S3: Determine a structured action plan based on the candidate multi-agent orchestration graph. The structured action plan includes multiple incremental actions ordered by dependencies.

[0054] Step S4: According to the structured action plan, execute incremental actions one by one in the Web visual editing interface to instantiate multiple agents, establish port connections, and configure running parameters.

[0055] Step S5: Perform a consistency check before and / or after each incremental action.

[0056] Step S6: When a blocking-level verification result occurs, the blocking-level verification result is converted into a local regeneration constraint, and based on the local regeneration constraint, a local structured action plan is regenerated only for the failed incremental action and its dependent subgraph.

[0057] Step S7: Convert the multi-agent network that has passed the consistency check into target code, firmware, or runtime configuration, and generate deployment consistency records.

[0058] Step S8: Verify the deployment consistency record through the edge gateway, and deploy the target code, firmware, or runtime configuration to the device runtime.

[0059] Step S9: During operation, the edge gateway performs local data collection, local rule judgment, local AI inference, and autonomous operation during network outages, and synchronizes the local trigger records to the cloud service after the network is restored.

[0060] The following uses a workshop temperature control system as an example to illustrate the implementation process of this invention. The engineer inputs the following into the natural language requirement entry: "Monitor the temperature of all motors in the workshop; if it exceeds 85℃ for 5 minutes, trigger an alarm and write the abnormal data to the training database." The platform reads the agent constraint descriptions from the multi-agent registry center for temperature acquisition, threshold judgment, alarm, database writing, and diagnostic agents, and simultaneously reads the device topology, current canvas status, and tenant permission information.

[0061] The natural language processing entry point invokes multiple generative models to output candidate multi-agent orchestration graphs. The multi-model consensus unit compares the agent set, port connection topology, parameter configuration, and target runtime consistency in the candidate graphs to determine the consensus orchestration graph that meets the consensus threshold. Subsequently, the natural language processing entry point breaks down the consensus orchestration graph into structured action plans. These structured action plans include incremental actions such as creating a temperature acquisition agent, binding motor temperature sensor resources, creating a threshold judgment agent, setting the threshold to 85℃, setting the duration to 5 minutes, creating an alarm agent, creating a database writing agent, establishing data port connections, establishing event port connections, and configuring failure rollback actions.

[0062] Before each node creation, connection establishment, or parameter modification, the consistency verification engine performs pre-verification. For example, when a structured action plan attempts to connect the data output port of the temperature acquisition agent to the event input port of the alarm agent, the pre-verification unit determines that the port types do not match and outputs a blocking-level error. The platform converts this blocking-level error into a local regeneration constraint, which prohibits directly connecting the temperature data port to the alarm event port and requires a threshold judgment agent or an event transformation agent to be selected between the temperature acquisition agent and the alarm agent. Based on this local regeneration constraint, the natural language requirement entry only regenerates the local structured action plan for the failed connection action and its dependent subgraph, without rebuilding the temperature acquisition agent and database writing agent that have passed the verification.

[0063] For example, when the threshold parameter is incorrectly set to 850℃ by the generated model, the pre-verification unit determines that the parameter is out of bounds based on the parameter range in the agent constraint description and outputs a blocking-level error. The platform converts this error into a local regeneration constraint that includes the upper limit, lower limit, and recommended default value of the parameter. The natural language requirement entry point regenerates the threshold parameter configuration action based on this local regeneration constraint, correcting the threshold to 85℃ or a parameter value within the allowable range.

[0064] After the wiring and parameter configuration are completed, the post-verification unit continues to verify the feedback loop depth, resource budget, communication latency budget, and target runtime compatibility. If the target edge gateway's memory budget is insufficient, the post-verification unit outputs a blocking-level error. Local regeneration constraints may include resource limits, lightweight algorithm agent scope, or deployment splitting suggestions. The natural language requirement entry point regenerates local actions involving the algorithm agent or deployment target based on the above constraints.

[0065] After successful verification, the model-driven generation module generates code, firmware, or runtime configuration adapted to the edge gateway or device runtime, and generates a deployment consistency record. The deployment consistency record binds the multi-agent network model, target runtime, resource budget, version information, and the hash value of the generated result. The deployment management module sends the generated result and the deployment consistency record to the edge gateway. Before deployment, the edge gateway calculates the hash value of the received content and compares it with the deployment consistency record; if the comparison is successful, the edge gateway deploys the generated result to the device runtime; if the comparison fails, the edge gateway rejects the deployment and reports the error to the cloud service.

[0066] After deployment, the edge gateway loads the corresponding multi-agent instances on-site. During operation, the edge gateway continuously collects temperature data and executes Watch rules. When the network is interrupted, the edge gateway continues to perform condition judgments, alarm outputs, control writes, and anomaly recording based on local cache rules; after the network is restored, the edge gateway will batch synchronize the triggered records to the cloud service according to the timestamp and rule version.

[0067] In another implementation, multiple physical edge gateways and a cross-platform virtual gateway are deployed together within the same factory network. After a firmware upgrade or configuration update task is released from the cloud, multiple gateways determine a master gateway through a master-slave gateway election mechanism. The master gateway downloads the generated results and deployment consistency records, and distributes them to the sub-gateways within the local area network after verification. When the master gateway loses connection, the sub-gateways re-elect a new master gateway and continue executing the distribution task. The cross-platform virtual gateway can participate in the election and task scheduling as a candidate gateway, thereby improving the continuity of deployment tasks in the industrial field.

[0068] In other implementations, the multi-agent runtime can employ sandboxed microservices, lightweight script runtimes, IEC 61499 function block runtimes, or embedded task runtimes; multi-agent inline inference can employ ONNX Runtime, TensorFlow Lite, OpenVINO, device NPU, or other model inference frameworks; edge gateway clusters can employ primary / standby hot standby, lightweight election, or load balancing methods; all of the above alternative methods should retain industrial security constraints such as multi-agent registration, structured action plans, consistency checks, local regeneration, access control, deployment consistency records, failure rollback, and network outage autonomy.

Claims

1. An industrial AI application design platform, characterized in that, This includes cloud services, edge gateways, and device runtime; The cloud service includes a natural language requirement entry point, a multi-agent registration center, a web-based visual editing interface, a consistency verification engine, a model-driven generation module, and a deployment management module. The multi-agent registration center is used to maintain the capability declarations, port definitions, parameter constraints, target runtime support information, resource consumption information, version information, and tenant permission information of multi-agents, and to form machine-readable agent constraint descriptions based on the capability declarations, port definitions, parameter constraints, target runtime support information, and resource consumption information. The natural language requirement entry is used to receive industrial application requirements, combine the agent constraint description, device topology, current canvas state and tenant permission information to generate at least one candidate multi-agent orchestration diagram, and convert the candidate multi-agent orchestration diagram into a structured action plan; The structured action plan includes multiple incremental actions ordered by dependencies. Each incremental action includes one or more of the following: action type, target agent identifier, port reference, parameter key value, target runtime, dependent action identifier, and failure rollback information. The Web-based visual editing interface is used to execute incremental actions one by one according to the structured action plan, in order to instantiate multiple agents, establish port connections, and configure running parameters. The consistency verification engine is used to verify the agent identifier, port direction, port type, data type, parameter range, permissions, resource budget, communication latency budget, and target runtime compatibility before and / or after each incremental action is executed, based on the agent constraint description, device topology, current canvas state, tenant permission information, and target runtime constraints. When the consistency verification engine generates a blocking-level verification result, the consistency verification engine converts the blocking-level verification result into a local regeneration constraint and feeds it back to the natural language requirement entry point. The local regeneration constraint includes one or more of the following: prohibited connection relationship, parameter boundary, alternative agent range, runtime filtering condition, resource limit or dependent subgraph range. The natural language requirement entry point, based on the local regeneration constraint, regenerates the local structured action plan only for the failed incremental action and its dependent subgraph corresponding to the blocking-level verification result, and merges the regenerated local structured action plan into the original structured action plan. The model-driven generation module is used to convert a multi-agent network that has passed consistency verification into target code, firmware, or runtime configuration, and generate a deployment consistency record including the multi-agent network model, target runtime, resource budget, version information, and generation result hash value. The deployment management module is used to deploy the target code, firmware or runtime configuration, and deployment consistency record to the device runtime via the edge gateway; The edge gateway is used to verify the deployment consistency record before deployment, and to load and run the multi-agent after deployment to perform local data collection, local rule judgment, local AI inference, on-site protocol bridging and network outage autonomy. The device runtime is used to carry multi-agent instances generated and deployed by the platform.

2. The industrial AI application design platform according to claim 1, characterized in that, The natural language requirement entry point includes a multi-model consensus unit; The multi-model consensus unit is used to encapsulate industrial application requirements, agent constraint descriptions, equipment topology, current canvas state and tenant permission information into industrial domain prompt information, and call at least two generative models to output candidate multi-agent orchestration diagrams respectively. The consensus orchestration graph is determined based on the similarity of agent sets, port connection topology, parameter configuration, and target runtime consistency among the candidate multi-agent orchestration graphs. When the consensus orchestration diagram meets the preset consensus threshold, the consensus orchestration diagram is converted into a structured action plan; when the preset consensus threshold is not met, regeneration or manual confirmation is triggered.

3. The industrial AI application design platform according to claim 1, characterized in that, The structured action plan is organized in the following order: creating an agent instance, binding resources, setting the target runtime, configuring parameters, establishing port connections, configuring the event state machine, and configuring rollback actions. The consistency verification engine selects the corresponding verification rule according to the action type of the incremental action, and when the subsequent incremental action depends on the previous incremental action, it uses the verification result of the previous incremental action as the execution condition of the subsequent incremental action.

4. The industrial AI application design platform according to claim 1, characterized in that, The consistency verification engine includes a pre-verification unit and a post-verification unit; The pre-verification unit is used to verify the uniqueness of the agent identifier, port direction matching, event port and data port matching, data type compatibility, parameter value range, tenant permissions, resource budget pre-occupancy, and target runtime support before the incremental action is executed. The post-verification unit is used to verify the feedback loop depth, resource conflicts, gateway memory budget, communication latency budget, deployment constraints, multi-agent network integrity, and rollbackability of executed incremental actions after the incremental action is executed.

5. The industrial AI application design platform according to claim 1, characterized in that, The multi-agent registration center divides multi-agents into task multi-agents, resource multi-agents, algorithm multi-agents, cooperative multi-agents, and diagnostic multi-agents. Each multi-agent system includes a perception interface, a decision engine, an execution interface, a goal and policy, and a cooperation interface. The perception interface is used to receive device data, message data, or model output data. The decision engine is used to generate control, diagnosis, or scheduling results based on objectives and strategies. The execution interface is used to output control commands, alarm information, or data writing commands. The collaboration interface is used to interact with other multi-agents through event ports, data ports, shared states, or message topics.

6. The industrial AI application design platform according to claim 1, characterized in that, The model-driven generation module includes a meta-model parsing unit, a template generation unit, a compilation and simulation unit, and a firmware verification unit. The meta-model parsing unit is used to parse the multi-agent network model that has passed the consistency check, and extract agent instances, port connections, parameter configurations, runtime goals and deployment constraints; The template generation unit is used to generate target code, firmware, or runtime configuration according to the target runtime. The compilation simulation unit is used to perform incremental compilation, simulation testing, and compilation log analysis. The firmware verification unit is used to generate a hash value of the generation result, and bind the hash value of the generation result with the multi-agent network model, target runtime, resource budget and version information to form the deployment consistency record.

7. The industrial AI application design platform according to claim 1, characterized in that, The edge gateway includes a distributed Watch engine; The distributed Watch engine is used to receive and cache Watch rules issued by the cloud service. The Watch rules include sampling objects, triggering conditions, duration, alarm actions, write actions, recording strategies, and rule versions. When the network is interrupted, the distributed Watch engine continues to perform data sampling, condition judgment, alarm output, control write value and trigger record storage based on locally cached Watch rules; Once the network is restored, the distributed Watch engine will trigger the batch synchronization of records to the cloud service according to the rule version and timestamp.

8. The industrial AI application design platform according to claim 1, characterized in that, The edge gateway includes physical edge gateways and cross-platform virtual gateways; The cross-platform virtual gateway has the same device discovery, field protocol bridging, local rule enforcement, local AI inference, MQTT communication and cluster election capabilities as the physical edge gateway; Multiple edge gateways determine the master gateway through a master-slave gateway election mechanism. The master gateway performs deployment task coordination, download of generation results, verification of deployment consistency records, and distribution of generation results within the local area network. When the master gateway loses connection, a new master gateway is re-elected.

9. An industrial AI application design method, characterized in that, The method, applied to the industrial AI application design platform according to any one of claims 1 to 8, comprises: Receive industrial application requirements and obtain agent constraint descriptions, device topology, current canvas status and tenant permission information from the multi-agent registry center; Based on the industrial application requirements, agent constraint description, device topology, current canvas state, and tenant permission information, at least one candidate multi-agent orchestration diagram is generated. A structured action plan is determined based on a candidate multi-agent orchestration graph, wherein the structured action plan includes multiple incremental actions ordered by dependencies; According to the structured action plan, incremental actions are executed one by one in the Web visual editing interface to instantiate multiple agents, establish port connections and configure running parameters; Perform consistency checks before and / or after each incremental action. When a blocking-level verification result is obtained, the blocking-level verification result is converted into a local regeneration constraint, and based on the local regeneration constraint, a local structured action plan is regenerated only for the failed incremental action and its dependent subgraph. The multi-agent network that passes the consistency check is converted into target code, firmware, or runtime configuration, and deployment consistency records are generated. The deployment consistency record is verified through the edge gateway, and the target code, firmware, or runtime configuration is deployed to the device for runtime. During operation, the edge gateway performs local data collection, local rule judgment, local AI inference, and autonomous operation during network outages, and synchronizes the local trigger records to the cloud service after the network is restored.

10. The industrial AI application design method according to claim 9, characterized in that, Generating a structured action plan includes: Call at least two generative models to output candidate multi-agent orchestration graphs; The consensus orchestration graph is determined based on the similarity of agent sets, port connection topology, parameter configuration, and target runtime consistency among candidate multi-agent orchestration graphs. When the consensus orchestration graph meets the preset consensus threshold, the consensus orchestration graph is split into a structured action plan that includes action type, target agent identifier, port reference, parameter key value, target runtime, dependent action identifier and failure rollback information. When the consensus orchestration diagram does not meet the preset consensus threshold, it is triggered to regenerate or manually confirm.

Images