Intelligent industrial operating system, control method and computer readable storage medium

Abstract

This invention provides an intelligent industrial operating system, a control method, and a computer-readable storage medium. The intelligent industrial operating system includes an interaction layer, an agent layer, a data layer, and an execution layer. The interaction layer receives raw instructions; the agent layer includes a main agent and multiple sub-agents, each sub-agent executing a sub-task dispatched by the main agent and returning a first execution result; the main agent's instruction decomposition module parses the raw instructions and, if it confirms that the raw instructions are inaccurate, queries the data layer for first feedback, guiding user interaction and converting the raw instructions into accurate instructions; the task decomposition module parses the accurate instructions into at least one sub-task and assigns it to the corresponding sub-agent, the sub-agent obtaining an execution result and returning it to the main agent; the comprehensive decision module integrates the first execution results of each sub-agent to generate a comprehensive execution result or a comprehensive decision instruction.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation and artificial intelligence, and more specifically, to an intelligent industrial operating system and its control method that enables menu-free operation through natural language interaction and multi-agent collaboration. Background Technology

[0002] In modern industrial environments, operators typically need to interact with various industrial information systems (such as MES, ERP, SCADA, and PLC) and internal enterprise knowledge bases. Users issue instructions, query information, or initiate processes through complex menu navigation, form filling, and / or clicks to complete tasks such as production monitoring, fault diagnosis, process adjustment, and data querying. This interaction method requires users to have high levels of system expertise and familiarity with the operating logic of specific software, resulting in high training costs, lengthy operation paths, low efficiency, and a tendency to make mistakes when dealing with urgent or complex tasks. Furthermore, the existence of data silos between systems is severe, and decision-making and execution require manual coordination across systems, making it difficult to achieve rapid and intelligent closed-loop responses.

[0003] Therefore, there is an urgent need for an intelligent industrial solution that can understand the natural language of the industrial field, automatically coordinate resources of multiple systems, and directly drive execution. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent industrial operating system, control method, and computer-readable storage medium to solve the aforementioned technical problems.

[0005] To achieve the above objectives, this invention proposes an intelligent industrial operating system, comprising an interaction layer, a data layer, and an intelligent agent layer. The interaction layer receives raw instructions input in natural language. The data layer connects multiple industrial information systems, an enterprise's internal knowledge base, and an industrial scenario knowledge graph. The intelligent agent layer connects to the interaction layer and the data layer, and includes a main intelligent agent and multiple sub-intelligent agents. The main intelligent agent includes an instruction parsing module, a task decomposition module, a comprehensive decision-making module, and a thought chain module. The instruction parsing module parses the original instruction, and when it is confirmed that the original instruction is inaccurate, it queries the data layer and provides initial feedback to guide user interaction in transforming the original instruction into an accurate one. The task decomposition module parses the accurate instruction into at least one sub-task and assigns it to the corresponding sub-agent based on execution priority. The corresponding sub-agent obtains the required information by calling the data layer and executes the task to obtain the first execution result, which is then returned to the main agent. The comprehensive decision-making module receives the first execution results from each sub-agent to generate a comprehensive execution result and sends it to the interaction layer. The thought chain module records and structures the internal reasoning state, decision basis, and data relationships of the instruction parsing module, the task decomposition module, and the comprehensive decision-making module in real time during the processing of user instructions, forming a traceable thought chain.

[0006] As an optional technical solution, when the instruction parsing module parses the original instruction, it extracts the key entities and intents from the original instruction and matches the key entities and intents with the standard terminology database and the industrial scenario knowledge graph in the knowledge base. When the matching confidence is lower than a first threshold, the instruction parsing module confirms that the original instruction is inaccurate and constructs a guiding question as the first feedback based on the matching candidates and related information in the knowledge base and the industrial scenario knowledge graph, and sends it to the interaction layer. After receiving the supplementary information input by the user based on the first feedback, the interaction layer updates the original instruction to form the accurate instruction, and matches the accurate instruction with the standard terminology database and the industrial scenario knowledge graph until the matching confidence reaches or exceeds the first threshold. The instruction parsing module then outputs the accurate instruction.

[0007] As an optional technical solution, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to the absence of key information, it automatically generates an interrogative instruction containing completion suggestions as the first feedback and sends it to the interaction layer based on the entity relationships and business rules in the industrial scenario knowledge graph; if the instruction parsing module identifies that the inaccuracy of the original instruction is due to the presence of semantic ambiguity, it performs semantic similarity matching based on the industrial pre-trained large model, retrieves the optimal parsing result from the industrial scenario knowledge graph, generates an interrogative instruction as the first feedback, and sends it to the interaction layer.

[0008] As an optional technical solution, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to the absence of key information, then based on the entity relationships and business rules in the industrial scenario knowledge graph, the instruction parsing module combines the default equipment status, standard range of process parameters, or historical operation records in the industrial scenario knowledge graph to complete one or more of the missing production line identifiers, equipment parameters, and process target values, and form the query-style instruction. If the instruction parsing module identifies that the inaccuracy of the original instruction is due to the presence of semantic ambiguity, then the instruction parsing module performs semantic similarity matching based on the industrial pre-trained large model, transforms the original instruction into standard terms and / or standard task templates stored in the standard terminology library and the industrial scenario knowledge graph, and forms the request confirmation instruction.

[0009] As an optional technical solution, the task decomposition module is used to analyze the dependencies between the subtasks after decomposition, determine the execution priority of each subtask based on the real-time production status and the predefined rules in the knowledge base, generate a task execution plan, and assign each subtask to the corresponding sub-agent in an orderly or parallel manner according to the task execution plan.

[0010] As an optional technical solution, the task decomposition module has a weighted scoring model. This weighted scoring model takes into account at least the criticality, resource dependency, and timeliness requirements of each subtask to rank the execution priority of each subtask in order to generate the task execution plan.

[0011] Furthermore, this invention also proposes a control method for an intelligent industrial operating system, which includes,

[0012] Step A: The interaction layer receives raw instructions input in natural language.

[0013] Step B: The instruction parsing module of the main intelligent agent in the intelligent agent layer parses the original instruction. When it is confirmed that the original instruction is inaccurate, it queries the data layer and provides the first feedback and guides the user interaction to transform the original instruction into an accurate instruction. The data layer connects multiple industrial information systems, the enterprise's internal knowledge base and industrial scenario knowledge graph.

[0014] Step C: The task decomposition module of the main agent parses the precise instruction into at least one subtask, and assigns the at least one subtask to the corresponding sub-agent in the agent layer based on the execution priority. The corresponding sub-agent obtains the required information by calling the data layer, executes the task, and returns the first execution result to the main agent; and

[0015] Step D: The main intelligent agent's integrated decision-making module receives the first execution results of each sub-intelligent agent to generate an integrated execution result and sends it to the interaction layer.

[0016] As an optional technical solution, in step B, when parsing the original instruction, the key entities and intents in the original instruction are extracted, and the key entities and intents are matched with the standard terminology database in the knowledge base and the industrial scenario knowledge graph. When the matching confidence is lower than the first threshold, the original instruction is confirmed to be inaccurate, and a guiding question is constructed as the first feedback based on the matching candidates and related information in the knowledge base and the industrial scenario knowledge graph and sent to the interaction layer. After receiving the supplementary information input by the user based on the first feedback, the instruction parsing module updates the original instruction to form the accurate instruction, and matches the accurate instruction with the labeled terminology database and the industrial scenario knowledge graph until the matching confidence reaches or exceeds the first threshold, and then outputs the accurate instruction.

[0017] As an optional technical solution, in step B, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to the absence of key information, it automatically generates an inquiry-style instruction containing completion suggestions as the first feedback and sends it to the interaction layer based on the entity associations and business rules in the industrial scenario knowledge graph; if the instruction parsing module identifies that the inaccuracy of the original instruction is due to semantic ambiguity, it performs semantic similarity matching based on the industrial pre-trained large model, retrieves the optimal parsing result from the industrial scenario knowledge graph, generates a request confirmation instruction as the first feedback, and sends it to the interaction layer.

[0018] As an optional technical solution, in step B, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to the absence of key information, it automatically completes one or more of the missing production line identifier, equipment parameters, or process target values ​​based on the entity relationships and business rules in the industrial scenario knowledge graph, combined with the default state of the equipment, the standard range of process parameters, or historical operation records in the industrial scenario knowledge graph, and forms the query sentence instruction; if it identifies that the inaccuracy of the original instruction is due to semantic ambiguity, it performs semantic similarity matching based on the industrial pre-trained large model, transforms the original instruction into the standard terminology library and the standard terms and / or standard task templates stored in the industrial scenario knowledge graph, and forms the request confirmation instruction.

[0019] As an optional technical solution, in step C, the task decomposition module analyzes the dependencies between the decomposed subtasks, determines the execution priority of each subtask based on the real-time production status and the predefined rules in the knowledge base, generates a task execution plan, and assigns each subtask to the corresponding sub-agent in an orderly or parallel manner according to the task execution plan.

[0020] As an optional technical solution, the task parsing module has a weighted scoring model. In step C, the weighted scoring model is used to sort the execution priorities of each subtask by taking into account at least the criticality, resource dependency and timeliness requirements of each subtask, so as to generate the task execution plan.

[0021] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed, implements the control method of the intelligent industrial operating system as described above.

[0022] The intelligent industrial operating system, control method, and computer-readable storage medium of the present invention can understand the potentially vague or inaccurate instructions given by the user through natural language (voice / text), guide the user's precise intentions through proactive interaction, and automatically complete the precise instructions after intelligently decomposing, optimizing and sorting them, thus eliminating the dependence on the menu hierarchy of traditional user interfaces. Attached Figure Description

[0023] Figure 1 This is a block diagram of the intelligent industrial operating system of the present invention;

[0024] Figure 2 This is a flowchart of the control method for the intelligent industrial operating system of the present invention. Detailed Implementation

[0025] To provide a further understanding of the purpose, structure, features and functions of the present invention, detailed descriptions are provided below with reference to embodiments.

[0026] Please refer to Figure 1 and Figure 2 , Figure 1 This is a block diagram of the intelligent industrial operating system of the present invention. Figure 2This is a flowchart of the control method for the intelligent industrial operating system of the present invention. The intelligent industrial interaction system 1000 of the present invention includes an interaction layer 100, a data layer 200, and an intelligent agent layer 300. The interaction layer 100 is used to receive raw instructions input in natural language. The data layer 200 connects multiple industrial information systems 210, an enterprise internal knowledge base 220, and an industrial scenario knowledge graph 230. The intelligent agent layer 300 connects the interaction layer 100 and the data layer 200, and includes a main intelligent agent 310 and multiple sub-intelligent agents 320 that interact with each other. The main intelligent agent 310 includes an instruction parsing module 311, a task decomposition module 312, a comprehensive decision-making module 313, and a thought chain module 314. The instruction parsing module 311 is used to parse the raw instructions, and when it is confirmed that the raw instructions are inaccurate, it queries the data layer 200 and provides first feedback, guiding user interaction to transform the raw instructions into accurate instructions. The task decomposition module 312 is used to parse precise instructions into at least one sub-task and assign at least one sub-task to the corresponding sub-agent 320 based on execution priority. The corresponding sub-agent 320 obtains the required information by calling the data layer 200, executes the task, obtains the first execution result, and then returns the first execution result to the main agent 310. The comprehensive decision-making module 313 is used to receive the first execution results of each sub-agent 320 to generate a comprehensive execution result or a comprehensive decision instruction and send it to the interaction layer 100. The thought chain module 314 is used to record and structure the internal reasoning state, decision basis, and data association relationship of the instruction parsing module 311, task decomposition module 312, and comprehensive decision-making module 313 during the processing of user instructions in real time, forming a traceable thought chain. The thought chain module 314 interacts with each module of the main agent 310 in real time, recording possible understanding branches and selection reasons during task parsing, recording decomposition logic and sorting basis during task decomposition, recording the reasoning path of result fusion during comprehensive decision-making, and providing explanations of the decision-making process to the user when necessary.

[0027] Furthermore, this invention also proposes a control method for an intelligent industrial operating system, which includes,

[0028] Step A (S110): The interaction layer 100 receives raw instructions input in natural language form;

[0029] Step B (S120): The instruction parsing module 311 of the main intelligent agent 310 in the intelligent agent layer 300 parses the original instruction. When it is confirmed that the original instruction is inaccurate, it queries the data layer 200 and provides the first feedback and guides the user interaction to convert the original instruction into an accurate instruction. The data layer 200 connects multiple industrial information systems, the enterprise's internal knowledge base and industrial scenario knowledge graph.

[0030] In step C (S130), the task decomposition module 312 of the main intelligent agent 310 parses the precise instruction into at least one subtask, and assigns at least one subtask to the corresponding sub-intelligent agent 320 in the intelligent agent layer 300 in combination with the execution priority. The corresponding sub-intelligent agent 320 obtains the required information by calling the data layer 200 and executes it to obtain the first execution result, and then returns the corresponding first execution result to the main intelligent agent 310.

[0031] Step D (S140): The comprehensive decision-making module 313 of the main intelligent agent 310 receives the first execution results of each sub-intelligent agent 320 to generate a comprehensive execution result and sends it to the interaction layer 100.

[0032] Furthermore, the main intelligent agent 310's thought chain module 314 records and structures the internal reasoning state, decision basis, and data correlation of the instruction parsing module 311, task decomposition module 312, and comprehensive decision-making module 313 during the processing of user instructions, forming a traceable thought chain.

[0033] In one embodiment, the interaction layer 100 may be an application or web interface (displaying a user interface) deployed on a mobile device, industrial control computer, or wearable device. It provides a voice or text input interface (e.g., the user interface includes input fields), supporting human-computer interaction in voice and / or text form to receive raw commands input by the user in natural language, such as: "Increase the production capacity of line A." Furthermore, the interaction layer 100 may display various raw command templates on the user interface based on the industrial information system 210, knowledge base 220, and industrial scenario knowledge graph 230 associated with the data layer 200, guiding the user to input corresponding raw commands by referring to the templates. A raw command template may be, for example, "Check the cause of the faulty equipment and provide a repair solution." In this invention, users do not need to memorize complex menu paths and can initiate control commands through natural language, significantly reducing the operational threshold and improving human-computer interaction efficiency. It should be noted that the descriptions of the control methods for the intelligent industrial interaction system and the intelligent industrial operating system in this invention are mutually applicable and will not be repeated here.

[0034] In one embodiment, the data layer 200 connects and aggregates information from multiple heterogeneous data sources (multiple heterogeneous industrial information systems 210). These industrial information systems 210 include, but are not limited to, manufacturing execution systems (MES), supervisory and data acquisition systems (SCADA), and enterprise resource planning systems (ERP). The manufacturing execution system provides real-time data such as production orders, work-in-process status, and equipment utilization; the supervisory and data acquisition system provides real-time equipment operating parameters (such as temperature, pressure, and speed) and alarm information; and the enterprise resource planning system provides information such as material inventory and customer order priority.

[0035] Furthermore, the data layer 200 integrates an internally built knowledge base 220 and an industrial scenario knowledge graph 230. The knowledge base 220 contains records related to people, machines, materials, methods, and environment in industrial production, such as structured knowledge like standard equipment naming, standard process parameters, and standard operating procedures. Industrial production data interacts and flows to form relationships. The industrial scenario knowledge graph 230 stores complex relationships (such as "belongs to," "connects," and "follows process") and business rules (such as "when equipment X is running at full load, the temperature must not exceed Y℃") between entities (such as equipment, products, and processes) in a graph structure. In this invention, by connecting the data layer 200 to various industrial information systems 210, the knowledge base 220, and the industrial scenario knowledge graph 230, unified data scheduling and utilization are achieved, enabling decisions to be based on comprehensive real-time and historical data, thus improving decision-making / execution quality.

[0036] In one embodiment, the agent layer 300 can be deployed on an enterprise server or in the cloud. The main agent 310 acts as the overall coordinator, and the instruction parsing module 311 can integrate industrial pre-trained large models based on the Transformer architecture and be used to parse the raw instructions input by the user in natural language. Each sub-agent 320 is a domain-specific execution unit, such as: production scheduling sub-agent, quality control sub-agent, equipment monitoring sub-agent, equipment control sub-agent, process simulation sub-agent, resource optimization sub-agent, data analysis sub-agent, data recording sub-agent, alarm handling sub-agent, report generation sub-agent, etc.

[0037] In one embodiment, the agent layer 300 further includes an agent management module 330. The agent management module 330 can maintain a dynamic agent registry. When the main agent 310 and each sub-agent 320 start up, they will register with the agent management module 330 and report their own metadata, such as: agent ID (e.g., Equipment Control Agent_V2.1), function description (e.g., "responsible for executing standard equipment speed control commands"), input / output interface protocol, current load and health status (heartbeat), etc.

[0038] Each of the aforementioned sub-agents is equipped with a dedicated analysis model corresponding to its respective domain, and can call the corresponding knowledge base 220, industrial scene knowledge graph 230 and corresponding industrial information system in the data layer 200. After receiving the sub-task assigned by the main agent 310, based on its built-in dedicated analysis model, it simultaneously calls the corresponding knowledge base 220, industrial scene knowledge graph 230 and corresponding industrial information system 210 or system interface in the data layer 200 to analyze and reason about the raw data obtained from the data layer 200, and generate a structured report containing causal analysis and / or optimization suggestions as the first execution result.

[0039] For example, the production scheduling sub-agent can be used for production planning, scheduling optimization, order progress tracking, and delivery risk assessment. It is configured with or can call advanced planning and scheduling algorithm models, and can access the first sub-knowledge base in knowledge base 220, which contains process flow, capacity data, and delivery date rules. The quality control sub-agent can be used for quality data statistical process control, anomaly pattern recognition, root cause analysis, and quality early warning. It is configured with an SPC analysis engine and defect pattern graph, and can call the second sub-knowledge base in knowledge base 220, which contains quality standards, inspection procedures, and historical failure cases. The equipment maintenance sub-agent can be used for equipment status monitoring, predictive maintenance analysis, fault diagnosis, and maintenance strategy recommendation. It is configured with an equipment health assessment model and a fault tree analysis model. It can call the third sub-knowledge base in knowledge base 220, which contains equipment manuals, inspection standards, and maintenance history. The resource optimization sub-agent can be used for consumption analysis, cost accounting, and optimization suggestions for resources such as materials, energy, and manpower. It is configured with a resource consumption model and a cost calculation model. It can call the fourth sub-knowledge base in knowledge base 220, which contains bill of materials, work quotas, and energy prices.

[0040] Based on the implementation of menu-free interaction, this invention further establishes a domain-specific sub-agent system. Through deep collaboration among domain-specific sub-agents, it achieves global optimization and decision-making across multiple dimensions such as production, quality, equipment, and resources.

[0041] In one embodiment, the original instruction is "improve overall operational efficiency." After parsing, the instruction parsing module 311 determines the original instruction to be a high-level comprehensive instruction. Simultaneously, recognizing that this high-level comprehensive instruction is imprecise, it combines the matching candidate options and related information from the enterprise's internal knowledge base 220 and the enterprise's internal industrial scenario knowledge graph 230 to decompose this high-level comprehensive instruction into, for example, "analyze production bottlenecks," "assess quality losses," "check equipment overall efficiency (OEE)," and "calculate key resource utilization." Based on this, it constructs guiding questions as the first feedback and sends them to the interaction layer 100. After the user confirms "yes" as supplementary information, the instruction parsing module 311 further analyzes the instruction in this direction. The system analyzes and breaks down precise instructions into four sub-tasks: "Analyze production bottlenecks," "Assess quality losses," "Check overall equipment efficiency (OEE)," and "Calculate key resource utilization." It then generates a task execution plan based on execution priorities and assigns each sub-task to its corresponding sub-agent. Specifically, the "Analyze and generate bottlenecks" sub-task is assigned to the production scheduling sub-agent, the "Assess quality losses" sub-task to the quality control sub-agent, the "Check overall equipment efficiency (OEE)" sub-task to the equipment monitoring sub-agent, and the "Calculate key resource utilization" sub-task to the resource optimization sub-agent. These are collaborative sub-tasks executed in parallel by sub-agents from different domains.

[0042] In one embodiment, the original instruction is "to organize the status and quantity of all devices, check the cause of failure of abnormal devices and provide solutions". The task decomposition module 312 decomposes the task into multiple sub-tasks and assigns them to the corresponding sub-agents for processing. Each sub-agent returns its first execution result to the comprehensive decision module 313. The tasks involved in this original instruction are mainly data viewing, organization and analysis, and do not involve the control of specific physical execution units or API interfaces. Therefore, the comprehensive decision module 313 receives the first execution results of each sub-agent to generate comprehensive execution results (such as corresponding chart analysis results) and sends them to the interaction layer 100 for user reference.

[0043] In one embodiment, the intelligent industrial interaction system further includes an execution layer 400, which connects to the intelligent agent layer 300 (including the main intelligent agent 310 and each sub-intelligent agent 320). The execution layer 400 consists of physical execution units such as programmable logic controllers (PLCs), industrial robots, and drivers, or API interfaces that can call other information systems. It is used to receive and execute control instructions transmitted by the intelligent agent layer 300. If the original instruction is "increase the assembly speed of assembly line A to 120 pieces / hour", which involves the control of specific physical execution units or API interfaces (e.g., controlling and adjusting the speed of a certain device), the task decomposition module 312 decomposes it into multiple sub-tasks and assigns them to the corresponding sub-intelligent agents for processing. Each sub-intelligent agent returns its corresponding first execution result to the comprehensive decision module 313. The comprehensive decision module 313 generates a comprehensive decision instruction based on this and sends it to the execution layer 400. The execution layer 400 receives and executes the comprehensive decision instruction. After the execution layer 400 executes the comprehensive decision instruction, it can send execution feedback to the main intelligent agent 310. The comprehensive decision module 313 of the main intelligent agent 310 can send the comprehensive execution result / comprehensive execution status to the interaction layer 100 based on this, such as "The speed of the assembly equipment of assembly line B has been adjusted, and the current assembly speed of assembly line A is 120 pieces / hour" for user reference.

[0044] In one embodiment, during step B, when parsing the original instruction, key entities and intents are extracted from the original instruction, and these key entities and intents are matched with the standard terminology database and industrial scenario knowledge graph in the knowledge base. When the matching confidence is lower than a first threshold, the original instruction is confirmed to be inaccurate. Based on the matching candidates and related information in the knowledge base and industrial scenario knowledge graph, a guiding question is constructed as the first feedback and transmitted to the interaction layer 100. The thought chain module 314 records the ambiguity of this match, the relevant candidate information retrieved from the knowledge base, and the logic for generating the guiding question (first feedback), forming a reasoning sub-chain of ambiguity point - retrieval information = clarification strategy. The interaction layer 100 receives supplementary information input by the user based on the first feedback. The instruction parsing module 311 parses the aforementioned supplementary information to update the original instruction and form an accurate instruction. The accurate instruction is then matched with the labeled terminology database and industrial scenario knowledge graph until the matching confidence reaches or exceeds the aforementioned first threshold, at which point the accurate instruction is output.

[0045] In one embodiment, in step B, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to missing key information, it automatically generates an interrogative instruction containing completion suggestions as the first feedback based on the entity relationships and business rules in the industrial scenario knowledge graph and transmits it to the interaction layer 100. Further, based on the entity relationships and business rules in the industrial scenario knowledge graph, and combined with the default equipment status, standard range of process parameters, or historical operation records in the industrial scenario knowledge graph, it automatically completes one or more of the missing production line identifiers, equipment parameters, and process target values, forming the aforementioned interrogative instruction.

[0046] In one embodiment, in step B, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to semantic ambiguity, then semantic similarity matching is performed based on the industrial pre-trained large model, and the optimal parsing result is retrieved from the industrial scenario knowledge graph to generate a request confirmation instruction as the first feedback and transmit it to the interaction layer 100. Further, based on the industrial pre-trained large model, semantic similarity matching is performed to transform the original instruction into standard terms and / or standard task templates stored in the standard terminology library and the industrial scenario knowledge graph, thus forming the aforementioned request confirmation instruction.

[0047] The following is an explanation with reference to a specific embodiment: Suppose a user inputs the original command via voice input through the interaction layer 100: "Increase the production capacity of line A by a little".

[0048] After receiving the original instruction, the instruction parsing module 311 performs a fault-tolerant parsing and guidance process. First, it converts the speech into text and performs natural language understanding to extract key entities and intents from the original instruction. The key entity is "Line A" (potentially referring to the production line), and the intent is "increase production capacity." The thought chain module 314 records the original instruction and marks key points at the parsing stage. Next, the instruction parsing module 311 matches the key entity "Line A" and the intent "increase production capacity" with the standard terminology database in the knowledge base 220 and the industrial scenario knowledge graph 230.

[0049] During the matching process, the instruction parsing module 311 discovered several potential ambiguities in the original instruction. "Line A" could correspond to "Assembly Line A" or "Inspection Line A," and "increasing capacity" is a vague objective, without specifying the exact amount of increase or which process parameter to adjust. Based on preset rules, the initial matching confidence level was 0.65, lower than the preset first threshold (e.g., 0.9). The instruction parsing module 311 determined the original instruction to be "inaccurate" and further analyzed / identified the type of inaccuracy. Combining the matching candidates and associated information in the industrial scenario knowledge graph 230, it was found that the "production line" entity is usually associated with adjustable parameters such as "speed," "efficiency," and "yield," but these were not specified in the original instruction. Therefore, the inaccuracy type was identified as the presence of missing key information. Simultaneously, "Line A" was ambiguous, and the inaccuracy type was identified as semantic ambiguity. The thought chain module 314 recorded the details of the matching failure and the relevant production line list, common operation types, and processing actions retrieved from the knowledge base 220 and the industrial scenario knowledge graph 230.

[0050] For inaccurate types involving semantic ambiguity, semantic similarity matching is performed based on a pre-trained industrial model. The optimal parsing result is retrieved from the industrial scenario knowledge graph 230. The original instruction is transformed into a standard terminology library and standard terms and / or standard task templates stored in the industrial scenario knowledge graph 230, and a request confirmation instruction is generated: "Do you want to adjust 'Assembly Line A' or 'Inspection Line A'?". For inaccurate types involving missing key information, the instruction parsing module 311 queries the entity relationships and business rules in the industrial scenario knowledge graph 230. Combining the historical operation records and standard process documents of "Assembly Line A" and "Inspection Line A" in the industrial scenario knowledge graph 230, it is found that "assembly speed" is a key parameter commonly used to adjust production capacity, and its standard range is 80-150 pieces / hour, with the current value being 100 pieces / hour. Based on this instruction parsing module 311, an inquiry-based instruction containing completion suggestions is generated: "If it is assembly line A, it is recommended to increase production capacity by adjusting the 'assembly speed.' The current speed is 100 pieces / hour, and the standard range is 80-150. What would you like to set it to?". Then, the request confirmation instruction and the inquiry-based instruction are integrated and sent as the first feedback to the interaction layer 100 to be presented to the user.

[0051] If the user provides supplementary information based on the initial feedback: "It's assembly line A, increase to 120," the instruction parsing module 311 integrates the original instruction and the supplementary information, updating the original instruction to form a precise instruction: "Increase the assembly speed of assembly line A from 100 pieces / hour to 120 pieces / hour." This precise instruction is then matched again with the knowledge base 220 and the industrial scenario knowledge graph 230. At this point, the key entities are clear (assembly line A, assembly speed), the intent is specific (setting the target value to 120), and the matching confidence level increases to 0.95, exceeding the first threshold. The instruction parsing module 311 then outputs this precise instruction (e.g., by sending it to the task decomposition module 312).

[0052] In one embodiment, if the matching confidence level is lower than a first threshold, the instruction parsing module 311 generates at least one guiding question based on the matching result, initiates multiple rounds of interaction with the user to clarify the instruction, until a precise instruction that meets the accuracy requirements is obtained. In other words, if the instruction parsing module 311 detects that the original instruction is inaccurate, after sending the first feedback to the interaction layer 100, it updates the original instruction based on the user's supplementary information to form a precise instruction. However, if the instruction parsing module 311 finds that there is still an inaccuracy problem when parsing the precise instruction, the instruction parsing module 311 sends the updated first feedback to the interaction layer 100 based on the current precise instruction so that the user can further supplement information. The instruction parsing module 311 is also used to maintain the coherence of the user's intent in the multi-round dialogue and perform contextual semantics based on the maintained session state to parse the final precise instruction. Furthermore, the thought chain module 314 associates and records the guiding question with the generation logic, and can also store the successful parsing of this interaction and the corresponding complete thought chain as new samples in the knowledge base to optimize the subsequent parsing model.

[0053] While Natural Language Processing (NLP) and intelligent agent technologies have been applied in the consumer sector, their direct transfer to industrial scenarios faces challenges. Industrial instructions are characterized by high specialization, precision, and strong context dependence. Original instructions often suffer from ambiguity, vagueness, or missing key information (e.g., "increase the speed of line A," "check recent faults"). In this invention, the intelligent agent layer 300 possesses instruction fault-tolerant parsing and guidance capabilities. By integrating the standard terminology database within the enterprise's internal knowledge base 220 with the industrial scenario knowledge graph 230, it can understand and tolerate the ambiguity of the original instructions. When the confidence level of the original instruction is insufficient, it automatically guides the user to provide supplementary information. Through one or more rounds of interaction, the ambiguous original instructions are transformed into precise instructions that the machine can execute accurately, ensuring the practicality of the intelligent industrial interaction system.

[0054] In one embodiment, the interaction layer 100 receives raw instructions input by the user in natural language. Natural language can be one of several languages, such as the user's native language (e.g., Chinese), English, Japanese, Spanish, etc. When the raw instructions input by the user are in a third-party language other than the user's native language, the main intelligent agent 310 first translates the third-party language into the user's native language, processes the instructions in the user's native language, and then translates the processing result back into the third-party language and sends it to the interaction layer 100 to provide feedback on the processing result.

[0055] In one embodiment, in step C, the task decomposition module 312 analyzes the dependencies between the decomposed subtasks, determines the execution priority of each subtask based on the real-time production status and predefined rules in the knowledge base, generates a task execution plan, and assigns each subtask to the corresponding sub-agent in an orderly or parallel manner according to the task execution plan. The task execution plan can be in the form of a directed acyclic graph, where nodes represent subtasks and edges represent dependencies or order relationships between tasks.

[0056] In one embodiment, the task parsing module has a weighted scoring model. In step C, the weighted scoring model is used to prioritize each subtask based on its criticality, resource dependency, and timeliness requirements, thereby generating a task execution plan. Furthermore, the criticality of each subtask has the highest weight in the weighted scoring model, so a subtask involving equipment safety, emergency production interruption, or handling of critical quality anomalies will be assigned the highest priority. During this process, the thought chain module 314 records the task decomposition logic and priority ranking criteria. Specifically, it records the process of mapping precise instructions to specific subtasks and provides a basis for the generation and priority score of each subtask.

[0057] The following explanation will continue using the example of a user inputting the original command "increase the production capacity of line A" through interaction layer 100.

[0058] After receiving the precise instruction "Increase the assembly speed of assembly line A from 100 pieces / hour to 120 pieces / hour", the task decomposition module 312 performs intelligent task decomposition and priority sorting. Specifically, the task decomposition module 312 queries the agent management module 330 to obtain the agent registry to confirm the callable sub-agents, and combines the equipment operation procedures in the knowledge base 220 and the process constraints in the industrial scenario knowledge graph 230 to parse the precise instruction into a series of logically related sub-tasks, such as:

[0059] Subtask T1: Verify the feasibility of increasing the speed of assembly line A to 120 pieces / hour under the current production order and material supply.

[0060] Subtask T2: Check whether the current status of assembly line A and related equipment (such as IC bonding machine, FPC bonding machine) allows for speed adjustment.

[0061] Subtask T3: Calculate the potential load impact on downstream processes (such as image inspection machines) after speed adjustment.

[0062] Subtask T4: Generate a specific speed control command sequence.

[0063] Subtask T5: Record the operation log and update the production performance data.

[0064] Furthermore, the task decomposition module 312 analyzes the dependencies between the decomposed subtasks. T1 and T2 are parallel checks, but T4 can only be executed after both succeed (i.e., the judgment condition is met); T3 depends on part of the result of T1 (such as the new output rate), but can be calculated in parallel while T2 is being performed; T5 is performed after all execution actions are completed.

[0065] Next, the task decomposition module 312 calls the data layer 200 (the corresponding industrial information system) to obtain the real-time production status. For example, there is currently a batch of high-priority orders being produced on assembly line A. Combining the predefined rules in the knowledge base 220 (such as "during the production of high-priority orders, parameter adjustments should prioritize ensuring continuous stability"), the task decomposition module 312 starts its weighted scoring model to prioritize each subtask. The scoring factors include: (1) Criticality: T1 (feasibility verification) and T2 (status check) are critical safety prerequisites and have high scores. (2) Resource dependence: T3 requires computational resources but is not a critical path and has a medium score. (3) Timeliness: The precise instructions provided by the user imply the expectation of quick execution, so all tasks have a high timeliness weight. After comprehensive scoring, the task decomposition module 312 sorts all subtasks in descending order, determines the execution order, and generates a task execution plan: T1 and T2 are executed in parallel immediately; T3 is started after T1 has a preliminary result; T4 is executed with high priority after both T1 and T2 return successful results; T5 is executed after T4 succeeds.

[0066] Based on the generated task execution plan, the task decomposition module 312 assigns each subtask to the corresponding sub-agent 320. Specifically, T1 is assigned to the "production scheduling sub-agent," T2 to the "equipment monitoring sub-agent," T3 to the "process simulation sub-agent," T4 to the "equipment control sub-agent," and T5 to the "data recording sub-agent." After receiving the task, each sub-agent actively calls the data layer 200 to obtain the required information (such as order data in MES and real-time status in SCADA) and executes it to obtain its first execution result. Each sub-agent 320 then returns its first execution result to the main agent 310 (the comprehensive decision-making module 313). For example, T1 returns "feasible," T2 returns "equipment ready," T3 returns "downstream can withstand, it is recommended to observe the labeling machine buffer position," T4 returns "control instructions are ready," and T5 is on standby.

[0067] Furthermore, the integrated decision-making module 313 receives these first execution results, especially the suggestion from T3, generates integrated decision-making instructions, and transmits them to the interaction layer 100 and the execution layer 400. The integrated decision-making instructions include issuing a speed adjustment command to the A assembly line controller, setting the speed to 120 pieces / hour. Simultaneously, it notifies the screen detection machine to pay attention to the status. The thought chain module 314 collaborates with the integrated decision-making module 313 to record the reasoning path of fusing the first execution results of each sub-agent and making decisions, including conflict analysis during fusion, the trade-off process, and the logical derivation steps of the final decision. In this embodiment, the integrated decision-making instructions are sent to the interaction layer 100 to inform the user of the decision and precautions, and also directly to the execution layer 400. The PLC in the execution layer 400 receives the integrated decision-making instructions and controls the motor of the A assembly line to adjust to the specified speed. After the A assembly line controller and the screen detection machine make adjustments, execution feedback is sent to the main agent 310, which then sends the integrated execution results to the interaction layer 100. In this way, the intelligent industrial interaction system 1000 automatically completes a closed loop from fuzzy language commands to physical device actions.

[0068] In existing technologies, simple dialogue systems using Natural Language Processing (NLP) lack the ability to deeply understand and utilize complex and heterogeneous industrial data (equipment status, process knowledge, historical records, etc.). They cannot automatically convert fuzzy instructions into precisely executable operation sequences, nor can they intelligently schedule multiple sub-tasks considering real-time production status and resource constraints. In this invention, the intelligent agent layer 300 not only decomposes tasks but also analyzes the logical dependencies between sub-tasks and dynamically prioritizes them based on real-time production data (such as equipment load and order urgency), generating optimal or near-optimal task execution plans. This achieves intelligent coordination between parallel and serial execution, improving overall task execution efficiency.

[0069] In one embodiment, when the interaction layer 100 presents the overall execution result and / or overall execution status, it may selectively synchronize or provide, upon request, a summary of part or all of the reasoning process generated by the thought chain module 314 for this user instruction, for the user's reference / confirmation.

[0070] In one embodiment, the user can input a new original instruction based on the comprehensive execution result fed back by the agent layer 300. For example, after receiving the instruction "organize the quantity and status of all devices" and obtaining the comprehensive execution result, the user can input a new original instruction "query the cause of the abnormal device and provide a solution". At this time, the context semantics can be performed according to the maintained session state to parse out the final precise instruction.

[0071] In one embodiment, the original instruction may be a periodic subscription instruction, such as "provide the previous day's output report to my email at 8:00 AM every day". The agent layer 300 then provides the output report to the designated email address daily according to this original instruction, and sends the comprehensive execution result (e.g., "The daily output report has been sent to your email address xxx@xxx.com") to the interaction layer 100.

[0072] In one embodiment, the control method of the intelligent industrial operating system further includes step E: collecting closed-loop data from the original instruction input, task execution to result feedback; analyzing and synthesizing the execution results; and based on the analysis results, optimizing the accuracy of fault-tolerant parsing in the instruction parsing module 311, optimizing the priority ranking rule model in the task decomposition module 312, and optimizing the processing strategies of each sub-agent. Furthermore, the data layer 200 also includes a historical interaction log library, which stores the complete thought chain formed by the aforementioned user instruction processing, along with the original instruction, system boot process, final precise instruction, task decomposition, final decision, and execution result, in the historical interaction log library.

[0073] Furthermore, based on the stored historical data of thought chains, the task parsing module 311, task decomposition module 312, and comprehensive decision-making module 313 can be optimized or provided with analogical reasoning support. In one embodiment, before the decision is generated by each module in the main intelligent agent 310, the logical consistency of the reasoning chain can be verified; or, when the decision is generated by each module in the main intelligent agent 310, the thought chains of similar historical cases can be traced back to assist the current decision.

[0074] In one embodiment, each agent in the agent layer 300 can be hot-swapped to flexibly adapt to rapidly changing technical needs and business scenarios in industrial settings. For example, after running according to comprehensive decision-making instructions for a period of time, the enterprise develops an advanced predictive maintenance sub-agent that can more accurately assess the wear condition of equipment mechanical components and predict risks before executing speed adjustment instructions. The (system administrator) deploys this newly developed predictive maintenance sub-agent to the system server. The predictive maintenance sub-agent starts up and automatically sends a registration request to the agent management module 330. After verifying its interface specifications, the agent management module 330 adds its information (identifier, version, and functional description: "predicting equipment speed adjustment risks based on vibration data") to the agent registry and sets its status to "available". The agent management module 330 also sends a notification to the main agent 310 (task decomposition module 312): "A new agent has been launched, and its functionality has been updated."

[0075] When a user inputs a new raw command in natural language, such as "increase the assembly speed of assembly line B," the task decomposition module 312, while decomposing the task, will query the agent management module 330 to obtain the latest agent registry. It will then find that for subtasks like "checking whether the equipment status allows speed adjustment," there are two candidate agents: the original "equipment monitoring sub-agent" (checking electrical and basic operating status) and the newly registered "predictive maintenance sub-agent" (assessing equipment risk). At this point, the task decomposition module 312 can dynamically adjust the task decomposition logic according to predefined strategies (such as prioritizing comprehensive risk assessment) or rules (such as forcibly adding predictive maintenance to high-speed adjustment tasks). For example, the original subtask T2 can be refined into T2a (basic status check) and T2b (predictive risk assessment), and dispatched to the original equipment monitoring sub-agent and the newly registered predictive maintenance sub-agent for parallel execution, respectively. The newly registered predictive maintenance sub-agent is invoked to obtain the required vibration history data in SCADA through the data layer 200, perform analysis, and return the first execution result (e.g., "The mechanical condition is good, and speed-up is allowed") to the main agent 310 (the integrated decision module 313).

[0076] In another embodiment, if a vulnerability is found in the existing report generation sub-agent that needs to be fixed, the agent management module 330 (via the system administrator) sends a "prepare to uninstall" command to the report generation sub-agent. Simultaneously, the agent management module 330 updates its status in the agent registry to "discontinued use" and sends a notification to the task decomposition module 312 to "temporarily not assign new tasks to the report generation sub-agent." Once the report generation sub-agent has completed its ongoing report task and safely exited, the agent management module 330 removes it from the agent registry. At the same time, the fixed new version of the report generation sub-agent can be deployed immediately and registered online.

[0077] In one embodiment, the intelligent industrial operating system 1000 can run two master agent instances, one master and one slave. When an upgrade is required, the system administrator starts a new version of the master agent instance (new instance / standby instance) through the agent management module 330. After the standby instance starts, it synchronizes the necessary runtime state (such as the ongoing multi-turn dialogue context) from the data layer 200 or the currently active master instance (old instance). After confirming that the synchronization is complete, the agent management module 330 switches the request routing of the interaction layer 100 and the data layer 200 to the standby instance, and then gradually drains and shuts down the master instance.

[0078] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed, implements the aforementioned control method of the intelligent industrial operating system.

[0079] In this invention, the intelligent parsing and proactive guidance mechanism of the instruction parsing module 311 effectively solves the ambiguity problem of natural language instructions in industrial scenarios; through the dependency analysis and dynamic priority sorting of the task decomposition module 312, efficient and reliable automated scheduling and execution of complex tasks are achieved. Throughout the process, users only need to engage in simple natural language dialogues without delving into any underlying system menus, greatly improving the intelligence level of industrial operations and the user experience.

[0080] Although the invention has been described in conjunction with the accompanying drawings, the embodiments disclosed in the drawings are intended to illustrate preferred embodiments of the invention and should not be construed as limiting the invention. The scale in the schematic drawings does not represent the actual proportions of the components, in order to clearly describe the required parts.

[0081] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. An intelligent industrial operating system, characterized in that... include: The interaction layer is used to receive raw instructions input in natural language. The data layer connects multiple industrial information systems, enterprise internal knowledge bases, and industrial scenario knowledge graphs. as well as The agent layer connects the interaction layer and the data layer. This agent layer includes a main agent and multiple sub-agents; the main agent includes... The instruction parsing module is used to parse the original instruction, and if it is confirmed that the original instruction is inaccurate, it queries the data layer and provides initial feedback and guides user interaction to convert the original instruction into an accurate instruction. The task decomposition module is used to parse the precise instruction into at least one subtask, and assign the at least one subtask to the corresponding sub-agent based on the execution priority. The corresponding sub-agent obtains the required information by calling the data layer and executes it to obtain the first execution result, and then returns the corresponding first execution result to the main agent. The integrated decision-making module is used to receive the first execution results of each sub-agent to generate an integrated execution result and / or an integrated execution state and send it to the interaction layer. as well as The thought chain module is used to record and structure the internal reasoning state, decision basis, and data relationship of the instruction parsing module, the task decomposition module, and the comprehensive decision-making module in real time during the processing of user instructions, forming a traceable thought chain.

2. The intelligent industrial interaction system according to claim 1, characterized in that, When parsing the original instruction, the instruction parsing module extracts the key entities and intents from the original instruction and matches them with the standard terminology database and the industrial scenario knowledge graph in the knowledge base. When the matching confidence is lower than a first threshold, the instruction parsing module confirms that the original instruction is inaccurate and constructs a guiding question as the first feedback based on the matching candidates and related information in the knowledge base and the industrial scenario knowledge graph, and sends it to the interaction layer. After receiving the supplementary information input by the user based on the first feedback, the interaction layer updates the original instruction to form the accurate instruction and matches the accurate instruction with the standard terminology database and the industrial scenario knowledge graph until the matching confidence reaches or exceeds the first threshold. The instruction parsing module then outputs the accurate instruction.

3. The intelligent industrial interaction system according to claim 2, characterized in that, If the instruction parsing module identifies that the inaccuracy of the original instruction is due to missing key information, it automatically generates an inquiry-style instruction containing completion suggestions as the first feedback based on the entity relationships and business rules in the industrial scenario knowledge graph and sends it to the interaction layer. If the instruction parsing module identifies that the inaccuracy of the original instruction is due to semantic ambiguity, it performs semantic similarity matching based on the industrial pre-trained large model, retrieves the optimal parsing result from the industrial scenario knowledge graph, generates a request confirmation instruction as the first feedback, and sends it to the interaction layer.

4. The intelligent industrial operating system according to claim 1, characterized in that, The task decomposition module is used to analyze the dependencies between the subtasks after decomposition, and determine the execution priority of each subtask based on the real-time production status and the predefined rules in the knowledge base, generate a task execution plan, and assign each subtask to the corresponding sub-agent in an orderly or parallel manner according to the task execution plan.

5. The intelligent industrial operating system according to claim 4, characterized in that, The task decomposition module has a weighted scoring model, which takes into account at least the criticality, resource dependency and timeliness requirements of each subtask to rank the execution priority of each subtask in order to generate the task execution plan.

6. A control method for an intelligent industrial operating system, characterized in that... include, Step A: The interaction layer receives raw instructions input in natural language. Step B: The instruction parsing module of the main intelligent agent in the intelligent agent layer parses the original instruction. When it is confirmed that the original instruction is inaccurate, it queries the data layer and provides the first feedback and guides the user interaction to transform the original instruction into an accurate instruction. The data layer connects multiple industrial information systems, the enterprise's internal knowledge base and industrial scenario knowledge graph. Step C: The task decomposition module of the main agent parses the precise instruction into at least one subtask, and assigns the at least one subtask to the corresponding sub-agent in the agent layer in combination with the execution priority. The corresponding sub-agent obtains the required information by calling the data layer and executes it to obtain the first execution result, and then returns the corresponding first execution result to the main agent. as well as Step D: The main intelligent agent's integrated decision-making module receives the first execution results of each sub-intelligent agent to generate an integrated execution result and sends it to the interaction layer; The main intelligent agent's thought chain module records and structures the internal reasoning state, decision basis, and data correlation of the instruction parsing module, task decomposition module, and comprehensive decision-making module during the processing of user instructions in real time, forming a traceable thought chain.

7. The control method according to claim 6, characterized in that, In step B, when parsing the original instruction, the key entities and intents in the original instruction are extracted, and the key entities and intents are matched with the standard terminology database in the knowledge base and the industrial scenario knowledge graph. When the matching confidence is lower than the first threshold, the original instruction is confirmed to be inaccurate, and a guiding question is constructed based on the matching candidates and related information in the knowledge base and the industrial scenario knowledge graph as the first feedback and sent to the interaction layer. After receiving the supplementary information input by the user based on the first feedback, the instruction parsing module updates the original instruction to form the accurate instruction, and matches the accurate instruction with the labeled terminology database and the industrial scenario knowledge graph until the matching confidence reaches or exceeds the first threshold, and then outputs the accurate instruction.

8. The control method according to claim 7, characterized in that, In step B, if the instruction parsing module identifies that the inaccuracy of the original instruction is due to missing key information, it automatically generates an inquiry-style instruction containing completion suggestions as the first feedback and sends it to the interaction layer based on the entity relationships and business rules in the industrial scenario knowledge graph; if the instruction parsing module identifies that the inaccuracy of the original instruction is due to semantic ambiguity, it performs semantic similarity matching based on the industrial pre-trained large model, retrieves the optimal parsing result from the industrial scenario knowledge graph, generates a request confirmation instruction as the first feedback, and sends it to the interaction layer.

9. The control method according to claim 6, characterized in that, In step C, the task decomposition module analyzes the dependencies between the decomposed subtasks, determines the execution priority of each subtask based on the real-time production status and the predefined rules in the knowledge base, generates a task execution plan, and assigns each subtask to the corresponding sub-agent in an orderly or parallel manner according to the task execution plan.

10. The control method according to claim 9, characterized in that, The task parsing module has a weighted scoring model. In step C, the weighted scoring model is used to sort the execution priorities of each subtask by taking into account at least the criticality, resource dependency and timeliness requirements of each subtask, so as to generate the task execution plan.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the control method of the intelligent industrial operating system as described in claims 6-10.

Images