An artificial intelligence-based factory intelligent scheduling method and system

Abstract

The application provides a kind of factory intelligent scheduling method and system based on artificial intelligence, which comprises the following steps: obtaining factory multi-source production data, converting it into standardized intermediate data, and constructing a dynamic knowledge graph;Response to scheduling task or disturbance event, retrieve scheduling context information from dynamic knowledge graph, generate candidate scheduling scheme, and call scheduling domain-specific language rule engine, based on rule base and real-time constraint parameter for hard constraint verification, generate structured violation report to drive targeted correction when verification fails, and convert to scheduling instruction and write into manufacturing execution system when verification is successful.Can unify the semantics of multi-source heterogeneous data, reduce the scheduling deviation caused by state lag and data island, and avoid the violation of production hard constraints by artificial intelligence generated scheme through deterministic constraint verification, reduce the risk of manufacturing execution system data pollution, improve the real-time, reliability and traceability of scheduling in dynamic disturbance scenarios.

Description

Technical Field

[0001] This invention relates to the field of factory scheduling technology, and in particular to an intelligent factory scheduling method and system based on artificial intelligence. Background Technology

[0002] As the manufacturing industry moves towards intelligence and flexibility, factory production scheduling needs to consider multiple constraints simultaneously, including work order delivery dates, process sequence, equipment availability, personnel skills and qualifications, material availability, and unforeseen disruptive events. Especially in discrete manufacturing scenarios such as automobile manufacturing, a single factory typically contains multiple production lines, various product models, multiple processes, and a large amount of equipment, personnel, and material resources. The scheduling process not only needs to balance multiple optimization objectives but also needs to adjust the original plan in a timely manner when dynamic disturbances such as equipment failures, emergency orders, material shortages, personnel absences, and rework orders occur.

[0003] Existing factories typically rely on Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), and Warehouse Management System (WMS) to manage data such as production execution, order planning, and material inventory, respectively. Because these systems are independent in terms of data structure, coding rules, field naming, and status representation, the same business object often has different data representations in different systems. This makes it difficult for scheduling systems to comprehensively model work orders, processes, equipment, materials, personnel, and constraints at a unified semantic level. Existing scheduling systems often rely on batch data extraction, transformation, and loading processes when acquiring cross-system data, resulting in data lag. The resource status, material status, and work order status used by the scheduling engine may not reflect the real-time production line status, easily leading to problems such as scheduling based on falsely available resources, inaccurate material completeness checks, and the inability to promptly transmit the impact of process dependencies.

[0004] Furthermore, existing advanced planning and scheduling systems mostly generate production plans based on static rules or static optimization models. When disturbances such as equipment failures, emergency orders, material shortages, or staff absences occur during production line operation, they are typically handled by manually triggering a full rescheduling or simply postponing the process. Full rescheduling is time-consuming, and the on-site execution lacks stable plan support during the rescheduling process; simple postponement makes it difficult to identify the indirect impacts of disturbances along the process dependency chain, material consumption chain, and resource occupation chain, which can easily cause delays to continue to propagate among downstream work orders. Therefore, existing scheduling methods cannot achieve rapid response to dynamic disturbances while ensuring the consistency of global constraints.

[0005] In recent years, artificial intelligence technologies such as large language models have been explored for use in complex task planning and production scheduling scenarios. While AI models can generate complex scheduling suggestions based on natural language tasks and contextual information, their outputs are inherently probabilistic and lack deterministic verification capabilities regarding hard production constraints such as equipment unavailability, changeover time, process sequence, material availability, personnel qualifications, and shift boundaries. If AI-generated scheduling plans are directly written into the Manufacturing Execution System (MES), issues may arise such as overlapping allocation of the same equipment, subsequent processes starting before preceding processes, scheduling work despite material shortages, and unqualified personnel being assigned to specific workstations. This can contaminate work order statuses in the MES and impact subsequent work reporting, material pull, and quality traceability. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a factory intelligent scheduling method and system based on artificial intelligence, which aims to solve the technical problems of low scheduling management efficiency and poor traceability in the existing technology.

[0007] To achieve the above objectives, in a first aspect, the present invention provides: an intelligent factory scheduling method based on artificial intelligence, comprising the following steps: Acquire multi-source production data during the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data into ontology instances based on a pre-constructed production scheduling domain ontology. A dynamic knowledge graph is constructed based on the ontology instances and the semantic relationships between them. In response to scheduling tasks or disturbance events, scheduling context information is retrieved from the dynamic knowledge graph and input into the artificial intelligence scheduling unit, which then generates candidate scheduling schemes. The production scheduling domain-specific language rule engine is invoked to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph. When the candidate production scheduling plan fails the hard constraint verification, a structured violation report is generated and fed back to the AI ​​production scheduling unit, so that the AI ​​production scheduling unit can make targeted corrections to the candidate production scheduling plan based on the structured violation report; When the candidate scheduling scheme passes the hard constraint verification, the verified candidate scheduling scheme is converted into a scheduling instruction that the manufacturing execution system can accept, and written into the manufacturing execution system in a transactional manner.

[0008] According to one aspect of the above technical solution, the step of converting the multi-source production data into standardized intermediate data with semantic annotation, and mapping the standardized intermediate data to ontology instances based on a pre-constructed production scheduling domain ontology, specifically includes: Data adapters are used to connect to the Manufacturing Execution System, Enterprise Resource Planning System, and Warehouse Management System, respectively. The acquired multi-source production data is deduplicated, filled with null values, and normalized in format. The processed multi-source production data is converted into standardized intermediate data in JSON-LD format; Based on the production scheduling domain ontology, entity extraction and entity alignment are performed on the standardized intermediate data to determine the correspondence between data records representing the same business object in different data sources. The extracted and aligned entity data is mapped to ontology instances, and the ontology instances and their semantic relationships are written into the graph database.

[0009] According to one aspect of the above technical solution, after the step of constructing a dynamic knowledge graph, the method further includes: Monitor production status change messages in the manufacturing execution system, wherein the production status change messages include at least one of equipment status change messages, work order status change messages, material status change messages, and personnel attendance status change messages; When a production status change message is detected, the affected nodes and edges in the dynamic knowledge graph are determined based on the production status change message. Incremental updates are performed on the affected nodes and edges; Based on ontology reasoning rules, reasoning is performed on the updated dynamic knowledge graph to deduce downstream processes and downstream work orders affected by equipment failure, material shortage, staff absence, or changes in work order status.

[0010] According to one aspect of the above technical solution, the step of retrieving scheduling context information from the dynamic knowledge graph in response to a disturbance event specifically includes: Obtain the event type, occurrence time, affected resources, expected duration, and severity level of the disturbance event, wherein the disturbance event includes at least one of the following: equipment failure event, emergency order insertion event, material shortage event, personnel absence event, and rework order insertion event; Based on the semantic relationships of the affected resources in the dynamic knowledge graph, the directly affected processes and work orders are determined; Deduce the indirectly affected downstream processes and downstream work orders by following the sequence of processes, material consumption relationships, and resource utilization relationships; The scope of the partial rearrangement is determined based on the directly affected processes and work orders, as well as the indirectly affected downstream processes and work orders. Local rearrangement is triggered for the local rearrangement range, and the scheduling domain-specific language rule engine is called to perform hard constraint verification before the local rearrangement result is written into the manufacturing execution system.

[0011] According to one aspect of the above technical solution, the production scheduling domain-specific language rule base includes at least one of the following statement structures: The CONSTRAINT statement is used to define hard constraint assertions that the production scheduling plan must meet; The RULE statement is used to define soft business rules that are recommended to be executed when triggering conditions are met; The RESOURCE statement is used to declare the resources participating in production scheduling and their initial state; The WINDOW statement is used to define the effective time range of a production schedule. The PRIORITY statement is used to define the weighting coefficients for multi-objective optimization.

[0012] According to one aspect of the above technical solution, a language-specific rule engine for production scheduling is invoked to perform hard constraint verification on the candidate production scheduling schemes, including: Load the currently valid rule set from the language-specific rule base for the production scheduling domain; The currently valid rule set is parsed into an internal constraint representation; Read the real-time constraint parameters corresponding to the internal constraint representation from the dynamic knowledge graph; Based on the internal constraint representation and the real-time constraint parameters, at least one of the following is performed on the candidate production scheduling scheme: equipment non-preemption constraint verification, changeover time constraint verification, process sequence constraint verification, delivery date constraint verification, material completeness constraint verification, personnel qualification constraint verification, shift boundary constraint verification, and buffer capacity constraint verification. When a constraint violation occurs, a structured violation report is generated. The structured violation report includes a violation rule identifier, the work order involved, the equipment involved, the violation time window, and correction guidelines.

[0013] According to one aspect of the above technical solution, the artificial intelligence scheduling unit includes a planning agent, an evaluation agent, and an execution agent; The planning agent is used to generate candidate production scheduling schemes based on the production scheduling context information; The evaluation agent is used to evaluate the candidate production scheduling scheme based on at least two of the following indicators: delivery date fulfillment rate, equipment utilization rate, changeover frequency, and material availability rate, to obtain the production scheduling evaluation result, and to generate improvement suggestions based on the production scheduling evaluation result. The planning agent is also used to make local or global corrections to the candidate production scheduling schemes based on the improvement suggestions. The execution agent is used to call the domain-specific language rule engine of the scheduling domain to perform hard constraint verification on the modified candidate scheduling scheme, and after the verification is passed, convert the modified candidate scheduling scheme into a scheduling instruction that the manufacturing execution system can accept.

[0014] According to one aspect of the above technical solution, the step of inputting the production scheduling context information into the artificial intelligence production scheduling unit, and having the artificial intelligence production scheduling unit generate candidate production scheduling schemes, specifically includes: The production scheduling task, the workshop status snapshot obtained from the dynamic knowledge graph, the constraint summary extracted from the production scheduling domain-specific language rule base, and historical high-quality production scheduling cases are constructed into structured prompt information; When the data volume of the workshop status snapshot exceeds the preset input capacity, the workshop status snapshot is compressed and truncated according to at least one of the following: work order priority, disturbance impact degree, and resource conflict risk. The compressed and priority-truncated structured prompt information is input into the planning agent to generate candidate scheduling schemes.

[0015] According to one aspect of the above technical solution, after the step of converting the verified candidate scheduling scheme into a scheduling instruction that the manufacturing execution system can accept, and writing it into the manufacturing execution system in a transactional manner, the method further includes: The final production scheduling plan, hard constraint verification results, and production scheduling evaluation results are written into the case library in the dynamic knowledge graph. In subsequent production scheduling tasks, based on the similarity between the current production scheduling task and historical production scheduling cases in the case library, high-quality historical production scheduling cases are retrieved and input into the artificial intelligence production scheduling unit. An audit log for generating production scheduling tasks is generated. The audit log includes at least one of the following: knowledge graph query records, AI production scheduling unit input and output records, candidate production scheduling scheme correction records, hard constraint verification records, structured violation reports, and manufacturing execution system write results.

[0016] On the other hand, this solution also provides an artificial intelligence-based intelligent production scheduling system for factories, including: The data module is used to acquire multi-source production data during the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data into ontology instances based on the pre-built production scheduling domain ontology. The graph module is used to construct a dynamic knowledge graph based on the ontology instances and the semantic relationships between them. The scheduling module is used to retrieve scheduling context information from the dynamic knowledge graph in response to scheduling tasks or disturbance events, and input the scheduling context information into the artificial intelligence scheduling unit, which then generates candidate scheduling schemes. The verification module is used to call the production scheduling domain-specific language rule engine to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph. The first execution module is used to generate a structured violation report when the candidate production scheduling plan fails the hard constraint verification, and to feed the structured violation report back to the artificial intelligence production scheduling unit, so that the artificial intelligence production scheduling unit can make targeted corrections to the candidate production scheduling plan based on the structured violation report; The second execution module is used to convert the candidate scheduling scheme that has passed the hard constraint verification into a scheduling instruction that can be received by the manufacturing execution system, and write it into the manufacturing execution system in a transaction manner when the candidate scheduling scheme passes the hard constraint verification.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: By converting multi-source production data from factories into standardized intermediate data with semantic annotation, and mapping it to ontology instances based on the scheduling domain ontology to construct a dynamic knowledge graph, work orders, processes, equipment, personnel, materials, constraints, and events in the Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), and Warehouse Management System (WMS) can form real-time associations in a unified semantic layer. This provides accurate scheduling context for the AI ​​scheduling unit, reducing scheduling deviations caused by data silos and state lags. Simultaneously, after the AI ​​scheduling unit generates candidate scheduling schemes, this invention further calls a scheduling domain-specific language rule engine to perform hard constraint checks on the candidate scheduling schemes based on the rule base and real-time constraint parameters obtained from the dynamic knowledge graph, such as equipment non-preemption, changeover time, process sequence, material availability, and personnel qualifications. This ensures that schemes that do not meet production physical constraints and business constraints are not directly written into the MES. When verification fails, a structured violation report is sent to the AI ​​scheduling unit, providing feedback on the violation rules, involved parties, violation time windows, and correction guidelines. This allows the AI ​​scheduling unit to make targeted corrections to the conflicting content, rather than regenerating the plan indiscriminately, thereby reducing the risk of AI scheduling illusions and data pollution in the manufacturing execution system. When disturbance events occur, this invention can also deduce the directly and indirectly affected processes and work orders based on a dynamic knowledge graph, determine the scope of local rescheduling, and perform hard constraint verification again before writing, improving the responsiveness to dynamic disturbances such as equipment failures, emergency order insertions, material shortages, personnel absences, and rework order insertions. By writing the verified scheduling plan into the manufacturing execution system in a transactional manner, and storing the final scheduling plan, verification results, evaluation results, and writing records in the case library and audit log, this invention can balance scheduling plan generation efficiency, execution reliability, experience reuse capability, and full-process traceability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent factory scheduling method based on artificial intelligence in the first embodiment of the present invention. Figure 2 This is a structural block diagram of the intelligent factory scheduling system based on artificial intelligence in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0021] Example 1 The first embodiment of the present invention provides an intelligent production scheduling method for factories based on artificial intelligence. This method can be applied to discrete manufacturing scenarios such as automobile manufacturing, machining, and electronic assembly, and is especially suitable for factory production scheduling scenarios with multiple work orders, multiple processes, multiple equipment, multiple personnel, multiple materials, and dynamic disturbance events.

[0022] Please see Figure 1 The figure shows a flowchart of the intelligent factory scheduling method based on artificial intelligence in the first embodiment of the present invention. As shown in the figure, the method includes the following steps: Step S100: Acquire multi-source production data during the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data to ontology instances based on a pre-constructed production scheduling domain ontology. The multi-source production data may include work order execution data and equipment status data from the Manufacturing Execution System, order delivery date data and material planning data from the Enterprise Resource Planning System, and inventory status data from the Warehouse Management System. Work order execution data in a Manufacturing Execution System (MES) can include work order number, product model, planned quantity, process route version, current operation sequence number, work order status, completed quantity, work-in-process location, planned start time, and planned completion time. Equipment status data in an MES can include equipment number, equipment type, production line, current status, overall equipment efficiency, fault level, estimated recovery time, currently occupied work orders, and equipment availability time window. Order delivery date data in an Enterprise Resource Planning (ERP) system can include sales order number, customer priority, agreed delivery date, emergency order flag, and order quantity. Material planning data in an ERP system can include material code, BOM version, planned demand, quantity in transit, and estimated arrival time. Inventory status data in a warehouse management system can include current material inventory, storage location, batch, locked quantity, available quantity, and safety stock threshold.

[0023] In some application scenarios of this embodiment, different data sources are accessed through data adapters. For manufacturing execution systems, RESTful APIs, OPC-UA interfaces, or message queues can be used; for enterprise resource planning systems, database views, SOAP interfaces, or standard business interfaces can be used; for warehouse management systems, inventory change messages, inbound / outbound log interfaces, or inventory query interfaces can be used. After acquiring the raw data, the multi-source production data is deduplicated, filled with null values, and normalized in format. For example, when the status of the same equipment is repeatedly reported within a short period of time, duplicate records can be removed based on the equipment number and status timestamp; when the estimated arrival time of a certain material is empty, the estimated arrival time can be filled in based on the supplier's delivery cycle and the purchase order creation time; when the time format, quantity unit, or status enumeration is inconsistent in different systems, it can be converted into a unified format, such as a standard timestamp, quantity unit, and status enumerations such as AVAILABLE, OCCUPIED, FAULT, and MAINTENANCE. After the above processing, the multi-source production data is converted into standardized intermediate data with semantic annotation, preferably in JSON-LD format. Each piece of standardized intermediate data includes not only field values, but also the ontology concept, data type, and semantic relationship corresponding to that field, so that it can be subsequently identified as belonging to which type of object among work orders, equipment, materials, personnel, processes, or events.

[0024] Based on a pre-constructed production scheduling ontology, standardized intermediate data is mapped to ontology instances. This production scheduling ontology describes the core objects and their relationships within the factory production scheduling domain, and may include resource classes, production object classes, constraint classes, and event classes. Resource classes may include equipment resources, personnel resources, and material resources; production object classes may include work orders, processes, and product models; constraint classes may include timing constraints, resource constraints, and business rule constraints; and event classes may include planned events and disturbance events. The attributes of equipment resources can include equipment number, equipment type, production line, rated capacity, current status, fault level, estimated recovery time, and changeover time matrix reference; the attributes of personnel resources can include personnel number, skill level, certified workstation list, and on-duty status; the attributes of material resources can include material code, material category, current inventory, safety stock threshold, in-transit purchase quantity, and estimated arrival time; the attributes of work orders can include work order number, sales order reference, vehicle model code, planned quantity, customer priority, agreed delivery date, current process progress, and work order status; the attributes of processes can include process sequence number, process name, standard working hours, required equipment type, required personnel skill level, and material consumption list; the attributes of disturbance events can include event type, occurrence time, affected resources, estimated duration, and severity level.

[0025] Step S200: Construct a dynamic knowledge graph based on the ontology instances and the semantic relationships between them.

[0026] When mapping ontology instances, entity extraction can be performed first based on field names, encoding rules, and preset mapping tables. Then, entity alignment determines the correspondence between data records representing the same business object in different data sources. For example, the work order number in the Manufacturing Execution System (MES) can be aligned with the sales order number in the Enterprise Resource Planning (ERP) system through the work order-sales order mapping relationship; the equipment number in the MES can be aligned with the work center number in the ERP system through the equipment-work center mapping table; and the material code in the warehouse management system can be aligned with the BOM material code in the ERP system. After entity alignment is completed, the corresponding data is mapped to ontology instances, and the ontology instances and their semantic relationships are written to the graph database. The semantic relationships can include the relationship between process and equipment demand, process and material consumption, process sequence, process and work order attribution, process and equipment allocation, process and personnel allocation, and the relationship between work orders and disturbance events. Thus, the dynamic knowledge graph can represent the factory production status in the form of nodes and edges, and provide real-time constraint parameters for subsequent production scheduling context retrieval and rule validation.

[0027] After constructing the dynamic knowledge graph, production status change messages in the manufacturing execution system can be continuously monitored. These messages can include at least one of equipment status change messages, work order status change messages, material status change messages, and personnel attendance status change messages. When a production status change message is detected, the affected nodes and edges in the dynamic knowledge graph are determined based on the message, and incremental updates are performed on these nodes and edges. For example, when a welding robot changes from AVAILABLE to FAULT, the current status of the equipment resource node is updated to FAULT, and the corresponding available time window for that equipment is updated simultaneously; when a work order changes from IN_PROGRESS to COMPLETED, the status of the work order node is updated, and the earliest start time for its subsequent processes is updated; when the inventory of a critical material falls below the safety stock threshold, the current inventory and material supply risk status of the material resource node are updated. After incremental updates, the updated dynamic knowledge graph is inferred based on ontology reasoning rules to deduce downstream processes and downstream work orders affected by equipment failure, material shortages, personnel absences, or work order status changes. For example, when a equipment failure affects a process being processed on that equipment, we can deduce whether subsequent processes are affected by the delay in the output of the preceding process by following the sequence of processes, and determine the set of affected work orders by following the relationship between processes and work orders.

[0028] In one alternative implementation, the ontology reasoning rules can be implemented using OWL rules, SWRL rules, graph traversal rules, or preset influence propagation rules. The specific implementation method does not affect the technical essence of the present invention inferring downstream influence links through knowledge graphs.

[0029] In step S300, in response to a scheduling task or disturbance event, scheduling context information is retrieved from the dynamic knowledge graph and input into the artificial intelligence scheduling unit, which then generates a candidate scheduling scheme.

[0030] Specifically, production scheduling tasks can be manually entered, such as "schedule production for the current shift's final assembly line," or they can be full-scale scheduling tasks triggered by the system at regular intervals, or rescheduling tasks triggered by the manufacturing execution system after detecting a disturbance event. The scheduling context information includes at least pending work order information, equipment availability status, material supply status, personnel qualification status, process dependencies, and delivery priority information. Pending work order information can include work order number, vehicle model, planned quantity, current process progress, and customer priority; equipment availability status can include a list of available equipment, equipment status, available time window, equipment rated capacity, and changeover time matrix; material supply status can include material inventory, quantity in transit, estimated arrival time, and material availability status; personnel qualification status can include a list of on-duty personnel, skill levels, and certified workstations; process dependencies can include the sequence of processes within each work order and minimum buffer time; and delivery priority information can include agreed delivery dates, emergency order flags, and customer priority levels.

[0031] When the trigger source is a disturbance event, the event type, occurrence time, affected resources, expected duration, and severity level of the disturbance event are first obtained. Disturbance events can include at least one of the following: equipment failure events, emergency order insertion events, material shortage events, personnel absence events, and rework order insertion events. Based on the semantic relationships of the affected resources in the dynamic knowledge graph, the directly affected processes and work orders are identified, and the indirectly affected downstream processes and work orders are derived along the process sequence, material consumption, and resource occupancy relationships. For example, when the affected resource of an equipment failure event is a painting machine, the process currently occupying that machine can be identified as the directly affected process, and the final assembly process, which depends on the output of that process, can be further identified as the indirectly affected process. When a material shortage event involves an engine assembly, all work orders that require that material and are planned to start within the shortage window can be identified as affected work orders based on the process and material consumption relationship. Subsequently, based on the directly affected processes and work orders, as well as the indirectly affected downstream processes and work orders, the scope of local reordering is determined. The scope of a partial rescheduling can include the set of work orders affected by the disturbance event, the set of alternative equipment, the set of adjustable personnel, and the adjustable time window. For this partial rescheduling scope, a partial rescheduling is triggered instead of a full rescheduling of all work orders, thereby reducing the scale of the rescheduling and minimizing disruption to normal production cycles. Before being written into the Manufacturing Execution System, the partial rescheduling results still need to undergo hard constraint verification using a scheduling domain-specific language rule engine to avoid introducing new equipment conflicts, material conflicts, or personnel qualification conflicts due to partial modifications.

[0032] The AI ​​scheduling unit can be implemented by a single AI model or by multiple agents with different roles working together. In a preferred embodiment, the AI ​​scheduling unit includes a planning agent, an evaluation agent, and an execution agent. The planning agent generates candidate scheduling schemes based on scheduling context information. Candidate scheduling schemes can be represented in JSON format and include work order number, process identifier, planned start time, planned completion time, allocated equipment, allocated personnel, and required material information. When generating candidate scheduling schemes, the planning agent can construct structured prompt information from the scheduling task, a workshop status snapshot retrieved from a dynamic knowledge graph, constraint summaries extracted from a scheduling domain-specific language rule base, and historical high-quality scheduling cases. The workshop status snapshot can include real-time equipment status, work order queue, material inventory, personnel skill certification, and process dependencies; the constraint summary can include natural language summaries of rules such as equipment non-preemption, changeover time, process sequence, material availability, personnel qualifications, and shift boundaries; historical high-quality scheduling cases can be retrieved from the case base in the dynamic knowledge graph and used as reference examples for candidate scheme generation. When the amount of data in the shop floor status snapshot exceeds the preset input capacity, the planning agent can compress and prioritize the shop floor status snapshot based on at least one of the following: work order priority, disturbance impact level, and resource conflict risk. For example, it can prioritize retaining urgent work orders, high customer priority work orders, work orders affected by disturbance events, work orders occupying bottleneck equipment, and work orders with material shortage risks, while deleting or summarizing low-priority historical states that have little impact on the current production schedule. Subsequently, the compressed and prioritized structured prompt information is input into the planning agent to generate candidate production scheduling schemes.

[0033] Preferably, after receiving the candidate production scheduling schemes generated by the planning agent, the evaluation agent evaluates the candidate production scheduling schemes based on at least two of the following indicators: delivery date fulfillment rate, equipment utilization rate, changeover frequency, and material availability rate, to obtain a production scheduling evaluation result. The delivery date fulfillment rate can be determined based on the ratio between the number of work orders completed on schedule and the total number of work orders participating in the scheduling; the equipment utilization rate can be determined based on the ratio between the effective processing time and available time of the equipment within the scheduling window; the changeover frequency can be statistically analyzed based on whether adjacent processes on the same equipment involve a change in corresponding vehicle models; and the material availability rate can be determined based on whether the materials required before the work order is scheduled to start meet the inventory and delivery time requirements. The evaluation agent generates improvement suggestions based on the production scheduling evaluation results. For example, for cases with low delivery date fulfillment rates, it suggests advancing high-priority work orders; for cases with low equipment utilization rates, it suggests adjusting process allocation to reduce equipment idle time; for cases with high changeover frequencies, it suggests merging the processing order of work orders for the same vehicle model; and for cases with low material availability rates, it suggests postponing work orders with incomplete material availability or replacing them with work orders with complete material availability. Based on the improvement suggestions from the evaluation agent, the planning agent makes partial or global adjustments to the candidate production scheduling schemes. Partial adjustments may only change the start time, equipment allocation, or personnel allocation for some work orders; global adjustments may regenerate the overall order and resource allocation relationships of the work orders participating in the production scheduling.

[0034] Step S400: Call the production scheduling domain-specific language rule engine to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph.

[0035] The domain-specific language rule engine performs deterministic verification on candidate scheduling schemes based on rules in the domain-specific language rule base and real-time constraint parameters obtained from a dynamic knowledge graph. The domain-specific language rule base may include at least one of the following statements: CONSTRAINT, RULE, RESOURCE, WINDOW, and PRIORITY. The CONSTRAINT statement defines hard constraint assertions that the scheduling scheme must meet, such as the same equipment not being occupied by two processes simultaneously within the same time period; the RULE statement defines soft business rules recommended for execution when trigger conditions are met, such as prioritizing high-priority customer orders; the RESOURCE statement declares the resources involved in scheduling and their initial states, such as equipment lists, personnel lists, and material lists; the WINDOW statement defines the effective time range of the scheduling plan, such as shift start time, shift end time, and allowed overtime; and the PRIORITY statement defines weight coefficients for multi-objective optimization, such as the weights of on-time delivery rate, equipment utilization rate, changeover frequency, and material availability rate. In one alternative implementation, the production scheduling domain-specific language rule base can be stored in the form of text rules, YAML files, JSON rule files, or database rule tables, as long as it can be parsed by the rule engine into an internal constraint representation and subjected to deterministic validation.

[0036] Specifically, the system loads the currently valid rule set from a language-specific rule base for production scheduling, parses the rule set into an internal constraint representation, and reads the real-time constraint parameters corresponding to the internal constraint representation from a dynamic knowledge graph. These real-time constraint parameters may include the current state of the equipment, the available time window for the equipment, the equipment changeover time matrix, the minimum buffer time for the process, the agreed delivery date of the work order, the material inventory, the estimated arrival time of the materials, the list of certified workstations, the start and end times of shifts, and the buffer capacity.

[0037] Based on internal constraint representation and real-time constraint parameters, at least one of the following constraints is applied to the candidate production scheduling scheme: non-preemptive equipment constraint verification, changeover time constraint verification, process sequence constraint verification, delivery date constraint verification, material completeness constraint verification, personnel qualification constraint verification, shift boundary constraint verification, and buffer capacity constraint verification.

[0038] Equipment non-preemption constraint verification is used to determine whether the planned processing times of two different processes on the same equipment overlap; changeover time constraint verification is used to determine whether, when the same equipment continuously processes different vehicle models, the planned start time of the subsequent process is later than the sum of the planned completion time and changeover preparation time of the previous process; process sequence constraint verification is used to determine whether the planned start time of the subsequent process is later than the sum of the planned completion time and minimum buffer time of the preceding process; delivery date constraint verification is used to determine whether high-priority work orders or work orders with hard delivery dates can be completed before the agreed delivery date; material availability constraint verification is used to determine whether the available inventory and expected arrival time of the materials required by the work order before the planned start time meet the start requirements; personnel qualification constraint verification is used to determine whether the assigned personnel have the skill certification required for the corresponding workstation or process; shift boundary constraint verification is used to determine whether the planned start time and planned completion time are within the effective scheduling window; buffer capacity constraint verification is used to determine whether the quantity of work-in-process between processes exceeds the corresponding buffer capacity. The above verifications are executed one by one based on rule expressions and real-time constraint parameters, and the verification results are pass, failure, or warning, thus forming a deterministic verification of the candidate scheduling scheme.

[0039] Step S500: When the candidate production scheduling scheme fails the hard constraint verification, a structured violation report is generated and fed back to the artificial intelligence production scheduling unit, so that the artificial intelligence production scheduling unit can make targeted corrections to the candidate production scheduling scheme based on the structured violation report.

[0040] When a candidate production schedule fails the hard constraint verification, a structured violation report is generated and fed back to the AI ​​production scheduling unit, enabling the AI ​​production scheduling unit to make targeted corrections to the candidate production schedule based on the structured violation report. The structured violation report can include violation rule identifiers, involved work orders, involved equipment, involved personnel, involved materials, violation time window, violation reason, and correction guidelines. For example, when the equipment non-preemption constraint verification fails, the structured violation report can record "Equipment E01 was simultaneously assigned to operation OP10 of work order WO001 and operation OP20 of work order WO002 from 10:00 to 10:30," and provide correction guidelines such as "Adjust the start time of one of the operations or reassign it to an alternative equipment"; when the personnel qualification constraint verification fails, the structured violation report can record "Personnel P01 does not have high-voltage cable connection workstation certification," and provide correction guidelines such as "Replace with personnel with the corresponding certification or adjust the operation to the certified personnel's shift time." After receiving a structured violation report, the AI-powered scheduling unit does not perform a general retry without a clear direction. Instead, it targets and corrects conflicting fields in candidate scheduling plans based on violation rule identifiers and correction guidelines. For example, it may adjust only the time windows, equipment allocation, or personnel allocation for conflicting processes. The corrected candidate scheduling plan is then submitted to a domain-specific language rule engine for hard constraint verification. In one optional implementation, a maximum number of corrections can be set after a verification failure. If the number of corrections exceeds a preset limit and the hard constraint verification still fails, the system retains the original scheduling plan and sends a prompt for manual intervention to the scheduling manager.

[0041] Step S600: When the candidate scheduling scheme passes the hard constraint verification, the candidate scheduling scheme that passes the verification is converted into a scheduling instruction that the manufacturing execution system can accept, and written into the manufacturing execution system in a transaction manner.

[0042] Specifically, the execution agent converts the verified candidate scheduling schemes into scheduling instructions that the Manufacturing Execution System (MES) can accept, and writes them into the MES in a transactional manner. The scheduling instructions may include work order number, process number, allocated equipment number, planned start time stamp, planned completion time stamp, operator employee number, required material list, and scheduling batch identifier. Transactional writing means submitting multiple scheduling instructions corresponding to the same candidate scheduling scheme as a single transaction to the MES. The transaction is committed when all scheduling instructions are successfully written, and rolled back when any scheduling instruction fails to be written, thus avoiding inconsistencies caused by some work orders being updated while others are not. After successful writing, the MES can update the work order status, process planned time, and resource occupancy relationships according to the scheduling instructions, and publish the scheduling results to the field execution terminal or scheduling management interface.

[0043] After the validated candidate scheduling schemes are written into the Manufacturing Execution System (MES), the system can also write the final scheduling scheme, hard constraint validation results, and scheduling evaluation results into a case library in a dynamic knowledge graph. Each historical scheduling case in the case library can include a scheduling task description, a snapshot of the workshop status before scheduling, the final scheduling scheme, scheduling evaluation indicators, hard constraint validation results, whether there are disturbance events, and the actual execution results. When a subsequent scheduling task occurs, the system can retrieve historical high-quality scheduling cases based on the similarity between the current scheduling task and historical scheduling cases in the case library and input them into the AI ​​scheduling unit. The similarity can be determined based on at least one of the following: vehicle model combination, work order quantity, bottleneck equipment, material supply status, personnel shift status, and disturbance event type. Thus, the AI ​​scheduling unit can generate new candidate scheduling schemes by referring to historical high-quality scheduling cases, improving the efficiency and quality of scheduling scheme generation. Simultaneously, the system generates an audit log for the scheduling task. The audit log can include at least one of the following: knowledge graph query records, AI scheduling unit input / output records, candidate scheduling scheme correction records, hard constraint validation records, structured violation reports, and MES write results. Audit logs allow for the tracing of data sources, scheme generation processes, violation correction processes, rule verification processes, and final write results for each production scheduling task, thus meeting the requirements of interpretability and traceability in industrial production processes.

[0044] In summary, the AI-based intelligent factory scheduling method described in the above embodiments of the present invention converts multi-source production data of the factory into standardized intermediate data with semantic annotation, and maps it to ontology instances based on the scheduling domain ontology to construct a dynamic knowledge graph. This enables work orders, processes, equipment, personnel, materials, constraints, and events in the Manufacturing Execution System (MES), Enterprise Resource Planning (ERP), and Warehouse Management System (WMS) to form real-time associations in a unified semantic layer. This provides the AI ​​scheduling unit with an accurate scheduling context, reducing scheduling deviations caused by data silos and state lag. Furthermore, after the AI ​​scheduling unit generates candidate scheduling schemes, the present invention further calls a scheduling domain-specific language rule engine. Based on the rule base and real-time constraint parameters obtained from the dynamic knowledge graph, the candidate scheduling schemes are checked for hard constraints such as equipment non-preemption, changeover time, process sequence, material availability, and personnel qualifications. This ensures that schemes that do not meet production physical constraints and business constraints are not directly written into the manufacturing system. The system executes the production process. When verification fails, a structured violation report is sent to the AI ​​scheduling unit, providing information on the violation rules, affected entities, violation time windows, and correction guidelines. This allows the AI ​​scheduling unit to make targeted corrections to the conflicting content, rather than regenerating the plan without direction, thereby reducing the risk of AI scheduling illusions and data contamination in the manufacturing execution system. When disturbance events occur, the invention can also deduce the directly and indirectly affected processes and work orders based on a dynamic knowledge graph, determine the scope of local rescheduling, and perform hard constraint verification again before writing, improving the responsiveness to dynamic disturbances such as equipment failures, emergency order insertions, material shortages, staff absences, and rework order insertions. By writing the verified scheduling plan into the manufacturing execution system in a transactional manner, and storing the final scheduling plan, verification results, evaluation results, and writing records in a case library and audit log, the invention can balance scheduling plan generation efficiency, execution reliability, experience reuse capability, and full-process traceability.

[0045] Example 2 like Figure 2 As shown, the second embodiment of the present invention provides an intelligent factory scheduling system based on artificial intelligence, including: The data module 100 is used to acquire multi-source production data in the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data into an ontology instance based on a pre-built production scheduling domain ontology. The graph module 200 is used to construct a dynamic knowledge graph based on the ontology instances and the semantic relationships between them. The scheduling module 300 is used to retrieve scheduling context information from the dynamic knowledge graph in response to scheduling tasks or disturbance events, and input the scheduling context information into the artificial intelligence scheduling unit, which then generates candidate scheduling schemes. The verification module 400 is used to call the production scheduling domain-specific language rule engine to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph. The first execution module 500 is used to generate a structured violation report when the candidate production scheduling scheme fails the hard constraint verification, and to feed the structured violation report back to the artificial intelligence production scheduling unit, so that the artificial intelligence production scheduling unit can make targeted corrections to the candidate production scheduling scheme based on the structured violation report; The second execution module 600 is used to convert the candidate scheduling scheme that has passed the hard constraint verification into a scheduling instruction that can be received by the manufacturing execution system, and write it into the manufacturing execution system in a transaction manner when the candidate scheduling scheme passes the hard constraint verification.

[0046] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A factory intelligent scheduling method based on artificial intelligence, characterized in that, Includes the following steps: Acquire multi-source production data during the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data into ontology instances based on a pre-constructed production scheduling domain ontology. A dynamic knowledge graph is constructed based on the ontology instances and the semantic relationships between them. In response to scheduling tasks or disturbance events, scheduling context information is retrieved from the dynamic knowledge graph and input into the artificial intelligence scheduling unit, which then generates candidate scheduling schemes. The production scheduling domain-specific language rule engine is invoked to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph. When the candidate production scheduling plan fails the hard constraint verification, a structured violation report is generated and fed back to the AI ​​production scheduling unit, so that the AI ​​production scheduling unit can make targeted corrections to the candidate production scheduling plan based on the structured violation report; When the candidate scheduling scheme passes the hard constraint verification, the verified candidate scheduling scheme is converted into a scheduling instruction that the manufacturing execution system can accept, and written into the manufacturing execution system in a transactional manner.

2. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, The steps of converting the multi-source production data into standardized intermediate data with semantic annotation, and mapping the standardized intermediate data to ontology instances based on a pre-constructed production scheduling domain ontology, specifically include: Data adapters are used to connect to the Manufacturing Execution System, Enterprise Resource Planning System, and Warehouse Management System, respectively. The acquired multi-source production data is deduplicated, filled with null values, and normalized in format. The processed multi-source production data is converted into standardized intermediate data in JSON-LD format; Based on the production scheduling domain ontology, entity extraction and entity alignment are performed on the standardized intermediate data to determine the correspondence between data records representing the same business object in different data sources. The extracted and aligned entity data is mapped to ontology instances, and the ontology instances and their semantic relationships are written into the graph database.

3. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, Following the step of constructing a dynamic knowledge graph, the method further includes: Monitor production status change messages in the manufacturing execution system, wherein the production status change messages include at least one of equipment status change messages, work order status change messages, material status change messages, and personnel attendance status change messages; When a production status change message is detected, the affected nodes and edges in the dynamic knowledge graph are determined based on the production status change message. Incremental updates are performed on the affected nodes and edges; Based on ontology reasoning rules, reasoning is performed on the updated dynamic knowledge graph to deduce downstream processes and downstream work orders affected by equipment failure, material shortage, staff absence, or changes in work order status.

4. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, The steps of retrieving scheduling context information from the dynamic knowledge graph in response to a disturbance event specifically include: Obtain the event type, occurrence time, affected resources, expected duration, and severity level of the disturbance event, wherein the disturbance event includes at least one of the following: equipment failure event, emergency order insertion event, material shortage event, personnel absence event, and rework order insertion event; Based on the semantic relationships of the affected resources in the dynamic knowledge graph, the directly affected processes and work orders are determined; Deduce the indirectly affected downstream processes and downstream work orders by following the sequence of processes, material consumption relationships, and resource utilization relationships; The scope of the partial rearrangement is determined based on the directly affected processes and work orders, as well as the indirectly affected downstream processes and work orders. Local rearrangement is triggered for the local rearrangement range, and the scheduling domain-specific language rule engine is called to perform hard constraint verification before the local rearrangement result is written into the manufacturing execution system.

5. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, The production scheduling domain-specific language rule base includes at least one of the following statement structures: The CONSTRAINT statement is used to define hard constraint assertions that the production scheduling plan must meet; The RULE statement is used to define soft business rules that are recommended to be executed when triggering conditions are met; The RESOURCE statement is used to declare the resources participating in production scheduling and their initial state; The WINDOW statement is used to define the effective time range of a production schedule. The PRIORITY statement is used to define the weighting coefficients for multi-objective optimization.

6. The intelligent factory scheduling method based on artificial intelligence according to claim 5, characterized in that, Calling a domain-specific language rule engine to perform hard constraint verification on the candidate production scheduling schemes includes: Load the currently valid rule set from the language-specific rule base for the production scheduling domain; The currently valid rule set is parsed into an internal constraint representation; Read the real-time constraint parameters corresponding to the internal constraint representation from the dynamic knowledge graph; Based on the internal constraint representation and the real-time constraint parameters, at least one of the following constraints is performed on the candidate production scheduling scheme: equipment non-preemption constraint verification, changeover time constraint verification, process sequence constraint verification, delivery date constraint verification, material completeness constraint verification, personnel qualification constraint verification, shift boundary constraint verification, and buffer capacity constraint verification. When a constraint violation occurs, a structured violation report is generated. The structured violation report includes a violation rule identifier, the work order involved, the equipment involved, the violation time window, and correction guidelines.

7. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, The AI-powered scheduling unit includes a planning agent, an evaluation agent, and an execution agent; The planning agent is used to generate candidate production scheduling schemes based on the production scheduling context information; The evaluation agent is used to evaluate the candidate production scheduling scheme based on at least two of the following indicators: delivery date fulfillment rate, equipment utilization rate, changeover frequency, and material availability rate, to obtain the production scheduling evaluation result, and to generate improvement suggestions based on the production scheduling evaluation result. The planning agent is also used to make local or global corrections to the candidate production scheduling schemes based on the improvement suggestions. The execution agent is used to call the domain-specific language rule engine of the scheduling field to perform hard constraint verification on the modified candidate scheduling scheme, and after the verification is passed, convert the modified candidate scheduling scheme into a scheduling instruction that the manufacturing execution system can accept.

8. The intelligent factory scheduling method based on artificial intelligence according to claim 7, characterized in that, The specific steps of inputting the production scheduling context information into the AI ​​production scheduling unit, and having the AI ​​production scheduling unit generate candidate production scheduling schemes, include: The production scheduling task, the workshop status snapshot obtained from the dynamic knowledge graph, the constraint summary extracted from the production scheduling domain-specific language rule base, and historical high-quality production scheduling cases are constructed into structured prompt information; When the data volume of the workshop status snapshot exceeds the preset input capacity, the workshop status snapshot is compressed and truncated according to at least one of the following: work order priority, disturbance impact degree, and resource conflict risk. The compressed and priority-truncated structured prompt information is input into the planning agent to generate candidate scheduling schemes.

9. The intelligent factory scheduling method based on artificial intelligence according to claim 1, characterized in that, After converting the validated candidate scheduling schemes into scheduling instructions that the Manufacturing Execution System (MES) can accept, and writing them into the MES in a transactional manner, the method further includes: The final production scheduling plan, hard constraint verification results, and production scheduling evaluation results are written into the case library in the dynamic knowledge graph. In subsequent production scheduling tasks, based on the similarity between the current production scheduling task and historical production scheduling cases in the case library, high-quality historical production scheduling cases are retrieved and input into the artificial intelligence production scheduling unit. An audit log for generating production scheduling tasks is generated. The audit log includes at least one of the following: knowledge graph query records, AI production scheduling unit input and output records, candidate production scheduling scheme correction records, hard constraint verification records, structured violation reports, and manufacturing execution system write results.

10. A system for implementing the artificial intelligence-based intelligent factory scheduling method according to any one of claims 1-9, characterized in that, include: The data module is used to acquire multi-source production data during the factory production process, convert the multi-source production data into standardized intermediate data with semantic annotation, and map the standardized intermediate data into ontology instances based on the pre-built production scheduling domain ontology. The graph module is used to construct a dynamic knowledge graph based on the ontology instances and the semantic relationships between them. The scheduling module is used to retrieve scheduling context information from the dynamic knowledge graph in response to scheduling tasks or disturbance events, and input the scheduling context information into the artificial intelligence scheduling unit, which then generates candidate scheduling schemes. The verification module is used to call the production scheduling domain-specific language rule engine to perform hard constraint verification on the candidate production scheduling scheme. The production scheduling domain-specific language rule engine performs deterministic verification on the candidate production scheduling scheme based on the rules in the production scheduling domain-specific language rule library and the real-time constraint parameters obtained from the dynamic knowledge graph. The first execution module is used to generate a structured violation report when the candidate production scheduling plan fails the hard constraint verification, and to feed the structured violation report back to the artificial intelligence production scheduling unit, so that the artificial intelligence production scheduling unit can make targeted corrections to the candidate production scheduling plan based on the structured violation report; The second execution module is used to convert the candidate scheduling scheme that has passed the hard constraint verification into a scheduling instruction that can be received by the manufacturing execution system, and write it into the manufacturing execution system in a transaction manner when the candidate scheduling scheme passes the hard constraint verification.

Images