Artificial Intelligence-Based Dynamic Scheduling Method and System for Industrial Production Lines

Abstract

This invention relates to the field of industrial scheduling technology, and discloses a dynamic scheduling method and system for industrial production lines based on artificial intelligence. The method includes: S1, State Awareness: converting real-time event streams into event semantic quadruples and updating the system snapshot to identify disturbance events; S2, Constraint Modeling: constructing a constraint fingerprint library containing process routes and resource information, storing hard and soft constraints in the production process; S3, Parameter Calibration: calibrating operation duration intervals online based on historical logs, and generating a scheduling permission set by combining risk budget and snapshots; S4, Impact Assessment: extracting rule-propagation duals from historical logs through symbolic induction; S5, Patch Synthesis; S6, Filtering and Execution. By constructing a constraint fingerprint library, hard and soft constraints are stored uniformly, relevant constraints are quickly retrieved, and differential constraints are generated, avoiding repeated processing of global constraints, thereby reducing scheduling computational complexity and improving response speed.

Description

Technical Field

[0001] This invention relates to the field of industrial scheduling technology, specifically to a method and system for dynamic scheduling of industrial production lines based on artificial intelligence. Background Technology

[0002] An industrial production line is a production system composed of multiple pieces of equipment, process units, and resources organized according to a predetermined technological route. It realizes the continuous or discrete transformation of products from input to output by sequentially processing, assembling, or handling raw materials. Due to the strict sequential relationship between different processes in an industrial production line, and the limited time and capacity of different equipment and personnel resources, the operation of the production line requires overall planning through scheduling to arrange tasks on the time axis, allocate resources, and execute processes in a coordinated manner. This ensures that all production tasks are completed efficiently and orderly while meeting technological requirements and resource constraints, thereby achieving a comprehensive balance between capacity utilization, delivery cycle, and operational stability.

[0003] Existing industrial production line scheduling typically relies on static planning models or global reordering methods based on rules and optimization algorithms. In scheduling models, constraints such as process sequence, resource mutual exclusion, resource availability time window, and resource capacity are often treated as a unified set of constraints for calculation. However, in practice, when disturbances such as equipment failure, execution delay, or task changes occur during production, the scheduling system usually needs to reload and process all relevant constraints to generate a new scheduling scheme. This leads to a rapid increase in computational load and a prolonged response time when disturbances are frequent or the production scale is large, thereby affecting the scheduling efficiency and real-time response capability of industrial production lines in dynamic operating environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based dynamic scheduling method and system for industrial production lines. This solves the problem that when disturbances occur during the production process, the computational load of new scheduling schemes is large, which affects the scheduling efficiency and real-time response capability of industrial production lines in dynamic operating environments.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a dynamic scheduling method for industrial production lines based on artificial intelligence, comprising the following steps:

[0006] S1, State Awareness: Converts real-time event streams into event semantic quadruples and updates system snapshots to identify disturbance events;

[0007] S2. Constraint Modeling: Construct a constraint fingerprint library containing process route and resource information to store hard and soft constraints in the production process;

[0008] S3, Parameter Calibration: Based on historical logs, online calibration operation duration range is performed, and a scheduling permit set is generated by combining risk budget and snapshot;

[0009] S4. Impact Assessment: Extract rule-propagation duality from historical logs through symbolic induction, and perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain;

[0010] S5. Patch Synthesis: Generate differential constraints and distinguish variables within the influence domain, and generate candidate patch graphs composed of atomic changes through the synthesizer;

[0011] S6. Execution Filtering: Perform incremental feasibility verification on the candidate patch graph, select the target patch, generate the target scheduling plan, issue it for execution, and maintain real-time closed-loop iteration.

[0012] Preferably, in step S1, the step of identifying the disturbance event is as follows:

[0013] The real-time event stream is parsed and standardized. Based on the resource identifier, operation identifier, event type identifier and timestamp of the event, an event semantic quadruple is constructed and written to the system snapshot.

[0014] Extract the baseline state corresponding to the event semantic quadruple from the system snapshot, calculate the perturbation feature vector based on the baseline state and the event semantic quadruple, match the perturbation feature vector with the preset perturbation judgment condition, and the threshold condition is whether the deviation exceeds the tolerance range.

[0015] When the disturbance determination condition is met, the corresponding event semantic quadruple is marked as a disturbance event, and a disturbance event identifier is generated;

[0016] The perturbation events marked within the preset time window are deduplicated and aggregated to obtain a set of perturbation events.

[0017] Preferably, in step S2, constructing a constraint fingerprint database containing process route and resource information includes the following steps:

[0018] Acquire process route data and resource information data, and parse the process route data into a process route diagram;

[0019] Resource information data is parsed into a resource capability matrix and resource calendar constraints, and constraint expressions for process-related hard constraints are generated based on the process roadmap.

[0020] Resource-class hard constraints are generated based on the resource capability matrix and resource calendar constraints, and soft constraint expressions are generated based on the production strategy and optimization objectives.

[0021] The constraint fingerprints are obtained by normalizing and encoding hard and soft constraints, and then associated and stored with process route diagram nodes and resource identifiers to form a constraint fingerprint library.

[0022] Preferably, in step S2, the hard constraints include: process sequence, resource mutual exclusion constraints, resource availability time window constraints, and resource capacity constraints.

[0023] The soft constraints include: delivery deviation penalty constraints, scheduling stability penalty constraints, and type change / switching penalty constraints.

[0024] Preferably, in step S3, the step of generating the scheduling permit set is as follows:

[0025] Online calibration is performed on each type of operation based on historical logs to obtain the corresponding operation duration range;

[0026] Extract the current set of tasks to be scheduled, resource status and time window information based on system snapshots, and determine the risk budget status related to disturbance events;

[0027] The permission threshold is determined based on the risk budget status. Under the constraint of the permission threshold, candidate scheduling actions for the set of tasks to be scheduled are enumerated. For each candidate scheduling action, the risk increment is calculated based on the operation duration interval and system snapshot. The candidate scheduling action is written into the scheduling permission set when the risk increment does not exceed the permission threshold.

[0028] Preferably, the step of performing closure iteration in step S4 is as follows:

[0029] The seed set of the influence domain is determined based on the set of disturbance events, the influence domain is initialized as the seed set of the influence domain, and the influence domain is used as the current set of nodes;

[0030] The rule-propagation dual is extracted from the historical log by symbolic induction, and its constraint expression is judged to meet the triggering condition under the constraint of the scheduling permission set.

[0031] For rules that meet the triggering conditions, invoke the propagation operator to generate a candidate set of extended nodes from the current set of nodes;

[0032] The new influence domain is obtained by merging the union of the candidate expanded node sets with the influence domain and updating it.

[0033] If the new influence domain is the same as the previous influence domain, the iteration terminates and the influence domain is output; otherwise, the node set is expanded as the new current node set and the iteration continues.

[0034] Preferably, in step S5, the steps of generating difference constraints and distinguishing variables are as follows:

[0035] Obtain the affected domain and extract the task identifier, resource identifier, and planned start time / planned completion time related to the affected domain based on the system snapshot;

[0036] Based on the constraint fingerprint database, the constraint expressions of hard and soft constraints associated with the influence domain are retrieved to obtain the set of influence domain constraints;

[0037] Based on the original scheduling plan corresponding to the system snapshot, mark the newly added, invalidated or replaced constraint expressions in the influence domain and generate corresponding differential constraints;

[0038] Variables outside the influence domain are categorized into the frozen variable set, and variables within the influence domain are categorized into the variable set. Difference constraints are then stored in association with the frozen variable set and the variable set.

[0039] Preferably, in step S5, the steps for constructing the candidate patch image are as follows:

[0040] Obtain the influence domain, differential constraints, set of frozen variables, set of variable variables, and set of scheduling permissions;

[0041] Under the constraints of the scheduling permission set, enumerate the candidate set of atomic changes that act on the set of variable variables, and perform local consistency checks based on the difference constraints and the set of frozen variables to filter out atomic changes that violate hard constraints, thus obtaining the set of available atomic changes.

[0042] A set of patch graph nodes is constructed based on the set of available atomic changes, and the dependencies between nodes are determined based on differential constraints and written into the dependency edge set;

[0043] For nodes with predecessor dependencies, complete the predecessor nodes, and mark the node pairs with mutual exclusion dependencies to obtain the candidate patch graph.

[0044] Preferably, in step S6, the feasibility verification step is as follows:

[0045] Obtain the candidate patch graph, differential constraints, set of frozen variables, set of variable variables, and the original scheduling plan and operation duration range from the system snapshot;

[0046] The atomic changes in the candidate patch graph are sorted topologically by dependency relationship to obtain the atomic change sequence;

[0047] Based on the original scheduling plan, atomic changes are applied one by one in the atomic change sequence only within the set of variable variables to form incrementally updated candidate scheduling states, while keeping the values ​​of the corresponding variables in the set of frozen variables unchanged;

[0048] After each atomic change is applied, an incremental consistency check is performed on the differential constraints based on the operation time interval. If any check fails, the candidate patch graph is determined to be infeasible and the verification is terminated. If all checks pass, the candidate patch graph is determined to be feasible and output as a feasible patch.

[0049] Preferably, an artificial intelligence-based dynamic scheduling system for industrial production lines includes the following modules:

[0050] The event semanticization and state snapshot module is used to acquire real-time event streams and convert them into event semantic quadruples to update system snapshots, identify disturbance events and obtain a set of disturbance events;

[0051] The constraint fingerprint library module is used to build a constraint fingerprint library containing process route and resource information and to store hard and soft constraints.

[0052] The parameter calibration and license generation module is used to calibrate the operation duration range online based on historical logs, and generate a set of scheduling licenses by combining risk budget and system snapshots;

[0053] The impact assessment module is used to extract rule-propagation duals from historical logs through symbolic induction, and to perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain.

[0054] The patch generation and execution module is used to generate differential constraints within the influence domain and distinguish between the set of frozen variables and the set of variable variables. Under the constraints of differential constraints, the set of frozen variables, the set of variable variables, and the set of scheduling permissions, it generates a candidate patch graph, performs incremental feasibility verification on the candidate patch graph, selects the target patch, generates the target scheduling plan, issues it for execution, and maintains real-time closed-loop iteration.

[0055] This invention provides a method and system for dynamic scheduling of industrial production lines based on artificial intelligence. It has the following beneficial effects:

[0056] 1. This invention constructs a constraint fingerprint library, which uniformly stores hard constraints such as process sequence, resource mutual exclusion, resource availability time window and resource capacity, as well as soft constraints related to delivery time and stability, in a standardized constraint expression form. After a disturbance occurs, relevant constraints can be quickly retrieved by process node and resource identifier and differential constraints can be generated, avoiding repeated processing of global constraints, thereby reducing scheduling computation complexity and improving response speed.

[0057] 2. This invention transforms real-time event streams into event semantic quadruples and combines them with system snapshots for disturbance identification, thereby transforming the production line status from unstructured events into calculable and traceable semantic states. This enables the accurate identification of disturbance events that truly affect scheduling feasibility based on the planning baseline and real-time execution deviation, avoiding false triggers caused by instantaneous fluctuations or repeated reporting, and improving the accuracy and stability of disturbance identification.

[0058] 3. This invention introduces a scheduling permission set mechanism based on operation duration intervals and risk budgets, which pre-constrains the allowable adjustment range during the scheduling action generation stage. This ensures that closure propagation and patch synthesis in the impact assessment are carried out only within the risk-controllable action space, effectively preventing excessive plan deviations from being introduced during scheduling repair and enhancing the safety and controllability of the scheduling process.

[0059] 4. This invention generates differential constraints within the influence domain and distinguishes between the set of frozen variables and the set of variable variables. Combined with a candidate patch graph composed of atomic changes and an incremental feasibility verification mechanism, it achieves local and incremental repair of production disturbances. While ensuring the feasibility of hard constraints on processes and resources, it maximizes the stability of the plan in unaffected areas, thereby reducing the plan oscillations caused by global reordering and improving the execution reliability of industrial production lines in real-time disturbance scenarios. Attached Figure Description

[0060] Figure 1 This is a flowchart of the method of the present invention;

[0061] Figure 2 This is a system framework diagram of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see the appendix Figure 1 This invention provides an artificial intelligence-based dynamic scheduling method for industrial production lines, comprising the following steps:

[0064] S1, State Awareness: Converts real-time event streams into event semantic quadruples and updates system snapshots to identify disturbance events;

[0065] S2. Constraint Modeling: Construct a constraint fingerprint library containing process route and resource information to store hard and soft constraints in the production process;

[0066] S3, Parameter Calibration: Based on historical logs, online calibration operation duration range is performed, and a scheduling permit set is generated by combining risk budget and snapshot;

[0067] S4. Impact Assessment: Extract rule-propagation duality from historical logs through symbolic induction, and perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain;

[0068] S5. Patch Synthesis: Generate differential constraints and distinguish variables within the influence domain, and generate candidate patch graphs composed of atomic changes through the synthesizer;

[0069] S6. Execution Filtering: Perform incremental feasibility verification on the candidate patch graph, select the target patch, generate the target scheduling plan, issue it for execution, and maintain real-time closed-loop iteration.

[0070] Specifically, in step S1, on an industrial production line, real-time event streams from equipment controllers and manufacturing execution are continuously connected. Each event is mapped to an event semantic quadruple containing resource identifier, operation identifier, event type identifier, and timestamp, and written to a system snapshot. The snapshot synchronously reflects changes in resource availability, in-process task status, and planned start and end times. When an event causes a deviation in planned time or resource status that exceeds the tolerance range, it is marked as a disturbance event and aggregated into a disturbance event set.

[0071] In step S2, a constraint fingerprint database is constructed based on the process route and resource information. Hard constraints such as process sequence, resource mutual exclusion, resource availability time window and resource capacity, as well as soft constraints such as delivery deviation penalty, scheduling stability penalty and type changeover penalty, are stored in a computable constraint expression form and associated with the corresponding process node and resource identifier, so that they can be quickly retrieved according to the scope of influence in subsequent calculations.

[0072] In step S3, the system further calibrates the operation duration range of various operations online from historical logs and generates a scheduling permission set by combining risk budget and system snapshot, so that the allowed scheduling actions and their scope of action meet the risk budget constraints. Subsequently, the system extracts rules from historical logs through symbolic induction—propagation duals—and performs closure iterative propagation on the set of disturbance events under the constraints of the scheduling permission set to obtain the influence domain. The propagation process takes the current set of nodes as input and outputs a candidate extended node set by the propagation operator, which is repeatedly incorporated into the influence domain until convergence, thereby limiting the disturbance impact to the task and resource range that can be actually propagated by process and resource constraints.

[0073] In step S4, within the system of influence domain, based on the original scheduling plan in the system snapshot, new failure replacement markers are added to the constraint expressions related to the influence domain in the constraint fingerprint library, and differential constraints are generated. At the same time, variables outside the influence domain are included in the frozen variable set, and variables within the influence domain are included in the variable set to ensure that the repair only applies to the locally adjustable part.

[0074] In step S5, the synthesizer then generates a candidate patch graph consisting of atomic changes under the constraints of difference constraints, frozen variable set, variable variable set and scheduling permission set. The nodes of the patch graph correspond to atomic changes and the dependent edges depict the preconditions and mutual exclusion relationships between atomic changes, so that the patches can be applied in an orderly manner and incompatible combinations are avoided.

[0075] In step S6, the system applies atomic changes to each candidate patch graph in the order determined by the dependency relationship within the set of variable variables, forming an incrementally updated candidate scheduling state while keeping the values ​​of the corresponding variables in the frozen variable set unchanged. After each application, an incremental consistency check is performed on the differential constraints based on the operation duration interval to determine whether hard constraints such as process sequence, resource mutual exclusion, resource availability time window, and resource capacity are met. If the check fails, the verification of the candidate patch graph is terminated. If all checks pass, a feasible patch is output, and a target scheduling plan is generated and issued for execution. New events returned by the execution layer continue to update the system snapshot in the form of event semantic quadruples and trigger a new round of disturbance identification and local patch iteration. This allows for rapid local repair of disturbances without compromising the feasibility of hard constraints and reduces unnecessary changes to the overall production line plan, thereby improving the scheduling's responsiveness to real-time disturbances and execution stability.

[0076] In step S1, the steps for identifying disturbance events are as follows:

[0077] The real-time event stream is parsed and standardized. Based on the resource identifier, operation identifier, event type identifier and timestamp of the event, an event semantic quadruple is constructed and written to the system snapshot.

[0078] Extract the baseline state corresponding to the event semantic quadruple from the system snapshot, calculate the perturbation feature vector based on the baseline state and the event semantic quadruple, match the perturbation feature vector with the preset perturbation judgment condition, and the threshold condition is whether the deviation exceeds the tolerance range.

[0079] When the disturbance determination condition is met, the corresponding event semantic quadruple is marked as a disturbance event, and a disturbance event identifier is generated;

[0080] The perturbation events marked within the preset time window are deduplicated and aggregated to obtain a set of perturbation events.

[0081] Specifically, in the real-time scheduling closed loop of the industrial production line, the original reported events are first transformed into computable semantic carriers and aligned with the system snapshot. This allows the deviation between the planned baseline and real-time observations to determine disturbances and form a unified entry point for subsequent propagation and patching. In practice, after parsing and standardizing the real-time event stream, the system writes each event into an event semantic quadruple and writes it into the system snapshot. The event semantic quadruple is denoted as... ,in, The resource identifier that generated the event. Indicates the associated operation identifier, Indicates the event type identifier. This indicates the timestamp of the event. The baseline state corresponding to this four-tuple in the system snapshot is used to provide a reference value for the same resource and operation under the original scheduling plan. The baseline state is represented as... in Indicates the availability status of resources. This indicates that the task is in progress. Indicates the queue buffer status. and These represent the planned start time and planned completion time of the operation in the original scheduling plan, respectively. When the event arrives, the system uses the timestamps in the four-tuple. Construct a perturbation eigenvector with the planned time in the baseline state. The most critical time deviation component can be defined as follows: It can also incorporate changes in resource status or queue buffer status as discrete quantities. To characterize disturbance patterns such as resource availability changing to unavailability or buffer congestion, the system then generates disturbance feature vectors. The disturbance is matched against preset disturbance judgment conditions, which use a tolerance range as a threshold and can be expressed as follows: ,in This represents the tolerance interval threshold. When this threshold condition is met, the system marks the corresponding event semantic quadruple as a disturbance event and generates a disturbance event identifier for use as a seed input for the impact domain in the subsequent impact assessment stage. Since the same physical disturbance may trigger multiple reported events from the same source in a short period of time, the system classifies the marked disturbance events according to resource identifiers within a preset time window. Operation label Event type identifier The deduplication and aggregation process outputs a set of disturbance events, thereby compressing a large number of low-level noise events into a small number of semantically clear and traceable disturbance entry points. This enables subsequent closure iterations under scheduling permission set constraints to accurately delineate the influence domain with a stable set of disturbance events as input and reduce the number of unnecessary patch triggers. Ultimately, this improves the response speed and scheduling stability of the production line in scenarios such as equipment failure, delayed arrival, or deviation from execution time.

[0082] In step S2, constructing a constraint fingerprint library containing process route and resource information includes the following steps:

[0083] Acquire process route data and resource information data, and parse the process route data into a process route diagram;

[0084] Resource information data is parsed into a resource capability matrix and resource calendar constraints, and constraint expressions for process-related hard constraints are generated based on the process roadmap.

[0085] Resource-class hard constraints are generated based on the resource capability matrix and resource calendar constraints, and soft constraint expressions are generated based on the production strategy and optimization objectives.

[0086] The constraint fingerprints are obtained by normalizing and encoding hard and soft constraints, and then associated and stored with process route diagram nodes and resource identifiers to form a constraint fingerprint library.

[0087] Specifically, firstly, process route data is acquired and parsed into a process route diagram to depict the sequential relationships between processes and the connection relationships between nodes. Then, resource information data is parsed into a resource capability matrix and resource calendar constraints, where the resource capability matrix can be denoted as... Its elements Indicates that the resource identifier is Does the resource have the capability to perform the operation? The system's operational capabilities include resource calendar constraints, which describe the available time window and capacity limits of resources on the timeline. Based on this, the system generates constraint expressions for process-related hard constraints according to the process route diagram and translates the most critical process sequence constraints into a computable form, such as pre-process constraints. With subsequent operations The start time variable is denoted as The lower bound of the operation duration interval is denoted as Then the sequential constraint can be written as This is used to ensure that a subsequent operation cannot begin before a preceding operation is completed. It also generates constraint expressions for resource-based hard constraints based on the resource capability matrix and resource calendar constraints, such as for any operation... Only allow selections that meet the requirements. Resource identifier To avoid unexecutable allocations, resource availability time windows and capacity limits are expressed as constraints that allow resources to carry tasks within the corresponding time windows and ensure that the number of parallel tasks does not exceed the capacity. The system also generates soft constraint expressions based on production strategies and optimization objectives, expressing factors such as delivery time deviations, scheduling stability, and type changeovers as comparable cost items in a penalized form for subsequent selection among feasible patches. Finally, the system standardizes and encodes the above hard and soft constraint expressions to form constraint fingerprints, which are then associated with process roadmap nodes and resource identifiers for storage. The resulting constraint fingerprint library provides the ability to quickly retrieve constraints by process node and resource identifier. This allows the system to directly extract the set of constraints in the influence domain from the constraint fingerprint library and generate differential constraints after a disturbance occurs and the influence domain is obtained. This reduces the overhead of global re-solution and ensures that incremental feasibility verification is only performed on a subset of constraints related to the influence domain. Ultimately, this improves the scheduling response speed, feasibility assurance capability, and stability of local repairs of industrial production lines under real-time disturbance conditions.

[0088] In step S2, hard constraints include: process sequence, resource mutual exclusion constraints, resource availability time window constraints, and resource capacity constraints.

[0089] Soft constraints include: delivery deviation penalty constraints, scheduling stability penalty constraints, and type change / switching penalty constraints.

[0090] Specifically, the constraint expression is formed by using the operation start time as the decision variable and combining it with the operation duration interval. The sequence constraint is used to reflect the sequential relationship of the process route, and can be written as follows: ,in and These represent the start time variables for the preceding and following operations, respectively. The lower bound of the duration interval of the preceding operation is used to ensure that the subsequent operation does not complete before the preceding operation. Resource mutual exclusion constraints are used to describe the fact that only one operation can be executed on the same resource at the same time interval. This can be expressed as two operations that are assigned to the same resource. The non-overlapping condition must be met. or ,in For the corresponding start time variable, To correspond to the lower bound of the duration and thus avoid resources processing two tasks simultaneously, the resource availability time window constraint incorporates equipment maintenance downtime and shift scheduling into the feasibility assessment, which can be written as follows: ,in For operation The start time variable, Indicates that the resource identifier is The set of available resource time windows ensures that scheduling actions do not place tasks during periods of equipment unavailability. Resource capacity constraints describe the number of in-process operations that a resource with limited parallel capacity can handle at any given time does not exceed its capacity limit, and can be written as... ,in Indicates the allocation of resources The set of operations, Indicates the maximum capacity. Indicates time operate The indicator value is set to determine whether the task is within the execution range, thus preventing overload. Under the premise of satisfying the above hard constraints, soft constraints incorporate the optimization objective into the patch selection through penalty terms. The delivery deviation penalty constraint can be implemented by measuring the deviation between the task completion time and the delivery date, setting the task completion time to... And the delivery date is Then the deviation penalty can be determined by The characterization aims to reduce late and early delivery deviations. The scheduling stability penalty constraint is implemented by measuring the change in the start time after the patch relative to the original scheduling plan, with the original start time set to... And after the patch, it is Then the stability penalty can be determined by Characterization is used to suppress unnecessary changes. Switchover penalty constraints are used to reflect the cost of switching adjacent operations on the same resource and are used as a comparison basis when screening candidate patch diagrams to reduce frequent switching. These hard and soft constraints are stored in the system as constraint fingerprints associated with process nodes and resource identifiers, so that relevant constraints can be quickly retrieved and differential constraints can be generated after the influence domain is determined. In the incremental feasibility verification stage, hard constraints are used to determine whether candidate patch diagrams are feasible, while soft constraints are used to select the best among feasible patches. Ultimately, rapid local repair of disturbances is achieved while maintaining the executability and delivery stability of the production line.

[0091] In step S3, the steps for generating the set of scheduling permits are as follows:

[0092] Online calibration is performed on each type of operation based on historical logs to obtain the corresponding operation duration range;

[0093] Extract the current set of tasks to be scheduled, resource status and time window information based on system snapshots, and determine the risk budget status related to disturbance events;

[0094] The permission threshold is determined based on the risk budget status. Under the constraint of the permission threshold, candidate scheduling actions for the set of tasks to be scheduled are enumerated. For each candidate scheduling action, the risk increment is calculated based on the operation duration interval and system snapshot. The candidate scheduling action is written into the scheduling permission set when the risk increment does not exceed the permission threshold.

[0095] Specifically, the operation duration range is obtained by online calibration of each type of operation from historical logs. in and These respectively represent the operation identifiers. The shortest and longest possible durations are used for conservative constraints and risk assessment under uncertain execution times. Simultaneously, information such as the current set of tasks to be scheduled, resource availability status, and resource availability time windows are extracted from system snapshots. Based on this, the risk budget status related to disturbance events is determined. The risk budget status characterizes the current allowable risk margin and sets the permissible threshold accordingly. ,in This represents the maximum allowable risk increment for this round of local repair. Subsequently, under the constraint of the allowable threshold, candidate scheduling actions for the set of tasks to be scheduled are enumerated and their risk increments are calculated one by one. ,in This represents a candidate scheduling action whose value corresponds to an adjustment of the start time, resource reallocation, or order adjustment of certain operations. This indicates the additional risk introduced by the action relative to the original scheduling plan in the system snapshot. The calculation of the risk increment is based on benchmark quantities such as the operation duration interval and the planned start and completion times in the system snapshot to estimate the potential time offset and resource usage changes caused by the action, and to form a comparable risk metric. The system uses discrimination conditions... As a permission gating mechanism, candidate scheduling actions that meet the conditions are written into the scheduling permission set. The resulting set of scheduling permissions serves two purposes: firstly, as a constraint on the closure iterative propagation during the impact assessment phase, ensuring that the propagation operator only expands the influence domain within the allowed scheduling actions; and secondly, as the boundary of the type and scope of atomic changes generated during the patch synthesis phase, guaranteeing that the subsequently generated candidate patch graph and the final target scheduling plan maintain a controllable change range under risk budget constraints and reduce the impact of disturbance repair on the overall stability of the production line.

[0096] The steps for performing closure iteration in step S4 are as follows:

[0097] The seed set of the influence domain is determined based on the set of disturbance events, the influence domain is initialized as the seed set of the influence domain, and the influence domain is used as the current set of nodes;

[0098] The rule-propagation dual is extracted from the historical log by symbolic induction, and its constraint expression is judged to meet the triggering condition under the constraint of the scheduling permission set.

[0099] For rules that meet the triggering conditions, invoke the propagation operator to generate a candidate set of extended nodes from the current set of nodes;

[0100] The new influence domain is obtained by merging the union of the candidate expanded node sets with the influence domain and updating it.

[0101] If the new influence domain is the same as the previous influence domain, the iteration terminates and the influence domain is output; otherwise, the node set is expanded as the new current node set and the iteration continues.

[0102] Specifically, the seed set of the influence domain is first determined from the set of disturbance events and used as the initial influence domain and the current node set. Let the initial set be... And the influence domain iteration result is The set elements correspond to task identifiers or resource identifiers associated with the disturbance event. The system then extracts several rule-propagation pairs from historical logs through symbolic induction. Each rule-propagation pair consists of a constraint expression and a propagation operator. The constraint expression describes under what technological relationships or resource states the disturbance will propagate to adjacent tasks or resources. The propagation operator generates a new candidate set of extended nodes from the current node set when the triggering condition is met. In each iteration, the system determines whether the constraint expression of the rule-propagation pair meets the triggering condition under the constraints of the scheduling permission set and calls the propagation operator for rules that meet the triggering condition, obtaining the set of extended nodes and merging and updating it with the current influence domain. The iterative process can be represented by a set recursive formula. in This indicates that all propagation operators that satisfy the triggering condition act together on the current set. and subject to the set of scheduling permissions The extended set of nodes obtained after constraints The constraint ensures that propagation only extends along the relationships covered by subsequent adjustable scheduling actions, avoiding the inclusion of practically unadjustable regions in the influence domain, which would lead to an expansion of the subsequent patch search space; the system uses fixed-point determination as the termination condition, when... At this point, it is assumed that the propagation closure has converged and the final influence domain is output. Otherwise, the newly added extended nodes will be used as the new current node set to continue iterating until convergence. This closure iteration mechanism can quickly and interpretably determine the boundaries of affected tasks and resources after a disturbance occurs. It ensures that the influence domain includes all objects that may need to be adjusted under the permission constraints, and limits the influence to a local range that is reachable from the disturbance propagation. This provides an accurate scope for subsequent differential constraint generation and incremental feasibility verification, and significantly reduces the computational overhead and plan oscillation of dynamic scheduling of industrial production lines in real-time disturbance scenarios.

[0103] In step S5, the steps for generating difference constraints and distinguishing variables are as follows:

[0104] Obtain the affected domain and extract the task identifier, resource identifier, and planned start time / planned completion time related to the affected domain based on the system snapshot;

[0105] Based on the constraint fingerprint database, the constraint expressions of hard and soft constraints associated with the influence domain are retrieved to obtain the set of influence domain constraints;

[0106] Based on the original scheduling plan corresponding to the system snapshot, mark the newly added, invalidated or replaced constraint expressions in the influence domain and generate corresponding differential constraints;

[0107] Variables outside the influence domain are categorized into the frozen variable set, and variables within the influence domain are categorized into the variable set. Difference constraints are then stored in association with the frozen variable set and the variable set.

[0108] Specifically, first, the affected domain is obtained, and the task identifiers and resource identifiers corresponding to the affected domain, as well as the planned start and completion times of these tasks in the original scheduling plan, are extracted from the system snapshot. The start time variable for each operation within the affected domain is then denoted as... And the corresponding start time in the original scheduling plan is recorded as The system then retrieves the constraint expressions of hard and soft constraints based on the constraint fingerprint database according to the process nodes and resource identifiers in the influence domain, forming an influence domain constraint set. Using the original scheduling plan in the system snapshot as a benchmark, these constraint expressions are differentially marked, and constraints that need to be added and enabled due to disturbances, constraints that have become invalid or disabled due to resource unavailability or task cancellation, and constraints that need to be replaced and updated due to parameter changes are marked and differential constraint sets are generated. ,in This represents the incremental change relative to the original scheduling plan constraint set and is used for subsequent consistency checks only on the changed parts; at the same time, the system distinguishes variables according to whether they are within the influence domain, and classifies start time variables and resource allocation variables outside the influence domain into the frozen variable set. And maintain its value equal to the original planned value, for example, for any frozen variable, it is required that... The relevant variables within the influence domain are categorized into the set of variable variables. Allows for atomic changes to be adjusted within a patch, thus only affecting subsequent candidate patch graphs when applied. Update candidate scheduling state within the range without changing The system will take the value of the difference constraint set. With frozen variable set and the set of variable variables Associative storage enables the patch synthesizer to automatically restrict changes to only apply to the set of variable variables and subject them to difference constraints when generating atomic changes. Simultaneously, incremental feasibility verification only requires checking the variables generated by the patch synthesizer. The affected subset of the marked constraints is checked and combined with the freeze condition. This ensures that plans in unaffected areas remain undisturbed, thereby enabling rapid local repair of disturbances to industrial production lines and significantly reducing plan oscillations and computational overhead caused by global reordering.

[0109] In step S5, the steps for constructing the candidate patch graph are as follows:

[0110] Obtain the influence domain, differential constraints, set of frozen variables, set of variable variables, and set of scheduling permissions;

[0111] Under the constraints of the scheduling permission set, enumerate the candidate set of atomic changes that act on the set of variable variables, and perform local consistency checks based on the difference constraints and the set of frozen variables to filter out atomic changes that violate hard constraints, thus obtaining the set of available atomic changes.

[0112] A set of patch graph nodes is constructed based on the set of available atomic changes, and the dependencies between nodes are determined based on differential constraints and written into the dependency edge set;

[0113] For nodes with predecessor dependencies, complete the predecessor nodes, and mark the node pairs with mutual exclusion dependencies to obtain the candidate patch graph.

[0114] Specifically, first obtain the influence domain and the set of difference constraints. Frozen variable set Set of variable variables and scheduling permit set ,in Give the range of scheduling variables that are allowed to be modified. Specify the freeze conditions that must remain equal to the original scheduling plan value. For example, for any freeze operation, the following conditions must be met. The system is Enumeration under constraints only applies to The atomic change candidate set is defined, and a local consistency check is performed on each candidate atomic change. The local consistency check is based on... The hard constraint subset marked in the middle and the freezing condition are used to determine whether an atomic change will directly cause the hard constraint to become unsatisfactory. For example, when an atomic change causes a pair of mutually exclusive operations to overlap on the same resource, thus failing to satisfy the non-overlapping condition. or When an atomic change is rejected, the remaining atomic changes form a set of available atomic changes. The system then uses the available atomic changes as the node set of the patch graph and determines the dependencies between nodes based on differential constraints to form a dependency edge set. The dependency relationship is used to reflect the precondition that one atomic change is required for another atomic change to be applied first, or the mutually exclusive condition that the two cannot be met at the same time, thereby ensuring that the patch application order and combination selection conform to the constraint logic. After the graph is constructed, the system fills in the preconditions of nodes with preconditions to ensure that the candidate patches have closed executable links, and marks the nodes with mutually exclusive dependencies to avoid generating patches containing mutually exclusive combinations during subsequent verification and selection. The final candidate patch graph provides clear application order and pruning information for subsequent incremental feasibility verification, enabling the dynamic scheduling system to quickly generate a small number of high-quality candidate patches within the influence domain and reduce the solution overhead caused by invalid combinations. At the same time, since all atomic changes are constrained by the scheduling permission set and only act on the set of variable variables, it can achieve local repair of production disturbances under the premise of controllable risk and stable plan.

[0115] In step S6, the feasibility verification steps are as follows:

[0116] Obtain the candidate patch graph, differential constraints, set of frozen variables, set of variable variables, and the original scheduling plan and operation duration range from the system snapshot;

[0117] The atomic changes in the candidate patch graph are sorted topologically by dependency relationship to obtain the atomic change sequence;

[0118] Based on the original scheduling plan, atomic changes are applied one by one in the atomic change sequence only within the set of variable variables to form incrementally updated candidate scheduling states, while keeping the values ​​of the corresponding variables in the set of frozen variables unchanged;

[0119] After each atomic change is applied, an incremental consistency check is performed on the differential constraints based on the operation time interval. If any check fails, the candidate patch graph is determined to be infeasible and the verification is terminated. If all checks pass, the candidate patch graph is determined to be feasible and output as a feasible patch.

[0120] Specifically, the first step is to obtain the candidate patch graph and its corresponding set of difference constraints. Frozen variable set Variable variable set The original scheduling plan is read from the system snapshot as the baseline state, and the original value of any scheduling variable in the original scheduling plan is denoted as... Among them, variables belonging to the frozen variable set must always satisfy The system then performs topological sorting of atomic changes based on the pre-dependencies between nodes in the candidate patch graph, obtaining a sequence of atomic changes that satisfy dependency constraints, which is used to determine a safe application order. During the verification process, the system uses the original scheduling plan as the initial state and only applies changes to the set of variable variables. Within the specified range, atomic changes are applied sequentially according to the atomic change sequence to form a progressively updated candidate scheduling state, and immediately after each application, the scheduling state is updated based on the operation duration interval. Perform incremental consistency checks on differential constraints, where the consistency checks include at least the process sequence constraints. Resource mutual exclusion constraint or The verification of resource availability time window constraints and resource capacity constraints is performed such that if any differential constraint cannot be satisfied in the current incremental state, the candidate patch graph is determined to be infeasible and the verification is terminated, thus avoiding the continued consumption of computing resources on invalid patches. If all atomic changes are applied sequentially and the incremental consistency check at each step passes, the candidate patch graph is determined to remain feasible within the freeze condition and operation duration uncertainty range and is output as a feasible patch. The role of this mechanism is to transform the patch feasibility judgment from a global one-time verification to an interpretable and pruning incremental verification process, enabling the system to quickly screen out inexecutable patches and prioritize scheduling and repair schemes that meet hard constraints and have a controlled scope of modification, thereby improving the scheduling response efficiency and execution reliability of industrial production lines under real-time disturbance scenarios.

[0121] Please see the appendix Figure 2 An AI-based dynamic scheduling system for industrial production lines, comprising the following modules:

[0122] The event semanticization and state snapshot module is used to acquire real-time event streams and convert them into event semantic quadruples to update system snapshots, identify disturbance events and obtain a set of disturbance events;

[0123] The constraint fingerprint library module is used to build a constraint fingerprint library containing process route and resource information and to store hard and soft constraints.

[0124] The parameter calibration and license generation module is used to calibrate the operation duration range online based on historical logs, and generate a set of scheduling licenses by combining risk budget and system snapshots;

[0125] The impact assessment module is used to extract rule-propagation duals from historical logs through symbolic induction, and to perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain.

[0126] The patch generation and execution module is used to generate differential constraints within the influence domain and distinguish between the set of frozen variables and the set of variable variables. Under the constraints of differential constraints, the set of frozen variables, the set of variable variables, and the set of scheduling permissions, it generates a candidate patch graph, performs incremental feasibility verification on the candidate patch graph, selects the target patch, generates the target scheduling plan, issues it for execution, and maintains real-time closed-loop iteration.

[0127] Specifically, the event semantics and state snapshot module continuously receives real-time event streams from equipment and job execution and transforms them into event semantic quadruplets containing resource identifiers, operation identifiers, event type identifiers, and timestamps, which are then written into the system snapshot. This allows the system snapshot to reflect key states such as resource availability and planned start and end times in real time, and to identify disturbance events to form a set of disturbance events as input for subsequent reasoning. The constraint fingerprint library module parses process routes and resource information and stores hard constraints such as process sequence, resource mutual exclusion, resource availability time windows, and resource capacity, as well as soft constraints such as delivery deviation penalties, scheduling stability penalties, and changeover penalties, in the form of standardized constraint fingerprints, associated with process nodes and resource identifiers. This enables the system to quickly retrieve relevant constraints according to their impact range and supports differential updates. The parameter calibration and permit generation module calibrates the operation duration range online from historical logs and generates a set of scheduling permits by combining risk budgets and system snapshots. This ensures that allowed scheduling actions and their scope of action meet risk budget constraints and provides gating boundaries for propagation and patch synthesis. The impact assessment module uses symbolic representation... The system extracts rules from historical logs, propagates their duality, and performs closure iterative propagation on the set of disturbance events under the constraints of the scheduling permission set to obtain the influence domain. This limits the potential chain reaction of disturbances to a subset of adjustable and reachable task resources. The patch generation and execution module generates differential constraints within the influence domain based on system snapshots and constraint fingerprint databases, freezes variables outside the influence domain, and sets variables within the influence domain as variable. This ensures that local repairs do not change the plan stability of unaffected areas. Under the constraints of differential constraints, frozen variable sets, variable variable sets, and scheduling permission sets, it generates a candidate patch graph composed of atomic changes and performs incremental feasibility verification to filter out patches that violate hard constraints. Finally, it selects feasible target patches, generates target scheduling plans, issues them for execution, and uses execution feedback to drive the event flow to continue updating the system snapshot, forming a real-time closed-loop iteration. Therefore, the system can quickly locate the affected area, reduce plan oscillations caused by global rescheduling, and improve the response speed, executability, and scheduling stability of the production line in disturbance scenarios such as equipment failure, delayed arrival, or deviation in execution time.

[0128] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic scheduling method for industrial production lines based on artificial intelligence, characterized in that, Includes the following steps: S1, State Awareness: Converts real-time event streams into event semantic quadruples and updates system snapshots to identify disturbance events; S2. Constraint Modeling: Construct a constraint fingerprint library containing process route and resource information to store hard and soft constraints in the production process; S3, Parameter Calibration: Based on historical logs, online calibration operation duration range is performed, and a scheduling permit set is generated by combining risk budget and snapshot; S4. Impact Assessment: Extract rule-propagation duality from historical logs through symbolic induction, and perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain; S5. Patch Synthesis: Generate differential constraints and distinguish variables within the influence domain, and generate candidate patch graphs composed of atomic changes through the synthesizer; S6. Execution Filtering: Perform incremental feasibility verification on the candidate patch graph, select the target patch, generate the target scheduling plan, issue it for execution, and maintain real-time closed-loop iteration.

2. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S1, the steps for identifying disturbance events are as follows: The real-time event stream is parsed and standardized. Based on the resource identifier, operation identifier, event type identifier and timestamp of the event, an event semantic quadruple is constructed and written to the system snapshot. Extract the baseline state corresponding to the event semantic quadruple from the system snapshot, calculate the perturbation feature vector based on the baseline state and the event semantic quadruple, match the perturbation feature vector with the preset perturbation judgment condition, and the threshold condition is whether the deviation exceeds the tolerance range. When the disturbance determination condition is met, the corresponding event semantic quadruple is marked as a disturbance event, and a disturbance event identifier is generated; The perturbation events marked within the preset time window are deduplicated and aggregated to obtain a set of perturbation events.

3. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S2, constructing a constraint fingerprint database containing process route and resource information includes the following steps: Acquire process route data and resource information data, and parse the process route data into a process route diagram; Resource information data is parsed into a resource capability matrix and resource calendar constraints, and constraint expressions for process-related hard constraints are generated based on the process roadmap. Resource-class hard constraints are generated based on the resource capability matrix and resource calendar constraints, and soft constraint expressions are generated based on the production strategy and optimization objectives. The constraint fingerprints are obtained by normalizing and encoding hard and soft constraints, and then associated and stored with process route diagram nodes and resource identifiers to form a constraint fingerprint library.

4. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S2, the hard constraints include: process sequence, resource mutual exclusion constraints, resource availability time window constraints, and resource capacity constraints. The soft constraints include: delivery deviation penalty constraints, scheduling stability penalty constraints, and type change / switching penalty constraints.

5. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S3, the steps for generating the scheduling permit set are as follows: Online calibration is performed on each type of operation based on historical logs to obtain the corresponding operation duration range; Extract the current set of tasks to be scheduled, resource status and time window information based on system snapshots, and determine the risk budget status related to disturbance events; The permission threshold is determined based on the risk budget status. Under the constraint of the permission threshold, candidate scheduling actions for the set of tasks to be scheduled are enumerated. For each candidate scheduling action, the risk increment is calculated based on the operation duration interval and system snapshot. The candidate scheduling action is written into the scheduling permission set when the risk increment does not exceed the permission threshold.

6. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, The steps for performing closure iteration in step S4 are as follows: The seed set of the influence domain is determined based on the set of disturbance events, the influence domain is initialized as the seed set of the influence domain, and the influence domain is used as the current set of nodes; The rule-propagation dual is extracted from the historical log by symbolic induction, and its constraint expression is judged to meet the triggering condition under the constraint of the scheduling permission set. For rules that meet the triggering conditions, invoke the propagation operator to generate a candidate set of extended nodes from the current set of nodes; The new influence domain is obtained by merging the union of the candidate expanded node sets with the influence domain and updating it. If the new influence domain is the same as the previous influence domain, the iteration terminates and the influence domain is output; otherwise, the node set is expanded as the new current node set and the iteration continues.

7. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S5, the steps for generating difference constraints and distinguishing variables are as follows: Obtain the affected domain and extract the task identifier, resource identifier, and planned start time / planned completion time related to the affected domain based on the system snapshot; Based on the constraint fingerprint database, the constraint expressions of hard and soft constraints associated with the influence domain are retrieved to obtain the set of influence domain constraints; Based on the original scheduling plan corresponding to the system snapshot, mark the newly added, invalidated or replaced constraint expressions in the influence domain and generate corresponding differential constraints; Variables outside the influence domain are categorized into the frozen variable set, and variables within the influence domain are categorized into the variable set. Difference constraints are then stored in association with the frozen variable set and the variable set.

8. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S5, the steps for constructing the candidate patch image are as follows: Obtain the influence domain, differential constraints, set of frozen variables, set of variable variables, and set of scheduling permissions; Under the constraints of the scheduling permission set, enumerate the candidate set of atomic changes that act on the set of variable variables, and perform local consistency checks based on the difference constraints and the set of frozen variables to filter out atomic changes that violate hard constraints, thus obtaining the set of available atomic changes. A set of patch graph nodes is constructed based on the set of available atomic changes, and the dependencies between nodes are determined based on differential constraints and written into the dependency edge set; For nodes with predecessor dependencies, complete the predecessor nodes, and mark the node pairs with mutual exclusion dependencies to obtain the candidate patch graph.

9. The method for dynamic scheduling of industrial production lines based on artificial intelligence according to claim 1, characterized in that, In step S6, the feasibility verification steps are as follows: Obtain the candidate patch graph, differential constraints, set of frozen variables, set of variable variables, and the original scheduling plan and operation duration range from the system snapshot; The atomic changes in the candidate patch graph are sorted topologically by dependency relationship to obtain the atomic change sequence; Based on the original scheduling plan, atomic changes are applied one by one in the atomic change sequence only within the set of variable variables to form incrementally updated candidate scheduling states, while keeping the values ​​of the corresponding variables in the set of frozen variables unchanged; After each atomic change is applied, an incremental consistency check is performed on the differential constraints based on the operation time interval. If any check fails, the candidate patch graph is determined to be infeasible and the verification is terminated. If all checks pass, the candidate patch graph is determined to be feasible and output as a feasible patch.

10. An artificial intelligence-based dynamic scheduling system for industrial production lines, characterized in that: The system for the AI-based dynamic scheduling method for industrial production lines according to any one of claims 1-9 includes the following modules: The event semanticization and state snapshot module is used to acquire real-time event streams and convert them into event semantic quadruples to update system snapshots, identify disturbance events and obtain a set of disturbance events; The constraint fingerprint library module is used to build a constraint fingerprint library containing process route and resource information and to store hard and soft constraints. The parameter calibration and license generation module is used to calibrate the operation duration range online based on historical logs, and generate a set of scheduling licenses by combining risk budget and system snapshots; The impact assessment module is used to extract rule-propagation duals from historical logs through symbolic induction, and to perform closure iteration on the set of disturbance events under the constraints of the scheduling permission set to obtain the impact domain. The patch generation and execution module is used to generate differential constraints within the influence domain and distinguish between the set of frozen variables and the set of variable variables. Under the constraints of differential constraints, the set of frozen variables, the set of variable variables, and the set of scheduling permissions, it generates a candidate patch graph, performs incremental feasibility verification on the candidate patch graph, selects the target patch, generates the target scheduling plan, issues it for execution, and maintains real-time closed-loop iteration.

Images