Manufacturing workshop production scheduling method and system based on multi-agent cooperation

Abstract

The application discloses a manufacturing workshop production scheduling method and system based on multi-agent cooperation, relates to the technical field of intelligent manufacturing and production scheduling, and comprises the following steps: acquiring workshop states, process resource requirements, resource time sequence dependencies and performance information, establishing process agents and resource agents, and generating resource initial realization domains; constructing resource dependency subgraphs, generating joint resource realization domains through release time point dependency propagation and time window closure; applying temporary resource tokens and generating conditional scheduling commitments; generating target scheduling schemes through an improved dynamic neighborhood search algorithm; establishing elastic margin accounts and performing disturbance absorption and transmission; generating a disturbance influence cone and performing local recovery when the conditional scheduling commitment fails; determining source failure resources, updating performance parameters, and regenerating resource initial realization domains. The application improves the realization capability of scheduling schemes, the disturbance recovery efficiency and the stability of workshop production scheduling under complex resource dependency conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production scheduling technology, and in particular to a manufacturing workshop production scheduling method and system based on multi-agent collaboration. Background Technology

[0002] As manufacturing workshops evolve towards multi-variety, small-batch, and flexible production, the collaborative scheduling and anomaly rescheduling of equipment, personnel, tooling, and materials have become crucial aspects of production execution. Existing technologies have employed multi-agent and dynamic scheduling methods. For example, CN120087650A discloses a method where a task agent, resource agent, personnel agent, and scheduling decision agent collaboratively process production status and perform rescheduling after abnormal disturbances; CN112308472B discloses a method to absorb delays and reduce delay propagation through process time lag relaxation; and CN108121305B discloses a method to establish a related task association tree starting from the disturbed task and dynamically adjust the affected tasks. These technologies demonstrate that multi-agent collaboration, time margin absorption of disturbances, and local rescheduling have been practically applied in the field of manufacturing scheduling.

[0003] However, existing scheduling typically organizes processes according to the availability of individual resources, making it difficult to handle the distortion of common execution conditions caused by the interdependence of multiple resource release times; the time margin of processes mainly serves as a static buffer, which poses a risk of duplicate occupation when multiple disturbance paths act simultaneously; after an anomaly occurs, the affected scope is mostly determined based on the preset task association relationship, making it difficult to dynamically shrink or expand with the actual remaining disturbance, and there is a lack of distinction between the source of the failed resource and the passively affected resource, which makes subsequent resource status corrections susceptible to propagation delay interference, affecting the stability and executability of continuous scheduling.

[0004] Therefore, how to provide a production scheduling method and system for manufacturing workshops based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a production scheduling method and system for manufacturing workshops based on multi-agent collaboration. This invention fully utilizes multi-agent collaboration, resource fulfillment modeling, improved dynamic neighborhood search, and disturbance propagation analysis methods to achieve collaborative scheduling, dynamic optimization, and local recovery of processes and resources. It also forms a closed-loop scheduling mechanism through performance parameter updates, possessing advantages such as strong resource coordination capabilities, high disturbance adaptability, high local recovery efficiency, and good scheduling scheme stability.

[0006] The manufacturing workshop production scheduling method based on multi-agent cooperation according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process Agent and resource Agent, and generate the initial resource fulfillment domain;

[0008] Step 2: Construct a resource dependency subgraph based on process resource requirements and resource timing dependencies, perform release time dependency propagation and time window closure on the initial resource fulfillment domain, and generate a joint resource fulfillment domain;

[0009] Step 3: The process agent applies for a temporary resource token based on the joint resource redemption domain, locks the resource after verifying the execution time window and pre-redemption conditions, determines the allowable elasticity margin, and generates a conditional scheduling commitment.

[0010] Step 4: Input the conditional scheduling commitment and joint resource fulfillment domain into the improved dynamic neighborhood search algorithm, filter the commitments to be adjusted according to the commitment status, perform time window shifting, resource replacement and reordering, expand the search range according to the remaining disturbance after the time disturbance is absorbed by the allowable elastic margin, and generate the target scheduling scheme after fulfillment constraint verification.

[0011] Step 5: Establish an elastic margin account based on the target scheduling scheme, and perform margin pre-occupancy, exclusive deduction, remaining disturbance transfer, and margin rollback for time disturbances;

[0012] Step 6: When the conditional scheduling commitment fails, generate a disturbance influence cone based on disturbance propagation and elastic margin consumption, determine the boundary anchor point, and call the improved dynamic neighborhood search algorithm to perform local recovery;

[0013] Step 7: Determine the source failure resource based on the disturbance propagation relationship, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource based on the update results.

[0014] Optionally, step one specifically includes:

[0015] Create a process agent based on the process identifier, create a resource agent based on the resource identifier, and create candidate associations between process agents and resource agents based on the process resource requirements;

[0016] The performance residual between the actual release time of resources and the planned release time of resources is calculated based on the performance information, and the historical performance residual and the realization calibration amount are determined based on the preset historical statistics window.

[0017] The reliable release time is determined based on the resource plan release time, the current occupied process plan end time, historical performance residuals, and fulfillment calibration quantities.

[0018] The available time window for resources is determined based on the trusted release time, resource work calendar, locked execution time window, and resource running status, and the fulfillment confidence level is determined based on historical fulfillment results.

[0019] By associating the trusted release time, resource availability time window, historical performance residuals, fulfillment calibration amount, and fulfillment confidence with the resource agent, an initial resource fulfillment domain is generated.

[0020] Optionally, step two specifically includes:

[0021] Based on the resource requirements of the process, resource nodes are determined. According to the execution sequence of resource preparation, type change, handling, installation, arrival, kitting, material loading and process handover, directed dependency edges are established between resource nodes with sequential dependencies and the duration of dependency actions is configured to generate a resource dependency subgraph.

[0022] The release time dependency propagation is performed according to the topological order of the resource dependency subgraph. The sum of the trusted release time of the preceding resource and the duration of the dependent action is compared with the trusted release time of the subsequent resource, and the latest time is determined as the trusted release time of the subsequent resource after correction.

[0023] Based on the corrected reliable release time, the available resource time window is extracted. The intersection calculation is performed on the available resource time windows of the resources required to complete the same process. The intersection interval with a duration shorter than the process processing time is deleted, and a joint available time window is generated.

[0024] The minimum value among the realization confidence of each resource participating in the time window closure is determined as the joint realization confidence, and it is combined with the corrected reliable release time of each resource and the joint available time window to generate the joint resource realization domain.

[0025] Optionally, step three specifically includes:

[0026] The process agent determines the candidate execution time window based on the combined available time window, generates a pre-reservation transaction identifier, and sends a resource pre-reservation request to the corresponding resource agent.

[0027] The resource agent verifies the occupancy relationship between the candidate execution time window and the available resource time window, the locked execution time window, and the valid temporary reserved time window. After the verification is successful, a temporary resource token is generated.

[0028] The pre-emption conditions are determined based on the completion status of the direct preceding process and the inclusion relationship between the candidate execution time window and the joint resource redemption domain. Consistency checks are performed on temporary resource tokens associated with the same pre-occupied transaction identifier.

[0029] After the consistency check passes, a commitment identifier is generated. The commitment identifier is associated with a temporary resource token and the temporary resource token and the candidate execution time window are converted to a locked state. The allowable elasticity margin is determined based on the time interval between the locked execution time window and the joint available time window, the direct subsequent process and the next conditional scheduling commitment on the same locked resource, and a conditional scheduling commitment is generated.

[0030] If any resource pre-allocation fails or consistency check fails, the corresponding temporary resource token and temporary reserved time window will be released.

[0031] Optionally, step four specifically includes:

[0032] The execution order of the corresponding processes is determined based on the conditional scheduling commitment. The subsequent resource occupancy relationship is determined based on the conditional scheduling commitment and execution order of the shared resources. The conditional scheduling commitment, the joint resource fulfillment domain, the process precedence relationship, and the resource subsequent occupancy relationship are input into the improved dynamic neighborhood search algorithm.

[0033] Based on the commitment status and joint resource fulfillment domain, commitments to be adjusted are screened, and the execution time window of the commitments to be adjusted is moved, resources are replaced, and the order is changed. Among them, the resource replacement selects alternative resources from the resource agents that have established candidate associations with the corresponding process agents, and generates candidate scheduling schemes.

[0034] The dynamic neighborhood is determined based on the time disturbance, allowable elastic margin and remaining disturbance generated by the candidate scheduling scheme, and the scheduling commitment execution fulfillment constraint verification is performed based on the joint resource fulfillment domain, pre-fulfillment conditions, locked execution time window and frozen state conditions.

[0035] The resource fulfillment risk is determined based on the joint fulfillment confidence level. The evaluation value of the candidate scheme is determined based on the cumulative delay time, the number of conditional scheduling commitment adjustments, the trial surplus consumption, and the resource fulfillment risk. The neighborhood search is iteratively executed to generate the target scheduling scheme.

[0036] Optionally, determining the dynamic neighborhood based on the time disturbance, allowable elastic margin, and remaining disturbance generated by the candidate scheduling scheme specifically includes:

[0037] The positive time offset generated after the adjustment of the execution time window of the commitment to be adjusted is determined as the initial time disturbance, and the allowable elasticity margin participating in the current candidate scheduling scheme is copied as the initial trial margin.

[0038] The time disturbance is propagated along the successive direction of the process sequence and the successive resource occupancy relationship. After receiving a non-zero time disturbance, the conditional scheduling commitment adds it to the dynamic neighborhood, determines the smaller value between the time disturbance and the remaining trial balance as the trial consumption, and performs the deduction.

[0039] The current propagation branch terminates when the remaining perturbation after consumption is zero; propagation continues when the remaining perturbation is greater than zero.

[0040] When multiple propagation paths arrive at the same conditional scheduling commitment, the larger of the recorded cumulative disturbance and the newly arrived disturbance will be determined as the updated cumulative disturbance, and only the newly added disturbance will be subject to trial balance deduction.

[0041] After all propagation branches terminate, the conditional scheduling commitments added to the dynamic neighborhood are determined as the search scope of the current candidate scheduling scheme.

[0042] Optionally, step five specifically includes:

[0043] Update the execution time window of the conditional scheduling commitment according to the target scheduling scheme, redetermine the allowable elasticity margin, and establish an elasticity margin account for the locked conditional scheduling commitment. The elasticity margin account includes the currently available margin, the reserved margin, the consumed margin, and the account version number.

[0044] After the time disturbance arrives and the scheduling commitment is made, the request quantity is determined based on the time disturbance amount and the current available balance. After the account version number consistency check is performed, the request quantity is deducted from the current available balance and written into the pre-occupied balance.

[0045] After the remaining balance is reserved, an exclusive deduction is performed, transferring the reserved balance to the consumed balance, so that the reserved balance and the consumed balance are no longer included in the current available balance;

[0046] The time amount remaining after deducting the actual exclusive deduction from the time disturbance amount is determined as the residual disturbance and continues to be propagated.

[0047] When the actual end time of a process is earlier than the expected end time after considering the current disturbance, the time difference between the two is determined as the rollback margin. After resource occupation and margin pre-occupancy verification, the smaller value between the rollback margin and the consumed margin is transferred back to the current available margin.

[0048] Optionally, step six specifically includes:

[0049] If the preconditions for execution are not met at the start of execution or the locked resources are not yet available, the time difference between the moment when the actual execution conditions are met and the start of the execution time window is determined as the root disturbance.

[0050] When the actual execution of a process exceeds the latest end time jointly defined by the end time of the execution time window and the current available margin, the time difference between the actual end time and the latest end time is determined as the root disturbance, and the corresponding conditional scheduling commitment is converted into a violation state.

[0051] The conditional scheduling commitment that violates the state is used as the root node of the disturbance and added to the disturbance influence cone. The root disturbance quantity is propagated along the successive direction of the process pre- and post-process relationship and the successive resource occupation relationship. The disturbance propagation relationship and the arrival time of the remaining disturbance are recorded.

[0052] After receiving a non-zero time disturbance, the conditional scheduling commitment adds it to the disturbance impact cone and calls the elastic margin account to perform margin pre-occupancy and exclusive deduction, so that the remaining disturbance after exclusive deduction can control the propagation to terminate or continue.

[0053] The first locked state conditional scheduling commitment that is directly adjacent to the disturbance influence cone and has not received a non-zero time disturbance is determined as the boundary anchor point. The locking resources and execution time window of the boundary anchor point remain unchanged, and the improved dynamic neighborhood search algorithm is called to perform local recovery of the conditional scheduling commitment inside the disturbance influence cone.

[0054] If no candidate scheduling scheme passes the constraint verification, the boundary anchor point that causes the constraint conflict will be added to the disturbance influence cone and the boundary anchor point will be redefined.

[0055] Optionally, step seven specifically includes:

[0056] Resources whose actual release time is later than the trusted release time and whose release delay reaches a preset failure deviation threshold are identified as abnormal resources.

[0057] Abnormal resources that have not received upstream residual disturbances or whose abnormal resource state occurs earlier than the earliest arrival time of upstream residual disturbances are identified as candidate source failure resources. From the candidate source failure resources, resources that do not have a preceding candidate source failure resource in the resource dependency subgraph are selected, and the candidate source failure resource with the earliest occurrence time of abnormal state is identified as the source failure resource.

[0058] Resources that have received upstream residual disturbances before the occurrence of abnormal resource status and whose actual release delay is no greater than the maximum upstream residual disturbance are identified as passively affected resources.

[0059] The historical performance residual is updated based on the time difference between the actual release time of the source failure resource and the planned release time of the resource. The fulfillment calibration amount is updated based on the calibration deviation between the actual release time and the credible release time and the preset calibration learning rate. The fulfillment confidence is updated based on the actual fulfillment result. The fulfillment calibration amount is not updated based on the release delay caused by the upstream residual disturbance for passively affected resources.

[0060] The updated historical performance residuals, performance calibration values, and performance confidence are used to form performance parameters, and the initial resource performance domain for the next scheduling cycle is regenerated using these performance parameters.

[0061] A manufacturing workshop production scheduling system based on multi-agent collaboration according to an embodiment of the present invention includes the following modules:

[0062] The resource fulfillment module is used to obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process agents and resource agents, and generate the initial resource fulfillment domain;

[0063] The joint realization module is used to construct resource dependency subgraphs, perform dependency propagation at release time and time window closure, and generate joint resource realization domains;

[0064] The conditional commitment module is used to apply for temporary resource tokens, lock resources after the execution time window and pre-fulfillment conditions are verified, determine the allowable elasticity margin, and generate conditional scheduling commitments.

[0065] The scheduling optimization module is used to perform time window shifting, resource replacement and reordering through an improved dynamic neighborhood search algorithm, and to determine the dynamic neighborhood based on the remaining disturbances to generate the target scheduling scheme.

[0066] The surplus management module is used to create flexible surplus accounts and perform surplus pre-occupancy, exclusive deduction, surplus disturbance transfer, and surplus rollback.

[0067] The local recovery module is used to generate a disturbance impact cone, determine boundary anchor points, and perform local recovery when the conditional scheduling commitment fails.

[0068] The fulfillment update module is used to identify the source failed resource, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource.

[0069] The beneficial effects of this invention are:

[0070] This invention transforms the independent availability states of equipment, personnel, tooling, and materials into a joint availability state constrained by release time dependence and a common time window through an initial resource availability domain, a resource dependency subgraph, and a joint resource availability domain. Based on the joint resource availability domain, it completes temporary resource token verification, resource locking, and conditional scheduling commitment generation, solving the problems of multiple resources being available but unable to be synchronized, and the disconnect between resource pre-occupancy and actual execution conditions. Furthermore, it utilizes an improved dynamic neighborhood search algorithm, where the commitment state determines the adjustment object, the trial consumption with allowable elasticity determines the search range, and the joint resource availability domain determines the executability of candidate solutions. This reduces invalid searches and unnecessary modifications to stable scheduling results, improving the matching degree between production scheduling and actual resource states.

[0071] This invention further transforms the allowable elastic margin into an elastic margin account with pre-occupancy, exclusive deduction, and rollback states, preventing the margin from being repeatedly consumed by multiple disturbance paths at the same time. It also controls the expansion range of the disturbance impact cone based on the remaining disturbance after the actual margin consumption. Under boundary anchor constraints, it performs local recovery only on scheduling commitments for affected conditions, solving the problems of repeated use of static time buffers and excessively large rearrangement ranges after anomalies. Simultaneously, it distinguishes between source-failed resources and passively affected resources based on the disturbance propagation relationship, updating fulfillment parameters only for source-failed resources and regenerating the initial resource fulfillment domain. This allows the execution results to continuously correct subsequent resource fulfillment judgments, forming a closed loop from resource fulfillment, scheduling commitment, disturbance recovery to fulfillment calibration, improving the executability, stability, and continuous scheduling capability of scheduling schemes in dynamic manufacturing environments. Attached Figure Description

[0072] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0073] Figure 1 This is a flowchart of the manufacturing workshop production scheduling method based on multi-agent collaboration proposed in this invention;

[0074] Figure 2 This is a schematic diagram of the improved dynamic neighborhood search algorithm in the manufacturing workshop production scheduling method based on multi-agent cooperation proposed in this invention;

[0075] Figure 3 This is the overall framework diagram of the manufacturing workshop production scheduling system based on multi-agent collaboration proposed in this invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0077] refer to Figures 1-2 A manufacturing workshop production scheduling method based on multi-agent collaboration includes the following steps:

[0078] Step 1: Obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process Agent and resource Agent, and generate the initial resource fulfillment domain;

[0079] Step 2: Construct a resource dependency subgraph based on process resource requirements and resource timing dependencies, perform release time dependency propagation and time window closure on the initial resource fulfillment domain, and generate a joint resource fulfillment domain;

[0080] Step 3: The process agent applies for a temporary resource token based on the joint resource redemption domain, locks the resource after verifying the execution time window and pre-redemption conditions, determines the allowable elasticity margin, and generates a conditional scheduling commitment.

[0081] Step 4: Input the conditional scheduling commitment and joint resource fulfillment domain into the improved dynamic neighborhood search algorithm, filter the commitments to be adjusted according to the commitment status, perform time window shifting, resource replacement and reordering, expand the search range according to the remaining disturbance after the time disturbance is absorbed by the allowable elastic margin, and generate the target scheduling scheme after fulfillment constraint verification.

[0082] Step 5: Establish an elastic margin account based on the target scheduling scheme, and perform margin pre-occupancy, exclusive deduction, remaining disturbance transfer, and margin rollback for time disturbances;

[0083] Step 6: When the conditional scheduling commitment fails, generate a disturbance influence cone based on disturbance propagation and elastic margin consumption, determine the boundary anchor point, and call the improved dynamic neighborhood search algorithm to perform local recovery;

[0084] Step 7: Determine the source failure resource based on the disturbance propagation relationship, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource based on the update results.

[0085] In this embodiment, step one specifically includes:

[0086] The system obtains production plans, process execution records, and resource occupancy records from the Manufacturing Execution System (MES); it also obtains resource operation records from the Equipment Management System (EMS), Personnel Scheduling System (MSS), Tooling Management System (MMS), and Material Management System (MMS); and it retrieves planned resource release times and actual resource release times from historical scheduling records. Using the MES clock as a unified time reference, the system associates information from different sources based on process identifiers, resource identifiers, and timestamps.

[0087] Create process agents based on process identifiers, and record process processing time, process sequence relationships, process resource requirements, and execution status. Create resource agents based on resource identifiers, and create resource agents for equipment, personnel, tooling, and materials respectively, and record the currently occupied process, the planned release time of the resource, the locked execution time window, and the resource operation status. Based on resource category, capability specification, and resource quantity, establish candidate associations between resource agents that meet the process resource requirements and the corresponding process agents.

[0088] The resource release calibration process for the performance information includes: extracting completed resource occupancy records from a preset historical statistics window according to resource identifiers, calculating the time difference between the actual resource release time and the planned resource release time in each record, and generating a performance residual sequence; determining the historical performance residual using the median of the performance residual sequence; extracting a preset calibration window from the end of the preset historical statistics window, and determining the calibration amount by using the difference between the median of the performance residuals in the preset calibration window and the historical performance residuals;

[0089] The reliable release time is determined by the following steps: comparing the planned release time of the resource with the planned end time of the currently occupied process, and determining the later time as the base release time; sequentially adding the historical performance residuals and fulfillment calibration amounts to the base release time to generate the performance correction release time; comparing the performance correction release time with the current time, and determining the later time as the reliable release time; when the resource is not occupied by a process, the current time is used instead of the planned end time of the currently occupied process in the calculation.

[0090] Based on the resource work calendar, the time period before the trusted release time, the locked execution time window, and the resource non-executable time period are deleted, and the remaining continuous time interval is determined as the resource available time window; the time periods corresponding to equipment failure or maintenance, personnel leaving their posts, tooling maintenance or being occupied by other processes, and materials not being fully set are determined as the resource non-executable time periods;

[0091] The fulfillment confidence level is determined by performing fulfillment confidence level determination on historical fulfillment records. Specifically, this includes comparing the actual resource release time of each historical fulfillment record with the corresponding credible release time, determining the records whose actual resource release time is not later than the credible release time as fulfillment success records, and determining the records whose actual resource release time is later than the credible release time as fulfillment failure records. The fulfillment confidence level is determined based on the proportion of the number of fulfillment success records to the number of valid fulfillment records.

[0092] Associate the trusted release time, resource availability time window, historical performance residual, fulfillment calibration amount and fulfillment confidence with the corresponding resource agent to generate the initial resource fulfillment domain;

[0093] Specifically, the preset historical statistics window is set to the most recent 30 valid performance records, and the preset calibration window is set to the most recent 10 valid performance records. When the number of valid performance records of the target resource is lower than the preset sample threshold, the performance records of the same type of resource are extracted according to the same resource category and capability specifications. The historical performance residual of the target resource is initialized with the median of the performance residual of the same type of resource. The preset sample threshold is set to 10 times.

[0094] In this embodiment, step two specifically includes:

[0095] Based on the resource requirements of the process, the process agent obtains the initial resource fulfillment domain from the resource agents that have established candidate associations with the process agent, and determines the execution sequence and dependent action duration of equipment preparation, changeover, tooling handling, tooling installation, personnel on-site, material preparation, material loading and process handover according to the process configuration.

[0096] Resource dependency subgraph construction is performed on resource time-series dependencies, specifically including: setting resources involved in completing the current process as resource nodes; when the current resource preparation operation is triggered by another resource reaching a release state or a preparation completion state, establishing directed dependency edges between the corresponding resource nodes from the preceding resource to the following resource, and configuring the dependency action duration for the directed dependency edges; performing topological sorting according to the directed dependency edges to generate a resource dependency subgraph; when a directed cycle exists in the resource dependency subgraph, writing the corresponding process into the resource dependency configuration exception queue and terminating the resource pre-occupancy of the corresponding process in this round;

[0097] The release time dependency propagation process is performed on the initial resource realization domain, specifically including: processing resource nodes sequentially according to the topological order of the resource dependency subgraph; keeping the trusted release time unchanged when there is no preceding resource node; when there is a preceding resource node, calculating the time after superimposing the trusted release time of each preceding resource with the corresponding dependency action duration, and determining the latest time among the obtained times and the trusted release time of the subsequent resources as the corrected trusted release time of the subsequent resources, and continuing to propagate the corrected trusted release time along the resource dependency subgraph;

[0098] The modified resource availability time windows are closed by time window closure, specifically including: deleting time intervals in each resource availability time window that are earlier than the modified reliable release time; performing intersection calculation on the remaining resource availability time windows for all resources required to complete the same process; determining the time intervals whose intersection duration reaches the process processing time as joint availability time windows; and deleting time intervals whose duration is less than the process processing time.

[0099] The minimum value among the fulfillment confidence of each resource participating in the time window closure is determined as the joint fulfillment confidence. The corrected reliable release time of each resource, the joint available time window, and the joint fulfillment confidence are combined to generate a joint resource fulfillment domain. When the joint available time window is empty, the corresponding process is kept in the waiting queue.

[0100] In this embodiment, step three specifically includes:

[0101] The process agent selects a time interval whose duration covers the process processing time as a candidate execution time window according to the order of the start time of the joint available time window, generates a pre-occupied transaction identifier, and sends a resource pre-occupancy request carrying the pre-occupied transaction identifier, process identifier and candidate execution time window to the resource agent corresponding to the resources required to complete the process.

[0102] The resource agent performs resource occupancy checks on candidate execution time windows, specifically including: determining whether the candidate execution time window is within a resource available time window, determining whether the candidate execution time window overlaps with a locked execution time window, and determining whether the candidate execution time window overlaps with a valid temporary reserved time window; after all checks pass, the candidate execution time window is written to the temporary reserved area and a temporary resource token is generated through atomic state update; if any check fails, a resource reservation failure status is returned.

[0103] The temporary resource token records the reserved transaction identifier, resource identifier, process identifier, candidate execution time window, generation time, expiration time, and token status; when the preset token validity period is reached and no locking instruction is received, the corresponding temporary reserved time window is deleted, and the temporary resource token is converted to a released state;

[0104] The pre-execution conditions are determined based on the completion status of the direct preceding process and the inclusion relationship between the candidate execution time window and the joint resource fulfillment domain. The pre-execution conditions include that the direct preceding process is completed before the start time of the candidate execution time window and that the candidate execution time window is within the joint available time window.

[0105] Perform consistency verification on all temporary resource tokens associated with the same pre-occupied transaction identifier. Specifically, this includes: verifying the pre-occupied transaction identifiers of all temporary resource tokens, finding the common intersection of the execution time windows corresponding to all temporary resource tokens, determining whether the duration of the common intersection covers the processing time of the process, and determining whether the pre-redemption conditions are met.

[0106] If any resource pre-allocation fails or consistency verification fails, the process Agent sends a release instruction based on the pre-allocation transaction identifier, converts the temporary resource tokens generated in this round to a released state, and deletes the corresponding temporary reserved time window;

[0107] When the consistency check passes, the process agent generates a unique commitment identifier, associates the commitment identifier with all temporary resource tokens corresponding to the same pre-occupied transaction identifier, and sends a lock instruction to the corresponding resource agent; the resource agent converts the temporary resource tokens to a locked state and converts the corresponding candidate execution time window into a locked execution time window;

[0108] The allowable elasticity margin is determined by the following steps: calculating the time interval from the end of the locked execution time window to the end of the joint available time window; when there is a directly subsequent locked operation, calculating the time interval from the end of the current operation to the start of the directly subsequent locked operation; when there is a next conditional scheduling commitment for the same locked resource, calculating the time interval from the end of the current execution time window to the start of the next conditional scheduling commitment execution time window; and determining the minimum non-negative value among all valid time intervals as the allowable elasticity margin.

[0109] The conditional scheduling commitment consists of the commitment identifier, corresponding process, locked resources, execution time window, pre-fulfillment conditions, allowed flexibility margin, and commitment status. The commitment status includes provisional status, locked status, frozen status, violated status, and completed status. After all resources are locked in a consistent manner, the conditional scheduling commitment is set to the locked status.

[0110] Specifically, the preset token validity period is set to 5 seconds.

[0111] In this embodiment, step four specifically includes:

[0112] The execution order of the corresponding processes is determined based on the conditional scheduling commitments. The subsequent resource occupancy relationship is determined based on the conditional scheduling commitments and execution order of the shared resources. The conditional scheduling commitments, the joint resource fulfillment domain, the process precedence relationship, and the resource subsequent occupancy relationship are input into the improved dynamic neighborhood search algorithm. The improved dynamic neighborhood search algorithm uses the conditional scheduling commitments as the search object and sets three types of neighborhood operations: time window movement, resource replacement, and reordering.

[0113] The conditional scheduling commitments are subjected to a screening process to be adjusted, which includes: conditional scheduling commitments in the frozen and completed states retaining their execution time windows and locked resources unchanged; conditional scheduling commitments in the provisional and violated states are identified as commitments to be adjusted; and conditional scheduling commitments in the locked state are identified as commitments to be adjusted when the joint resource fulfillment domain no longer covers the original execution time window, resource time conflicts occur, or prior time disturbances are received.

[0114] The specific operations for adjusting the execution of commitments include: changing the execution time window within the jointly available time window according to the preset scheduling time granularity, forming a time window movement result; selecting alternative resources from resource agents that have established candidate associations with the corresponding process agent and can cover the candidate execution time window, forming a resource replacement result; exchanging the resource occupancy order of adjacent conditional scheduling commitments that share the same resource and do not have a direct process precedence relationship, forming a reordering result; and generating candidate scheduling schemes based on the results of each neighborhood operation.

[0115] The candidate scheduling scheme is subjected to disturbance trial calculation processing, which specifically includes: calculating the positive time offset generated by the execution end time of the commitment to be adjusted after adjustment relative to the execution end time before adjustment, and determining the positive time offset as the initial time disturbance; copying the allowable elasticity margin of each conditional scheduling commitment participating in the current candidate scheduling scheme into the corresponding initial trial calculation margin, and forming the remaining trial calculation margin based on the margin value after trial consumption.

[0116] Propagate time disturbances along the successive direction of the process sequence and the successive resource occupancy relationship; after receiving a non-zero time disturbance, the conditional scheduling commitment that received the time disturbance is added to the dynamic neighborhood, the smaller value between the time disturbance and the corresponding remaining trial balance is determined as the trial consumption, and the trial consumption is deducted from the remaining trial balance; when the remaining disturbance after trial consumption is zero, the current propagation branch is terminated, and when the remaining disturbance is greater than zero, it continues to propagate to the next stage;

[0117] When multiple propagation paths reach the same conditional scheduling commitment, the larger of the recorded cumulative disturbance and the newly arrived disturbance is determined as the updated cumulative disturbance. The difference between the updated cumulative disturbance and the previous cumulative disturbance is determined as the new disturbance. Only the new disturbance is subject to trial balance deduction. The trial balance that has already been consumed is not consumed again by other propagation paths in the current candidate scheduling scheme. Dynamic neighborhood expansion stops when the remaining disturbances of all propagation branches are zero or when the remaining disturbances reach the end of a path without a subsequent conditional scheduling commitment.

[0118] The candidate scheduling schemes formed by the dynamic neighborhood are subjected to fulfillment constraint verification processing, which specifically includes: determining whether the adjusted execution time window is within the joint available time window of the corresponding joint resource fulfillment domain, determining whether the pre-fulfillment conditions are met, determining whether the alternative resources meet the process resource requirements, determining whether the execution time windows of the same resource overlap, and determining whether the execution time window and locked resources of the frozen state condition scheduling commitment remain unchanged; if any condition is not met, the corresponding candidate scheduling scheme is deleted.

[0119] The resource fulfillment risk is determined by the following steps: the resource fulfillment risk of the corresponding process is determined by subtracting the joint fulfillment confidence level of the conditional scheduling commitment from 1; when the current candidate scheduling scheme involves multiple processes, the maximum value of the resource fulfillment risk of each process is determined as the resource fulfillment risk of the current candidate scheduling scheme.

[0120] The candidate scheduling schemes that pass the fulfillment constraint verification are evaluated, including: calculating the cumulative delay time, the number of conditional scheduling commitment adjustments, the trial surplus consumption, and the resource fulfillment risk; using Min-Max normalization to unify each evaluation quantity; and performing weighted summation according to preset evaluation weights to generate candidate scheme evaluation values; when the evaluation value of a candidate scheme is lower than the current candidate scheduling scheme evaluation value, the current candidate scheduling scheme is updated, and the neighborhood search continues according to preset neighborhood switching conditions; the search ends after reaching the preset maximum search rounds or the preset number of consecutive rounds without improvement, and the candidate scheduling scheme with the smallest evaluation value during the search process is determined as the target scheduling scheme;

[0121] Specifically, the preset scheduling time granularity is set to 1 minute; the evaluation weights for cumulative delay duration, conditional scheduling commitment adjustment quantity, trial surplus consumption quantity, and resource fulfillment risk are set to 0.35, 0.30, 0.20, and 0.15, respectively; when no candidate scheduling scheme with a reduced evaluation value is generated in the current neighborhood for 3 consecutive rounds, the neighborhood operation is switched; the maximum number of search rounds is set to 100 rounds; the search ends when no candidate scheduling scheme with a reduced evaluation value is generated in any of the three types of neighborhoods for 5 consecutive rounds.

[0122] In this embodiment, step five specifically includes:

[0123] Update the execution time window of the conditional scheduling commitment according to the target scheduling scheme, and redetermine the allowable elasticity in accordance with the allowable elasticity determination method in step three; establish an elasticity account for each conditional scheduling commitment in the locked state, the elasticity account includes the current available elasticity, the reserved elasticity, the consumed elasticity and the account version number; write the redetermined allowable elasticity into the current available elasticity, and initialize the reserved elasticity and the consumed elasticity to zero;

[0124] The process of pre-allocating reserve space for time disturbances includes: after the time disturbance reaches the scheduling commitment, obtaining the account version number and current available reserve space of the elastic reserve space account, and determining the smaller value between the time disturbance amount and the current available reserve space as the requested amount; performing atomic comparison update, if the account version number is consistent with the account version number used in calculating the requested amount, deducting the requested amount from the current available reserve space, writing the requested amount into the pre-allocated reserve space, and updating the account version number; if the account version number is inconsistent, re-determining the requested amount based on the updated account status.

[0125] After the reserve is pre-allocated, an exclusive deduction process is performed, transferring the requested quantity from the pre-allocated reserve to the consumed reserve. The pre-allocated and consumed reserves are not included in the current available reserve. Subsequent time disturbances will only perform reserve pre-allocation based on the updated current available reserve, so that the same elastic reserve is not consumed repeatedly by different disturbances.

[0126] The remaining disturbance is processed by propagation, specifically including: determining the remaining disturbance as the time amount after deducting the actual exclusive deduction from the time disturbance amount; terminating the corresponding propagation branch when the remaining disturbance is zero, and continuing to propagate along the successor direction of the process sequence and the successor resource occupation relationship when the remaining disturbance is greater than zero.

[0127] Perform rollback processing on consumed reserves, specifically including: when the actual end time of the process is earlier than the expected end time after considering the current disturbance, the time difference between the two is determined as the rollback reserve; check whether there are other conditional scheduling commitments in a locked state occupying the time interval corresponding to the rollback reserve, and whether there are other valid reserves pre-occupied; if neither of the two is found, the smaller value between the rollback reserve and the consumed reserve is transferred from the consumed reserve back to the currently available reserve;

[0128] Each instance of pre-allocation, exclusive deduction, and rollback of remaining balance is associated with a commitment identifier, disturbance identifier, account version number, current available remaining balance before processing, amount of remaining balance change, current available remaining balance after processing, and processing time.

[0129] In this embodiment, step six specifically includes:

[0130] During the execution of conditional scheduling commitments, the actual start time of the process, the actual end time of the process, the actual release time of locked resources, and the status of the pre-fulfillment conditions are obtained.

[0131] If the preconditions for fulfillment are not met at the start of the execution time window, or if the locked resources are not yet available at the start of the execution time window, the corresponding conditional scheduling commitment is converted to a violated state. After all preconditions for fulfillment are met and all locked resources are available, the time difference between the moment when the process has the actual execution conditions and the start of the execution time window is determined as the root disturbance. Before the actual execution conditions are formed, the root disturbance is updated according to the time difference between the current moment and the start of the execution time window.

[0132] The latest end time is determined based on the end time of the execution time window and the current available margin in the elastic margin account; when the actual end time of the process is later than the latest end time, the time difference between the actual end time and the latest end time is determined as the root disturbance, and the corresponding conditional scheduling commitment is converted into a violation state.

[0133] Using the conditional scheduling commitment that violates the state as the root node of the disturbance, the root node of the disturbance is added to the disturbance influence cone, and the root disturbance quantity is propagated along the successive direction of the process pre- and post-process relationship and the resource successive occupation relationship; each time the remaining disturbance is propagated, the pre-process conditional scheduling commitment, the post-process conditional scheduling commitment, the remaining disturbance quantity and the arrival time of the remaining disturbance are recorded to generate the disturbance propagation relationship;

[0134] After receiving a non-zero time disturbance, the subsequent conditional scheduling commitment adds the received time disturbance to the disturbance impact cone and calls the corresponding elastic margin account to perform margin pre-occupancy and exclusive deduction according to step five. When the remaining disturbance after exclusive deduction is zero, the current propagation branch is terminated. When the remaining disturbance is greater than zero, it continues to propagate along the successor direction of the process pre- and post-process relationship and the resource successor occupation relationship. When all propagation branches terminate or the remaining disturbance reaches the end of the path where there is no successor conditional scheduling commitment, the disturbance impact cone is generated.

[0135] The boundary anchor point determination process for the disturbance impact cone includes: determining the first locked conditional scheduling commitment that is directly adjacent to the conditional scheduling commitments outside and inside the disturbance impact cone and has not received a non-zero time disturbance, along the disturbance propagation relationship, and determining the corresponding conditional scheduling commitment as the boundary anchor point; keeping the locked resources and execution time window of the boundary anchor point unchanged;

[0136] Unaffected conditional scheduling commitments outside the disturbance impact cone and not belonging to the boundary anchor point are converted to a frozen state. Conditional scheduling commitments within the disturbance impact cone that are in a violated state, a provisional state, or a locked state are taken as local recovery objects. The corresponding initial resource fulfillment domain and joint resource fulfillment domain are regenerated according to the current workshop state. The updated joint resource fulfillment domain and boundary anchor point constraints are input into the improved dynamic neighborhood search algorithm to perform time window shifting, resource replacement, and reordering to generate local recovery candidate scheduling schemes.

[0137] The partial recovery candidate scheduling scheme must not change the locked resources and execution time window of the boundary anchor point, and shall perform the fulfillment constraint verification and scheme evaluation in step four; when there is a partial recovery candidate scheduling scheme that passes the fulfillment constraint verification, the partial recovery candidate scheduling scheme with the smallest evaluation value shall be determined as the partial recovery scheduling scheme;

[0138] If there is no candidate scheduling scheme for local recovery that has passed the constraint verification within the disturbance impact cone, the boundary anchor point that caused the time boundary conflict or resource occupation conflict is determined, the corresponding boundary anchor point is added to the disturbance impact cone, and the boundary anchor point is re-determined along the successive direction of the process pre- and post-process relationship and the successive resource occupation relationship, and local recovery is performed again.

[0139] In this embodiment, step seven specifically includes:

[0140] After the conditional scheduling commitment enters a violation state, the planned release time, reliable release time, actual release time, occurrence time of abnormal resource status, upstream remaining disturbance amount, and arrival time of upstream remaining disturbance are obtained based on the disturbance propagation relationship.

[0141] The abnormal resource is subjected to a determination process, which specifically includes: calculating the delay between the actual release time of the resource and the trusted release time, and determining the resource whose actual release time is later than the trusted release time and whose delay reaches the preset failure deviation threshold as an abnormal resource.

[0142] The process of determining source failure for abnormal resources includes: when an abnormal resource does not receive an upstream residual disturbance, the corresponding abnormal resource is identified as a candidate source failure resource; when an abnormal resource receives an upstream residual disturbance, the time when the abnormal state of the resource occurs is compared with the time when the earliest upstream residual disturbance arrives, and the abnormal resource whose abnormal state occurs earlier than the time when the earliest upstream residual disturbance arrives is identified as a candidate source failure resource; from the candidate source failure resources, resources that do not have a preceding candidate source failure resource in the resource dependency subgraph are selected, and the candidate source failure resource whose abnormal state occurs earliest is identified as the source failure resource.

[0143] The determination process for passively affected resources includes: comparing the upstream remaining disturbance amounts that arrive at the same abnormal resource and determining the maximum value as the maximum upstream remaining disturbance amount; if the upstream remaining disturbance has been received before the abnormal state of the resource occurs and the actual release delay of the resource is not greater than the maximum upstream remaining disturbance amount, the corresponding abnormal resource is determined as a passively affected resource.

[0144] The process of updating performance parameters for failed resources includes: calculating the time difference between the actual release time and the planned release time of the failed resource, generating the current performance residual, adding the current performance residual to the performance information, and re-determining the historical performance residual based on a preset historical statistics window; calculating the time difference between the actual release time and the credible release time of the failed resource, generating a calibration deviation, multiplying the calibration deviation by a preset calibration learning rate, and adding it to the previous performance calibration amount to generate the updated performance calibration amount; comparing the actual release time and the credible release time of the failed resource, determining successful performance if the actual release time is not later than the credible release time, and determining failed performance if the actual release time is later than the credible release time, adding the current performance result to the performance information, and updating the performance confidence based on the ratio of the updated number of successful performance records to the number of valid performance records;

[0145] For passively affected resources, the release delay caused by the remaining disturbance upstream is marked as a propagation delay, and the realization calibration amount is not updated using the propagation delay;

[0146] Specifically, the preset failure deviation threshold is set to 60 seconds, the preset historical statistics window is set to the 30 most recent valid performance records, and the preset calibration learning rate is set to 0.20.

[0147] The updated historical performance residuals, fulfillment calibration amounts, and fulfillment confidence are combined to form performance parameters and written back to the corresponding source failure resource agent. After the next scheduling cycle begins, each resource agent, in conjunction with the latest workshop status and performance parameters, re-executes resource release calibration and resource availability time window determination, generates a new initial resource fulfillment domain, and continues to execute release time dependency propagation, time window closure, conditional scheduling commitment generation, improved dynamic neighborhood search, elastic margin account management, disturbance impact cone generation, and local recovery, forming a closed loop for manufacturing workshop production scheduling.

[0148] refer to Figure 3 A manufacturing workshop production scheduling system based on multi-agent collaboration includes the following modules:

[0149] The resource fulfillment module is used to obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process agents and resource agents, and generate the initial resource fulfillment domain;

[0150] The joint realization module is used to construct resource dependency subgraphs, perform dependency propagation at release time and time window closure, and generate joint resource realization domains;

[0151] The conditional commitment module is used to apply for temporary resource tokens, lock resources after the execution time window and pre-fulfillment conditions are verified, determine the allowable elasticity margin, and generate conditional scheduling commitments.

[0152] The scheduling optimization module is used to perform time window shifting, resource replacement and reordering through an improved dynamic neighborhood search algorithm, and to determine the dynamic neighborhood based on the remaining disturbances to generate the target scheduling scheme.

[0153] The surplus management module is used to create flexible surplus accounts and perform surplus pre-occupancy, exclusive deduction, surplus disturbance transfer, and surplus rollback.

[0154] The local recovery module is used to generate a disturbance impact cone, determine boundary anchor points, and perform local recovery when the conditional scheduling commitment fails.

[0155] The fulfillment update module is used to identify the source failed resource, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource.

[0156] Example 1: To verify the feasibility of this invention in practice, it was applied to the daily production scheduling of a precision parts manufacturing workshop. The workshop had 10 processing machines, 4 operators, and 6 sets of replaceable tooling. There were 36 test orders, corresponding to 114 processes. Each process involved at least two types of resources: equipment, personnel, tooling, and materials. The test cycle was 12 hours. Three types of disturbances were introduced during the test: an 18-minute delay in equipment release, a 12-minute delay in personnel arrival, and a 25-minute delay in material availability. The scheduling cycle was repeated 20 times under the same order, resource status, and disturbance times, and the average result was statistically analyzed.

[0157] The process agent establishes candidate associations with resource agents that meet the process resource requirements and capability specifications. The resource fulfillment statistics window is set to the 30 most recent valid fulfillment records, the calibration window is set to the 10 most recent records, and the sample threshold is set to 10 records. The reliable release time is determined based on historical fulfillment residuals and fulfillment calibration amounts, and a joint resource fulfillment domain is generated through release time dependency propagation and time window closure in the resource dependency subgraph. The validity period of the temporary resource token is set to 5 seconds, and a conditional scheduling commitment is formed after verification through a common execution time window and pre-fulfillment conditions.

[0158] The improved dynamic neighborhood search algorithm has a scheduling time granularity of 1 minute. The evaluation weights for cumulative delay time, conditional scheduling commitment adjustment quantity, trial surplus consumption, and resource fulfillment risk are set to 0.35, 0.30, 0.20, and 0.15, respectively, with a maximum search round of 100 rounds. During the execution phase, elastic surplus will be allowed to be converted into elastic surplus accounts. Once the same surplus has been pre-occupied and exclusively deducted, it will no longer participate in other disturbance consumption. Remaining disturbances propagate along the successive direction of the process's preceding and following relationships and the subsequent resource occupancy relationships, thus forming a disturbance influence cone. The failure deviation threshold is set to 60 seconds, and the performance parameter calibration learning rate is set to 0.20.

[0159] Three comparison methods are set up. Comparison method A uses rule-based scheduling that combines delivery time priority and resource availability priority, and postpones the affected processes after an anomaly occurs; comparison method B uses multi-agent reinforcement learning to complete equipment and process matching, and regenerates the scheduling results after disturbance; comparison method C uses the unimproved dynamic neighborhood search algorithm, using the same time window movement, resource replacement and reordering operations, but does not set conditional scheduling commitment screening, trial surplus exclusive consumption and joint resource fulfillment constraints.

[0160] Table 1 Test Results of Different Production Scheduling Methods

[0161] This invention 95.6 1.6 8.4 12.8 87.4 Comparison Method A 84.1 6.9 27.4 18.9 77.6 Comparison Method B 91.6 4.0 16.2 28.5 85.8 Comparison Method C 92.8 3.1 11.5 21.7 84.2

[0162] Test results show that Comparison Method A, using fixed rules, has relatively limited impact on the original schedule, with an average rescheduling rate of 18.9%, lower than Comparison Methods B and C. However, even when equipment, personnel, and tooling meet availability conditions, they may still not be available simultaneously within the same time window, resulting in a resource time conflict rate of 6.9%. After equipment release delays, the process mainly relies on postponement, with an average recovery time of 27.4 minutes. Comparison Method B can reallocate resources based on workshop status, achieving a comprehensive resource utilization rate of 85.8%, the highest among the three methods. However, re-decision-making after disturbances involves many processes, with an average rescheduling rate of 28.5%. Comparison Method C limits global rescheduling through neighborhood search, reducing the average recovery time to 11.5 minutes. However, fixed neighborhoods and static time margins cannot reflect the occupation of the same margin by multiple disturbances, and the resource time conflict rate remains at 3.1%.

[0163] This invention achieves an on-time completion rate of 95.6%, an improvement of 2.8 percentage points compared to comparative method C. Simultaneously, the resource time conflict rate decreases from 3.1% to 1.6%, a reduction of approximately 48.4%. The joint resource fulfillment domain performs closed-loop verification of the shared execution time of multiple resources before the formation of conditional scheduling commitments, reducing scheduling results where individual resources are available but the overall process is unexecutable. After allowing elastic margins to participate in neighborhood trial calculations, the average reordering ratio of this invention is only 12.8%, a reduction of 8.9 percentage points compared to comparative method C. During the execution phase, the time margin is pre-occupied and exclusively deducted through the elastic margin account, ensuring that remaining disturbances accurately reflect unabsorbed delays, and generating a disturbance influence cone accordingly, reducing the average disturbance recovery time to 8.4 minutes.

[0164] Meanwhile, the overall resource utilization rate of this invention reaches 87.4%, which is still higher than the 85.8% of the comparative method B even with a reduced scheduling adjustment range. This indicates that the boundary anchor point constraint on local recovery does not come at the cost of a large amount of idle resources. The fulfillment parameters of the source failed resources are used to regenerate the initial resource fulfillment domain in subsequent scheduling cycles, so that the reliable release state of resources is continuously calibrated according to the actual fulfillment results. This achieves a closed loop between joint resource fulfillment, conditional scheduling commitment, exclusive consumption of elastic margin, local disturbance recovery, and fulfillment parameter updates, thereby improving the executability and stability of dynamic production scheduling in the manufacturing workshop.

[0165] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A manufacturing workshop production scheduling method based on multi-agent collaboration, characterized in that, Includes the following steps: Step 1: Obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process Agent and resource Agent, and generate the initial resource fulfillment domain; Step 2: Construct a resource dependency subgraph based on process resource requirements and resource timing dependencies, perform release time dependency propagation and time window closure on the initial resource fulfillment domain, and generate a joint resource fulfillment domain; Step 3: The process agent applies for a temporary resource token based on the joint resource redemption domain, locks the resource after verifying the execution time window and pre-redemption conditions, determines the allowable elasticity margin, and generates a conditional scheduling commitment. Step 4: Input the conditional scheduling commitment and joint resource fulfillment domain into the improved dynamic neighborhood search algorithm, filter the commitments to be adjusted according to the commitment status, perform time window shifting, resource replacement and reordering, expand the search range according to the remaining disturbance after the time disturbance is absorbed by the allowable elastic margin, and generate the target scheduling scheme after fulfillment constraint verification. Step 5: Establish an elastic margin account based on the target scheduling scheme, and perform margin pre-occupancy, exclusive deduction, remaining disturbance transfer, and margin rollback for time disturbances; Step 6: When the conditional scheduling commitment fails, generate a disturbance influence cone based on disturbance propagation and elastic margin consumption, determine the boundary anchor point, and call the improved dynamic neighborhood search algorithm to perform local recovery; Step 7: Determine the source failure resource based on the disturbance propagation relationship, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource based on the update results.

2. The manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step one specifically includes: Create a process agent based on the process identifier, create a resource agent based on the resource identifier, and create candidate associations between process agents and resource agents based on the process resource requirements; The performance residual between the actual release time of resources and the planned release time of resources is calculated based on the performance information, and the historical performance residual and the realization calibration amount are determined based on the preset historical statistics window. The reliable release time is determined based on the resource plan release time, the current occupied process plan end time, historical performance residuals, and fulfillment calibration quantities. The available time window for resources is determined based on the trusted release time, resource work calendar, locked execution time window, and resource running status, and the fulfillment confidence level is determined based on historical fulfillment results; By associating the trusted release time, resource availability time window, historical performance residuals, fulfillment calibration amount, and fulfillment confidence with the resource agent, an initial resource fulfillment domain is generated.

3. The manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step two specifically includes: Based on the resource requirements of the process, resource nodes are determined. According to the execution sequence of resource preparation, type change, handling, installation, arrival, kitting, material loading and process handover, directed dependency edges are established between resource nodes with sequential dependencies and the duration of dependency actions is configured to generate a resource dependency subgraph. The release time dependency propagation is performed according to the topological order of the resource dependency subgraph. The sum of the trusted release time of the preceding resource and the duration of the dependent action is compared with the trusted release time of the subsequent resource, and the latest time is determined as the trusted release time of the subsequent resource after correction. Based on the corrected reliable release time, the available resource time window is extracted. The intersection calculation is performed on the available resource time windows of the resources required to complete the same process. The intersection interval with a duration shorter than the process processing time is deleted, and a joint available time window is generated. The minimum value among the realization confidence of each resource participating in the time window closure is determined as the joint realization confidence, and it is combined with the corrected reliable release time of each resource and the joint available time window to generate the joint resource realization domain.

4. A manufacturing workshop production scheduling method based on multi-agent collaboration as described in claim 1, characterized in that, Step three specifically includes: The process agent determines the candidate execution time window based on the combined available time window, generates a pre-reservation transaction identifier, and sends a resource pre-reservation request to the corresponding resource agent. The resource agent verifies the occupancy relationship between the candidate execution time window and the available resource time window, the locked execution time window, and the valid temporary reserved time window. After the verification is successful, a temporary resource token is generated. The pre-emption conditions are determined based on the completion status of the direct preceding process and the inclusion relationship between the candidate execution time window and the joint resource redemption domain. Consistency checks are performed on temporary resource tokens associated with the same pre-occupied transaction identifier. After the consistency check passes, a commitment identifier is generated. The commitment identifier is associated with a temporary resource token and the temporary resource token and the candidate execution time window are converted to a locked state. The allowable elasticity margin is determined based on the time interval between the locked execution time window and the joint available time window, the direct subsequent process and the next conditional scheduling commitment on the same locked resource, and a conditional scheduling commitment is generated. If any resource pre-allocation fails or consistency check fails, the corresponding temporary resource token and temporary reserved time window will be released.

5. A manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step four specifically includes: The execution order of the corresponding processes is determined based on the conditional scheduling commitment. The subsequent resource occupancy relationship is determined based on the conditional scheduling commitment and execution order of the shared resources. The conditional scheduling commitment, the joint resource fulfillment domain, the process precedence relationship, and the resource subsequent occupancy relationship are input into the improved dynamic neighborhood search algorithm. Based on the commitment status and joint resource fulfillment domain, commitments to be adjusted are screened, and the execution time window of the commitments to be adjusted is moved, resources are replaced, and the order is changed. Among them, the resource replacement selects alternative resources from the resource agents that have established candidate associations with the corresponding process agents, and generates candidate scheduling schemes. The dynamic neighborhood is determined based on the time disturbance, allowable elastic margin and remaining disturbance generated by the candidate scheduling scheme, and the scheduling commitment execution fulfillment constraint verification is performed based on the joint resource fulfillment domain, pre-fulfillment conditions, locked execution time window and frozen state conditions. The resource fulfillment risk is determined based on the joint fulfillment confidence level. The evaluation value of the candidate scheme is determined based on the cumulative delay time, the number of conditional scheduling commitment adjustments, the trial surplus consumption, and the resource fulfillment risk. The neighborhood search is iteratively executed to generate the target scheduling scheme.

6. A manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, The process of determining the dynamic neighborhood based on the time disturbance, allowable elastic margin, and remaining disturbance generated by the candidate scheduling scheme specifically includes: The positive time offset generated after the adjustment of the execution time window of the commitment to be adjusted is determined as the initial time disturbance, and the allowable elasticity margin participating in the current candidate scheduling scheme is copied as the initial trial margin. The time disturbance is propagated along the successive direction of the process sequence and the successive resource occupancy relationship. After receiving a non-zero time disturbance, the conditional scheduling commitment adds it to the dynamic neighborhood, determines the smaller value between the time disturbance and the remaining trial balance as the trial consumption, and performs the deduction. The current propagation branch terminates when the remaining perturbation after consumption is zero; propagation continues when the remaining perturbation is greater than zero. When multiple propagation paths arrive at the same conditional scheduling commitment, the larger of the recorded cumulative disturbance and the newly arrived disturbance will be determined as the updated cumulative disturbance, and only the newly added disturbance will be subject to trial balance deduction. After all propagation branches terminate, the conditional scheduling commitments added to the dynamic neighborhood are determined as the search scope of the current candidate scheduling scheme.

7. A manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step five specifically includes: Update the execution time window of the conditional scheduling commitment according to the target scheduling scheme, redetermine the allowable elasticity margin, and establish an elasticity margin account for the locked conditional scheduling commitment. The elasticity margin account includes the currently available margin, the reserved margin, the consumed margin, and the account version number. After the time disturbance arrives and the scheduling commitment is made, the request quantity is determined based on the time disturbance amount and the current available balance. After the account version number consistency check is performed, the request quantity is deducted from the current available balance and written into the pre-occupied balance. After the remaining balance is reserved, an exclusive deduction is performed, transferring the reserved balance to the consumed balance, so that the reserved balance and the consumed balance are no longer included in the current available balance; The time amount remaining after deducting the actual exclusive deduction from the time disturbance amount is determined as the residual disturbance and continues to be propagated. When the actual end time of a process is earlier than the expected end time after considering the current disturbance, the time difference between the two is determined as the rollback margin. After resource occupation and margin pre-occupancy verification, the smaller value between the rollback margin and the consumed margin is transferred back to the current available margin.

8. A manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step six specifically includes: If the preconditions for execution are not met at the start of execution or the locked resources are not yet available, the time difference between the moment when the actual execution conditions are met and the start of the execution time window is determined as the root disturbance. When the actual execution of a process exceeds the latest end time jointly defined by the end time of the execution time window and the current available margin, the time difference between the actual end time and the latest end time is determined as the root disturbance, and the corresponding conditional scheduling commitment is converted into a violation state. The conditional scheduling commitment that violates the state is used as the root node of the disturbance and added to the disturbance influence cone. The root disturbance quantity is propagated along the successive direction of the process pre- and post-process relationship and the successive resource occupation relationship. The disturbance propagation relationship and the arrival time of the remaining disturbance are recorded. After receiving a non-zero time disturbance, the conditional scheduling commitment adds it to the disturbance impact cone and calls the elastic margin account to perform margin pre-occupancy and exclusive deduction, so that the remaining disturbance after exclusive deduction controls the termination or continuation of the propagation. The first locked state conditional scheduling commitment that is directly adjacent to the disturbance influence cone and has not received a non-zero time disturbance is determined as the boundary anchor point. The locking resources and execution time window of the boundary anchor point remain unchanged, and the improved dynamic neighborhood search algorithm is called to perform local recovery of the conditional scheduling commitment inside the disturbance influence cone. If no candidate scheduling scheme passes the constraint verification, the boundary anchor point that causes the constraint conflict will be added to the disturbance influence cone and the boundary anchor point will be redefined.

9. A manufacturing workshop production scheduling method based on multi-agent collaboration according to claim 1, characterized in that, Step seven specifically includes: Resources whose actual release time is later than the trusted release time and whose release delay reaches a preset failure deviation threshold are identified as abnormal resources. Abnormal resources that have not received upstream residual disturbances or whose abnormal resource state occurs earlier than the earliest arrival time of upstream residual disturbances are identified as candidate source failure resources. From the candidate source failure resources, resources that do not have a preceding candidate source failure resource in the resource dependency subgraph are selected, and the candidate source failure resource with the earliest occurrence time of abnormal state is identified as the source failure resource. Resources that have received upstream residual disturbances before the occurrence of abnormal resource status and whose actual release delay is no greater than the maximum upstream residual disturbance are identified as passively affected resources. The historical performance residual is updated based on the time difference between the actual release time of the source failure resource and the planned release time of the resource. The fulfillment calibration amount is updated based on the calibration deviation between the actual release time and the credible release time and the preset calibration learning rate. The fulfillment confidence is updated based on the actual fulfillment result. The fulfillment calibration amount is not updated based on the release delay caused by the upstream residual disturbance for passively affected resources. The updated historical performance residuals, performance calibration values, and performance confidence are used to form performance parameters, and the initial resource performance domain for the next scheduling cycle is regenerated using these performance parameters.

10. A manufacturing workshop production scheduling system based on multi-agent collaboration, applied to the manufacturing workshop production scheduling method based on multi-agent collaboration as described in any one of claims 1-9, characterized in that, Includes the following modules: The resource fulfillment module is used to obtain workshop status, process resource requirements, resource timing dependencies and fulfillment information, establish process agents and resource agents, and generate the initial resource fulfillment domain; The joint realization module is used to construct resource dependency subgraphs, perform dependency propagation at release time and time window closure, and generate joint resource realization domains; The conditional commitment module is used to apply for temporary resource tokens, lock resources after the execution time window and pre-fulfillment conditions are verified, determine the allowable elasticity margin, and generate conditional scheduling commitments. The scheduling optimization module is used to perform time window shifting, resource replacement and reordering through an improved dynamic neighborhood search algorithm, and to determine the dynamic neighborhood based on the remaining disturbances to generate the target scheduling scheme. The surplus management module is used to create flexible surplus accounts and perform surplus pre-occupancy, exclusive deduction, surplus disturbance transfer, and surplus rollback. The local recovery module is used to generate a disturbance impact cone, determine boundary anchor points, and perform local recovery when the conditional scheduling commitment fails. The fulfillment update module is used to identify the source failed resource, update the fulfillment parameters, and regenerate the initial fulfillment domain of the resource.

Images