Helicopter component assembly scheduling method and system based on large language model

Abstract

The application discloses a helicopter component assembly scheduling method and system based on a large language model. The system comprises a unit state acquisition module, a task data acquisition module, an allocation decision module, a verification and rollback module, a local observation construction module, a graph structure scheduling module, a scheduling item generation and rolling revision module, a reward generation and learning update module, and a scheduling result output module. The application expresses the scheduling item field and updates the resource calendar occupation information synchronously. The output scheduling result meets the field hard constraint and can be directly issued for execution. At the same time, the benchmark-tolerance-gating-measurement closed loop is included in the scheduling constraint, which can improve the scheduling executability and assembly quality stability, reduce rework and resource conflict. The application triggers rework insertion and rolling revision of the scheduling scheme dataset when the measurement is out of tolerance, so that the plan can be adjusted in time according to the field state change, and the delivery stability is improved. The resource conflict of the method is significantly reduced, and the system has strong scalability and reusability.

Description

Technical Field

[0001] This invention belongs to the field of aviation equipment manufacturing and intelligent scheduling, specifically involving a helicopter component assembly scheduling method and system based on a large language model. Background Technology

[0002] The assembly of critical helicopter components is typically a complex discrete manufacturing process involving multiple workstations, equipment, fixtures, and measurement resources. This is especially true in the docking scenarios of large components such as the fuselage section, main rotor hub, tail rotor hub, and landing gear. The assembly process is characterized by long assembly reference chains, strong spatial accessibility constraints, numerous quality gate points, and frequent measurement and rework. Existing assembly scheduling schemes can be mainly divided into three categories: rule-driven scheduling, heuristic / optimization algorithm scheduling, and data-driven scheduling. However, the following problems still exist in the helicopter assembly scenario:

[0003] 1) Insufficient consideration of assembly datum, tolerance grades, and quality gate constraints.

[0004] Traditional production scheduling often focuses on core constraints such as process time, resource availability, and delivery date. It frequently fails to incorporate quality-related information, such as assembly datum binding status, critical quality control point gating status, and measurement completion / out-of-tolerance indicators, into the scheduling decision-making process in a structured manner. Because helicopter assembly is highly dependent on docking datum surfaces, datum holes / edges, and the set of docking feature points, if scheduling does not explicitly consider these datum and gating statuses, processes may be assigned to resource combinations that do not meet datum conditions or fail gating, leading to unexecuted operations or increased quality risks after execution.

[0005] 2) The descriptions of the window occupancy and resource calendar conflict constraints for tooling fixtures and measuring equipment are coarse.

[0006] Current production scheduling often simplifies resources into a binary "available / unavailable" state in resource modeling, lacking a detailed depiction of "resource calendar occupancy information" such as fixture assembly / disassembly, shared measuring equipment, measurement window scheduling, and equipment maintenance / shifts. Helicopter assembly often involves multiple tasks competing for the same tooling fixture or measuring equipment. If the resource calendar is not updated synchronously during scheduling and conflicts are not masked or avoided, the scheduling results are prone to resource contention during the execution phase, leading to frequent plan adjustments.

[0007] 3) Lack of a rolling revision mechanism for measurement closure and rework insertion.

[0008] In actual assembly, after a critical quality control point process is executed, measuring equipment is usually needed to measure the set of docking feature points or the corresponding features of the critical quality control points. If the deviation indicated by the measurement results exceeds the threshold corresponding to the assembly tolerance level, a rework process needs to be generated and inserted into the assembly process set. Then, the process sequence constraints and gating states must be satisfied again before execution can continue. Traditional static scheduling often struggles to consistently update the insertion of rework processes, the reallocation of resource occupancy windows, and the postponement of subsequent processes in the scheduling scheme dataset, leading to problems such as "inconsistency between plan and actual work" or "inability to propagate the impact of rework."

[0009] 4) Uncertainty in the arrival of assembly tasks and differences in assembly unit capabilities lead to unstable allocation.

[0010] Helicopter assembly sites are subject to dynamic disturbances such as mission arrival, changes in resource status, and measurement anomalies. Furthermore, different assembly units exhibit varying capabilities in docking, positioning, and fastening, necessitating matching based on unit capability identifiers. Existing solutions either rely on manual experience for unit allocation, offering strong interpretability but lacking consistency and reproducibility; or they depend on fixed rules or single indicators, failing to consider multi-dimensional constraints such as capability matching, load balancing, fixture / measurement availability, and accessibility, easily leading to unstable allocation results or frequent rollbacks.

[0011] 5) Production scheduling results lack a structured expression that can be directly issued and executed.

[0012] Some solutions focus on generating "priority / score" or "next action suggestions" instead of outputting a production scheduling dataset that includes fields such as process identifier, resource identifier, planned start / end time, fixture / measurement window, gate status, and rework flag. This makes it difficult to interface with assembly execution systems or manufacturing execution systems, and also makes it difficult to form a visual Gantt chart production scheduling table and traceable resource calendar occupancy information.

[0013] In recent years, large language models have demonstrated the ability to semantically understand, reason, and generate multi-source textual / structured information, making them suitable for interpreting assembly line conditions and invoking process knowledge. However, directly applying large language models to production scheduling decisions can lead to uncontrolled outputs and inconsistencies with on-site constraints, resulting in unexecutable plans or resource conflicts. Therefore, it is necessary to limit the output of large language models to a preset structured format and establish a feasible production scheduling loop through mechanisms such as format and feasibility verification, rollback allocation, and resource calendar conflict determination.

[0014] Therefore, there is an urgent need for a dynamic process decision-making and scheduling technology for the assembly of key helicopter components. This technology should be able to allocate assembly units when assembly tasks arrive, and generate executable scheduling entries within each assembly unit by combining constraints such as process sequence constraints, assembly datum information, tolerance levels, key quality control point gating, fixture / measuring equipment occupancy windows, and resource calendar conflicts. Simultaneously, when measurement deviations occur, rework processes can be automatically inserted and the scheduling plan dataset can be continuously revised. Finally, an executable scheduling result can be output, thereby improving scheduling executability, reducing plan fluctuations caused by resource conflicts and rework, and enhancing assembly quality and delivery stability. Summary of the Invention

[0015] To address the problems in existing technologies, such as the difficulty in incorporating assembly datum, tolerance levels, and key quality control points into scheduling constraints; the difficulty in uniformly modeling conflicts between tooling fixtures and measuring equipment occupancy windows and resource calendars; the difficulty in consistently revising scheduling plans after measurement closure-loop triggers rework; the instability in allocation due to differences in assembly unit capabilities; and the difficulty in forming structured data that can be directly deployed and executed from scheduling results, this invention provides a helicopter component assembly scheduling method and system based on a large language model. The specific technical solution is as follows:

[0016] A helicopter component assembly scheduling system based on a large language model includes a unit status acquisition module, a mission data acquisition module, and:

[0017] The allocation decision module is used to construct a structured assembly context based on the collected unit status data and task data, input it into a semantic reasoning model based on a large language model, and output the assembly unit allocation output.

[0018] The verification and rollback module is used to perform constraint verification on the output of the assembly unit allocation. If the verification fails, the rollback allocation is performed to determine the target assembly unit.

[0019] The local observation construction module is used to determine the set of ready processes within the target assembly unit based on process sequence constraints, and to construct a local observation tensor and a corresponding validity mask based on the set of ready processes, resource status and resource calendar occupancy information.

[0020] The graph structure scheduling module is used to construct a bipartite graph of "process node - resource node". It inputs the local observation tensor, the bipartite graph and the validity mask into the graph structure policy network, and outputs the action probability distribution of each feasible process-resource pair; and determines the target action under the constraint of the validity mask; at the same time, it determines the planned start time and planned end time of the target process according to the constraint conditions.

[0021] The production scheduling item generation and rolling revision module is used to generate production scheduling items based on target processes, target resources, planned start time and planned end time, and write them into the production scheduling scheme dataset, and synchronously update resource calendar occupancy information.

[0022] The reward generation and learning update module is used to calculate intermediate evaluation vectors for candidate actions and their corresponding candidate successor assembly states using an executable evaluation function, and then map them into step rewards through a proxy reward model.

[0023] The production scheduling result output module is used to output the production scheduling result when the preset output conditions are met.

[0024] A helicopter component assembly scheduling method based on a large language model includes:

[0025] S1: Obtain the unit status data of each assembly unit;

[0026] S2: Obtain task data for newly arrived assembly tasks;

[0027] S3: Construct a structured assembly context based on unit state data and task data, input a semantic reasoning model based on a large language model, and output the assembly unit allocation output.

[0028] S4: Perform constraint verification on the output of the assembly unit allocation. If it fails, perform a rollback allocation to determine the target assembly unit.

[0029] S5: Within the target assembly unit, determine the set of ready processes based on the process sequence constraints, and construct a local observation tensor and the corresponding validity mask based on the set of ready processes, resource status and resource calendar occupancy information;

[0030] S6: Construct a bipartite graph of "process node - resource node", and input the local observation tensor, bipartite graph, and validity mask into the graph structure strategy network to output the action probability distribution of each feasible process-resource pair; under the constraint of validity mask, determine the target action; and determine the planned start time and planned end time of the target process according to the constraint conditions.

[0031] S7: Based on the target process, target resources, the planned start time and planned end time of the target process, generate production scheduling items, write them into the production scheduling dataset, and synchronously update the resource calendar occupancy information;

[0032] S8: Output production scheduling results when preset output conditions are met.

[0033] The beneficial effects of this invention are as follows:

[0034] (1) The production scheduling results are more executable: This invention expresses the production scheduling items in a field and updates the resource calendar occupancy information synchronously. The output production scheduling results can be directly used to generate work instructions or Gantt chart production schedules, reducing the risk of "plans being difficult to implement". By performing structured constraints and verification rollback on the output of the large language model, it utilizes its semantic reasoning ability and ensures that the production scheduling results meet the hard constraints on site and can be directly issued for execution.

[0035] (2) Quality constraints are incorporated into the scheduling closed loop: This invention integrates the assembly benchmark binding status, key quality control point gating status, and measurement completion / out-of-tolerance flags into local observation and validity mask, so that production scheduling naturally follows quality gating and measurement constraints, reducing unexecutable production scheduling and quality rework.

[0036] (3) Strong adaptability to dynamic disturbances: The present invention measures out-of-tolerance triggers rework insertion and rolling revision of the production scheduling data set, so that the plan can be adjusted in a timely manner according to changes in the on-site status, thereby improving delivery stability.

[0037] (4) Resource conflicts are significantly reduced: The present invention explicitly introduces the fixture / measuring equipment occupancy window and resource calendar conflict constraints in the scheduling decision, and implements action shielding for infeasible combinations, thereby reducing the waiting time caused by fixture contention, measurement queuing and resource conflicts.

[0038] (5) More stable allocation of assembly units: The allocation output of this invention adopts structured fields and undergoes dual verification of format and feasibility. When the verification fails, the allocation is rolled back through the cost function, which reduces the allocation fluctuation caused by the instability of the model output.

[0039] (6) Scalability and reusability: Without changing the unified data structure of production scheduling items and resource calendar, the present invention can be extended to assembly scenarios with different key components, different numbers of assembly units and different resource calendar rules, which facilitates engineering deployment and maintenance. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a helicopter component assembly scheduling system based on a large language model, according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the execution logic of the graph structure scheduling module in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram illustrating the execution logic of the production scheduling entry generation and rolling revision module in an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the execution logic of the reward generation and learning update module in an embodiment of the present invention.

[0044] Figure 5This is a flowchart of a helicopter component assembly scheduling method based on a large language model, which is an embodiment of the invention. Detailed Implementation

[0045] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0046] This embodiment uses the assembly of key helicopter components as an application scenario. These key components include the fuselage section, main rotor hub, tail rotor hub, and landing gear. Two assembly units (preferably at least two, but can also be a single assembly unit) are set up at the assembly site, denoted as Assembly Unit U1 and Assembly Unit U2. Each assembly unit includes an assembly station, assembly equipment, tooling fixtures, and measuring equipment, and possesses resource calendar management capabilities.

[0047] On the one hand, such as Figure 1 As shown, the helicopter component assembly scheduling system based on a large language model of the present invention includes at least a processor, a memory, and program instructions executed by the processor, and implements the following functional modules: unit status acquisition module, task data acquisition module, allocation decision module, verification and rollback module, local observation construction module, graph structure scheduling module, scheduling item generation and rolling revision module, reward generation and learning update module, and scheduling result output module.

[0048] 1. Unit Status Acquisition Module

[0049] The unit status acquisition module is used to obtain unit status data for each assembly unit from the assembly execution system or resource management system. This unit status data includes: the resource availability status of the assembly unit, the in-process task queue, the earliest available time for each resource, resource calendar occupancy information, and capability identifiers.

[0050] The resource availability status includes the occupancy status of assembly stations, the availability status of assembly equipment, the availability status of tooling fixtures, and the availability status of measuring equipment. Capability identifiers characterize the assembly unit's ability to dock, position, and fasten critical helicopter components. Resource calendar occupancy information records the occupancy intervals of each assembly station, assembly equipment, tooling fixture, and measuring equipment on the timeline. Resource calendar occupancy information includes at least the resource identifier, occupancy start time, occupancy end time, and occupancy reason code to support subsequent conflict detection and rolling revisions.

[0051] 2. Task Data Acquisition Module

[0052] The task data acquisition module is used to read task data from the process database or manufacturing execution system. The task data includes task identifier, key component type, assembly process set, process sequence constraints, assembly datum information, assembly tolerance level, delivery date, set of key quality control points, and standard or estimated process time; the assembly datum information includes mating datum surface, datum hole, datum edge, and set of mating feature points.

[0053] 3. Allocation Decision Module

[0054] The allocation decision module is used to construct a structured assembly context based on unit state data and task data, input it into a semantic reasoning model based on a large language model, and output the assembly unit allocation output.

[0055] The structured assembly context includes:

[0056] a) Field constraint section, used to limit the set of output fields, field types, value ranges, and required / optional field rules;

[0057] b) Unit Status Section, used to describe the resource availability status, earliest available time, resource calendar occupancy information and capability identifier of each assembly unit in key-value pairs;

[0058] c) Task data segment, used to describe key component types, assembly tolerance levels, delivery dates, assembly datum information, key quality control points, process sequence constraints, and rework insertion rules after measurement deviations in key-value pairs;

[0059] d) Process knowledge segment, which is retrieved from the assembly process knowledge base; the assembly process knowledge base includes assembly sequence constraints, assembly datum selection rules, gating rules for key quality control points, and measurement deviation handling rules related to key helicopter components.

[0060] The assembly unit allocation output is limited to a preset structured format.

[0061] In one implementation, the assembly unit allocation output satisfies the following JSON structure: { "unit_id": unit identifier, value is an integer; "confidence": confidence level, value is a real number in the range [0,1], "reason_code": indicates the reason for allocation to this unit, which is an enumeration code}. The reason_code is selected from a preset set of reason codes, which includes at least two or more of the following: capability matching, low load, fixture availability, measurement availability, gate pass, and reachability satisfaction.

[0062] The semantic reasoning model based on the large language model is a pre-trained language model that is constrained by instructions to output a JSON structure.

[0063] The process knowledge segment is retrieved from the assembly process knowledge base through enhanced retrieval and concatenated into the structured assembly context to ensure that the value of reason_code in the JSON structure is consistent with that in the assembly process knowledge base.

[0064] 4. Verification and rollback module

[0065] The verification and rollback module is used to perform constraint verification on the assembly unit allocation output. When the assembly unit allocation output fails the constraint verification, rollback allocation is performed to determine the target assembly unit.

[0066] The constraint verification includes verification of the pre-defined structured format output of the semantic reasoning model based on a large language model, as well as feasibility verification. Feasibility verification includes: verification of the existence of the target assembly unit, verification of the executability of process sequence constraints, verification of assembly unit capability identifier matching, verification of the compatibility between tooling fixtures and measuring equipment, and verification of installation accessibility.

[0067] The rollback assignment selects the target assembly unit based on the cost function, and chooses the candidate assembly unit that minimizes the cost function J(u) as the target assembly unit. The expression for the cost function is as follows:

[0068]

[0069] Where u is the candidate assembly unit identifier; The estimated completion cost is calculated based on the earliest available time, the in-process task queue, and resource calendar occupancy information; Q(u) is the quality cost calculated based on the assembly tolerance level, the gating status of key quality control points, and historical measurement deviations; L(u) is the load cost calculated based on the ratio of occupied time to available time of candidate assembly units; R(u) is the resource conflict cost calculated based on tooling and fixture occupancy conflicts, measuring equipment occupancy conflicts, and capacity identifier mismatch conflicts; α1~α4 are preset weighting coefficients.

[0070] 5. Local Observation Construction Module

[0071] The local observation construction module is used to determine the set of ready processes within the target assembly unit based on process sequence constraints, and to construct a fixed-dimensional local observation tensor and the corresponding validity mask based on the set of ready processes, resource status and resource calendar occupancy information.

[0072] The local observation tensor includes the assembly datum binding status field, the gated status field of key quality control points, the measurement completion field, the out-of-tolerance flag field, the fixture occupied window field, and the measuring equipment occupied window field.

[0073] The local observation tensor is a real matrix of dimension m×d, and the validity mask is a 0 / 1 vector of length m; where u is the target assembly unit identifier, m is the number of resources participating in scheduling within the target assembly unit, and d is the feature dimension.

[0074] When the number of ready operations is greater than m, select the top-m ready operations based on the operation priority score S(·) to construct the local observation tensor; when the number of ready operations is less than m, in addition to selecting all ready operations, fill the local observation tensor and mark invalid positions in the validity mask.

[0075] The formula for calculating the process priority score S(⋅) is as follows:

[0076]

[0077] Where slack(o) is the slack of the delivery date corresponding to the process, w(o) is the task priority weight, g(o) is the key quality control point hit mark, ρ(o) is the rework risk index, and β1~β4 are preset coefficients.

[0078] When different processes have the same score, they are eliminated in parallel according to the order of smaller delivery slack, true hit flag of key quality control points, and higher task priority weight.

[0079] 6. Graph Structure Scheduling Module

[0080] like Figure 2 As shown, the graph structure scheduling module is used to construct a bipartite graph of "process node - resource node". It inputs the local observation tensor, the bipartite graph and the validity mask into the graph structure strategy network, and outputs the action probability distribution of each feasible process-resource pair. Under the constraint of the validity mask, the target action is determined. At the same time, based on the process sequence constraint, the earliest available time of the target resource, the occupancy window of tooling fixtures and measuring equipment and the resource calendar occupancy information, the planned start time and planned end time of the target process are determined.

[0081] In the bipartite graph, process nodes represent ready processes within the current target assembly unit, while resource nodes represent assembly stations, assembly equipment, tooling fixtures, measuring equipment, or combinations thereof that can participate in scheduling. Edges in the bipartite graph represent the feasibility relationship that "resources can execute the process".

[0082] Feasibility relationships are determined by the following constraints: tooling compatibility constraints, measuring equipment availability constraints, process-required tool constraints, process-required capability matching constraints, installation accessibility constraints, critical quality control point gating constraints, and resource calendar occupancy conflict constraints.

[0083] The graph structure scheduling module applies action masking to process-resource combinations that do not meet the above constraints, making the probability of the masked action zero, thereby ensuring that the target action selected subsequently meets the on-site execution constraints.

[0084] The target action includes at least a target process and a target resource, used to indicate the allocation decision of assigning the target process to the target resource. In one implementation, the graph structure scheduling module can directly select the process-resource pair with the highest action probability as the target action; in another implementation, the graph structure scheduling module can also output multiple candidate process-resource pairs, calculate the corresponding candidate plan start time and candidate plan end time to form multiple candidate scheduling entries, and receive the candidate step reward obtained by the reward generation and learning update module from the intermediate evaluation vector mapping of the candidate subsequent assembly state and candidate action corresponding to each candidate scheduling entry, and select the candidate action with the highest candidate step reward as the target action.

[0085] In this way, the step reward is not only used for strategy optimization during the training phase, but also for selecting the best candidate scheduling items during the online scheduling phase.

[0086] The planned start time of the target process is the maximum value of the start time that satisfies constraints including process sequence constraints, the earliest available time of the target resources, the occupancy window of tooling and measuring equipment, and resource calendar occupancy information; the planned end time of the target process is determined by the planned start time of the target process and the standard working time or estimated working time of the process.

[0087] The graph structure scheduling module outputs the planned start time and planned end time of the target action and target process to the production scheduling item generation and rolling revision module, so as to generate production scheduling items and write them into the production scheduling scheme dataset.

[0088] 7. Production scheduling item generation and rolling revision module

[0089] like Figure 3 As shown, the production scheduling item generation and rolling revision module is used to generate production scheduling items based on target processes, target resources, planned start time, and planned end time, and write them into the production scheduling scheme dataset, while synchronously updating resource calendar occupancy information. Specifically, after executing the target process corresponding to the key quality control point, the measuring equipment is triggered to measure the set of docking feature points or the features corresponding to the key quality control point to obtain measurement results. When the deviation represented by the measurement results exceeds the threshold corresponding to the assembly tolerance level, a rework process is generated and inserted into the assembly process set. At the same time, the process sequence constraints and gating status fields are updated so that the rework process enters the ready process set after satisfying the constraints, and the production scheduling scheme dataset is rolled over. Finally, the measurement results are written into the assembly status.

[0090] The production scheduling items include: process identifier, assembly unit identifier, resource identifier, planned start time, planned end time, tooling fixture identifier, measuring equipment identifier or measuring window identifier, gate control status, and rework indicator.

[0091] 8. Reward Generation and Learning Update Module

[0092] like Figure 4 As shown, the reward generation and learning update module uses an executable evaluation function to calculate intermediate evaluation vectors for candidate successor assembly states and candidate actions, and maps these intermediate evaluation vectors to step rewards through a proxy reward model. The step rewards are sent to the graph structure scheduling module to guide policy learning in the graph structure policy network and to help the graph structure scheduling module select candidate actions.

[0093] As one implementation method, the reward generation and learning update module includes: an evaluation function generation and verification correction submodule, a reward mapping submodule, and a policy learning update submodule.

[0094] The evaluation function generation and verification correction submodule is used to provide evaluation dimension definitions, input and output dimension constraints and boundary check rules to the semantic reasoning model based on the large language model, and to provide key quality control point gating rules, measurement closed-loop rules and measurement deviation handling rules related to helicopter assembly, so as to generate an executable definition of the evaluation function.

[0095] Specifically, the evaluation function generation and verification / correction submodule is used to perform runtime verification on the evaluation function using randomly selected assembly state-action pairs in an isolated runtime environment. When the runtime verification fails, a structured error log is recorded and fed back to the semantic reasoning model based on a large language model to generate a corrected evaluation function until the runtime verification passes. The runtime verification includes dimensional consistency verification, index out-of-bounds verification, exception handling, and timeout limit verification. The structured error log includes at least the error type, trigger input summary, and error location information.

[0096] The reward mapping submodule is used to calculate intermediate evaluation vectors for candidate successor assembly states and candidate actions corresponding to candidate actions using a validated executable evaluation function, and then maps the intermediate evaluation vectors to step rewards through a proxy reward model.

[0097] Among them, the intermediate evaluation vector can represent at least one or more of the following: cycle time cost, load cost, quality cost, gate failure penalty, measurement deviation penalty, and rework insertion penalty.

[0098] The agent reward model is trained using a reward alignment loss function. As one implementation, the reward alignment loss function is expressed as:

[0099]

[0100] in, The assembly trajectory is T, where T is the trajectory length. For agent reward model, For model parameters; z t This is the intermediate evaluation vector; Represents the expected assembly trajectory; The trajectory report for the assembly trajectory is obtained by weighting at least two of the following based on the production scheduling dataset: total assembly completion time, pass rate of key quality control points, number of measurement deviations, number of reworks, and rework duration.

[0101] The policy learning update submodule constructs or estimates the advantage function based on the step reward and updates the policy parameters of the graph-structured policy network according to the policy learning objective function of the graph-structured policy network. The updated policy parameters are output to the graph-structured scheduling module, which then recalculates the action probability distribution and determines the target action in subsequent rounds. Thus, the step reward guides the policy learning of the graph-structured policy network.

[0102] As one implementation, the policy parameter of the graph-structured policy network is θ, and it is updated using a policy learning objective function based on the pruning probability ratio. The policy learning objective function is:

[0103]

[0104] in, A is the cutting factor. t The dominant function;

[0105]

[0106] Where, π θ (a t |s t ) indicates the assembly state s t Choose action a t The probability of the strategy; The assembly state before the update s t Choose action a t The strategy probability; the advantage function A t Rewards from steps r t Obtained through construction or estimation.

[0107] 9. Production Scheduling Result Output Module

[0108] The production scheduling result output module is used to output the production scheduling result when the preset output conditions are met, and to send the production scheduling result to the assembly execution system or manufacturing execution system to form assembly operation instructions or Gantt chart production schedule.

[0109] The preset output conditions are reaching the rolling planning time domain boundary, the number of times the production scheduling data set is updated reaches a threshold, or a new assembly task arrives.

[0110] On the other hand, this invention also proposes a helicopter component assembly scheduling method based on a large language model. This method includes allocating assembly units to the assembly task when it arrives, allocating resources for assembly processes within the assembly unit, and outputting the scheduling results. Figure 5 As shown, the specific steps include the following:

[0111] S1: Obtain the unit status data of each assembly unit;

[0112] S2: Obtain task data for newly arrived assembly tasks;

[0113] S3: Construct a structured assembly context based on unit state data and task data, input a semantic reasoning model based on a large language model, and output the assembly unit allocation output.

[0114] S4: Perform constraint verification on the assembly unit allocation output. If the assembly unit allocation output fails the constraint verification, perform a rollback allocation to determine the target assembly unit.

[0115] S5: Within the target assembly unit, determine the set of ready processes based on the process sequence constraints, and construct a fixed-dimensional local observation tensor and the corresponding validity mask based on the set of ready processes, resource status and resource calendar occupancy information;

[0116] S6: Construct a bipartite graph of "process node - resource node", and input the local observation tensor, bipartite graph and validity mask into the graph structure strategy network, output the action probability distribution of each feasible process - resource pair; under the constraint of validity mask, determine the target action; based on the process sequence constraint, the earliest available time of the target resource, the occupancy window of tooling fixtures and measuring equipment, and the resource calendar occupancy information, determine the planned start time and planned end time of the target process;

[0117] In the process of determining the target action, the executable evaluation function is used to calculate the intermediate evaluation vector for the candidate successor assembly state and candidate action corresponding to the candidate action, and the intermediate evaluation vector is mapped to the step reward through the agent reward model; the policy parameters of the policy network are updated according to the step reward guidance graph structure, and the target action is selected according to the step reward.

[0118] S7: Based on the target process, target resources, the planned start time and planned end time of the target process, generate production scheduling items, write them into the production scheduling dataset, and synchronously update the resource calendar occupancy information;

[0119] S8: After executing the target process corresponding to the key quality control point, the measuring equipment is triggered to measure the set of docking feature points or the features corresponding to the key quality control point to obtain the measurement results. When the deviation represented by the measurement results exceeds the threshold corresponding to the assembly tolerance level, a rework process is generated and inserted into the assembly process set. At the same time, the process sequence constraints and gating status fields are updated so that the rework process enters the ready process set after satisfying the constraints. The production scheduling plan dataset is then rolled over and revised. Finally, the measurement results are written to the assembly status.

[0120] S9: When the preset output conditions are met, output the production scheduling results. The production scheduling results include the production scheduling plan dataset and the corresponding resource calendar occupancy information. The production scheduling results are then sent to the assembly execution system or manufacturing execution system to form assembly operation instructions or Gantt chart production schedules.

[0121] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A helicopter component assembly scheduling system based on a large language model, characterized in that, This includes a unit status acquisition module, a task data acquisition module, and: The allocation decision module is used to construct a structured assembly context based on the collected unit status data and task data, input it into a semantic reasoning model based on a large language model, and output the assembly unit allocation output. The verification and rollback module is used to perform constraint verification on the output of the assembly unit allocation. If the verification fails, the rollback allocation is performed to determine the target assembly unit. The local observation construction module is used to determine the set of ready processes within the target assembly unit based on process sequence constraints, and to construct a local observation tensor and a corresponding validity mask based on the set of ready processes, resource status, and resource calendar occupancy information. The local observation tensor includes assembly datum binding status fields, gated status fields for key quality control points, measurement completion fields, out-of-tolerance flag fields, fixture occupancy window fields, and measurement equipment occupancy window fields. The local observation tensor has a dimension of [missing information]. m × d A real matrix, the validity mask is of length . m A 0 / 1 vector; where, m This refers to the number of resources within the target assembly unit that are involved in scheduling. d For feature dimensions; The graph structure scheduling module is used to construct a bipartite graph of "process node - resource node". It inputs the local observation tensor, the bipartite graph and the validity mask into the graph structure policy network, and outputs the action probability distribution of each feasible process-resource pair; and determines the target action under the constraint of the validity mask; at the same time, it determines the planned start time and planned end time of the target process according to the constraint conditions. The production scheduling item generation and rolling revision module is used to generate production scheduling items based on target processes, target resources, planned start time and planned end time, and write them into the production scheduling scheme dataset, and synchronously update resource calendar occupancy information. The reward generation and learning update module is used to calculate intermediate evaluation vectors for candidate actions and their corresponding candidate successor assembly states using an executable evaluation function, and then map them into step rewards through a proxy reward model. The production scheduling result output module is used to output the production scheduling result when the preset output conditions are met.

2. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, After executing the target process corresponding to the key quality control point, the production scheduling entry generation and rolling revision module triggers the measuring equipment to measure the set of docking feature points or the features corresponding to the key quality control point to obtain the measurement results. When the deviation represented by the measurement results exceeds the threshold corresponding to the assembly tolerance level, a rework process is generated and inserted into the assembly process set. At the same time, the process sequence constraints and gating status fields are updated so that the rework process enters the ready process set after satisfying the constraints. The production scheduling scheme dataset is then rolled over and revised. Finally, the measurement results are written into the assembly status.

3. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, The unit status acquisition module is used to collect unit status data, which includes: the resource availability status of the assembly unit, the in-process task queue, the earliest available time of each resource, resource calendar occupancy information, and capability identifiers; the resource availability status includes the assembly station occupancy status, assembly equipment availability status, tooling fixture availability status, and measuring equipment availability status; the capability identifier is used to characterize the assembly unit's ability to dock, position, and fasten key helicopter components; the resource calendar occupancy information is used to record the occupancy interval of each assembly station, assembly equipment, tooling fixture, and measuring equipment on the time axis.

4. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, The task data acquisition module is used to read task data, which includes task identifier, key component type, assembly process set, process sequence constraints, assembly datum information, assembly tolerance grade, delivery date, key quality control point set, and standard or estimated process time; the assembly datum information includes docking datum surface, datum hole, datum edge, and docking feature point set.

5. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, The constraint verification includes verification of the preset structured format output of the semantic reasoning model and feasibility verification.

6. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, When determining the target action, the graph structure scheduling module outputs multiple candidate process-resource pairs and calculates the corresponding candidate plan start time and candidate plan end time to form multiple candidate production schedule entries. It also receives the candidate step reward obtained by the reward generation and learning update module through the intermediate evaluation vector mapping of the candidate successor assembly status and candidate action corresponding to each candidate production schedule entry, and selects the candidate action with the largest candidate step reward as the target action.

7. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, The reward generation and learning update module includes an evaluation function generation and verification correction submodule, a reward mapping submodule, and a policy learning update submodule. The evaluation function generation and verification correction submodule is used to provide the semantic reasoning model with evaluation dimension definitions, input and output dimension constraints and boundary check rules, and to provide key quality control point gating rules, measurement closed-loop rules and measurement deviation handling rules related to helicopter assembly, so as to generate an executable definition of the evaluation function; The reward mapping submodule is used to calculate intermediate evaluation vectors for candidate successor assembly states and candidate actions corresponding to candidate actions using a verified executable evaluation function, and to map the intermediate evaluation vectors to step rewards through a proxy reward model. The policy learning update submodule is used to construct or estimate the advantage function based on the step reward, and update the policy parameters of the graph structure policy network according to the policy learning objective function of the graph structure policy network. The updated policy parameters are then output to the graph structure scheduling module.

8. The helicopter component assembly scheduling system based on a large language model according to claim 1, characterized in that, The production scheduling items include process identifier, assembly unit identifier, resource identifier, planned start time, planned end time, tooling fixture identifier, measuring equipment identifier or measuring window identifier, gate status, and rework indicator.

9. A helicopter component assembly scheduling method based on a large language model, characterized in that, include: S1: Obtain the unit status data of each assembly unit; S2: Obtain task data for newly arrived assembly tasks; S3: Construct a structured assembly context based on unit state data and task data, input a semantic reasoning model based on a large language model, and output the assembly unit allocation output. S4: Perform constraint verification on the output of the assembly unit allocation. If it fails, perform a rollback allocation to determine the target assembly unit. S5: Within the target assembly unit, determine the set of ready processes based on process sequence constraints, and construct a local observation tensor and its corresponding validity mask based on the set of ready processes, resource status, and resource calendar occupancy information; the local observation tensor includes an assembly datum binding status field, a gated status field for key quality control points, a measurement completion field, an out-of-tolerance flag field, a fixture occupancy window field, and a measuring equipment occupancy window field; the local observation tensor has a dimension of... m × d A real matrix, the validity mask is of length . m A 0 / 1 vector; where, m This refers to the number of resources within the target assembly unit that are involved in scheduling. d For feature dimensions; S6: Construct a bipartite graph of "process node - resource node", and input the local observation tensor, bipartite graph and validity mask into the graph structure strategy network, output the action probability distribution of each feasible process - resource pair; under the constraint of validity mask, determine the target action; and determine the planned start time and planned end time of the target process according to the constraint conditions. S7: Based on the target process, target resources, the planned start time and planned end time of the target process, generate production scheduling items, write them into the production scheduling dataset, and synchronously update the resource calendar occupancy information; S8: Output production scheduling results when preset output conditions are met.

10. The helicopter component assembly scheduling method based on a large language model according to claim 9, characterized in that, Before outputting the production scheduling results when the preset output conditions are met, the following steps are also included: After executing the target process corresponding to the key quality control point, the measuring equipment is triggered to measure the set of docking feature points or the features corresponding to the key quality control point to obtain the measurement results. When the deviation represented by the measurement results exceeds the threshold corresponding to the assembly tolerance level, a rework process is generated and inserted into the assembly process set. At the same time, the process sequence constraints and gating status fields are updated so that the rework process enters the ready process set after satisfying the constraints. The production scheduling plan dataset is then rolled over and revised. Finally, the measurement results are written into the assembly status.

Images