Multi-agent aircraft general assembly dynamic scheduling iterative optimization method based on digital twinning

Abstract

The invention relates to a multi-agent aircraft general assembly dynamic scheduling iterative optimization method based on digital twinning, and the method employs a production line-station two-stage regulation and control method under an active / reaction type scheduling strategy, and combines a knowledge-driven improved / learning type aircraft general assembly scheduling algorithm. A multi-agent reinforcement learning algorithm based on experience sharing is used for model training, and a dynamic scheduling iterative optimization method based on a multi-Agent system is used for timely responding to an abnormal disturbance event in an aircraft final assembly production line. Compared with the prior art, the method has the advantages that the dynamic scheduling and iterative evolution mechanism of the assembly outline between pulsation stations and in the stations is deeply explored and clarified, and a series of problems that abnormal disturbance response is not timely, scheduling is complex and coordination is unbalanced and the like in a complex production environment are solved through the dynamic scheduling iterative optimization process; and a new mode of dynamic scheduling and intelligent management and control of the aircraft general assembly workshop is realized.

Description

Technical Field

[0001] The present invention relates to the field of aircraft manufacturing and assembly, and specifically to a dynamic scheduling iterative optimization method for multi-agent aircraft final assembly based on digital twin. Background Technique

[0002] In the manufacturing of large aircraft, the workload of aircraft assembly accounts for 50% - 70% of its direct manufacturing workload. Automated assembly and flexible assembly have always been the directions of aircraft assembly efforts. Taking a certain large airliner in China as an example, it mainly consists of the nose, front fuselage, middle fuselage - center wing, rear fuselage, vertical tail, wings, tail wings, etc., with a total of about more than 1 million parts. These parts involve the entire manufacturing industrial chain, and the manufacturing of most parts is simultaneously undertaken by Xi'an Aircraft Industry (Group) Company, Shenyang Aircraft Corporation, Chengdu Aircraft Industry (Group) Company, Hongdu Aviation Industry Corporation, etc. It has the characteristics of numerous parts, large size, poor rigidity, high precision, and complex shape and structure. The main processes of aircraft assembly include the sub-assembly stage and the final assembly and commissioning stage. The sub-assembly stage is a segmented assembly, mainly for fuselage section assembly, wing section assembly, and tail wing assembly. The difference between the final assembly stage and the sub-assembly stage is that the whole aircraft needs to be docked. Due to safety reasons, the accuracy requirements are extremely strict. At the same time, system integration and commissioning are carried out, such as the integration and comprehensive commissioning of the fuel system, hydraulic system, flight control system, electrical system, etc. In the final assembly stage, a large number of complex processes are involved. Each process usually intersects, and it is extremely difficult to allocate personnel. It requires overall planning and scheduling capabilities. At the same time, due to the large volume and complex processes of the aircraft, most of the work is manual assembly by workers, the working conditions are complex and changeable, and dynamic response and commissioning are required. The safety requirements are extremely strict, and it is also necessary to trace the origin, record a large amount of quality and process information to form complete resume information. At the same time, it is difficult to divide and regulate the scheduling within the station among the stations in the pulsating production line, and it is difficult to quickly and efficiently regulate and adapt to the production site. Summary of the Invention

[0003] The purpose of the present invention is to overcome the traditional scheduling mode of "experience + static + algorithm", solve a series of problems such as complex aircraft final assembly operation scheduling, low completion rate of the daily plan of the assembly outline, difficult personnel management and control, untimely handling of abnormal events, quality inspection and control, etc., and design a dynamic scheduling iterative optimization method for multi-agent aircraft final assembly based on digital twin.

[0004] The purpose of the present invention can be achieved through the following technical solutions:

[0005] A dynamic scheduling iterative optimization method for multi-agent aircraft final assembly based on digital, the method includes the following steps:

[0006] S1: Under the digital twin workshop, perform virtual-real matching modeling for aircraft final assembly dynamic scheduling;

[0007] S2: Design a multi-agent system (MAS) to obtain a scheduling solution;

[0008] S3: The scheduling decision agent evaluates and synchronizes its plan, simulates the twin workshop and updates the data;

[0009] S4: If an abnormal disturbance event occurs, the disturbance event is identified through the feature library built by the system, and the attention mechanism deep long short-term memory network is used to predict the working hours under the current conditions, and its predicted data is updated to the scheduling system;

[0010] S5: The scheduling decision agent evaluates the current scheduling plan. If the current plan cannot be executed, the job rescheduling strategy is used for multiple iterations of optimization. If it still cannot be executed, the scheduling decision agent provides data feedback.

[0011] S6: After receiving data feedback from S5, a two-level control strategy of "production line-station" based on reactive scheduling is adopted, and iterative optimization is continuously performed to finally obtain a better scheduling plan, which is the final execution plan.

[0012] S7: After responding to the disturbance event, the final scheduling plan obtained through multiple iterative optimizations is applied to the aircraft assembly dynamic scheduling twin model. If an abnormal disturbance event occurs, S4 to S7 are repeated to achieve iterative optimization of the entire aircraft assembly dynamic scheduling system to obtain an efficient and robust scheduling plan.

[0013] Furthermore, step S1 is specifically as follows:

[0014] In the database of the aircraft assembly dynamic scheduling system, the production factor data with pertinence, accuracy and real-time performance are screened out to provide a model basis and data support for the subsequent iterative optimization of aircraft assembly dynamic scheduling;

[0015] A two-level planning model and a triple Euler model are used to construct a dynamic scheduling twin for aircraft final assembly. Data mining and machine learning techniques are used to obtain historical data, real-time data, and deduced data of the aircraft final assembly workshop, and the data of the dynamic scheduling twin for aircraft final assembly is updated.

[0016] Furthermore, step S2 is specifically as follows:

[0017] S21: Map the physical layer data to the twin layer one by one, build the aircraft assembly dynamic scheduling system and its database in the virtual environment, and continuously update the data;

[0018] Collect multi-source heterogeneous information such as the operation information of aircraft final assembly, the real-time status of resources, the technical levels and proficiency of personnel, and the environment, and construct a large database. Design a Multi-Agent System (MAS) for iterative optimization. MAS includes operation Agent, resource Agent, personnel Agent, computing Agent, and scheduling decision Agent; the operation Agent, resource Agent, and personnel Agent respectively screen and process the data of the aircraft final assembly dynamic scheduling system to obtain the corresponding production factor data;

[0019] S211: Obtain the operation data of the aircraft final assembly line and continuously iterate and update the operation data pool

[0020] The operation Agent mainly contains the total operation information of each aircraft, that is, the Assembly Outline (AO). According to the number of aircraft to be assembled currently, select the total number of AOs that need to be executed for aircraft assembly within this week, formulate the operation data pool for this week, and upload this data to the scheduling decision sub-Agent for data evaluation;

[0021] Screen and process the data of the above aircraft final assembly dynamic scheduling system to obtain the operation data of the aircraft final assembly line. The Assembly Outline (AO) of the aircraft pulsating assembly line is the smallest unit of the currently assigned task and is usually scheduled by AO. Each AO includes specific AO number, AO name, successor AO set (OSP process constraint, specifying the precedence relationship of AOs), section code, section limit, man-hour, required worker type, required number of workers, required tool type, required number of tools, required types of special workers or skilled workers, and other relevant data. The operation Agent first selects the unexecuted AOs within this week (five days) according to the OSP process constraint and puts them into the operation data pool, uploads this data to the scheduling decision sub-Agent, interacts and judges with the data of the resource Agent and personnel Agent to determine whether the AOs in the current operation data pool meet the execution conditions, and continuously iterates and updates the operation data pool.

[0022] S212: Obtain all the resource information in the aircraft final assembly stage and update and iterate the resource database

[0023] The resource Agent interacts with the aircraft final assembly dynamic scheduling system in real time, obtains all the information of resources, stores it in the resource database, uploads the data to the scheduling decision sub-Agent, analyzes the material completeness according to experience and knowledge rules, judges whether the current AO in the operation Agent meets the material resource constraint. If it meets, put it into the processable AOs as the operation to be executed. If it does not meet the part resource constraint, mark it and wait for the resource Agent to interact with the total system for update and then re-judge;

[0024] Screen and process the data of the above aircraft final assembly dynamic scheduling system to obtain the specific data of the resources required for the aircraft final assembly operation AO, mainly including material distribution information and material distribution notice information. Since the amount of operation resources required for aircraft final assembly is extremely large and the material parts involve the entire manufacturing industry chain, the material completeness is extremely low, which has a very serious impact on the aircraft final assembly dynamic scheduling, and the data update and iteration are extremely rapid and frequent. The resource Agent efficiently processes its data information. The material distribution information is that the current material has reached the material library of the assembly line, and AO meets the startable conditions; the material distribution notice information is that the current material has not arrived yet, but has been distributed, and the arrival time can almost be predicted. Mark this AO, and once the material arrives, immediately update it to a startable AO. For a series of problems such as tooling equipment damage and part processing quality, follow the same operation method. Every time the resource Agent iterates and updates, it uploads the data to the scheduling decision sub-Agent to interact with the data of the operation Agent, so that the operation data pool is updated and iterated.

[0025] S213: Obtain the specific information of the personnel in the current aircraft team and process the information

[0026] The personnel Agent contains information such as the work skills, work levels, proficiency levels, and current work status of all workers. The personnel Agent screens the employees, judges whether various types of work are saturated, conducts personnel data statistics, and transfers the personnel information of the executable workers to the scheduling decision sub-Agent for data interaction;

[0027] Filter and process the data of the above aircraft final assembly dynamic scheduling system to obtain the specific information of the personnel in the current aircraft team. Since the aircraft final assembly line involves professional knowledge and skills in multiple fields and has extremely strict requirements for quality and safety, the personnel information is complex and the scheduling is very difficult. The Personnel Agent analyzes and processes the personnel qualifications and personnel information. Each worker includes specific data information such as personnel qualifications, skill levels, work proficiency, personnel work status, and personnel workstations. The personnel qualification is that "among all the dispatched personnel, at least one person has the qualifications required for any one AO", that is, the current AO must be assigned personnel with corresponding qualifications to perform the operation; the skill level is the skill level of the person. The higher the skill level, the more types of operations the worker can perform; the work proficiency is the proficiency shown by the worker in performing a certain operation. The higher the work proficiency, the shorter the time for the worker to perform this operation. The Personnel Agent obtains the work status of the worker at the current moment, marks the idle workers and their information, and uploads it to the Scheduling Decision Sub-Agent to assign workers to the executable AOs in the Job Agent. If no worker with this qualification can be found in the current state, wait until the corresponding worker completes the operation and the human resources are released. After the data of the Personnel Agent is updated, the scheduling is carried out again; or if not enough workers can be found to execute the AO within the same time period, consider using operators. The number of operators is fixed and they do not have special qualifications and skill levels. If the start-up conditions are still not met, the workers need to complete the operation and release the human resources. After the data of the Personnel Agent is updated, the scheduling is carried out again.

[0028] S22: The Scheduling Decision Agent interacts and updates the data information of all sub-Agents to determine the material completeness information, executable personnel information, and executable operation information, and transmits the current data to the Computing Agent through the sub-Agent.

[0029] The sub-Agents in the Scheduling Decision Agent receive the data from the Job Agent, Resource Agent, and Personnel Agent, determine the material completeness information, executable AO list, and executable personnel information under the current working conditions, aggregate the current data, and use it as the scheduling data set to be transmitted to the Computing Agent to obtain a feasible solution for the static scheduling under the current working conditions. The scheduling data set is iterated as the data in the Job Agent, Resource Agent, and Personnel Agent is updated.

[0030] S23: The Computing Agent calls the system integration algorithm, dynamically constructs a mathematical model, calculates and selects the best scheduling plan for the next evaluation.

[0031] The computing agent calls the knowledge-driven "improvement / learning" type aircraft final assembly scheduling algorithm, trains the model of the multi-agent reinforcement learning algorithm based on experience sharing, dynamically constructs the aircraft final assembly dynamic scheduling mathematical model in combination with the current system data and working conditions, calculates the decision-making scheduling plan, and continuously compares it with the existing plan. Finally, an optimal decision-making scheduling plan is selected and sent to the scheduling decision agent for application;

[0032] S231: Transmit the scheduling data set to the computing agent and dynamically construct the aircraft final assembly dynamic scheduling mathematical model. For the convenience of modeling, some variable symbols are defined as follows:

[0033] CT: Assembly line beat

[0034] SI: Assembly line smoothness index

[0035] T: Maximum completion time of the assembly line

[0036] M: Number of workers on the aircraft assembly line

[0037] N: Total number of operations on the aircraft assembly line

[0038] Q: Aircraft section space, Q = A, B,... Z

[0039] l: Assembly line station index, l = 1, 2,... L

[0040] q: Assembly section index, j = 1, 2,... Q

[0041] j: Assembly operation index, j = 1, 2,... N

[0042] i: Assembly worker index, i = 1, 2,... M

[0043] k: Assembly resource index, k = 1, 2,... K

[0044] o: Assembly work type index, o = 1, 2,... O

[0045] S: Special worker index, s = 1, 2,... S

[0046] d: Job time discrete node, d = 1, 2,... T

[0047] t j : Continuous working time of job j

[0048] ST j : Start execution time of job j

[0049] ET j : Completion time of job j

[0050] Pj : Set of immediate predecessors of job j

[0051] S j : Set of immediate successors of job j

[0052] R k : Total maximum amount of resource k available

[0053] r jk : Requirement of job j for resource k

[0054] r jo : Requirement of job j for ordinary worker o

[0055] r js : Requirement of job j for special worker s

[0056] E lq : Maximum space capacity of the middle section q of station l

[0057] e jq : Requirement of job j for section q

[0058] u jPj If job j and the previous job Pj are completed by the same worker, it is 1; otherwise, it is 0

[0059] x jlq If job j is assigned to be completed in the q-th section of station l, it is 1; otherwise, it is 0

[0060] y joi If job j is assigned to be completed by the i-th person of ordinary type o, it is 1; otherwise, it is 0

[0061] z js If job j is assigned to be completed by the s-th person of special type, it is 1; otherwise, it is 0

[0062] Based on the above variable symbols, the objective function is defined as follows:

[0063] (1) Minimize the cycle time (CT) of the assembly line

[0064] min CT = max(T l ) l ∈ [1, L]

[0065] Minimizing the cycle time of the assembly line means that the aircraft moves pulsatingly to the next station according to a certain assembly line cycle time. In the aircraft final assembly dynamic scheduling problem of this patent, it is the second type of assembly line balancing problem, that is, when the number of stations is determined, the smaller the assembly line cycle time CT, the smaller the total assembly time;

[0066] (2) Minimize the smooth index (SI) of the assembly line

[0067]

[0068] The minimum assembly line smoothness index refers to the degree of dispersion of the station times on the assembly line, indicating the balance of the workload within each station. Generally speaking, the smaller the assembly line smoothness index SI, the more balanced the workload within each station.

[0069] Construct a mathematical model and establish constraints:

[0070] (1) Operations on the aircraft assembly line need to be completed within the station beat and shall not exceed the station beat time.

[0071] ST j +t j ≤CT, j ∈ [1, L]

[0072] (2) Each operation needs to be completed within the specified time and shall not be overdue.

[0073] ET j -ST j ≤t j , j ∈ [1, L]

[0074] (3) Each operation can only be assigned to be executed in a section space within one station.

[0075]

[0076] (4) Constraints on the precedence relationship of operations. The start time of the subsequent operation shall not exceed the end time of the previous operation.

[0077]

[0078] (5) Constraint on the number of workers, that is, the total number of assigned workers shall not exceed the maximum number of provided workers. The former represents ordinary workers and the latter represents special workers.

[0079]

[0080] (6) At the same time, each special worker can only execute one operation.

[0081]

[0082] (7) Resource constraint, that is, the number of resources required for parallel operations shall not exceed the total amount of provided resources. The former represents the resource constraint of ordinary workers and the latter represents the resource constraint of special workers.

[0083] ∑x jlq y joi r jk ≤R k, where \(j\in[1,N]\), \(i\in[1,M]\), \(l\in[1,L]\), \(o\in[1,O]\), \(k\in[1,K]\), \(q\in Q\)

[0084] \(\sum_{x}\) jlq z js r jk \(\leq R\) k , where \(j\in[1,N]\), \(l\in[1,L]\), \(o\in[1,O]\), \(k\in[1,K]\), \(q\in Q\)

[0085] (8) Spatial constraint: Neither ordinary workers nor special workers shall exceed the maximum capacity of the section space

[0086] \(\sum_{x}\) jlq y joi e jq +x jlq z js e jq \(\leq E\), where \(j\in[1,N]\), \(i\in[1,M]\), \(l\in[1,L]\), \(o\in[1,O]\), \(q\in Q\)

[0087] After the model is determined, the multi-agent reinforcement learning algorithm based on experience sharing is called for training. First, initialize the computing agents in the aircraft final assembly dynamic scheduling system, observe the current state of the digital twin model, obtain production factor data, use the multi-agent reinforcement learning algorithm based on experience sharing for model training, store the data and the model in the computing agents, generate a scheduling plan based on the data in the current database, transmit it to the intelligent decision-making agent, evaluate the current plan, including multiple indicators, conduct a comprehensive evaluation, and record the optimal scheduling plan and evaluation results in the historical database. By continuously updating the production state, historical database, and learning model, a production scheduling plan can be obtained quickly and efficiently.

[0088] S232: Obtain the resource data, personnel data, and operation data under the current working conditions, and convert the aircraft final assembly dynamic scheduling problem into a Markov decision process

[0089] Obtain the production factor data from the aircraft final assembly dynamic scheduling system, including various resource data, ordinary worker and special worker data, and specific data such as the precedence relationship, required resources, required personnel, and operation time of each operation. Abstract the aircraft final assembly dynamic scheduling problem into a Markov decision process (MDP) to provide data for subsequent algorithms.

[0090] S233: Design an agent for each job individual and initialize the local Q-value table \(Q\) of all agents i (s t ,a t ) and the public Q-value table \(Q\) c (s t ,at )

[0091] Design an agent for each job, and each agent defines a 5-tuple (S, A, α, γ, R). Among them, S represents the set of states, s t represents the state corresponding to the agent at time t, A represents the set of actions, a t represents the action selected by the agent at time t, α represents the learning rate, γ represents the discount rate, R represents the reward obtained by the agent. At the same time, improve the update mechanism of the Q-value table of each agent, including the local Q-value table Q i (s t , a t ), the local experience pool E i and the public Q-value table Q c (s t , a t ), the public experience pool E c , so that experience sharing and cooperative scheduling can be carried out among agents.

[0092] S234: Judge the current state and select an action according to the ε-greedy strategy

[0093]

[0094] S235: The agent executes the current action, returns the next state, and calculates the reward

[0095] r = w 1 R time + w 2 R resource + w 3 R worker

[0096]

[0097] S236: Each agent stores the current data experience e = (s, a, r, s′) into its own experience pool and the public experience pool, executes the experience sharing strategy, and updates the public Q-value table Q c (s t , a t )

[0098] E i ← E i ∪{e}

[0099]

[0100] S237: Distribute the public experience pool to each agent and update the local experience pool E of the agent i

[0101] Ei ←E i ∪E c

[0102] S238: Each agent learns from the local experience pool, updates the policy network, and makes the best choice

[0103]

[0104] S239: Loop operation, repeating steps S234 to S238 until the number of loops is reached

[0105] Each intelligent agent repeats the above steps and continuously optimizes collaboratively until the final result is obtained as a scheduling plan under the current working conditions.

[0106] Furthermore, step S3 is specifically as follows:

[0107] A feasible job scheduling plan is initially determined through S2 and uploaded to the scheduling decision agent. The current plan is evaluated by comparing it with historical data or experience. If it is the optimal value in historical data, or is not much different from previous experience data, it is determined as the current scheduling plan and uploaded to the aircraft dynamic scheduling system for simulation verification and effect analysis. The pros and cons of the current plan are analyzed through simulation. If the effect is good, the current scheduling plan is assigned to each agent, the data is updated, and it is uploaded to the aircraft assembly workshop at the same time, and the aircraft is assembled according to this plan;

[0108] The better scheduling scheme obtained in the calculation agent is transmitted to the scheduling decision agent for evaluation. As an intelligent agent, the scheduling decision agent stores a large amount of historical data in the database, and combines historical data with workers' experience for large-scale training. It is the core of the entire MAS collaborative scheduling. The scheduling decision agent obtains a scheduling scheme under the current working conditions, and first evaluates whether there is a resource conflict or a personnel conflict. If no conflict occurs, it is evaluated with the historical data, that is, the historical data of each station of an aircraft after assembly. If the overall data is not much different from the historical data, or is better than the historical data, it is determined as the scheduling scheme executed in the current state, and the scheme is uploaded to the aircraft assembly dynamic scheduling system for simulation. The scheme is applied to the twin model of dynamic scheduling of aircraft assembly for model simulation. If the scheme can complete the simulation without any job, conflicts between resources and personnel make it impossible to proceed, or the completion time of the current scheme is relatively small and can meet the expectations of the scheduler, the scheme will be assigned to each Agent to complete a job, record and update the data, and upload it to the aircraft assembly production line for production application in the aircraft assembly workshop.

[0109] Further, step S4 is specifically as follows:

[0110] S41: Identify and train the current abnormal event using the method for identifying uncertain disturbance events based on a convolutional neural network to obtain the recognition result of the disturbance event.

[0111] When an abnormal disturbance occurs in the aircraft final assembly production line, the relevant data is updated to the aircraft final assembly dynamic scheduling system. Through the disturbance event type feature library in the system, the method for identifying uncertain disturbance events based on a convolutional neural network is used to identify the current abnormal event, and the result after neural network training is transmitted to the job Agent, resource Agent, and personnel Agent for data synchronization and update.

[0112] When an abnormal disturbance occurs in the aircraft final assembly workshop, it is mapped to the aircraft final assembly dynamic scheduling twin model for analysis of the type and characterization of the assembly environment disturbance event. Data extraction is performed on it. According to the different data forms, the operation data of the assembly workshop is divided into time series data and grayscale image data. Three convolutional layers, two pooling layers, and a fully connected layer with Softmax activation are constructed using a one-dimensional convolutional kernel to train the neural network, and the recognition result of the disturbance event is obtained.

[0113] S42: Extract and analyze the features of the results identified in S41, use the attention mechanism deep long short-term memory network to predict the assembly working hours, and perform data update.

[0114] Most of these abnormal disturbances will cause changes in the aircraft assembly operation working hours, thus having a significant impact on the entire production line. Extract and analyze the features of the results identified in S41, use the attention mechanism deep long short-term memory network to predict the assembly working hours, and input the predicted working hours into the job Agent for data update.

[0115] Further, step S5 is specifically as follows:

[0116] S51: Update the obtained disturbance recognition data to the job Agent, resource Agent, and personnel Agent, and iterate the data information in the previous working condition.

[0117] Update the obtained disturbance recognition data to the job Agent, resource Agent, and personnel Agent. Among them, the working hours after the abnormal disturbance are predicted through step S42, and the predicted working hours are iterated into the job Agent. The personnel Agent updates the worker data, marks the workers who are working, and iteratively updates the worker data again. The resource Agent deletes the parts that have been completed, marks the information of the parts and tooling that are being used, adds the newly arrived parts, and performs material completeness analysis.

[0118] Based on the completeness of the resource Agent and the executable worker data of the personnel Agent, the job Agent re-evaluates the startable AOs. According to the data pool of the system, it iterates the jobs that have been processed, adds new executable jobs, and uploads the synchronized data to the scheduling decision-making Agent; it conducts data interaction between the updated data through the job Agent, resource Agent, personnel Agent and the scheduling decision-making sub-Agent, and iterates the data information in the previous working condition according to the operations from step S211 to step S213 again.

[0119] S52: Obtain the optimal solution,

[0120] The scheduling decision-making Agents conduct information interaction and sharing through sub-Agents, call the computing Agent to obtain the executable solutions under the current working condition, obtain the station scheduling solution and the job scheduling solution, which is a feasible planning solution;

[0121] Transmit the scheduling data set after abnormal disturbances to the computing Agent, call the specific operation steps of S23, obtain the job scheduling solution under the current working condition, specifically including the station scheduling solution and the job scheduling solution, use this solution as the planning solution, conduct iterative optimization, and obtain the optimal solution.

[0122] S53: The scheduling decision-making Agent judges the job scheduling solution,

[0123] First, perform rescheduling on the job scheduling solution. The scheduling decision-making Agent judges the current solution. By comparing with the historical data and the completion time in the station scheduling, if the completion time of the current job scheduling solution exceeds the planned time in the station scheduling solution, affecting the job scheduling within the subsequent stations and causing it to be unable to be executed according to the expected time, or the scheduling result is relatively poor and fails to meet the expectations of the scheduling personnel, the data is returned to the computing Agent for re-job scheduling. If the result is still relatively poor after multiple schedulings, the data is fed back to the scheduling decision-making Agent, and the scheduling decision-making Agent changes the scheduling strategy.

[0124] Furthermore, step S6 is specifically as follows:

[0125] S61: Change the scheduling strategy and adopt a two-level control strategy to obtain a better station scheduling solution,

[0126] After receiving the feedback data, since job rescheduling can no longer handle this disturbance event, the scheduling decision-making Agent adopts a "production line - station" two-level control strategy based on reactive scheduling. First, perform station rescheduling at the production line level, re-call the computing Agent, change the set parameters, re-divide the assembly outline, and conduct iterative optimization to obtain a better station scheduling solution;

[0127] S62: Based on the station scheduling plan, further calculate the job scheduling plan.

[0128] After obtaining the updated station scheduling plan, the calculation Agent changes the set parameters, calculates the job scheduling plan within each station, repeats S5, continuously performs job rescheduling, obtains a relatively optimal job scheduling plan, and uploads it to the scheduling decision Agent for evaluation and decision-making.

[0129] S63: Repeat step S61 and step S62, continuously iterate the two-level scheduling plan of "production line - station", obtain a relatively optimal scheduling plan as the final execution plan.

[0130] Upload the obtained job scheduling plan to the scheduling decision Agent. The scheduling decision Agent makes a judgment, decides to enable job rescheduling or the two-level control of "production line - station" for reactive scheduling, repeats S61 and S62 until a relatively ideal scheduling plan is obtained as the final execution plan.

[0131] Compared with the prior art, the present invention has the following beneficial effects:

[0132] 1) The present invention applies the digital twin technology to the aircraft final assembly workshop, mapping the huge production status to the database in the aircraft final assembly dynamic scheduling system. The aircraft assembly line involves the station division and job scheduling of tens of thousands of AOs, and the front and rear constraints are very complex. The resource demand is huge, the personnel scheduling is complicated and difficult, and the entire data volume is very large. At the same time, these data are associated in various ways, and the coupling mechanism is complex. Driven by the data in the digital twin workshop, the iterative operation of production element management, production activity planning, and production process control is realized in the twin workshop, and the adaptive scheduling and self-organizing iterative optimization of dynamic time-varying scenarios are realized.

[0133] 2) The present invention adopts a two-level control method of "production line - station" based on the "active / reactive" scheduling strategy to iteratively optimize the scheduling plan. At present, most studies separate the production line level - station beat and the station level - knowledge beat and consider them separately. The two-level control method adopted by the present invention provides an initial scheduling plan through active scheduling, and the reactive scheduling strategy corrects it on this basis to adapt to the occurrence of abnormal disturbances. At the same time, the station scheduling plan and the job scheduling plan are comprehensively considered. The job scheduling plan is scheduled after being issued by the station scheduling plan, and the job scheduling plan will also affect the beat of the station scheduling plan. Abnormal disturbances and the overall plan are comprehensively considered during the scheduling process, so as to maximize the efficiency and robustness of scheduling. This self-organizing and adaptive scheduling mechanism can not only cope with known rules and constraints, but also timely respond to unknown changes and abnormal situations, making the scheduling process more flexible and efficient.

[0134] 3) The present invention uses a multi-agent reinforcement learning algorithm based on experience sharing for the dynamic scheduling of aircraft final assembly. In view of the characteristics of flexible and variable large aircraft assembly processes, complex fuselage structures, numerous parallel operations, extremely heavy assembly tasks, complex division of assembly outlines, and difficult job scheduling, the ESMAQ algorithm is integrated into the computing agent of the aircraft final assembly dynamic scheduling system. The computing agent obtains data from the system, including job information, resource status, spatial constraints, etc. Then, combined with the computing power of the ESMAQ algorithm, the system will be able to quickly generate an optimized scheduling plan in a dynamic environment. This integration process will inject higher-level intelligence and flexibility into the dynamic scheduling system, enabling it to more accurately respond to the changing environment and requirements.

[0135] 4) The dynamic scheduling iterative optimization method based on the multi-agent system proposed by the present invention collects multi-source heterogeneous information such as job information of aircraft final assembly, real-time status of resources, technical levels and proficiency of personnel, and the environment, constructs a large database for the entire system to provide data support, and uses MAS to achieve the iterative optimization of aircraft final assembly dynamic scheduling. It mainly uses the "production line - station" two-level control method of the scheduling decision agent based on the "active / reactive" scheduling strategy to timely respond to abnormal disturbance events in the aircraft final assembly production line, solve the problem that the division of the station beat in the assembly outline and the knowledge beat within the station is unreasonable, which makes it difficult to execute the closed-loop iterative optimization of the system, realizes the iterative optimization of aircraft final assembly dynamic scheduling, obtains the best real-time production plan, and improves the robustness and fault tolerance of the aircraft final assembly production line. Description of the Drawings

[0136] Figure 1 It is a schematic diagram of the technical route of dynamic scheduling iterative optimization based on multi-agent reinforcement learning in the digital twin workshop of the present invention;

[0137] Figure 2 It is a framework diagram of the reinforcement learning algorithm of the computing agent of the present invention;

[0138] Figure 3 It is a training flow chart of the multi-agent reinforcement learning algorithm based on experience sharing of the present invention;

[0139] Figure 4 It is a two-level control flow chart based on the "active / reactive" scheduling strategy of the present invention;

[0140] Figure 5 It is a schematic diagram of the dynamic scheduling iterative optimization process based on the multi-agent system of the present invention. Detailed Embodiments

[0141] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives the detailed implementation manner and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0142] The present invention provides a digital-twin-based iterative optimization method for dynamic scheduling of multi-agent aircraft final assembly. It mainly includes a two-level regulation method of "production line - station" based on an "active / reactive" scheduling strategy, and a knowledge-driven "improve / learn" type aircraft final assembly scheduling algorithm, mainly a scheduling method based on an experience-sharing multi-agent reinforcement learning algorithm and a dynamic scheduling iterative optimization method based on a multi-Agent system, as shown in the attached Figures 1-5 figure. The method includes the following steps:

[0143] S1: Under the digital-twin workshop, perform virtual-real matching modeling for dynamic scheduling of aircraft final assembly;

[0144] In the database of the aircraft final assembly dynamic scheduling system, screen out production factor data with pertinence, accuracy, and real-time nature, providing a model basis and data support for subsequent iterative optimization of aircraft final assembly dynamic scheduling;

[0145] Due to reasons such as the complex assembly outline and production conditions of aircraft final assembly dynamic scheduling, the complex and changeable multi-source heterogeneous data during the assembly process, the difficulty in predicting and tracing disturbances, etc., perform virtual-real matching modeling for final assembly dynamic scheduling under the digital-twin workshop. Adopt a two-layer programming model and a triple Euler model to construct a digital twin of aircraft final assembly dynamic scheduling. Use data mining and machine learning technologies to obtain historical data, real-time data, and deduced data of the aircraft final assembly workshop, and update the data of the aircraft final assembly dynamic scheduling digital twin.

[0146] S2: Design a multi-Agent system (Multi-Agent System, MAS) to obtain a scheduling plan;

[0147] Respectively, the job Agent, resource Agent, and personnel Agent obtain and screen out the required production factor data, conduct data interaction with the scheduling decision Agent, transmit the data information under the current working conditions to the computing Agent, dynamically construct a mathematical model for aircraft final assembly dynamic scheduling, and call the knowledge-driven "improve / learn" type aircraft final assembly scheduling algorithm, mainly using an experience-sharing multi-agent reinforcement learning algorithm for model training to obtain an initial scheduling plan;

[0148] S21: Map the physical layer data to the twin layer one by one, construct an aircraft final assembly dynamic scheduling system and its database in the virtual environment, and continuously update the data;

[0149] Collect multi-source heterogeneous information such as the operation information of aircraft final assembly, the real-time status of resources, the technical levels and proficiency of personnel, and the environment, and construct a large database. And design a Multi-Agent System (MAS) for iterative optimization. MAS includes operation Agent, resource Agent, personnel Agent, computing Agent, and scheduling decision Agent; the operation Agent, resource Agent, and personnel Agent respectively screen and process the data of the aircraft final assembly dynamic scheduling system to obtain their corresponding production factor data;

[0150] S211: Obtain the operation data of the aircraft final assembly line and continuously iterate and update the operation data pool

[0151] The operation Agent mainly contains the total operation information of each aircraft, that is, the Assembly Outline AO. According to the number of aircraft to be assembled currently, select the total number of AOs that need to be executed for aircraft assembly within this week, formulate the operation data pool for this week, and upload this data to the scheduling decision sub-Agent for data evaluation;

[0152] Screen and process the data of the above aircraft final assembly dynamic scheduling system to obtain the operation data of the aircraft final assembly line. The Assembly Outline AO of the aircraft pulsating assembly line is the smallest unit of the currently assigned task, and usually scheduling is carried out based on AO. Each AO includes specific AO number, AO name, successor AO set (OSP process constraint, specifying the precedence relationship of AOs), section code, section limit, man-hour, required worker type, required number of workers, required tool type, required number of tools, required special workers or skilled worker types, and other relevant data. The operation Agent first selects the unexecuted AOs within this week (five days) according to the OSP process constraint and puts them into the operation data pool, uploads this data to the scheduling decision sub-Agent, interacts and judges with the data of the resource Agent and personnel Agent to determine whether the AOs in the current operation data pool meet the execution conditions, and continuously iterates and updates the operation data pool.

[0153] S212: Obtain all the resource information of the aircraft final assembly stage and update and iterate the resource database

[0154] The resource Agent interacts with the aircraft final assembly dynamic scheduling system in real time, obtains all the information of the resources, stores it in the resource database, uploads the data to the scheduling decision sub-Agent, analyzes the material completeness according to experience and knowledge rules, judges whether the current AO in the operation Agent meets the material resource constraint. If it meets, put it into the processable AO as the operation to be executed. If it does not meet the part resource constraint, mark it and wait for the resource Agent to interact with the total system for update and then re-judge;

[0155] Filter and process the data of the above aircraft final assembly dynamic scheduling system to obtain the specific data of the resources required for the aircraft final assembly operation AO, mainly including material distribution information and material distribution notice information. Since the amount of operation resources required for aircraft final assembly is extremely large and the material parts involve the entire manufacturing industry chain, the material completeness is extremely low, which has a very serious impact on the aircraft final assembly dynamic scheduling, and the data update and iteration are extremely rapid and frequent. The resource Agent processes its data information efficiently. The material distribution information is that the current material has reached the material warehouse of the assembly line, and AO meets the startable condition; the material distribution notice information is that the current material has not arrived yet, but it has been distributed and the arrival time can almost be predicted. Mark this AO, and once the material arrives, immediately update it to a startable AO. For a series of problems such as tooling equipment damage and part processing quality, follow the same operation method. Every time the resource Agent iterates and updates, it uploads the data to the scheduling decision sub-Agent and interacts with the data of the operation Agent to update and iterate the operation data pool.

[0156] S213: Obtain the specific information of the personnel in the current aircraft team and process the information

[0157] The personnel Agent contains information such as the work skills, work levels, proficiency levels, and current work status of all workers. The personnel Agent screens the employees, judges whether various types of work are saturated, conducts personnel data statistics, and transmits the personnel information of the executable workers to the scheduling decision sub-Agent for data interaction;

[0158] Filter and process the data of the above aircraft final assembly dynamic scheduling system to obtain the specific information of the personnel in the current aircraft team. Since the aircraft final assembly line involves professional knowledge and skills in multiple fields and has extremely strict requirements for quality and safety, the personnel information is complex and the scheduling is very difficult. The Personnel Agent analyzes and processes the personnel qualifications and personnel information. Each worker includes specific data information such as personnel qualifications, skill levels, work proficiency, personnel work status, and personnel workstations. The personnel qualification is that among all the dispatched personnel, at least one person has the qualification required for any one AO, that is, the current AO must be assigned personnel with corresponding qualifications to perform the operation; the skill level is the skill level of the person. The higher the skill level, the more types of operations the worker can perform; the work proficiency is the proficiency shown by the worker in performing a certain operation. The higher the work proficiency, the shorter the time for the worker to perform this operation. The Personnel Agent obtains the work status of the worker at the current moment, marks the idle workers and their information, and uploads it to the Scheduling Decision Sub-Agent to assign workers to the executable AOs in the Job Agent. If no worker with this qualification can be found in the current state, wait until the corresponding worker completes the operation and the human resources are released. After the data of the Personnel Agent is updated, the scheduling is carried out again; or if not enough workers can be found to execute the AO within the same time period, consider using operators. The number of operators is fixed and they do not have special qualifications and skill levels. If the starting conditions are still not met, the workers need to complete the operation and release the human resources. After the data of the Personnel Agent is updated, the scheduling is carried out again.

[0159] S22: The Scheduling Decision Agent interacts and updates the data information of all sub-Agents to determine the material completeness information, executable personnel information, and executable operation information, and transmits the current data to the Computing Agent through the sub-Agent.

[0160] The sub-Agents in the Scheduling Decision Agent receive the data from the Job Agent, Resource Agent, and Personnel Agent, determine the material completeness information, executable AO list, and executable personnel information under the current working conditions, aggregate the current data as the scheduling data set, and transmit it to the Computing Agent to obtain a feasible solution for the static scheduling under the current working conditions. The scheduling data set is iterated as the data in the Job Agent, Resource Agent, and Personnel Agent is updated.

[0161] S23: The Computing Agent calls the system integration algorithm, dynamically constructs a mathematical model, calculates and selects the best scheduling plan for the next evaluation.

[0162] The computing Agent calls the knowledge-driven "improvement / learning" type aircraft final assembly scheduling algorithm, trains the model of the multi-agent reinforcement learning algorithm based on experience sharing, dynamically constructs the aircraft final assembly dynamic scheduling mathematical model in combination with the current system data and working conditions, calculates the decision-making scheduling plan, continuously compares it with the existing plans, and finally selects an optimal decision-making scheduling plan to send to the scheduling decision Agent for application;

[0163] S231: Transmit the scheduling data set to the computing Agent and dynamically construct the aircraft final assembly dynamic scheduling mathematical model

[0164] For the convenience of modeling, define some variable symbols:

[0165] CT: Assembly line beat

[0166] SI: Assembly line smoothness index

[0167] T: Maximum completion time of the assembly line

[0168] M: Number of workers on the aircraft assembly line

[0169] N: Total number of operations on the aircraft assembly line

[0170] Q: Aircraft section space, Q = A, B,... Z

[0171] l: Assembly line station index, l = 1, 2,... L

[0172] q: Assembly section index, j = 1, 2,... Q

[0173] j: Assembly operation index, j = 1, 2,... N

[0174] i: Assembly worker index, i = 1, 2,... M

[0175] k: Assembly resource index, k = 1, 2,... K

[0176] o: Assembly work type index, o = 1, 2,... O

[0177] S: Special worker index, s = 1, 2,... S

[0178] d: Job time discrete node, d = 1, 2,... T

[0179] t j : Continuous working time of job j

[0180] ST j : Start execution time of job j

[0181] ET j : Completion time of job j

[0182] P j : The set of immediate predecessors of job j

[0183] S j : The set of immediate successors of job j

[0184] R k : The maximum total amount of resources provided by the k-th type of resource

[0185] r jk : The demand of job j for resource k

[0186] r jo : The demand of job j for ordinary worker o

[0187] r js : The demand of job j for special worker s

[0188] E lq : The maximum space capacity of the middle section q of the station l

[0189] e jq : The demand of job j for section q

[0190] If job j and the previous job Pj are completed by the same worker, it is 1; otherwise, it is 0

[0191] x jlq If job j is assigned to be completed in the q-th section of station l, it is 1; otherwise, it is 0

[0192] y joi If job j is assigned to be completed by the i-th person of the ordinary work type o, it is 1; otherwise, it is 0

[0193] z js If job j is assigned to be completed by the s-th person of the special work type, it is 1; otherwise, it is 0

[0194] Based on the above variable symbols, define the objective function:

[0195] (1) Minimize the assembly line cycle time (Cycle Time, CT)

[0196] min CT = max(T l ) l ∈ [1, L]

[0197] Minimizing the assembly line cycle time means that the aircraft moves pulsatingly to the next station according to a certain assembly line cycle time. In the aircraft final assembly dynamic scheduling problem of this patent, it is the second type of assembly line balancing problem, that is, when the number of stations is determined, the smaller the assembly line cycle time CT, the smaller the total assembly time;

[0198] (2) Minimize the Smooth Index (SI) of the assembly line

[0199]

[0200] Minimizing the Smooth Index of the assembly line refers to the degree of dispersion of the station times on the assembly line, indicating the balance of the workload within each station. Generally speaking, the smaller the Smooth Index SI of the assembly line, the more balanced the workload within each station.

[0201] Construct a mathematical model and establish constraints:

[0202] (1) Aircraft assembly line operations need to be completed within the station cycle time and shall not exceed the station cycle time

[0203] ST j +t j ≤CT, j∈[1,L]

[0204] (2) Each operation needs to be completed within the specified time and shall not be overdue

[0205] ET j -ST j ≤t j , j∈[1,L]

[0206] (3) Each operation can only be assigned to a section space within one station for execution

[0207]

[0208] (4) Constraints on the precedence relationship of operations. The start time of the subsequent operation shall not exceed the end time of the previous operation

[0209]

[0210] (5) Constraint on the number of workers, that is, the total number of assigned workers shall not exceed the maximum number of provided workers. The former represents ordinary workers and the latter represents special workers

[0211]

[0212] (6) At the same time, each special worker can only execute one operation

[0213]

[0214] (7) Resource constraints, that is, the number of resources required for parallel operations shall not exceed the total amount of provided resources. The former represents the resource constraint of ordinary workers and the latter represents the resource constraint of special workers

[0215] ∑x jlq y joi rjk ≤R k , where \(j\in[1,N]\), \(i\in[1,M]\), \(l\in[1,L]\), \(o\in[1,O]\), \(k\in[1,K]\), \(q\in Q\)

[0216] \(\sum x\) jlq z js r jk ≤R k , where \(j\in[1,N]\), \(l\in[1,L]\), \(o\in[1,O]\), \(k\in[1,K]\), \(q\in Q\)

[0217] (8) Spatial constraint: Neither ordinary workers nor special workers shall exceed the maximum capacity of the section space.

[0218] \(\sum x\) jlq y joi e jq +x jlq z js e jq ≤E, where \(j\in[1,N]\), \(i\in[1,M]\), \(l\in[1,L]\), \(o\in[1,O]\), \(q\in Q\)

[0219] After the model is determined, the multi-agent reinforcement learning algorithm based on experience sharing is called for training. First, initialize the computing agents in the aircraft final assembly dynamic scheduling system, observe the current state of the digital twin model, obtain the production factor data, use the multi-agent reinforcement learning algorithm based on experience sharing for model training, store the data and the model in the computing agents, generate a scheduling plan based on the data in the current database, transmit it to the intelligent decision-making agent, evaluate the current plan, including multiple indicators, conduct a comprehensive evaluation, and record the optimal scheduling plan and evaluation results in the historical database. By continuously updating the production state, historical database, and learning model, a production scheduling plan can be obtained quickly and efficiently. The specific framework is as Figure 2 shown. The specific algorithm training process is as follows:

[0220] S232: Obtain the resource data, personnel data, and operation data under the current working conditions, and convert the aircraft final assembly dynamic scheduling problem into a Markov decision process.

[0221] Obtain the production factor data from the aircraft final assembly dynamic scheduling system, including various resource data, ordinary worker and special worker data, the precedence relationship, required resources, required personnel, operation time, etc. of each operation, and abstract the aircraft final assembly dynamic scheduling problem into a Markov decision process (MDP) to provide data for the subsequent algorithm.

[0222] S233: Design an agent for each operation individual and initialize the local Q-value table Q of all agents. i (st , a t ), and the public Q - value table Q c (s t , a t )

[0223] Design an agent for each job. Each agent defines a 5 - tuple (S, A, α, γ, R), where S represents the set of states, s t represents the state corresponding to the agent at time t, A represents the set of actions, a t represents the action selected by the agent at time t, α represents the learning rate, γ represents the discount rate, R represents the reward obtained by the agent. At the same time, improve the update mechanism of each agent's Q - value table, including the local Q - value table Q i (s t , a t ), the local experience pool E i and the public Q - value table Q c (s t , a t ), the public experience pool E c , so that experience sharing and collaborative scheduling can be carried out among agents.

[0224] S234: Judge the current state and select an action according to the ε - greedy strategy

[0225]

[0226] S235: The agent executes the current action, returns the next state, and calculates the reward

[0227] r = w 1 R time + w 2 R resource + w 3 R worker

[0228]

[0229] S236: Each agent stores the current data experience e=(s, a, r, s′) into its own experience pool and the public experience pool, executes the experience sharing strategy, and updates the public Q - value table Q c (s t , a t )

[0230] E i ← E i ∪ {e}

[0231]

[0232] S237: Distribute the common experience pool to each agent and update the local experience pool E of the agent i

[0233] E i ←E i ∪E c

[0234] S238: Each agent learns from the local experience pool, updates the policy network, and makes an optimal choice

[0235]

[0236] S239: Perform loop operations, repeat steps S234 to S238 until the number of loops is reached

[0237] Each agent repeats the above steps, continuously collaborates and optimizes until the final result is obtained, which is used as a scheduling plan under the current working conditions. The specific algorithm flow chart is as shown in the appendix Figure 3 as follows. Its pseudocode is as follows:

[0238]

[0239]

[0240] S3: The scheduling decision Agent evaluates and synchronizes its plan, conducts twin workshop simulation and updates data;

[0241] A feasible job scheduling plan is initially determined through S2 and uploaded to the scheduling decision Agent. It is compared with historical data or experience to evaluate the current plan. If it is the optimal value in the historical data or is not much different from the previous experience data, it is determined as the current scheduling plan and uploaded to the aircraft dynamic scheduling system for simulation verification and effect analysis. Analyze the advantages and disadvantages of the current plan through simulation. If the effect is relatively good, the current scheduling plan is distributed to each Agent for data update, and at the same time uploaded to the aircraft final assembly workshop for aircraft assembly according to this plan;

[0242] The better scheduling scheme obtained in the calculation agent is transmitted to the scheduling decision agent for evaluation. As an intelligent agent, the scheduling decision agent stores a large amount of historical data in the database, and combines historical data with workers' experience for large-scale training. It is the core of the entire MAS collaborative scheduling. The scheduling decision agent obtains a scheduling scheme under the current working conditions, and first evaluates whether there is a resource conflict or a personnel conflict. If no conflict occurs, it is evaluated with the historical data, that is, the historical data of each station of an aircraft after assembly. If the overall data is not much different from the historical data, or is better than the historical data, it is determined as the scheduling scheme executed in the current state, and the scheme is uploaded to the aircraft assembly dynamic scheduling system for simulation. The scheme is applied to the twin model of dynamic scheduling of aircraft assembly for model simulation. If the scheme can complete the simulation without any job, conflicts between resources and personnel make it impossible to proceed, or the completion time of the current scheme is relatively small and can meet the expectations of the scheduler, the scheme will be assigned to each Agent to complete a job, record and update the data, and upload it to the aircraft assembly production line for production application in the aircraft assembly workshop.

[0243] S4: If an abnormal disturbance event occurs, the disturbance event is identified through the feature library built by the system, and the attention mechanism deep long short-term memory network is used to predict the working hours under the current conditions, and its predicted data is updated to the scheduling system;

[0244] S41: The uncertain disturbance event recognition method based on the convolutional neural network is used to identify and train the current abnormal event, and the recognition result of the disturbance event is obtained.

[0245] When an abnormal disturbance occurs in the aircraft assembly production line, the relevant data is updated to the aircraft assembly dynamic scheduling system. The current abnormal event is identified through the disturbance event type feature library in the system and the uncertain disturbance event identification method based on the convolutional neural network. The results after neural network training are transmitted to the job agent, resource agent, and personnel agent for data synchronization update;

[0246] Abnormal disturbances occur in the aircraft assembly workshop and are mapped to the aircraft assembly dynamic scheduling twin model. The types of disturbance events in the assembly environment and their characterization are analyzed, and data is extracted. According to the different data forms, the assembly workshop operation data is divided into time series data and grayscale image data. A one-dimensional convolution kernel is used to construct three convolution layers, two pooling layers and a fully connected layer with Softmax activation. After training the neural network, the recognition results of the disturbance events are obtained.

[0247] S42: Extract and analyze the features of the results identified in S41, use the attention mechanism deep long short-term memory network to predict the assembly man-hours, and update the data.

[0248] Most of these abnormal disturbances will cause changes in the man-hours of aircraft assembly operations, and then have a significant impact on the entire production line. Extract and analyze the features of the results identified in S41, use the attention mechanism deep long short-term memory network to predict the assembly man-hours, and input the predicted man-hours into the operation Agent for data update.

[0249] Extract the data features of the disturbance event, use the recognition results obtained in step S41, establish a man-hour prediction model based on the attention mechanism deep long short-term memory network, comprehensively consider the data information of the assembly operation, the completeness information of parts and tooling resources, the scheduling data information of personnel, and the temporal correlation relationship between the real-time data and the disturbance event, effectively predict the man-hours of the assembly operation, and update the predicted man-hours to the operation Agent as the operation scheduling information after the disturbance.

[0250] S5: The scheduling decision Agent judges the current scheduling plan. If the current plan cannot be executed, the operation rescheduling strategy is used for multiple iterative optimizations. If it still cannot be executed, the scheduling decision Agent gives a data feedback on it.

[0251] Synchronize the updated operation data, resource data, and personnel data, and adopt a two-level control process based on the "active / reactive" scheduling strategy, as shown in the appendix Figure 3 First, perform the station scheduling plan, divide the assembly outline according to the current station scheduling plan, and then perform the operation scheduling within each station to generate a preliminary overall scheduling plan, which is judged and executed by the scheduling decision Agent. If it cannot be executed, first perform the operation scheduling rearrangement. If the operation scheduling plan does not reach the better data in history or conflicts with the station scheduling plan, then perform the station scheduling plan rearrangement, and continuously iterate and update to obtain a better scheduling plan.

[0252] S51: Update the obtained disturbance recognition data to the operation Agent, resource Agent, and personnel Agent, and iterate the data information in the previous working condition.

[0253] Update the obtained disturbance identification data to the job Agent, resource Agent, and personnel Agent. Among them, predict the operation man-hours after abnormal disturbance through step S42, and iterate the predicted man-hours into the job Agent. The personnel Agent updates the worker data, marks the workers who are working, and iteratively updates the worker data again. The resource Agent deletes the parts that have been completed, marks the information of the parts and tooling that are being used, adds the newly arrived parts, and conducts material sufficiency analysis.

[0254] Based on the sufficiency of the resource Agent and the executable worker data of the personnel Agent, the job Agent re-evaluates the startable AO. According to the data pool of the system, iterate the completed operations, add new executable operations, and upload the synchronized data to the scheduling decision Agent; interact the updated data through the job Agent, resource Agent, personnel Agent, and scheduling decision sub-Agent, and iterate the data information in the previous working condition according to the operations from step S211 to step S213 again.

[0255] S52: Obtain the optimal solution.

[0256] The scheduling decision Agents interact and share information through sub-Agents, call the calculation Agent to obtain the executable solutions under the current working condition, and obtain the station scheduling solution and operation scheduling solution, which is a feasible planning solution.

[0257] Transmit the scheduling data set after abnormal disturbance to the calculation Agent, call the specific operation steps of S23 to obtain the operation scheduling solution under the current working condition, which specifically includes the station scheduling solution and operation scheduling solution. Take this solution as the planning solution and conduct iterative optimization to obtain the optimal solution.

[0258] S53: The scheduling decision Agent judges the operation scheduling solution.

[0259] First, perform rescheduling on the operation scheduling solution. The scheduling decision Agent judges the current solution. By comparing with the historical data and the completion time in the station scheduling, if the completion time of the current operation scheduling solution exceeds the planned time in the station scheduling solution, affecting the operation scheduling within the subsequent stations and causing it to be unable to be executed as expected, or the scheduling result is relatively poor and cannot meet the expectations of the scheduling personnel, return the data to the calculation Agent to re-perform operation scheduling. If the result is still relatively poor after multiple schedulings, feedback this data to the scheduling decision Agent, and the scheduling decision Agent changes the scheduling strategy.

[0260] S6: After the data feedback in S5, adopt a two-level regulation strategy of "production line - station" based on reactive scheduling, continuously iterate and optimize, and finally obtain a relatively optimal scheduling plan as the final execution plan;

[0261] S61: Replace the scheduling strategy, adopt a two-level regulation strategy to obtain a better station scheduling plan,

[0262] The scheduling decision Agent receives the feedback data. Since job rescheduling can no longer handle this disturbance event, adopt a two-level regulation strategy of "production line - station" based on reactive scheduling. First, perform station rescheduling at the production line level, re-call the calculation Agent, change the set parameters, re-divide the assembly outline, and iterate and optimize to obtain a better station scheduling plan;

[0263] Since job rescheduling can no longer handle this disturbance event, the scheduling decision Agent decides to adopt a two-level regulation of "production line - station", transmit the changed strategy to the calculation Agent, and the calculation Agent changes the set parameters and re-divides the assembly outline. With the minimum completion time and smoothness as the optimization objectives, re-calculate the station scheduling plan, and continuously iterate and optimize to obtain a better station scheduling plan for further job scheduling.

[0264] S62: On the basis of the station scheduling plan, further calculate the job scheduling plan,

[0265] After obtaining the updated station scheduling plan, the calculation Agent changes the set parameters, calculates the job scheduling plan within each station, repeats S5, continuously performs job rescheduling, obtains a relatively optimal job scheduling plan, and uploads it to the scheduling decision Agent for evaluation and decision-making.

[0266] S63: Repeat Step S61 and Step S62, continuously iterate the two-level scheduling plan of "production line - station", obtain a relatively optimal scheduling plan as the final execution plan,

[0267] Upload the obtained job scheduling plan to the scheduling decision Agent, and the scheduling decision Agent makes a judgment to decide whether to enable job rescheduling or the two-level regulation of "production line - station" based on reactive scheduling, and repeat S61 and S62 until a relatively ideal scheduling plan is obtained as the final execution plan.

[0268] S7: After responding to the disturbance event, apply the final scheduling plan obtained through multiple iterations and optimizations to the dynamic scheduling twin model of aircraft final assembly. If an abnormal disturbance event occurs, repeat S4 to S7 to realize the iterative optimization of the entire aircraft final assembly dynamic scheduling system to obtain an efficient and robust scheduling plan;

[0269] The final implementation plan is imported into the digital twin model of aircraft final assembly for simulation verification. The twin model accurately simulates all key elements in the workshop environment, including the real-time status of equipment, personnel, and materials, enabling the plan to be truly restored in the virtual environment. During the simulation process, the model will execute step by step according to the plan steps and dynamically monitor the effects and efficiencies of each link. The plan verified by the twin model is officially applied to the actual aircraft final assembly workshop, executed in the real production environment, and the real-time production data is synchronously updated into the digital twin model. Whenever a new abnormal disturbance occurs in the workshop, the real-time information is input into the dynamic scheduling system of aircraft final assembly, and the system will automatically trigger the above steps to execute the dynamic scheduling optimization process based on the multi-agent system, generate a new optimization plan, apply it to the workshop after simulation verification again, as shown in the appendix Figure 4 As shown, this iterative optimization mechanism continuously improves the stability and efficiency of the scheduling plan through the feedback and real-time response of the twin model, enabling the final assembly process to always maintain efficient and stable operation in a dynamic environment.

[0270] The specific embodiments of the present invention and the technical principles involved have been described in detail above. Those skilled in the art should understand that the scope of the present invention is not limited to the specific combination of the above technical features. Therefore, any technical solutions that can be obtained by those skilled in the art according to the concept of the present invention, combined with the prior art, through logical analysis, reasoning, or limited experiments shall fall within the protection scope defined by the claims.

Claims

1. A multi-agent aircraft assembly dynamic scheduling iterative optimization method based on digital twins, characterized in that: The method comprises the following steps: S1: In the digital twin workshop, virtual-real matching modeling of dynamic scheduling of aircraft assembly is carried out; S2: Design a multi-agent system (MAS) to obtain a scheduling solution; S3: The scheduling decision agent evaluates and synchronizes its plan, simulates the twin workshop and updates the data; S4: If an abnormal disturbance event occurs, the disturbance event is identified through the feature library built by the system, and the attention mechanism deep long short-term memory network is used to predict the working hours under the current conditions, and its predicted data is updated to the scheduling system; S5: The scheduling decision agent evaluates the current scheduling plan. If the current plan cannot be executed, the job rescheduling strategy is used for multiple iterations of optimization. If it still cannot be executed, the scheduling decision agent provides data feedback. S6: After receiving data feedback from S5, a two-level control strategy of "production line-station" based on reactive scheduling is adopted, and iterative optimization is continuously performed to finally obtain a better scheduling solution, which is the final execution solution. S7: After responding to the disturbance event, the final scheduling plan obtained through multiple iterative optimizations is applied to the aircraft assembly dynamic scheduling twin model. If an abnormal disturbance event occurs, S4 to S7 are repeated to achieve iterative optimization of the entire aircraft assembly dynamic scheduling system to obtain an efficient and robust scheduling plan.

2. The method for dynamic scheduling iterative optimization of multi-agent aircraft assembly based on digital twins according to claim 1, characterized in that: Step S2 is specifically as follows: S21: Map the physical layer data to the twin layer one by one, build the aircraft assembly dynamic scheduling system and its database in the virtual environment, and continuously update the data; S22: The scheduling decision agent interactively updates the data information of all sub-agents, determines the material completeness information, executable personnel information, executable operation information, and transmits the current data to the calculation agent through the sub-agent. S23: The computing agent calls the system integration algorithm, dynamically builds a mathematical model, calculates and selects the best scheduling solution for the next step of evaluation.

3. The method for dynamic scheduling iterative optimization of multi-agent aircraft assembly based on digital twins according to claim 1, characterized in that: Step S4 is specifically as follows: S41: The uncertain disturbance event recognition method based on the convolutional neural network is used to identify and train the current abnormal event, and the recognition result of the disturbance event is obtained. S42: Extract and analyze features of the results identified in S41, use the attention mechanism deep long short-term memory network to predict the assembly time, and update the data.

Images