CNP-NSGA-II-based aircraft pulse assembly line distributed rescheduling method

Abstract

The invention relates to an aircraft pulse assembly line distributed rescheduling method based on CNP-NSGA-II, and the method comprises the following steps: obtaining real-time disturbance state information, and obtaining a rescheduling scene type through a trained scene recognition model based on a gating circulation unit; in response to the rescheduling scene type being a first type, performing right-shift rescheduling on the aircraft pulsation assembly line; in response to the rescheduling scene type being a second type, optimizing the assembly process sorting and / or assembly worker allocation of the aircraft pulsation assembly line station by constructing a non-dominated sorting genetic algorithm II, and realizing in-station collaborative scheduling of the aircraft pulsation assembly line; and in response to the rescheduling scene type being a third type, generating bid invitation information including human resources needing assistance of other stations through a contract network algorithm, optimizing assembly worker allocation of the stations, and realizing inter-station collaborative scheduling of the aircraft pulsation assembly line.

Description

Technical Field

[0001] The invention relates to the field of assembly manufacturing, and in particular to a distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II (Contract Net-Non-Dominated Sorting Genetic Algorithm II). Background Art

[0002] In the aircraft manufacturing process, the assembly stage is the final key link, which includes many important operations such as large component docking, cable laying, key component installation and system testing. The execution of these operations directly determines the final output efficiency and supply capacity of the aircraft.

[0003] The meaning of pulsating assembly is that each assembly station starts the assembly task from the current station in the order of assembly instructions, completes the specified assembly process within a fixed cycle, and then enters the next station at the same time to carry out the assembly task of the next assembly cycle, and finally completes all assembly processes. Compared with the traditional fixed assembly line, the pulsating assembly line has a detailed and clear division of labor, a high degree of specialization, and a stable production cycle. With the changes in the domestic and international situation, the demand for various types of aviation aircraft such as commercial aircraft, military aircraft, and general aircraft is increasing. At the same time, the number of aircraft product components is increasing, and the process flow is more complicated. These changes have brought great challenges to the aircraft manufacturing industry. Therefore, optimizing the management and process flow of the assembly stage and improving work efficiency and quality are the key to improving the level of aircraft manufacturing.

[0004] In the actual assembly process of the aircraft pulse assembly line, some unpredictable emergencies often occur. In an uncertain environment, the manufacturing system generates a new scheduling strategy based on the original scheduling strategy according to the real-time changing system state. The scheduling method that adapts to the current state is called rescheduling. At present, the actual assembly line has not yet established an effective identification mechanism and response measures for emergencies. When disturbances occur, they are usually solved by overload methods such as overtime work, resulting in low assembly efficiency and high worker fatigue. Therefore, when an emergency occurs, how to efficiently identify the source of the emergency and adjust the scheduling strategy in time according to the characteristics and impact of the emergency has become an urgent problem to be solved.

[0005] In the Chinese patent application "An online simulation method for aircraft assembly line driven by working condition data" (publication number CN118898115A), Zhang Jie et al. proposed an online simulation method for aircraft assembly line driven by working condition data, which collects real-time on-site production information of aircraft assembly line, builds an online simulation model and outputs simulation results to predict the planned operation status of the aircraft assembly line's ongoing operations and unstarted operations, and makes abnormal judgments and abnormal traceability for delayed operations in the planned operation status, providing strong support for the decision-making of on-site control personnel. In the Chinese patent application "An intelligent monitoring system for aircraft assembly process based on simulation" (publication number CN106647336A), Guo Yu et al. proposed an intelligent monitoring system for aircraft assembly process based on simulation, which monitors the real-time assembly process visually, and based on real-time data updates, combines intelligent optimization algorithms and assembly knowledge, realizes manufacturing behavior prediction, abnormal manufacturing behavior identification, fault analysis, and material status tracking, effectively improving assembly management efficiency and assembly quality. Its advantage is that it takes some abnormal situations into consideration and makes predictions and tracking, but it does not take into account situations such as worker absenteeism, and does not propose corresponding re-scheduling plans, so it still depends on the decision of on-site control personnel.

[0006] In summary, there is currently a lack of a rescheduling method for aircraft pulse assembly lines to solve or partially solve the aforementioned problems. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II, so as to solve or partially solve the problem of how to efficiently identify the source of emergencies and how to reschedule for emergencies such as worker absence.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] The present invention provides a distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II, comprising the following steps:

[0010] Obtain real-time disturbance state information and use the trained gated recurrent unit-based scene recognition model to obtain the rescheduling scene type;

[0011] In response to the rescheduling scenario type being the first type, rescheduling the aircraft pulse assembly line to the right;

[0012] In response to the rescheduling scenario type being the second type, constructing a non-dominated sorting genetic algorithm II to optimize the assembly process sorting and / or assembly worker allocation of the aircraft pulse assembly line stations, so as to achieve intra-station collaborative scheduling of the aircraft pulse assembly line;

[0013] In response to the rescheduling scenario type being the third type, a contract network algorithm is used to generate bidding information including human resources that require assistance from other stations, optimize the allocation of assembly workers at the stations, and achieve coordinated scheduling between stations on the aircraft pulse assembly line.

[0014] As a preferred technical solution, the training process of the scene recognition model includes the following steps:

[0015] By simulating the rescheduling scenario, the disturbance state information under different rescheduling scenarios is obtained, the corresponding labels are obtained by marking, the disturbance state information is normalized, and training samples are constructed;

[0016] Construct a densely connected network with sequential connections, based on a gated recurrent neural network and a normalized exponential function layer as a scene recognition model;

[0017] The scene recognition model is trained using the training samples.

[0018] As a preferred technical solution, the optimization of the process sorting and / or worker allocation of the aircraft pulse assembly line station by constructing a non-dominated sorting genetic algorithm II comprises the following steps:

[0019] In response to the rescheduling scenario type being the second type, for the stations of the aircraft pulse assembly line, the rule weights corresponding to the assembly process optimization and the assembly worker allocation are used as chromosome encoding to initialize the population;

[0020] Through contract collaboration decoding, the optimization target values ​​of the actual completion time of the station and the time for dispatching workers to the bidding station are calculated. Subgroups are generated through fast non-dominated sorting, selection, crossover and mutation. The parent group and the subgroup are merged to minimize the actual completion time of the station and the time for dispatching workers to the bidding station. A new parent generation is generated through fast non-dominated sorting and congestion calculation. This step is repeated until the preset iteration round is reached.

[0021] Based on the Pareto front solution set of the iterated population, a rescheduling plan is obtained to determine whether the rescheduling plan exceeds the beat. If so, the scheduling process corresponding to the third type of rescheduling scenario type is executed. If not, the collaborative scheduling within the station of the aircraft pulse assembly line is realized based on the rescheduling plan.

[0022] As a preferred technical solution, the rules corresponding to the optimization of assembly processes include: the more subsequent processes the current process has, the higher the priority; the longer the sum of the time of the subsequent processes of the current process, the higher the priority; and the longer the time from the current process to the process with no subsequent processes, the higher the priority.

[0023] As a preferred technical solution, the corresponding rules for allocating assembly workers include: workers with shorter actual assembly time have higher priority; workers with shorter completion time have higher priority; and workers with higher assembly efficiency have higher priority.

[0024] As a preferred technical solution, the process of calculating the optimization target value of the rule weight includes the following steps:

[0025] For all assembly processes of the disturbing worker's team, they are divided into completed process set, assembly process set and to-be-assembled process set according to the disturbance;

[0026] For each assembly process in the set of to-be-assembled processes, based on the current population, the assembly process with the highest priority is calculated according to the rules corresponding to the assembly process optimization, the worker with the highest priority is calculated according to the rules corresponding to the assembly worker allocation, and the assembly process set and the worker status are updated;

[0027] After all assembly processes of the disturbing worker's team are updated, the optimization target value is calculated.

[0028] As a preferred technical solution, the process of generating bidding information including human resources requiring assistance from other stations through the contract network algorithm and optimizing the allocation of assembly workers at the stations includes the following steps:

[0029] Determine whether the negotiation resource lock in the current booth is accessible. If not, suspend and wait, and repeat this step until the negotiation resource lock is accessible;

[0030] In response to the negotiation resource lock being accessible, locking the negotiation resource lock;

[0031] Obtain multiple sets of candidate scheduling solutions for process sequencing and assembly worker allocation through the NSGA-II algorithm;

[0032] The multiple groups of candidate scheduling solutions are sorted, and the bidding information corresponding to the candidate scheduling solutions is constructed according to the sorting, and the bidding information corresponding to the candidate scheduling solutions with higher sorting is sent to other stations of the machine pulse assembly line. If other stations all refuse to bid, the bidding information corresponding to the next candidate scheduling solution is sent to other stations of the machine pulse assembly line, and the bidding is repeated until the information accepting the bidding is received or the negotiation rounds reach a preset upper limit, and the negotiation resource lock is released.

[0033] As a preferred technical solution, in response to the negotiation rounds reaching a preset upper limit, the assembly process optimization and assembly worker allocation corresponding rules of all stations of the machine pulse assembly line are retrained.

[0034] As a preferred technical solution, the bidding information includes the bidding station, the tendering station, the bidding virtual worker number and the virtual worker number in the tendering station that meets the requirements of the bidding document.

[0035] As a preferred technical solution, the aircraft pulse assembly line includes multiple stations, and a station agent for distributed decision-making is constructed for each station. Each of the station agents includes a scheduling agent, and a task agent and a worker agent that respectively optimize the assembly process and allocate assembly workers under the guidance of the scheduling agent. The distributed rescheduling method is applied to each station agent.

[0036] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0037] (1) Realize efficient identification of the source of emergencies: The present invention obtains real-time disturbance state information and uses a trained scene recognition model based on gated recurrent units to obtain the rescheduling scene type, thereby achieving efficient and accurate scene identification and preparing for subsequent rescheduling.

[0038] (2) Rescheduling for emergencies such as worker absence: The present invention adopts a distributed scheduling method, and each station can make decisions and execute independently, which reduces the communication burden and improves scheduling efficiency. At the same time, unlike the existing static scheduling, the present invention is based on the GRU neural network model, which can accurately identify the impact of disturbance events on the pulsation rhythm and classify them, preparing for subsequent rescheduling. The improved CNP algorithm and NSGA-II algorithm are used to realize the dynamic adjustment of resources between stations and the optimization of process sorting and worker allocation within the station, improve the coordination efficiency, and achieve global optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of a distributed rescheduling method of an aircraft pulse assembly line based on CNP-NSGA-II in an embodiment;

[0040] Figure 2 It is a framework diagram of the rescheduling scene recognition process in the embodiment;

[0041] Figure 3 This is a framework diagram of the intra-station collaborative rescheduling process based on NSGA-II in the embodiment;

[0042] Figure 4 It is a framework diagram of the rescheduling process based on improved negotiation between contractual websites in the embodiment;

[0043] Figure 5 Schematic diagram of an electronic device in an embodiment. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0045] Example 1

[0046] In view of the problems existing in the aforementioned prior art, the present embodiment provides a distributed rescheduling method for an aircraft pulsating assembly line based on CNP-NSGA-II (Contract Net-Non-Dominated Sorting Genetic Algorithm II), which aims to address the shortcomings of existing scheduling methods in dealing with dynamic disturbance events, achieve accurate identification and effective response to sudden worker absence events, and ensure the stable operation and production efficiency of the assembly line in a dynamic environment.

[0047] Step S0: Set up a distributed decision-making station agent at each station on the assembly line, which includes three sub-agents: task agent, worker agent, and scheduling agent. The task agent and worker agent solve two sub-tasks, namely, the optimization of assembly sequence and the allocation of assembly workers. These two types of agents achieve overall station performance optimization through real-time collaborative decision-making based on the guidance of the scheduling agent. After the task agent or worker agent perceives the sudden disturbance event information, it inputs the disturbance information into the information perception module of the scheduling agent to identify the rescheduling scene.

[0048] Step S1, see Figure 1 The scheduling agent uses the pre-trained gated recurrent unit (GRU) to analyze real-time information, identify the impact of disturbance events on the pulse beat, and divide it into three scenarios: Scenario 1 (i.e. the first type of rescheduling scenario): the impact on the station is not significant, and only right shift rescheduling is required; Scenario 2 (i.e. the second type of rescheduling scenario): after right shift rescheduling, the original beat is still exceeded, and collaborative rescheduling within the station is required. Scenario 3 (i.e. the third type of rescheduling scenario): after right shift rescheduling and collaborative rescheduling within the station, the beat is still exceeded, and inter-station negotiation rescheduling is required.

[0049] For details, see Figure 2 , based on GRU rescheduling scene recognition. Specifically including steps S101-S103:

[0050] Step S101, sample generation and processing. First, a simulation method is used to simulate various rescheduling scenarios, collect status data in different rescheduling scenarios, and label each sample. That is, for each type of scenario: if it can be solved by directly shifting right and rescheduling, it is marked as scenario 1 that does not require rescheduling; if it exceeds the beat, try to reassign to other workers according to the original rules; if it does not exceed the beat, it is marked as scenario 2 that requires rescheduling within the station; if it exceeds the beat, it is marked as category 3 that requires assistance from other stations.

[0051] Each sample finally formed contains state X:x1,x2,…,x n With labels y=1,2,3. X describes the disturbance information from three aspects: process, worker, and station. The state includes information on the process, worker, and station. Specifically, the process includes two state quantities: the sum of the subsequent process time of the disturbed worker and the number of subsequent processes of the disturbed worker; the worker aspect includes four state quantities: the disturbance team number, the expected disturbance duration, the remaining task time of the normal workers in the same team of the disturbed worker, and the remaining time of the process of the disturbed worker; the station aspect includes two state quantities: the current distance beat remaining time and the disturbance station ID. Next, the sample label is processed by one-hot encoding. This embodiment divides the rescheduling scenario into three scenarios. In order to eliminate the influence of the category value, one-hot encoding is used to convert labels 1, 2, and 3 into [1, 0, 0], [0, 1, 0], and [0, 0, 1] respectively. Finally, in order to eliminate the influence of dimensional imbalance on training, the state variables are normalized. Let

[0052] Step S102, network training. First, design the network. The input layer uses a dense network. The middle layer adds a GRU neural network. The last layer adds a normalized exponential function (softmax) layer for classification. Then, randomly initialize the parameters and set the loss function as follows: Iteratively calculate the loss function and update the parameters for training. Save the trained network to the scene recognition module of the scheduling agent.

[0053] Step S103, when performing rescheduling scene recognition, the real-time disturbance state information is input into the trained model, and the rescheduling scene category is output.

[0054] Step S2: After identifying the rescheduling scenario 1, the scheduling agent directly issues a right shift rescheduling instruction and instructs the task agent and worker agent to make local adjustments. It does not involve changes in the process sequence and assigned personnel. The scheduling agent issues the finalized scheduling plan to the task agent and worker agent to instruct them to perform the actual scheduling execution.

[0055] Specifically, for rescheduling scenario 1, the impact of the disturbance in this scenario is small, and the right shift rescheduling does not exceed the predetermined beat and does not affect subsequent normal operations, so no other adjustments are required.

[0056] Step S3, after identifying the rescheduling scenario 2, the scheduling agent instructs the task agent to use the Non-dominated Sorting Genetic Algorithms-II (NSGA-II) algorithm to optimize the process sorting and worker allocation within the station to adapt to the impact of the disturbance event. The decision execution module of the scheduling agent directly issues the adjusted scheduling strategy to the scheduling execution modules of the task agent and the worker agent, and the agents in the two stations convert it into an actual scheduling plan through collaborative execution.

[0057] See also Figure 3 , this step provides a rescheduling method for intra-station collaboration based on NSGA-II for rescheduling scenario 2. The intra-station collaborative rescheduling problem is: at station z, team b has a set of processes to be assembled, and some workers of team b are absent. Station z adjusts the process sequencing rules and personnel allocation rules of team b so that team b can complete the assembly within the specified beat under disturbance. This problem needs to consider the optimization goals of both production performance and stability, that is, it is necessary to ensure that the performance after rescheduling is optimized, and to minimize the difference between the changed scheduling plan and the original scheduling plan as much as possible to avoid excessive costs in resource reallocation. The optimization objectives include minimizing the actual completion time of the team and the number of changes in the worker's process. This part combines NSGA-II with the assembly rules in the pulse line to realize the adjustment of the intra-station scheduling plan, and improves the encoding and decoding methods based on the specific characteristics of the research object.

[0058] Improvement of the encoding method: The encoding method combined with assembly rules is adopted. In the research object, since the number of rescheduled processes is small, but the process number varies widely, and the workers to be assigned also have multiple attributes such as station and team, if an end-to-end encoding method is adopted, it is necessary to re-encode according to the process set, worker team type, and worker number of each rescheduled process, which will cause the complexity of encoding. Therefore, the chromosome encoding is improved by the combined rule method. Each chromosome encoding w represents the weight of the adopted rule. [w1,w2,w3] belongs to the rule weight of the task agent, [w4,w5,w6] belongs to the rule weight of the worker agent, and w i∈(0,1). The rules of the task agent include that the more subsequent processes the process has, the higher the priority, the longer the sum of the subsequent process times, and the longer the time from the process to the process without subsequent processes. The rules of the worker agent include that the shorter the actual assembly time of the worker, the higher the priority, the shorter the completion time of the worker, and the higher the assembly efficiency of the worker. In the combined scheduling rule, the priority CP of the scheduling object o is o The mathematical expression of can be summarized as follows, P ow The priority of the scheduling object o calculated by a single rule.

[0059]

[0060] The actions based on the combined scheduling rules can be expressed as: Each value is the weight of the corresponding single scheduling rule, W i is the number of scheduling rules, represents the weight of the task agent rule w, represents the weight of the worker agent rule w.

[0061] Improvement of decoding method: cooperative decoding method is adopted. After the task agent and the worker agent work together, the chromosome can be decoded to obtain the optimization target value. First, the task agent divides the process set, checks all the assembly process tasks of the disturbed worker's team, and divides them into three process sets according to the different disturbance conditions, namely the completed process set, the assembly process set and the to-be-assembled process set. There is no need to adjust the completed processes, keep the original scheduling order for the assembly process set, and reschedule the to-be-assembled process set based on the rules output by NSGA-II. Then, the population output by the scheduling agent is decoded, that is, the process with the highest priority is calculated according to the task sorting rules, and the process is input into the worker agent, and then the to-be-assembled process set is updated; the worker agent then calculates the worker with the highest priority according to the allocation rules in the chromosome, and updates the worker status. After all processes are allocated, two optimization target values ​​are calculated.

[0062] Specifically, the intra-station collaborative rescheduling method based on NSGA-II may include steps S301-S304:

[0063] Step S301, population initialization: First, the scene recognition part of the scheduling agent outputs the intra-station adjustment result, and the intra-station collaborative rescheduling module initializes the population of the NSGA-II algorithm.

[0064] Step S302: Input the population into the task agent and the worker agent for decoding, and calculate the optimization target value.

[0065] Step S303, iterative optimization. After fast non-dominated sorting, selection, crossover, and mutation, a subgroup is generated, and then the parent generation and the child generation are merged, and a new parent generation is generated through fast non-dominated sorting and crowding calculation. Among them, the selection operator is based on the binary tournament strategy, selecting individuals with small non-dominated order and large crowding distance; the crossover operator randomly generates two crossover points on the two parent chromosomes for crossover in each round; the mutation operator sets the mutation distribution parameter μ, generates a random number ε in each round, and mutates the parent generation according to the polynomial mutation. The mutation of each chromosome is shown in the following formula

[0066]

[0067] Step S304, comprehensively consider stability and efficiency, and select the final rescheduling solution from the Pareto front solution set of the last generation population. Considering the accuracy limit of the rescheduling scenario prediction, if the scheduling solution obtained at this time still exceeds the beat, execute step S4 and enter the rescheduling scenario 3 based on the improved CNP station-to-station negotiation rescheduling.

[0068] Step S4, after identifying the rescheduling scenario 3, the scheduling agent uses the Contract Net Protocol (CNP) algorithm to generate a tender document containing human resource information that requires assistance from other stations, inputs it into the information perception module of the scheduling agent of other stations, and selects the tender documents returned by other stations to adjust the human resources between stations.

[0069] See also Figure 4 For rescheduling scenario 3, it is necessary to adjust the resources between stations. Since the information of each station in the distributed scheduling system is asymmetric and each station does not have the resource information of other stations, it is necessary for the station agents to adjust the resources through negotiation interaction. Among the existing negotiation algorithms, the contract network algorithm is a classic algorithm for solving the dynamic interaction of various parts in the distributed system. The logical structure of CNP is relatively simple, which can effectively decompose large-scale production tasks and dynamically find the optimal allocation plan. This section realizes the negotiation rescheduling between stations by improving the contract network algorithm.

[0070] First, we make the following assumptions. Due to the randomness of emergencies and the efficiency of algorithm processing, there is only one negotiation thread at each moment; each negotiation thread considers a team, that is, the workers who are absent each time belong to the same team, but they can be multiple workers in the same team.

[0071] The CNP method includes two key points: agent structure design and bidding strategy. Based on the distributed station agent, the dynamic role-playing mechanism of the bidding agent and the win-win oriented bidding negotiation strategy are designed respectively, and an improved CNP method is proposed.

[0072] Dynamic role-playing mechanism of bidding agent: Regarding the structure of bidding agent, traditional contract network usually requires a unified bidding party to manage tasks. In the research object of this paper, considering the need to adjust the personnel structure of the station, if such a mode is adopted, it is necessary to count the real-time data of all station-related teams and workers, and perform global rescheduling and simulation calculations, which greatly increases the workload, does not fully utilize the advantages of distributed scheduling, and is not conducive to the timely adjustment of scheduling schemes. In view of this problem, in order to improve the decision-making efficiency, a dynamic role-playing mechanism of bidding agent is designed, that is, each station agent can become a bidding party to prepare bidding documents, and can also become a bidding party to accept biddings from other stations. Human resource adjustment can be achieved through negotiation between station agents. In order to avoid conflicts among negotiation threads, a negotiation resource lock is first set. When the rescheduling scenario recognition module of the station scheduling agent outputs the rescheduling scenario 3, it checks whether the negotiation resource lock is accessible. If it is not accessible, the resource is suspended and waited; if it is accessible, the bidding process is entered, and the resource will be locked during the bidding period until the resource is released after the end of this bidding.

[0073] Win-win oriented bidding negotiation strategy: Traditional contract network algorithms often only consider the optimal performance of the tendering party. Since the agents at each station in this paper dynamically play the role of bidding agents, the relationship between each agent is equal. Therefore, when a station requests resource assistance from other stations, it should also consider the performance of other stations and try not to affect the normal assembly of other stations. This section focuses on the three most important steps in the negotiation process, namely, tender document design, bid document design, and tender document selection. Three sub-strategies are designed, namely, a tender document design method based on NSGA-II, a bid document preparation method based on idle time matching, and a bid document selection method based on worker combination efficiency scoring.

[0074] First, the tender document is designed based on NSGA-II. In order to reduce invalid bidding and ensure that the performance of other stations is not affected as much as possible, the specific information of the assistance required by the disturbed station is first calculated, including the number of assists and the assistance time. The calculation is still based on the NSGA-II algorithm. Through the collaboration between the scheduling agent, the task agent, and the worker agent, the final rescheduling plan is obtained to determine the tender document. Considering the minimization of the completion time of the bidding station and the minimization of the dispatch time of other stations, the objective function of the NSGA-II algorithm is designed to minimize the actual completion time f1 of the station and minimize the time f2 of the dispatched workers of the bidding station. The selection operator with preference is adopted. The preference means that when selecting two parent chromosomes based on the binary tournament strategy, the optimization target f2 of the two chromosomes is preferentially compared. In the chromosome decoding calculation of the worker agent, considering the role of the dispatched workers, virtual workers need to be set: if NoW workers are absent, NoW virtual workers are set. The Pareto front solutions are sorted in ascending order according to the sum of the assembly time of the dispatched workers of the external stations (more than one worker may be required to be dispatched externally), and tender documents are prepared separately. The tender document corresponding to each set of solutions is NoW dimension, and each dimension contains a tender information, represented by a five-tuple <Z 0 ,B,W 0 ,St,Ft> indicates that Z 0 represents the bidding station number, B represents the team number to be dispatched, that is, the team number of the absent worker; W 0 Indicates the number of the virtual worker to be dispatched, St represents the worker W 0 The expected start time of dispatch, Ft represents worker W 0 The estimated end time of the dispatch.

[0075] Next, the bid that matches the free time is prepared. After receiving the tender, other stations query the current status of the team workers and prepare the bid with the information of the workers who meet the requirements of the tender. Each bid is represented by a six-tuple <Z 0 ,Z,Agr,W 0 ,W,Effi>:Z 0 represents the bidding position, Z represents the tendering position; W 0 is the virtual worker number of the tender document, W is the worker number of the bidding station that meets the requirements of the tender document; Agr represents whether the bidding station agrees to bid. The bidding station responds to each piece of tender information〈Z 0 ,B,W 0 ,St,Ft〉, traverse the status of all workers in team B. If there is no idle worker in [St,Ft], Agr is 0, and W and Effi information are directly empty; if there is a worker who meets the bidding information, Agr is 1. W, Effi are the worker numbers and efficiencies that meet the requirements.

[0076] Finally, workers select bids based on their combined efficiency scores. i After receiving bids from other stations, find the best combination of dispatched workers through traversal.

[0077] Specifically, the inter-station negotiation rescheduling method based on the improved VNP may include steps S401-S407:

[0078] Step S401, set a negotiation resource lock, when the station agent i When the rescheduling scenario identification module outputs the rescheduling scenario 3, the negotiation module first checks whether the negotiation resource lock is accessible. If not, the resource is suspended and waited; if accessible, it enters 32) and locks the negotiation resource.

[0079] Step S402, station agent i The negotiation module and the task agent and worker agent obtain several sets of rescheduling solutions based on the NSGA-II algorithm.

[0080] Step S403: sort the scheduling solutions according to certain preferences, and design tender documents for the solutions in sequence according to the sorting order, and first broadcast the tender document corresponding to the solution ranked first to other stations.

[0081] Step S404, station agent j The scheduling agent information processing module receives the information from the agent i The tender is matched based on the free time and then sent back to the agent i .

[0082] Step S405, station agent i The scheduling agent information processing module receives bids from other stations and inputs them into the negotiation module. The negotiation module selects a suitable bidding station based on the worker combination efficiency score and notifies it, signs a contract with it, dispatches workers to the corresponding bidding station, terminates this negotiation, and releases the negotiation resource lock; if other stations refuse to bid, it enters the next round of bidding and goes to 46).

[0083] Step S406, station agent i The negotiation module selects the next solution to prepare the tender document according to the order of the solutions in Step 3, and repeats Steps S403 to S406 until the negotiation round reaches the upper limit. If no bidding position is found, go to Step S407.

[0084] Step S407, start global network training, retrain and generate assembly rules for all stations, and release the negotiation resource lock.

[0085] The effectiveness of this method will be illustrated below using a real aircraft pulse assembly line, which can perform the assembly tasks of various aircraft models. The assembly line layout and total human resources have been determined. The assembly line layout is linear and divided into 5 stations. Each station is responsible for the assembly of different parts. The aircraft completes the assembly task in each station in turn and is transported to the next station at a predetermined beat.

[0086] The final assembly line currently receives final assembly orders for various types of aircraft. This embodiment is based on a certain model for research. The process route of this model has been planned in the early stage, with a total of more than 3,000 processes. For the convenience of management, the processes are roughly divided into 12 process packages AL according to the assembly parts. The process packages strictly follow a linear relationship of close precedence and close succession in ascending alphabetical order, and the sub-processes within the process package present a nonlinear, sparse relationship of close precedence and close succession.

[0087] Table 1 shows the basic information of some processes. The standard assembly time, the required assembly team and the number of workers for each process are determined by the process characteristics. Each process requires a number of workers from a team to assemble. The distribution of assembly time is relatively uneven, with 94% of the processes taking less than 10 hours to complete. The number of assembly workers ranges from 1 to 5, and most of the assembly processes require the cooperation of 2 or more workers. In addition, the process set presents a sparse relationship between the preceding and the following processes, with 80% of the processes having no preceding process and 84% of the processes having no following process. Based on production experience, the initial cycle time is set to 290 hours.

[0088] The assembly line assembly teams have also been planned in advance: divided into 17 teams according to profession, each team has an assembly skill, each team has 5 workers, and the proficiency of the workers' skills is evenly distributed between [0.8,1.2]. For the same process, workers with higher proficiency values ​​have shorter assembly time. The actual assembly time of the process is the standard assembly time multiplied by the proficiency.

[0089] Table 1 Schematic diagram of some processes

[0090]

[0091] The actual assembly line is adopted, random worker absence disturbance events are added, and the method of the present invention is adopted to perform rescheduling optimization. The specific steps are as follows:

[0092] Step S1, such as Figure 2 , and perform GRU-based rescheduling scene recognition. The specific process is as follows:

[0093] Step S101, randomly set disturbances of worker absence in different shifts at each station, and collect multiple sets of disturbance state data during the simulation process. And add a label to each sample. That is, for each type of scenario: if it can be solved by directly shifting right and rescheduling, it is marked as scenario 1 that does not require rescheduling; if it exceeds the beat, try to reallocate to other workers according to the original rules; if it does not exceed the beat, it is marked as scenario 2 that requires rescheduling within the station; if it exceeds the beat, it is marked as scenario 3 that requires assistance from other stations. Each sample finally formed contains state X:x1,x2,…,x n With labels y=1,2,3. X describes the disturbance information from three aspects: process, worker, and station. The process aspect includes two state quantities: the sum of the subsequent process time of the disturbed worker and the number of subsequent processes of the disturbed worker; the worker aspect includes four state quantities: the disturbance team number, the expected disturbance duration, the remaining task time of normal workers in the same team of the disturbed worker, and the remaining time of the disturbed worker process; the station aspect includes two state quantities: the current distance beat remaining time and the disturbance station ID. Next, perform One-Hot Encoding on the sample labels. Change labels 1, 2, and 3 to [1, 0, 0], [0, 1, 0], and [0, 0, 1] respectively. Next, normalize the state variables and let

[0094] Step S102, network training. Design the network, use a dense network for the input layer, add a GRU neural network in the middle layer, and add a softmax layer for classification in the last layer. Then, initialize the parameters randomly and set the loss function as follows: Iteratively calculate the loss function and update the parameters for training. Save the trained network to the scene recognition module of the scheduling agent.

[0095] Step S103, when performing rescheduling scene recognition, the real-time disturbance state information is input into the trained model, and the rescheduling scene category is output. The results of the comparison between the GRU rescheduling recognition results of this embodiment and the general classification method KNN are shown in Table 2. The deep learning network constructed in this embodiment has a higher accuracy than the general classification algorithm and is more suitable for the classification of disturbance events in aircraft pulsation assembly.

[0096] Table 2 Comparison of accuracy between GRU method and KNN method

[0097] GRU KNN Test set accuracy 0.90 0.77 Validation set accuracy 0.80 0.54

[0098] Step S2, the disturbance handling method under rescheduling scenario 1 is implemented, that is, the implementation of right-shift rescheduling. Assume that a disturbance occurs at 187h, and workers 3 and 5 of team 8 at station 1 are absent until 237h. At this time, right-shift rescheduling is adopted, and the actual completion time of the team is moved to 246.4h, but it does not exceed the beat, so right-shift rescheduling can be used to complete the rescheduling.

[0099] Step S3, such as Figure 3 , the disturbance handling method under rescheduling scenario 2 is realized, that is, the realization of intra-station collaborative rescheduling based on NSGA-II. At 350h, worker 2 of team 1 at station 2 leaves his post, and the processes that need to be scheduled include: [712 822 786750 765 763 688 800 828 760 715 824 795 731 713 808 816 719 820 722 852 779825 739 771 735]. The specific process of intra-station collaborative rescheduling based on NSGA-II is as follows:

[0100] Step S301, population initialization. First, the scene recognition part of the scheduling agent outputs the adjustment results within the station, and the collaborative rescheduling module within the station initializes the population of the NSGA-II algorithm. The chromosome encoding is improved by using a combined rule method. Each chromosome encoding w represents the weight of the adopted rule. [w1, w2, w3] belongs to the rule weight of the task agent, [w4, w5, w6] belongs to the rule weight of the worker agent, and w i ∈(0,1). In the combined scheduling rule, the priority CP of the scheduling object o is o The mathematical expression of can be summarized as follows, P ow The priority of the scheduling object o calculated by a single rule.

[0101]

[0102] The actions based on the combined scheduling rules can be expressed as: Each value is the weight of the corresponding single scheduling rule, W i is the number of scheduling rules, represents the weight of the task agent rule w, represents the weight of the worker agent rule w.

[0103] Step S302, input the population into the task agent and the worker agent for decoding, and calculate the optimization target value. The task agent divides the process set, checks all the assembly process tasks of the team where the disturbed worker is located, and divides them into three process sets according to the different disturbance conditions, namely, the completed process set, the assembly process set and the to-be-assembled process set. There is no need to adjust the completed process, and the original scheduling order of the assembly process set is maintained. For the to-be-assembled process set, it is rescheduled based on the rules output by NSGA-II. After that, the population output by the scheduling agent is decoded, that is, the process with the highest priority is calculated according to the task sorting rules, and the process is input to the worker agent, and then the to-be-assembled process set is updated; the worker agent then calculates the worker with the highest priority according to the allocation rules in the chromosome, and updates the worker status. After all the process allocations are completed, two optimization target values ​​are calculated. Considering the optimization goals of production performance and stability, that is, it is necessary to ensure that the performance after rescheduling is optimized, and to minimize the difference between the changed scheduling plan and the original scheduling plan to avoid excessive costs in resource reallocation, two types of optimization goals are designed as follows:

[0104] Optimization target 1 is the actual completion time f1 of the team, which is defined as follows: L For the process that needs to be rescheduled:

[0105] f1=min(t O ),o∈O L

[0106] Optimization goal 2 is the number of worker process changes f2, which is defined as follows: zbw represents the process set of worker w in team b at station z, O ′ zbw represents the worker's process set after adjustment, and f2 represents the change in the worker's assembly process set before and after adjustment:

[0107]

[0108] Step S303, iterative optimization. After fast non-dominated sorting, selection, crossover, and mutation, a subgroup is generated, and then the parent generation and the child generation are merged, and a new parent generation is generated through fast non-dominated sorting and crowding calculation. Among them, the selection operator is based on the binary tournament strategy, selecting individuals with small non-dominated order and large crowding distance; the crossover operator randomly generates two crossover points on the two parent chromosomes for crossover in each round; the mutation operator sets the mutation distribution parameter μ, generates a random number ε in each round, and mutates the parent generation according to the polynomial mutation. The mutation of each chromosome is shown in the following formula

[0109]

[0110] Step S304, comprehensively considering stability and efficiency, select the final rescheduling solution from the Pareto front solution set of the last generation population. The optimization target values ​​are shown in Table 3. Comprehensively considering stability and assembly efficiency, select the rule corresponding to f1=449.1, f2=155 as the final solution. At this time, the weights of the two types of rules after adjustment are [0.41, 0.31, 0.10, 0.16, 0, 0.16]. The final scheduling result still does not exceed the beat (560h), which proves the effectiveness of the scheduling.

[0111] Table 3 Rescheduling scenario 2 - optimization target value

[0112] f1 f2 453.7 146 452.7 148 449.4 149 449.1 155 449.1 159

[0113] Step S4, such as Figure 4 , the disturbance under rescheduling scenario 3 is realized, that is, the inter-station negotiation rescheduling method based on the improved CNP is realized. Design a disturbance event, that is, assume that 2 workers from station 1 team 7 leave their posts 250 hours after a completion cycle, namely worker 2 and worker 3. The remaining process set of worker 2 is: [554, 639, 649], and the remaining process set of worker 3 is [557, 639, 649]. The rescheduling scenario is identified as scenario 3. At this time, it is necessary to adjust the resources between stations through the improved CNP inter-station negotiation rescheduling method and negotiate with other stations through bidding. The specific process is as follows:

[0114] Step S401, set a negotiation resource lock, when the station agent i When the rescheduling scenario identification module outputs rescheduling scenario 3, the negotiation module first checks whether the negotiation resource lock is accessible. If not, the resource is suspended and waited; if accessible, it enters 2) and locks the negotiation resource.

[0115] Step S402, station agent i The negotiation module and the task agent and worker agent calculate the specific information of the disturbed station that needs assistance, including the number of assists and the assistance time, and obtain several sets of rescheduling solutions based on the NSGA-II algorithm. Considering the minimization of the completion time of the bidding station and the minimization of the dispatch time of other stations, the objective function of the NSGA-II algorithm is designed to minimize the actual completion time f1 of the station and minimize the time f2 of the dispatched workers of the bidding station. In the chromosome decoding calculation of the worker agent, considering the role of the dispatched workers, it is necessary to set up virtual workers: if NoW workers are absent, NoW virtual workers are set.

[0116] Step S403, sort these scheduling solutions according to certain preferences, and design tender documents for these solutions in turn according to the sorting order, and broadcast the tender documents corresponding to the solution ranked first to other stations. When selecting two parent chromosomes based on the binary tournament strategy using a selection operator with preferences, the optimization objectives f2 of the two chromosomes are preferentially compared. Sort the Pareto front solutions in ascending order according to the sum of the assembly time of workers dispatched from external stations (more than one worker may need to be dispatched externally), and prepare tender documents for each set of solutions. The tender document corresponding to each set of solutions is a NoW dimension, and each dimension contains a tender information, expressed as a five-tuple. <Z 0 ,B,W 0 ,St,Ft> indicates that Z 0 represents the bidding station number, B represents the team number to be dispatched, that is, the team number of the absent worker; W 0 Indicates the number of the virtual worker to be dispatched, St represents the worker W 0 The expected start time of dispatch, Ft represents worker W 0 Estimated end time of dispatch. As shown in Table 4, in order to reduce the dispatch cost of other stations, the assembly rule corresponding to the solution of f1 = 280, f2 = 18.6 is taken, and the tender documents [[1,7,1,268,277],[1,7,2,268,273]] are calculated.

[0117] Table 4. Tender preparation - Pareto frontier solution

[0118] f1 f2 268.8 36.8 280.0 18.6

[0119] Step S404, station agent j The scheduling agent information processing module receives the information from the agent i The tender is matched based on the free time and then sent back to the agent i After receiving the tender, other stations will query the current status of the team workers and prepare the tender with the information of the workers who meet the requirements of the tender. Each tender is represented by a six-tuple <Z 0 ,Z,Agr,W 0 ,W,Effi>:Z 0 represents the bidding position, Z represents the tendering position; W 0 is the virtual worker number of the tender document, W is the worker number of the bidding station that meets the requirements of the tender document; Agr represents whether the bidding station agrees to bid. The bidding station responds to each piece of tender information〈Z 0 ,B,W 0,St,Ft〉, traverse the states of all workers in team B. If there is no idle worker in [St,Ft], Agr is 0, and W and Effi information are directly empty; if there is a worker who meets the bidding information, Agr is 1. W and Effi are the worker numbers and efficiencies that meet the requirements. The returned bids are shown in Table 5.

[0120] Table 5 Bidding Preparation

[0121] <![CDATA[Z 0 ]]> Z Agr <![CDATA[W 0 ]]> W Effi 1 2 0 0 0 0 1 3 1 1 3 0.8 1 3 1 2 4 0.9 1 4 1 1 3 0.8 1 4 1 2 4 0.9 1 5 1 1 1 0.8 1 5 1 2 3 1

[0122] Step S405, station agent i The dispatch agent information processing module receives bids from other stations and inputs them into the negotiation module. The negotiation module selects the appropriate station for bidding based on the worker combination efficiency score and notifies it, signs a contract with it, dispatches workers to the corresponding bidding station, terminates this negotiation, and releases the negotiation resource lock. According to the traversal principle, the two most efficient workers in the bid are selected for calculation. The final completion time is 280.0h. The first two bids with the highest efficiency, namely worker 4 at station 3 and worker 4 at station 4, are selected.

[0123] Step S406, start global network training, retrain and generate assembly rules for all stations, and release the negotiation resource lock.

[0124] In summary, this method adopts a distributed scheduling method, and each station can make decisions and execute independently, which reduces the communication burden and improves scheduling efficiency. At the same time, unlike the existing static scheduling, the present invention is based on the GRU neural network model, which can accurately identify the impact of disturbance events on the pulsation rhythm and classify them, preparing for subsequent rescheduling. The improved CNP algorithm and NSGA-II algorithm are used to realize the dynamic adjustment of resources between stations and the optimization of process sorting and worker allocation within the station, improve the coordination efficiency, and achieve global optimization.

[0125] Example 2

[0126] Based on Example 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the distributed rescheduling method of the aircraft pulse assembly line based on CNP-NSGA-II as described in Example 1.

[0127] like Figure 5 As mentioned above, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1Of course, in addition to the software implementation, the present invention does not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0128] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0129] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0130] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II, characterized in that: The steps include: Obtain real-time disturbance state information and use the trained gated recurrent unit-based scene recognition model to obtain the rescheduling scene type; In response to the rescheduling scenario type being the first type, rescheduling the aircraft pulse assembly line to the right; In response to the rescheduling scenario type being the second type, constructing a non-dominated sorting genetic algorithm II to optimize the assembly process sorting and / or assembly worker allocation of the aircraft pulse assembly line stations, so as to achieve intra-station collaborative scheduling of the aircraft pulse assembly line; In response to the rescheduling scenario type being the third type, a contract network algorithm is used to generate bidding information including human resources that require assistance from other stations, optimize the allocation of assembly workers at the stations, and achieve coordinated scheduling between stations on the aircraft pulse assembly line.

2. The distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II according to claim 1, characterized in that: The training process of the scene recognition model includes the following steps: By simulating the rescheduling scenario, the disturbance state information under different rescheduling scenarios is obtained, the corresponding labels are obtained by marking, the disturbance state information is normalized, and training samples are constructed; Construct a densely connected network with sequential connections, based on a gated recurrent neural network and a normalized exponential function layer as a scene recognition model; The scene recognition model is trained using the training samples.

3. The distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II according to claim 1, characterized in that: The method of optimizing the process sorting and / or worker allocation of the aircraft pulse assembly line station by constructing a non-dominated sorting genetic algorithm II comprises the following steps: In response to the rescheduling scenario type being the second type, for the stations of the aircraft pulse assembly line, the rule weights corresponding to the assembly process optimization and the assembly worker allocation are used as chromosome encoding to initialize the population; Through contract collaboration decoding, the optimization target values ​​of the actual completion time of the station and the time for dispatching workers to the bidding station are calculated. Subgroups are generated through fast non-dominated sorting, selection, crossover and mutation. The parent group and the subgroup are merged to minimize the actual completion time of the station and the time for dispatching workers to the bidding station. A new parent generation is generated through fast non-dominated sorting and congestion calculation. This step is repeated until the preset iteration round is reached. Based on the Pareto front solution set of the iterated population, a rescheduling plan is obtained to determine whether the rescheduling plan exceeds the beat. If so, the scheduling process corresponding to the third type of rescheduling scenario type is executed. If not, the collaborative scheduling within the station of the aircraft pulse assembly line is realized based on the rescheduling plan.

4. The distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II according to claim 3 is characterized in that: The corresponding rules for assembly process optimization include: the more subsequent processes the current process has, the higher the priority; the longer the sum of the time of the subsequent processes of the current process, the higher the priority; and the longer the time from the current process to the process with no subsequent processes, the higher the priority.

5. The distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II according to claim 3 is characterized in that: The corresponding rules for allocating assembly workers include: workers with shorter actual assembly time have higher priority; workers with shorter completion time have higher priority; and workers with higher assembly efficiency have higher priority.

6. The distributed rescheduling method for an aircraft pulse assembly line based on CNP-NSGA-II according to claim 3 is characterized in that: The process of calculating the optimization target value of the rule weight includes the following steps: For all assembly processes of the disturbing worker's team, they are divided into completed process set, assembly process set and to-be-assembled process set according to the disturbance; For each assembly process in the set of to-be-assembled processes, based on the current population, the assembly process with the highest priority is calculated according to the rules corresponding to the assembly process optimization, the worker with the highest priority is calculated according to the rules corresponding to the assembly worker allocation, and the assembly process set and the worker status are updated; After all assembly processes of the disturbing worker's team are updated, the optimization target value is calculated.

7. The distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II according to claim 1, characterized in that: The process of generating bidding information including human resources requiring assistance from other stations through the contract network algorithm and optimizing the allocation of assembly workers to the stations includes the following steps: Determine whether the negotiation resource lock in the current booth is accessible. If not, suspend and wait, and repeat this step until the negotiation resource lock is accessible; In response to the negotiation resource lock being accessible, locking the negotiation resource lock; Obtain multiple sets of candidate scheduling solutions for process sequencing and assembly worker allocation through the NSGA-II algorithm; The multiple groups of candidate scheduling solutions are sorted, and the bidding information corresponding to the candidate scheduling solutions is constructed according to the sorting, and the bidding information corresponding to the candidate scheduling solutions with higher sorting is sent to other stations of the machine pulse assembly line. If other stations all refuse to bid, the bidding information corresponding to the next candidate scheduling solution is sent to other stations of the machine pulse assembly line, and the bidding is repeated until the information accepting the bidding is received or the negotiation rounds reach a preset upper limit, and the negotiation resource lock is released.

8. The distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II according to claim 7 is characterized in that: In response to the negotiation rounds reaching a preset upper limit, the assembly process optimization and assembly worker allocation corresponding rules of all stations of the machine pulse assembly line are retrained.

9. The distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II according to claim 7, characterized in that: The bidding information includes the bidding station, the tendering station, the bidding virtual worker number and the virtual worker number in the tendering station that meets the requirements of the bidding document.

10. The distributed rescheduling method for aircraft pulse assembly line based on CNP-NSGA-II according to claim 1, characterized in that: The aircraft pulse assembly line includes multiple stations, and a station agent for distributed decision-making is constructed for each station. Each station agent includes a scheduling agent, and a task agent and a worker agent that respectively optimize the assembly process and allocate assembly workers under the guidance of the scheduling agent. The distributed rescheduling method is applied to each station agent.

Images