A generative dynamic production resource scheduling and balancing method and system

Abstract

The present invention belongs to the technical field of production scheduling optimization, and discloses a generative dynamic production resource scheduling balancing method and system. The method uses a linear programming algorithm to abstract the daily scheduling results into the working hours to be completed for each process task every day, and then calls the solver to solve the scheduling completion hours of the day in sequence, and obtains the objective function of the scheduling plan, until the day, the sum of the completion hours of each process is the rated working hours corresponding to the process; introduces a hyperparameter coefficient into the objective function, and outputs the hyperparameter coefficient value through deep neural network model training and learning, and brings it into the objective function of the linear programming algorithm to obtain the same scheduling result as the planner; the deep neural network model obtained after training simulates the planner to make a correct balanced scheduling plan without manual intervention. The present invention can effectively reduce labor costs, thereby further realizing intelligent scheduling.

Description

Technical Field

[0001] The present invention belongs to the technical field of production scheduling optimization, and in particular relates to a generative dynamic production resource scheduling balancing method and system. Background Art

[0002] Production scheduling optimization has always been a key issue in the manufacturing industry. The traditional strategy based on time minimization produces scheduling results in an overly mechanized manner, and these optimal scheduling plans are scheduling results under ideal conditions. The actual production may deviate from the ideal conditions. In the large-scale discrete manufacturing industry, especially in the field of rail transit assembly, the assembly process is complex, large-scale, and long working hours. Compared with automated production and manufacturing, it relies more on the role of people in production. In addition, there are often unstable orders, which leads to some production personnel having no work plans after intensive work, and their working time and idle time are extremely unbalanced. The result of manual scheduling / adjustment by the planner is that the planner fully considers various production factors and balances the interests of all parties to obtain a balanced scheduling result. If this behavior pattern of the planner can be learned and solidified, then the ideal balanced scheduling result of production resources will be obtained during automatic scheduling, thereby achieving the ultimate goal of intelligent scheduling and intelligent manufacturing. Summary of the invention

[0003] In order to overcome the problems existing in the related art, the disclosed embodiment of the present invention provides a generative dynamic production resource scheduling and balancing method and system, and specifically relates to a dynamic production resource balancing algorithm based on a generative model. The method proposed by the present invention preliminarily obtains the ideal scheduling result through linear programming, and then uses the ideal result as input to call the machine learning model to obtain the final production resource balancing scheduling result. This model is trained by learning the historical scheduling results of the planner, and learns the production adjustment behavior pattern of the planner, which is equivalent to solidifying the planner's adjustment / scheduling behavior into an algorithm model.

[0004] The technical solution is as follows: a generative dynamic production resource scheduling balancing method, which uses linear programming combined with machine learning production adjustment behavior, adjusts the production scheduling plan through a deep neural network model after learning and training, and outputs a scheduling matrix that meets the production resource balance, so that production efficiency, cost control and resource allocation are balanced; specifically, it includes the following steps:

[0005] S1, according to the constraints between production resources and production resources, the set of process tasks that need to be started is obtained according to the order and BOM, the capacity list of the position is obtained according to the position situation of the factory, and the daily material distribution matrix is ​​obtained according to the material distribution plan; based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production completion time of the day is calculated to obtain the objective function of the production schedule until the The sum of the completion time of each process is the rated working time corresponding to the process.

[0006] S2, introduces hyperparameter coefficients into the objective function, outputs the hyperparameter coefficient values ​​through deep neural network model training and learning, and brings them into the objective function of the linear programming algorithm to obtain the production scheduling results;

[0007] S3, in the deep neural network model, by expanding the loss function calculation, additionally calling the linear programming algorithm to obtain the predicted value, the output of the deep neural network model is adjusted by the linear programming algorithm and then compared with the teacher value to obtain the loss value, so that the trained deep neural network model can make a balanced production schedule without human intervention.

[0008] In step S1, the production resources include: process, process task, station, station capacity, material, production rhythm, and the material tree Bom of the complete product;

[0009] The constraint relationship includes: the required machine capacity of the process is less than or equal to the machine capacity of the day, the required materials are less than or equal to the planned material delivery quantity of the day, and the start time of the process task in the production rhythm cycle is not less than the start time of the cycle in which the production rhythm is located;

[0010] The set of process tasks required to start is This collection includes work orders added according to the production rhythm, with the required capacity, required materials, process sequence, and earliest start time information of the process to which it belongs attached;

[0011] The capacity list of the stations obtained according to the station situation of the factory is as follows: ;

[0012] The material distribution matrix obtained every day according to the material distribution plan is: , For the number of days, is the material number, The quantity of the material.

[0013] In step S1, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production completion time of the day is calculated, and the objective function of the production schedule is obtained, including:

[0014] The daily production scheduling results are abstracted into the working hours required to complete each process task every day. The completion time on the first day is , subscript 1 corresponds to the current first day, Corresponding process task number, there are The eight process tasks, after calling the solver, get the scheduling results as follows:

[0015] ;

[0016] In the formula, The number of days for production scheduling, is the process task number, The value is the working hours of a certain process on that day. ; for The minimum effective value of

[0017] The production scheduling target of the production scheduling result is to find the maximum value of the sum of the daily work completion volume. The target formula is:

[0018] ;

[0019] In the formula, The maximum value of the sum of the completed work.

[0020] Furthermore, when the required station capacity of the process is less than or equal to the station capacity of the day, the capacity constraints on the first day need to be met, including: the constraints of the process itself and the process with materials available for start-up;

[0021] The process itself has constraints, including:

[0022] Seats The processes that can be completed in one day's production capacity, ;

[0023] Seats The processes that can be completed in one day's production capacity, ;

[0024] Seats The processes that can be completed in one day's production capacity, ;

[0025] Seats The processes that can be completed in one day's production capacity, ;

[0026] The processes for which the materials can be used to start production include:

[0027] materials Available processes to start, ;

[0028] materials Available processes to start, ;

[0029] materials Available processes to start, ;

[0030] materials Available processes to start, ;

[0031] materials Available processes to start, ;

[0032] The start-up percentage of each process task cannot exceed the rated working hours, including:

[0033] ;

[0034] in, Represents the 1st to 8th processes of a day;

[0035] Process Tasks The subsequent process tasks are tasks that start one week after the start of production according to the production rhythm and cannot be started on the first day, including: ;

[0036] Process Tasks Waiting for process tasks Construction starts after completion, including:

[0037] ;

[0038] In the formula, Is it a process task? After the task, For the Working hours for the first process per day;

[0039] Translates to:

[0040] ;

[0041] In the formula, For the front The day cannot occupy working hours;

[0042] In the Day, before Daily process tasks The sum of the completion percentages is equal to 100%;

[0043] Input the above constraints and objective formulas into the solver to obtain the planning results for the first day.

[0044] Furthermore, on the second day, the target formula for the evolution of the production target working hours over time is:

[0045] ;

[0046] In the formula, The completion time for the second day;

[0047] The constraints on function scheduling must be met, including the following constraints:

[0048] Table capacity constraints: ;

[0049] Material supply constraints: ;

[0050] Working hours constraints: ;in, It indicates the 1st to 8th processes on the second day;

[0051] The completed working hours on the second day are less than or equal to the required rated working hours minus the working hours completed previously;

[0052] Find the first The result of the scheduled completion time of the day is , The number of days for production scheduling, Number the process task.

[0053] Until Days, the sum of the completion hours of each process is the rated working hours corresponding to the process, satisfying:

[0054] ;

[0055] In the formula, The total number of days for production scheduling. The process number is The process tasks, For the The daily process number is The production hours of process tasks are scheduled. The process number is The sum of all scheduled working hours for the process.

[0056] In step S2, the hyperparameter coefficient is introduced into the objective function, and the expression is:

[0057] ;

[0058] In the formula, is the first Day process number The working time parameters, and The production schedule hours are multiplied to achieve the purpose of using the coefficient adjustment result.

[0059] In step S2, the hyperparameter coefficient value is output through deep neural network model training and learning, and is brought into the objective function of the linear programming algorithm to obtain the production scheduling result; including: according to the production scheduling result Initialize a coefficient matrix , the initialization data is all 1, As an input data, standardize the number of rows and columns of the output result; split the model input layer data into time, elements and quantity, unify them into vector representations with three dimensions and merge them together; the specific steps are:

[0060] Step 1: Feature encoding. The time dimension is represented by 24 hours and encoded as a 24-dimensional one-hot encoding vector. The element dimension uses one-hot encoding to represent each element as a vector, and the original value of the quantity dimension is used.

[0061] Step 2: Vector fusion: For each element dimension, multiply the unique hot encoding vector of each element by the number of elements to obtain the weighted vector of the element; add the weighted vectors of all elements to obtain the final vector representation of the dimension; then normalize, using the L2 normalization method to normalize each dimension vector;

[0062] Work order basic information matrix input: The work order basic information is used as a constraint description for other materials, production capacity, and worker scheduling basic necessary conditions, and participates in the calculation of all feature extraction neurons in the first layer of the hidden layer;

[0063] Input of the station capacity matrix: During feature extraction, the basic information of the work order is converted into a matrix of work orders corresponding to stations;

[0064] Worker scheduling matrix input: the expected schedule of workers in the future, including the team to which the worker belongs, the type of work order that can be started, the off-duty situation, the available working hours, and the introduction of activation function adjustment;

[0065] Material supply matrix input: processed as a two-dimensional matrix of the supply quantity of each material per day.

[0066] In step S3, the deep neural network model consists of a fully connected layer and an activation layer. The fully connected layer is used to memorize the deviation characteristics between the ideal linear production scheduling and the actual production scheduling, and the activation layer is used to bring nonlinear characteristics.

[0067] The loss function uses the mean square error function MSE to measure the average of the squared difference between the predicted value and the true value, and distinguishes positive and negative errors for outliers;

[0068] The loss value calculated by the loss function uses the gradient descent algorithm to update the neural network parameters.

[0069] Another object of the present invention is to provide a generative dynamic production resource scheduling and balancing system, which is implemented by the generative dynamic production resource scheduling and balancing method, and the system includes:

[0070] The production scheduling objective function acquisition module is used to obtain the set of process tasks that need to be started according to the constraints between production resources and production resources, orders and BOM, obtain the capacity list of the positions according to the position situation of the factory, and obtain the daily material distribution matrix according to the material distribution plan; based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production completion time of the day is calculated to obtain the objective function of the production schedule until the The sum of the completion time of each process is the rated working time corresponding to the process.

[0071] The deep neural network model adjustment module is used to introduce hyperparameter coefficients into the objective function. Through deep neural network model training and learning, the hyperparameter coefficient values ​​are output and brought into the objective function of the linear programming algorithm to obtain the production adjustment and scheduling results.

[0072] In the deep neural network model training module, it is used to expand the loss function calculation in the deep neural network model, additionally call the linear programming algorithm to obtain the prediction value, and adjust the output of the deep neural network model through the linear programming algorithm and then compare it with the teacher value to obtain the loss value, so that the trained deep neural network model can make a balanced production schedule without human intervention.

[0073] Combining all the above-mentioned technical solutions, the beneficial effects of the present invention are as follows: the present invention solves the limitations faced by the production scheduling scheme of the assembly production plan of the rail transit industry in the actual production environment, especially the situation where the employee working time is too intensive or sparse due to the neglect of the balanced distribution of the plan. An innovative generative dynamic model is proposed, which not only considers time efficiency, but also comprehensively considers multi-dimensional factors such as the capacity limitation of the unit, the rhythm of material supply, and the distribution of human working hours to achieve the optimal balance of production resources. The core is to use linear programming combined with machine learning to learn the production adjustment behavior of the planner, and intelligently adjust the production schedule through the model after learning and training, and output a production scheduling matrix that meets the balance of production resources, ensuring a win-win situation of production efficiency, cost control and balanced resource allocation.

[0074] The present invention improves the efficiency of enterprise scheduling and production scheduling based on production resources, greatly improves the precision of production resource utilization, thereby avoiding waste of production resources due to scheduling and reducing enterprise production costs; on the other hand, refined scheduling places very high demands on the enterprise's production planners. This method combines standardized algorithms and deep network models to enable enterprises that do not have senior schedulers to achieve professional and refined scheduling, empowering enterprises to address key links in large-scale production.

[0075] Enterprise production scheduling often faces the difficulty of "plans cannot keep up with changes", which is mainly due to the narrow dimensions that traditional methods can handle and the rigidity of algorithms. This paper solves this fundamental problem by deeply exploring the essence of production scheduling and proposing a standardized algorithm, and then combining a deep neural network model to provide scenario generalization capabilities.

[0076] Many companies and individuals rely on past experience and often use inherent templates, refer to inherent dimensional data, or use unchanging algorithms in production scheduling. This can bring them a sense of security of expected results, and they are also biased against technologies such as neural networks that have generalization characteristics. However, corporate production activities are affected by many factors, such as rapidly changing market demand or the rapid update and iteration of the company itself. Technologies that stick to a sense of security are often difficult to adapt to environmental changes. This method is based on the essential scheduling algorithm and the generalization ability of neural networks, which keeps the basic logic of scheduling unchanged and can fully adapt to the changing scenario requirements, overcoming people's technical prejudices against the risks that changes may bring. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0078] Figure 1 It is a flow chart of a generative dynamic production resource scheduling and balancing method provided by an embodiment of the present invention;

[0079] Figure 2 It is a linear planning production scheduling flow chart provided by an embodiment of the present invention;

[0080] Figure 3 This is a flow chart of the generative AI production adjustment provided by an embodiment of the present invention;

[0081] Figure 4 It is a visualization diagram of the production scheduling results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0083] The innovation of this invention is that, on the one hand, the invention invents a standard linear programming algorithm by deeply exploring the essence of the production scheduling business logic, ensuring that the foundation of the production scheduling algorithm is solid and retaining the flexibility of repetition for algorithm expansion; on the other hand, the generalization ability of the deep learning network model provides the algorithm with the ability to adapt to scene changes, and the big data of the model can fully refine the production scheduling granularity. The combination of the two solves the problems of the rigidity, roughness and difficulty of dynamic planning of traditional production scheduling algorithms.

[0084] Embodiment 1, Figure 1 The embodiment of the present invention provides a generative dynamic production resource scheduling and balancing method, comprising:

[0085] S1, according to the constraints between production resources and production resources, the set of process tasks that need to be started is obtained according to the order and BOM, the capacity list of the position is obtained according to the position situation of the factory, and the daily material distribution matrix is ​​obtained according to the material distribution plan; based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production completion time of the day is calculated to obtain the objective function of the production schedule until the The sum of the completion time of each process is the rated working time corresponding to the process.

[0086] S2, introduces hyperparameter coefficients into the objective function, outputs the hyperparameter coefficient values ​​through deep neural network model training and learning, and brings them into the objective function of the linear programming algorithm to obtain the production scheduling results;

[0087] S3, in the deep neural network model, by expanding the loss function calculation, additionally calling the linear programming algorithm to obtain the predicted value, the output of the deep neural network model is adjusted by the linear programming algorithm and then compared with the teacher value to obtain the loss value, so that the trained deep neural network model can make a balanced production schedule without human intervention.

[0088] Embodiment 2: The generative dynamic production resource scheduling and balancing method provided in the embodiment of the present invention is further described in detail below with reference to the accompanying drawings.

[0089] 1. Optimal production scheduling algorithm.

[0090] First, the optimal production scheduling result in a mathematical sense is obtained by finding the shortest path. The present invention uses linear programming to solve this optimization problem. In the rail transit industry, the production scheduling plan is generally the arrangement of work orders for each day within a time period. After being abstracted into a mathematical model, it is expressed as the completion ratio of a certain work order on a certain day. The completion ratio of each work order in a day represents the production scheduling plan for that day. If the completion ratio of each day is the largest, it means that the production efficiency is the highest, and the denser the production scheduling plan, the better the optimal production scheduling result required by the present invention. In linear programming, the optimal problem required by the present invention becomes the sum of the maximum completion ratios of each day.

[0091] 1.1, production resources.

[0092] In the rail transit industry, the basic production factors of users can be summarized as: people (manpower), machines (machines and equipment), materials (materials), methods (process methods), and environment (production environment).

[0093] When scheduling production, you need to fully consider these production factors and production resources. Since they are production resources, production must be carried out within the scope of the resources provided. You cannot produce without equipment or plan to spray two buckets of car paint when only one bucket of car paint is provided.

[0094] The following is the production resource definition:

[0095] Process: The smallest operating unit in the production process that cannot be further decomposed.

[0096] Process task: a concrete task of a process.

[0097] Position: The work location where production process tasks are performed. Generally speaking, one position can only correspond to the production of one process.

[0098] Station capacity: For a process, how many steps can this station complete in one day.

[0099] Materials: generally refers to input materials, material resources required to start process tasks.

[0100] Production rhythm: determines the rhythm of the production process, for example, after production starts, a new carriage will be produced every other week.

[0101] Bom: A complete product material tree.

[0102] 1.2, Constraint relationship.

[0103] The purpose of optimal production scheduling is to produce the most products with limited production resources on a micro level. On a macro level, it is to shorten the time from the start to the end of production to the shortest possible level and maximize production efficiency.

[0104] In a day's production plan, the available capacity is limited, so the capacity required by the scheduled process must be less than or equal to the capacity of the day. The materials required by the scheduled process must be less than or equal to the planned material delivery quantity for the day. Due to the existence of production rhythm, the start time of the process task in the rhythm cycle cannot be less than the start time of the cycle.

[0105] 1.3, data preparation.

[0106] First, according to the order and BOM, we get the set of process tasks (work orders) that need to be started. i ,i∈{1,2,3...}, this set contains all the work orders of an order, including the work orders added according to the production rhythm, with the required capacity, required materials, process sequence, earliest start time and other information of the process. The earliest start time of the work order put into production according to the production rhythm cannot be earlier than its production start time.

[0107] The capacity list of the stations obtained according to the station situation of the factory is as follows: ;

[0108] The material distribution matrix obtained every day according to the material distribution plan is: , For the number of days, is the material number, The quantity of the material.

[0109] 1.4, Linear programming algorithm.

[0110] Based on the above production resources and their mutual constraints, the following data tables 1-3 are obtained.

[0111] Table 1 Correspondence table of production resources required for process tasks:

[0112]

[0113] Table 2 Daily station capacity:

[0114]

[0115] Table 3 Daily arrival quantity:

[0116]

[0117] In this example, the materials are simplified into a one-dimensional matrix, representing the unchanged daily material distribution plan.

[0118] Using linear programming, the daily production scheduling results are abstracted into the working hours required for each process task every day. The completion time on the first day is , the first subscript corresponds to the current first day, Corresponding to the process task number, there are a total of , eight process tasks, so The array length is 8. After calling the solver, the final production scheduling result expected by the present invention is:

[0119] ;

[0120] In the formula, The number of days for production scheduling, is the process task number, The value is the working hours of a certain process on that day. ; for The minimum effective value of

[0121] The production scheduling goal is to produce as much as possible within a day and complete enough work orders. The goal converted to linear programming is to find the maximum sum of the daily work completion, which is the goal formula used to define the production scheduling task goal:

[0122] ;

[0123] In the formula, The maximum value of the sum of the completed work.

[0124] On the first day, capacity constraints need to be met, including: process constraints and the availability of materials for the process to start;

[0125] The process itself has constraints, including:

[0126] Seats The processes that can be completed in one day's production capacity, ;

[0127] Seats The processes that can be completed in one day's production capacity, ;

[0128] Seats The processes that can be completed in one day's production capacity, ;

[0129] Seats The processes that can be completed in one day's production capacity, ;

[0130] The processes for which the materials can be started include:

[0131] materials Available processes to start, ;

[0132] materials Available processes to start, ;

[0133] materials Available processes to start, ;

[0134] materials Available processes to start, ;

[0135] materials Available processes to start, ;

[0136] By material For example, the process that can be started is 5. The quantity of materials required to complete the work order. is the proportion of processes completed on the day, In order to calculate the material quantity required for the work order according to the completion ratio of the day, The work order also requires the same material, so they are combined together as constraints.

[0137] The start-up percentage of each process task cannot exceed the rated working hours, including:

[0138] ;

[0139] in, Represents the 1st to 8th processes of a day;

[0140] Process Tasks The subsequent process tasks are tasks that start one week after the start of production according to the production rhythm and cannot be started on the first day, including: ;

[0141] Process Tasks Waiting for process tasks Construction starts after completion, including:

[0142] ;

[0143] In the formula, Is it a process task? After the task, For the Working hours for the first process per day;

[0144] Translates to:

[0145] ;

[0146] In the formula, For the front The day cannot occupy working hours;

[0147] In the Day, before Daily process tasks The sum of the completion percentages is equal to 100%;

[0148] Input the above constraints and objective formulas into the solver to obtain the planning results for the first day.

[0149] if To complete it on the same day, you need to meet Sufficient working time to complete after completion The proportion of work completed as scheduled. If there are only 8 working hours per day, then .

[0150] By inputting the above constraints and objective formulas into the solver, we can get the planning results for the first day.

[0151] On the second day, the target formula used to describe the evolution of the scheduling target hours over time is:

[0152] ;

[0153] In the formula, The completion time for the second day;

[0154] The present invention innovatively proposes that the constraints imposed on the function scheduling must be met, including the following constraints:

[0155] Table capacity constraints: ;

[0156] Material supply constraints: ;

[0157] Working hours constraints: ;in, It indicates the 1st to 8th processes on the second day;

[0158] The completed working hours on the second day must be less than or equal to the required rated working hours minus the working hours completed previously.

[0159] Find the first The result of the scheduled completion time of the day is , The number of days for production scheduling, Number the process task, forming a ( OK Until The sum of the completion time of each process is the rated working time corresponding to the process. That is, it satisfies:

[0160] ;

[0161] In the formula, The total number of days for production scheduling. The process number is The process tasks, For the The daily process number is The production hours of process tasks are scheduled. The process number is The sum of all scheduled working hours for the process.

[0162] The present invention innovatively proposes formula (5), which can make the process scheduling time and rated time ultimately consistent;

[0163] 1.5, hyperparameter. According to the process sequence number in the process route, the present invention can obtain the order relationship between the processes. In order to obtain a production schedule that conforms to the process sequence, the present invention introduces coefficients into the objective function. Each work order corresponds to a separate coefficient:

[0164] ;

[0165] In the formula, is the first Day process number The working time parameters, and The production schedule hours are multiplied to achieve the purpose of using the coefficient adjustment result.

[0166] Formula (6) is used to describe the deformation after adding the coefficient on the basis of production scheduling.

[0167] It can be known from basic mathematical knowledge that the higher the completion ratio assigned to the work order with the higher coefficient, the higher the value of this objective function, and the final solution of the optimization problem will change accordingly due to the increase or decrease of this coefficient. By adjusting this coefficient, the present invention can adjust the entire production scheduling result. When this coefficient is adjusted to the most appropriate, a production scheduling result similar to that of the planner can be achieved.

[0168] In this example, the initialization coefficient is set to 1, and the coefficient is then brought into the calculation after the subsequent generative AI generates the coefficient matrix.

[0169] When adjusting the coefficients, the objective function of each day corresponds to a separate one-dimensional coefficient vector, so the coefficient of the entire production scheduling plan is a matrix with the same number of rows and columns as the production scheduling plan result matrix.

[0170] 1.6, Linear programming production planning process, such as Figure 2 shown.

[0171] This process explains how to use linear programming to perform preliminary production scheduling and produce the optimal production scheduling result in a mathematical sense. This process produces a production schedule every day until formula (5) is satisfied and production scheduling is terminated. Specifically, it includes:

[0172] First, prepare the production factor data, which can be considered as the context of the entire production scheduling activity; then calculate the daily complete working hours based on factors such as production capacity and material supply, that is, the production capacity data measured from the perspective of working hours; next, call the solver based on production factors and capacity to get the production schedule for the day; compare the production schedule for the day with the capacity relationship, and if there are remaining production resources, construct the plan in a loop until it meets the production requirements; finally, output the production scheduling results in matrix form.

[0173] 2. Generative AI production adjustment.

[0174] 2.1, Input layer data and relationships.

[0175] Since the ultimate goal of the present invention is to produce a balanced production schedule, the result value of the plan is a two-dimensional vector matrix: Prior to this, the present invention already has the production scheduling results of the planner as the teacher value, so according to the production scheduling results of the planner Initialize a coefficient matrix , its initialization data is all 1. As an input data, it is used to standardize the number of rows and columns of the output results.

[0176] The input layer data of the deep neural network model mainly includes factors that affect production scheduling, including work order reference, station capacity, worker scheduling, material supply, etc. Structurally, these influencing factors can be divided into three parts: time (24 hours), elements (work order, station, worker scheduling, material) and quantity (number of work orders, number of station capacity, worker scheduling hours, and material supply). Therefore, they can be unified into a vector representation with three dimensions and integrated together.

[0177] The specific method is as follows:

[0178] Step 1: Feature encoding.

[0179] The time dimension is represented by 24 hours and encoded as a 24-dimensional one-hot encoding vector. The element dimension uses one-hot encoding to represent each element as a vector. For example, if there are 10 stations, a 10-dimensional one-hot encoding vector is used to represent each station. For the quantity dimension, the original value of the quantity is used.

[0180] Step 2: Vector fusion.

[0181] First, weighted summation is performed. For each element dimension, the unique hot encoding vector of each element is first multiplied by the number of elements to obtain the weighted vector of the element. Then, the weighted vectors of all elements are added together to obtain the final vector representation of the dimension.

[0182] Then normalize. In order to ensure that vectors of different dimensions have comparable scales, each dimension vector needs to be normalized. Here, the L2 normalization method is used.

[0183] Work order basic information matrix input: The work order basic information, as a constraint description of other basic necessary conditions such as materials, production capacity, and worker scheduling, participates in the calculation of all feature extraction neurons in the first layer of the hidden layer.

[0184] Input of station capacity matrix: The station capacity matrix is ​​a capacity statistic for a foreseeable future period of time, which includes the capacity that each work station can provide every day. The capacity of a station on a certain day positively affects the coefficient of the process that needs the station. When extracting features, the present invention needs to convert the basic information of the work order into a matrix of work orders corresponding to stations. Assume Represents a collection of work orders, Represents a set of processes, represents the set of stations, then a matrix element The matrix is ​​therefore:

[0185] ;

[0186] Worker scheduling matrix input: The worker scheduling matrix is ​​the scheduling of workers in a foreseeable future period of time, including the team to which the worker belongs, the type of work order that can be started, the off-duty situation, the available working hours, etc. The start of a work order requires a worker with matching process skills, so if a worker is not scheduled on that day, the corresponding work order cannot be scheduled, and its coefficient value is close to 0. The more working hours it provides, the higher the coefficient in theory, but it cannot be infinitely high. The upper limit needs to refer to the working hours required for the work order, and a suitable activation function needs to be introduced to adjust it.

[0187] Material supply matrix input: Generally, a factory will have a material supply arrival schedule. The present invention can process it into a two-dimensional matrix of the daily supply quantity of each material. If the work order lacks input materials, it will lead to the inability to start work. Therefore, the supply of materials also positively affects the production scheduling coefficient of the work order.

[0188] 2.2, Deep neural network model process design.

[0189] The neural network structure consists of a fully connected layer and an activation layer. The fully connected layer is mainly used to memorize the deviation characteristics between the ideal linear production schedule and the actual production adjustment (teacher value), and the activation layer brings nonlinear characteristics (the adjustment of the production schedule by the planner is usually not linear).

[0190] Combined with the construction of the input layer of the deep neural network model in step 2.1, what the deep neural network model needs to learn is the adjustment coefficient between the scheduling factors and the manual scheduling results, which can also be considered as the adjustment coefficient between the theoretical scheduling plan generated by the scheduling factors and the manual scheduling plan. Therefore, this coefficient can be used to adjust the optimal scheduling algorithm, so that the scheduling results are closer to the manual scheduling effect.

[0191] For this adjustment coefficient, i.e. the model output, it is a two-dimensional vector Pij of the same size as the planner's scheduling result. Through back propagation during model training, this vector is adjusted until it is close enough to the planner's scheduling result.

[0192] For model training, in addition to training data, a loss function needs to be defined. Here, the mature mean square error function MSE is used. MES can measure the average square difference between the predicted value and the true value. It is sensitive to outliers but can distinguish between positive and negative errors. In the production scheduling problem, MSE can be used to measure the difference between the production volume predicted by the model and the actual production volume, and to penalize large errors.

[0193] The loss value calculated by the loss function uses the industry's existing gradient descent algorithm to update the neural network parameters, and the learning rate uses a value of about 0.2, which is adjusted depending on the specific training situation.

[0194] The ultimate capability of the neural network model of the present invention is to take the preliminary production scheduling plan as input, and be able to output the correct production scheduling model coefficient vector matrix, and use this production scheduling model coefficient vector matrix as the hyperparameter in the model of the present invention (step 1.5 hyperparameter) to re-schedule and obtain the desired production scheduling result of the present invention.

[0195] Based on the above objectives, the present invention designs a neural network model that meets the requirements of the present invention. The focus is on the output of the model and the calculation of the loss function. Compared with the general neural network model, the present invention requires additional steps at this step, because the output result is a coefficient vector matrix and a teacher value (manual production scheduling plan), and it is difficult to calculate the loss function. Therefore, when calculating the loss function, the present invention needs to first call "Step 1.6 Linear Programming Production Scheduling Process", bring the output result (coefficient vector matrix) into the model for calculation, and the output verification production schedule is used to calculate the loss value with the teacher value. This deep neural network model performs back propagation based on the loss value and uses gradient descent to update the weight matrix of the attention layer. After iterative training, it is expected that the target model and production scheduling model coefficient vector matrix required by the present invention will be obtained.

[0196] In the embodiment of the present invention, the process of generative AI production adjustment is as follows: Figure 3 shown.

[0197] 2.3, deep neural network model training.

[0198] Based on historical production scheduling data, a production adjustment model that conforms to the planner's behavioral logic can be trained to adjust the initial production scheduling results to a balanced production scheduling plan.

[0199] First, you need to prepare training data and obtain a preliminary production schedule based on the historical orders and their corresponding production resource data according to the steps of "optimal production scheduling modeling" (steps 1.1-1.6).

[0200] Run the preliminary schedule as input Figure 3 Perform iterative training, during which you can adjust the model by adjusting the learning rate, regularization rate, regularization function, etc. until the training results tend to be stable, so that the deep neural network model can converge quickly and effectively.

[0201] Embodiment 3, the present invention provides a generative dynamic production resource scheduling and balancing system, comprising:

[0202] The production scheduling objective function acquisition module is used to obtain the set of process tasks that need to be started according to the order and BOM according to the constraints between production resources and production resources, obtain the capacity list of the positions according to the position situation of the factory, and obtain the daily material distribution matrix according to the material distribution plan. Based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day. The production completion hours of the Nth day are solved in sequence by calling the solver to obtain the objective function of the production scheduling plan until the i-th day, when the sum of the completion hours of each process is the rated working hours corresponding to the process;

[0203] The deep neural network model adjustment module is used to introduce hyperparameter coefficients into the objective function. Through deep neural network model training and learning, the hyperparameter coefficient values ​​are output and brought into the objective function of the linear programming algorithm to obtain the same production scheduling results as the planner.

[0204] In the deep neural network model training module, it is used to expand the loss function calculation and additionally call the linear programming algorithm to obtain the predicted value, so as to obtain the loss value using the loss function and perform back propagation, so that the trained deep neural network model can simulate the planner to make the correct balanced production schedule without human intervention.

[0205] To further illustrate the effects of the embodiments of the present invention, the following experiment was conducted.

[0206] 3. Experiment.

[0207] 3.1, means of production.

[0208] In this scenario, the production materials include 25 stations, 50 workers, two shifts (8 hours per shift), 200 materials involved in the production schedule, and an average material supply rate of 0.95. Six months of production schedule results for planners are available for training.

[0209] 3.2, model construction.

[0210] According to the basic idea of ​​step 2.2, a deep learning neural network model is constructed, using 1 input layer, 3 hidden layers, 1 activation layer and 1 output layer. The mean square error function MSE is used as the loss function.

[0211] The above model can be built based on the mature Pytorch framework.

[0212] 3.3, Training data.

[0213] The 6-month production scheduling result data is divided into 6 sets of training data for each month. The corresponding production scheduling factor data is integrated into the input vector according to step 2.1.

[0214] 3.4, Model training.

[0215] Train the model based on the trainer in the model framework until the loss reaches the expected value. Save the model and the last layer parameter matrix of the model as adjustment coefficients.

[0216] 3.5, Application scheduling.

[0217] The optimal production scheduling algorithm is implemented using traditional coding methods, and the algorithm results are calculated with the adjustment coefficient to obtain the final production scheduling result.

[0218] The above functions are encapsulated as external APIs and connected to the existing information systems of enterprises (ERP, MES, etc.). The production scheduling request (including production materials, production scheduling parameters, etc.) is received from the enterprise information system, and the production scheduling results are obtained after scheduling. The production scheduling results are converted into the format required by the enterprise information system, fed back to the enterprise information system, and the production scheduling results are displayed in an appropriate visual way. The production scheduling results are visualized as follows: Figure 4 shown.

[0219] In short, in order to meet the actual production situation, balance the interests of all parties, improve production efficiency and save costs, the present invention proposes a generative production resource balancing algorithm for the production planning and scheduling needs of discrete manufacturing industries, especially the rail transit industry. By learning the production adjustment behaviors and reasons of planners and solidifying them into a production adjustment model, the preliminary production scheduling results generated by the present invention are automatically adjusted to production scheduling results similar to those of the planners.

[0220] It mainly includes two modules: the optimal production scheduling algorithm and the generative AI production adjustment. The linear programming algorithm model is a necessary component tool for the output plan and is also the subsystem on which the generative AI production adjustment depends. In the linear programming model, the coefficient is added to the objective function. After training and learning in the subsequent AI model, the correct coefficient is output and brought into the objective function of the linear programming model to obtain a production adjustment result similar to that of the planner. In the AI ​​model, since the output result cannot be compared with the teacher value, the loss function calculation is expanded and the linear programming model is additionally called to obtain the predicted value, so as to achieve the purpose of using the loss function to obtain the loss value and perform back propagation to achieve effective learning. The stable and convergent AI model obtained after training can simulate the planner to make the correct balanced production scheduling plan without human intervention, further effectively reduce labor costs, and further realize intelligent scheduling. Help the manufacturing industry, especially the large-scale discrete manufacturing industry, to transform to intelligent manufacturing.

[0221] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A generative dynamic production resource scheduling and balancing method, characterized in that: This method uses linear programming combined with machine learning to adjust production schedules, adjusts production schedules through a deep neural network model trained after learning, and outputs a production schedule matrix that is consistent with the balance of production resources, so that production efficiency is balanced with cost control and resource allocation; The specific steps include: S1, according to the constraints between production resources and production resources, the set of process tasks that need to be started is obtained according to the order and BOM, the capacity list of the position is obtained according to the position situation of the factory, and the daily material distribution matrix is ​​obtained according to the material distribution plan; based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production schedule completion time of the day is used to obtain the objective function of the production schedule until the The sum of the completion time of each process is the rated working time corresponding to the process. S2, introduces the hyperparameter coefficient into the objective function, outputs the hyperparameter coefficient value through deep neural network model training and learning, and brings it into the objective function of the linear programming algorithm to obtain the production scheduling result; S3, in the deep neural network model, by expanding the loss function calculation, additionally calling the linear programming algorithm to obtain the predicted value, the output of the deep neural network model is adjusted by the linear programming algorithm and then compared with the teacher value to obtain the loss value, so that the trained deep neural network model can make a balanced production schedule without human intervention; The production resources include: process, process task, station, station capacity, material, production rhythm, and the material tree Bom of the complete product; The constraint relationship includes: the required machine capacity of the process is less than or equal to the machine capacity of the day, the required materials are less than or equal to the planned material delivery quantity of the day, and the start time of the process task in the production rhythm cycle is not less than the start time of the cycle in which the production rhythm is located; The set of process tasks required to start is This collection includes work orders added according to the production rhythm, with the required capacity, required materials, process sequence, and earliest start time information of the process to which it belongs attached; The table capacity list obtained according to the table situation of the factory is ; The material distribution matrix obtained every day according to the material distribution plan is: , For the number of days, is the material number, The quantity of the material.

2. The generative dynamic production resource scheduling and balancing method according to claim 1 is characterized in that: In step S1, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production completion time of the day is calculated, and the objective function of the production schedule is obtained, including: The daily production scheduling results are abstracted into the working hours required to complete each process task every day. The completion time on the first day is , subscript 1 corresponds to the current first day, Corresponding process task number, there are The eight process tasks, after calling the solver, get the scheduling results as follows: ; In the formula, The number of days for production scheduling, is the process task number, The value is the working hours of a certain process on that day. The working hours required to complete the work order; for The minimum effective value of The production scheduling target of the production scheduling result is to find the maximum value of the sum of the daily work completion volume. The target formula is: ； In the formula, The maximum value of the sum of the completed work.

3. The generative dynamic production resource scheduling and balancing method according to claim 1 is characterized in that: When the required station capacity of the process is less than or equal to the station capacity of the day, the capacity constraints on the first day need to be met, including: the constraints of the process itself and the process with materials available for start; The process itself has constraints, including: Seats The processes that can be completed in one day's production capacity, ; Seats The processes that can be completed in one day's production capacity, ; Seats The processes that can be completed in one day's production capacity, ; Seats The processes that can be completed in one day's production capacity, ; The processes for which the materials can be started include: materials Available processes to start, ; materials Available processes to start, ; materials Available processes to start, ; materials Available processes to start, ; materials Available processes to start, ; The start-up percentage of each process task cannot exceed the rated working hours, including: ； in, Represents the 1st to 8th processes of a day; Process Tasks The subsequent process tasks are tasks that start one week after the start of production according to the production rhythm and cannot be started on the first day, including: ; Process Tasks Waiting for process tasks Construction starts after completion, including: ； In the formula, Is it a process task? After the task, For the Working hours for the first process per day; Converts to: ; In the formula, For the front The day cannot occupy working hours; In the Day, before Daily process tasks The sum of the completion percentages is equal to 100%; Input the above constraints and objective formulas into the solver to obtain the planning results for the first day.

4. The generative dynamic production resource scheduling and balancing method according to claim 3 is characterized in that: On the second day, the target formula for the evolution of the production target working hours over time is: ； In the formula, The completion time for the second day; The constraints on function scheduling must be met, including the following constraints: Table capacity constraints: ; Material supply constraints: ； Working hours constraints: ； in It indicates the 1st to 8th processes on the second day; The completed working hours on the second day are less than or equal to the required rated working hours minus the working hours completed previously; Find the first The result of the scheduled completion time of the day is , The number of days for production scheduling, Number the process task.

5. The generative dynamic production resource scheduling and balancing method according to claim 4 is characterized in that: Until Days, the sum of the completion hours of each process is the rated working hours corresponding to the process, satisfying: ； In the formula, The total number of days for production scheduling. The process number is The process tasks, For the The daily process number is The production hours of process tasks are scheduled. The process number is The sum of all scheduled working hours for the process.

6. The generative dynamic production resource scheduling and balancing method according to claim 1, characterized in that: In step S2, the hyperparameter coefficient is introduced into the objective function, and the expression is: ； In the formula, is the first Day process number Working time parameters, The production schedule hours are multiplied to achieve the purpose of using the coefficient adjustment result.

7. The generative dynamic production resource scheduling and balancing method according to claim 1, characterized in that: In step S2, the hyperparameter coefficient value is output through deep neural network model training and learning, and is brought into the objective function of the linear programming algorithm to obtain the production scheduling result; including: according to the production scheduling result Initialize a coefficient matrix , the initialization data is all 1, As an input data, standardize the number of rows and columns of the output result; split the model input layer data into time, elements and quantity, unify them into vector representations with three dimensions and merge them together; the specific steps are: Step 1: Feature encoding. The time dimension is represented by 24 hours and encoded as a 24-dimensional one-hot encoding vector. The element dimension uses one-hot encoding to represent each element as a vector, and the original value of the quantity dimension is used. Step 2: Vector fusion: For each element dimension, multiply the unique hot encoding vector of each element by the number of elements to obtain the weighted vector of the element; add the weighted vectors of all elements to obtain the final vector representation of the dimension; then normalize, using the L2 normalization method to normalize each dimension vector; Work order basic information matrix input: The work order basic information is used as a constraint description for other materials, production capacity, and worker scheduling basic necessary conditions, and participates in the calculation of all feature extraction neurons in the first layer of the hidden layer; Input of the station capacity matrix: During feature extraction, the basic information of the work order is converted into a matrix of work orders corresponding to stations; Worker scheduling matrix input: the expected schedule of workers in the future, including the team to which the worker belongs, the type of work order that can be started, the off-duty situation, the available working hours, and the introduction of activation function adjustment; Material supply matrix input: processed as a two-dimensional matrix of the supply quantity of each material per day.

8. The generative dynamic production resource scheduling and balancing method according to claim 1, characterized in that: In step S3, the deep neural network model consists of a fully connected layer and an activation layer. The fully connected layer is used to memorize the deviation characteristics between the ideal linear production scheduling and the actual production scheduling, and the activation layer is used to bring nonlinear characteristics. The loss function uses the mean square error function MSE to measure the average of the squared difference between the predicted value and the true value, and distinguishes positive and negative errors for outliers; The loss value calculated by the loss function uses the gradient descent algorithm to update the neural network parameters.

9. A generative dynamic production resource scheduling and balancing system, characterized in that: The system is implemented by the generative dynamic production resource scheduling and balancing method according to any one of claims 1 to 8, and the system includes: The production scheduling objective function acquisition module is used to obtain the set of process tasks that need to be started according to the constraints between production resources and production resources, orders and BOM, obtain the capacity list of the positions according to the position situation of the factory, and obtain the daily material distribution matrix according to the material distribution plan; based on the above information, the linear programming algorithm is used to abstract the daily production scheduling results into the working hours to be completed for each process task every day, and the solver is called to solve the first The production schedule completion time of the day is used to obtain the objective function of the production schedule until the The sum of the completion time of each process is the rated working time corresponding to the process. The deep neural network model adjustment module is used to introduce hyperparameter coefficients into the objective function. Through deep neural network model training and learning, the hyperparameter coefficient values ​​are output and brought into the objective function of the linear programming algorithm to obtain the production adjustment and scheduling results. The deep neural network model training module expands the loss function calculation and additionally calls the linear programming algorithm to obtain the predicted value. The output of the deep neural network model is adjusted by the linear programming algorithm and then compared with the teacher value to obtain the loss value. The trained deep neural network model can make a balanced production schedule without human intervention.

Images