APS dynamic production planning and scheduling system

Abstract

The present invention relates to an APS dynamic production planning and scheduling system, which includes the following scheduling process: S1, inputting production plan information; S2, handing over data modeling to APS, and formulating a production plan and scheduling through a production scheduling algorithm; S3, outputting the scheduling result and providing feedback; S4, based on the feedback of the result, if satisfied, the production plan can be issued, integrated with a third-party system, and the flow is completed, and the process ends here; S5, based on the feedback of the result, if dissatisfied, the input information is modified, the order is adjusted, and steps S1 to S4 are repeated. The present invention realizes the dynamic adjustment of digital factory planning and scheduling by constructing a dynamic production planning and scheduling method; for fluctuations in manufacturing time, machine failures, preventive maintenance, dynamic order insertion, etc., the probability distribution of various uncertainties is learned through a neural network, and at the same time, the online learning mechanism provided by reinforcement learning is used to dynamically respond to various uncertainties, providing a complete solution for the factory's intelligent manufacturing upgrade.

Description

Technical Field

[0001] The present invention relates to the technical field of production planning and scheduling, and in particular to an APS dynamic production planning and scheduling system. Background Art

[0002] Advanced Planning and Scheduling (APS) is the process of assigning production tasks to production resources. Taking into account capacity and equipment, and given a fixed material quantity, APS arranges the production sequence for each task, optimizes the production sequence, and selects the right equipment to reduce waiting times and balance the production load between machines and workers. This optimizes production capacity, improves production efficiency, and shortens production lead time.

[0003] However, production planning and scheduling is a complex issue due to its inherently complex logic and numerous variables in practical application, especially for discrete production involving high-variety, small-batch production and complex process flows. Factors such as equipment failure and delayed raw material supply make the planner's work complex and error-prone.

[0004] At present, the existing production planning and scheduling mostly remains in the static planning and scheduling stage, and has not yet achieved a closed-loop feedback with actual production information. It is unable to deal with unexpected events such as process delays, machine failures, and inserted orders in a timely manner, which affects the application effect of production planning and scheduling. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to propose an APS dynamic production planning and scheduling system, which can improve productivity by accurately controlling production and rationally formulating production plans.

[0006] The technical solution adopted by the present invention is: an APS dynamic production planning and scheduling system, including the following scheduling process,

[0007] S1. Input production plan information, including raw material information, process information and order information;

[0008] S2. APS completes data modeling and formulates production plans and schedules through production scheduling algorithms;

[0009] S3. Output scheduling results and provide feedback;

[0010] S4. Based on the feedback, if the results are satisfactory, the production plan can be issued and integrated with the third-party system to complete the process.

[0011] S5. Based on the feedback, if you are not satisfied, modify the input information, adjust the order, and repeat steps S1 to S4.

[0012] Furthermore, in step S1 of the present invention, the raw material information, process information and order information have a matching relationship.

[0013] Furthermore, in step S2 of the present invention, the production scheduling algorithm includes the following steps:

[0014] 1) Using GA to generate static schedule without considering dynamic characteristics;

[0015] 2) When dynamic characteristics appear, the work order release rules and key parameters in the GA scheduling are adjusted online through reinforcement learning, and the GA scheduling is re-performed;

[0016] 3) Through the neural network, the probability distribution of uncertainty is carried, and this probability distribution is used as a feedforward prediction to guide and adjust the scheduling results of the next GA genetic algorithm;

[0017] 4) Repeat steps 1) to 3) continuously, thus achieving a positive feedback loop from the emergence of individual intelligence to group intelligence, and then from group intelligence to the evolution of individual intelligence, until intelligence converges.

[0018] Furthermore, step 1) of the present invention includes the following steps:

[0019] A. Problem initialization;

[0020] B. Use genetic operators to generate optimal solutions through swarm intelligence;

[0021] Genetic operators include,

[0022] Selection operator, that is, judging the robustness of the individual through the Fitness robustness function;

[0023] The crossover operator is to cross over and recombine the parent chromosomes of two individuals to generate offspring, where the crossover position can be set in multiple ways;

[0024] as well as,

[0025] Mutation operator, that is, the random change of chromosomes in offspring;

[0026] Through genetic operators, the evolution process from parent generation to offspring generation is completed;

[0027] C. Repeat the evolution process in step B until the end condition is met; the end condition can be a period of time, meeting the expectation of the objective function, or reaching a certain number of iterations;

[0028] D. Based on the generated optimal solution, decoding is performed in the opposite process of encoding to generate the actual scheduling plan.

[0029] Furthermore, in step A of the present invention, the problem initialization includes the following steps:

[0030] a) Encode the PPM model according to the PPM modeling results;

[0031] b) According to the set objective function, convert it into a fitness robust function;

[0032] c) Generate the initial population randomly.

[0033] Furthermore, in step 2) of the present invention, the dynamic characteristics are classified in advance into two types: uncertainty caused by random factors and uncertainty caused by system dynamic characteristics; and a work order release rule pool is pre-established.

[0034] Furthermore, when uncertainty occurs, the present invention includes the following steps:

[0035] I. The dynamic scheduling problem is regarded as an MDP Markov time series model, that is, the schedule generated by the GA genetic algorithm is the state at time t-1, and the current state after the uncertainty is generated is the state at time t;

[0036] II. Using the Q-learning method of reinforcement learning, the model takes the environment parameters, the original objective function, and the dynamic characteristics as inputs. It is trained through trial and error and matched with the specific rules in the work order release rule pool.

[0037] III. Use the original objective function as the reward function, learn the mapping method with the maximum reward value, and store the learning results in the Q-learning query table.

[0038] Furthermore, in step II of the present invention, an epsilon-greedy algorithm is used to increase the randomness of the search.

[0039] Furthermore, in step 3) of the present invention, if the dynamic characteristics are uncertainties caused by random factors, then the modeling parameters of the physical model are adjusted in advance by setting buffers and reserving production capacity according to the predicted probability of the occurrence of the characteristics; if the dynamic characteristics are uncertainties caused by the system dynamics characteristics, then the parameters in the modeling are adjusted according to the predicted probability of the occurrence of the characteristics, and used as reference indicators for preventive maintenance.

[0040] Furthermore, in step S3 of the present invention, the production plan is displayed in the form of a Gantt chart; by combining the production plan and inventory quantity, the consumption of raw materials and the inventory changes throughout the production process can be known; and the resource utilization rate can be estimated based on the production plan.

[0041] The beneficial effects of the present invention are:

[0042] By building a dynamic production planning and scheduling method, dynamic adjustment of digital factory planning and scheduling is achieved; in response to fluctuations in manufacturing time, machine failures, preventive maintenance, dynamic order insertion, etc., planners use neural networks to learn the probability distribution of various types of uncertainties. At the same time, through the online learning mechanism provided by reinforcement learning, they can dynamically respond to various types of uncertainties, providing a complete solution for the factory's intelligent manufacturing upgrade. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the APS dynamic production planning and scheduling system of the present invention;

[0044] Figure 2 It is a schematic diagram of the APS dynamic production planning and scheduling algorithm of the present invention. DETAILED DESCRIPTION

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0046] like Figure 1 The APS dynamic production planning and scheduling system shown includes the following scheduling processes:

[0047] S1. Input production plan information, including raw material information, process information and order information;

[0048] S2. APS completes data modeling and formulates production plans and schedules through production scheduling algorithms;

[0049] S3. Output scheduling results and provide feedback;

[0050] S4. Based on the feedback, if the results are satisfactory, the production plan can be issued and integrated with the third-party system to complete the process.

[0051] S5. Based on the feedback, if you are not satisfied, modify the input information, adjust the order, and repeat steps S1 to S4.

[0052] In step S1, raw material information, process information, and order information are matched. Order elements include finished product, delivery date, and quantity, specifically, when and in what quantity the product needs to be produced. A process is the process of transforming one product into another through a specific operation. A process takes time, for example, transforming materials A and B into semi-finished product C.

[0053] In addition, the following aspects are also involved:

[0054] The ratio between raw materials (before processing) and "finished products" (after processing);

[0055] The number of resources (equipment, personnel) that the processing operation depends on;

[0056] The time consumed by a process to produce a unit quantity of product;

[0057] Process type: positively correlated with production quantity (piecework process), unrelated to production quantity (fixed process), replaceable process);

[0058] Raw materials and semi-finished products flow in different workshops (locations), and products are distributed;

[0059] Raw materials, inventory, purchase of raw materials from suppliers;

[0060] In step S3, the production plan is displayed in the form of a Gantt chart; combining the production plan and inventory quantity, the consumption of raw materials and the inventory changes throughout the production process can be known; and the utilization rate of resources (equipment, etc.) can be calculated based on the production plan.

[0061] Problem feedback includes: 1. Whether a certain order is delayed in delivery; 2. Whether the raw materials required for a certain product are insufficient; and the logistics scheduling plan for product distribution.

[0062] The core of APS dynamic production planning and scheduling is the APS dynamic production planning and scheduling algorithm, such as Figure 2 The production scheduling algorithm shown includes the following steps:

[0063] 1) Using GA to generate static schedule without considering dynamic characteristics;

[0064] 2) When dynamic characteristics appear, the work order release rules and key parameters in the GA scheduling are adjusted online through reinforcement learning, and the GA scheduling is re-performed;

[0065] 3) Through the neural network, the probability distribution of uncertainty is carried, and this probability distribution is used as a feedforward prediction to guide and adjust the scheduling results of the next GA genetic algorithm;

[0066] 4) Repeat steps 1) to 3) continuously, thus achieving a positive feedback loop from the emergence of individual intelligence to group intelligence, and then from group intelligence to the evolution of individual intelligence, until intelligence converges.

[0067] The GA genetic algorithm generally includes four processes:

[0068] 1. Problem initialization, which includes three key steps:

[0069] 1) Based on the PPM modeling results, the original FJSSP problem is encoded to generate a "chromosome";

[0070] 2) According to the objective function set by the user, convert it into a Fitness function (usually they can be set to the same, mainly depending on the definition of the encoding method);

[0071] 3) The initial population is usually generated randomly, or the solution generated by the moving bottleneck heuristic algorithm can be used as the initial solution.

[0072] 2. Genetic operators are used to generate optimal solutions through swarm intelligence. These operators specifically include three: 1) a selection operator, which uses the Fitness function to determine the robustness of individuals; 2) a crossover operator, which crosses and recombines the parent chromosomes of two individuals to generate offspring, with a variety of crossover positions; and 3) a mutation operator, which modifies the offspring's chromosomes with a certain probability (usually randomized), such as through real-valued mutations or binary mutations. Through genetic operators, the evolutionary process from parent to offspring is completed.

[0073] 3. Repeat the evolution process in step 2 until the end condition is met. The end condition can be running for a period of time (such as 1 hour), meeting the expectation of the objective function, or reaching a certain number of iterations.

[0074] 4. Based on the generated optimal solution, decoding is performed in the exact opposite process of encoding to generate the actual schedule.

[0075] When faced with multi-objective optimization problems, the NSGA-II algorithm is used for multi-objective optimization solutions. The essence of NSGA-II is to introduce an elitist strategy (to ensure that offspring do not perform worse than their parents), a fast non-dominated strategy (to reduce computational complexity and select offspring from a doubled space), and a crowding comparison operator (to ensure that individuals are evenly distributed across the entire Pareto domain) based on GA.

[0076] However, real-world production scenarios often involve numerous uncertainties, such as fluctuations in manufacturing time, machine failures, preventative maintenance, and dynamic order insertion. Therefore, the actual production scheduling problem can be described as a dynamic FJSSP problem (Dynamic / Fuzzy FJSSP). To address these uncertainties, neural networks are needed to learn the probability distribution of these uncertainties. Furthermore, reinforcement learning provides an online learning mechanism to dynamically address these uncertainties. This approach solves the Dynamic FJSSP problem by combining GA, neural networks, and reinforcement learning—the swarm intelligence evolutionary algorithm.

[0077] In the swarm intelligence evolution algorithm, the functions of each part are as follows:

[0078] 1. Neural networks are used to carry the probability distribution of uncertainty, that is, to carry learned intelligence;

[0079] 2. GA provides a collective intelligence emergence mechanism from individual intelligence to group intelligence;

[0080] 3. Reinforcement learning provides GA with an evolutionary mechanism for uncertainty.

[0081] The process of solving the dynamic FJSSP scheduling problem with the swarm intelligence evolutionary algorithm is as follows:

[0082] 1. Classify dynamic characteristics in advance into uncertainties caused by random factors, including emergency orders and machine failures; and uncertainties caused by system dynamics, including fluctuations in manufacturing time and preventive maintenance.

[0083] 2. Pre-establish a work order release rule pool, which includes commonly used work order release rules, such as EDD, LIFO, LPT, etc. As shown in the following table:

[0084]

[0085] 3. When uncertainty occurs, the dynamic scheduling problem is regarded as an MDP Markov time series model, that is, the scheduling plan generated by the genetic algorithm is the state at time t-1, and the current state after the uncertainty occurs is the state at time t.

[0086] 4. Using the reinforcement learning Q-learning method, the model takes environmental parameters (including capacity utilization, completion time, and number of delayed work orders), the original objective function, and dynamic characteristics (including category and urgency (represented by a priority level of 0-10)) as inputs. The model is trained through trial and error and matched with specific rules in the work order release rule pool.

[0087] In order to enhance the long-term performance of the algorithm, the epsilon-greedy method is used to increase the randomness of the search.

[0088] 5. Use the original objective function as the reward function, learn the mapping method that maximizes the reward value, and store the learning results in the Q-learning lookup table.

[0089] In actual use, the dynamic characteristics that appear will be processed according to the mapping relationship learned in the query table; at the same time, the online learning characteristics of the algorithm will be used to continuously optimize the matching accuracy.

[0090] The neural network learns and carries the probability distribution of dynamic characteristics of various categories and different importance. The probability distribution learned by the neural network is used as a feedforward prediction to guide and adjust the scheduling results of the next genetic algorithm.

[0091] If the dynamic characteristics are uncertain due to random factors, the modeling parameters of the physical model will be adjusted in advance by setting buffers and reserving production capacity based on the predicted probability of the characteristics to cope with uncertainties such as machine failures and emergency orders.

[0092] If the dynamic characteristics are uncertainties caused by the system dynamics, the parameters such as manufacturing time and production capacity in the modeling will be adjusted according to the predicted probability of the occurrence of the characteristics, and they will be used as reference indicators for preventive maintenance.

[0093] Therefore, if the user accepts the APS scheduling plan generated by the above algorithm, the production plan can be sent to MES to complete the business flow.

[0094] By making comprehensive use of the above information, users can also:

[0095] 1. Analysis of production capacity bottlenecks

[0096] Some orders cannot be delivered on time. By combining production plans with equipment load conditions, we can identify bottleneck resources and then promote on-time delivery of orders by increasing resources or outsourcing production to third parties.

[0097] 2. Simulation Production

[0098] Based on the current schedule, if a new order is received, can the factory deliver on time? You can use APS to simulate production conditions and assess whether delivery can be met.

[0099] 3. Production Adjustment

[0100] According to the actual production information fed back by the production management module, the scheduling results are adjusted dynamically, and secondary scheduling can be performed to form a closed-loop feedback from advanced planning and scheduling to actual production.

[0101] The above description only describes specific embodiments of the present invention. Various examples do not limit the essential content of the present invention. After reading the description, ordinary technicians in the relevant technical field can make modifications or variations to the specific embodiments described above without departing from the essence and scope of the present invention.

Claims

1. An APS dynamic production planning and scheduling system, characterized by: The following scheduling process is included: S1. Input production plan information, including raw material information, process information and order information; S2. APS completes data modeling and formulates production plans and schedules through production scheduling algorithms; The production scheduling algorithm includes the following steps: 1) Using GA to generate static schedule without considering dynamic characteristics; The following steps are involved: A. Problem initialization; B. Use genetic operators to generate optimal solutions through swarm intelligence; Genetic operators include, Selection operator, that is, judging the robustness of the individual through the Fitness robustness function; The crossover operator is to cross over and recombine the parent chromosomes of two individuals to generate offspring, where the crossover position can be set in multiple ways; as well as, Mutation operator, that is, the random change of chromosomes in offspring; Through genetic operators, the evolution process from parent generation to offspring generation is completed; C. Repeat the evolution process in step B until the end condition is met; the end condition can be a period of time, meeting the expectation of the objective function, or reaching a certain number of iterations; D. Based on the generated optimal solution, decoding is performed in the opposite process of encoding to generate the actual schedule; 2) When dynamic characteristics appear, the work order release rules and key parameters in the GA scheduling are adjusted online through reinforcement learning, and the GA scheduling is re-performed; In step 2), the dynamic characteristics are classified in advance into two types: uncertainty caused by random factors and uncertainty caused by system dynamic characteristics; and a work order release rule pool is pre-established; When uncertainty arises, the following steps are included: I. The dynamic scheduling problem is regarded as an MDP Markov time series model, that is, the schedule generated by the GA genetic algorithm is the state at time t-1, and the current state after the uncertainty is generated is the state at time t; II. Using the Q-learning method of reinforcement learning, the model takes the environment parameters, the original objective function, and the dynamic characteristics as inputs. It is trained through trial and error and matched with the specific rules in the work order release rule pool. III. Use the original objective function as the reward function, learn the mapping method that maximizes the reward value, and store the learning results in the Q-learning query table; In step II, an epsilon-greedy algorithm is used to increase the randomness of the search; 3) Through the neural network, the probability distribution of uncertainty is carried, and this probability distribution is used as a feedforward prediction to guide and adjust the scheduling results of the next GA genetic algorithm; 4) Repeat steps 1) to 3) continuously, thus achieving a positive feedback loop from the emergence of individual intelligence to group intelligence, and then from group intelligence to the evolution of individual intelligence, until intelligence converges; S3. Output scheduling results and provide feedback; S4. Based on the feedback, if the results are satisfactory, the production plan can be issued and integrated with the third-party system to complete the process. S5. Based on the feedback, if you are not satisfied, modify the input information, adjust the order, and repeat steps S1 to S4.

2. The APS dynamic production planning and scheduling system according to claim 1, wherein: In step S1, the raw material information, process information and order information have a matching relationship.

3. The APS dynamic production planning and scheduling system according to claim 1, wherein: In step A, the problem initialization includes the following steps: a) Encode the PPM model according to the PPM modeling results; b) According to the set objective function, convert it into a fitness robust function; c) Generate the initial population randomly.

4. The APS dynamic production planning and scheduling system according to claim 1, wherein: In step 3), if the dynamic characteristics are uncertainties caused by random factors, the modeling parameters of the physical model are adjusted in advance by setting buffers and reserving production capacity based on the predicted probability of the occurrence of the characteristics; if the dynamic characteristics are uncertainties caused by the system dynamics characteristics, the parameters in the modeling are adjusted based on the predicted probability of the occurrence of the characteristics, and used as reference indicators for preventive maintenance.

5. The APS dynamic production planning and scheduling system according to claim 1, wherein: In step S3, the production plan is displayed in the form of a Gantt chart; by combining the production plan and inventory quantity, the consumption of raw materials and the inventory changes throughout the production process can be known; And calculate resource utilization based on production plan.

Images