Dynamic flexible operation semiconductor workshop scheduling method of knowledge-driven evolutionary algorithm

Abstract

The invention is suitable for the technical field of workshop scheduling, and particularly relates to a dynamic flexible operation semiconductor workshop scheduling method of a knowledge-driven evolutionary algorithm, and the method comprises the steps: obtaining workshop working condition data which at least comprises workpiece data and machine data; constructing a workshop scheduling mathematical model based on the workshop working condition data; and solving the workshop scheduling data model based on a genetic algorithm, Q-learning and a variable neighborhood search algorithm based on knowledge driving to generate a workshop scheduling scheme. According to the method, workpiece transportation time, machine adjustment time and machine fault uncertainty factors are introduced on the basis of traditional flexible job shop scheduling, energy-saving scheduling is also considered, and the energy-saving purpose is achieved while the production requirement is met by reducing energy consumption. While the efficiency is improved, the energy consumption can be reduced, and the market competitiveness and sustainable development capability of enterprises are enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of workshop scheduling, and in particular to a dynamic flexible operation semiconductor workshop scheduling method based on a knowledge-driven evolutionary algorithm. Background Art

[0002] Flexible Shop Scheduling refers to the process of flexibly allocating and scheduling resources (such as machines, personnel, and materials) in a manufacturing environment to adapt to changes and uncertainties in production tasks, thereby optimizing production efficiency, shortening delivery times, reducing inventory backlogs, and improving equipment utilization. It allows jobs to choose from multiple possible processing paths, not limited to fixed machines or process sequences, which increases the ability to respond to emergencies and optimize the overall production process. Through intelligent algorithms and technologies (such as heuristic algorithms, genetic algorithms, simulated annealing, etc.), flexible shop scheduling can achieve the best allocation of resources under dynamically changing conditions, ensuring production flexibility and responsiveness.

[0003] Flexible job shop scheduling is gradually developing towards green and intelligent. However, in actual production, various dynamic disturbances often occur, resulting in reduced feasibility of scheduling solutions. Summary of the invention

[0004] The purpose of the present invention is to provide a dynamic flexible operation semiconductor workshop scheduling method based on a knowledge-driven evolutionary algorithm, aiming to solve the problem that various dynamic disturbances often occur in actual production, resulting in reduced executability of scheduling schemes.

[0005] The present invention is implemented by a knowledge-driven evolutionary algorithm dynamic flexible operation semiconductor workshop scheduling method, the method comprising:

[0006] Acquiring workshop working condition data, wherein the working condition data at least includes workpiece data and machine data;

[0007] Construct a mathematical model for workshop scheduling based on workshop working condition data;

[0008] The workshop scheduling data model is solved based on genetic algorithm, Q-learning and knowledge-driven variable neighborhood search algorithm to generate a workshop scheduling plan.

[0009] Preferably, the optimization objective of the workshop scheduling mathematical model is expressed as:

[0010]

[0011] minTEC=PEC+SEC+REC+IEC+CEC

[0012]

[0013] Among them, in the workshop scheduling mathematical model, processing energy consumption, adjustment energy consumption, transportation energy consumption, idle energy consumption and general energy consumption are expressed as:

[0014]

[0015]

[0016] In the mathematical model of shop scheduling, the constraints are expressed as:

[0017]

[0018]

[0019] Among them, I and i' are workpiece indices, j and j' are process indices, K and k' are machine indices, I represents the workpiece set, I = {1, 2, ..., n}, J i represents the process set of workpiece i, J i ={1,2,…,h i}, K represents the machine set, K = {1, 2, ..., m}, JJ represents the Nth j -1 process set, JJ={1,2,…,N j -1},M i,j Indicates process O i,j Optional machine set, M i,j ={1,2,…,g ij}, L k represents the location set of machine k, L k ={1,2,…,l k}, L k * The first L of machine k k -1 location set, L k * ={1,2,…,l k -1}, n represents the number of workpieces, h i represents the number of processes of workpiece i, m represents the total number of machines, and N j represents the total number of processes, l represents the location of the machine, O i,j represents the jth process of workpiece i, g i,j Indicates process O i,j The number of machines that can be selected, l k represents the maximum number of positions of machine k, PT i,j,k Indicates process O i,j Processing time on machine k, AP i,j,k Indicates process O i,j Unit processing energy consumption on machine k, AI kIndicates that machine k is in process O i,j Unit idle energy consumption, ST i,j,k Indicates process O i,j Adjustment time on machine k, AS i,j,k Indicates process O i,j Unit preparation energy consumption on machine k, LI represents a positive infinite number, TT i,j,k,k' Indicates process O i,j The transportation time from machine k to machine k', AC represents the unit general energy consumption in the workshop, UT represents the unit transportation energy consumption for transporting equipment, C i represents the completion time of job i, X i,j,k represents a binary constant, if process O i,j If it is processed on machine k, the value is 1; otherwise, it is 0. i,j,k,l represents a binary constant, if process O i,j If the machine k is processing at position l, the value is 1; otherwise, it is 0. i,j Indicates process O i,j Start time, ST k,l represents the start time of location l of machine k, CT k,l represents the makespan of machine k at location l, MCT represents the maximum makespan, TEC represents the total energy consumption, and WCM represents the critical load of the machine.

[0020] Preferably, the step of solving the workshop scheduling data model based on genetic algorithm, Q-learning and knowledge-driven variable neighborhood search algorithm includes:

[0021] Obtain the initial population based on the initialization strategy;

[0022] Initialize the Q-learning algorithm and set the crossover probability P c and mutation probability P m ;

[0023] Obtain a new population by performing crossover and mutation operations, calculate the state and reward of the new population, and update the Q table;

[0024] The solution is solved by using a knowledge-driven variable neighborhood search algorithm, which non-dominated sorts the target values ​​of all selected individuals and calculates the crowding distance.

[0025] Perform multiple iterations until the iteration stop condition is met and output the workshop scheduling plan.

[0026] Preferably, in the step of obtaining the initial population based on the initialization strategy, four initialization rules are used to obtain the initial population, and the initialization rules are respectively generating the initial solution completely randomly, generating the initial solution according to the minimum time, generating the initial solution preferentially according to the maximum remaining processing time, and preferentially allocating the machine with the smallest workload among the key machines for processing.

[0027] Preferably, in the step of obtaining a new population by performing crossover and mutation operations, when selecting machines, a multi-point crossover method is used for selection, and when sorting the processes, a workpiece exchange method is used for sorting. When performing mutation operations, a partially random method is used to select available machines, and when sorting the processes, a method of randomly exchanging process positions is used.

[0028] Preferably, in the step of obtaining a new population by performing crossover and mutation operations, the agent will take different actions to obtain the crossover probability and the mutation probability, and select multiple actions as the action space of the crossover probability and the mutation probability respectively. For the crossover probability, the range value is 0.4 to 0.9, and the interval between each action is 0.05; for the mutation probability, the range value is 0.01 to 0.21, and the interval between each action is 0.02, and a random number from 0 to 1 is added to each action.

[0029] Preferably, in the step of calculating the state and reward of the new population,

[0030] Compare the average fitness value and population diversity of the population before and after crossover mutation. If the weighted average fitness value and population diversity of the population increase, a positive reward is given, otherwise a negative reward is given. 1 and a 2 are all set to 0.5,

[0031]

[0032] Where R is the weighted value of the average fitness value of the population and the diversity of the population, r is the reward, and when R is greater than 1, the agent gives a positive reward of 1; otherwise, a negative reward of -1 is given.

[0033] Preferably, in the step of calculating the state and reward of the new population,

[0034] Compare the average fitness value and population diversity of the population before and after crossover mutation. If the weighted average fitness value and population diversity of the population increase, a positive reward is given, otherwise a negative reward is given. 1 and a 2 are all set to 0.5,

[0035]

[0036] Where R is the weighted value of the average fitness value of the population and the diversity of the population, r is the reward, and when R is greater than 1, the agent gives a positive reward of 1; otherwise, a negative reward of -1 is given.

[0037] Preferably, in the step of solving the problem by a knowledge-driven variable neighborhood search algorithm, four neighborhood operators are set, namely neighborhood operator 1, neighborhood operator 2, neighborhood operator 3 and neighborhood operator 4, neighborhood operator 1 is used to move the position of each process on the critical path to another optional machine set; neighborhood operator 2 is used to randomly select the key block position on the critical path, find the middle process in the key block and move it to the end of the key block, neighborhood operator 3 is used to randomly select the key block position on the critical path, find the middle process in the key block and move it to the beginning of the key block, and neighborhood operator 4 is used to increase the diversity of process positions on the critical path by randomly exchanging the process positions on the critical path.

[0038] Preferably, in the step of solving the problem by a knowledge-driven variable neighborhood search algorithm, four neighborhood operators are set, namely neighborhood operator 1, neighborhood operator 2, neighborhood operator 3 and neighborhood operator 4, neighborhood operator 1 is used to move the position of each process on the critical path to another optional machine set; neighborhood operator 2 is used to randomly select the key block position on the critical path, find the middle process in the key block and move it to the end of the key block, neighborhood operator 3 is used to randomly select the key block position on the critical path, find the middle process in the key block and move it to the beginning of the key block, and neighborhood operator 4 is used to increase the diversity of process positions on the critical path by randomly exchanging the process positions on the critical path.

[0039] The knowledge-driven evolutionary algorithm dynamic flexible semiconductor workshop scheduling method provided by the present invention introduces the uncertainty factors of workpiece transportation time, machine adjustment time and machine failure on the basis of traditional flexible workshop scheduling, and also considers energy-saving scheduling, so as to achieve energy-saving purpose while meeting production demand by reducing energy consumption. While improving efficiency, it can reduce energy consumption and enhance the market competitiveness and sustainable development ability of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 An example of a rescheduling Gantt chart provided in an embodiment of the present invention;

[0041] Figure 2 A process framework of a dynamic flexible operation semiconductor workshop scheduling method based on a knowledge-driven evolutionary algorithm provided in an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of two-stage encoding provided by an embodiment of the present invention;

[0043] Figure 4A schematic diagram of crossover and mutation operations provided by an embodiment of the present invention;

[0044] Figure 5 A schematic diagram of a neighborhood operator 1 provided in an embodiment of the present invention;

[0045] Figure 6 A schematic diagram of a neighborhood operator 2 provided in an embodiment of the present invention;

[0046] Figure 7 A schematic diagram of a neighborhood operator 3 provided in an embodiment of the present invention;

[0047] Figure 8 A schematic diagram of a neighborhood operator 4 provided in an embodiment of the present invention;

[0048] Fig. 9 A four-factor main effect diagram provided by an embodiment of the present invention;

[0049] Fig.10 An interaction diagram provided for an embodiment of the present invention.

[0050] Fig.11 A schematic diagram of the distribution of evaluation indicators provided by an embodiment of the present invention;

[0051] Fig.12 A schematic diagram of the distribution of non-dominated solutions of the algorithm provided in an embodiment of the present invention;

[0052] Fig.13 A Gantt chart before rescheduling under the MK02 instance provided in an embodiment of the present invention;

[0053] Fig.14 A Gantt chart after rescheduling under the MK02 instance provided in an embodiment of the present invention;

[0054] Fig.15 An improved and rescheduled Gantt chart under the MK02 instance provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0057] like Figure 2 As shown, it is a process framework of a dynamic flexible operation semiconductor workshop scheduling method of a knowledge-driven evolutionary algorithm provided in an embodiment of the present invention, and the method includes:

[0058] Acquiring workshop working condition data, wherein the working condition data at least includes workpiece data and machine data;

[0059] Construct a mathematical model for workshop scheduling based on workshop working condition data;

[0060] The workshop scheduling data model is solved based on genetic algorithm, Q-learning and knowledge-driven variable neighborhood search algorithm to generate a workshop scheduling plan.

[0061] In the present invention, taking aerospace parts as an example, the energy efficient dynamic flexible job shop scheduling problem with machine failure (EEDFJSP) is studied. The present invention proposes a knowledge-driven evolutionary algorithm for dynamic flexible job shop scheduling of semiconductors (KEARL), whose optimization objectives are maximum completion time (MCT), total energy consumption (TEC) and workload of critical machines (WCM).

[0062] In the EEDFJSP problem, multiple constraints need to be considered. Machines are assigned to multiple jobs for scheduling and the following three objectives are optimized: maximum completion time, total energy consumption, and critical load. An n×m EEDFJSP can be represented as n jobs J = {J 1 ,J 2 ,...,J n} and m machines M={M 1 ,M 2 ,...,M m}, each workpiece J i Contains several processes O ij , where O ij is the j-th process of workpiece i (j=1,2,3,...,h).

[0063] Machine failure generally includes the failed machine, the time of failure, and the time of repair. When a machine failure occurs, the failed machine will not perform any operations until it is repaired. Once a machine failure occurs, all unstarted operations will be rescheduled according to the rescheduling strategy. To simplify the problem, the following assumptions and constraints are made:

[0064] (1) All workpieces and all machines are ready at time zero.

[0065] (2) All machines have settling time when they start processing the first workpiece.

[0066] (3) Unlimited transportation resources allow workpieces to be transported directly to the next machine without waiting.

[0067] (4) The buffer in front of each machine is large enough.

[0068] (5) Strictly follow the priority order between processes.

[0069] (6) There are no priority constraints between the processes of different workpieces.

[0070] (7) Each machine can only process one workpiece at a time.

[0071] (8) Each process is processed only once using one machine.

[0072] (9) Once started, processing is not allowed to be interrupted unless a machine failure occurs.

[0073] (10) When a machine fails, the faulty machine shall be stopped and shut down immediately.

[0074] (11) The process being processed when the machine failure occurs will be reprocessed.

[0075] (12) After a machine fails, the faulty machine will be repaired within a certain period of time.

[0076] Machine failure is one of the most common dynamic events in a workshop, including the failed machine, the time when the failure started, and the time to repair. The probability of machine failure is usually related to the workload of the machine. Machines that have been in use for a long time are more likely to fail. This paper adopts the machine failure formula proposed by N.Al-Hinai.

[0077] The failure probability of machine k is as follows:

[0078]

[0079] Among them, T k : represents the working time of machine k, and R represents the total working time of all machines.

[0080] The machine failure start time is generated from the uniform distribution of formula (2):

[0081] SBT k =[α 1 OMT,α 2 OMT] (2)

[0082] Among them, OMT represents the maximum completion time of the original schedule. 1 and α 2 is generated between the interval [0,1]. 1 or α 2 When α is between 0 and 0.5, the failure occurs in the first half of the dispatch; 1 or α 2 When it is between 0.5 and 1, the failure occurs in the second half of the scheduling, as shown in Table 1.

[0083] For each machine M k , assuming that the repair time after failure is RT k , and follows the uniform distribution as shown in formula (3).

[0084] RT k =[β 1 T k ,β 2 T k ] (3)

[0085] Among them, β 1 and β 2 In the interval [0,1], T k Indicates machine M k . Working hours. 1 or β 2 The value of is limited to 0.1 to 0.15 to keep the failure level low. The machine failure level is determined by SBT k and RT k The present invention takes β 1 =0.15, β 2 =0.25.

[0086] Table 1 Machine failure levels

[0087]

[0088] The core idea of ​​the rescheduling strategy is to evenly distribute the affected workpieces to ensure similar processing time on each machine. This can reduce the probability of failure and shorten the completion time. This paper designs a new rescheduling strategy based on the traditional right-shift rescheduling to reduce the probability of secondary failure. The unprocessed workpieces are allocated to candidate machines according to the principle of minimum machine workload, aiming to balance the workload, shorten the time, and effectively reduce the occurrence of machine failures.

[0089] Table 2 gives a scheduling example with 3 workpieces and 4 machines, and Table 3 shows the transportation time between machines. Figure 1(a) shows the Gantt chart of this scheduling example. In order to clearly represent the corresponding process, i0j is used to represent the jth process of the i-th workpiece. At time 7, machine 3 fails and the failure duration is 1. The Gantt chart formed by right shift rescheduling is as follows: Figure 1 (b) In order to effectively reduce the probability of secondary failure, process 301 is moved to another candidate machine with a smaller workload, which not only reduces the maximum completion time but also makes the machine workload more balanced, as shown in Figure 1 (c) as shown.

[0090] Table 2 Processing and adjustment time

[0091]

[0092] Table 3 Transportation time

[0093]

[0094] The optimization objective of the workshop scheduling mathematical model is expressed as:

[0095]

[0096] min TEC=PEC+SEC+REC+IEC+CEC (5)

[0097]

[0098] Among them, in the workshop scheduling mathematical model, processing energy consumption, adjustment energy consumption, transportation energy consumption, idle energy consumption and general energy consumption are expressed as:

[0099]

[0100] In the mathematical model of shop scheduling, the constraints are expressed as:

[0101]

[0102] Among them, I and i' are workpiece indices, j and j' are process indices, K and k' are machine indices, I represents the workpiece set, I = {1, 2, ..., n}, J i represents the process set of workpiece i, J i ={1,2,…,h i}, K represents the machine set, K = {1, 2, ..., m}, JJ represents the Nth j -1 process set, JJ={1,2,…,N j -1},M i,j Indicates process O i,j Optional machine set, M i,j ={1,2,…,g ij}, Lk represents the location set of machine k, L k ={1,2,…,l k}, L k * The first L of machine k k -1 location set, L k * ={1,2,…,l k -1}, n represents the number of workpieces, h i represents the number of processes of workpiece i, m represents the total number of machines, and N j represents the total number of processes, l represents the location of the machine, O i,j represents the jth process of workpiece i, g i,j Indicates process O i,j The number of machines that can be selected, l k represents the maximum number of positions of machine k, PT i,j,k Indicates process O i,j Processing time on machine k, AP i,j,k Indicates process O i,j Unit processing energy consumption on machine k, AI k Indicates that machine k is in process O i,j Unit idle energy consumption, ST i,j,k Indicates process O i,j Adjustment time on machine k, AS i,j,k Indicates process O i,j Unit preparation energy consumption on machine k, LI represents a positive infinite number, TT i,j,k,k' Indicates process O i,j The transportation time from machine k to machine k', AC represents the unit general energy consumption in the workshop, UT represents the unit transportation energy consumption for transporting equipment, C i represents the completion time of job i, X i,j,k represents a binary constant, if process O i,j If it is processed on machine k, the value is 1; otherwise, it is 0. i,j,k,l represents a binary constant, if process O i,j If the machine k is processing at position l, the value is 1; otherwise, it is 0. i,j Indicates process O i,j Start time, ST k,l represents the start time of location l of machine k, CT k,l represents the makespan of machine k at location l, MCT represents the maximum makespan, TEC represents the total energy consumption, and WCM represents the critical load of the machine.

[0103] Equations (4)–(6) are the optimization objectives, including the maximum completion time, total energy consumption, and machine critical load.

[0104] Equations (7)–(11) represent processing energy consumption (PEC), adjustment energy consumption (SEC), transport energy consumption (REC), idle energy consumption (IEC), and common energy consumption (CEC).

[0105] Constraints (12)–(13) ensure that each process of each job can be assigned to only one available machine.

[0106] Constraint (14) requires that there can be only one process at a processing position of the machine, and two processes are not allowed to be assigned to the same position.

[0107] Constraint (15) states that a process can only be processed at a unique location.

[0108] Constraints (16) and (17) show the association between these variables.

[0109] Constraint (18) restricts the allocation of the current position to be made only after the previous free position has been allocated.

[0110] Constraint (19) indicates that the current process can only be started after the previous process of the workpiece is completed and transported to the corresponding machine.

[0111] Constraint (20) ensures that its start time is not less than the start time of the previous position.

[0112] Constraints (21) and (22) restrict SOi,j and STk,l to be equal.

[0113] Constraint (23) defines the maximum completion time.

[0114] Constraint (24) restricts continuous decision variables to be positive.

[0115] In the present invention, if Figure 2 As shown in Figure 2, four initialization strategies are used to obtain the initial population. Then, the Q-learning learning process is initialized and P is randomly selected. c and P m Within a certain range. Next, a new population is obtained through crossover-mutation operations, and their states and rewards are calculated. A suitable algorithm is selected to update the Q table. Subsequently, a knowledge-driven variable neighborhood search is performed to find a better solution. Finally, the target values ​​of all selected individuals are non-dominated and the crowding distance is calculated. Figure 2 The KEARL framework is presented.

[0116] In the present invention, two-part coding is used to implement the machine selection (MS) part and the operation sequencing (OS) part, such as Figure 3As shown. In the machine selection part, we arrange all the processes corresponding to each workpiece in order to complete the encoding and find the optional machines corresponding to each process. Then, each number represents that the corresponding process will be processed on the first machine in the optional machine set of the process; in the process sorting part, the number represents the workpiece number, and the first appearance of the number in the process sorting chromosome represents the first process of the corresponding workpiece. When decoding, the processes in the OS chromosome are decoded in sequence to generate the scheduling process, and then the corresponding selected machine number is assigned to each process in the OS. In this paper, insertion decoding is used to perform decoding.

[0117] In the initialization strategy set by the present invention, four initialization rules are designed to ensure the convergence speed of the fitness value and the quality of the initial population. These initialization strategies include: (1) completely randomly generating the initial solution, (2) generating the initial solution according to the minimum time, (3) preferentially generating the initial solution according to the maximum remaining processing time, and (4) preferentially allocating the machine with the smallest workload among the key machines for processing. In the initial population, in order to ensure the effectiveness of each strategy for the initial solution, the proportion of the four initialization strategies is set to 25%.

[0118] (1) Completely random generation of initial solutions. The purpose of this initialization strategy is to expand the diversity of the population. In the machine selection and process sorting parts, genes are randomly generated within a certain range.

[0119] (2) Generate the initial solution based on the minimum time. The purpose of this strategy is to effectively reduce the maximum completion time. In the machine selection part, the machine with the minimum time is selected; in the process sorting part, the gene is randomly generated within a certain range.

[0120] (3) Generate the initial solution based on the maximum remaining processing time. The purpose of this initialization strategy is to improve the quality of the solution by balancing the processing time of the processes in the workpiece. For the machine selection part, the gene is randomly generated within a certain range; in the process sorting part, the process corresponding to the workpiece with the longest remaining processing time is selected for processing first.

[0121] (4) Prioritize the key machines with the smallest workload for processing. The purpose of this initialization strategy is to reduce the workload of the machines so that the load of each machine is as balanced as possible. Reducing the workload of the machines means that the working time of each machine is shortened, thereby achieving the goal of minimizing the maximum completion time and total energy consumption. In the machine selection part, the machine with the smallest workload is selected from the available machines for processing. If the workload of multiple machines is the same, a machine is randomly selected for processing. In the process sorting part, the processes are randomly sorted.

[0122] Crossover and mutation are two core components. Crossover is used to increase the diversity of the population and discover new solutions, while mutation is used to increase the randomness of the optimization process. For the crossover method, it is divided into a machine selection part and a process sorting part. The machine selection part uses multi-point crossover, and the process sorting part uses workpiece exchange, such as Figure 4 As shown in (a). For the mutation method, it is also divided into two parts: machine selection and process sorting. The machine selection part randomly selects available machines, and the process sorting part randomly exchanges the positions of two processes, as shown in Figure 4 (b) as shown.

[0123] The present invention also adopts a parameter adaptation strategy based on RL. Both the Q-learning algorithm and the SARSA algorithm are RL methods, which belong to a mode of strategy selection. They continuously iterate and update the Q table through a limited number of state and action combinations, and can eventually converge, that is, the intelligent agent can judge the action it wants to select according to the current state by looking up the table. The Q-table update formulas of the Q-learning algorithm and the SARSA algorithm are shown in formulas (25) and (26) respectively.

[0124]

[0125]

[0126] Among them, Q(s t ,a t ) means in state s t Next, perform action a t Q value. max Q(s t+1 ,a) means in the next state s t+1 The highest Q value among all actions. represents the learning rate, which determines the impact of the new value on the current value; γ represents the discount factor, which determines the decay rate of future rewards.

[0127] In order to construct the state space of EEDFJSP, the fitness value needs to be calculated. Considering that the roulette wheel selection method is usually applied to single-objective optimization methods, the three-dimensional state quantity in the state space is converted into a one-dimensional state quantity by weighting, and then discretized and mapped. The fitness value function mainly considers the following aspects: (1) average fitness value, as shown in formula (27); (2) population diversity, as shown in formula (28); (3) optimal individual fitness value, as shown in formula (29); Formula (30) is the weighted addition of formulas (27), (28), and (29).

[0128]

[0129] Oi t represents the target value of the ith individual in the tth generation, minOi t represents the optimal target value of the ith individual in the tth generation, A(t) is the average fitness value of the tth generation, A(1) is the average fitness value of the first generation, and D(t), D(1), B(t), and B(1) have similar meanings to A(t) and A(1). 1 ,w 2 , and w 3 is the weight value, and w 1 +w 2 +w 3 =1. In this paper, after many experiments, it is found that the average fitness value and diversity of the population can reflect the state of the entire population, and it is easier to obtain excellent individuals, so w 1 ,w 2 , and w 3 Set them to 0.35, 0.35 and 0.3 respectively.

[0130] According to the overall fitness value, this paper divides the state space into 21 sub-states, where s = (s 1 ,s 2 ,...,s 21 ), when the weighted value is between [0,1], the interval value is 0.05, and every 0.05 is a state. When the fitness value is greater than 1, it is the 21st state.

[0131] For each generation, the agent will take different actions to obtain the crossover rate (Pc) and mutation rate (Pm), and select 10 actions as the action space of Pc and Pm respectively. For Pc, the commonly used range is 0.4 to 0.9, and the interval between each action is 0.05; for Pm, the commonly used range is 0.01 to 0.21, and the interval between each action is 0.02. A random number from 0 to 1 is added to each action to increase the flexibility of action selection.

[0132] In KEARL, the average fitness value and diversity of the population have a great impact on obtaining excellent individuals. Therefore, by comparing the average fitness value and population diversity of the population before and after crossover, if the weighted value of the average fitness value and population diversity of the population increases, a positive reward is given; otherwise, a negative reward is given. Since the average fitness value and diversity have little impact on excellent individuals, a 1 and a 2 Both are set to 0.5, as shown in equations (31) and (32):

[0133]

[0134] Where R is the weighted value of the average fitness value of the population and the diversity of the population. When R is greater than 1, the agent gives a positive reward of 1; otherwise, a negative reward of -1 is given.

[0135] Action selection strategy

[0136] The improved ε-greedy strategy is adopted, as shown in formula (33). In order to make the ε-greedy strategy converge better in the later stage of iteration, this paper designs an ε iterative decline strategy, in which ε will gradually decay with the increase of the number of iterations, and stop decaying after decaying to the set minimum value. This makes the algorithm tend to explore new solutions in the early stage, and tend to retain and use the learning results in the later stage. See formula (34) for details.

[0137]

[0138] ε=ε 1 -t((ε 1 -0.05) / max t) (13)

[0139] where rand is a random number in the range [0,1], ε 1 is the initial value of ε, and max t is the maximum number of iterations.

[0140] Variable neighborhood search avoids falling into local optimality by changing the neighborhood structure and exploring different areas in the solution space. How to design an effective variable neighborhood search strategy is crucial to the scheduling problem. The present invention optimizes the three main objectives of maximum completion time, total energy consumption and machine critical load for EEDFJSP. By introducing problem-based knowledge into variable neighborhood search, four knowledge-driven variable neighborhood search operators are designed, mainly considering the order of processes on the critical path. The critical path has the longest total duration in the scheduling plan, has the greatest impact on the completion time of the entire scheduling plan, and may have multiple critical paths. The first step is to find the critical process, and the search for the critical process is shown in Algorithm 1.

[0141]

[0142] Neighborhood operator 1 moves the position of each process on the critical path to another optional set of machines. Figure 5 As shown in the figure, the critical path consists of processes 201, 101, 301, and 302. Process 101 is randomly selected and moved to M3 for processing. After the move, process 301 on M2 is processed directly after process 201, and the start processing time of process 302 is also advanced accordingly, thereby reducing the maximum completion time by 2.

[0143] Neighborhood operator 2 randomly selects the key block position on the critical path, finds the intermediate process in the key block and moves it to the end of the key block. Figure 6As shown in Figure 1, the critical path consists of processes 201, 301, 101, and 102. First, find the critical block (201, 301, and 101), then find the intermediate process and move it to the end of the critical block. After rescheduling, the maximum completion time is reduced by 3.

[0144] Neighborhood operator 3 randomly selects the key block position on the critical path, finds the intermediate process in the key block and moves it to the beginning of the key block. Figure 7 As shown in the figure, the critical path consists of processes 301, 201, 202, 203, and 102. First, find the critical block (301, 201, and 101), then find its intermediate process and move it to the beginning of the critical block. After rescheduling, the maximum completion time is reduced by 1.

[0145] Neighborhood operator 4 increases the diversity of process positions on the critical path by randomly exchanging the process positions on the critical path. Figure 8 As shown, the critical path consists of processes 201, 301, 203, 102, and 302. After randomly exchanging the processing order of processes 203 and 102 on M3, the maximum completion time is reduced by 6.

[0146] In the present invention, an energy-saving scheduling strategy is adopted. The implementation of an effective energy-saving strategy can reduce energy consumption with minimal impact on other optimization objectives. Therefore, three energy-saving scheduling strategies based on three optimization objectives are designed as follows:

[0147] ES1: Move the last process to the machine with the shortest completion time among all the processing machines.

[0148] ES2: Move the final process to the machine with the lowest energy consumption

[0149] ES3: Move the last process to the machine with the smallest machine load among all processing machines.

[0150] Algorithm 2 gives the pseudo code of the energy-saving scheduling strategy.

[0151]

[0152] Experimental comparison and analysis:

[0153] Experimental setup:

[0154] The present invention evaluates the performance of KEARL in solving the EEDFJSP problem through three groups of experiments. All algorithms in the present invention are written in the MATLAB2023a environment, and the operating environment is 12th Gen Intel(R)Core(TM)i5-12600KF3.70GHz and 16.0GB RAM. To ensure the accuracy and effectiveness of the experiment, each algorithm is run independently 10 times on all instances, and the maximum number of iterations is set to 200. First, a multi-factor Taguchi orthogonal test is performed on the algorithm to determine the optimal parameter combination of the factors, and then the effectiveness of the four key strategies in KEARL is analyzed. Finally, KEARL is experimentally compared with five other cutting-edge algorithms to further verify the effectiveness of the algorithm. This paper uses three evaluation indicators: IGD, HV and SC to evaluate the performance of the above-mentioned related algorithms.

[0155] This paper uses a total of 20 FJSP benchmark instances: 10 instances (MK01-10), 5 instances (DP01-5), and 5 large-scale instances (la01-05). Among them, DP and la instances are randomly generated based on the 10 FJSP benchmark test instances MK01~MK10. The relevant test data can be extracted through the link https: / / pan.baidu.com / s / 1BOx2p0dj8UMd5nIjS__6UQ, and the extraction code is yzxx.

[0156] Parameter setting is crucial to the performance of KEARL in solving the EEDFJSP problem. This paper focuses on four key parameters, namely population size S, learning rate α, discount factor γ and initial greed factor ε. A four-factor four-level Taguchi orthogonal experimental design (DOE) is performed on these four parameters, and IGD is used as the evaluation index to determine the best combination of these parameters. Each parameter is set to S∈{60,80,100,120},α∈{0.6,0.7,0.8,0.9},γ∈{0.6,0.7,0.8,0.9},ε∈{0.6,0.7,0.8,0.9}, forming L16(4 4 )DOE orthogonal table, as shown in Table 4, a total of 16 groups of experiments were conducted, each group of experiments was conducted three times and the average value was calculated. The smaller the IGD value, the better the measurement parameter. Based on the data in Table 4, the following is obtained: Fig. 9 The results shown in Figure 2 show that the best parameter combination is S = 120, α = 0.6, γ = 0.9, and ε = 0.6. Fig.10 Analyze the interaction between the four parameters, Fig.10 It is a three-dimensional interaction diagram of the four parameters combined in pairs. The z-axis represents the evaluation index IGD. The smaller the IGD value, the stronger the interaction between the two.

[0157] KEARL vs. CPLEX: The mathematical model is solved using the CPLEX 12.9 solver, and the results are compared with those obtained by the KEARL algorithm. The present invention uses the benchmark test instance of MK01-10 for solving. The maximum running time of each instance is set to 1 hour (3600 seconds). Once this time is exceeded, CPLEX will stop running. Since the CPLEX solver is not suitable for solving multi-objective optimization problems, total energy consumption (TEC) is selected as the optimization target, which is more complex in the model. Selecting it as the optimization target not only represents the other two targets, but also objectively reflects the performance of the solver.

[0158] The comparison results of KEARL and CPLEX in 10 benchmark test instances are shown in Table 5. From the comparison results in the table, it can be seen that the results obtained by KEARL are better than those of CPLEX, and CPLEX cannot reach the optimal solution within 3600 seconds. The results of CPLEX verify the correctness of the KEARL algorithm proposed in this invention.

[0159] Table 5 Comparison of KEARL and CPLEX results

[0160]

[0161]

[0162] In order to verify the effectiveness of the four initialization strategies proposed in the present invention, KEARL was compared with KEARL (NO-IS) which did not adopt the initialization strategy. To avoid the randomness of the algorithm, each algorithm instance was run 10 times and the average value was taken. Table 6 shows the results of using the IGD evaluation index and its average value after KEARL and NO-IS were run 10 times under the MK01 instance, so that we can understand more clearly how the final data of each instance is obtained; Table 7 gives the results of IGD, HV and SC values ​​of KEARL and NO-IS under all instances. The bold font in the table indicates the best value under the same instance. The results show that KEARL outperforms NO-IS in the three evaluation indicators in most instances.

[0163] Table 6 IGD values ​​obtained by KEARL and NO-IS

[0164]

[0165] Table 7 Mean values ​​of IGD, HV and SC obtained by KEARL and NO-IS

[0166]

[0167]

[0168] In order to demonstrate the effectiveness of the parameter adaptation strategy proposed in the present invention, KEARL is compared with KEARL (NO-PAS) without the parameter adaptation strategy. Table 8 shows the results of IGD, HV and SC values ​​of KEARL and NO-PAS in all instances, and the bold fonts in the table represent the best values ​​in the same instance. KEARL outperforms NO-PAS in all three evaluation indicators in most cases. These results illustrate the effectiveness of the proposed parameter adaptation strategy.

[0169] Table 8 Mean values ​​of IGD, HV and SC obtained by KEARL and NO-PAS

[0170]

[0171] To demonstrate the effectiveness of the knowledge-driven variable neighborhood search strategy proposed in the present invention, KEARL was compared with KEARL (NO-VNSS) which does not adopt the knowledge-driven variable neighborhood search strategy. To avoid the randomness of the algorithm, each algorithm instance was run 10 times and the average value was taken. Table 9 gives the results of IGD, HV and SC values ​​of KEARL and NO-VNSS in all instances. The bold fonts in the table represent the best values ​​in the same instance. The results show that KEARL is better than NO-VNSS in most cases.

[0172] Table 9 Mean values ​​of IGD, HV and SC obtained by KEARL and NO-VNSS

[0173]

[0174] In order to verify the effectiveness of the energy-saving strategy proposed in this invention, we compared KEARL with KEARL (NO-EES) without adopting the energy-saving strategy. To avoid the randomness of the algorithm, each algorithm instance was run 10 times and the average value was taken. Table 10 gives the results of IGD, HV and SC values ​​of KEARL and NO-EES in all instances. The bold font in the table indicates the best value in the same instance. The results show that KEARL is better than NO-EES in all three evaluation indicators in most cases.

[0175] Table 10 Mean values ​​of IGD, HV and SC obtained by KEARL and NO-EES

[0176]

[0177]

[0178] In order to verify the effectiveness of the proposed KEARL algorithm, it is compared with five cutting-edge algorithms studied in the existing literature, namely DQNSGA, MOEA / D, MOICA, NSGA-II and NSGA-III. The performance of these six algorithms is evaluated by using three evaluation indicators (IGD, HV and SC). Tables 11, 12 and 13 show the comparison of the results using IGD, HV and SC as evaluation indicators under 20 instance conditions, respectively. The best values ​​in the tables are shown in bold. In order to better display the results, violin plots are drawn using the data in Table 11, as shown in Fig.11 (a); a violin plot was drawn using the data in Table 12, as shown in Fig.11 (b) shows that a violin plot was drawn using the data in Table 13. Fig.11 (c) As shown in the three figures, KEARL outperforms other comparison algorithms in the distribution of 20 instances.

[0179] In order to intuitively demonstrate the effect of each algorithm, a three-dimensional scatter plot of the six algorithms is provided. Due to limited space, this study selected three representative examples: DP01, la01, and MK06. Fig.12 (a) Fig.12 (b) and Fig.12 (c) shows the results of the six algorithms under three examples DP01, la01 and MK06. It can be seen from the figure that the KEARL algorithm proposed in the present invention is better than other comparison algorithms.

[0180] In order to clearly observe the improvement of rescheduling, this paper takes MK02 as an example. Fig.13 The Gantt chart of the initial scheduling plan for the MK02 instance is shown. The disturbance event occurrence time is set to 12, and the machine failure maintenance time is set to 6. The Gantt charts of the right shift rescheduling and improved rescheduling under the influence of the machine failure disturbance event are shown in Figure 1. Fig.14 and Fig.15 The red grid indicates the time period when the machine failed.

[0181] Table 11 Mean IGD values ​​obtained by KEARL, DQNSGA, MOEA / D, MOICA, NSGA-II, and NSGA-III

[0182]

[0183]

[0184] Table 12 Mean HV values ​​obtained by KEARL, DQNSGA, MOEA / D, MOICA, NSGA-II and NSGA-III

[0185]

[0186] Table 13 Mean SC values ​​obtained by KEARL, DQNSGA, MOEA / D, MOICA, NSGA-II and NSGA-III

[0187]

[0188]

[0189] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A dynamic flexible operation semiconductor workshop scheduling method based on knowledge-driven evolutionary algorithm, characterized in that: The method comprises: Acquiring workshop working condition data, wherein the working condition data at least includes workpiece data and machine data; Construct a mathematical model for workshop scheduling based on workshop working condition data; The workshop scheduling data model is solved based on genetic algorithm, Q-learning and knowledge-driven variable neighborhood search algorithm to generate a workshop scheduling plan.

2. The dynamic flexible operation semiconductor workshop scheduling method based on the knowledge-driven evolutionary algorithm according to claim 1 is characterized in that: The steps of solving the workshop scheduling data model based on genetic algorithm, Q-learning and knowledge-driven variable neighborhood search algorithm include: Obtain the initial population based on the initialization strategy; Initialize the Q-learning algorithm and set the crossover probability P c and mutation probability P m ; Obtain a new population by performing crossover and mutation operations, calculate the state and reward of the new population, and update the Q table; The solution is solved by using a knowledge-driven variable neighborhood search algorithm, which non-dominated sorts the target values ​​of all selected individuals and calculates the crowding distance. Perform multiple iterations until the iteration stop condition is met and output the workshop scheduling plan.

3. The dynamic flexible operation semiconductor workshop scheduling method based on the knowledge-driven evolutionary algorithm according to claim 2 is characterized in that: In the step of obtaining the initial population based on the initialization strategy, four initialization rules are used to obtain the initial population. The initialization rules are: completely random generation of initial solutions, generation of initial solutions based on minimum time, priority generation of initial solutions based on maximum remaining processing time, and priority allocation of key machines with the smallest workload for processing.

4. The dynamic flexible operation semiconductor workshop scheduling method based on the knowledge-driven evolutionary algorithm according to claim 3 is characterized in that: In the step of obtaining a new population by performing crossover and mutation operations, when selecting machines, a multi-point crossover method is used for selection, and when sorting processes, a workpiece exchange method is used for sorting. When performing mutation operations, a partially random method is used to select available machines, and when sorting processes, a random exchange of process positions is used.

5. The dynamic flexible operation semiconductor workshop scheduling method based on the knowledge-driven evolutionary algorithm according to claim 4 is characterized in that: In the steps of obtaining a new population by performing crossover and mutation operations, the agent will take different actions to obtain the crossover probability and mutation probability, and select multiple actions as the action space of the crossover probability and mutation probability respectively. For the crossover probability, the range value is 0.4 to 0.9, and the interval between each action is 0.05; for the mutation probability, the range value is 0.01 to 0.21, and the interval between each action is 0.02, and a random number from 0 to 1 is added to each action.

Images