Dynamic optimal operational control for complex systems with nonlinear external loop disturbances

1.   Introduction

The flexible job-shop scheduling problem (FJSP) is an extension of the traditional job-shop scheduling problem ^[1,2] and has been proven to be an NP-hard problem. Due to limitations of production equipment and costs, the majority of production workshops in China still rely on manual scheduling, resulting in low scheduling efficiency and difficulty in handling complex situations. On the other hand, companies that utilize intelligent scheduling methods ^[3,4,5] for production processing demonstrate superior efficiency and performance in highly complex production scenarios. Therefore, intelligent scheduling becomes particularly important, and the quality of algorithms plays a crucial role. In recent years, several metaheuristic algorithms have been proposed, such as particle swarm optimization (PSO) ^[6], grey wolf optimization (GWO) ^[7], and whale optimization algorithm (WOA) ^[8]. Scholars have achieved significant progress in solving shop scheduling problems by employing these emerging metaheuristic algorithms ^[9,10,11]. Shen et al. ^[12] improved the global and local optimization capabilities and diversity of a bird swarm by modifying the structure of the bird swarm, by introducing crossover and mutation operators, and by incorporating a variable neighborhood search algorithm in different bird swarms, thereby enhancing the algorithm's ability to escape local optima and improving convergence. Chen et al. ^[13] proposed a self-learning genetic algorithm based on reinforcement learning (RL) to intelligently adjust key parameters of the genetic algorithm (GA), leading to outstanding performance in solving flexible job-shop scheduling problems. Wang et al. ^[14] introduced a population improvement strategy to the genetic algorithm, significantly enhancing its effectiveness in flexible job-shop scheduling problems. Lin et al. ^[15] proposed a discrete PSO algorithm with adaptive inertia weight and verified its effectiveness and implementation through comparative experiments. N.I. Anuar et al. ^[16] presented an improved PSO algorithm with enhanced search capability through particle reinitialization, systematic switching of the best solution, and taboo search-based mutation. Zhao et al. ^[17] proposed an improved ant colony algorithm by enhancing the ant colony's pheromone updating method, thereby improving the algorithm's efficiency and optimization results. Luan et al. ^[18] proposed a novel metaheuristic algorithm, whale swarm algorithm (WSA), for flexible job-shop scheduling problems and verified its effectiveness. Yuan et al. ^[19] proposed a new hybrid harmony search algorithm (HHS), which enhances the local search capability of the harmony search (HS) algorithm. Li et al. ^[20] introduced a hybrid artificial bee colony algorithm based on taboo search, which improves the effectiveness of the algorithm by enhancing the quality of the population. In summary, each algorithm has its own advantages and disadvantages. Therefore, several improved algorithms that are proposed for the flexible job-shop scheduling problem have focused on addressing the shortcomings of the algorithms themselves, aiming to achieve more desirable results. However, there have been few attempts to improve the connection between the algorithm and the problem, specifically the encoding and decoding schemes. This situation may lead to a scenario where algorithm improvements reach their limits but there is no improvement in the flexible job-shop scheduling problem itself. This paper deals with that by proposing simultaneous improvements to the encoding and decoding approaches as well as the algorithm itself.

In this paper, a flexible job-shop scheduling model with the objective of minimizing the maximum completion time is established, and a novel algorithm that combines the dung beetle optimization algorithm and simulated annealing algorithm is used to solve this model. To address the issues of low convergence accuracy and susceptibility to local optima in flexible job-shop scheduling algorithms, a series of improvements are made to the dung beetle optimization. Considering the high complexity of dual-layer encoding for operations and machines, which increases the difficulty of optimization, this paper adopts a single-layer approach for process encoding and incorporates machine selection during decoding, significantly improving the quality of the initial solutions. During population update, the dung beetle optimization algorithm is applied for optimization, followed by simulated annealing operations on the population to enhance the algorithm's convergence speed and solution quality. Finally, through simulation experiments and comparisons with other algorithms, the effectiveness and superiority of the proposed algorithm are validated. Additionally, the feasibility of the algorithm is demonstrated through a case study of factory production.

2.   Problem description

The description of the flexible job-shop scheduling problem (FJSP) ^[21,22,23] is as follows: There are n jobs {J₁, J₂…J_n} that need to be processed on m machines {M₁, M₂, …M_m}. Each job J_i consists of one or more operations, and the sequence of operations is predetermined. O_ij represents the j-th operation of job J_i, which can be processed on one of the different machines, denoted as M_ij. The processing time for each operation varies depending on the machine used. The scheduling objective of FJSP is to select the most suitable machine for each operation O_ij, to determine the optimal processing sequence of operations for each machine, and to determine the start time for each operation, in order to optimize the performance metrics of the entire system. Therefore, the flexible job-shop scheduling problem consists of two sub-problems: determining the machine assignments for each job and determining the processing sequence on each machine.

Some flexible job-shop scheduling problem mathematical models are shown in Table 1.

In Eq (1), Ci represents the completion time of job J_i. Equation (2) indicates that the operation j of job J_i is processed on machine k, that X_ijk = 1, otherwise X_ijk = 0. This equation ensures that each operation of a job is processed on exactly one machine. Equation (3) introduces the variables S_ijk, which represents the start time of the operation j of job J_i on machine k, and t_ijk, which represents the processing time of the operation j of job J_i on machine k. S_i(j+1)k represents the start time of the operation j+1 of job J_i on machine k. Equation (3) enforces the constraint that the processing sequence of different operations of the same job must be respected. In Eq (4), E_ijk represents the completion time of the operation j of job J_i on machine k. This equation indicates that once an operation starts, it cannot be interrupted, and it also implies that a job can be processed at most on one machine at any given time. Equation (5) states that all jobs can start processing at time zero, meaning there are no restrictions on the initial start times of the jobs.

3.   Basic algorithm description

3.1.   Dung beetle optimizer

The dung beetle optimizer (DBO) ^[24], proposed by Professor Shen Bo's team at Donghua University in November 2022, is a novel swarm intelligence optimization algorithm. It follows their previous work on the sparrow search algorithm (SSA) ^[25]. The DBO algorithm is inspired by the behaviors of dung beetles, including rolling ball, dancing, breeding, foraging, and stealing. It was specifically designed for global optimization problems and first introduced in the 79th issue of the Journal of Supercomputing in the United States.

Five behavior patterns ^[24] and corresponding position update formulations are shown in Table 2.

In Eq (6), the symbol t represents the current iteration number, and x_i(t) represents the position of the dung beetle i in the population at the iteration t. The constant value k∈(0, 0.2] denotes the divergence factor, b is a constant value between 0 and 1, and α is a natural coefficient set to -1 or 1, where 1 represents no bias and -1 represents a deviation from the original direction. X^w represents the worst position in the current population, and ∆x is used to simulate the variation of light intensity.

Equation (7) introduces the variable θ, which is an angle between 0 and π that represents the divergence angle. |x_i(t)-x_i(t-1)| represents the difference between the position of the dung beetle i at the iteration t and its position at the iteration t-1. It is important to note that if θ equals 0, π/2, or π, the position of the dung beetle is not updated.

In Eq (8), X^* represents the current local best position, and Lb^* and Ub^* represent the lower and upper bounds of the breeding region, respectively. R = 1-t/T_max, where T_max represents the maximum number of iterations, and Lb and Ub are the upper and lower bounds of the optimization problem.

Equation (9) introduces B_i(t), which represents the position information of the dung beetle i at the iteration t. b₁ and b₂ are two independent random vectors of size 1×D, where D represents the dimension of the optimization problem.

In Eq (10), X^b represents the global best position, while Lb^b and Ub^b represent the lower and upper limits of the optimal foraging region.

Equation (11) involves the variable C₁, which is a random number following a normal distribution, and C₂ is a random vector belonging to the interval (0, 1).

Equation (12) introduces the variable g, which is a random vector of size 1×D following a normal distribution, and S represents a constant value.

The algorithm partitions the population into five segments and updates their positions based on the five behavioral patterns by using corresponding formulations presented in Table 2. Finally, the global population is updated to determine the optimal position. Figure 1 depicts the direction of dung beetle rolling, guided by sunlight navigation. Figure 2 showcases the dance behaviors that are performed by dung beetles when encountering obstacles, contrasting it with the tangent function. Figure 2(a) is the tangent function curve and 2(b) is a conceptual model of dung beetle dance behavior.

3.2.   Simulated annealing

The simulated annealing (SA) algorithm (Vecchi and Kirkpatrick 1983), that was proposed in 1983, is a well-established algorithm that has been widely used in the field of combinatorial optimization. This algorithm was inspired by the annealing process of solid materials. When the temperature of a solid is high, its particles move rapidly, and as the temperature decreases, the particles gradually reach a stable state. Due to the similarities between the SA algorithm and combinatorial optimization problems, it has been increasingly applied in the engineering domain. For solving minimization optimization problems, the main steps of SA are as follows:

Step 1: Set the initial parameters: a sufficiently large temperature T₀, cooling rate α, and the number of iterations k. Let T₀ = T.

Step 2: Initialize the iteration counter i = 1. Generate an initial solution X₀ randomly. Set X_best = X₀ and calculate the objective function value E(X_best).

Step 3: Generate a neighboring solution X_new based on X_best. Calculate E(X_new) and the objective function increment E = E(X_new)-E(X_best).

Step 4: If ∆E < 0, set X_best = X_new and accept the new solution as the current solution.

Step 5: If ∆E > 0, calculate the acceptance probability p = e^-∆E/T. Generate a random number r between 0 and 1. If r < p, accept the new solution by setting X_best = X_new. Otherwise, reject the new solution, and the original solution remains the best. Increment i by 1.

Step 6: Set T = α × T.

Step 7: Check the relationship between i and k. If i is less than k, repeat steps 3 to 5. Otherwise, exit the loop.

4.   Improved dung beetle optimization algorithm based on simulated annealing algorithm

Due to the limited performance of the original DBO algorithm in the context of flexible job-shop scheduling models, particularly in terms of low convergence accuracy and susceptibility to local optima, the simulated annealing (SA) algorithm is considered to be integrated with the DBO algorithm. SA can avoid local optimization to a large extent in the early stage and has high convergence accuracy in the later stage. An enhanced algorithm called dung beetle optimization with simulated annealing (DBO + SA) is proposed for the flexible job-shop scheduling problem. In the SA component, perturbed solutions are generated by applying different crossover methods to the initial solutions. Furthermore, the algorithm incorporates simulated annealing operations into each iteration, allowing for a higher probability of accepting suboptimal solutions during the early stages of iteration. This approach helps prevent the algorithm from getting stuck in local optima. As the number of iterations increases, the probability of accepting suboptimal solutions gradually decreases, leading to improved convergence accuracy in the final solution.

4.1.   Encoding and decoding

In current research, most scholars employ a two-stage encoding approach for flexible job-shop scheduling problems, involving work order sequencing and machine assignment selection ^[26,27,28]. Two-stage encoding requires longer lengths, which gives twice the total number of work orders when compared to single-layer encoding. This exponentially increases the number of possible permutations, posing challenges for optimization. Additionally, the initial solutions that are generated using two-stage encoding are based on completely random machine assignments, which results in poor initial solutions and slower convergence rates. To address these issues, this paper proposes a single-layer encoding approach that only encodes work orders without considering machine assignments. A unique decoding scheme is developed to correspond to the single-layer encoding. Integer encoding ^[29,30,31] is employed during the encoding process, while the decoded solutions are obtained by rounding real numbers to integers and then performing the decoding process.

Different from the majority of two-stage double-layer encoding approaches, this proposal introduces a change in the encoding scheme by using a single-layer encoding. Only the work orders of the jobs are encoded, and the length of the encoding is the total number of work orders across all jobs. Random initialization is employed to enhance the global search capability of the initial population, and machine assignments are determined during the decoding process. By adopting the single-layer encoding, the encoding workload can be significantly reduced, making the encoding process more concise. Besides, single-layer encoding has better pedagogical results for beginners. The specific steps of the encoding process are formulated as follows:

Step 1: L represents the total number of work orders, Li represents the total number of work orders for job i, 1≤i≤total number of jobs, and i is an integer. The numbers 1 to L1 represent the number of work order for job 1, the numbers L1+1 to L2 represent the number of work order for job 2, and so on, until all work orders for all jobs have corresponding numbers.

Step 2: Generate a zero vector, x, of length L. Fill the numbers 1 through L randomly into L positions in the vector. At this stage, x represents a complete individual of a dung beetle.

Step 3: Repeat Step 2 multiple times until the predetermined population size is achieved.

Since only the work orders are encoded without encoding the machines, decoding requires selecting the machines corresponding to the work orders before proceeding with subsequent operations. Machine selection involves considering various factors, including the completion time of the previous work order, whether there are other work orders to be processed on the selected machine, the earliest available start time if the machine is chosen with other work orders, and whether it affects the subsequent work orders. Additionally, attention should be given to situations where multiple machines are available, such as how to make the selection.

To address the aforementioned issue, the solution proposed in this paper involves recording the involvement of each machine in the processing. Each machine has its dedicated timetable, which records the time slots when it is engaged in processing. Similarly, each work order has a progress table that keeps track of the current operation number for that work order. Whenever a work order completes an operation, the corresponding entry in the progress table is incremented. Additionally, a decoding table is created, consisting of five columns: work order number, machine number, operation number, start time, and end time. The number of rows in the decoding table is equal to the total number of operations. Following these steps, when decoding an encoded element, the work order number and operation number can be obtained. By referencing the decoding table entries corresponding to that work order, the completion time of the previous operation can be determined. Combining this with the timetables of all available machines, the earliest start time for the current operation can be identified. The machines are then selected in chronological order until a suitable machine that can accommodate the operation is found. The accompanying decoding strategy yields relatively higher-quality initial solutions, facilitating subsequent optimization processes by accelerating convergence without compromising solution quality. The specific decoding rules are as follows:

Step 1: Identify the work order number and operation number corresponding to the encoded element.

Step 2: Utilizing the machine timetable, work-order operation progress table, and decoding table, retrieve the available machine codes for the current operation and their respective scheduling schemes based on the machines' availability.

Step 3: Based on the scheduled schemes for each machine, determine the available time slots for the current operation on each machine and arrange them in ascending order.

Step 4: Iterate through the available time slots and check if the current operation can fit within each interval. If a suitable slot is found, insert this operation. Then, proceed to Step 5.

Step 5: Update the machine timetable, work-order operation progress table, and decoding table accordingly. Then, return to Step 1 and repeat the process until all the encoding elements have been decoded. The process is shown in Figure 3.

4.2.   Algorithm and implementation

The specific steps of the DBO + SA algorithm are summarized in the flowchart in Figure 3.

Step 1: Set parameters and initialize the population: Set t = 1, define population size, maximum iteration count T_max, proportion of each behavior in the total population, deflection coefficient k, natural coefficient α₁, constant b and S, initial temperature T0, and cooling coefficient α₂. Create initial solutions and the initial population based on the encoding method described in Section 3.1.1.

Step 2: Calculate the objective function value: Decode the solutions using the decoding method described in Section 3.1.2 and retrieve the maximum value from the fifth column of the decoding table, which represents the objective function value.

Step 3: Position update: Update the positions of the individuals in the population according to their respective behavior patterns and the corresponding formulations. Note that the result of the certificate code calculated by the formula may be decimal. Before calculating the objective function value, the decimal values should be rounded to the nearest integers and then decoded.

Step 4: Simulated annealing operation on the population: Each individual in the population has a one-third probability of generating a perturbed solution using one of the three methods that are illustrated in Figures 4(a)–4(c). Assuming the original encoding is 1–5, random positions X1 and X2 are selected for subsequent operations. The inversion operation reverses the elements between X1 and X2, the exchange operation swaps the elements at positions X1 and X2, and the insertion operation inserts the element at X2 after the element at position X1. Each individual in the population generates a perturbed solution through these operations. If the perturbed solution is better than the original solution, it is accepted directly. If the perturbed solution is worse, it is accepted with a probability of p = e^-∆E/T, where ∆E is the difference in objective function value between the perturbed solution and the original solution. At the beginning of the iteration, when T has a large value, p is close to 1, resulting in a higher probability of accepting worse solutions. As the iteration progresses and T decreases, p approaches 0, reducing the probability of accepting worse solutions.

Step 5: Update the population, replace the global best individual and its fitness value, and set t = t + 1.

Step 6: If t > T or the stopping criteria are met, end the algorithm and output the current best solution. Otherwise, return to Step 3.

After multiple tests, it was found that the proportion of each behavior can determine the emphasis of the algorithm. Rolling behavior helps increase the algorithm's global search capability, while dancing behavior accelerates the algorithm's convergence speed, enhancing its efficiency in the local search space. Breeding, foraging, and stealing behaviors contribute to refining and improving the already found solutions, thus enhancing the algorithm's convergence precision. After multiple comparisons, the proportions of rolling, dancing, breeding, foraging, and stealing behavioral patterns were determined to be 27%, 3%, 30%, 20%, and 20%, respectively. Additionally, x was set to 0.9, y to 0.1, initial temperature to 3000, and cooling coefficient to 0.97.

5.   Comparative analysis of experiments

5.1.   Comparison of standard examples

To validate the effectiveness of the integrated algorithm proposed in this study, we selected the widely-used Brandimarte benchmark instance (MK01-MK10) ^[32] from the FJSP standard scheduling problem. For FJSP-related algorithms, there are numerous test cases available, with the most commonly used ones categorized into three types: Brandimarte cases, Kacem cases, and custom-built cases. Brandimarte cases are semi-flexible, Kacem cases are fully flexible, and custom-built cases are not convenient for comparing multiple algorithms. As semi-flexible scenarios better reflect the current real-world workshop conditions, this paper opts for Brandimarte cases as the testbed for algorithm evaluation.

The DBO+SA algorithm was implemented using Matlab2017b and executed on a system running Windows 10, with an Intel(R) Core(TM) i5-6300HQ CPU @ 2.30GHz and 16 GB of memory. Set the population size to 100 and the maximum iteration count to 100.

To validate the effectiveness of the improvements made by incorporating the simulated annealing approach in the DBO+SA algorithm, a comparative experiment was conducted between the basic DBO and DBO+SA. In order to mitigate the influence of randomness, the algorithm was executed 10 times. The term "Best" represents the best result obtained among the 10 simulation experiments, while "Avg" denotes the average optimization outcome over the 10 simulations. The experimental results are summarized in Table 3, where the problem instances are defined by their dimensions, with "m" representing the number of jobs and "n" representing the number of machines. The results indicate the total processing time required to complete all jobs. The best results are highlighted in bold red in Table 3.

Taking the MK03 instance from the aforementioned cases as an example, in Figure 6(a), it can be observed that the optimal value converges around the 20th generation and achieves a favorable convergence value. Moreover, Figure 6(b) illustrates that the average objective function value of the initial population is around 223. This demonstrates the effectiveness of the proposed encoding scheme in significantly improving the quality of the initial solutions. This further validates the superior convergence speed of the DBO+SA algorithm.

Figure 7 presents the Gantt chart generated based on the scheduling results for the MK03 instance. The chart uses the start time of the first operation as the reference zero time point, with the horizontal axis representing the total processing time in hours and the vertical axis denoting the machine number. Each rectangular block represents an operation, and the text within the block indicates the job number, operation number, and processing duration. For instance, "P(5, 6)" denotes that operation 6 of job 5. The vertical position of the block indicates the machine number on which the operation is scheduled, while the left and right coordinates of the block represent the start and end times of the operation, respectively.

Simultaneously, the DBO+SA algorithm was compared with other metaheuristic algorithms, including the basic genetic algorithm, the quantum whale optimization algorithm ^[33], the hybrid grey wolf optimization algorithm ^[34], and the hybrid genetic optimization algorithm proposed in ^[35]. The optimal results are highlighted in bold red in Table 4.

From Table 3, it can be observed that the fused DBO+SA algorithm outperforms the initial DBO algorithm in terms of both the optimal and average values. This confirms the effectiveness of the DBO+SA algorithm. In Table 4, when comparing the DBO+SA algorithm with other algorithms, it is found that the DBO+SA algorithm outperforms the algorithms mentioned in the literature in 6 out of 10 standard cases. This further demonstrates the superiority of the DBO+SA algorithm.

Finally, the convergence curves of each algorithm on instances MK01, MK02, MK06, and MK10 are depicted in Figure 8. Additionally, the Gantt charts of the scheduling solutions for DBO+SA on instances MK01, MK02, MK06, and MK10 are presented in Figures 9–12.

From Figure 8, it can be observed that the algorithm proposed in this paper exhibits a superior initial population than other algorithms. Consequently, it significantly increases the convergence speed without compromising convergence precision, thanks to the encoding and decoding strategies proposed in this paper. This finding provides a valuable insight for researchers, suggesting the potential for improvement through enhancements to encoding and decoding methods.

As can be seen from Figures 9–12, although the overall scheduling scheme has achieved the goal of minimizing the maximum completion time, the machine load varies greatly. Some machines have almost no downtime events, while others have a lot of idle time, which will cause machines with high loads to malfunction more frequently and increase costs. Therefore, it is necessary for relevant scholars to work on multi-objective FJSP to make FJSP more relevant to the real situation in the workshop.

5.2.   Simulation of a workshop study case

Although the algorithm presented in this paper performs well on standard test cases, validating its feasibility requires real-world workshop scheduling cases. To this end, we collected instance information from company X's precision workshop, which is a partially flexible job shop. The workshop is represented by a scale of 28 × 24, indicating 28 jobs, 24 machines, and a total of 130 operations. Due to space limitations, the instance information is divided into two parts and presented in the appendix. It has also been uploaded to the cloud for reference by readers ^[36]. Table 5 displays the job numbers, operation numbers, machine numbers available for use, and the processing times for the first 11 machines for each corresponding operation. Table 6 shows the processing times for the remaining 13 machines for each corresponding operation.

Figure 13 illustrates the layout of company X's precision workshop, which is divided into three main sections. The upper end in Figure 13 represents the office area, while the middle section corresponds to the processing area, housing a total of 24 machining tools. The letters and numbers in Figure 13 denote the machine models, such as "DMU60" referring to a five-axis CNC machine tool from the German company DMG Mori. The left lower corner represents Machine 1, with subsequent machines numbered sequentially from bottom to top. The above side is dedicated to inspection, storage, and polishing of products. The green cross-shaped area in the middle represents the pedestrian logistics pathway within the workshop. Additionally, the workshop is equipped with three overhead cranes with a lifting capacity of 5 tons each, responsible for handling heavier workpieces. Figures 14(a) and 14(b) present the actual view of the workshop's interior.

The 222 products, developed and prototyped recently by company X, are processed using a CNC 5-axis lathe. Each product typically undergoes more than 20 machining operations, 4–6 positioning steps, and 3 inspection procedures, resulting in a complex and intricate production process. Figure 15(a) and 15(b) show two three-dimensional models of a 222 product. Figure 16(a) and 16(b) depict a real scene where Machine 14 and Machine 3 are processing workpiece 8 and workpiece 4, respectively. Machine 3 is a 5-axis machining center DMU60. It is responsible for processing the side surface of workpiece 4. Machine 14 is a vertical lathe. It is responsible for processing the side surface of workpiece 8. Currently, the workshop adopts a manual scheduling approach, where human intervention is required to determine the next machining tool after completing each operation. This method is inefficient and labor-intensive. Taking the collected instance of 222 products mentioned in this paper as an example, excluding scheduling and handling time, the processing time for this instance is approximately 300 labor hours.

By applying the proposed single-layer encoding and DBO+SA algorithm optimization method presented in this paper, the flexible scheduling problem in the workshop can significantly improve production efficiency. With a population size of 200 and a maximum iteration count of 100 generations, the final result yields a maximum processing time of 146 labor hours for the given production tasks. Compared to the manual scheduling approach with a processing time of 300 labor hours, the production efficiency has increased by over 50%. Figure 17(a) illustrates the convergence curve of the global optimal solution's objective, and Figure 17(b) illustrates the convergence curve of the population's average objective function. In Figure 17(a), it can be observed that the optimal value converges after 15 iterations, demonstrating a relatively rapid convergence rate. In Figure 17(b), it can be seen that the average value also enters a convergence phase after around 80 generations. Figure 18 displays the Gantt chart of the scheduling plan for the 222 products. From the chart, it is evident that the last operation to be completed is the 6th operation of Job 4, processed by Machine 22. Therefore, the maximum completion time for the 222 jobs is determined by the completion time of the 6th operation of Job 4, which is 146 labor hours. The successful scheduling of this instance demonstrates the feasibility and superiority of the DBO+SA algorithm in flexible job-shop scheduling.

5.3.   Discussion

The students training details include:

(1) Algorithm improvement: using CEC2021 test functions, the performance of a variety of algorithms were all tested, and the possibility of fusion of different algorithms was analyzed. Finally, DBO, which has greater development potential, and SA, which is suitable for use in fusing various algorithms, were selected from the multiple algorithms.

(2) Coding and decoding: We tried the two-layer coding and decoding strategy based on machine allocation and process ordering and the unified coding strategy and found that they have their own advantages and disadvantages. Finally, we tried to combine the two coding strategies to form a single-layer coding and decoding strategy; after comparison, we found that the effect is good, and we finally determined the coding and decoding strategy in this paper.

(3) Simulation experiments: Students were sent to Company X for simulation data collection. They recorded the basic details of the parts in an order and followed each part through the entire workshop scheduling production process. Collected data was then organized for simulation experiments.

The students were able to gain a better understanding of real workshop conditions through this project, going beyond the sole goal of minimizing completion time. They began to consider more complex situations in the machining process, such as machine breakdowns, changes in machining schedules, and the load of the bottleneck machine and the total machine load, which allowed them to have a holistic view of a complex production shop. This further enhanced the students' view of a complex production plant. In terms of education, the project improved students' ability to consider practical aspects when solving problems and provided them with a way to solve problems.

6.   Conclusions

In this paper, the flexible job-shop scheduling problem (FJSP) was addressed with the objective of minimizing the maximum completion time in a manufacturing process. To tackle the issues of complex encoding and susceptibility to local optima in existing algorithms, an algorithm was proposed as DBO+SA by integrating the dung beetle optimization (DBO) and simulated annealing (SA) algorithms. The approach was achieved as follows:

First, a one-dimensional encoding scheme was employed to discretely initialize the population for the scheduling of operations in the flexible job-shop scheduling problem. This allows for the application of the algorithm to the flexible job-shop scheduling problem. Second, an adaptive decoding method was applied to decode the encoding. This involved selecting the processing machine for each operation and calculating the objective function value. Next, the dung beetle optimization and simulated annealing algorithms were fused together, aiming to enhance both global and local search capabilities while ensuring effective exploration of the search space. Test results were analyzed with comparisons using the improved algorithms from other literature on the MK series instances proposed by Brandimarte. The results demonstrated the effectiveness of the algorithm. Furthermore, real-world production instances were collected from Company X's precision workshop. By applying the scheduling algorithm proposed in this paper, the production efficiency improved 50% when compared with the existing manual scheduling method employed in the workshop. It should be noted that the algorithm and model proposed in this paper are only applicable to the single-objective FJSP with the objective function of minimizing the maximum completion time. While it can accommodate various scales of production scenarios, the considerations and objective function are relatively singular, and thus may not address complex production scenarios comprehensively.

This further validates the feasibility of the proposed encoding method and algorithm. Future work could involve the application of this algorithm to more complex workshop scheduling problems to expand its range of applications. In conclusion, the improvements made to the algorithms and scheduling models in this paper offer valuable insights into educational aspects of job-shop scheduling. For instance, the single-layer encoding proposed herein simplifies the coding process, making it more accessible for beginners. Additionally, the accompanying decoding strategy yields relatively higher-quality initial solutions, facilitating subsequent optimization processes by accelerating convergence without compromising solution quality.

The students were able to gain a better understanding of real workshop conditions through the exercise of this project and were not limited to the pursuit of completion time as a goal. Instead, they began to consider more complex situations in the machining process, such as machine breakdowns, changes in machining schedules, and the load of the bottleneck machine and the total machine load, providing them with a holistic view of a complex production shop. The students' view of the big picture in a complex production plant was further enhanced. In terms of education, the project improves students' ability to consider practical aspects when solving problems and provides them with a way to solve them.

Author contributions

Shuangji Yao: Conceptualization, Methodology creation, Investigation, Resources; Yunfei Guo: Writing - Original Draft, Data curation, Result Analysis; Botao Yang: Methodology, Writing ‒ review & editing; You Lv: Investigation, Result Analysis; Marco Ceccarelli: Conceptualization; Xiaoshuang Dai and Giuseppe Carbone: Investigation, Supervision. All authors have read and approved the final version of the manuscript for publication.

Acknowledgments

Project supported by the Foundation for Innovative Research Groups of the Natural Science Foundation of Hebei Province (Grant No. E2020203174) and Key Project of National Natural Science Foundation of China, Project (Grant No. U20A20332).

Conflict of interest

The authors declare that they have no conflicts of interest.

Ethics declaration

The authors declare that the ethics committee approval was waived for the study.

Appendixes

Appendix A