Multi-scene robot mixed flow assembly line balancing method and system
Through the mixed integer linear programming model and reinforcement learning technology, combined with multi-evolution operator selection and goal-oriented operator design, the flexibility and local optimization problems of robot mixed flow assembly line balancing optimization in multiple scenarios are solved, and production efficiency and energy consumption are achieved The dual optimization improves the flexibility and efficiency of the system.
Patent Information
- Application Number
- CN202411127400.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-26
AI Technical Summary
The existing technology of robot mixed-flow assembly line balance optimization methods lack flexible adaptability in multiple scenarios, easily fall into local optimal solutions, and it is difficult to achieve dual optimization of production efficiency and energy consumption in complex production environments.
Build a hybrid integer linear programming model and design a multi-objective multi-factor evolution algorithm based on reinforcement learning. By accurately integrating multi-objective optimization and reinforcement learning technology, the balance optimization of production efficiency and energy consumption is achieved, and multi-evolution operator selection and goal orientation are adopted. Neighborhood search operator design, combined with adaptive operator selection of reinforcement learning, dynamically adjust optimization strategies.
It significantly improves the algorithm's global search capability and solution effect, ensures that the dual optimization of production efficiency and energy consumption is achieved in multiple scenarios, avoids falling into local optimization, and improves the flexibility and efficiency of the system.
Smart Images

Figure CN119024791A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of robot mixed-flow production, and in particular to a robot mixed-flow assembly line balancing method and system in multiple scenarios. Background Art
[0002] After the introduction of robots, the production efficiency of mixed-flow assembly lines has been significantly improved; however, current research focuses on design and optimization under ideal conditions, but often ignores unexpected situations that may occur in the actual production process, such as robot maintenance, sudden failures, and product model switching; these neglected factors make it difficult for the system to achieve the expected efficiency and stability when facing complex production environments.
[0003] Currently, the technologies that are most similar to the present invention are those mixed-flow assembly line balance optimization methods for a single scenario; these methods mainly focus on optimizing the assembly line production cycle; for example, traditional optimization algorithms, such as the genetic algorithm for a mixed-flow assembly line balance allocation method and device under uncertain demand with publication number CN114415615 A, the flying bird algorithm for a mixed-flow assembly line balance sorting optimization method and device with publication number CN114415615 A, as well as the ant colony algorithm and particle swarm optimization, have been widely used in the mixed-flow assembly line balance problem.
[0004] However, these algorithms are generally based on the assumption that the production environment is stable and unchanged, and lack the ability to flexibly adapt to multiple scenarios. First, existing optimization algorithms are often based on a single scenario assumption, that is, it is assumed that the production environment always remains stable. However, this assumption ignores the unavailability of robots due to maintenance or failures, and also ignores the complexity that may be caused by switching between different product models. Secondly, it is easy to fall into the dilemma of local optimal solutions during the solution process, especially in complex production environments, where it is difficult to ensure that the global optimal solution is found. Thirdly, there are limitations in adaptability. Faced with diverse production scenarios, these methods lack adaptive capabilities and it is difficult to flexibly adjust optimization strategies according to actual conditions. Finally, they focus on improving production efficiency, while relatively insufficient consideration is given to energy consumption optimization. However, in actual production, energy consumption is also a crucial consideration. Summary of the invention
[0005] In view of the shortcomings of the existing methods, the improvement of the present invention lies in that a mixed integer linear programming model is constructed for the balancing problem of multi-scenario robot mixed-flow assembly lines, and a multi-objective and multi-factor evolutionary algorithm based on reinforcement learning is designed. By accurately integrating multi-objective optimization and reinforcement learning technology, the dual optimization goals of production efficiency and energy consumption in diversified production scenarios are successfully achieved; at the same time, in the optimization process, attention is always paid to the balance between production efficiency and energy consumption, and the multi-objective optimization method is used to ensure that energy consumption is effectively reduced while improving production efficiency.
[0006] The technical solution adopted by the present invention is: a robot mixed-flow assembly line balancing method under multiple scenarios comprises the following steps:
[0007] Step 1: Construct a mixed integer linear programming model for the robot's product beat, energy consumption of processed products, and product process changes in a certain scenario, and perform target optimization and constraints on the mixed integer linear programming model;
[0008] In a preferred embodiment of the present invention, the target optimization formula for the energy consumption of the processed product is:
[0009] MinObj 2 =∑ c∈c ∑ m∈M EC mc (2)
[0010]
[0011] Among them, R is the robot set, R c is the set of unavailable robots in scenario c, J is the set of workstations, I is the set of processes, C is the set of scenarios, m is the product number, c is the scenario number, i is the process number, j is the workstation number, M is the product set, and t imr is the time required for robot r to process product m in the i-th process, PEC r is the energy consumption per unit time of robot r in processing state, CT mc is the beat of product m in scenario c, x ijc In scenario c, process i is assigned to workstation j, y rj For robot r to be assigned to workstation j, IEC r is the energy consumption per unit time of robot r in idle state.
[0012] Step 2: Encode and decode normal production scenarios and unavailable scenarios;
[0013] In a preferred embodiment of the present invention, encoding and decoding a normal production scenario includes:
[0014] Coding: Construct process sorting vector, process allocation vector and robot allocation vector;
[0015] Decoding: The robot allocation plan is obtained by sorting the robot weights in descending order.
[0016] Step 3: Calculate individual fitness for normal production scenarios and unavailable scenarios;
[0017] In a preferred embodiment of the present invention, the individual fitness calculation of the normal production scenario includes:
[0018] According to the process allocation results, the total processing time of each product in each workstation is calculated, and the maximum processing time is selected as the beat. The processing energy consumption is obtained by multiplying the robot's processing energy consumption coefficient per unit time and the processing time; the idle energy consumption is calculated based on the difference between the robot's processing time and the beat, multiplied by the idle energy consumption coefficient per unit time; the total energy consumption is obtained by summing the processing energy consumption and the idle energy consumption.
[0019] In a preferred embodiment of the present invention, the individual fitness calculation of the unavailable scenario includes:
[0020] The best individual is selected from the normal production scenario and combined with the individuals in the unavailable scenario for comprehensive evaluation; for the calculation of process changes, the workstation numbers assigned to the same process in the two scenarios are first identified, and the absolute value of the difference between the two numbers is calculated.
[0021] Step 4: Select multiple evolution operators;
[0022] In a preferred embodiment of the present invention, step 4 specifically includes:
[0023] First, the order of each evolutionary operator is randomly shuffled; second, each evolutionary operator is assigned to the corresponding population in turn; the evolutionary operator is used to update the corresponding population and generate the corresponding sub-population.
[0024] Step 5: Design goal-oriented neighborhood search operators for different scenarios including production cycle and robot allocation plan;
[0025] In a preferred embodiment of the present invention, the design of the target-oriented neighborhood search operator in different scenarios of the robot allocation plan includes:
[0026] Evaluate the energy consumption of each product at each workstation, reallocate the processes at the workstation with the highest energy consumption to available workstations; move the processes at a workstation to the workstation with the lowest energy consumption; calculate the total energy consumption of each workstation, and swap the robots at the workstation with the highest and lowest energy consumption.
[0027] Step 6: Adaptive neighborhood search operator selection based on reinforcement learning;
[0028] In a preferred embodiment of the present invention, step six specifically includes:
[0029] Step 61, designing an action including a plurality of states and a plurality of neighborhood search operators;
[0030] Step 62: Use formula
[0031] Dynamically update the state and action value; among them, Q t+1 (S t ,A t ) and Q(S t,A t ) represent the state S t Next action A t The updated value and current value of the corresponding element, R t+1 In state S t Next, execute action A t The reward value obtained after α and β represent the learning and decay coefficients respectively.
[0032] A preferred embodiment of the present invention is a robot mixed-flow assembly line balancing system in multiple scenarios, comprising: a memory for storing instructions executable by a processor; a processor for executing the instructions to implement a robot mixed-flow assembly line balancing method in multiple scenarios.
[0033] A preferred embodiment of the present invention is a computer-readable medium storing computer program codes, which, when executed by a processor, implement a robot mixed-flow assembly line balancing method in multiple scenarios.
[0034] Beneficial effects of the present invention:
[0035] 1. Based on the mixed integer linear programming (MILP) model, the robot mixed-flow assembly line balancing problem in multiple scenarios is accurately defined. This model not only takes into account production efficiency but also reduces the energy consumption of the robot;
[0036] 2. A unique multi-evolution operator selection mechanism is designed, which effectively prevents the algorithm from falling into local optimality by randomly selecting evolution operators in each iteration. This strategy increases the diversity of the search space and significantly improves the global search capability of the algorithm.
[0037] 3. 12 goal-oriented neighborhood search operators are carefully designed for various production scenarios and optimization goals (efficiency and energy consumption); these operators can efficiently explore the solution space, significantly improve the algorithm's solution effect, and ensure that the algorithm is more accurate and efficient in finding the optimal solution;
[0038] 4. Adaptive operator selection mechanism based on reinforcement learning. The performance of the neighborhood search operator is evaluated in real time, and the operator selection strategy is adjusted dynamically accordingly. That is, the search operator can be adaptively selected for each scenario based on real-time performance indicators. This mechanism ensures that the algorithm can flexibly adjust the optimization strategy in different scenarios, thereby greatly improving the optimization effect and efficiency.
[0039] 5. The reinforcement learning-based multi-objective multi-factor evolutionary algorithm (RLMOMFEA) can optimize multiple objectives (beat, energy consumption, process changes) at the same time, and effectively respond to multi-scenario challenges in complex production environments through multi-factor optimization methods;
[0040] 6. Comprehensively consider the multi-objective optimization of production efficiency and energy consumption. In the process of model or algorithm optimization, production efficiency and energy consumption are always taken as the core considerations. Through multi-objective optimization methods, efforts are made to find the best balance between the two to achieve the dual goals of energy conservation and emission reduction and efficient production;
[0041] 7. Through a large number of numerical experiments, the excellent performance of the RLMOMFEA algorithm in solving the robot mixed-flow assembly line balancing problem in multiple scenarios has been verified. The experimental results show that the algorithm surpasses the existing single-task and multi-task evolutionary algorithms in terms of convergence speed and solution diversity, and has significant practical application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a logic diagram of the robot mixed-flow assembly line balancing method under multiple scenarios of the present invention;
[0043] Figure 2 is a specific example of the encoding strategy of the present invention;
[0044] Figure 3 This is an example of individual encoding and decoding in the normal production scenario of the present invention;
[0045] Figure 4 It is an example of individual encoding and decoding in the unusable scenario of the present invention;
[0046] Figure 5 is an example of a feasible scheduling scheme of the present invention;
[0047] Figure 6 is the mean value diagram of RLMOMFEA and three variant algorithms GD, IGD and HVR when the confidence interval of the present invention is 95%;
[0048] Figure 7 It is the mean graph of RLMOMFEA and six single-task evolutionary algorithms GD, IGD and HVR when the confidence interval of the present invention is 95%;
[0049] Figure 8 It is the mean graph of RLMOMFEA and six multi-task evolutionary algorithms GD, IGD and HVR when the confidence interval of the present invention is 95%. DETAILED DESCRIPTION
[0050] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.
[0051] like Figure 1 As shown, the robot mixed-flow assembly line balancing method under multiple scenarios includes the following steps:
[0052] Case data: covers the number of processes, number of workstations, number of robots, priority relationships between processes, set of unavailable robots, and the time required for each robot to process the process. It also includes detailed records of the energy consumption per unit of processing time and idle time of the robot.
[0053] Algorithm parameters: including the number of individuals in the population corresponding to each scenario, as well as the core parameters of the multi-factor evolutionary algorithm, such as random mating probability (rmp), and the parameter P in reinforcement learning Q , α and β;
[0054] Population initialization: For each scenario, an initial population is generated to ensure that the number of individuals in each population is equal, providing a basis for subsequent algorithm iterations.
[0055] Generate code: For individuals in normal production scenarios, randomly generate codes containing the number of processes, the number of workstations, and the number of robots; in unavailable scenarios, randomly generate codes containing only the number of processes and the number of workstations. Each element in the code is a random number between 0 and 1, providing a data basis for the subsequent decoding process.
[0056] Decoding: Based on the codes of each individual, a detailed process sequence and the number of processes assigned to each workstation are generated. Furthermore, a corresponding process allocation plan and robot allocation plan are generated for each individual.
[0057] Archive set generation: extract non-dominated individuals from the corresponding population of each scene, and save these high-quality individuals in the archive set for subsequent analysis reference.
[0058] Fitness calculation: For individuals in normal production scenarios, the load of each product at each workstation is calculated based on the process and robot allocation plan obtained, and then the corresponding sum of beats and energy consumption is obtained. For individuals in unavailable scenarios, the optimal individual in normal production scenarios is used as a reference, combined with its process allocation and robot allocation information, to calculate the individual's beat, energy consumption and process changes.
[0059] Population decomposition: Divide a unified population into multiple small populations, each of which independently stores individuals suitable for each scenario to achieve more efficient and accurate optimization.
[0060] Randomly assign evolution operators: Before each iteration, the order of evolution operators is shuffled and randomly assigned to each population to increase the diversity and flexibility of the algorithm.
[0061] Neighborhood search operator selection: By comparing P Q The relationship between the size of the random number and the selection of a suitable neighborhood search operator; if P QIf P is greater than the random number, the neighborhood search operator with the largest Q-value is selected based on the current scene state; Q If it is less than the random number, a neighborhood search operator is randomly selected.
[0062] Population update: Update the population in each scenario and generate sub-populations to provide new candidate solutions for the next round of iteration.
[0063] Fitness calculation (sub-population): The fitness of individuals in the sub-population is calculated to evaluate their quality.
[0064] Q-table update: According to the difference in IGD value between the sub-population and the original population in each scenario, the corresponding Q-table is updated to provide guidance for the subsequent neighborhood search operator selection.
[0065] Population combination: Combine the original population and sub-population in each scenario to enrich the diversity of the population and retain high-quality individuals.
[0066] Form a new population: Select better individuals from the combined population to form a new population, providing high-quality candidate solutions for the next round of iteration.
[0067] Update of non-dominated solution set: extract non-dominated individuals from the combined population and save them in the archive set, and delete the dominated individuals in each scenario in the archive set to maintain the efficiency and accuracy of the archive set.
[0068] Termination condition judgment: judge whether the algorithm running time exceeds the given threshold. If it exceeds, output the calculation result; otherwise, enter the next cycle to continue optimization.
[0069] Result output: The corresponding archive set under each scenario is finally output;
[0070] Step 1: Construct a mixed integer linear programming model;
[0071] The multi-scenario robot mixed-flow assembly line balancing problem covers a variety of production environments, mainly covering a normal production scenario and several potential unavailable scenarios; in normal production scenarios, all robots can process products normally; however, when encountering unavailable scenarios, one or more robots will temporarily lose their ability to work due to maintenance or failure; at this time, when the production environment switches to these unavailable scenarios, the workstations assigned to the affected robots will not be able to continue processing products, and the processes originally assigned to them will need to be flexibly deployed to other workstations; after the robots are repaired, the previously adjusted processes will be able to return to their original workstations for continued processing.
[0072] Parameter Definition
[0073]
[0074]
[0075] Mixed Integer Linear Programming Model:
[0076] MinObj 1 =∑ c∈C ∑ m∈M CT mc (1)
[0077] MinObj 2 =∑ c∈C ∑ m∈M EC mc (2)
[0078] MinObj 3 =∑ c>1 ∑ i∈I ∑ m∈M Ta imc (3)
[0079] Subject to:
[0080]
[0081]
[0082] Among them, formulas (1-3) clearly define the core objectives of mixed integer scale model optimization; specifically, formula (1) aims to minimize the sum of beats of all products in all scenarios; formula (2) focuses on minimizing the total energy consumption of all products in all scenarios; and formula (3) focuses on minimizing the sum of process changes of all products in all unavailable scenarios.
[0083] Regarding the setting of constraint conditions, formula (4) ensures that each process can only be assigned to a specific workstation in all scenarios; formula (5) ensures that the process will not be assigned to a workstation that cannot be executed due to robot unavailability in a specific scenario; formula (6) stipulates that each available robot must process at least one process in each scenario; in addition, formulas (7) and (8) together ensure that only one robot is assigned to each workstation, and each robot is only assigned to one workstation; formula (9) stipulates that each process can only be processed after its predecessor process is completed; formula (10) limits the total processing time of each product at each workstation to ensure that the cycle time of the product is not exceeded.
[0084] Formula (11) calculates in detail the total energy consumption of each product in each scenario, including the energy consumption in working state and idle state; while formulas (12) and (13) focus on calculating the process changes of each process of each product in a specific scenario; in addition, formulas (14) to (16) impose non-negative constraints on variables such as cycle time, energy consumption and process changes to ensure the effectiveness and rationality of the model; finally, formulas (17) and (18) clearly define the binary variables used in the model, providing a basis for the construction and solution of the model.
[0085] By flexibly adapting to and switching between a variety of production scenarios, the robot's production capacity can be maximized. The core of the robot mixed-flow assembly line balance problem in multiple scenarios lies in how to effectively allocate the processes of various products to each workstation in different scenarios and select the most matching robot for each workstation, so as to achieve multiple goals such as optimizing the production efficiency of the mixed-flow assembly line and reducing the robot's energy consumption.
[0086] Construct a multi-objective and multi-factor evolutionary algorithm based on reinforcement learning, including:
[0087] Step 2: Encoding and decoding based on scene decomposition;
[0088] By evolving each scenario independently and effectively transmitting information between scenarios, the problem can be effectively solved;
[0089] The balancing problem of robot mixed-flow assembly lines in multiple scenarios includes both normal production situations and multiple unavailable situations. Therefore, in order to cope with this complexity, a complete scheduling solution should be composed of multiple independent parts, each of which is accurately mapped to a process and robot allocation plan in a specific scenario, that is, the total number of scenarios. In view of the relative independence between the various production scenarios, the present invention adopts an efficient strategy, that is, to generate a population for each production scenario, and each individual in the population focuses on optimizing the scheduling solution in its corresponding scenario. Figure 2 Provide specific examples of this coding strategy;
[0090] Under normal production conditions, two key decision points, process sorting and robot allocation, need to be considered comprehensively. Therefore, three vectors are used to express them clearly: process sorting, process allocation, and robot allocation. In contrast, for unavailable scenarios, we only need to focus on optimizing process allocation. Therefore, only two variables, process sorting and process allocation, are used to represent them.
[0091] The process sorting vector is used to clarify the execution order of each process. The vector contains N elements, where N is the total number of processes. Each element is a random number between 0 and 1, representing the weight of the corresponding process. The larger the weight, the earlier the processing time of the process will be.
[0092] The process allocation vector is responsible for determining the number of processes undertaken by each workstation; the vector contains J elements, each element represents the weight of a workstation; the higher the weight, the more processes the workstation will obtain.
[0093] The robot allocation vector focuses on assigning an appropriate workstation to each robot. The vector contains R elements, where R represents the total number of robots. Each element represents the weight of the corresponding robot. The larger the weight, the closer the assigned workstation will be.
[0094] Subsequently, the individual codes are decoded to obtain detailed process allocation and robot allocation plans; the robot allocation plan is generated by sorting the robot weights in descending order; and the process allocation plan is obtained by combining the information of the process sorting vector and the process allocation vector. The former mainly generates the processing sequence based on the process weights and the priority relationship between processes, and the latter determines the number of processes for each workstation by calculating the proportion of each workstation weight in the total weights of all workstations. For example, assuming that the weights of the four workstations are [0.54, 0.42, 0.69, 0.20], and the total number of processes is 9, then workstation 1 can be allocated about 3 processes, workstation 2 can be allocated about 2 processes, 3 processes are allocated to workstation 3, and the remaining 1 process will be allocated to the last workstation to ensure that all processes are allocated.
[0095] For individuals in unavailable scenarios, it is difficult to determine which workstations are available or unavailable because the specific locations of unavailable robots are unclear. In this case, it is necessary to estimate the number of unavailable robots in advance and calculate the number of available workstations accordingly. Then, the processes are allocated to each workstation in advance. Once the unavailable workstations are identified, the process allocation plan will be adjusted to adapt to the actual availability of each workstation.
[0096] Step 3: Calculate individual fitness in each scenario;
[0097] Based on the scenario decomposition, for the normal production scenario, the individual only needs to calculate two core goals: the sum of the cycle times of all products and the sum of the energy consumption; however, in the unavailable scenario, in addition to the above two goals, the individual also needs to calculate the sum of the changes in the process when transitioning from the normal production scenario to the unavailable scenario.
[0098] In the population of normal production scenarios, each individual contains detailed key information such as process allocation and robot allocation; this enables individuals to independently perform fitness assessments; specifically, based on the process allocation results, the total processing time of each product in each workstation is calculated, and the maximum value is selected as the beat; the energy consumption calculation includes two parts: processing energy consumption and robot idle energy consumption; the processing energy consumption is obtained by multiplying the robot's processing energy consumption coefficient per unit time and the processing time, while the idle energy consumption is calculated based on the difference between the robot's processing time and the beat, multiplied by the idle energy consumption coefficient per unit time; adding the two together can give the total energy consumption of the individual in a normal production scenario.
[0099] For the population in the unavailable scenario, each individual only contains basic information about the process allocation, and detailed information such as the robots configured at each workstation cannot be directly obtained; therefore, when evaluating the fitness of individual x in the unavailable scenario, it is necessary to select the best individual y from the normal production scenario and combine it with x for comprehensive evaluation; the calculation method of cycle time and energy consumption follows the established process in the normal production scenario; for the calculation of process changes, first identify the serial numbers of the workstations assigned to the same process in the two scenarios, and calculate the absolute value of the difference between the two serial numbers, so as to obtain the change in the process between the two scenarios.
[0100] Step 4: Select multiple evolution operators;
[0101] The existing multi-factor evolutionary algorithm mainly relies on a single evolutionary operator to update the population during the evolution process. However, due to the differences in different scenarios, the applicability of a single operator may be limited, which may affect the overall performance of the algorithm. In order to overcome this limitation, a multi-evolutionary operator selection strategy is proposed; that is, before each iteration starts, a different evolutionary operator is randomly selected for each scenario for iteration to promote its evolutionary efficiency and prevent it from falling into the local optimum; in this embodiment, only three production scenarios are considered, so three evolutionary operators are taken as an example; the first evolutionary operator is to randomly select an individual from all populations, the second evolutionary operator is to select an individual from the first level of the population under all scenarios, and the third evolutionary operator is to select an individual from the archive set; these selected evolutionary operators will be used as parents together with each individual in the population, and offspring will be generated after crossover and saved in the sub-population.
[0102] Table 1 The process of allocating evolutionary operators to each sub-population
[0103]
[0104]
[0105] Table 1 describes in detail the process of assigning evolution operators to each sub-population. First, the order of each evolution operator is randomly shuffled, and then the first operator is assigned to the first population, and the second operator is assigned to the second population in turn until all the assignments are completed. It is worth noting that this operator assignment is not a one-time operation, but is repeated at the beginning of each iteration to optimize the diversity and exploration ability of each sub-population.
[0106] This allocation strategy has significant advantages: first, it ensures that all scenarios can benefit from the application of multiple evolutionary operators, thereby reducing the limitations of a single operator on the algorithm performance; second, by assigning different operators to different scenarios, the diversity within each population can be enhanced, and it is possible to discover more diverse and more effective scheduling solutions; finally, the different characteristics of evolutionary operators help to fully explore and find scheduling solutions suitable for each specific scenario.
[0107] Step 5: Design a goal-oriented neighborhood search operator for the robot allocation plan;
[0108] In order to prevent the algorithm from falling into the dilemma of local optimal solution, this algorithm carefully designs 12 goal-oriented neighborhood search operators for various scenarios; among them, the first eight operators have wide applicability, the first six operators focus on optimizing production rhythm, and the last two focus on reducing robot energy consumption; operators 9 and 10 focus on optimizing the robot allocation strategy in normal production scenarios to ensure the smoothness of the production process; and operators 11 and 12 optimize the target by adjusting process changes for unavailable scenarios;
[0109] The following are detailed descriptions of these operators:
[0110] Operator 1: Determine the load of each workstation by calculating the processing time of each product at each workstation; randomly select a process from the heaviest-loaded workstation and randomly assign it to other available workstations; during the assignment process, ensure that the order of process assignment conforms to the logic of the production process, and that other processes cannot be assigned to workstations assigned to unavailable robots; for individuals in abnormal operation scenarios, due to the lack of robot assignment information, the best individual will be selected from the normal production scenario as a reference.
[0111] Operator 2: Randomly select a process from the heaviest loaded workstation and reallocate it to the lightest loaded workstation to balance the loads of each workstation.
[0112] Operator 3: Sort the workstations according to their load from large to small, and try to assign the process to the workstation with the lightest load to minimize the load until the best allocation solution is found.
[0113] Operator 4: Randomly select a process and try to reallocate it to the lightest-loaded workstation until a feasible allocation strategy is obtained.
[0114] Operator 5: Swap the processes at the heaviest load station with those at the randomly selected station to adjust the load distribution.
[0115] Operator 6: Directly swap the processes at the heaviest and lightest workstations to achieve load balance.
[0116] Operator 7: Evaluate the energy consumption of each product at each workstation and reallocate the processes at the workstation with the highest energy consumption to other available workstations.
[0117] Operator 8: Move the processes of other stations to the station with the lowest energy consumption to reduce the overall energy consumption.
[0118] Operator 9: Calculate the total energy consumption of each workstation and swap the robots at the workstations with the highest and lowest energy consumption to optimize the energy consumption distribution.
[0119] Operator 10: Randomly select two robots and swap their positions to explore different production configurations.
[0120] Operator 11: Analyze the changes in the process and randomly assign the processes with large changes to other workstations to adapt to changes in the production environment.
[0121] Operator 12: Based on the process variation analysis, the processes with significant variations are assigned to the workstations closest to the individuals in the normal production environment to maintain the continuity and stability of production.
[0122] Step 6: Adaptive neighborhood search operator selection based on reinforcement learning;
[0123] Neighborhood search operators play a crucial role in efficiently exploring the solution space and preventing evolutionary algorithms from falling into local optimality too early. However, in the algorithm, given that each neighborhood search operator is optimized for a specific goal or scenario, it becomes impractical to manually select a suitable operator. Therefore, an adaptive operator selection method based on reinforcement learning is designed, which can accurately select the appropriate neighborhood search operator for each scenario according to the evolutionary progress of the algorithm and the evolutionary state of each scenario.
[0124] To achieve this goal, the Q learning method in reinforcement learning is used to select neighborhood operators, but before applying this method, two key issues need to be solved:
[0125] 1. Q-table design: How to design the states and behaviors in the Q-table to more effectively guide operator selection is the key to ensuring algorithm performance.
[0126] 2. Design of reward mechanism: We need to reward operators appropriately based on their application results to promote more efficient and accurate operator selection.
[0127] Considering that the robot mixed-flow assembly line in multiple scenarios includes normal production scenarios and multiple unavailable scenarios, a single Q-table may not be able to select suitable neighborhood search operators for all scenarios. Therefore, an innovative strategy is proposed, which is to generate an independent Q-table for each production scenario and use the operators available for each scenario as its actions.
[0128] For the states in the Q-table, an indicator-based method is used to evaluate the effectiveness of the operator. Inverted generational distance (IGD) has been widely used in existing multi-objective evaluation indicators because it can comprehensively evaluate the performance of the non-dominated solutions obtained by the algorithm. Therefore, the inverted generational distance is used to define the state. The smaller the IGD value, the better the performance of the algorithm. The specific state definition is as follows:
[0129] S 1 : IGD of the offspring population under a given scenario o The value is not greater than the IGD of the archive set A value;
[0130] S 2 :IGD o The value is greater than the IGD of the archive set A value, and there are individuals in the offspring population that are not dominated by individuals in the archive set;
[0131] S 3 :IGD o The value is greater than the IGD of the archive set A value, and all individuals in the offspring population are dominated by individuals in the archive set.
[0132] Among them, the archive set is mainly used to save the non-dominated solution sets obtained by various groups during the evolution of the algorithm, and is continuously updated as the algorithm evolves.
[0133] Therefore, the Q-table in this method contains a total of 3 states and 12 actions representing the neighborhood search operator; when the algorithm is initialized, all elements in the Q-table are set to 0, and a specific update formula is used to dynamically adjust the value of each element; this method makes full use of the ability of reinforcement learning, enabling the algorithm to dynamically select the appropriate neighborhood search operator according to the evolutionary process, thereby greatly improving the robustness and effectiveness of the algorithm in solving complex multi-scenario optimization problems.
[0134]
[0135] Among them, Q t+1 (St ,A t ) and Q(S t ,A t ) represent the state S t Next action A t The updated value and current value of the corresponding element, R t+1 In state S t Next, execute action A t The reward value obtained after α and β represent the learning and decay coefficients respectively.
[0136] Table 2Q - Table Example
[0137]
[0138] After applying different evolution operators to update the population, it is necessary to assign corresponding reward values to them according to the comparison results between the sub-population and the archive set individuals. This method sets three reward values for three different states based on the comparison results of the IGD values between the sub-population individuals and the archive set individuals. When in state S 1 When the subpopulation individual is better than the archive set individual, a positive reward value is given; in state S 2 In the state S, even if the subpopulation performs slightly worse, it still helps to promote the Pareto frontier and is given a neutral reward value; 3 In the above example, if the subpopulation performs worse than the individuals in the archive set and fails to advance the Pareto frontier, a negative reward value is assigned; specifically, the reward values of these three states are set to 2, 0, and -2, respectively.
[0139] In addition, during the experiment, it was observed that the Q-table sometimes falls into the local optimum during the evolution process, which limits the effective selection of the neighborhood search operator. To deal with this problem, the parameter P is introduced. Q , used to represent the probability of selecting an operator using reinforcement learning, when P Q When it is greater than the random number, reinforcement learning will be used to select the operator; otherwise, an operator will be randomly selected from all optional operators.
[0140] Based on the above ideas, a reinforcement learning-based adaptive operator selection method is proposed. The specific method flow is shown in Table 3.
[0141] Table 3 Adaptive operator selection based on reinforcement learning
[0142]
[0143] The overall flow of the algorithm is shown in Table 4. In the initialization stage of the algorithm, the system creates a population for each scenario. The number of individuals in each population remains consistent, which is PS. Then, according to the differences in the problems corresponding to each scenario, the system generates the encoding of the individuals in each scenario and performs corresponding decoding operations. Then, the fitness values of the individuals in each population are calculated using step three, and the individuals are sorted by fast non-dominated sorting, and the non-dominated individuals are stored in the archive set.
[0144] During the evolution of the algorithm, the system will first determine whether the termination condition has been reached; if the condition is met, the result will be output directly; otherwise, the algorithm will enter the next cycle; before each cycle starts, the system will randomly shuffle the positions of the evolution operators according to step four and redistribute them to various populations; then, according to step five, the population is updated using the evolution operator and the neighborhood search operator to generate a new sub-population; then, the original population and sub-population in each scenario are merged to form a combined population; individuals with excellent performance are selected from the combined population to form the next generation population, and the non-dominated individuals in it are updated to the archive set.
[0145] Table 4 Overall algorithm flow
[0146] Specific embodiment:
[0148] It explains in detail how to carry out effective management in a production environment with two products, nine processes, four workstations and four robots. The case clearly lists the processing time of each process, the priority relationship between processes and the energy consumption of the robot. The specific data are shown in Table 5. Among them, if the processing time of a process is 0, it means that the product does not need to go through this process. For example, the processing time of the first process of product A is 0, which means that it does not need to go through the first process.
[0149] In addition, it specifically points out that two robots (Robot 2 and 3) require regular maintenance; therefore, this case covers three production scenarios: normal production scenario, Robot 2 maintenance scenario, and Robot 3 maintenance scenario; the settings of these scenarios are of practical guiding significance and provide a comprehensive reference for production management.
[0150] Table 5 Case related information
[0151]
[0152] In normal production scenarios, individuals in the population need to make simultaneous decisions on process allocation and robot allocation. This process involves three stages of coding: Figure 3 An efficient and feasible encoding and decoding process is presented in detail.
[0153] exist Figure 3 In the process sorting vector, each element accurately reflects the weight of the corresponding process. For example, the weight of process 1 is 0.41, while the weight of process 2 is 0.72. In the decoding process, since process 1 has no preceding process, it naturally becomes the first process to be put into the process processing sequence. Subsequently, process 2 and process 3 also enter the processing candidate. Since the weight of process 2 is 0.72, which is higher than that of process 3, process 2 is placed in the second place in the process processing sequence. According to this logic, the complete process processing sequence corresponding to this segment of code can be clearly determined.
[0154] The elements in the process allocation vector directly correspond to the weight of each workstation, and the size of the weight directly determines the number of processes allocated to the workstation; similarly, the elements in the robot allocation vector also reflect the weight of each robot, and robots with larger weights will be given priority in workstation allocation; after calculation, the number of process allocations for each workstation and the robot allocation plan can be clearly obtained.
[0155] Combining the above information, a detailed scheduling plan can be formed, such as Figure 3 As shown in the lower part, in this solution, processes 2 and 3 of product A are accurately assigned to workstation 1 and processed by robot 1; processes 5, 8, and 6 are assigned to workstation 3 and processed by robot 4; therefore, through this series of encoding and decoding processes, accurate process allocation plans and robot allocation plans under normal production scenarios can be obtained.
[0156] In the unavailable scenario, individuals in the population only need to make decisions on process allocation, so their coding structure is relatively simple and only contains two sections of coding. Figure 4 A feasible encoding and decoding process is demonstrated.
[0157] When decoding the individual, we cannot determine the exact location of the unavailable robot, so it is difficult to accurately distinguish which workstations are idle and which are occupied. However, it is known that in any given scenario, only one robot is unavailable. Therefore, in the decoding stage, an approximate strategy is adopted to preliminarily allocate the processes to three workstations, such as Figure 4 As shown in the middle part of , process 2 of product A is assigned to the first station, while processes 3 and 6 are assigned to the second station.
[0158] Once it is determined which workstation is unavailable, the preliminary scheduling plan will be adjusted accordingly to obtain a complete and optimized scheduling plan; assuming that robot 3 in the scenario fails and is assigned to workstation 2, the specific process adjustment will be as follows: Figure 4 The last one is shown in a dotted box.
[0159] After careful calculation and optimization, a feasible scheduling scheme was obtained, and its detailed layout is as follows: Figure 5 shown.
[0160] In a normal production environment, when all robots are available, the process allocation and robot deployment plan corresponding to Situation 1 is used to organize the production process; once Robot 3 is unavailable for some reason, production is reorganized according to the scheduling plan formulated in Situation 2 to maximize the production capacity of maintenance-free robots.
[0161] When robot 3 completes maintenance and resumes working status, it can quickly switch to the corresponding scheduling plan under the normal production environment to maintain the continuity and efficiency of production. It can be seen that the present invention not only further emphasizes the importance of studying the balance problem of multi-scenario robot mixed-flow assembly lines, but also effectively proves the feasibility and effectiveness of the method of the present invention in practical applications.
[0162] Algorithm parameter verification and performance testing:
[0163] 39 cases were designed to test the effectiveness of the algorithm. The basic information in these cases, such as the number of processes, the number of workstations, and the priority relationship between processes, comes from the assembly line benchmark case website: https: / / assembly-line-balancing.de; each case contains two unavailable scenarios, in each scenario, one robot is unavailable; the operation time of different robots in the same process is [t i ×0.8,t i ×1.2] range, t i represents the standard processing time of process i. The robot’s processing energy consumption per unit time and idle energy consumption are also randomly generated. The detailed information is shown in Table 6.
[0164] The 39 cases are designed to verify the actual effectiveness of the algorithm. The basic data of these cases, such as the number of processes, the number of workstations, and the priority relationship between processes, are all from the authoritative assembly line benchmark case website: https: / / assembly-line-balancing.de. Each case cleverly incorporates two unavailable scenarios to simulate robot failures that may occur in actual production environments. In these scenarios, one robot is unavailable; in order to be closer to reality, the operation time of different robots in the same process is within the standard processing time t i is randomly generated in the range of 0.8 to 1.2 times of i represents the standard processing time of process i; in addition, the robot's processing energy consumption per unit time and idle energy consumption are also randomly generated to more fully simulate the actual production environment. The detailed information of these carefully designed cases has been organized in Table 6;
[0165] Table 639 Test case information
[0166]
[0167]
[0168] In Table 6, N represents the number of processes in each case, J is the number of workstations, M refers to the number of products, R represents the number of robots, and R c It specifically refers to the number of unavailable robots in each instance. In order to make a fair and comprehensive evaluation of the algorithm performance, three widely recognized evaluation criteria are adopted: generation distance (GD), inverse generation distance (IGD) and hyper volume ratio (HVR).
[0169] First, the generation distance (GD) is used to evaluate the algorithm performance by calculating the average minimum distance from each point in the non-dominated solution set obtained by the algorithm to the true Pareto frontier; the smaller the GD value, the better the algorithm performance.
[0170] Secondly, the inverse generation distance (IGD) evaluates the convergence and diversity of the algorithm by calculating the distance from the true Pareto front to the non-dominated solution set obtained by each algorithm. A smaller IGD value means that the algorithm performs better in approaching the true Pareto front. It is worth noting that the true Pareto front is constructed based on the non-dominated solutions selected from all algorithm solution sets.
[0171] Finally, the Hyper Volume Ratio (HVR) evaluates algorithm performance by comparing the ratio of the solution space covered by each algorithm to the true Pareto frontier. The larger the HVR value, the better the algorithm performance.
[0172] The RLMOMFEA algorithm and its three variants - MFEA, MFEA_EOA and MFEA_AOS - will be compared and analyzed one by one to evaluate the effectiveness of the proposed improvements. The MFEA algorithm, as the basic algorithm without any improvements, will be used as the comparison benchmark. MFEA_EOA is the optimized version of the MFEA algorithm after incorporating the multi-evolution operator allocation strategy. And MFEA_AOS is the improved algorithm formed by further combining the MFEA algorithm with 12 goal-oriented neighborhood search operators and the operator selection method based on reinforcement learning. To ensure the fairness of the comparison, all algorithms use the same encoding and decoding methods, and only the evolutionary strategies are different. Each algorithm will independently solve 39 cases 10 times and terminate the calculation when the CPU running time exceeds N×N×20 milliseconds.
[0173] Table 7 lists the calculation results of the four algorithms in detail, and bolds the optimal value of each case. It can be seen from the data in the table that the MFEA algorithm only achieved the best GD, IGD and HVR values in the 27th case. In contrast, the MFEA_EOA algorithm achieved the optimal values in 8 cases, especially in the ninth case, where its GD, IGD and HVR values were the best. The MFEA_AOS algorithm showed even better performance, achieving 35 optimal values and the best GD, IGD and HVR values in 8 instances. These results fully demonstrate the effect of the proposed improvement strategy on improving the efficiency of the algorithm. When the two improvement strategies are combined to form the RLMOMFEA algorithm, its performance is further improved. The RLMOMFEA algorithm achieved 71 optimal values. This remarkable achievement shows that the proposed improvements can promote each other, thereby greatly enhancing the solution efficiency of the algorithm.
[0174] In the last row of Table 7, the average GD, IGD, and HVR values of all algorithms are summarized; among them, the average values of MFEA_EOA and MFEA_AOS algorithms are (0.061, 0.051, 0.779) and (0.049, 0.042, 0.841), respectively. Compared with the MFEA algorithm, their performance has been significantly improved, but still slightly inferior to the RLMOMFEA algorithm. This once again verifies the effectiveness and superiority of the proposed improvement strategy.
[0175] Table 7 Comparison results between RLMOMFEA and three variant algorithms
[0176]
[0177]
[0178] Figure 6 The mean performance of RLMOMFEA and its variants in GD, IGD and HVR under a 95% confidence interval is clearly demonstrated; the results show that compared with the basic MFEA algorithm, MFEA_EOA and MFEA_AOS both exhibit lower GD and IGD values, or higher HVR values, which fully verifies the effectiveness of the proposed improvement strategy in solving the RMALBP_MC problem; more significantly, RLMOMFEA always maintains the lowest GD and IGD values, and the highest HVR value among the three, which further emphasizes the significant promoting effect of the fusion improvement strategy on the algorithm performance.
[0179] In order to comprehensively evaluate the effectiveness of multi-factor evolutionary methods in solving multi-scenario robot mixed-flow assembly line balancing problems, RLMOMFEA was deeply compared with six mainstream single-task evolutionary algorithms, including differential evolution algorithm (DE), decomposition-based multi-objective evolutionary algorithm (MOEA / D), multi-objective grey wolf optimization algorithm (MOGWO), multi-objective whale optimization algorithm (WOA), non-dominated sorting genetic algorithm II (NSGA-II) and reference point non-dominated sorting multi-objective optimization algorithm (NSGAIII). In the comparison process, all algorithms adopted the same encoding and decoding methods as RLMOMFEA, but they evolved in independent scenarios respectively, and there was no information exchange between scenarios.
[0180] Ten independent algorithm solutions were performed for 39 cases, and a unified termination condition was set; Table 8 lists the comparison results of the seven algorithms in detail; it can be seen from the table that RLMOMFEA achieved better GD, IGD and HVR values in most cases, fully demonstrating the excellent performance of multi-factor evolution; the only exception is case 9, where the MOGWO algorithm achieved the smallest GD value; in addition, the GD, IGD and HVR values obtained by RLMOMFEA in 8 instances were 0, 0 and 1, indicating that the non-dominated solution of RLMOMFEA has a significant advantage in the comparison and cannot be surpassed by other algorithms.
[0181] These comparative results not only highlight the high efficiency of multi-factor evolution in solving multi-scenario robot mixed-flow assembly line balancing problems, but also verify the powerful performance of the RLMOMFEA algorithm itself.
[0182] Table 8 Comparison results between RLMOMFEA and six single-task evolutionary algorithms
[0183]
[0184]
[0185] Figure 7 The figure clearly shows the mean comparison of GD, IGD and HVR indicators between RLMOMFEA and six single-task evolutionary algorithms under 95% confidence interval. It can be observed from the figure that RLMOMFEA not only always achieves the smallest GD and IGD values, but also achieves the largest HVR value. In addition, compared with the other six algorithms, the confidence interval of RLMOMFEA is significantly narrower, which fully demonstrates its excellent performance in convergence and diversity. These significant experimental results further verify the effectiveness of multi-factor evolutionary algorithms in solving the balancing problem of multi-scenario robot mixed-flow assembly lines, and can achieve better solutions.
[0186] The effectiveness of the RLMOMFEA algorithm is explored in depth through comparative analysis with six multi-task evolutionary algorithms; these comparative algorithms cover cognitive evolutionary multi-task method (MO-MFEA-II), explicit self-encoding evolutionary multi-task (EMT-EA), effective knowledge transfer method for multi-objective multi-task optimization (EMT / et), multi-objective evolutionary algorithm based on dual neighborhood decomposition (MTEA / DD-DN), two-stage adaptive knowledge transfer multi-task optimization algorithm based on population distribution (EMT-PD), and multi-objective multi-task optimization with multiple knowledge types and transfer adaptation (MTDE-MKTA); to ensure fairness, all algorithms independently solved 39 cases 10 times under the same conditions, and adopted the same termination criteria.
[0187] By comparison, Table 9 clearly shows the performance data of the seven algorithms; it can be seen that RLMOMFEA performs well in multiple key indicators; specifically, RLMOMFEA has achieved significant advantages in the three key evaluation indicators of GD, IGD and HVR; it obtained the smallest GD value in 34 cases, proving the high accuracy of its solution quality; in the other 34 cases, RLMOMFEA also achieved the smallest IGD value, which further confirms the close fit between its non-dominated solution set and the true Pareto frontier; in addition, RLMOMFEA obtained the largest HVR value in 35 cases, which fully demonstrated the breadth and diversity of its non-dominated solutions.
[0188] In comparison, the other six multi-task evolutionary algorithms are inferior in performance; they achieved the best results in 0, 0, 2, 7, 1 and 4 instances respectively, which is far from comparable to RLMOMFEA; therefore, through this series of comparative analyses, it can be clearly concluded that the RLMOMFEA algorithm is significantly better than the other six multi-task evolutionary algorithms in performance.
[0189] Table 9. Comparison results of RLMOMFEA and six multi-task evolutionary algorithms
[0190]
[0191]
[0192] Figure 8 The mean performance of each algorithm in the three indicators of GD, IGD and HVR under the 95% confidence interval is shown. This data significantly highlights the advantages of RLMOMFEA compared with other algorithms; whether in GD, IGD, or HVR, RLMOMFEA has shown performance that is superior to all other task evolutionary algorithms, further demonstrating its excellent convergence and diversity.
[0193] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.
Claims
1. A robot mixed-flow assembly line balancing method in multiple scenarios, characterized in that: The following steps are involved: Step 1: Construct a mixed integer linear programming model for the robot's product beat, energy consumption of processed products, and product process changes in a certain scenario, and perform target optimization and constraints on the mixed integer linear programming model; Step 2: Encode and decode normal production scenarios and unavailable scenarios; Step 3: Calculate individual fitness for normal production scenarios and unavailable scenarios; Step 4: Select multiple evolution operators; Step 5: Design goal-oriented neighborhood search operators for different scenarios including production cycle and robot allocation plan; Step 6: Adaptive neighborhood search operator selection based on reinforcement learning.
2. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: The target optimization formula for energy consumption of processed products is: MinObj2=∑ c∈C ∑ m∈M EC mc (2) Among them, R is the robot set, R c is the set of unavailable robots in scenario c, J is the set of workstations, I is the set of processes, C is the set of scenarios, m is the product number, c is the scenario number, i is the process number, j is the workstation number, M is the product set, and t imr is the time required for robot r to process product m in the i-th process, PEC r is the energy consumption per unit time of robot r in processing state, CT mc is the beat of product m in scenario c, x ijc In scenario c, process i is assigned to workstation j, y rj For robot r to be assigned to workstation j, IEC r is the energy consumption per unit time of robot r in idle state.
3. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: Encoding and decoding for normal production scenarios include: Coding: Construct process sorting vector, process allocation vector and robot allocation vector; Decoding: Sort the robot weights in descending order to obtain the robot allocation plan.
4. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: The individual fitness calculation for normal production scenarios includes: According to the process allocation results, the total processing time of each product in each workstation is calculated, and the maximum processing time is selected as the beat. The processing energy consumption is obtained by multiplying the robot's processing energy consumption coefficient per unit time and the processing time; the idle energy consumption is calculated based on the difference between the robot's processing time and the beat, multiplied by the idle energy consumption coefficient per unit time; the total energy consumption is obtained by summing the processing energy consumption and the idle energy consumption.
5. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: The individual fitness calculation for unavailable scenarios includes: The best individual is selected from the normal production scenario and combined with the individuals in the unavailable scenario for comprehensive evaluation; first, the workstation serial numbers assigned to the same process in the two scenarios are identified, and the absolute value of the difference between the two serial numbers is calculated to obtain the process change.
6. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: Step 4 specifically includes: First, the order of each evolutionary operator is randomly shuffled; second, each evolutionary operator is assigned to the corresponding population in turn; the corresponding population is updated using the evolutionary operator to generate the corresponding sub-population.
7. The multi-scenario robot mixed-flow assembly line balancing method according to claim 1 is characterized in that: The goal-oriented neighborhood search operator design for different scenarios of robot allocation planning includes: Evaluate the energy consumption of each product at each workstation and reallocate the processes at the workstation with the highest energy consumption to available workstations; Move the work process to the one with the lowest energy consumption; Calculate the total energy consumption of each workstation and swap the robots at the workstations with the highest and lowest energy consumption.
8. The robot mixed-flow assembly line balancing method in multiple scenarios according to claim 1 is characterized in that: Step 6 specifically includes: Step 61, designing an action including a plurality of states and a plurality of neighborhood search operators; Step 62: Use formula Dynamically update the state and action value; among them, Q t+1 (S t ,A t ) and Q(S t ,A t ) represent the state S t Next action A t The updated value and current value of the corresponding element, R t+1 In state S t Next, execute action A t The reward value obtained after α and β represent the learning and decay coefficients respectively.
9. The robot mixed-flow assembly line balancing system in multiple scenarios is characterized by: include: a memory for storing instructions executable by a processor; A processor, used to execute instructions to implement the robot mixed-flow assembly line balancing method in multiple scenarios as described in any one of claims 1-8.
10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements the robot mixed-flow assembly line balancing method in multiple scenarios as described in any one of claims 1 to 8.
