Line balance optimization algorithm automatic design method driven by large language model

Abstract

The invention discloses an automatic design method for a line balance optimization algorithm driven by a large language model, and belongs to the technical field of mixed flow assembly lines. Establishing a mixed flow assembly line balance optimization mathematical model according to the acquired information and optimization requirements of the mixed flow assembly line; and performing line balance optimization algorithm automatic design on the mixed flow assembly line balance optimization mathematical model through the large language model, including solving the mathematical model through the automatically generated line balance optimization algorithm according to the obtained mixed flow assembly line information to obtain a target function value, and evaluating the performance of the line balance optimization algorithm through the target function value. And performing double evolution according to a feedback evaluation result to obtain a double-evolved line balance optimization algorithm, performing evaluation again, comparing with the line balance optimization algorithm obtained in the previous round, performing iteration continuously until a stop condition is met, and finally obtaining an optimal line balance optimization algorithm. According to the method, automatic design and evaluation can be realized, the optimal solution of the balance optimization problem of the mixed-flow assembly line is obtained, and the design efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of mixed-flow assembly line technology, and in particular relates to an automatic design method for line balancing optimization algorithm driven by a large language model. Background Technology

[0002] Mixed-model assembly lines, as a crucial component of industrial production systems, are widely used in various industrial sectors such as automobile manufacturing and electronics production. They refer to assembly lines that produce multiple different product models on the same line. Such assembly lines can effectively allocate the assembly tasks required for different product types to a series of ordered workstations while meeting task priority constraints and cycle time limitations. The core of mixed-model assembly lines is to achieve balanced optimization of the assembly line while producing different product types, thereby improving production efficiency and reducing costs.

[0003] Mixed-flow assembly line balancing optimization refers to the problem of rationally allocating a series of tasks to workstations while meeting task priority constraints and cycle time limits. The quality of the line balancing optimization scheme (i.e., the task allocation scheme) has a crucial impact on the performance of the entire production line. By effectively optimizing the line balancing scheme, the balancing rate and workstation load balance of the production line can be improved, thereby reducing the number of workstations required, and ultimately achieving the goals of reducing production costs and increasing production line capacity.

[0004] Currently, various optimization algorithms exist for solving line balancing optimization problems, such as exact algorithms, line balancing optimization algorithms, local search, and meta-line balancing optimization algorithms. Among them, the line balancing optimization algorithm (Heuristic Algorithm) is a heuristic optimization algorithm based on predefined rules. It employs approximate solutions or greedy search strategies to quickly find feasible solutions. The line balancing optimization algorithm assigns tasks to workstations according to priority by setting task allocation rules, gradually constructing feasible solutions. Classic heuristic rules include maximizing task time (T), maximizing the ratio of task time to the number of available workstations (TdS), and maximizing the ratio of task time to the latest workstation (TdL). The line balancing optimization algorithm is widely used because it can balance solution speed and solution quality.

[0005] However, existing heuristic rules typically rely on manual design and are highly dependent on expert experience, resulting in inefficient algorithm design and a lack of generalization ability across different problem scenarios. These shortcomings have spurred the emergence of technologies related to automated algorithm design, aiming to reduce reliance on manual intervention and improve production line balancing by automatically generating better heuristic rules. For example, in recent years, hyperheuristic methods such as Genetic Programming (GP) have been proposed to automate the design of line balancing optimization algorithms. GP typically encodes heuristic rules into a tree structure. This method borrows the iterative evolutionary process of Genetic Algorithms (GA), simulating natural selection. Subsequently, new tree structures are generated through processes such as crossover and mutation. Finally, after multiple iterations, a superior line balancing optimization algorithm is produced, achieving the goal of automated line balancing optimization algorithm design. Similar technologies to GP include Gene Expression Programming (GEP), which improves upon GP by refining the encoding method to enhance search efficiency.

[0006] GP and GEP can automate algorithm design to some extent, reducing the workload of manual design and improving design efficiency. However, this tree-based expression method has certain drawbacks: the huge search space leads to slow convergence of the algorithm in the later stages and consumes a lot of computing resources. In addition, it still relies on expert knowledge and experience and lacks reasoning ability.

[0007] In recent years, Large Language Models (LLMs) have demonstrated tremendous potential in the field of automated algorithm design. The Large Language Models for Algorithm Design (LLM4AD) paradigm integrates the iterative optimization methods of Evolutionary Algorithms (EAs) to find optimal solutions with the knowledge reasoning capabilities of LLMs. This paradigm utilizes EAs for iterative search, endowing LLMs with flexible global search capabilities. Simultaneously, it fully leverages the rich domain knowledge and reasoning generalization abilities of LLMs to generate and improve algorithms. By embedding LLMs into the iterative optimization process of EAs, this method achieves a more intelligent and efficient search process, resulting in superior algorithms that balance breadth and depth—i.e., better line-balanced optimization algorithms.

[0008] However, such methods have not yet been studied in the field of assembly line balancing. Therefore, how to automatically design line balancing optimization algorithms for mixed-flow assembly lines has become a major focus. Furthermore, the line balancing optimization algorithms generated by LLM often exhibit strong randomness, making it difficult to converge along the optimal search direction during iteration, resulting in low search efficiency. Moreover, the repetitive nature of LLM-generated line balancing optimization algorithms leads to slow or even non-convergence in later stages. Therefore, how to guide the design of better algorithms for large models, achieving a balance between diversity and convergence, and improving the efficiency of algorithm design is also a key challenge.

[0009] The current design methods for assembly line balancing optimization algorithms have the following drawbacks: 1. Limitations of manually designed heuristic rules: Existing heuristic rule design relies on human experience, which is time-consuming and difficult to adapt to complex and ever-changing production environments.

[0010] 2. Limitations of automatic design of heuristic rules: Although methods such as genetic programming (GP) attempt to achieve automatic generation of heuristic rules, their search space is huge, convergence is difficult in the later stages, and the generalization ability of the generated rules is limited.

[0011] 3. Limitations of applying large language models: When large language models are directly applied to algorithm design, the generated solutions (heuristic algorithms) have strong randomness and repetitiveness, lack clear optimization direction, and result in low search efficiency. Summary of the Invention

[0012] To address the aforementioned technical problems, this invention provides an automatic design method for line balancing optimization algorithms driven by a large language model. The core of this method is the proposed Dual Evolutionary Heuristic Algorithm Automatic Design Method (DEH). This method utilizes the reasoning capabilities of a large language model and the iterative optimization method of evolutionary computation to automatically generate heuristic algorithms (heuristic rules) suitable for mixed-flow assembly line balancing problems. This significantly reduces reliance on manual design, expert knowledge, and design efficiency.

[0013] The objective of this invention is achieved through the following technical solution: This invention provides an automatic design method for line balancing optimization algorithms driven by a large language model, which establishes a mathematical model for balancing optimization of mixed-flow assembly lines based on the acquired information of the mixed-flow assembly line and the line balancing optimization requirements. The large language model is used to automatically design a line balancing optimization algorithm for the mathematical model of the mixed-flow assembly line balancing optimization. This includes solving the mathematical model using an automatically generated line balancing optimization algorithm based on the acquired mixed-flow assembly line information to obtain the objective function value, evaluating the performance of the line balancing optimization algorithm using the objective function value, and performing a double evolution based on the feedback evaluation results to obtain a double-evolved line balancing optimization algorithm. The double-evolved line balancing optimization algorithm is evaluated again and compared with the line balancing optimization algorithm obtained in the previous round. This process is repeated until a stopping condition is met, and finally, the optimal line balancing optimization algorithm is obtained.

[0014] Furthermore, the large language model undergoes automatic design of a line balancing optimization algorithm, with the following steps: During the initialization phase, the large language model automatically generates multiple initial line balancing optimization algorithms based on the acquired information of the mixed-flow assembly line and the line balancing optimization requirements. Algorithm evaluation phase: Solve the established mixed-flow assembly line balancing optimization mathematical model using the generated initial line balancing optimization algorithms respectively, obtain the objective function value of the mixed-flow assembly line balancing optimization mathematical model, and evaluate the performance of the generated initial line balancing optimization algorithms by using the objective function value; Evaluation and feedback phase: The code and language description of line balancing optimization algorithms that do not meet the stopping conditions during the algorithm evaluation phase and subsequent iterations to the final stage are analyzed and evaluated, and the evaluation results, including the reasons for the good or bad performance, evaluation indicators and data, improvement directions and suggestions, are stored in the experience pool. Dual evolution optimization phase: Based on the evaluation and feedback phase, the code population and language description population of the line balancing optimization algorithm are subjected to dual evolution to generate dual-evolved code population and language description population; The line balancing optimization algorithms after dual evolution were obtained: Based on the above analysis results, code population and language description population, several line balancing optimization algorithms after dual evolution were obtained; Environment selection phase: The obtained double-evolution line balancing optimization algorithms are evaluated and compared with multiple line balancing optimization algorithms obtained after the previous round of double evolution. The fewer workstations, the better the algorithm performance. The better line balancing optimization algorithm is retained. The steps from the evaluation feedback phase to the environment selection phase are iterated continuously until the stopping condition is met, resulting in the optimal line balancing optimization algorithm.

[0015] Furthermore, the mathematical model for the balanced optimization of the mixed-flow assembly line is as follows: Objective function: , Constraints: , , , , Equation (1-1) represents the objective function of the mixed-flow assembly line balancing problem, which indicates that the goal of the problem is to minimize the number of workstations; Equation (1-2) provides constraints to ensure that each workstation... The weighted processing time of all tasks at this workstation does not exceed the cycle time. If the workstation If no task is assigned, then =0, the constraint holds; Equation (1-3) guarantees that each task The task is assigned to one and only one workstation; Equation (1-4) constraint ensures that the task is assigned to one and only one workstation. The workstation index is not greater than the task. The workstation index, where Indicates task It is a task Prerequisite tasks ensure the priority of tasks; , For binary decision variables, in the constraints of equation (1-5) Indicates task Assigned to a workstation Indicates whether workstation k is in use. Indicates workstation Enabled, otherwise 0; Model In terms of production demand ratio, meet , Model Assign task Processing time, For the target beat.

[0016] Furthermore, in the initialization phase, the large language model generates multiple initial line balancing optimization algorithm language descriptions that conform to the mixed-flow line balancing problem based on the domain knowledge pre-stored in the experience pool and the information of the acquired mixed-flow assembly line and the line balancing optimization requirements. Based on the generated descriptions of multiple line balancing optimization algorithms and the information in the experience pool, an initial description of the line balancing optimization algorithm with additional experience value information is generated, and multiple initial line balancing optimization algorithms with additional experience value information are obtained based on the description, which are used to solve the mathematical model of the mixed-flow assembly line balancing optimization.

[0017] Furthermore, the specific method of the evaluation feedback phase is as follows: the large language model constructs a thought chain based on the evaluation results of the algorithm evaluation phase, analyzes the performance of the rule language description and code, and learns from the experience of past iterative feedback to guide the optimization direction of the rule language description population and code population generated in the next round. The rule language description and code that generate the line balancing optimization algorithm that meets the constraints are retained, the language descriptions that do not meet the constraints are eliminated, and the analysis results are stored in the experience pool.

[0018] Furthermore, the aforementioned dual evolutionary optimization stage, based on the evaluation feedback stage, evolves the rule-based language description population and code population of the line-balanced optimization algorithm to generate new rule-based language description populations and code populations. The specific steps are as follows: Evolutionary operation: During the first round of evolution and each subsequent generation of the online equilibrium optimization algorithm, crossover and mutation operations are performed on both the language description population and the code population to generate the evolved combination of the online equilibrium optimization algorithm's language description and code. Co-evolution: Pairing mutually matching and constraint-satisfied evolved language descriptions and codes to form a co-evolutionary combination, resulting in a double-evolutionary line-balanced optimization algorithm; at the same time, eliminating rule-based language descriptions and code combinations that do not meet the constraints.

[0019] The beneficial effects of this invention are as follows: 1. This invention utilizes the reasoning capabilities of large language models and the iterative optimization method of evolutionary computation to automatically generate heuristic algorithms (heuristic rules) applicable to mixed-flow assembly line balancing problems, and obtains the optimal solution to the mixed-flow assembly line balancing optimization problem. This largely eliminates the reliance on manual design, reduces dependence on expert knowledge, and improves design efficiency.

[0020] 2. The dual evolution method and evaluation feedback method employed in this invention enhance the generalization ability of the generated rules and improve the algorithm's solution efficiency. Specifically, the dual evolution method achieves co-evolution of the rule language description (Prompt) and the code, reducing the repetitiveness of generated rules. The evaluation feedback method guides the iteration direction through the thought chain constructed by the large language model, reducing randomness and accelerating convergence, thus solving the defects of repetitive and random rules generated when using only the large language model. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the design method of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] Example: This invention discloses an automatic design method for line balancing optimization algorithms driven by a large language model. Based on the acquired information of the mixed-flow assembly line and the line balancing optimization requirements, a mathematical model for the balanced optimization of the mixed-flow assembly line is established. The large language model is then used to automatically design a line balancing optimization algorithm for this mathematical model. This includes solving the mathematical model using an automatically generated line balancing optimization algorithm based on the acquired mixed-flow assembly line information to obtain an objective function value; evaluating the performance of the line balancing optimization algorithm using the objective function value; and performing a double evolution based on the feedback evaluation results to obtain a double-evolved line balancing optimization algorithm. The double-evolved line balancing optimization algorithm is then evaluated again and compared with the line balancing optimization algorithm obtained in the previous round. This process is iterated until a stopping condition is met, ultimately yielding the optimal line balancing optimization algorithm.

[0024] like Figure 1 As shown, the automatic design method of the line balancing optimization algorithm driven by the large language model of the present invention includes the following steps: S1: First, obtain information about the mixed-flow assembly line and the line balancing optimization requirements, and establish a mathematical model for the mixed-flow assembly line balancing optimization, specifically: Objective function: , Constraints: , , , , Equation (1-1) represents the objective function of the mixed-flow assembly line balancing problem, which indicates that the goal of the problem is to minimize the number of workstations; Equation (1-2) provides constraints to ensure that each workstation... The weighted processing time of all tasks at this workstation does not exceed the cycle time. If the workstation If no task is assigned, then =0, the constraint holds; Equation (1-3) guarantees that each task The task is assigned to one and only one workstation; Equation (1-4) constraint ensures that the task is assigned to one and only one workstation. The workstation index is not greater than the task. The workstation index, where Indicates task It is a task The prerequisite tasks ensure the priority of tasks; in the constraints of equation (1-5) Indicates task Assigned to a workstation Indicates workstation Enabled, otherwise 0; The symbols are defined in the table below:

[0025] S2. Automatic Design Phase: S21: Initialization Phase. The Large Language Model (LLM) is based on the domain knowledge (i.e., the description of line balancing optimization algorithms, such as the maximum time first line balancing optimization algorithm) pre-stored in the experience pool. According to the information on mixed-flow line balancing obtained in step S1 and the line balancing optimization requirements, it automatically generates multiple initial line balancing optimization algorithm language descriptions (Prompts) that conform to the mixed-flow line balancing problem. This stage yields a language description that can be adjusted and optimized based on evaluation feedback in subsequent iterations. This helps the LLM understand the basic logic of the line-balanced optimization algorithm to be generated (e.g., describing the basic logic of task allocation, the main steps of the algorithm, and the optimization objective, providing a foundation for the LLM to generate specific code), thus laying the groundwork for generating the line-balanced optimization algorithm in subsequent iterations.

[0026] S22: Line Balancing Optimization Algorithm Generation Stage: Based on the generated initial line balancing optimization algorithm language descriptions and information in the experience pool (which includes algorithm language descriptions, code, the formed line balancing optimization algorithms, evaluation feedback results, reasons for performance quality, etc.), the large language model generates multiple initial line balancing optimization algorithm language descriptions with additional experience value information. These descriptions specifically include task allocation strategies, algorithm flow methods, and detailed algorithm steps. Based on these descriptions, multiple initial line balancing optimization algorithms with additional experience value information are obtained and used to solve the mixed-flow line balancing optimization model.

[0027] S23: Line balancing optimization algorithm evaluation stage: The generated initial line balancing optimization algorithms solve the mixed-flow assembly line balancing optimization mathematical model established in step S1. The Large Language Model (LLM) calculates the objective function value of the mixed-flow assembly line balancing optimization mathematical model, i.e., the number of workstations, through the objective function and constraints of the mixed-flow assembly line balancing optimization mathematical model mentioned at the beginning. The performance of the generated initial line balancing optimization algorithms is verified by the obtained objective function values. The fewer the number of workstations, the better the performance of the algorithm. S24: Evaluation and Feedback Phase: The Large Language Model (LLM) analyzes and evaluates the code and rule language description (Prompt) of line balancing optimization algorithms that did not meet the stopping condition in step S23 (line balancing optimization algorithm evaluation phase) and subsequent iterations to the final phase. The evaluation results, including the reasons for performance differences, evaluation metrics and data, improvement directions, and suggestions, are stored in the experience pool. The specific method is as follows: The large language model constructs a thought chain based on the evaluation results of the line balancing optimization algorithm evaluation phase, analyzes the performance of the rule language descriptions and code, and learns from the experience of past iterative feedback to guide the optimization direction of the rule language description population and code population generated in the next round. It retains the rule language descriptions and code that generate line balancing optimization algorithms that meet the constraints, removes the language descriptions and code that do not meet the constraints, and stores the evaluation results in the experience pool.

[0028] This invention constructs a thought chain, enabling large language models to learn from feedback, avoiding the generation of inefficient algorithms and reinforcing the design principles of efficient algorithms. This feedback method helps large language models gradually converge during the iteration process, generating better line-balanced optimization algorithms and improving design efficiency.

[0029] The aforementioned thought chain is a logical reasoning process inherent in large language models. LLM analyzes and evaluates the results, specifically: analyzing the strengths and weaknesses of the current algorithm, identifying which parts perform well and which are deficient (e.g., an algorithm performs well on high-priority tasks but poorly on low-priority tasks); extracting the reasons for the performance differences (e.g., the task priority calculation formula might be overly sensitive to high-priority tasks while neglecting low-priority tasks); its core lies in breaking down complex problems into multiple intermediate reasoning steps, forming a logical chain, ultimately deriving the answer, and proposing directions for improvement (e.g., adjusting the task priority calculation formula to increase the weight of low-priority tasks). Based on information from the experience pool, the evolutionary direction is adjusted, guiding the next round of evolution towards a smaller objective function value that satisfies the constraints. For example, if the current algorithm performs poorly on certain datasets, LLM can adjust the logic of the generation rules, introduce new strategies, or optimize existing strategies.

[0030] S25: Dual Evolution Optimization Phase: Based on the evaluation and feedback phase, a dual evolution is simultaneously performed on the code population and the rule-based language description population of the line-balanced optimization algorithm, generating dual-evolved code and language description populations. The dual evolution includes: Evolutionary Operations: During each subsequent generation of the online balanced optimization algorithm, crossover and mutation operations are performed simultaneously on the language description population and the code population. Crossover involves exchanging parts of the structure of two individuals to generate new individuals. These parts must be structurally compatible, semantically consistent, and contribute to achieving the optimization objective; otherwise, they will be illogical or grammatically incorrect. Mutation involves randomly altering certain parts of an individual to ensure that the generated new parts are semantically reasonable, free of basic logical errors, meet the constraints, and contribute to achieving the optimization objective. Through these operations, multiple combinations of language descriptions and codes for the online balanced optimization algorithm, having undergone dual evolution, are generated. Co-evolution: Evolved language descriptions and codes that match each other and satisfy constraints (i.e., possessing better logic, adaptability, and other performance characteristics; for example, a language describes how to allocate tasks based on processing time and priority, and the generated code effectively implements this strategy and performs well on multiple test datasets, significantly reducing the number of workstations; in this case, the language description and code combination is considered to meet the conditions) are paired to form a co-evolved combination; this yields a double-evolved line-balanced optimization algorithm. Simultaneously, language descriptions and code combinations that do not meet the constraints (if a language description and code combination performs better on some datasets but clearly fails to meet the constraints on others, or generates too many workstations, which is unrealistic, then this combination is considered unqualified) are eliminated after reflection (i.e., analysis of its shortcomings) through the LLM thought process. This co-evolutionary approach allows language descriptions and code to mutually promote and evolve together.

[0031] The large language model itself has a thought chain reflection function. This thought chain reflection is the process by which LLM evaluates the line-balanced optimization algorithms generated in the automatic algorithm generation stage, as well as optimization algorithms that have re-iterated but failed to meet the stopping conditions, based on the current evolutionary state and existing knowledge and experience in the experience pool. This evaluation is done by calculating the objective function value and verifying the constraints. The evolutionary direction is adjusted according to the evaluation results, guiding the next round of evolution towards a smaller objective function value that meets the constraints, so as to ultimately generate multiple line-balanced optimization algorithms after double evolution. The existing knowledge includes past evaluation results, algorithm descriptions, and code.

[0032] S26: Obtaining the dual-evolutionary line-balanced optimization algorithm: Based on the above analysis results, code population and language description population, the large language model obtains multiple dual-evolutionary line-balanced optimization algorithms; S27: Environment Selection Stage: The algorithm evaluation of the multiple line balancing optimization algorithms obtained by dual evolution will be carried out separately and compared with the multiple line balancing optimization algorithms obtained by dual evolution in the previous round. The fewer the number of workstations, the better the performance of the algorithm. The line balancing optimization algorithm with fewer workstations will be retained. To obtain the optimal solution, multiple better solutions can be retained at the same time to increase the diversity of solutions (avoiding local optima). The process iterates continuously from the evaluation and feedback phase in step S24 to the environment selection phase in step S27 until the stopping condition is met (which can be set such as the objective function value reaching a threshold, the number of iterations reaching an upper limit, or sufficient diversity of solutions), thus obtaining the optimal line-balanced optimization algorithm. If the stopping condition is not met, the process restarts from the evaluation and feedback phase in step S24.

[0033] The parts not described in detail in this application are all existing conventional technologies and will not be elaborated here.

[0034] It is understood that the above specific description of the present invention is only for illustrating the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention to achieve the same technical effect; as long as the use needs are met, they are all within the protection scope of the present invention.

Claims

1. A large language model driven line balancing optimization algorithm automatic design method, characterized by: According to the information and line balancing optimization requirements of the mixed flow assembly line obtained, a mixed flow assembly line balancing optimization mathematical model is established; Through the large language model, the line balancing optimization algorithm automatic design of the mixed flow assembly line balancing optimization mathematical model is carried out, including that the large language model obtains the target function value by solving the mathematical model through the automatically generated line balancing optimization algorithm according to the obtained mixed flow assembly line information, evaluates the performance of the line balancing optimization algorithm through the target function value, and obtains the double-evolved line balancing optimization algorithm according to the feedback evaluation result, re-evaluates the obtained double-evolved line balancing optimization algorithm, compares it with the line balancing optimization algorithm obtained in the last round, and iterates continuously until the stop condition is met, and finally the optimal line balancing optimization algorithm is obtained.

2. The large language model driven line balancing optimization algorithm automatic design method of claim 1, wherein: The large language model carries out line balancing optimization algorithm automatic design, and the steps are as follows: In the initialization stage, the large language model automatically generates a plurality of initial line balancing optimization algorithms according to the information and line balancing optimization requirements of the mixed flow assembly line obtained; Algorithm evaluation stage: solve the mixed flow assembly line balancing optimization mathematical model by using the generated plurality of initial line balancing optimization algorithms, obtain the target function value of the mixed flow assembly line balancing optimization mathematical model, and evaluate the performance of the generated initial line balancing optimization algorithm through the target function value; Evaluation feedback stage: analyze and evaluate the code and language description of the line balancing optimization algorithm that does not meet the stop condition in the algorithm evaluation stage and the iteration to the final stage, and store the evaluation results including the reasons for the performance, evaluation indexes and data, improvement direction and suggestions in the experience pool; Double evolution optimization stage: on the basis of the evaluation feedback stage, the code population and language description population of the line balancing optimization algorithm are respectively double-evolved to generate double-evolved code population and language description population; Get the double-evolved line balancing optimization algorithm: according to the above analysis results, code population and language description population, a plurality of double-evolved line balancing optimization algorithms are obtained; Environment selection stage: the obtained double-evolved line balancing optimization algorithm is evaluated, and compared with the plurality of line balancing optimization algorithms obtained after the last round of double evolution, the fewer the number of stations, the better the performance of the algorithm, and the better line balancing optimization algorithm is reserved; The steps from the evaluation feedback stage to the environment selection stage are iterated continuously until the stop condition is met, and the optimal line balancing optimization algorithm is obtained.

3. The large language model driven line balancing optimization algorithm automatic design method according to claim 1 or 2, characterized in that: The mixed flow assembly line balancing optimization mathematical model is: target function: ， Constraint condition: ， ， ， ， Equation (1-1) represents the objective function of the mixed-flow assembly line balancing problem, which indicates that the goal of the problem is to minimize the number of workstations; Equation (1-2) provides constraints to ensure that each workstation... The weighted processing time of all tasks at this workstation does not exceed the cycle time. If the workstation If no task is assigned, then =0, the constraint holds; Equation (1-3) guarantees that each task The task is assigned to one and only one workstation; Equation (1-4) constraint ensures that the task is assigned to one and only one workstation. The workstation index is not greater than the task. The workstation index, where Indicates task It is a task Prerequisite tasks ensure the priority of tasks; , For binary decision variables, in the constraints of equation (1-5) Indicates task Assigned to a workstation Indicates whether workstation k is in use. Indicates workstation Enabled, otherwise 0; Model In terms of production demand ratio, meet , Model Assign task Processing time, For the target beat.

4. The large language model driven line balancing optimization algorithm automatic design method of claim 2, wherein: In the initialization stage, the large language model is based on the domain knowledge stored in the experience pool in advance, and generates a plurality of initial line balancing optimization algorithm language descriptions conforming to the mixed flow line balancing problem according to the obtained mixed flow assembly line information and line balancing optimization requirements. According to the generated multiple line balance optimization algorithm description and the information in the experience pool, an initial line balance optimization algorithm description of additional experience value information is generated, and an initial multiple line balance optimization algorithm of additional experience value information is obtained according to the description, which is used for solving the mixed flow assembly line balance optimization mathematical model.

5. The large language model driven line balancing optimization algorithm automatic design method of claim 1, wherein: The specific method of the evaluation feedback stage is that the large language model constructs a thinking chain according to the evaluation result of the algorithm evaluation stage, analyzes the performance of the rule language description and the code, and learns from the past iterative feedback experience to guide the optimization direction of the rule language description population and the code population generated in the next round, retains the rule language description and the code of the line balance optimization algorithm that meets the constraint condition, eliminates the language description that does not meet the constraint condition, and stores the analysis result in the experience pool.

6. The large language model driven line balancing optimization algorithm automatic design method of claim 1, wherein: The double evolution optimization stage is to evolve the rule language description population and the code population of the line balance optimization algorithm on the basis of the evaluation feedback stage to generate a new rule language description population and a code population, and the specific steps are as follows: Evolution operation: in the first round to each generation evolution process of the line balance optimization algorithm, the language description population and the code population are simultaneously subjected to cross and mutation operations to generate the evolved line balance optimization algorithm language description and code combination; Co-evolution: the evolved language description and the code that match each other and meet the constraint condition are paired to form a co-evolved combination to obtain the double-evolved line balance optimization algorithm; at the same time, the rule language description and the code combination that do not meet the constraint condition are eliminated.

Images