An industrial large model-based generative design and production scheduling optimization method and system

Abstract

A method for generative design and production scheduling optimization based on an industrial large model, comprising the following steps: step S1: obtaining a user's request and inputting the user's request into an industrial large model; step S2: determining whether the user's request is an optimization problem, if not, generating a corresponding reply through the industrial large model, if it is an optimization problem, constructing a target function and a constraint condition based on the optimization problem; step S3: based on the target function and the constraint condition, the global search of the solution of the optimization problem is carried out by using the Lévy flight iteration method, and the optimal solution is output. The system finally overcomes the core pain points in the prior art, such as dependence on artificial experience, limited optimization capability, fragmentation of each link, and difficulty in reusing knowledge, and thus produces revolutionary progress from experience-driven to data and model-driven, providing an unprecedented efficient, intelligent and reliable solution for industrial design and production operation.

Description

Technical Field

[0001] This invention relates to the field of production planning optimization technology, and in particular to a generative design and scheduling optimization method and system based on a large industrial model. Background Technology

[0002] As global manufacturing evolves towards intelligence and flexibility, enterprises face increasing production complexity and uncertainty. Production line layout design and production scheduling, as two core planning components of a manufacturing system, fully reflect the level of intelligence and flexibility in manufacturing and directly determine production efficiency, resource utilization, and order fulfillment capabilities. Currently, most enterprises adopt the following solutions for production line layout design and production scheduling: 1. Design and Scheduling Based on Human Experience: This method relies on the experience of senior engineers or planners for production line layout planning and production scheduling. This approach is highly dependent on individual skills. Even with professional training, senior engineers and technicians still face significant subjectivity in the decision-making process and struggle to handle complex optimization problems with multiple objectives and constraints. Severe information silos exist between different departments (such as process engineering and production planning), preventing a comprehensive consideration of the entire production process. This leads to a disconnect between layout design and actual scheduling, making it difficult to achieve global optimization.

[0003] 2. Based on traditional optimization algorithms and simulation software: This approach employs traditional operations research optimization algorithms such as linear programming, genetic algorithms, and simulated annealing, or utilizes discrete event simulation software to verify specific solutions. While this method improves scientific rigor to some extent, its drawbacks are: 1) Lack of "generation" capability: The algorithm itself cannot creatively generate new solutions and relies heavily on the input of the initial solution. In essence, it is a kind of "selection" and "optimization" within a given range.

[0004] 2) The “combinatorial explosion” problem: For large-scale, multi-resource production line layout and scheduling problems, the solution space is extremely large. Traditional algorithms are prone to getting stuck in local optima and the calculation is time-consuming, which cannot meet the needs of rapidly responding to changes in market demand.

[0005] 3. Methods based on early AI models: In recent years, some studies have attempted to apply machine learning models for predictive maintenance or local optimization. However, these models are often "narrow AI," with limited functionality, only able to solve specific point problems (such as predicting equipment failures or optimizing single parameters), lacking the ability to understand non-standard constraints described in natural language, and even less possessing advanced cognitive abilities for generative design and multi-stage collaborative planning.

[0006] Therefore, there is an urgent need in this field for a new generation of intelligent systems that can deeply integrate industrial knowledge, possess creative generation capabilities, and achieve global optimization of layout and production scheduling. Summary of the Invention

[0007] To address the aforementioned shortcomings, the present invention aims to propose a generative design and scheduling optimization method and system based on a large industrial model.

[0008] To achieve this objective, the present invention adopts the following technical solution: a generative design and scheduling optimization method based on a large industrial model, comprising the following steps: Step S1: Obtain the user's request and input the user's request into the industrial big data model; Step S2: Determine whether the user request is an optimization problem. If it is not an optimization problem, generate the corresponding response using the industrial big data model. If it is an optimization problem, construct the objective function and constraints based on the optimization problem. Step S3: Based on the objective function and constraints, perform a global search for the solution to the optimization problem using the Lévy flight iteration method, and output the optimal solution.

[0009] The preferred training process for the large industrial model is as follows: Obtain the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents, and use the engineering design dataset as the structural basis of the knowledge base; The knowledge base construction data is preprocessed to obtain the first data, and the first data is converted into a data vector using a text embedding model. Deploy the Qdrant vector database and store the data vectors in the Qdrant vector database as a knowledge base for the industrial large model; The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

[0010] Preferably, the following rules should also be followed when training large industrial models: The original weights of the input parameters during the training of the large industrial model are frozen and do not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated. The formula for obtaining the final weight is as follows: ;in For the original weights that are frozen, is the scaling factor, r is the rank of the low-rank decomposition, matrix A is a d×r real matrix, the initialization of matrix A is obtained by random initialization using a Gaussian distribution, matrix B is an r×k real matrix, the initial value of matrix B is 0, k=d, and r is less than k and d.

[0011] Preferably, the step S3, which involves a global search for the solution to the optimization problem using the Lévy flight iteration method, is as follows: Step S31: Set the preliminary solution given by the industrial large model as the parameters to be optimized. , where d is the dimension of the vector; Step S32: Determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; Step S33: Randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); Step S34: Perform Lévy flight iteration on the initial candidate solutions to generate the optimal solution.

[0012] Preferably, the step of generating the optimal solution through Lévy flight iteration in step S34 is as follows: Step S341: Generate a random step size vector s that follows a Lévy distribution; Where u and v follow a normal distribution, , , , ; Step S342: Select the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: ; Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; Step S343: Obtain the updated objective function value at the current position as the first function value, obtain the objective function value of the optimal solution as the second function value, and determine whether the first function value is better than the second function value. If it is better, update the search result after this position update to the optimal solution; if it is not better, keep the optimal solution. Step S344: Determine whether the current iteration has converged. If it has converged, output the optimal solution as the optimal solution. If it has not converged, repeat steps S342 to S343. The convergence criterion is that any of the following convergence conditions are met: The number of iterations equals the preset number; In K consecutive iterations, the difference between the second function value and the first function value is less than the difference threshold. ; in , Where K is 5% to 20% of the preset number of times.

[0013] A generative design and scheduling optimization system based on a large industrial model, using the aforementioned generative design and scheduling optimization method based on a large industrial model, includes: an input module, a processing module, and a solution module; The input module is used to obtain user requests and input them into the industrial big data model; The processing module is used to determine whether the user request is an optimization problem. If it is not an optimization problem, it generates a corresponding response through the industrial big data model. If it is an optimization problem, it constructs an objective function and constraints based on the optimization problem. The solution module is used to perform a global search for the solution to the optimization problem based on the objective function and constraints, and output the optimal solution by using the Lévy flight iteration method.

[0014] In some embodiments, a training module is also included; The training module is used to acquire the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents. The engineering design dataset is used as the structural basis of the knowledge base. The knowledge base construction data is preprocessed to obtain the first data, and the first data is converted into a data vector using a text embedding model. Deploy the Qdrant vector database and store the data vectors in the Qdrant vector database as a knowledge base for the industrial large model; The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

[0015] Preferably, the training module includes an update unit; The update unit is used to freeze the original weights of the input parameters during the training of the large industrial model and does not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated.

[0016] In some embodiments, the solving module includes a setting unit, a solution space acquisition unit, an initial solution acquisition unit, and an optimal solution acquisition unit; The setting unit is used to set the preliminary solution given by the industrial large model as the parameters to be optimized; The solution space acquisition unit is used to determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; The initial solution acquisition unit is used to randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); The optimal solution acquisition unit is used to generate the optimal solution by performing Lévy flight iteration on the initial candidate solutions.

[0017] In some embodiments, the optimal solution acquisition unit includes a first subunit, a second subunit, a third subunit, and a fourth subunit; The first sub-unit is used to generate a random step-size vector s that follows a Lévy distribution; The second sub-unit selects the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: ; Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; The third subunit is used to obtain the objective function value after the current position is updated, as the first function value, obtain the objective function value of the optimal solution, as the second function value, and determine whether the first function value is better than the second function value. If it is better, the search result after the current position update is updated to the optimal solution; if it is not better, the optimal solution is kept. The third sub-unit is used to determine whether the current iteration has converged. If it has converged, the optimal solution is output as the optimal solution. If it has not converged, the second and third sub-units are called again.

[0018] One of the above technical solutions has the following advantages or beneficial effects: This patent does not simply introduce industrial large-scale model technology, but rather organically combines the cognitive ability of industrial large-scale models, the search capability of optimization algorithms, and the engineering requirements of industrial systems through a complete and systematic innovation. It is precisely because of the adoption of the above-mentioned synergistic technical means that this system ultimately overcomes the core pain points of existing technologies, such as reliance on human experience, limited optimization capabilities, fragmented processes, and difficulty in knowledge reuse. This results in revolutionary progress from automation to intelligence, from local optimization to global collaboration, and from experience-driven to data and model-driven approaches, providing unprecedentedly efficient, intelligent, and reliable solutions for industrial design and production operations. Attached Figure Description

[0019] Figure 1 This is a flowchart of one embodiment of the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of one embodiment of the system of the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] like Figures 1-2 As shown, a generative design and scheduling optimization method based on a large industrial model includes the following steps: Step S1: Obtain the user's request and input the user's request into the industrial big data model; Step S2: Determine whether the user request is an optimization problem. If it is not an optimization problem, generate the corresponding response using the industrial big data model. If it is an optimization problem, construct the objective function and constraints based on the optimization problem. Step S3: Based on the objective function and constraints, perform a global search for the solution to the optimization problem using the Lévy flight iteration method, and output the optimal solution.

[0025] This invention systematically solves the problems in existing industrial optimization solutions through the deep integration and synergy of multiple technologies. Its key technical points are as follows: 1. Because it adopts the technique of fine-tuning to adapt the open-source industrial big model to the domain, it overcomes the technical problems of lack of professional knowledge, unreliable output, and high cost of full parameter fine-tuning in the industrial vertical domain of the general industrial big model. As a result, it has the advantage of transforming general AI capabilities into "industrial experts" who are proficient in production line layout and production scheduling at low cost and high efficiency, and realizes the accurate injection and accumulation of domain knowledge.

[0026] 2. Because it adopts a collaborative workflow that deeply integrates the generative capabilities of large languages ​​with metaheuristic optimization algorithms such as Lévy flight (i.e., the industrial large model is responsible for the initial conception and problem definition, and the optimization algorithm is responsible for the precise solution), it overcomes the technical problems of traditional optimization algorithms lacking creativity and relying on initial solutions, as well as the insufficient accuracy of industrial large models in complex mathematical optimization and the existence of "illusions". Thus, it has the advantages of combining creative divergent thinking and precise mathematical solution capabilities, and can automatically generate practical and feasible global optimal or near-optimal solutions.

[0027] 3. Because it adopts the random walk strategy unique to the Lévy fly algorithm (mainly short-distance fine search with occasional long-distance jumps), it overcomes the technical problems of traditional optimization algorithms (such as genetic algorithms and simulated annealing) that are prone to getting stuck in local optima and premature convergence when solving high-dimensional, multi-peak industrial layout and production scheduling problems. As a result, it has stronger global exploration capabilities and a mechanism to escape local optima, significantly improving the probability and quality of finding better solutions in a huge solution space.

[0028] 4. Because it adopts a system architecture centered on an intelligent agent and using Docker Compose to uniformly orchestrate industrial large-scale model services, optimization algorithm services, and various functional modules, it overcomes the technical problems of traditional industrial software such as functional environment dependency conflicts, module isolation, data fragmentation, difficulty in collaborative optimization, and complex deployment. As a result, it has the advantages of achieving closed-loop linkage optimization of production line layout and production scheduling, elastic resource scaling, and one-click standardized deployment, opening up the digital link from design to execution, and improving the system's integrity and ease of use.

[0029] 5. Because it adopts the technical means of constructing an industrial domain vector knowledge base and equipping the large industrial model with an external "memory system", it overcomes the technical problems of limited static knowledge in the model parameters and the inability to effectively utilize massive unstructured industrial data (such as CAD drawings and process documents) to assist decision-making in real time. As a result, it has the advantages of having evidence-based and traceable answers and decisions, which greatly improves the accuracy and reliability of the generated solutions and avoids relying on imagination.

[0030] This patent does not simply introduce industrial large-scale modeling technology, but rather organically combines the cognitive capabilities of industrial large-scale models, the search capabilities of optimization algorithms, and the engineering requirements of industrial systems through a complete and systematic innovation. It is precisely because of this synergistic approach that this system ultimately overcomes the core pain points of existing technologies, such as reliance on human experience, limited optimization capabilities, fragmented processes, and difficulty in knowledge reuse. This results in revolutionary progress from automation to intelligence, from local optimization to global collaboration, and from experience-driven to data- and model-driven approaches, providing unprecedentedly efficient, intelligent, and reliable solutions for industrial design and production operations.

[0031] The preferred training process for the large industrial model is as follows: Obtain the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents, and use the engineering design dataset as the structural basis of the knowledge base; First and foremost, it must be emphasized that data is an indispensable and absolutely crucial element in the process of fine-tuning industrial large-scale models. The quantity and quality of data will affect the final result. In order to train a high-quality model and make accurate responses, it is necessary to ensure that the data is standardized and structurally rigorous.

[0032] The specific data requirements are as follows:

[0033] Data source: The data mainly comes from the company's PDM, MES, ERP and other systems. Some document data should be archived on the company's file server.

[0034] Data quality control: After data collection is completed, a dedicated team or personnel should systematically evaluate the usability, authenticity, and quality of the data, removing any data that does not meet the requirements to ensure the high quality of the final data.

[0035] The knowledge base construction data is preprocessed to obtain the first data, which includes data cleaning and annotation: standardizing the data format, handling missing values, and allowing experienced engineers and experts to annotate key data in the knowledge base construction data. A text embedding model is then used to convert the first data into data vectors. Deploy the Qdrant vector database and store the data vectors within it as a knowledge base for the large industrial model. Qdrant is perfectly compatible with Python, boasts excellent memory optimization, and is easy to deploy. Data converted into vector form at the embedding layer is stored in the vector database, serving as the knowledge base for the large model—essentially, its own "library."

[0036] The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

[0037] The meaning of the "Command-Output" data group is as follows: Instruction (problem): Describes a specific task.

[0038] Output (answer): Show how this task should be completed.

[0039] This process is equivalent to transforming data into a series of questions and providing standardized answers to each question.

[0040] Data source for "Command-Output": Extract instructions and outputs from the data information collected in the knowledge base.

[0041] If the data is fragmented and lacks distinct features, a workshop should be convened with domain experts or experienced engineers from within the company. The workshop will pose questions (instructions) for different scenarios that may be encountered, and the experts and engineers will provide different solutions (outputs) based on the different questions. The data will then be organized into a "instruction-output" data set.

[0042] Ensuring the quality of training data is crucial: Ensure that the "instruction-output" data sets cover most scenarios, such as normal situations, emergency orders, equipment failures, and new product trials. Create a unified template for the instruction-output data sets, for example: [Command Template] Please develop a plan for the following situations: Constraints: {Site / Equipment / Orders / Delivery Date}; Optimization objectives: {cost / time / efficiency}; Special requirements: {Other points to note}; Output Template Solution Overview: {Brief Description}; Specific arrangements: {detailed description}; Expected results: {quantitative indicators}; Important Notes: {Special Reminder} After all the data has been organized, the following should be inserted at the beginning of the output section: "action": call_optimization_tool, "tool_name": enter the tool name.

[0043] This is to inject tool-calling capabilities during the model fine-tuning phase, subsequently calling the Lévy Flight Optimization algorithm for solution. If the amount of data in the knowledge base construction information is sufficient, 100-200 high-quality samples should be prepared initially, with the number increased as needed.

[0044] In some embodiments, the following rules also need to be followed when training large industrial models: The original weights of the input parameters during the training of the large industrial model are frozen and do not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated. The formula for obtaining the final weight is as follows: ;in For the original weights that are frozen, is the scaling factor, r is the rank of the low-rank decomposition, matrix A is a d×r real matrix, the initialization of matrix A is obtained by random initialization using a Gaussian distribution, matrix B is an r×k real matrix, the initial value of matrix B is 0, k=d, and r is less than k and d.

[0045] This invention employs efficient fine-tuning techniques to fine-tune large industrial models, particularly suitable for adaptive tuning of large pre-trained language models. The core idea is to add a subset of parameters to the frozen base model and update only this subset during tuning, without modifying the original parameters of the base model. This method significantly reduces the number of trainable parameters for downstream tasks, substantially improving training speed and memory efficiency.

[0046] The specific steps are as follows: Write a training script (Python), use vLLM technology to accelerate the process and then use it to validate the model after fine-tuning, load a large industrial model, and set the parameters. The script loads the prepared instruction-output training data set. A Dockerfile is written to declare how to build the image, a requirements document is written to declare the necessary dependencies, and the vLLM library is added. A `docker-compose.training.yml` file is written in the project folder to manage all service dependencies, ports, volumes, etc. Finally, the training service is started using the `docker-compose` command, entering the container to run the training script. If the data volume is large, the training time will be very long, possibly several days. After training, the industrial large-scale model will be merged and saved to the specified location according to the settings in the training script.

[0047] Preferably, the step S3, which involves a global search for the solution to the optimization problem using the Lévy flight iteration method, is as follows: Step S31: Set the preliminary solution given by the industrial large model as the parameters to be optimized. , where d is the dimension of the vector; Step S32: Determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; Step S33: Randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); Step S34: Perform Lévy flight iteration on the initial candidate solutions to generate the optimal solution.

[0048] Preferably, the step of generating the optimal solution through Lévy flight iteration in step S34 is as follows: Step S341: Generate a random step size vector s that follows a Lévy distribution; Where u and v follow a normal distribution, , , , ; Step S342: Select the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: The probability p can be set to less than 10%, and can be adjusted by the large model during iteration, allowing the solution to jump randomly over a large range, leaving the current small region. Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; Step S343: Obtain the updated objective function value at the current position as the first function value, obtain the objective function value of the optimal solution as the second function value, and determine whether the first function value is better than the second function value. If it is better, update the search result after this position update to the optimal solution; if it is not better, keep the optimal solution. Step S344: Determine whether the current iteration has converged. If it has converged, output the optimal solution as the optimal solution. If it has not converged, repeat steps S342 to S343. The convergence criterion is that any of the following convergence conditions are met: The number of iterations equals the preset number; In K consecutive iterations, the difference between the second function value and the first function value is less than the difference threshold. ; in , Where K is 5% to 20% of the preset number of times.

[0049] The industrial large model provides an initial solution to the problem. Then, the Lévy Flight Optimization Algorithm is called to perform small-scale local searches in the solution space for most of the time, and occasionally to conduct long-distance global explorations to continuously obtain solutions. These solutions are substituted into the objective function, and the constraints are handled with a penalty function. The process is iterated continuously to finally find the optimal solution.

[0050] The advantages and innovations of the Lévy flight optimization algorithm in this patent are as follows: The solution space is provided by an industrial large-scale model, which constructs the solution space and adjusts the objective function in real time according to user instructions, rather than the traditional manual, precise, static definition. This ensures that the final solution better meets the user's goals and is more suitable for practical application scenarios. When the industrial large-scale model determines that the current optimization has reached a deadlock and multiple iterations have not improved the situation, it can proactively adjust the probability p, for example, by temporarily increasing p to increase the probability of escaping the current small region, thus better avoiding getting trapped in local optima.

[0051] Implementation: The following operations are performed in the inference environment. First, an Agent service needs to be written to determine whether the current problem requires invoking the Lévy flight optimization algorithm. If it is determined to be a production line layout optimization problem or a production scheduling optimization problem, then the algorithm is invoked; otherwise, the industrial big data model directly answers the question. Next, the Lévy_optimizer flight optimization algorithm service is written, and the algorithm formula is substituted. Dockerfiles for both the Agent and Lévy_optimizer are written to create images, and requirement documents for both services are written to declare the required dependencies.

[0052] Create a Docker-Compose.Inference.yml file in the project folder to centrally manage the dependencies, ports, and volumes of services such as Lévy_optimizer, Agent, LLM, and vLLM. Finally, all services can be run directly using Docker-Compose commands.

[0053] Compared to existing Lévy flight optimization algorithms, this invention explicitly separates two search modes: The global search employs a "complete space reset" strategy, allowing solutions to be randomly regenerated within the search space. This provides the algorithm with the ability to completely escape local optima, far superior to traditional Lévy flight algorithms that rely solely on occasional long-step escape mechanisms. Secondly, the local search innovatively introduces "vectorization guidance," where the update step size is multiplied by the vector difference between the current solution and the historical best solution. This allows the search process to intelligently apply adjustments of varying intensities across different dimensions, with greater corrections obtained in dimensions far from the optimum, thus accelerating the refined development of local regions. Finally, a dual judgment method combining "dynamic relative threshold" and "continuous trend observation" is employed. The threshold is not a fixed value but a proportion based on the objective function value of the historical best solution. This allows the convergence accuracy requirement to automatically adjust with the optimization process. Larger fluctuations are allowed in the early stages for thorough exploration, while the standard is increased in the later stages for precise convergence. Simultaneously, convergence is only determined after multiple consecutive occurrences of improvement below the threshold, effectively avoiding misjudgments caused by single random fluctuations and enhancing the algorithm's robustness.

[0054] A generative design and scheduling optimization system based on a large industrial model, using the aforementioned generative design and scheduling optimization method based on a large industrial model, includes: an input module, a processing module, and a solution module; The input module is used to obtain user requests and input them into the industrial big data model; The processing module is used to determine whether the user request is an optimization problem. If it is not an optimization problem, it generates a corresponding response through the industrial big data model. If it is an optimization problem, it constructs an objective function and constraints based on the optimization problem. The solution module is used to perform a global search for the solution to the optimization problem based on the objective function and constraints, and output the optimal solution by using the Lévy flight iteration method.

[0055] In some embodiments, a training module is also included; The training module is used to acquire the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents. The engineering design dataset is used as the structural basis of the knowledge base. The knowledge base construction data is preprocessed to obtain the first data, and the first data is converted into a data vector using a text embedding model. Deploy the Qdrant vector database and store the data vectors in the Qdrant vector database as a knowledge base for the industrial large model; The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

[0056] Preferably, the training module includes an update unit; The update unit is used to freeze the original weights of the input parameters during the training of the large industrial model and does not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated.

[0057] Preferably, the solution module includes a setting unit, a solution space acquisition unit, an initial solution acquisition unit, and an optimal solution acquisition unit; The setting unit is used to set the preliminary solution given by the industrial large model as the parameters to be optimized; The solution space acquisition unit is used to determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; The initial solution acquisition unit is used to randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); The optimal solution acquisition unit is used to generate the optimal solution by performing Lévy flight iteration on the initial candidate solutions.

[0058] In some embodiments, the optimal solution acquisition unit includes a first subunit, a second subunit, a third subunit, and a fourth subunit; The first sub-unit is used to generate a random step-size vector s that follows a Lévy distribution; The second sub-unit selects the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: ; Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; The third subunit is used to obtain the objective function value after the current position is updated, as the first function value, obtain the objective function value of the optimal solution, as the second function value, and determine whether the first function value is better than the second function value. If it is better, the search result after the current position update is updated to the optimal solution; if it is not better, the optimal solution is kept. The third sub-unit is used to determine whether the current iteration has converged. If it has converged, the optimal solution is output as the optimal solution. If it has not converged, the second and third sub-units are called again.

[0059] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0060] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A generative design and scheduling optimization method based on a large industrial model, characterized in that, Includes the following steps: Step S1: Obtain the user's request and input the user's request into the industrial big data model; Step S2: Determine whether the user request is an optimization problem. If it is not an optimization problem, generate the corresponding response using the industrial big data model. If it is an optimization problem, construct the objective function and constraints based on the optimization problem. Step S3: Based on the objective function and constraints, perform a global search for the solution to the optimization problem using the Lévy flight iteration method, and output the optimal solution.

2. The generative design and scheduling optimization method based on a large industrial model according to claim 1, characterized in that, The training process for the large industrial model is as follows: Obtain the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents, and use the engineering design dataset as the structural basis of the knowledge base; The knowledge base construction data is preprocessed to obtain the first data, and the first data is converted into a data vector using a text embedding model. Deploy the Qdrant vector database and store the data vectors in the Qdrant vector database as a knowledge base for the industrial large model; The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

3. The generative design and scheduling optimization method based on a large industrial model according to claim 2, characterized in that, The following rules also need to be followed when training large industrial models: The original weights of the input parameters during the training of the large industrial model are frozen and do not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated. The formula for obtaining the final weight is as follows: ;in For the frozen original weights, is the scaling factor, r is the rank of the low-rank decomposition, matrix A is a d×r real matrix, the initialization of matrix A is obtained by random initialization using a Gaussian distribution, matrix B is an r×k real matrix, the initial value of matrix B is 0, k=d, and r is less than k and d.

4. The generative design and scheduling optimization method based on a large industrial model according to claim 1, characterized in that, The steps in step S3, which involve a global search for the solution to the optimization problem using the Lévy flight iteration method, are as follows: Step S31: Set the preliminary solution given by the industrial large model as the parameters to be optimized. , where d is the dimension of the vector; Step S32: Determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; Step S33: Randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); Step S34: Perform Lévy flight iteration on the initial candidate solutions to generate the optimal solution.

5. The generative design and scheduling optimization method based on a large industrial model according to claim 4, characterized in that, The steps in step S34 for generating the optimal solution through Lévy flight iteration are as follows: Step S341: Generate a random step size vector s that follows a Lévy distribution; Where u and v follow a normal distribution, , , ， ； Step S342: Select the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: ； Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; Step S343: Obtain the updated objective function value at the current position as the first function value, obtain the objective function value of the optimal solution as the second function value, and determine whether the first function value is better than the second function value. If it is better, update the search result after this position update to the optimal solution; otherwise, keep the optimal solution. Step S344: Determine whether the current iteration has converged. If it has converged, output the optimal solution as the optimal solution. If it has not converged, repeat steps S342 to S343. The convergence criterion is that any of the following convergence conditions are met: The number of iterations equals the preset number; In K consecutive iterations, the difference between the second function value and the first function value is less than the difference threshold. ; in , Where K is 5% to 20% of the preset number of times.

6. A generative design and scheduling optimization system based on a large industrial model, characterized in that, The generative design and scheduling optimization method based on an industrial large model as described in any one of claims 1 to 5 includes: an input module, a processing module, and a solution module; The input module is used to obtain user requests and input them into the industrial big data model; The processing module is used to determine whether the user request is an optimization problem. If it is not an optimization problem, it generates a corresponding response through the industrial big data model. If it is an optimization problem, it constructs an objective function and constraints based on the optimization problem. The solution module is used to perform a global search for the solution to the optimization problem based on the objective function and constraints, and output the optimal solution by using the Lévy flight iteration method.

7. The generative design and scheduling optimization system based on a large industrial model according to claim 6, characterized in that, It also includes a training module; The training module is used to acquire the enterprise's production industrial dataset, which includes production line layout data, equipment parameter data, production process data, production order data, implementation status data, and knowledge documents. The engineering design dataset is used as the structural basis of the knowledge base. The knowledge base construction data is preprocessed to obtain the first data, and the first data is converted into a data vector using a text embedding model. Deploy the Qdrant vector database and store the data vectors in the Qdrant vector database as a knowledge base for the industrial large model; The pre-set knowledge base is used to construct a "command-output" data set, which is then used to train the large industrial model.

8. The generative design and scheduling optimization system based on a large industrial model according to claim 7, characterized in that, The training module includes an update unit; The update unit is used to freeze the original weights of the input parameters during the training of the large industrial model and does not participate in gradient updates. Backpropagation is performed on the input "instruction-output" data set to update matrices A and B. Based on the frozen original weights and the updated matrices A and B, the final weights of the current input parameters of the industrial large model are generated.

9. The generative design and scheduling optimization system based on a large industrial model according to claim 6, characterized in that, The solution module includes a setup unit, a solution space acquisition unit, an initial solution acquisition unit, and an optimal solution acquisition unit; The setting unit is used to set the preliminary solution given by the industrial large model as the parameters to be optimized; The solution space acquisition unit is used to determine each parameter in the parameters to be optimized based on the objective function and constraints. The range of values ​​for forms the solution space; The initial solution acquisition unit is used to randomly generate N initial candidate solutions, which satisfy the following distribution in the solution space: ; in L is the dimension The lower bound vector in the solution space, where U is a vector of dimension . The upper bound vector in the solution space. To generate A random number in the interval [0,1); The optimal solution acquisition unit is used to generate the optimal solution by performing Lévy flight iteration on the initial candidate solutions.

10. A generative design and scheduling optimization system based on a large industrial model according to claim 6, characterized in that, The optimal solution acquisition unit includes a first subunit, a second subunit, a third subunit, and a fourth subunit; The first sub-unit is used to generate a random step-size vector s that follows a Lévy distribution; The second sub-unit selects the update position of the initial candidate solution based on probability; A global search update is performed with probability p, where the global search update method is as follows: ； Local search updates are performed with probability 1-p, where the local search update method is as follows: ,in Step scaling factor For element-wise multiplication, Let be one of the candidate solutions in the t-th iteration. This is the optimal solution for the t-th iteration; The third subunit is used to obtain the objective function value after the current position is updated, as the first function value, obtain the objective function value of the optimal solution, as the second function value, and determine whether the first function value is better than the second function value. If it is better, the search result after the current position update is updated to the optimal solution; if it is not better, the optimal solution is kept. The third sub-unit is used to determine whether the current iteration has converged. If it has converged, the optimal solution is output as the optimal solution. If it has not converged, the second and third sub-units are called again.

Images