Production scheduling method and system applied to multi-objective optimization in waterproof roll manufacturing

Abstract

The embodiment of the invention discloses a production scheduling method and system applied to multi-objective optimization in waterproof roll manufacturing, and the method comprises the steps: obtaining the state constraint information of a waterproof roll production line in a preset time period according to a preset multi-objective optimization label, and enabling the information to reflect the limiting conditions which need to be satisfied by the production of the production line; based on the determined target function type and constraint condition set of the multi-target production scheduling model, performing association mapping to obtain the multi-target production scheduling model; solving the model by adopting an improved particle swarm algorithm with an adjusted particle coding mode and a fitness function, and adjusting a particle search behavior to obtain a multi-target optimization optimal solution; and generating a multi-target production scheduling strategy for the production line according to the weight distribution relationship of each target in the optimal solution and the priority ranking result, and guiding the actual production arrangement.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a production scheduling method and system for multi-objective optimization in the manufacturing of waterproof membranes. Background Technology

[0002] In the waterproof membrane manufacturing industry, production scheduling is a crucial link in ensuring efficient and stable production. Traditional production scheduling methods often focus only on a single objective, such as maximizing production efficiency or minimizing production costs, neglecting the impact of other important factors on production. With increasing market competition and diversified customer demands, companies have higher and higher requirements for waterproof membrane production, needing to not only ensure production efficiency but also consider product quality, cost control, and other aspects.

[0003] In recent years, some studies have begun to explore multi-objective optimization methods for production scheduling. Some methods construct multi-objective models that comprehensively consider multiple objectives in the production process. However, during model construction, the acquisition of actual constraints on the production line is often incomplete and inaccurate, leading to discrepancies between the model and actual production conditions. When solving multi-objective models, traditional optimization algorithms suffer from low search efficiency and a tendency to get trapped in local optima, making it difficult to quickly and effectively find the optimal solution for multi-objective optimization. Furthermore, the generated production scheduling strategies often fail to adequately balance the relationships between various objectives and cannot be dynamically adjusted according to actual production conditions, resulting in poor flexibility and adaptability in production scheduling.

[0004] Therefore, the existing model construction lacks a precise grasp of the production line state constraint information, resulting in a disconnect between the model and actual production. Furthermore, the existing solution algorithms are inefficient and prone to getting trapped in local optima, making it difficult to obtain ideal multi-objective optimization solutions, which in turn affects the scientificity and effectiveness of production scheduling strategies. Summary of the Invention

[0005] This application provides a production scheduling method and system for multi-objective optimization in the manufacturing of waterproof membranes.

[0006] This application provides, in one aspect, a multi-objective optimization production scheduling method and a multi-objective optimization production scheduling system for waterproof membrane manufacturing. The method includes: Based on the preset multi-objective optimization labels, the state constraint information of the waterproof membrane production line within a preset time period is obtained. The state constraint information reflects the various restrictions that the production line needs to meet during the production process. Based on the state constraint information, the objective function type and corresponding constraint set of the multi-objective production scheduling model are determined. By associating and mapping the objective function type with the constraint set, the multi-objective production scheduling model is obtained. The multi-objective production scheduling model is solved using an improved particle swarm optimization algorithm with adjustments to the particle encoding method and fitness function. By adjusting the search behavior of the particles during the solution process, the optimal solution for the multi-objective optimization is obtained. Based on the weight distribution relationship and priority ranking results of each objective in the optimal solution of the multi-objective optimization, a multi-objective production scheduling strategy is generated for the waterproof membrane production line. The multi-objective production scheduling strategy is used to guide the actual production arrangement of the production line.

[0007] This application provides a multi-objective optimized production scheduling system, including: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the aforementioned multi-objective optimization production scheduling methods applied in the manufacture of waterproof membranes.

[0008] One embodiment of this application provides a readable storage medium storing a program or instructions, which, when executed by a processor, implements the steps of the production scheduling method for multi-objective optimization in the manufacture of waterproof membranes.

[0009] Therefore, the embodiments of this application achieve multi-objective optimization of waterproof membrane production scheduling. By obtaining state constraint information based on preset multi-objective optimization labels, scheduling inefficiencies caused by missing or inaccurate information are avoided. Based on the state constraint information, the objective function type and constraint condition set are determined, and a multi-objective production scheduling model is obtained through correlation mapping. This allows the model to closely align with actual production, fully considering various factors and objectives in the production process. Compared to traditional single-objective models, it can more comprehensively balance the relationship between multiple objectives such as production efficiency, product quality, and production cost.

[0010] This application also employs an improved particle swarm optimization algorithm with adjusted particle encoding and fitness functions to solve the multi-objective production scheduling model. By optimizing the particle encoding method, the production scheduling scheme is represented more effectively. Simultaneously, the improved fitness function comprehensively considers the optimization of multiple objectives, flexibly adjusting the particle search behavior during the solution process. This allows the algorithm to adaptively switch between global exploration and local development, improving search efficiency and solution accuracy, and enabling it to quickly and accurately find the optimal solution for multi-objective optimization. Finally, a multi-objective production scheduling strategy is generated based on the optimal solution. This strategy fully considers the weight distribution and priority ranking results of each objective, providing scientific and reasonable guidance for the actual production arrangement of waterproof membrane production lines. It ensures that the production line achieves collaborative optimization of multiple objectives while meeting various constraints, thereby improving production efficiency and intelligence. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a multi-objective optimization production scheduling method for the manufacture of waterproof membranes, provided as an embodiment of this application.

[0013] Figure 2 This is a schematic diagram of the basic structure of a multi-objective optimization production scheduling system provided in an embodiment of this application.

[0014] Figure 3 This is a functional block diagram of a multi-objective optimized production scheduling device provided in an embodiment of this application.

[0015] Figure 4 This is a schematic diagram illustrating a technical approach to multi-objective optimization production scheduling provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] Please see Figure 1 , Figure 1 This is a flowchart of a multi-objective optimization production scheduling method for waterproof membrane manufacturing provided in this application embodiment. The method can be executed by a multi-objective optimization production scheduling system, or it can be jointly executed by a multi-objective optimization production scheduling system and a server. The method may include steps S110-S140.

[0018] Step S110: Based on the preset multi-objective optimization label, obtain the state constraint information of the waterproof membrane production line within a preset time period. The state constraint information reflects the various restrictions that the production line needs to meet during the production process.

[0019] In this embodiment, the preset multi-objective optimization labels are pre-set based on the actual needs and objectives of waterproof membrane production, such as improving production efficiency, ensuring product quality, and reducing production costs. The preset time period can be determined according to the needs of production planning and scheduling, such as one day or one week. Through the preset multi-objective optimization labels, state constraint information within the preset time period is obtained from various data sources of the waterproof membrane production line. These data sources include, but are not limited to, the monitoring system of production equipment, raw material supply records, and parameter settings of the production process.

[0020] State constraint information reflects various restrictions that a production line needs to meet during production, such as the maximum production capacity of the equipment, the quantity and quality requirements of raw materials, and the temperature and pressure ranges of the production process. For example, a production equipment monitoring system can provide information such as the equipment's operating status, production speed, and fault records. This information can be used to determine the equipment's maximum production capacity and maintenance needs, thus serving as part of the state constraint information.

[0021] Raw material supply records can provide information such as the quantity of raw materials purchased, delivery time, and quality inspection results. This information can be used to determine the supply quantity and quality requirements of raw materials, thus serving as part of the state constraint information. Production process parameter settings can provide parameters such as temperature, pressure, and time during the production process. These parameters can be used to determine the limiting conditions of the production process, also serving as part of the state constraint information.

[0022] Step S120: Based on the state constraint information, determine the objective function type and corresponding constraint set of the multi-objective production scheduling model. By associating and mapping the objective function type with the constraint set, the multi-objective production scheduling model is obtained.

[0023] In this embodiment, based on the acquired state constraint information, the objective function type and corresponding constraint set of the multi-objective production scheduling model are further determined. The objective function type reflects the goal of production scheduling, such as maximizing production rate, minimizing product defect rate, and reducing raw material consumption. The constraint set is determined based on the state constraint information and is used to limit the range of values ​​for the objective function. By associating and mapping the objective function type with the constraint set, a correspondence is established between them, thereby obtaining the multi-objective production scheduling model. For example, if the state constraint information indicates that the maximum production capacity of the equipment is limited, then when determining the objective function type, maximizing the production rate can be considered as an objective function type, and at the same time, a limit on the equipment's production capacity can be set in the constraint set. In this way, when solving the multi-objective production scheduling model, the production rate can be maximized while satisfying the equipment's production capacity limit.

[0024] Step S121: Classify the state constraint information to distinguish between efficiency-related constraints affecting production efficiency, quality-related constraints affecting product quality, and cost-related constraints affecting production costs.

[0025] In this embodiment, the acquired state constraint information is categorized into efficiency-related constraints, quality-related constraints, and cost-related constraints. Efficiency-related constraints primarily affect production efficiency, such as equipment production speed and maintenance time. Quality-related constraints primarily affect product quality, such as raw material quality requirements and the precision of production process parameters. Cost-related constraints primarily affect production costs, such as raw material purchase prices and equipment energy consumption. This categorization allows for a more comprehensive understanding of the impact of different types of constraints on production scheduling. For example, for efficiency-related constraints, if equipment production speed is slow, maximizing production rate can be considered as an objective function type when determining the objective function type, while setting restrictions on equipment production speed in the constraint set. For quality-related constraints, if raw material quality requirements are high, minimizing product defect rate can be considered as an objective function type when determining the objective function type, while setting restrictions on raw material quality in the constraint set. For cost-related constraints, if the purchase price of raw materials is high, then when determining the objective function type, reducing raw material consumption can be considered as an objective function type, and restrictions on the purchase price of raw materials can be set in the constraint set.

[0026] Step S122: Determine the first type of objective function with the goal of maximizing production rate based on efficiency constraints, determine the second type of objective function with the goal of minimizing product defect rate based on quality constraints, and determine the third type of objective function with the goal of reducing raw material consumption based on cost constraints.

[0027] In this embodiment, different types of objective functions are determined based on the categorized efficiency, quality, and cost constraints. The first type of objective function, based on the efficiency constraints, aims to maximize production rate; in production scheduling, this means maximizing production speed to meet market demand. The second type of objective function, based on the quality constraints, aims to minimize the product defect rate; in production scheduling, this means ensuring product quality and reducing the number of defective products. The third type of objective function, based on the cost constraints, aims to reduce raw material consumption; in production scheduling, this means using raw materials rationally to reduce production costs.

[0028] For example, for efficiency constraints, if the equipment's production speed is slow, the objective function type for the first category can be to maximize the production rate, which can be achieved by optimizing equipment operating parameters and adjusting the production process. For quality constraints, if the raw material quality requirements are high, the objective function type for the second category can be to minimize the product defect rate, which can be achieved by strengthening raw material quality inspection and optimizing production processes. For cost constraints, if the raw material purchase price is high, the objective function type for the third category can be to reduce raw material consumption, which can be achieved by optimizing raw material usage and improving raw material utilization rates to reduce production costs.

[0029] Step S123: For each type of objective function, extract the corresponding limitation range and related factors from the corresponding constraint information to form a set of constraints corresponding to each type of objective function.

[0030] In this embodiment, for each type of objective function, corresponding limitations and related factors are extracted from the corresponding constraint information to form a set of constraints corresponding to each type of objective function. For the first type of objective function, which aims to maximize production rate, limitations and related factors such as the maximum production capacity of equipment and the maintenance time of equipment are extracted from efficiency-related constraint information to form a corresponding set of constraints. For the second type of objective function, which aims to minimize product defect rate, limitations and related factors such as the quality requirements of raw materials and the precision of production process parameters are extracted from quality-related constraint information to form a corresponding set of constraints. For the third type of objective function, which aims to reduce raw material consumption, limitations and related factors such as the purchase price of raw materials and the supply quantity of raw materials are extracted from cost-related constraint information to form a corresponding set of constraints.

[0031] For example, for the first type of objective function, if efficiency constraints indicate that the equipment's maximum production capacity is a certain quantity of waterproof membrane per hour, then when forming the constraint set, the equipment's maximum production capacity can be used as a limiting range, while the equipment's maintenance time can be considered as a related factor, taking into account the impact on production rate during equipment maintenance. For the second type of objective function, if quality constraints indicate that the raw material's quality requirements are specific chemical composition and physical properties, then when forming the constraint set, the raw material's quality requirements can be used as a limiting range, while the precision of the production process parameters can be considered as a related factor, taking into account the impact of production process stability on product defect rate. For the third type of objective function, if cost constraints indicate that the raw material's purchase price is high, then when forming the constraint set, the raw material's purchase price can be used as a limiting range, while the raw material's supply quantity can be considered as a related factor, taking into account the impact of raw material supply stability on raw material consumption.

[0032] Step S124: Construct the association mapping rules between the objective function type and the set of constraints, bind different types of objective functions with their corresponding constraints, and integrate the binding relationship into the model structure by establishing mathematical relational expressions to obtain a multi-objective production scheduling model.

[0033] In this embodiment, an association mapping rule is constructed between objective function types and constraint sets to bind different types of objective functions to their corresponding constraints. The association mapping rule defines the correspondence between objective function types and constraint sets; for example, the first type of objective function corresponds to the efficiency constraint set, the second type corresponds to the quality constraint set, and the third type corresponds to the cost constraint set. By establishing a mathematical relational expression, the binding relationship is integrated into the model structure, resulting in a multi-objective production scheduling model. This mathematical relational expression describes the interaction and constraint relationships between objective function types and constraint sets, enabling the simultaneous consideration of multiple objectives and constraints when solving the multi-objective production scheduling model.

[0034] For example, mathematical relationships can be established using methods such as linear programming and nonlinear programming. For linear programming, the objective function type and the set of constraints can be represented as linear equations or inequalities, and the optimal solution is obtained by solving the linear programming problem. For nonlinear programming, the objective function type and the set of constraints can be represented as nonlinear equations or inequalities, and the optimal solution is obtained by solving the nonlinear programming problem. When establishing mathematical relationships, it is crucial to ensure that the dimensions of the objective function type and the set of constraints are consistent to avoid unreasonable calculation results.

[0035] Step S130: Solve the multi-objective production scheduling model based on the improved particle swarm optimization algorithm with adjusted particle encoding method and fitness function. By adjusting the search behavior of particles during the solution process, the optimal solution for multi-objective optimization is obtained.

[0036] In this embodiment, an improved particle swarm optimization (PSO) algorithm with adjusted particle encoding and fitness functions is used to solve a multi-objective production scheduling model. PSO is a swarm intelligence-based optimization algorithm that finds the optimal solution by simulating the collective behavior of flocks of birds or schools of fish. In this embodiment, the particle encoding and fitness functions are adjusted to suit the characteristics of the multi-objective production scheduling model.

[0037] During the solution process, the search behavior of the particles is adjusted to enable them to search for the optimal solution more efficiently in the solution space. For example, by adjusting the inertia weight and learning factor of the particles, the particles can adaptively switch between global exploration and local exploitation during the search process, improving search efficiency. Finally, through iterative solving of the improved particle swarm optimization algorithm, a multi-objective optimization optimal solution is obtained. This optimal solution satisfies all constraints and simultaneously optimizes multiple objective functions.

[0038] Step S131: The particles are encoded using a hybrid coding method based on process sequence and equipment allocation. Each particle represents a production scheduling scheme. The process sequence coding reflects the execution order of each process, and the equipment allocation coding reflects the production equipment used in each process.

[0039] In this embodiment, a hybrid coding method based on process sequence and equipment allocation is used to encode particles. Each particle represents a production scheduling scheme, which is specifically described by process sequence coding and equipment allocation coding. The process sequence coding reflects the order in which each process is executed. For example, for waterproof membrane production, it may include processes such as raw material preparation, mixing, coating, and drying. The process sequence coding can determine the execution order of these processes. The equipment allocation coding reflects the production equipment used in each process. For example, the raw material preparation process may use specific mixing equipment, and the coating process may use specific coating machines. Through this hybrid coding method, the production scheduling scheme can be represented as particles in an efficient way.

[0040] For example, in a waterproof membrane production process with five steps, the step sequence code can be a sequence of five elements, where each element represents the step number, and the steps are executed in the order of this sequence. Similarly, the equipment allocation code can be a sequence of five elements, where each element represents the identifier of the production equipment used in the corresponding step. Thus, a single element can completely represent a production scheduling scheme.

[0041] Step S1311: For all processes in the production of waterproof membrane, determine the range of process execution sequence according to the production process requirements, and assign a unique number to each process within the range of process execution sequence to indicate its position in the process sequence coding.

[0042] In this embodiment, for all processes in the production of waterproof membrane, the execution sequence range of each process is determined according to the production process requirements. The production process requirements specify the sequential relationship and logical order between each process; for example, the raw material preparation process must be completed before the mixing process, and the mixing process must be completed before the coating process. Within the determined execution sequence range, a unique sequence number is assigned to each process, indicating its position in the process sequence code. For example, for the five processes in the production of waterproof membrane, their execution sequence range is determined according to the production process requirements, and then these five processes are assigned sequence numbers 1-5 respectively. Thus, in the process sequence code, the order of the sequence numbers represents the order in which each process is executed. This method ensures that the process sequence code accurately reflects the requirements of the production process, guaranteeing the feasibility of the production scheduling plan.

[0043] Step S1312: For each process, compile a list of production equipment that can be used in that process, assign a unique identifier to each piece of equipment, and use the identifier to represent the production equipment selected for the corresponding process in the equipment assignment code.

[0044] In this embodiment, a list of production equipment that can be used in each step of the waterproof membrane production process is compiled. Different steps may require different types of production equipment; for example, the raw material preparation step may require mixing equipment, and the coating step may require a coating machine. A unique identifier is assigned to each piece of equipment that can be used in each step. These identifiers are used in the equipment allocation code to represent the production equipment selected for the corresponding step. For example, for the raw material preparation step, the available production equipment includes mixing equipment A and mixing equipment B. Identifier 1 is assigned to mixing equipment A, and identifier 2 is assigned to mixing equipment B. If mixing equipment A is selected in a certain production scheduling scheme, then the position of the raw material preparation step is represented by identifier 1 in the equipment allocation code. Through this method, the production equipment used in each step can be accurately recorded, providing specific equipment information for the implementation of the production scheduling scheme.

[0045] Step S1313: Combine the process sequence code and the equipment allocation code in a preset order to form a complete particle code. The length of the process sequence code is equal to the total number of production processes, and the length of the equipment allocation code is also equal to the total number of production processes.

[0046] In this embodiment, the process sequence code and equipment allocation code are combined in a preset order to form a complete particle code. The preset order can be determined according to actual needs; for example, the process sequence code can be placed first, followed by the equipment allocation code. The length of the process sequence code is equal to the total number of production processes, because each process has a corresponding sequence number in the process sequence code. The length of the equipment allocation code is also equal to the total number of production processes, because each process needs to select a production device, which is represented by an identifier in the equipment allocation code. For example, for a waterproof membrane production process with 5 processes, the process sequence code is a sequence containing 5 elements, and the equipment allocation code is also a sequence containing 5 elements. Combining them in a preset order forms a complete particle code containing 10 elements, which represents a specific production scheduling scheme.

[0047] Step S132: Design a new fitness function based on the type of objective function in the multi-objective production scheduling model. The new fitness function comprehensively considers the optimization of multiple objectives. By weighting the degree of achievement of each objective, the fitness value of the particle is obtained.

[0048] In this embodiment, a new fitness function is designed based on the type of objective function in the multi-objective production scheduling model. Since the multi-objective production scheduling model has multiple objective functions, the new fitness function needs to comprehensively consider the optimization of multiple objectives. The fitness value of a particle is obtained by weighting the achievement degree of each objective. This weighting calculation considers the importance of each objective, and different objectives can be assigned different weights.

[0049] For example, for the three objective functions of maximizing production rate, minimizing product defect rate, and reducing raw material consumption, different weights can be assigned to them according to actual needs. If more emphasis is placed on production rate, a higher weight can be assigned to maximizing production rate; if more emphasis is placed on product quality, a higher weight can be assigned to minimizing product defect rate. In this way, the new fitness function can more accurately evaluate the merits of the production scheduling scheme represented by each particle.

[0050] Step S1321: Construct a multi-task learning framework, mapping the objective function type in the multi-objective production scheduling model to the parallel task branches of the framework, with each task branch corresponding to a feature extraction channel of a type of objective function.

[0051] In this embodiment, a multi-task learning framework is constructed, mapping the objective function types in the multi-objective production scheduling model to parallel task branches of the framework. The multi-task learning framework can process multiple related tasks simultaneously. In this embodiment, each objective function type corresponds to one parallel task branch. Each task branch has a corresponding feature extraction channel for extracting features related to that objective function type.

[0052] For example, for the objective function type of maximizing production rate, the feature extraction channel of the corresponding task branch can extract features related to production rate, such as equipment production speed and process execution time. For the objective function type of minimizing product defect rate, the feature extraction channel of the corresponding task branch can extract features related to product quality, such as raw material quality and production process parameters. Through this method, the multi-task learning framework can process multiple objective function types in parallel, improving computational efficiency.

[0053] Step S1322: In each task branch, the correlation features between the objective function type and the set of constraints are dynamically captured through an attention mechanism to generate an objective correlation feature vector, which includes the objective achievement difficulty coefficient and constraint sensitivity parameters.

[0054] In this embodiment, an attention mechanism is used to dynamically capture the correlation features between the objective function type and the constraint set in each task branch. The attention mechanism automatically focuses on important features related to the objective function type and ignores irrelevant features. Through the attention mechanism, the correlation between the objective function type and the constraint set can be captured more accurately. The generated objective-related feature vector includes the objective achievement difficulty coefficient and constraint sensitivity parameters.

[0055] The objective achievement difficulty coefficient reflects the ease or difficulty of achieving the objective function. For example, if the equipment's production capacity is limited, maximizing the production rate will have a higher objective achievement difficulty coefficient. The constraint sensitivity parameter reflects the sensitivity of the objective function to constraints. For instance, if the quantity of raw materials supplied has a significant impact on product quality, minimizing the product defect rate will be highly sensitive to the constraint of the quantity of raw materials supplied. By using objective-related feature vectors, a deeper understanding of the relationship between each type of objective function and the set of constraints can be gained.

[0056] Step S1323: Perform hierarchical processing on the particle population based on the non-dominated sorting mechanism, calculate the crowding distance of particles in each layer, and construct a dynamic weight generation network based on the crowding distance and the target association feature vector.

[0057] In this embodiment, the particle population is hierarchically processed based on a non-dominated ranking mechanism. This mechanism divides the particle population into different levels, where particles in each level are not dominated by particles above that level. Within each level, the crowding distance of particles is calculated, reflecting the particle distribution density within that level. A dynamic weight generation network is constructed based on the crowding distance and the target association feature vector. This network dynamically generates weights for each objective function type based on the particle crowding distance and the target association feature vector.

[0058] For example, if the objective function type has a high achievement difficulty coefficient and the corresponding particles have a large crowding distance at a certain level, the dynamic weight generation network can assign higher weights to this objective function type to encourage particles to focus more on optimizing this objective during the search process. In this way, the dynamic weight generation network can adaptively adjust the weights of each objective function type based on the particle's search behavior, improving the algorithm's search efficiency.

[0059] Step S1324: Output the real-time weight coefficients of each objective function type through the dynamic weight generation network. The real-time weight coefficients are adaptively adjusted as the degree of objective achievement changes during the particle search process.

[0060] In this embodiment, a dynamic weight generation network outputs real-time weight coefficients for each objective function type. These real-time weight coefficients adaptively adjust as the degree of objective achievement changes during the particle search process. In the early stages of the particle search, global exploration may be more necessary. The dynamic weight generation network can assign relatively balanced weights to each objective function type, allowing particles to search extensively in the solution space. As the search progresses, if the objective achievement of a certain objective function type is relatively good, the dynamic weight generation network can appropriately reduce the weight of that objective function type, shifting more attention to other objective function types that have not yet been well optimized.

[0061] For example, if maximizing production rate has been well optimized during the search process, but minimizing product defect rate requires further optimization, the dynamic weight generation network can reduce the weight of the objective function type maximizing production rate and increase the weight of the objective function type minimizing product defect rate. Through this adaptive adjustment of real-time weight coefficients, particles can more specifically optimize each objective function during the search process, improving the effectiveness of multi-objective optimization.

[0062] Step S1325: The deviation between the actual optimized value and the ideal optimized value of each objective function type is weighted and fused using real-time weighting coefficients to generate a comprehensive fitness value. The actual optimized value and the ideal optimized value have the same dimensions, and the deviation is calculated using Euclidean distance in the feature space.

[0063] In this embodiment, the deviation between the actual and ideal optimized values ​​of each objective function type is weighted and fused using real-time weighting coefficients to generate a comprehensive fitness value. The actual optimized value refers to the actual optimization effect achieved by each objective function type under the current production scheduling scheme. The ideal optimized value refers to the optimal effect achievable by each objective function type. The deviation is calculated using Euclidean distance in the feature space, reflecting the difference between the actual and ideal optimized values. Weighting and fusion of the deviation using real-time weighting coefficients takes into account the importance of each objective function type.

[0064] For example, if the real-time weight coefficient of a certain objective function type is high, then the proportion of the bias of that objective function type in the weighted fusion will be larger. In this way, the comprehensive fitness value can more comprehensively reflect the optimization effect of the current production scheduling scheme on multiple objective function types, providing a more accurate evaluation standard for particle search.

[0065] Step S133: Initialize the position and velocity of the particle swarm. During the particle swarm iteration, adjust the inertia weight and learning factor according to the fitness value of the particles, so that the particles can adaptively switch between global exploration and local development during the search process.

[0066] In this embodiment, the position and velocity of the particle swarm are initialized. The position of the particle swarm represents the production scheduling scheme represented by each particle, and the velocity represents the direction and speed of movement of the particle in the solution space. During the particle swarm iteration process, the inertia weight and learning factor are adjusted according to the fitness value of the particles. The inertia weight affects the particle's ability to maintain its current velocity, and the learning factor affects the particle's ability to learn towards the individual optimal solution and the global optimal solution. By adjusting the inertia weight and learning factor, the particles adaptively switch between global exploration and local development during the search process.

[0067] In the early stages of iteration, larger inertia weights and learning factors enable particles to have strong global exploration capabilities, allowing them to search extensively within the solution space. As iteration progresses, the inertia weights and learning factors are gradually reduced, allowing particles to focus more on local exploration to find better solutions. For example, in the early stages of iteration, both the inertia weights and learning factors can be set to large values, enabling particles to move quickly within the solution space and explore different production scheduling schemes. In the later stages of iteration, the inertia weights can be gradually reduced, and the learning factor can also be appropriately decreased, allowing particles to focus more on searching the current local region and improving search accuracy.

[0068] Step S134: Stop the iteration operation when the number of particle swarm iterations reaches the preset maximum number of iterations or the fitness value of the population converges after multiple consecutive iterations, and obtain the optimal solution for multi-objective optimization based on the particle swarm that has completed the iterations.

[0069] In this embodiment, a preset maximum number of iterations and convergence conditions are set. The iteration operation stops when the particle swarm optimization reaches the preset maximum number of iterations. Alternatively, the iteration operation also stops when the fitness value of the population converges after multiple consecutive iterations. Fitness convergence indicates that the particle swarm has found a relatively stable region in the solution space, and further iteration may not yield a better solution. The optimal solution for multi-objective optimization is obtained by filtering the particle swarm that has completed iterations. The filtering process can be based on the comprehensive fitness value, selecting the production scheduling scheme represented by the particle with the optimal comprehensive fitness value as the optimal solution for multi-objective optimization. For example, if a smaller comprehensive fitness value indicates a better production scheduling scheme, then the particle with the smallest comprehensive fitness value can be selected as the optimal solution for multi-objective optimization. Through the above method, and after iterative solving using the improved particle swarm optimization algorithm, a production scheduling scheme that satisfies multiple objective functions simultaneously is finally obtained.

[0070] Step S140: Based on the weight allocation relationship and priority ranking result of each objective in the optimal solution of the multi-objective optimization, a multi-objective production scheduling strategy for the waterproof membrane production line is generated. The multi-objective production scheduling strategy is used to guide the actual production arrangement of the production line.

[0071] In this embodiment, a multi-objective production scheduling strategy for a waterproof membrane production line is generated based on the weight allocation relationship and priority ranking results of each objective in the optimal solution of the multi-objective optimization. The optimal solution of the multi-objective optimization includes the weight allocation relationship and priority ranking results of each objective, which reflects the importance and order of each objective under the current production conditions. Using this information, combined with the actual situation of the waterproof membrane production line, a multi-objective production scheduling strategy is generated. This strategy guides the actual production arrangement of the production line, including the execution sequence of processes, equipment allocation, and raw material supply.

[0072] For example, if maximizing production rate has a higher weight and priority in the optimal solution of a multi-objective optimization, then when generating a multi-objective production scheduling strategy, measures to improve production rate, such as optimizing process flows and increasing equipment investment, can be given priority. At the same time, the achievement of other objectives should also be considered to ensure that product quality and production costs are maintained while increasing production rate.

[0073] Step S141: Perform feature analysis on the optimal solution of the multi-objective optimization, extract the target weight allocation matrix and priority ranking sequence to construct a target decision directed graph, where the nodes in the target decision directed graph represent production targets and the directed edges represent the priority dependencies between targets.

[0074] In this embodiment, feature analysis is performed on the optimal solution of multi-objective optimization to extract the objective weight allocation matrix and priority ranking sequence. The objective weight allocation matrix records the weight distribution of each objective, and the priority ranking sequence records the priority order of each objective. A directed graph of objective decision is constructed using the objective weight allocation matrix and the priority ranking sequence. The directed graph of objective decision includes nodes and directed edges used to connect the nodes. The nodes in the directed graph of objective decision represent production objectives, such as maximizing production rate, minimizing product defect rate, and reducing raw material consumption. The directed edges represent the priority dependencies between objectives. For example, if maximizing production rate has a higher priority than minimizing product defect rate, then in the directed graph of objective decision, there will be a directed edge from the node that maximizes production rate to the node that minimizes product defect rate. The directed graph of objective decision provides a more intuitive representation of the relationships between various production objectives.

[0075] Step S142: Based on the target decision directed graph, a graph neural network is used to propagate node features and generate graph embedding vectors containing target cooperation relationships. Potential target conflict pairs are identified by calculating the similarity of the graph embedding vectors.

[0076] In this embodiment, a graph neural network is used for node feature propagation based on a directed graph of target decisions. Graph neural networks can process graph-structured data and propagate features between nodes through a message passing mechanism. In the directed graph of target decisions, node features include target weight values, priority coefficients, and resource requirement vectors. Through node feature propagation in the graph neural network, graph embedding vectors containing target collaboration relationships are generated. Graph embedding vectors map node information in the directed graph of target decisions to a low-dimensional vector space, facilitating computation and analysis. By calculating the similarity of graph embedding vectors, potential target conflict pairs are identified. Nodes corresponding to graph embedding vectors with high similarity may have target conflicts; for example, two targets may compete for resource requirements or constrain each other during implementation. By identifying potential target conflict pairs, potential problems in production scheduling can be detected in advance.

[0077] Step S1421: Initialize the node feature vector of the target decision directed graph, wherein the node feature vector includes the target weight value, priority coefficient and resource requirement vector.

[0078] In this embodiment, the node feature vectors of the directed graph for target decision are initialized. The node feature vectors include target weight values, priority coefficients, and resource requirement vectors. The target weight values ​​reflect the importance of each target in multi-objective optimization, the priority coefficients reflect the priority order of each target, and the resource requirement vectors reflect the resources required for each target to be implemented.

[0079] For example, for the objective of maximizing production rate, the objective weights can be determined based on the weight distribution relationship in the optimal solution of multi-objective optimization, the priority coefficients can be determined based on the priority ranking sequence, and the resource demand vector can include equipment usage time, raw material consumption, etc. By initializing the node feature vectors, initial information can be provided for the node feature propagation of the graph neural network, ensuring that the graph neural network can accurately process the directed graph of objective decision.

[0080] Step S1422: Construct the message passing layer of the graph neural network, define the propagation path between nodes through the adjacency matrix, and fuse the node's own features and the features of its neighboring nodes during the propagation process to generate an intermediate feature vector.

[0081] In this embodiment, a message passing layer of a graph neural network is constructed. The message passing layer is the core layer of the graph neural network, defining the propagation path between nodes through an adjacency matrix. The adjacency matrix records the connection relationships between nodes in the directed graph of the target decision, and the neighboring nodes of each node can be determined through the adjacency matrix. During message passing, the node's own features and the features of its neighboring nodes are fused to generate an intermediate feature vector. For example, for a given node, it receives feature information from its neighboring nodes and fuses its own features with the features of its neighboring nodes. Through this method, the intermediate feature vector can contain more information about the relationships between nodes.

[0082] Step S1423: Perform graph attention aggregation on the intermediate feature vector, dynamically allocate attention weights according to the correlation strength between nodes, and enhance the feature representation of key target nodes.

[0083] In this embodiment, graph attention aggregation is performed on the intermediate feature vector. Graph attention aggregation dynamically allocates attention weights based on the correlation strength between nodes. Correlation strength reflects the importance and relevance between nodes. For example, in a directed graph of target decision, if there is a strong priority dependency between two targets, their correlation strength is high. By dynamically allocating attention weights, the feature representation of key target nodes is strengthened. Key target nodes refer to target nodes that play an important role in production scheduling, such as target nodes with high priority. By strengthening the feature representation of key target nodes, the graph embedding vector can more accurately reflect the structure and information of the directed graph of target decision, improving the identification accuracy of potential target conflict pairs.

[0084] Step S1424: Generate the graph embedding vector of the node through the output layer of the graph neural network, calculate the cosine similarity between the graph embedding vectors of any two nodes, and set a similarity threshold to filter potential target conflict pairs.

[0085] In this embodiment, graph embedding vectors of nodes are generated through the output layer of a graph neural network. These graph embedding vectors map node information in the directed graph of the target decision to a low-dimensional vector space, facilitating similarity calculation. The cosine similarity between any two node graph embedding vectors is calculated; cosine similarity reflects the angle between the two vectors, with a smaller angle indicating higher similarity. A similarity threshold is set, and node pairs with cosine similarity higher than the threshold are identified as potential target conflict pairs. For example, if the similarity threshold is set to a certain value, when the cosine similarity between two node graph embedding vectors is higher than that threshold, it is considered that the targets corresponding to these two nodes may conflict. Through this method, potential target conflict pairs can be quickly identified.

[0086] Step S1425: Classify the selected target conflict pairs by conflict type, and generate a conflict severity score based on the conflict type. The score is used to guide the priority of conflict resolution processing.

[0087] In this embodiment, the selected target conflict pairs are categorized by conflict type. Conflict types can be divided into resource competition conflicts, goal achievement conflicts, etc. Resource competition conflicts refer to a competition between two goals for resource needs; for example, both goals need to use the same equipment. Goal achievement conflicts refer to mutual constraints between two goals during the achievement process; for example, the achievement of one goal may increase the difficulty of achieving the other goal. A conflict severity score is generated based on the conflict type, reflecting the degree of impact of the conflict on production scheduling. For example, for resource competition conflicts, if the competing resource is a critical resource, the conflict severity score will be higher. The conflict severity score is used to guide the priority of conflict resolution processing; conflict pairs with higher scores are processed first to ensure the smooth operation of production scheduling.

[0088] Step S143: Perform conflict resolution processing on the target conflict pair, and generate a conflict resolution action sequence through a reinforcement learning policy network. The conflict resolution action sequence includes resource reallocation rules and task splitting strategies.

[0089] In this embodiment, conflict resolution processing is performed on target conflict pairs. A conflict resolution action sequence is generated using a reinforcement learning policy network. A reinforcement learning policy network is a neural network based on reinforcement learning, which can learn the optimal action policy based on the current state and reward signal. The conflict resolution action sequence includes resource reallocation rules and task splitting strategies. Resource reallocation rules are used to adjust the allocation of resources; for example, allocating resources needed by one target to another to resolve resource contention conflicts. Task splitting strategies are used to break down a task into multiple sub-tasks to alleviate target achievement conflicts. For example, if two targets conflict in time, one task can be split into multiple sub-tasks, and the execution time of the sub-tasks can be adjusted to avoid conflict. The conflict resolution action sequence generated by the reinforcement learning policy network can effectively resolve target conflict pairs, improving the feasibility and efficiency of production scheduling.

[0090] Step S1431: Use the conflict feature vector of the target conflict pair and the current resource state vector as input to the reinforcement learning policy network. The conflict feature vector includes conflict type, severity score and target correlation.

[0091] In this embodiment, the conflict feature vector of the target conflict pair and the current resource state vector are used as inputs to the reinforcement learning policy network. The conflict feature vector includes conflict type, severity score, and target correlation. The conflict type reflects the specific type of the target conflict pair, such as resource competition conflict or target achievement conflict. The severity score reflects the degree of impact of the conflict on production scheduling. The target correlation reflects the correlation between two targets. The current resource state vector reflects the resource usage in the current production process, such as equipment idle time and raw material inventory. By using the conflict feature vector and the current resource state vector as inputs, the reinforcement learning policy network can generate a suitable conflict resolution action sequence based on the current conflict situation and resource state.

[0092] Step S1432: In the exploration phase of reinforcement learning, a goal selection strategy is adopted to select conflict resolution actions. The action space includes basic operations such as resource transfer, task delay, and process adjustment.

[0093] In this embodiment, during the exploration phase of reinforcement learning, a target selection strategy is employed to choose conflict resolution actions. The exploration phase is the initial stage of reinforcement learning, where different actions are tried to find the optimal action strategy. The action space includes basic operations such as resource transfer, task delay, and process adjustment. Resource transfer refers to transferring resources needed for one objective to another to resolve resource contention conflicts. Task delay refers to postponing the execution time of a task to avoid time conflicts. Process adjustment refers to adjusting the parameters of the production process to improve the achievement of the objective. Through the target selection strategy, appropriate conflict resolution actions can be selected from the action space.

[0094] Step S1433: In the reinforcement learning utilization phase, the selection probability of each action is output through the policy network, and a conflict resolution action sequence is generated based on probability sampling. The length of the conflict resolution action sequence is adjusted according to the severity of the conflict.

[0095] In this embodiment, during the utilization phase of reinforcement learning, the policy network outputs the selection probability of each action. The utilization phase is a later stage of reinforcement learning, where experience has been accumulated, allowing for the selection of better actions based on that experience. The policy network outputs the selection probability of each action based on the input conflict feature vector and the current resource state vector. A conflict resolution action sequence is generated based on probability sampling, where probability sampling refers to randomly selecting actions according to their selection probabilities. The length of the conflict resolution action sequence is adjusted according to the conflict severity; higher conflict severity may result in a longer sequence, requiring more actions to resolve the conflict. This method allows for more efficient generation of conflict resolution action sequences during the utilization phase, improving the efficiency of conflict resolution.

[0096] Step S1434: Construct a reward function. The inputs to the reward function include the target achievement rate after conflict resolution, resource utilization rate, and task delay time. The output is a comprehensive reward value.

[0097] In this embodiment, a reward function is constructed. The inputs to the reward function include the goal achievement rate after conflict resolution, resource utilization rate, and task delay time. The goal achievement rate reflects the degree to which each goal is achieved after executing the conflict resolution action sequence. Resource utilization rate reflects the efficiency of resource use, such as whether equipment usage time is sufficient and whether raw material consumption is reasonable. Task delay time reflects whether the task is completed on time. The reward function outputs a comprehensive reward value based on these inputs, which reflects the effectiveness of conflict resolution. For example, if the goal achievement rate, resource utilization rate, and task delay time are high after conflict resolution, the comprehensive reward value will be high. The reward function provides feedback to the reinforcement learning policy network, guiding its learning and optimization.

[0098] Step S1435: Optimize the policy network parameters using the deep deterministic policy gradient algorithm so that the generated conflict resolution action sequence is used to maximize the comprehensive reward value and achieve the resolution of the target conflict.

[0099] In this embodiment, the policy network parameters are optimized using a deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm is a reinforcement learning algorithm used to optimize a continuous action space. It continuously adjusts the parameters of the policy network based on feedback from the reward function. By optimizing the policy network parameters, the generated conflict resolution action sequence is used to maximize the overall reward value. When the overall reward value reaches its maximum, it indicates that the conflict resolution action sequence can most effectively resolve the target conflict, thus achieving conflict resolution. Through this method, the deep deterministic policy gradient algorithm can continuously improve the performance of the reinforcement learning policy network, ensuring that the conflict resolution action sequence can effectively resolve target conflict pairs, thereby improving the feasibility and efficiency of production scheduling.

[0100] Step S144: Input the resolved target weight allocation matrix and priority sorting sequence into the production task scheduler, and use the production task scheduler to predict the task execution time window based on the long short-term memory network to generate the target production task sequence.

[0101] In this embodiment, the resolved target weight allocation matrix and priority ranking sequence are input into the production task scheduler. The production task scheduler predicts the task execution time window based on a Long Short-Term Memory (LSTM) network. A LSTM network is a recurrent neural network that can process sequential data and capture long-term dependencies within the data. In production scheduling, the task execution time window includes the earliest start time, latest end time, and duration of the task. Predicting the task execution time window using a LSTM network allows for a more accurate arrangement of task execution order. A target production task sequence is generated based on the task execution time window, which is the order in which tasks are executed according to the task execution time window and priority ranking sequence. For example, if the earliest start time and latest end time of a task are determined, and the task has a high priority, then that task will be placed in a suitable position in the target production task sequence. Through this method, the production task scheduler can generate a reasonable target production task sequence based on the resolved target weight allocation matrix and priority ranking sequence, providing specific task arrangements for production scheduling.

[0102] Step S1441: Perform matrix decomposition on the resolved target weight allocation matrix to obtain the task resource requirement matrix and equipment capability matrix, and calculate the task-equipment matching degree through matrix multiplication.

[0103] In this embodiment, the resolved target weight allocation matrix is ​​decomposed to obtain a task resource requirement matrix and an equipment capability matrix. The task resource requirement matrix records the resource requirements of each task, such as the quantity of equipment and raw materials needed for each task. The equipment capability matrix records the production capacity of each piece of equipment, such as the maximum production speed and maximum load capacity. The task-equipment matching degree is calculated through matrix multiplication, reflecting the degree of matching between each task and each piece of equipment. For example, if the production capacity of the equipment required for a task matches the actual production capacity of a piece of equipment, then the task-equipment matching degree between them is high. By calculating the task-equipment matching degree, suitable equipment can be allocated to tasks, improving production efficiency.

[0104] Step S1442: Input the task-device matching degree and priority ranking sequence into the encoder of the long short-term memory network, and use the encoder to selectively memorize key task features through a gating mechanism to generate task temporal feature vectors.

[0105] In this embodiment, the task-device matching degree and priority ranking sequence are input into the encoder of the Long Short-Term Memory (LSTM) network. The encoder, part of the LTM network, selectively memorizes key task features through a gating mechanism. This gating mechanism controls the inflow and outflow of information, enabling the LTM network to capture long-term dependencies in sequence data. In this embodiment, the encoder selectively memorizes key task features, such as task execution time and device usage, based on the task-device matching degree and priority ranking sequence. The encoder generates a task temporal feature vector, which contains the task's temporal order and feature information.

[0106] Step S1443: The decoder of the long short-term memory network predicts the earliest start time, latest end time and duration of each task based on the task temporal feature vector, forming a task execution time window.

[0107] In this embodiment, the decoder of the Long Short-Term Memory (LSTM) network predicts the earliest start time, latest end time, and duration of each task based on the task's temporal feature vector. The decoder, another part of the LSM network, outputs the task's time information based on the task's temporal feature vector. The earliest start time refers to the earliest time a task can begin execution, the latest end time refers to the latest time a task must complete, and the duration refers to the time required for the task to complete from start to finish. By predicting this time information, a task execution time window is formed.

[0108] For example, for a given task, the decoder predicts the earliest start time (a certain moment), the latest end time (another moment), and the duration (a certain period) of the task based on the task's temporal feature vector. This temporal information constitutes the task's execution time window. The decoder of a Long Short-Term Memory (LSTM) network can accurately predict this task execution time window.

[0109] Step S1444: Allocate tasks according to the task execution time window and equipment capacity matrix, and use a greedy algorithm to construct the original production task sequence so that the task allocation meets the equipment load constraints.

[0110] In this embodiment, tasks are allocated based on the task execution time window and the equipment capacity matrix. The task execution time window defines the execution time range for each task, and the equipment capacity matrix defines the production capacity of each device. A greedy algorithm is used to construct the original production task sequence. A greedy algorithm is an algorithm that makes the optimal choice in each step based on the current state. During task allocation, the greedy algorithm prioritizes tasks that satisfy equipment load constraints. Equipment load constraints refer to the production capacity limitations of a device within a certain time period; for example, a device cannot execute multiple tasks simultaneously within a certain time period. By constructing the original production task sequence using a greedy algorithm, the utilization rate of the equipment can be maximized and the execution order of tasks can be rationally arranged while satisfying equipment load constraints.

[0111] Step S1445: Verify the original production task sequence by including resource conflict detection, time window overlap calculation and target achievement rate evaluation, and take the sequence that passes the verification as the target production task sequence.

[0112] In this embodiment, the original production task sequence is validated. The validation includes resource conflict detection, time window overlap calculation, and target achievement rate assessment. Resource conflict detection checks for resource contention within the original production task sequence; for example, multiple tasks may simultaneously require the use of the same equipment. Time window overlap calculation measures the degree of overlap between the task execution time windows in the original production task sequence; excessive overlap may lead to task execution chaos. Target achievement rate assessment evaluates the degree to which each target of the original production task sequence is achieved during execution, such as whether it can maximize production rate or minimize product defect rate. The validated sequence is used as the target production task sequence; only validated sequences ensure the feasibility and effectiveness of production scheduling. For example, if resource conflicts exist in the original production task sequence, task allocation needs to be adjusted until the conflicts are resolved. This validation process improves the quality of the target production task sequence, providing more reliable task arrangements for production scheduling.

[0113] Step S145: Focus on key task nodes through an attention mechanism to enhance the temporal features of the target production task sequence, adjust the task execution order and resource allocation scheme, and generate a multi-objective production scheduling strategy.

[0114] In this embodiment, an attention mechanism is used to focus on critical task nodes to enhance the temporal characteristics of the target production task sequence. The attention mechanism automatically focuses on critical task nodes in the target production task sequence. Critical task nodes are those that have a significant impact on production scheduling, such as high-priority or resource-intensive nodes. By focusing on critical task nodes and enhancing their temporal characteristics, these critical task nodes are given more consideration in production scheduling. Based on the enhanced temporal characteristics, the task execution order and resource allocation scheme are adjusted. For example, if the temporal characteristics of a critical task node indicate that it needs to be executed earlier, the execution order of that task in the target production task sequence can be adjusted. By adjusting the task execution order and resource allocation scheme, a multi-objective production scheduling strategy is generated. This multi-objective production scheduling strategy comprehensively considers the optimization of multiple objectives, and can rationally arrange the execution order of production tasks and resource allocation while satisfying multiple objectives, thereby improving the efficiency and quality of production scheduling.

[0115] In a non-limiting embodiment, the method further includes: Step 210: Construct an abnormal test scenario library. The scenario library includes sub-scenarios such as equipment failure, raw material fluctuation, and process drift. Each sub-scenarios is described by a feature vector, which defines the probability of occurrence, the scope of impact, and the duration of the abnormality.

[0116] In this embodiment, an anomaly test scenario library is constructed. This library includes sub-scenarios such as equipment failure, raw material fluctuations, and process drift, representing potential anomalies in the production of waterproof membrane rolls. Each sub-scenarios is described by a feature vector, defining its probability of occurrence, scope of impact, and duration. The probability of occurrence reflects the likelihood of the anomaly, the scope of impact reflects its influence on the production process, and the duration reflects the length of time the anomaly lasts. For example, in the equipment failure sub-scenarios, the probability of occurrence can be determined based on the equipment's historical failure records and maintenance history; the scope of impact can include the impact of equipment failure on production speed and product quality; and the duration can be determined based on the equipment's repair time. By constructing this anomaly test scenario library, comprehensive anomaly testing of multi-objective production scheduling strategies can be performed, evaluating the strategy's performance under different anomaly conditions.

[0117] Step 220: Input the multi-objective production scheduling strategy into the scenario simulator, and reproduce the production process through the scenario simulator based on the discrete event simulation engine. Record the production interruption duration, quality fluctuation range and cost increase ratio under abnormal scenarios.

[0118] In this embodiment, a multi-objective production scheduling strategy is input into a scenario simulator. The scenario simulator reproduces the production process based on a discrete event simulation engine. The discrete event simulation engine can simulate various events in the production process, such as equipment startup, shutdown, task start-up, and end. During the production reproduction process, abnormal situations from the abnormal test scenario library are introduced, and the production interruption duration, quality fluctuation amplitude, and cost increase percentage under abnormal scenarios are recorded. The production interruption duration reflects the length of time production is stopped due to the abnormal situation, the quality fluctuation amplitude reflects the degree of impact of the abnormal situation on product quality, and the cost increase percentage reflects the proportion of production cost increase caused by the abnormal situation. For example, when equipment fails, the production interruption time, as well as the changes in product quality and the increase in production costs before and after the failure, are recorded. By recording this data, the performance of the multi-objective production scheduling strategy under abnormal scenarios can be evaluated.

[0119] Step 230: Use the anomaly impact assessment model to perform feature dimensionality reduction on the recorded data, extract key influencing factors, and construct an anomaly-strategy response correlation map.

[0120] In this embodiment, an anomaly impact assessment model is used to perform feature dimensionality reduction on the recorded data. The recorded data includes multiple features such as production interruption duration, quality fluctuation amplitude, and cost increase percentage. Feature dimensionality reduction can reduce the dimensionality of the data and remove redundant information. Key influencing factors are extracted using the anomaly impact assessment model. These key influencing factors are those that significantly impact the response to abnormal situations and multi-objective production scheduling strategies, such as equipment failure type and raw material fluctuation degree. An anomaly-strategy response correlation graph is constructed based on the key influencing factors. This graph records the relationship between abnormal situations and multi-objective production scheduling strategy responses. For example, the graph can show the relationship between a certain equipment failure and the corresponding production scheduling strategy adjustment. By constructing the anomaly-strategy response correlation graph, a more comprehensive understanding of the impact of abnormal situations on multi-objective production scheduling strategies can be achieved.

[0121] Step 240: Based on the association graph, identify vulnerable nodes of the scheduling strategy, use generative adversarial networks to generate enhanced test samples, and perform reinforcement testing on the vulnerable nodes.

[0122] In this embodiment, vulnerable nodes of the scheduling strategy are identified based on anomaly-policy response correlation graphs. Vulnerable nodes are those where multi-objective production scheduling strategies are prone to problems under abnormal conditions. For example, if a task is prone to production interruption due to equipment failure, then the node corresponding to that task is a vulnerable node. Generative adversarial networks (GANs) are used to generate enhanced test samples. A GAN consists of a generator and a discriminator. The generator can generate samples similar to real data, and the discriminator can determine the authenticity of the samples. Enhanced test samples generated by the GAN can simulate more abnormal situations, thus strengthening the testing of vulnerable nodes. For example, more severe equipment failures and larger fluctuations in raw material costs can be generated to test vulnerable nodes and evaluate the performance of the scheduling strategy under these extreme conditions. Through enhanced testing, the shortcomings of the scheduling strategy can be more comprehensively discovered.

[0123] Step 250: Generate strategy optimization suggestions, including redundant resource configuration schemes and dynamic scheduling rules, based on the test results. Feed the strategy optimization suggestions back to the multi-objective production scheduling model and iteratively update the scheduling strategy until it passes a preset number of abnormal scenario tests.

[0124] In this embodiment, strategy optimization suggestions are generated based on the results of reinforcement testing. These suggestions include redundant resource configuration schemes and dynamic scheduling rules. Redundant resource configuration schemes refer to adding backup resources during production to cope with abnormal situations, such as adding backup equipment or raw materials. Dynamic scheduling rules refer to how to dynamically adjust the production scheduling strategy under abnormal conditions, such as how to quickly reallocate tasks when equipment fails. The strategy optimization suggestions are fed back to the multi-objective production scheduling model, which iteratively updates the scheduling strategy based on these suggestions. The iterative update process continuously adjusts the scheduling strategy until it passes a preset number of abnormal scenario tests. This preset number of abnormal scenario tests ensures that the scheduling strategy performs well under various abnormal conditions, improving its robustness. For example, if the preset number of abnormal scenario tests requires the scheduling strategy to operate normally under 10 different abnormal conditions, the scheduling strategy is iteratively updated until this requirement is met. Through this method, the multi-objective production scheduling strategy can be continuously optimized, better ensuring production continuity under abnormal conditions.

[0125] As a non-limiting embodiment, the method further includes: Step 310: Collect multimodal real-time data during the production process of waterproof membrane. The multimodal real-time data includes equipment operating status data, raw material characteristic data, and environmental perception data.

[0126] In this embodiment, multimodal real-time data is collected during the production process of waterproof membrane. This multimodal real-time data includes equipment operating status data, raw material characteristic data, and environmental sensing data. Equipment operating status data can be collected through the production equipment monitoring system, including parameters such as equipment operating speed, temperature, and pressure. Raw material characteristic data can be collected through raw material testing equipment, including parameters such as the chemical composition and physical properties of the raw materials. Environmental sensing data can be collected through environmental sensors, including parameters such as the temperature, humidity, and light intensity of the production environment. For example, the production equipment monitoring system can record the equipment operating speed in real time, the raw material testing equipment can detect the chemical composition of the raw materials in real time, and the environmental sensors can monitor the temperature of the production environment in real time. By collecting multimodal real-time data, a comprehensive understanding of the status of the waterproof membrane production process can be achieved.

[0127] Step 320: Remove noise interference and perform feature standardization through data preprocessing.

[0128] In this embodiment, a data preprocessing process is performed on the collected multimodal real-time data. The data preprocessing process includes noise removal and feature standardization. Noise interference can affect the accuracy and reliability of the data; it is removed through methods such as filtering and smoothing. Feature standardization standardizes the feature values ​​of the data, ensuring that data with different features have the same dimensions and distribution range. For example, temperature and pressure parameters in equipment operating status data may have different dimensions and value ranges; feature standardization can transform them to the same range, facilitating model processing. This data preprocessing process improves the quality of the data.

[0129] Step 330: Construct a Transformer-based temporal feature extraction model, perform hierarchical feature encoding on the preprocessed multimodal real-time data, and generate a multi-scale temporal feature map containing long-term dependencies.

[0130] In this embodiment, a Transformer-based temporal feature extraction model is constructed. Transformer is a deep learning model based on an attention mechanism, capable of processing sequential data and capturing long-term dependencies within the data. Hierarchical feature encoding is performed on the preprocessed multimodal real-time data, allowing for multi-level extraction and representation of data features. The Transformer-based temporal feature extraction model generates a multi-scale temporal feature map containing long-term dependencies. This multi-scale temporal feature map includes data feature information at different time scales, such as short-term fluctuations and long-term trends. For example, for equipment operating status data in the waterproof membrane production process, the multi-scale temporal feature map can simultaneously include short-term fluctuations in equipment operating speed and long-term operating trends. By generating a multi-scale temporal feature map, a more comprehensive understanding of the temporal characteristics of the waterproof membrane production process can be achieved.

[0131] Step 340: Input the multi-scale temporal feature map into the anomaly detection model, identify potential production anomalies through the integration strategy of autoencoder reconstruction error and isolated forest algorithm, and output the anomaly feature vector and anomaly confidence.

[0132] In this embodiment, a multi-scale temporal feature map is input into the anomaly detection model. The anomaly detection model employs an integrated strategy of autoencoder reconstruction error and the isolated forest algorithm to identify potential production anomalies. An autoencoder is an unsupervised learning model that learns feature representations of data and detects anomalies by reconstructing the data. Reconstruction error refers to the difference between the data reconstructed by the autoencoder and the original data; a large reconstruction error may indicate the presence of anomalies. The isolated forest algorithm is a tree-based anomaly detection algorithm that identifies anomalous points in the data by constructing isolated trees. By integrating autoencoder reconstruction error and the isolated forest algorithm, the accuracy and reliability of anomaly detection can be improved. The anomaly detection model outputs anomaly feature vectors and anomaly confidence scores. The anomaly feature vectors record the characteristic information of the anomaly, such as the time of occurrence and the type of anomaly. The anomaly confidence score reflects the credibility of the anomaly detection results. For example, a high anomaly confidence score indicates that the detected anomaly is more likely to be a genuine anomaly. Through the anomaly detection model, potential anomalies in the waterproof membrane production process can be detected in a timely manner.

[0133] Step 350: Retrieve historical anomaly case database based on anomaly feature vectors, generate preliminary response plans through case reasoning, and simultaneously start reinforcement learning response model to optimize response plans with anomaly mitigation effect as reward signal.

[0134] In this embodiment, a historical anomaly case database is retrieved based on anomaly feature vectors. This database records past anomalies and corresponding solutions that occurred during the production of waterproof membrane rolls. Through case-based reasoning, similar anomalies are searched in the database based on their feature vectors, and preliminary solutions are generated. Case-based reasoning is a method for solving current problems based on past experience, drawing on successful experiences from historical cases to generate solutions. A reinforcement learning response model is simultaneously initiated, using the anomaly mitigation effect as a reward signal to optimize the solution. Reinforcement learning is a learning method that learns the optimal strategy through interaction with the environment and based on reward signals. In this embodiment, the reinforcement learning response model provides reward signals based on the anomaly mitigation effect, continuously adjusting the solution to better mitigate the anomaly. For example, if a solution effectively reduces the impact of anomalies on production, the reinforcement learning response model will provide a higher reward signal, encouraging the use of that solution. By combining case-based reasoning and the reinforcement learning response model, solutions can be quickly generated and optimized, improving the ability to handle anomalies.

[0135] Step 360: Transform the optimized response plan into production parameter adjustment instructions, and send them to the production execution system in real time via the industrial bus to dynamically correct the task execution parameters in the multi-objective production scheduling strategy, thereby ensuring production continuity under abnormal operating conditions.

[0136] In this embodiment, the optimized response plan is transformed into production parameter adjustment instructions. These instructions include adjusting equipment operating parameters and raw material supply parameters. These instructions are sent in real-time to the Production Execution System (MES) via an industrial bus, a type of bus used for communication between devices in industrial automation systems. The MES dynamically corrects the task execution parameters in the multi-objective production scheduling strategy based on the production parameter adjustment instructions. For example, if the abnormal situation is equipment failure, the MES can adjust the operating parameters of other equipment according to the optimized response plan to ensure production continuity. By dynamically correcting the task execution parameters in the multi-objective production scheduling strategy, production continuity can be guaranteed under abnormal operating conditions, reducing the impact of abnormal situations on production. For example, in the event of equipment failure, adjusting production parameters can allow production to continue to a certain extent, avoiding prolonged production interruptions.

[0137] As a non-limiting embodiment, the method further includes: Step 410: Construct a cross-scenario production scheduling knowledge graph. The graph nodes of the cross-scenario production scheduling knowledge graph include production process characteristics, equipment performance characteristics, and target optimization characteristics. The graph nodes are linked to multi-source scheduling experience data through entity linking technology.

[0138] In this embodiment, a cross-scenario production scheduling knowledge graph is constructed. The graph nodes of the cross-scenario production scheduling knowledge graph include production process characteristics, equipment performance characteristics, and target optimization characteristics. Production process characteristics include information such as the process flow and process parameters in the production of waterproof membranes. Equipment performance characteristics include information such as the production capacity and operational stability of the production equipment. Target optimization characteristics include information such as the type of objective function and weight allocation in multi-objective production scheduling. Graph nodes are associated with multi-source scheduling experience data through entity linking technology. Entity linking technology can associate graph nodes with entities in multi-source scheduling experience data; for example, it can associate production process characteristic nodes with historical production process data. By constructing a cross-scenario production scheduling knowledge graph, multi-source scheduling experience data can be integrated, providing rich knowledge support for cross-scenario production scheduling.

[0139] Step 420: Use a meta-learning framework to perform feature transfer processing on the cross-scenario production scheduling knowledge graph, transfer the scheduling strategy of mature production scenarios to new production scenarios, and minimize the feature distribution difference between the source domain and the target domain through metric learning methods.

[0140] In this embodiment, a meta-learning framework is used to perform feature transfer processing on a cross-scenario production scheduling knowledge graph. Meta-learning is a learning method that can quickly learn new knowledge and adapt to new environments. In this embodiment, the meta-learning framework is used to transfer scheduling strategies from mature production scenarios to new production scenarios. Mature production scenarios have rich production experience and optimized scheduling strategies, while new production scenarios may lack this experience. Through the meta-learning framework, useful features and knowledge can be extracted from mature production scenarios and applied to new production scenarios.

[0141] This study employs metric learning to minimize the feature distribution differences between the source domain (mature production scenario) and the target domain (newly commissioned production scenario). Metric learning is a method for learning distance metrics between data. By minimizing the feature distribution differences between the source and target domains, the scheduling strategy of the source domain can be better adapted to the characteristics of the target domain. For example, metric learning can identify similar production process features and equipment performance features in the source and target domains. This allows for the appropriate adjustment of scheduling strategies from mature production scenarios and their application in new production scenarios. Consequently, it reduces the time and cost of re-evaluating scheduling strategies in new production scenarios, thereby improving production efficiency.

[0142] Step 430: Train the domain-adaptive model, eliminate cross-scene feature shifts through adversarial learning, and generate a scheduling strategy prototype adapted to the target scene.

[0143] In this embodiment, a domain-adaptive model is trained to address the problem of cross-scene feature shift. Cross-scene feature shift refers to the fact that the same feature may have different distributions and meanings in different production scenarios. The domain-adaptive model eliminates this feature shift through adversarial learning. Adversarial learning consists of a generator and a discriminator. The generator attempts to generate data with a feature distribution similar to that of the target scene, while the discriminator attempts to distinguish between the generated data and the real target scene data. Through continuous adversarial training, the generator can learn the feature distribution of the target scene, thereby eliminating the feature shift between the source and target domains.

[0144] During training, data from both the source and target domains are simultaneously input into the domain adaptation model. The generator produces simulated target domain data based on the source domain data, while the discriminator distinguishes between the generated data and the real target domain data. Based on the discrimination result, the parameters of the generator and discriminator are adjusted so that the data generated by the generator increasingly approximates the real target domain data. After multiple iterations of training, the domain adaptation model learns the mapping relationship between the source and target domains, generating a prototype scheduling strategy adapted to the target scenario. This prototype strategy combines scheduling experience from the source domain with the characteristics of the target scenario.

[0145] Step 440: Construct a reinforcement learning exploration space based on the scheduling strategy prototype, and use the near-end policy optimization algorithm to adjust the task priority weights and resource allocation ratios to generate a scheduling strategy specific to the target scenario.

[0146] In this embodiment, a reinforcement learning exploration space is constructed based on a scheduling policy prototype. Reinforcement learning is a learning method that learns the optimal policy by interacting with the environment and based on reward signals. The scheduling policy prototype provides an initial policy framework for reinforcement learning, upon which an exploration space is constructed to explore different task priority weights and resource allocation ratios.

[0147] The proximal policy optimization algorithm is used to adjust task priority weights and resource allocation ratios. Proximal policy optimization is an algorithm used to optimize policy networks, which can quickly learn better policies while ensuring that policy updates are not too drastic. In reinforcement learning, the agent tries different task priority weights and resource allocation ratios in the exploration space and evaluates the effectiveness of these attempts based on environmental feedback (reward signals). Reward signals can be designed based on multiple objectives such as production efficiency, product quality, and cost. For example, if a certain task priority weight and resource allocation ratio can improve production efficiency while ensuring product quality and reducing costs, a higher reward signal is given.

[0148] Through continuous exploration and optimization, the near-end strategy optimization algorithm can find the optimal task priority weight and resource allocation ratio in the target scenario, thereby generating a target scenario-specific scheduling strategy. This specific scheduling strategy fully considers the characteristics of the target scenario and the requirements of multi-objective optimization, and can improve the production scheduling efficiency and quality in the target scenario.

[0149] Step 450: Deploy the strategy verification module, simulate the production process of the target scenario through the digital twin system, collect feature feedback data during the execution of the target scenario-specific scheduling strategy, and iteratively optimize the transfer learning parameters of the domain adaptive model to improve the robustness of the scheduling strategy in the new scenario.

[0150] In this embodiment, a strategy verification module is deployed to verify the effectiveness and robustness of the target scenario-specific scheduling strategy. A digital twin system is a technology that simulates the operation of a real system using a virtual model. In this embodiment, a digital twin system is used to simulate the production process of the target scenario. The target scenario-specific scheduling strategy is input into the digital twin system to simulate the operation of the production process, and feature feedback data is collected during the strategy execution process. This feature feedback data includes information on production efficiency, product quality, equipment status, etc., reflecting the execution effect of the scheduling strategy in actual production.

[0151] Based on the collected feature feedback data, the transfer learning parameters of the domain adaptive model are iteratively optimized. These parameters influence the application effectiveness of scheduling experience from the source domain in the target scenario. By continuously adjusting these parameters based on the feature feedback data, the domain adaptive model can better adapt to the characteristics of the target scenario. For example, if feature feedback data shows that the execution effect of the target scenario-specific scheduling strategy is poor in a certain production process, the transfer learning parameters can be adjusted to enable the domain adaptive model to better handle production scheduling in that process.

[0152] Through multiple iterative optimizations, the robustness of the scheduling strategy in new scenarios is continuously improved. Robustness refers to the ability of the scheduling strategy to maintain good performance when facing various uncertainties and abnormal situations. After verification and optimization by the strategy verification module, the target scenario-specific scheduling strategy can better cope with various challenges in new scenarios, improving the stability and reliability of production. For example, in the event of temporary equipment failures or fluctuations in raw material supply, the optimized scheduling strategy can quickly adjust the task execution order and resource allocation to ensure production continuity and product quality.

[0153] Please see Figure 2The figure is a schematic diagram of the basic structure of a multi-objective optimization production scheduling system 200 provided in an embodiment of this application. The multi-objective optimization production scheduling system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the multi-objective optimization production scheduling methods applied to the manufacturing of waterproof membranes.

[0154] Please see Figure 3 This application provides a functional block diagram of a multi-objective optimized production scheduling device, which includes: The information acquisition module is used to acquire the state constraint information of the waterproof membrane production line within a preset time period based on the preset multi-objective optimization labels. The state constraint information reflects the various restrictions that the production line needs to meet during the production process. The model building module is used to determine the objective function type and corresponding constraint set of the multi-objective production scheduling model based on the state constraint information, and to obtain the multi-objective production scheduling model by associating and mapping the objective function type with the constraint set. The model solving module is used to solve the multi-objective production scheduling model based on an improved particle swarm algorithm with adjusted particle encoding method and fitness function. By adjusting the search behavior of particles during the solving process, the optimal solution for multi-objective optimization is obtained. The strategy generation module is used to generate a multi-objective production scheduling strategy for the waterproof membrane production line based on the weight allocation relationship and priority ranking result of each objective in the optimal solution of the multi-objective optimization. The multi-objective production scheduling strategy is used to guide the actual production arrangement of the production line.

[0155] Please refer to the following: Figure 4 This application implements multi-objective optimization of waterproof membrane production scheduling. By obtaining state constraint information based on preset multi-objective optimization labels, scheduling inefficiencies caused by missing or inaccurate information are avoided. Based on the state constraint information, the objective function type and constraint set are determined, and a multi-objective production scheduling model is obtained through correlation mapping. This model closely reflects actual production and fully considers various factors and objectives in the production process. Compared to traditional single-objective models, it can more comprehensively balance the relationship between multiple objectives such as production efficiency, product quality, and production cost.

[0156] This application also employs an improved particle swarm optimization algorithm with adjusted particle encoding and fitness functions to solve the multi-objective production scheduling model. By optimizing the particle encoding method, the production scheduling scheme is represented more effectively. Simultaneously, the improved fitness function comprehensively considers the optimization of multiple objectives, flexibly adjusting the particle search behavior during the solution process. This allows the algorithm to adaptively switch between global exploration and local development, improving search efficiency and solution accuracy, and enabling it to quickly and accurately find the optimal solution for multi-objective optimization. Finally, a multi-objective production scheduling strategy is generated based on the optimal solution. This strategy fully considers the weight distribution and priority ranking results of each objective, providing scientific and reasonable guidance for the actual production arrangement of waterproof membrane production lines. It ensures that the production line achieves collaborative optimization of multiple objectives while meeting various constraints, thereby improving production efficiency and intelligence.

[0157] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0158] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the multi-objective optimized production scheduling system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the multi-objective optimized production scheduling system to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0159] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A production scheduling method for multi-objective optimization applied in waterproofing membrane manufacturing, characterized in that, The application relates to a multi-objective optimization method for a waterproof roll production line. According to preset multi-objective optimization labels, state constraint information of the waterproof roll production line in a preset time period is acquired, and the state constraint information reflects various limiting conditions that need to be met in the production process of the production line. Based on the state constraint information, a target function type of a multi-objective production scheduling model and a corresponding constraint condition set are determined, the target function type is associated with the constraint condition set for mapping processing, and the multi-objective production scheduling model is obtained. An improved particle swarm algorithm with adjusted particle coding mode and fitness function is used to solve the multi-objective production scheduling model, the search behavior of particles is adjusted in the solving process, and a multi-objective optimization optimal solution is obtained. According to the weight distribution relationship and priority order result of each target in the multi-objective optimization optimal solution, a multi-objective production scheduling strategy for the waterproof roll production line is generated, and the multi-objective production scheduling strategy is used to guide actual production arrangement of the production line.

2. The method of claim 1, wherein, The method comprises the following steps: The state constraint information is classified and processed to distinguish efficiency type constraint information affecting production efficiency, quality type constraint information affecting product quality and cost type constraint information affecting production cost. A first type target function type aiming at maximizing production rate is determined according to the efficiency type constraint information, a second type target function type aiming at minimizing product unqualified rate is determined according to the quality type constraint information, and a third type target function type aiming at reducing raw material consumption is determined according to the cost type constraint information. For each type of target function type, a corresponding constraint condition set corresponding to each type of target function type is formed by extracting a corresponding limit range and associated factors from the corresponding constraint information. An associated mapping rule between the target function type and the constraint condition set is constructed, different types of target functions are bound with corresponding constraint conditions, a binding relationship is integrated into a model structure by establishing a mathematical relationship expression, and the multi-objective production scheduling model is obtained.

3. The method of claim 1, wherein, The improved particle swarm algorithm with adjusted particle coding mode and fitness function is used to solve the multi-objective production scheduling model, the search behavior of particles is adjusted in the solving process, and the multi-objective optimization optimal solution is obtained. A hybrid coding mode based on process sequence and equipment allocation is used to code the particles, each particle is used to represent a production scheduling scheme, process sequence coding reflects the execution sequence of each process, and equipment allocation coding reflects the production equipment used by each process. A new fitness function is designed according to the target function type in the multi-objective production scheduling model, the new fitness function comprehensively considers the optimization of multiple targets, the implementation degree of each target is weighted and calculated, and the fitness value of the particle is obtained. The positions and speeds of the particle swarm are initialized, and during the iteration of the particle swarm, the inertia weight and the learning factor are adjusted according to the fitness value of the particle, so that the particle can adaptively switch between global exploration and local development during the search process; When the number of iterations of the particle swarm reaches the preset maximum number of iterations or the fitness value of the population converges after continuous iterations, the iteration operation is stopped, and the multi-objective optimization optimal solution is obtained by screening the particle swarm that completes the iteration.

4. The method of claim 3, wherein, The particle is encoded by using a hybrid encoding mode based on the process sequence and the device allocation, each particle is used to represent a production scheduling scheme, the process sequence encoding reflects the execution order of each process, and the device allocation encoding reflects the production device used by each process, and the device allocation encoding includes: For all processes of the waterproof coiled material production, the process execution sequence range is determined according to the production process requirement, and a unique serial number is allocated to each process in the process execution sequence range to represent the position of the process in the process sequence encoding; For each process, a list of production devices available for the process is counted, and a unique identification number is allocated to each device to represent the selected production device for the corresponding process in the device allocation encoding; The process sequence encoding and the device allocation encoding are combined in a preset order to form a complete particle encoding, the length of the process sequence encoding is equal to the total number of production processes, and the length of the device allocation encoding is also equal to the total number of production processes.

5. The method of claim 3, wherein, The new fitness function is designed according to the type of the objective function in the multi-objective production scheduling model, the new fitness function comprehensively considers the optimization of multiple objectives, and the fitness value of the particle is obtained by weighted calculation of the implementation degree of each objective, and the new fitness function includes: A multi-task learning framework is constructed, and the type of the objective function in the multi-objective production scheduling model is mapped to a parallel task branch of the framework, and each task branch corresponds to a feature extraction channel of a type of objective function; In each task branch, the associated features of the type of the objective function and the set of constraint conditions are dynamically captured through an attention mechanism to generate an objective associated feature vector, and the objective associated feature vector includes an objective implementation difficulty coefficient and a constraint sensitivity parameter; The particle population is processed in layers based on a non-dominated sorting mechanism, the crowding distance of each layer of particles is calculated, and a dynamic weight generation network is constructed based on the crowding distance and the objective associated feature vector; Real-time weight coefficients of each type of objective function are output by the dynamic weight generation network, and the real-time weight coefficients are adaptively adjusted with the change of the implementation degree of the objective in the particle search process; The deviation degree of the actual optimization value and the ideal optimization value of each type of objective function is weighted and fused by the real-time weight coefficient to generate a comprehensive fitness value, the actual optimization value and the ideal optimization value have the same dimension, and the deviation degree is calculated by the Euclidean distance in the feature space.

6. The method of claim 1, wherein, The multi-objective production scheduling strategy for the waterproof coiled material production line is generated according to the weight allocation relationship and the priority sorting result of each objective in the multi-objective optimization optimal solution, and the multi-objective production scheduling strategy includes: The multi-objective optimization optimal solution is subjected to feature analysis, a target weight distribution matrix and a priority order sequence are extracted, and a target decision directed graph is constructed, wherein nodes in the target decision directed graph represent production targets, and directed edges represent priority dependency relationships between targets; Based on the target decision directed graph, a graph neural network is used for node feature propagation to generate a graph embedding vector containing target coordination relationships, and potential target conflict pairs are identified through similarity calculation of the graph embedding vector; Conflict resolution processing is performed on the target conflict pairs, and a conflict resolution action sequence is generated through a reinforcement learning strategy network, which contains resource reallocation rules and task splitting strategies; The resolved target weight distribution matrix and priority order sequence are input into a production task scheduler, and a long short-term memory network is used to predict a task execution time window based on the production task scheduler to generate a target production task sequence; Through attention mechanism focusing on key task nodes, time series feature enhancement processing is performed on the target production task sequence to adjust the task execution order and resource allocation scheme, and a multi-objective production scheduling strategy is generated.

7. The method of claim 6, wherein, The target decision directed graph is initialized, and the node feature vector contains target weight values, priority coefficients, and resource demand vectors; A message passing layer of the graph neural network is constructed, and the propagation path between nodes is defined through an adjacency matrix. In the propagation process, the node's own features and neighbor node features are fused to generate an intermediate feature vector; The intermediate feature vector is subjected to graph attention aggregation processing, and the attention weight is dynamically allocated according to the correlation strength between nodes to strengthen the feature expression of key target nodes; The graph embedding vector of the node is generated through the output layer of the graph neural network, the cosine similarity of the graph embedding vectors of any two nodes is calculated, and a similarity threshold is set to screen potential target conflict pairs; The screened target conflict pairs are subjected to conflict type classification, and a conflict severity score is generated based on the conflict type, which is used to guide the priority of conflict resolution processing. The target conflict pairs are subjected to conflict resolution processing, and a conflict resolution action sequence is generated through a reinforcement learning strategy network, which contains resource reallocation rules and task splitting strategies, including:

8. The method of claim 6, wherein, The conflict feature vector of the target conflict pair and the current resource state vector are input into the reinforcement learning strategy network as inputs, and the conflict feature vector contains conflict type, severity score, and target correlation degree; In the exploration stage of reinforcement learning, a target selection strategy is used to select conflict resolution actions, and the action space contains basic operations such as resource transfer, task delay, and process adjustment; In the utilization stage of reinforcement learning, the selection probability of each action is output through the strategy network, and a conflict resolution action sequence is generated based on probability sampling, and the length of the conflict resolution action sequence is adjusted according to the conflict severity. ​ A reward function is constructed, the input of the reward function includes the target achievement rate after conflict resolution, resource utilization rate and task delay time, and the output is a comprehensive reward value; The policy network parameters are optimized through a deep deterministic policy gradient algorithm, so that the generated conflict resolution action sequence is used to maximize the comprehensive reward value, and the target conflict is resolved; The target weight allocation matrix after resolution and the priority sequence are input into a production task scheduler, the production task scheduler is used to predict a task execution time window based on a long short-term memory network, and a target production task sequence is generated, including: The target weight allocation matrix after resolution is subjected to matrix decomposition to obtain a task resource demand matrix and a device capacity matrix, and a task-device matching degree is calculated through matrix multiplication; The task-device matching degree and the priority sequence are input into an encoder of the long short-term memory network, the encoder is used to selectively remember key task features through a gating mechanism, and a task timing feature vector is generated; An earliest start time, a latest end time and a duration of each task are predicted based on the task timing feature vector through a decoder of the long short-term memory network, and a task execution time window is formed; Task allocation is performed according to the task execution time window and the device capacity matrix, a greedy algorithm is used to construct an original production task sequence, and the task allocation satisfies a device load constraint; The original production task sequence is checked, including resource conflict detection, time window overlap calculation and target achievement rate evaluation, and the sequence that passes the check is taken as the target production task sequence.

9. A multi-objective optimization production scheduling system, characterized by, It comprises: a processor, a storage device having a computer program stored thereon, and a network interface for providing network communication function; when the computer program is executed by the processor, the processor implements the production scheduling method for multi-objective optimization applied to waterproofing membrane manufacturing according to any one of claims 1-8.

10. A readable storage medium, characterized by, The readable storage medium stores programs or instructions, and the programs or instructions are executed by the processor to implement the production scheduling method for multi-objective optimization applied to waterproofing membrane manufacturing according to any one of claims 1-8.

Images