Casting production scheduling method and system based on digital twinning

Abstract

The invention relates to the technical field of casting production management, in particular to a casting production scheduling method and system based on digital twinning. According to the method, a dynamic disturbance event is judged according to a comprehensive quality feature vector, reinforcement learning adjustment is adopted for the disturbance event, and if no dynamic disturbance exists or disturbance is eliminated, a multi-objective optimization algorithm is directly adopted to accurately obtain a global optimal production scheduling method. According to the method, the global optimal production scheduling method is accurately obtained through the multi-objective optimization algorithm, dynamic disturbance in casting production scheduling is actively and flexibly responded through constraint condition judgment and reinforcement learning adjustment, the global optimal accuracy and robustness of the casting production scheduling method are further improved, and therefore the casting production efficiency is effectively improved, and the production cost is reduced.

Description

Technical Field

[0001] This application relates to the technical field of foundry production management, and in particular, to a foundry production scheduling method and system based on digital twin. Background Art

[0002] As the core link of production management, foundry production scheduling essentially constructs a time series plan and resource allocation plan for production tasks through multi-dimensional coupling analysis of customer orders, production capacity, and resource constraints. In the current industrial background of surging personalized customization demands and improved equipment intelligence levels, enterprises are facing higher production efficiency requirements and shorter production delivery cycles. Therefore, how to scientifically and reasonably arrange production plans and complete production requirements with quality and quantity guaranteed has become the key technical breakthrough for improving the core competitiveness of foundry enterprises.

[0003] Currently, some technical solutions for optimizing foundry production scheduling have been proposed. For example, Chinese Patent CN202411681338.8 discloses an intelligent production scheduling and dispatching system and method based on digital twin. This invention patent generates corresponding simulation logs by simulating the processes of different production scheduling methods, and then determines the optimal solution by comparing the simulation logs. However, this method is difficult to cope with sudden equipment failures and emergency order insertions in foundry production, and it is difficult to accurately obtain the optimal production scheduling method under dynamic disturbances.

[0004] Chinese Patent CN202411496857.7 discloses a production scheduling method, a production scheduling device, and an electronic device for foundry production. First, a multi-dimensional array of sand box arrays, site arrays, bottleneck process arrays, and material arrays is established, and then a genetic algorithm is used to integrate the production resources of the entire foundry process, and then each process is optimized. However, the production scheduling method obtained by this method is prone to falling into local optimal solutions and cannot accurately obtain the global optimal production scheduling method. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a foundry production scheduling method and system based on digital twin, which accurately obtains the global optimal production scheduling method through a multi-objective optimization algorithm, and at the same time actively and flexibly responds to dynamic disturbances in foundry production scheduling through constraint condition determination and reinforcement learning adjustment, further improving the accuracy and robustness of the global optimality of the foundry production scheduling method, thereby effectively improving foundry production efficiency and reducing production costs.

[0006] In the first aspect, the present invention provides a foundry production scheduling method based on digital twin, including the following steps: The foundry's parameter information is collected through the workshop's IoT devices, and the foundry is mapped to the digital twin layer based on the parameter information. The digital twin layer builds a three-dimensional visualization model to map the equipment status, process parameters, and resource data in real time, and reduces the dimension of the mapped parameter information. The reduced-dimensional comprehensive quality feature vector is imported into the production scheduling integrated database. The production scheduling behavior is dynamically disturbed according to the production scheduling integrated database. If there is a disturbance in the production scheduling behavior, reinforcement learning adjustment is initiated, and the production scheduling strategy is dynamically adjusted in real time according to the disturbance type. If there is no dynamic disturbance or the disturbance is eliminated, the multi-objective optimization algorithm is directly used to accurately obtain the global optimal scheduling method.

[0007] As a preferred technical solution of the present invention, the parameter information dimensionality reduction comprises the steps of: The parameter information includes equipment parameters, casting finished product parameters, order delivery cycle and material logistics information; Eliminate the data that exceeds the physical range or is invalid in the parameter information; Normalize the parameter information and use the random forest algorithm to retain the parameters whose importance scores are greater than the importance score threshold; Extract N1 groups of original space-time snapshots from the parameter information and calculate the covariance matrix; perform eigenvalue decomposition based on the covariance matrix to obtain the eigenvector matrix, select and retain the dominant modes of N2 groups whose energy proportion exceeds the threshold, and project the original space-time snapshots to the low-dimensional space spanned by the dominant modes, so that the residual paradigm accuracy error is less than the accuracy error threshold, and obtain the modal coefficient matrix after dimensionality reduction; Extract each modal coefficient by column from the modal coefficient matrix Y obtained by dimensionality reduction , get the original coefficient of the kth dominant mode, and calculate the weighted coefficient , superimpose the weighted modal coefficient matrix along the row direction to obtain the comprehensive quality characteristic vector; The comprehensive quality feature vector includes equipment parameter index, casting finished product parameter index, order delivery cycle index and material logistics information index; The projection formula is:

[0008] The covariance matrix is ​​expressed as: ; The weighted coefficient calculation formula is: ; Among them, X is the original space-time snapshot, is the dominant mode, Y is the modal coefficient matrix after dimensionality reduction, For the snapshots, is the mean field, C is the covariance matrix, is the modal weight.

[0009] As a preferred technical solution of the present invention, the dynamic disturbance determination of the scheduling behavior includes: Set constraint conditions: The constraint conditions include the range of equipment index during normal operation of the equipment, the range of expected parameter index of the cast product, the range of normal delivery cycle index of the order, and the range of material downstream state index; Determination of disturbance type: Determine equipment failure disturbance according to the equipment parameter index and the range of normal operation of the equipment; Determine the quality abnormality of the casting according to the casting product parameter index and the range of expected parameter index of the casting product; Determine emergency order insertion according to the order delivery cycle index and the range of normal delivery cycle index of the order; Determine order cancellation according to the material logistics information index and the range of material downstream state index.

[0010] As a preferred technical solution of the present invention, the reinforcement learning is used to dynamically adjust the scheduling strategy in real time according to the disturbance type, and includes the steps of: Set the state space according to the disturbance type; The state space includes equipment status, quality abnormality, order queue, and resource constraints; The equipment status includes equipment running / failure status; The quality abnormality includes casting composition deviation and inclusion content; The order queue includes order priority, delivery date, and process route; The resource constraints include die availability, raw material inventory, and energy supply; Set the adjustment strategy according to the state space; The adjustment strategy includes sequential adjustment, equipment switching, resource reallocation, and process parameter optimization; The sequential adjustment includes inserting / removing orders and rearranging priorities; The equipment switching includes migrating the tasks of the faulty equipment to the standby equipment; The resource reallocation includes die allocation, personnel shift adjustment, and energy dynamic scheduling; The process parameter optimization includes melting temperature adjustment and cooling rate adjustment; Set the action space: Output the adjustment strategy through the action space; Define the incentive function according to the state space and the action space to guide the scheduling system to learn the optimal scheduling strategy in the disturbance environment; The incentive function is used to quantify the penalty and reward mechanisms; The incentive function includes on-time delivery rate: , resource utilization rate: , quality qualification rate: , default cost: , equipment idle loss: , quality rework cost: ; Among them, represents the actual completion time, represents the planned delivery time, is the actual production capacity, is the theoretical maximum production capacity, is the number of qualified products, is the total output, is the order liquidated damages, is the delay compensation, is the equipment idle time, is the energy consumption cost per unit time, is the number of defective products, is the rework cost per piece.

[0011] As a preferred technical solution of the present invention, the reinforcement learning further includes: When the equipment fails, the state space adjusts the equipment state to equipment failure according to the disturbance type. The adjustment strategy provides a strategy for migrating the tasks of the failed equipment to the standby equipment and optimizing the process parameters according to the equipment failure type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The resource utilization rate encourages the restoration of production capacity after migration, and the equipment idle loss punishes the original equipment idle loss; When the quality is abnormal, the state space adjusts the quality abnormality to the casting composition deviation or inclusion content according to the disturbance type. The adjustment strategy provides a strategy for adjusting the melting temperature or cooling rate according to the quality abnormality type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The quality pass rate ensures the quality pass rate, and the quality rework cost suppresses the rework caused by excessive adjustment; When there is an emergency order insertion / order cancellation, the state space adjusts the order queue or resource constraints according to the disturbance type. The adjustment strategy provides a strategy for rearranging the priority or removing the order according to the emergency order insertion / order cancellation type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The on-time delivery rate encourages the timely completion of emergency orders, and the default cost restricts the liquidated damages cost of order cancellation.

[0012] As a preferred technical solution of the present invention, the multi-objective optimization algorithm is used to accurately generate a global optimal production scheduling method, including: Initializing the population: Using three-dimensional matrix encoding, the processes, equipment parameters, and time parameters after disturbance elimination or without disturbance intervention are randomly generated into chromosomes according to the three-dimensional matrix encoding to form the initial population of the algorithm. The initial population represents the initial casting production scheduling plan; Calculating the fitness: Calculating the fitness function for each chromosome; Selection: Selecting chromosomes from the initial population according to the fitness function for generating the next generation of chromosomes; Crossover: Generating new chromosomes by crossing the equipment and time sequences of the selected chromosomes; Mutation: Randomly adjusting the process, equipment, or time control for mutation to generate new chromosomes; Iteration: Repeat the selection, crossover, and mutation operations until the maximum number of iterations is reached or the fitness no longer improves significantly; Local optimum determination: If the fitness variance of 90% of the individuals in the population is less than 0.1, it is determined that the algorithm has fallen into a local optimum; After falling into the local optimum, enter the simulated annealing operation to expand the search space and finely search for a better solution within the local area; Output the global Pareto front solution set. According to the specific scenario requirements, including the shortest production cycle, the lowest equipment energy consumption, or the least material loss, optimize the equipment load distribution and obtain the global optimal solution; The time parameters include the time consumption and time constraints of each process on the equipment; The fitness function is used to quantify the quality of the chromosome and measure the comprehensive performance of the scheduling scheme.

[0013] As a preferred technical solution of the present invention, the simulated annealing operation includes: Select the top 20% of the individuals with the highest fitness in the current population for local optimization; Convert the fitness function into an energy function , and , is the maximum fitness of the current population; Set the initial temperature according to the energy fluctuation of the current population , where N is the number of iterations to stop, is the average energy, is the th individual's energy; Generate a neighborhood using Gaussian perturbation , where , is a normal distribution with a mean of 0 and a variance of , and are the neighborhood solutions before and after Gaussian perturbation processing, respectively, is the cooling rate, L is the number of annealing times at each temperature, is the initial step size; Accept the inferior solution according to the Metropolis criterion. For the generated neighborhood solution , calculate the energy change . When , directly accept to replace the current solution; when , accept with a probability of ; Set the energy drop amplitude threshold. If the energy drop amplitude is less than the energy drop amplitude threshold, accelerate the cooling , otherwise , where and are the temperature values after the k-th and (k + 1)-th iterations respectively; until the temperature drops to or reaches the preset maximum number of iteration times, and then jump out of the simulated annealing.

[0014] As a preferred technical solution of the present invention, the optimization of equipment load distribution includes: Generating an equipment load adjustment formula through weighted summation: ; where , and are all weight coefficients, and their value ranges are from 0 to 1, and , is the total production cycle, is the total energy consumption of the equipment, is the material cost loss.

[0015] In the second aspect, the present invention also provides a casting production scheduling system based on digital twin, which executes the casting production scheduling method as described above, including a physical entity layer, a digital twin layer, and a production scheduling optimization layer. The physical entity layer includes melting equipment, pouring equipment, molding equipment, and cooling equipment, which are used to execute metal casting production. The digital twin layer includes a three-dimensional visualization model, which is used to map the equipment status, process parameters, and resource data of the physical layer in real time. Among them, the production scheduling optimization layer includes a dynamic disturbance determination unit, a reinforcement learning adjustment unit, a multi-objective optimization unit, and a human-computer interaction interface, which are used to accurately generate a globally optimal production scheduling method and actively respond to dynamic disturbance events. The dynamic disturbance determination unit is used to determine the type of disturbance according to the constraint conditions. The reinforcement learning adjustment unit adjusts the production scheduling strategy in real time according to the determination result of the dynamic disturbance unit. The multi-objective optimization unit takes the parameters after disturbance elimination or without disturbance intervention as input, and uses a multi-objective optimization algorithm to accurately generate a globally optimal production scheduling method and optimize the equipment load distribution at the same time.

[0016] As a preferred technical solution of the present invention, the human-computer interaction interface includes a production scheduling Gantt chart and a multi-dimensional index dashboard. The production scheduling Gantt chart is used for process dragging adjustment and version management, and records the traces of reinforcement learning adjustment. The multi-dimensional index dashboard is used to display the equipment utilization rate, energy consumption curve, and dynamic disturbance in real time.

[0017] The beneficial effects of the present invention include: The present invention can accurately obtain the global optimal solution set of the production scheduling method through a multi-objective optimization algorithm, and satisfy the global optimal production scheduling method under specific requirements according to the optimized equipment load distribution method; at the same time, it combines constraint conditions and the reduced-dimensional comprehensive quality feature vector for dynamic disturbance determination, and uses reinforcement learning to adjust for dynamic disturbance events, continuously adjusting the production scheduling strategy through the incentive function to eliminate disturbance interference, further improving the accuracy and robustness of the global optimality of the casting production scheduling method, thereby effectively improving the casting production efficiency and reducing the production cost.

[0018] Aiming at the existing technology, the present invention can flexibly cope with four types of typical dynamic disturbances in casting production scheduling (equipment failure, quality abnormality, emergency order insertion, order cancellation), and specifically eliminate the interference of dynamic disturbances on the optimality of the production scheduling method; at the same time, the present invention can avoid falling into local optimality during the production scheduling process and accurately obtain the global optimal production scheduling method. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic diagram of a casting production scheduling method based on digital twin provided by an embodiment of the present invention; Figure 2 It is a schematic flow diagram of the reinforcement learning and multi-objective optimization algorithm in the present invention; Figure 3 It is a schematic structural diagram of a casting production scheduling system based on digital twin provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all of them. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts also belong to the scope of protection of the present application.

[0022] The following further describes the optimal embodiments of the present invention with reference to the drawings; Please refer to Figure 1 and 2 , this embodiment provides a casting production scheduling method based on digital twin, including the following steps: Collect the parameter information of the foundry workshop through the Internet of Things devices in the workshop. Map the foundry workshop to the digital twin layer based on the parameter information. The digital twin layer constructs a three-dimensional visualization model to map the equipment status, process parameters, and resource data in real time, and reduces the dimension of the mapped parameter information. Import the reduced comprehensive quality feature vector into the production scheduling integration database; Among them, the three-dimensional visualization model includes: Modeling of foundry equipment: Use SolidWorks to construct the three-dimensional geometric models of melting equipment, pouring equipment, cooling equipment, cleaning equipment, and inspection equipment. Construct the physical behavior models of the heating curve of the melting equipment and the vibration frequency of the molding equipment through MATLAB / Simulink, and embed the equipment operation parameters and status constraint parameters; The equipment operation parameters include rated power, maximum capacity, and maintenance cycle, and the status constraint parameters include mold replacement time and temperature threshold; Modeling of the process flow: Define the process logic chain of foundry pouring. Through a hybrid modeling method combining Petri nets and directed graphs, abstract each process node into a state machine containing event trigger rules and event dynamic parameters, which can be set as: the trigger condition for the completion of melting is that the temperature of the molten alloy is greater than or equal to and the component spectral analysis is qualified; the trigger condition for the start of pouring is that the pouring positioning accuracy is less than and the temperature is greater than or equal to ; the trigger condition for cooling to switch is that the surface temperature of the casting is less than or equal to (phase change point); Modeling of material resources: Record the mold specifications, remaining life, occupancy status, and the inventory and consumption rates of molding sand, molten alloy, and waste; Among them, the dimension reduction of parameter information includes the steps of: The parameter information includes equipment parameters, casting finished product parameters, order delivery cycle, and material logistics information; The equipment parameters include the liquid level temperature of the pouring ladle, the surface temperature of the mold, the gas pressure of the melting furnace, and the pressure of the cooling water system, etc. The casting finished product parameters include the composition deviation of the casting, the inclusion content, the shrinkage cavity volume ratio, and the tensile strength; the order delivery cycle includes the order entry time, the completion time of the first-piece inspection, the emergency order insertion buffer time, and the delivery date; the material logistics information includes the mold inventory parameters, the unit energy consumption of the melting process, and the energy supply consumption; Eliminate the data that exceeds the physical range or is invalid in the parameter information; Normalize the parameter information. Set several feature points in the random forest. Each feature point sets the feature score downward according to the importance level, and retain the parameters whose importance score is greater than the importance score threshold; Extract N1 groups of original spatio-temporal snapshots from the parameter information, calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix to obtain the eigenvector matrix, screen and retain N2 groups of dominant modes (N2 < N1) with energy proportion exceeding the threshold, project the original spatio-temporal snapshots into the low-dimensional space spanned by the dominant modes, and make the residual norm precision error less than the precision error threshold to obtain the reduced-dimensional modal coefficient matrix; Extract each modal coefficient by column from the reduced-dimensional modal coefficient matrix Y , obtain the original coefficient of the k-th dominant mode, and calculate the weighted coefficient , stack the weighted modal coefficient matrix along the row direction to obtain the comprehensive quality feature vector; The comprehensive quality feature vector includes the equipment parameter index, the casting finished product parameter index, the order delivery cycle index, and the material logistics information index; The projection formula is:

[0023] The covariance matrix is expressed as: ; The calculation formula for the weighted coefficient is: ; Among them, X is the original spatio-temporal snapshot, is the dominant mode, Y is the reduced-dimensional modal coefficient matrix, is the th snapshot, is the mean field, C is the covariance matrix, is the modal weight.

[0024] Perform dynamic disturbance determination on the scheduling behavior according to the scheduling integration database. If there is a disturbance in the scheduling behavior, start the reinforcement learning adjustment and dynamically adjust the scheduling strategy in real time according to the disturbance type; if there is no dynamic disturbance or the disturbance is eliminated, directly use the multi-objective optimization algorithm to accurately obtain the global optimal scheduling method; Among them, the dynamic disturbance determination of the scheduling behavior includes: Set constraint conditions: The constraint conditions include the interval range of the equipment index when the equipment is operating normally, the interval range of the expected parameter index of the casting finished product, the interval range of the order normal delivery cycle index, and the interval range of the material downstream state index; Disturbance type determination: Determine equipment failure disturbance according to the equipment parameter index and the interval range of the equipment parameter index when the equipment is operating normally; Determine that there is a quality abnormality in the casting according to the interval range of the casting finished product parameter index and the expected parameter index of the casting finished product; Determine emergency order insertion according to the order delivery cycle index and the interval range of the order normal delivery cycle index; Determine order cancellation according to the material logistics information index and the interval range of the material downstream state index; Among them, the reinforcement learning is used to dynamically adjust the production scheduling strategy in real time according to the disturbance type, including the steps of: Setting the state space according to the disturbance type; the state space includes equipment status, quality anomalies, order queue, and resource constraints; the equipment status includes equipment running / fault status; the quality anomalies include casting component deviation and inclusion content; the order queue includes order priority, delivery date, and process route; the resource constraints include die availability, raw material inventory, and energy supply; Setting the adjustment strategy according to the state space; the adjustment strategy includes sequential adjustment, equipment switching, resource reallocation, and process parameter optimization; the sequential adjustment includes inserting / removing orders and rearranging priorities; the equipment switching includes migrating the tasks of faulty equipment to standby equipment; the resource reallocation includes die allocation, personnel shift adjustment, and dynamic energy scheduling; the process parameter optimization includes melting temperature adjustment and cooling rate adjustment; Setting the action space: outputting the adjustment strategy through the action space; Defining an incentive function according to the state space and the action space to guide the production scheduling system to learn the optimal scheduling strategy in a disturbance environment; the incentive function is used to quantify the penalty and reward mechanisms; the incentive function includes on-time delivery rate: , resource utilization rate: , qualified product rate: , default cost: , equipment idle loss: , quality rework cost: ; Among them, represents the actual completion time, represents the planned delivery time, is the actual production capacity, is the theoretical maximum production capacity, is the number of qualified products, is the total output, is the order liquidated damages, is the delay compensation, is the equipment idle time, is the unit time energy consumption cost, is the number of defective products, is the single-piece rework cost.

[0025] Among them, the reinforcement learning further includes the steps of: When a device fails, the state space adjusts the device state to a device failure according to the type of disturbance. The adjustment strategy provides a strategy for migrating the tasks of the faulty device to a standby device and optimizing process parameters according to the type of device failure, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The resource utilization rate encourages the recovery of production capacity after migration, and the equipment idle loss punishes the original equipment idle loss; When quality anomalies occur, the state space adjusts the quality anomalies to casting composition deviations or inclusion contents according to the type of disturbance. The adjustment strategy provides a strategy for adjusting the melting temperature or cooling rate according to the type of quality anomaly, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The quality qualification rate ensures the quality qualification rate, and the quality rework cost suppresses rework caused by excessive adjustment; When an emergency order insertion / order cancellation occurs, the state space adjusts the order queue or resource constraints according to the type of disturbance. The adjustment strategy provides a strategy for rearranging priorities or removing orders according to the type of emergency order insertion / order cancellation, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function for calculation. The on-time delivery rate encourages the timely completion of emergency orders, and the default cost restricts the liquidated damages cost for order cancellation.

[0026] Among them, the multi-objective optimization algorithm is used to accurately generate a global optimal production scheduling method, including: Initializing the population: Encoding with a three-dimensional matrix, randomly generating chromosomes based on the three-dimensional matrix encoding of the processes, equipment parameters, and time parameters after disturbance elimination or without disturbance intervention, and forming the initial population of the algorithm. The initial population represents the initial casting production scheduling plan; Calculating the fitness: Calculating the fitness function for each chromosome; Selection: Selecting chromosomes from the initial population according to the fitness function for generating the next generation of chromosomes; Crossover: Generating new chromosomes by crossing the equipment and time sequences of the selected chromosomes; Mutation: Randomly adjusting the process, equipment, or time control for mutation to generate new chromosomes; Iteration: Repeating the selection, crossover, and mutation operations until the maximum iteration number is reached or the fitness no longer improves significantly; Local optimum determination: If the fitness variance of 90% of the individuals in the population is less than 0.1, it is determined that the algorithm has fallen into a local optimum; After falling into the local optimum, enter the simulated annealing operation to expand the search space and finely search for a better solution within the local domain; Output the global Pareto front solution set, optimize the equipment load distribution according to specific scenario requirements, including the shortest production cycle, the lowest equipment energy consumption, or the least material loss, and obtain the global optimal solution; The time parameters include the time consumption and time constraints of each process on the equipment; The fitness function is used to quantify the quality of chromosomes and measure the comprehensive performance of the production scheduling plan; The simulated annealing operation includes: Select the top 20% of the individuals with the highest fitness in the current population for local optimization; Convert the fitness function into an energy function , and , is the maximum fitness of the current population; Set the initial temperature according to the energy fluctuation of the current population , where N is the number of stopping iterations, is the average energy, is the energy of the i-th individual; Generate a neighborhood using Gaussian perturbation , where , is a normal distribution with a mean of 0 and a variance of , and are the neighborhood solutions before and after Gaussian perturbation processing respectively, is the cooling rate, L is the number of annealing times at each temperature, is the initial step size; Accept the inferior solution according to the Metropolis criterion. For the generated neighborhood solution , calculate the energy change . When , directly accept to replace the current solution; when , accept with a probability of ; Set the energy drop amplitude threshold. If the energy drop amplitude is less than the energy drop amplitude threshold, accelerate the cooling , otherwise , where and are the temperature values after the k-th and k + 1-th iterations respectively; Until the temperature drops to or reaches the preset maximum number of iterations, jump out of the simulated annealing.

[0027] The optimization of equipment load distribution includes: Generate an equipment load adjustment formula through weighted summation: ; Among them, , and are both weight coefficients, with a value range of 0 to 1, and , is the total production cycle, is the total energy consumption of the equipment, is the material cost loss.

[0028] Please refer to Figure 3 , this embodiment provides a casting scheduling system based on digital twin, which executes the casting scheduling method described above, including a physical entity layer, a digital twin layer, and a scheduling optimization layer. The physical entity layer includes melting equipment, pouring equipment, molding equipment, and cooling equipment, which are used to execute metal casting production. The digital twin layer includes a three-dimensional visualization model, which is used to map the equipment status, process parameters, and resource data of the physical layer in real time. The system includes a scheduling optimization layer, which is used to accurately generate a globally optimal scheduling method and actively respond to dynamic disturbance events. When the system executes, it realizes the steps of the method described above. Among them, the scheduling optimization layer includes a dynamic disturbance determination unit, a reinforcement learning adjustment unit, a multi-objective optimization unit, and a human-computer interaction interface, which are used to accurately generate a globally optimal scheduling method and actively respond to dynamic disturbance events. The dynamic disturbance determination unit is used to determine the type of disturbance according to the constraint conditions. The reinforcement learning adjustment unit adjusts the scheduling strategy in real time according to the determination result of the dynamic disturbance unit. The multi-objective optimization unit takes the parameters after disturbance elimination or without disturbance intervention as input, and uses a multi-objective optimization algorithm to accurately generate a globally optimal scheduling method and optimize the equipment load distribution at the same time. The human-computer interaction interface includes a scheduling Gantt chart and a multi-dimensional index dashboard. The scheduling Gantt chart is used for process dragging adjustment and version management, and records the traces of reinforcement learning adjustment. The multi-dimensional index dashboard is used to display the equipment utilization rate, energy consumption curve, and dynamic disturbance in real time.

[0029] The above is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A casting production scheduling method based on digital twin, characterized by: The following steps are involved: The foundry's parameter information is collected through the workshop's IoT devices, and the foundry is mapped to the digital twin layer based on the parameter information. The digital twin layer builds a three-dimensional visualization model to map the equipment status, process parameters, and resource data in real time, and reduces the dimension of the mapped parameter information. The reduced-dimensional comprehensive quality feature vector is imported into the production scheduling integrated database. Dynamically perturb the production scheduling behavior based on the production scheduling integrated database. If there is a disturbance in the production scheduling behavior, the reinforcement learning adjustment is initiated to dynamically adjust the production scheduling strategy in real time based on the disturbance type. If there is no dynamic disturbance or the disturbance is eliminated, the multi-objective optimization algorithm is directly used to accurately obtain the global optimal production scheduling method.

2. The casting production scheduling method based on digital twin according to claim 1, characterized in that: The parameter information dimension reduction includes: The parameter information includes equipment parameters, casting finished product parameters, order delivery cycle and material logistics information; Eliminate the data that exceeds the physical range or is invalid in the parameter information; Normalize the parameter information and use the random forest algorithm to retain the parameters whose importance scores are greater than the importance score threshold; Extract the original spatiotemporal snapshots of group N1 from the parameter information and calculate the covariance matrix; perform eigenvalue decomposition based on the covariance matrix to obtain the eigenvector matrix, and select and retain the dominant modes of group N2 whose energy proportion exceeds the threshold. , project the original space-time snapshot into the low-dimensional space spanned by the dominant mode, so that the residual norm accuracy error is less than the accuracy error threshold, and obtain the modal coefficient matrix after dimensionality reduction; Extract each modal coefficient by column from the modal coefficient matrix Y obtained by dimensionality reduction , get the original coefficient of the kth dominant mode, and calculate the weighted coefficient , superimpose the weighted modal coefficient matrix along the row direction to obtain the comprehensive quality characteristic vector; The comprehensive quality feature vector includes equipment parameter index, casting finished product parameter index, order delivery cycle index and material logistics information index; The projection formula is: ； The covariance matrix is ​​expressed as: ； The weighted coefficient calculation formula is: ； Among them, X is the original space-time snapshot, is the dominant mode, Y is the modal coefficient matrix obtained by dimensionality reduction, For the snapshots, is the mean field, C is the covariance matrix, is the modal weight.

3. The casting production scheduling method based on digital twin according to claim 1, characterized in that: The dynamic disturbance determination of the production scheduling behavior includes: Set constraints: Constraints include the range of equipment index when the equipment is operating normally, the range of expected parameters of finished castings, the range of normal order delivery cycle index, and the range of material downstream status index; Disturbance type determination: Determine equipment failure disturbance based on the equipment parameter index and the range of the equipment index when the equipment is operating normally; determine quality abnormalities in castings based on the range of the finished casting parameter index and the expected finished casting parameter index; determine emergency insertion orders based on the order delivery cycle index and the range of the normal order delivery cycle index; determine order cancellation based on the material logistics information index and the range of the material downstream status index.

4. The casting production scheduling method based on digital twin according to claim 1, characterized in that: The reinforcement learning is used to dynamically adjust the production scheduling strategy in real time according to the disturbance type, including the following steps: Set the state space according to the disturbance type; the state space includes equipment status, quality anomalies, order queues and resource constraints; the equipment status includes equipment operation / fault status; Quality anomalies include casting composition deviation and inclusion content; order queues include order priority, delivery time and process path; resource constraints include mold availability, raw material inventory and energy supply; According to the state space, set the adjustment strategy; Adjustment strategies include sequence adjustment, equipment switching, resource reallocation and process parameter optimization; sequence adjustment includes inserting / removing orders and re-prioritization; equipment switching includes migrating tasks from faulty equipment to spare equipment; resource reallocation includes mold transfer, personnel shift adjustment and dynamic energy scheduling; process parameter optimization includes melting temperature adjustment and cooling rate adjustment; Set the action space: Output the adjustment strategy through the action space; The incentive function is defined according to the state space and action space to guide the production scheduling system to learn the optimal scheduling strategy under the disturbance environment; the incentive function is used to quantify the penalty and reward mechanism; the incentive function includes the on-time delivery rate: , resource utilization: , quality pass rate: , default cost: , equipment idling loss: , quality rework cost: ; in, Indicates the actual completion time. Indicates the planned delivery time, is the actual production capacity, is the theoretical maximum capacity, is the number of qualified products, is the total output, For order liquidated damages, For delay compensation, The device idle time. is the energy cost per unit time, is the number of defective items, is the rework cost per piece.

5. The casting production scheduling method based on digital twin according to claim 4 is characterized in that: The reinforcement learning further includes the steps of: When equipment fails, the state space adjusts the equipment state to equipment failure according to the disturbance type. The adjustment strategy provides the task migration of the faulty equipment to the backup equipment and the process parameter optimization strategy according to the equipment failure type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function calculation. The resource utilization rate encourages the recovery of production capacity after migration, and the equipment idling loss punishes the idling loss of the original equipment. When the quality is abnormal, the state space will adjust the quality abnormality to casting composition deviation or inclusion content according to the disturbance type. The adjustment strategy provides melting temperature adjustment or cooling rate adjustment strategy according to the quality abnormality type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function calculation. The quality pass rate ensures the quality pass rate, and the quality rework cost suppresses rework caused by excessive adjustment; When an emergency order is inserted / an order is canceled, the state space adjusts the order queue or resource constraints according to the disturbance type. The adjustment strategy provides a re-prioritization or order removal strategy based on the emergency order / order cancellation type, and actively responds through the action space. The parameter information after the action space response is imported into the incentive function calculation. The on-time delivery rate encourages emergency orders to be completed on time, and the breach of contract cost constrains the penalty cost for order cancellation.

6. The casting production scheduling method based on digital twin according to claim 1, characterized in that: The multi-objective optimization algorithm is used to accurately generate the global optimal production scheduling method, including: Initialize the population: use three-dimensional matrix coding to randomly generate chromosomes based on the three-dimensional matrix coding of the process, equipment parameters and time parameters after the disturbance is eliminated or without disturbance intervention, to form the initial population of the algorithm, and the initial population represents the initial casting production schedule; calculate fitness: calculate the fitness function for each chromosome; Selection: Select chromosomes from the initial population according to the fitness function to generate the next generation of chromosomes; Crossover: Generate new chromosomes by crossing the devices and time series of the selected chromosomes; Mutation: Randomly adjust the process, equipment or time to control the variation and generate new chromosomes; Iteration: Repeat the selection, crossover and mutation operations until the maximum number of iterations is reached or the fitness is no longer significantly improved; Local optimum determination: If the fitness variance of 90% of the individuals in the population is less than 0.1, it is determined to be trapped in the local optimum; After falling into a local optimum, the simulated annealing operation is performed to expand the search space and search for a better solution in the local area; Output the global Pareto frontier solution set, optimize the equipment load distribution and obtain the global optimal solution according to the specific scenario requirements, including the shortest production cycle, the lowest equipment energy consumption or the least material loss; The time parameters include the time consumption and time constraints of each process on the equipment; The fitness function is used to quantify the quality of the chromosome and measure the comprehensive performance of the production scheduling plan.

7. The casting production scheduling method based on digital twin according to claim 6 is characterized in that: The simulated annealing operation comprises: Select the top 20% individuals with the highest individual fitness in the current population for local optimization; The fitness function Convert to energy function ,and , is the maximum fitness of the current population; Set the initial temperature according to the current population energy fluctuation , where N is the number of stop iterations, is the average energy, For the The energy of an individual; Use Gaussian perturbation to generate the neighborhood ,in , The mean is 0 and the variance is The normal distribution of and are the neighborhood solutions before and after Gaussian perturbation processing, is the cooling rate, L is the number of annealing times at each temperature, is the initial step length; The Metropolis criterion accepts inferior solutions, and for the generated neighborhood solutions , calculate the energy change ,when , directly accept Replace the current solution; when , with probability accept ; Set the energy drop threshold. If the energy drop is less than the energy drop threshold, the temperature will be lowered faster. ,otherwise ,in, are the temperature values ​​after the kth and k+1th iterations respectively; Until the temperature drops to Or the maximum number of iterations is reached and simulated annealing is exited.

8. The casting production scheduling method based on digital twin according to claim 6, characterized in that: The optimizing equipment load distribution comprises: The equipment load regulation formula is generated by weighted summation: ； in, , and are weight coefficients, ranging from 0 to 1, and , is the total production cycle, is the total energy consumption of the equipment, Loss of material cost.

9. A casting production scheduling system based on digital twin, executing the casting production scheduling method according to any one of claims 1 to 8, characterized in that: Including physical entity layer, digital twin layer and production scheduling optimization layer, The physical entity layer includes smelting equipment, pouring equipment, molding equipment and cooling equipment, which are used to perform metal casting production; The digital twin layer includes a three-dimensional visualization model for real-time mapping of physical layer equipment status, process parameters and resource data; The production scheduling optimization layer includes a dynamic disturbance determination unit, a reinforcement learning adjustment unit, a multi-objective optimization unit and a human-computer interaction interface, which are used to accurately generate the global optimal production scheduling method and actively respond to dynamic disturbance events; The dynamic disturbance determination unit is used to determine the disturbance type according to the constraint conditions; The reinforcement learning adjustment unit dynamically adjusts the production scheduling strategy in real time according to the determination result of the dynamic disturbance unit; The multi-objective optimization unit takes the parameters after the disturbance is eliminated or without disturbance intervention as input, and uses the multi-objective optimization algorithm to accurately generate the global optimal production scheduling method, while optimizing the equipment load distribution.

10. The casting production scheduling system based on digital twin according to claim 9, characterized in that: The human-computer interaction interface includes a production scheduling Gantt chart and a multi-dimensional indicator dashboard; The production scheduling Gantt chart is used for process drag-and-drop adjustment and version management, and records the traces of reinforcement learning adjustments; The multi-dimensional indicator dashboard is used to display equipment utilization, energy consumption curve and dynamic disturbance in real time.

Images