Multi-objective optimization deep drawing variable blank holder force prediction method and system

Abstract

The invention provides a multi-objective optimized deep stamping variable blank holder force prediction method and system, and belongs to the technical field of metal plate plastic forming manufacturing process optimizing.The method comprises the steps that a deep stamping analysis model of a to-be-predicted part is established, and process data are obtained through finite element simulation according to the established model; training a deep neural network model based on the obtained process data to obtain a trained deep neural network model; taking the trained deep neural network model as a fitness evaluator, and carrying out multi-objective optimization solution to obtain a Pareto optimal solution set of variable blank holder force parameters; and performing decision analysis on the optimal solution set, and selecting the solution with the maximum relative closeness as a final implementation scheme.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of optimization of metal sheet plastic forming manufacturing process, and particularly relates to a multi-objective optimization deep drawing variable blank holder force prediction method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art. The essence of deep drawing is to drive the flange area material of the flat plate blank to produce radial flow and tangential shrinkage, forming a hollow part with a depth much greater than the diameter. During the process, the material undergoes a large amount of tensile, bending and compression combined deformation, which puts a very high requirement on the forming limit of the material. In the stamping process, different areas of the sheet are in different stress states: the flange area is mainly subjected to radial tensile stress and tangential compressive stress, which is prone to wrinkling; the concave die fillet area is subjected to bending and stretching, which is prone to thinning and cracking; the wall area is subjected to unidirectional stretching; the bottom area is subjected to less force. This complexity makes the process control and defect prediction very difficult. Therefore, in the field of sheet deep drawing forming, the blank holder force is a crucial process parameter, and its control strategy directly affects the forming quality of the product and plays a decisive role in suppressing defects such as cracking and wrinkling. In the stamping process, the force on the blank holder is adjusted in real time and dynamically according to the stamping stroke or time, i.e. the variable blank holder force, which is a program control process based on pre-set rules.

[0003] Traditional blank holder force setting relies on the experience of engineers, simplified mathematical models or a large number of physical trial and error experiments. These methods have significant limitations: it is difficult to accurately describe the material nonlinearity, complex friction contact conditions and the highly coupled relationship between multiple process variables, resulting in conservative or non-optimal process parameters, causing product qualification rate fluctuations, low material utilization, and long research and development cycle and high cost.

[0004] To improve optimization efficiency, existing research attempts to introduce optimization algorithms and machine learning models. For example, standard genetic algorithms can be used for parameter search, and neural networks can be used to build process proxy models. However, traditional single-objective optimization algorithms have difficulty in obtaining a balanced set of Pareto optimal solutions at one time when dealing with multiple objectives such as minimizing cracking and minimizing wrinkling, which often require multiple single-objective optimizations. At the same time, the prediction accuracy and generalization ability of neural network models are heavily dependent on the size and quality of the training data.

[0005] In summary, the above schemes usually aggregate multiple objectives into a single objective, and cannot reveal the inherent trade-off relationship between the objectives, and cannot obtain concave solutions. For non-convex Pareto front, the weighted sum method cannot find the optimal solution on the concave surface. In addition, the optimization method is inefficient, requiring multiple runs and high computational cost. SUMMARY

[0006] To overcome the above deficiencies of the prior art, the present application provides a multi-objective optimization deep drawing variable blank holder force prediction method and system, which can efficiently and automatically predict and find a set of optimal variable blank holder force curves, achieving the best balance in multiple key quality indicators.

[0007] To achieve the above object, one or more embodiments of the present application provide the following technical solutions: In a first aspect, a multi-objective optimization deep drawing variable blank holder force prediction method is disclosed, comprising: A deep drawing analysis model of a part to be predicted is established, and process data is obtained through finite element simulation for the established model; A deep neural network model is trained based on the obtained process data to obtain a trained deep neural network model; the model is learned to replace finite element simulation, realizing rapid mapping from process parameters to forming quality indicators; The trained deep neural network model is used as a fitness evaluator to perform multi-objective optimization solving to obtain a Pareto optimal solution set of variable blank holder force parameters; Decision analysis is performed on the optimal solution set, and the solution with the largest relative closeness is selected as the final implementation scheme.

[0008] As a further technical solution, process data is obtained through finite element simulation for the established model, specifically comprising: In the deep drawing process, the sheet metal is placed on the concave die, then the blank holder is pressed down and clamps the outer edge of the sheet metal, and the material flow is accurately controlled through the drawing rib force and the segmented variable blank holder force; Then the convex die of the rigid body is started to move downward to draw the sheet metal into the concave die to form the product; After each simulation, the principal strain and secondary strain data of all elements on the sheet metal are extracted, and the strain state distribution is determined on the forming limit diagram; By analyzing the position of the element strain point relative to the safety area, the areas in different strain states of uniaxial tension, biaxial tension and pure shear are identified; On this basis, the evaluation value reflecting the risk of rupture and the evaluation value characterizing the risk of wrinkling in each simulation case are calculated, thereby realizing the calculation of the rupture objective function value and the wrinkling objective function value of each simulation case.

[0009] As a further technical solution, the variable blank holder force curve is parameterized as the force values FV1, FV2 and FV3 of the 3 key control points; The stroke of the variable blank holder force is divided into S1, S2 and S3 of the 3 key control points.

[0010] As a further technical solution, the deep neural network model comprises 1 input layer, 3 hidden layers and 1 output layer. The obtained process data is divided into a training set and a test set, a composite loss function combining MSE and MAE is used, an Adam optimizer is used for training, and after training, the correlation between the predicted value and the simulation value is verified.

[0011] The number of input nodes in the input layer is 6, including 3 pressure ring force size control point parameters and 3 pressure ring force stroke control point parameters. The number of output nodes in the output layer is 2, which are the rupture target function value and the wrinkling target function value respectively.

[0012] As a further technical solution, the trained deep neural network model is used as a fitness evaluator for the NSGA-II algorithm, and the optimization objectives are set as minimizing the rupture target function value and minimizing the wrinkling target function value. During the operation of the NSGA-II algorithm, the deep neural network model quickly predicts the multi-objective performance of each candidate solution, and the NSGA-II algorithm performs non-dominated sorting and congestion calculation according to the prediction results, guiding the population to evolve towards the Pareto optimal frontier, and obtaining the Pareto optimal solution set of the variable pressure ring force parameters.

[0013] As a further technical solution, after obtaining the Pareto optimal solution set, the TOPSIS method is used for final decision making: first, normalize the two target function values, and set the weight of the rupture risk and the weight of the wrinkling risk according to the actual engineering requirements; Calculate the Euclidean distance of each Pareto solution from the positive ideal solution and the negative ideal solution, and finally select the solution with the largest relative closeness as the optimal variable pressure ring force implementation scheme.

[0014] In the second aspect, an intelligent prediction and decision system for manufacturing process parameters based on deep learning and multi-objective optimization is disclosed, comprising: The process data acquisition module is configured to establish a deep drawing analysis model of the part to be predicted, and obtain process data through finite element simulation for the established model; The deep neural network model training module is configured to train the deep neural network model based on the obtained process data, and obtain the trained deep neural network model; The Pareto optimal solution set acquisition module is configured to use the trained deep neural network model as a fitness evaluator, perform multi-objective optimization, and obtain the Pareto optimal solution set of the variable pressure ring force parameters; The solution module is configured to perform decision analysis on the optimal solution set and select the solution with the largest relative closeness as the final implementation scheme.

[0015] The one or more technical solutions have the following beneficial effects: The technical solution of the application uses a deep neural network model to replace time-consuming finite element simulation for rapid performance prediction, uses the trained deep neural network model as a fitness evaluator, performs multi-objective optimization solving, obtains a Pareto optimal solution set of variable blank holder force parameters, has high global search capability, significantly shortens the process development cycle, reduces the research and development cost, and greatly improves the optimization efficiency and accuracy.

[0016] The technical solution of the application realizes the collaborative optimization and scientific decision of multiple objectives, can not only provide a series of process schemes that achieve the best balance among multiple quality indicators such as cracking and wrinkling, but also automatically determine a unique optimal solution through the TOPSIS method, and greatly improves the decision efficiency.

[0017] The technical solution of the application provides a flexible and adjustable decision mechanism, and through the weight setting in the TOPSIS method, the relative importance of each target can be flexibly adjusted according to the quality requirements of different products, and the applicability of the method is enhanced.

[0018] The technical solution of the application has self-learning and continuous optimization capability, and through feedback of optimization data to the DNN model, a closed-loop optimization system is formed, which can continuously adapt to new production conditions and quality requirements.

[0019] The advantages of the additional aspects of the application will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their description serve to explain the application and do not constitute an improper limitation of the application.

[0021] Figure 1 It is a deep drawing finite element analysis model used in the embodiment of the application, which shows the geometric configuration and boundary condition setting of key components such as sheet metal, blank holder, punch and die, and is the data source basis of the whole method.

[0022] Figure 2 It is a whole work flow and data interaction schematic diagram of the multi-objective optimization strategy described in the application, which specifically shows the complete technical route from data preparation, DNN training, NSGA-II optimization to TOPSIS decision, and is a panoramic view of the core method of the application.

[0023] Figure 3The principle diagram of the NSGA-II algorithm and the DNN deep neural network model collaborative optimization, which illustrates the population initialization, non-dominated sorting, crowding degree calculation and fitness evaluation process based on DNN prediction. The structure diagram of the DNN model shows the neuron configuration of the input layer, hidden layer and output layer of the network.

[0024] Figure 4 The workflow diagram of weight setting and relative closeness calculation in the TOPSIS multi-attribute decision-making method, which shows the decision-making process from the Pareto solution set normalization to the final solution selection.

[0025] Figure 5 The correlation verification diagram of the DNN model prediction value and the finite element simulation actual value, which intuitively shows the prediction accuracy of the model through the comparison of scatter distribution and fitting line.

[0026] Figure 6 The principle diagram of the rupture and wrinkling evaluation based on the forming limit diagram, which marks the boundaries of the safe zone, the rupture zone and the wrinkling zone, and the distribution of the typical strain path, is the key basis for the quality evaluation standard in this method. DETAILED DESCRIPTION

[0027] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0028] It should be noted that the terms used herein are only for the purpose of describing the specific embodiments, and are not intended to limit the exemplary embodiments according to the present application.

[0029] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0030] Based on the content of the background art, the prior art has not effectively integrated high-precision prediction models with advanced multi-objective optimization algorithms to form a systematic solution that can automatically and efficiently find the best variable blank holder force curve that meets multiple quality objectives. Therefore, under the premise of ensuring processing quality and efficiency, realizing multi-objective collaborative optimization of deep drawing process parameters is still a technical problem to be solved in the field.

[0031] Term explanation part: Deep drawing process: refers to the plastic forming process of sheet metal under the combined action of punch, die and blank holder to form deep cavity parts; Variable blank holder force: refers to the force exerted on the blank holder by the multi-axis servo motor changing with time or position according to a specific rule during deep drawing; Pareto optimal solution set: refers to the distribution of limited resources among conflicting objectives in multi-objective optimization problems, all possible optimal allocation schemes; TOPSIS method: is a multi-attribute decision-making method by approaching ideal solution ranking technology.

[0032] DNN: Deep Neural Network, deep neural network; NSGA-II: Non-dominated Sorting Genetic Algorithm II: second generation non-dominated sorting genetic algorithm; TOPSIS: Technique for Order Preference by Similarity to Ideal Solution: ideal solution ranking method.

[0033] Example one The embodiment discloses a kind of variable blank holder force prediction methods of multi-objective optimization deep drawing, with the development of certain circular cup deep drawing process as an example, by fusing DNN and NSGA-II, and the variable blank holder force prediction of multi-objective optimization deep drawing of ideal solution ranking method, this method can efficiently, automatically predict and find a group of optimal variable blank holder force curve, realize best balance on multiple key quality indicators, refer to Figure 2 , including: Step one: build DNN prediction model, obtain process data by finite element simulation and train network; Step two: the trained DNN model is used as fitness evaluator and integrated with NSGA-II algorithm, multi-objective optimization is solved, and the Pareto optimal solution set of variable blank holder force parameter is obtained; Step three: TOPSIS method is used to analyze the solution set, and the optimal solution is selected as the final process scheme, in this step, the specific physical quantity of optimal solution is the variable blank holder force corresponding to the predicted forming quality index, wherein forming quality index and variable blank holder force are read out as a combination by prediction.

[0034] In step one, a high-precision deep neural network prediction model is constructed: a large number of numerical simulations are performed on the deep drawing process by finite element simulation software, and the variable blank holder force curve is systematically changed as a representative complex process parameter. The strain distribution data of the sheet metal are extracted from the simulation results, and the fracture and wrinkling evaluation indexes based on the forming limit diagram are calculated accordingly. On the basis of this data set, a deep neural network model is constructed and trained, which takes the process parameters as the input and the predicted forming quality index as the output. In this example, the variable blank holder force is the independent variable, and the predicted forming quality index is the dependent variable. In this machine learning model, the process parameters (variable blank holder force) are input, and the predicted forming quality (fracture, wrinkling) index is output.

[0035] More specifically, first, in the data preparation stage, the Abaqus finite element software is used to establish a deep drawing analysis model of the circular cup part, as shown in Figure 1 During deep drawing, the sheet metal is placed on the die, then the blank holder is pressed down and clamps the outer edge of the sheet metal, and the material flow is precisely controlled by the drawing lug force and the segmented variable blank holder force. Finally, the convex die of the rigid body moves downward to draw the sheet metal into the concave die to form the product.

[0036] In the deep drawing simulation model of the circular cup part established by the Abaqus finite element software as shown in Figure 1 The variable blank holder force curve is parameterized as the force values FV1, FV2 and FV3 of the 3 key control points, and the stroke of the variable blank holder force is divided into S1, S2 and S3 of the 3 key control points. The Latin hypercube experimental design method is used to arrange 256 simulation schemes. By adjusting the pressure on the blank holder in real time, the resistance of the material flowing into the die is actively balanced, thereby suppressing the two major defects of wrinkling and fracture, and obtaining high-quality products. The wrinkling and fracture evaluation values after simulation are used as a set of data for prediction and optimization.

[0037] After each simulation, the main and secondary strain data of all elements on the sheet metal are extracted, and the strain state distribution is determined on the forming limit diagram (as shown in Figure 6 By analyzing the position of the element strain point relative to the safety area, the areas in different strain states such as uniaxial tension, biaxial tension and pure shear are identified. On this basis, the evaluation value reflecting the fracture risk and the evaluation value representing the degree of wrinkling in each simulation case are calculated, so as to realize the calculation of the fracture objective function value and the wrinkling objective function value of each simulation case.

[0038] The evaluation value is a comprehensive value of the two function values of wrinkling and fracture, and the weight values of the two are usually defined as a one-to-one relationship, that is, the vertical distance of the main and secondary strain points to the pure shear line and the fracture line is added.

[0039] When extracting the principal and secondary strain data, at any point inside the object, the strain state is usually a complex second-order tensor, containing normal strain and shear strain. The size and direction of the strain clearly reveal the limit of deformation and the active direction of deformation at that point. The specific steps are as follows: after completing the finite element calculation, enter the post-processing module of the software, find the strain-related options in the variable or result output list, and export the data.

[0040] Figure 6 As shown, the forming limit diagram is a curve obtained by fitting the measured data after a series of experimental tests (ISO 12004 or GB / T 15825.8, etc.). The forming limit diagram of each material is the same, and its specific data can be obtained from the standard database or the material supplier.

[0041] The strain state distribution describes the size and direction of deformation (elongation or compression) of each point inside the material after forming. The strain distribution after forming directly affects the strength and stiffness of the part. The strain state distribution, i.e., the position of the point composed of the principal strain (used to construct the y-axis coordinate of the point) and the secondary strain (used to construct the x-axis coordinate of the point) on the forming limit diagram, can be used to determine whether the finite element unit is in a state of rupture or wrinkling by judging the position of the point.

[0042] To analyze the strain distribution, the post-processing function of the finite element software is usually used for qualitative and quantitative analysis. The principal strain and secondary strain of each finite element unit are taken as a data point. All data points are plotted on the forming limit diagram (FLD). The data points located in the rupture zone, safe zone, and wrinkling zone are counted to quantitatively evaluate the overall safety margin of the product and determine whether the finite element is in a state of rupture or wrinkling, and the forming quality of the product is judged by the simulated finite element state.

[0043] When identifying, the position of the point composed of the principal strain (y-axis coordinate of the point) and the secondary strain (x-axis coordinate of the point) is analyzed. If the data point is located above the rupture line (gray zone), it indicates that there is a serious strain concentration locally, and the thickness direction strain (thinning) is very serious. If the data points are concentrated in the compression instability region (gray zone) on the left side of the pure shear line in the lower left corner of the FLD, it indicates that the material is unstable under compressive stress, and the thickness may increase and wrinkle.

[0044] When calculating the evaluation value reflecting the risk of rupture and the evaluation value representing the degree of wrinkling in each simulation case, the objective function value is calculated by the forming limit diagram, i.e., the strain state of each point of the sheet obtained by simulation or experiment is compared with the inherent forming limit curve of the material. The process is the vertical distance from the point composed of the principal and secondary strains of the finite element to the pure shear line and the rupture line, which is denoted as the risk value of possible wrinkling or rupture. The smaller the value, the smaller the risk.

[0045] By evaluating the state of each finite element, a precise bridge connecting the "simulation results" and "physical failure" is built in the virtual environment, realizing the quantitative and visual prediction of the stamping forming defects, thereby guiding the design optimization, and ultimately achieving the final purpose of reducing research and development costs, shortening development cycles, and ensuring and improving product quality.

[0046] Then, a deep neural network model (structure as shown in Figure 3 ) is built, which includes 1 input layer (input node number is 6, including 3 blank holder force size control point parameters and 3 blank holder force stroke control point parameters, the specific setting of these parameters is determined according to network public information and industry general experience data), 3 hidden layers and 1 output layer, and the output node number is 2, which is the rupture target function value and the wrinkling target function value respectively. 256 groups of data are divided into training set and test set according to the proportion of 8:2, a composite loss function combining mean square error MSE and mean absolute error MAE is used, and Adam optimizer is used for training, and the learning rate is set to 0.001. After training, the correlation verification of the predicted value and the simulation value is shown in Figure 5 , the determination coefficient R² of the model reaches 0.986 and 0.999, indicating that the prediction accuracy meets the engineering requirements. The training of the above deep neural network uses a composite loss function combining mean square error and mean absolute error to ensure the accuracy and stability of the model prediction.

[0047] In this embodiment, the expressions of mean square error and mean absolute error are as follows:

[0048] Where N represents the total number of samples, y j is the true value of the jth sample, and is the predicted value of the same sample by the model. The present application uses MSE and MAE as evaluation indexes of model performance, which are respectively used to quantitatively represent the square average level and the absolute average level of the error between the predicted value and the true value.

[0049] It should be noted that in the construction process of the deep neural network model, the process parameters including the variable blank holder force curve, 3 blank holder force size parameters and 3 blank holder force stroke parameters are taken as the model input, and the network structure is designed accordingly, so that the input layer node number matches the total number of process parameters, and the input node number is 6, including 3 blank holder force size parameters and 3 blank holder force stroke parameters, and the output layer node number is consistent with the number of forming quality indexes to be predicted.

[0050] In this embodiment, regarding step two, the integrated DNN and NSGA-II perform multi-objective optimization: the trained DNN model is used as a fast and accurate fitness evaluator, integrated with the second-generation non-dominated sorting genetic algorithm. Multiple optimization objectives are set, including minimizing the risk of cracking and minimizing the risk of wrinkling.

[0051] In this embodiment, based on the high-dimensional nonlinear mapping relationship learned by the training-converged deep neural network, potential new samples can be predicted through an inference process. The predicted results (objective function values, dependent variables) and their corresponding predicted inputs (process parameters, independent variables) form a combination. A large number of data combinations, set to 1000, can be read in the learned mapping. This data combination can be used as the input value of NSGA, which can improve the calculation accuracy of NSGA.

[0052] The core objective of the NSGA-II algorithm is to automatically explore and discover process parameter combinations that can achieve the best balance between multiple optimization objectives (cracking and wrinkling). The generation method is to randomly initialize a process parameter population, i.e., a large number of different parameter combinations, and simulate selection (prefer good parameter combinations), crossover (mix different parameter combinations), and mutation (randomly fine-tune parameters) operations in biological evolution to iteratively generate a new generation of candidate solutions. The specific steps include: the algorithm repeatedly performs "evaluate objective function (calculate the specific performance indicators of each parameter combination), non-dominated sorting (classify according to advantages and disadvantages), calculate crowding degree (measure the difference between solutions), elite preservation (preserve good solutions in each generation), and genetic operation (simulate gene recombination and mutation)" until convergence, and finally output the Pareto optimal solution set, i.e., a set of balanced solutions that cannot be optimized simultaneously. Through the above method, the parameter space can be automatically and efficiently explored, directly providing a set of uniformly distributed trade-off solutions, specifically a set of solutions that achieve the best balance between multiple objectives, providing a rich and scientific decision basis for multi-objective process optimization.

[0053] The NSGA-II algorithm generates a candidate solution population for the variable blank holder force parameter, uses the DNN model to quickly predict the multi-objective performance of each candidate solution, and performs non-dominated sorting and crowding degree calculation based on this to guide the population to evolve towards the true Pareto optimal frontier. The expression of non-dominated sorting is as follows:

[0054] where x and y are two different candidate solutions to be compared. k is the index of the objective function, and M is the total number of objective functions to be optimized. k (x) is the function value of solution x on the kth objective function (cracking objective, wrinkling objective).

[0055] The expression of crowding degree calculation is as follows:

[0056] where I[i] distance f is the crowding degree of individual i, k is the index of the objective function currently being processed, and f k (i) is the function value of individual i on the m-th objective function, f max It is the maximum value of the entire population on the k-th objective function, f min is the minimum value of the entire population on the m-th objective function, and L is the total number of individuals contained in the current non-dominated layer F.

[0057] Non-dominated sorting can automatically generate a set of Pareto optimal solutions in a single optimization process, without relying on prior weight settings. This overcomes the shortcomings of traditional weighted aggregation methods, which are highly subjective and can only obtain a single solution per run. By stratifying solutions based on their dominance relationships, it provides clear guidance for evolutionary selection, driving the population to converge efficiently to the true Pareto front. Crowding calculation, by quantifying the distribution density among individuals, prioritizes retaining solutions with large crowding distances, promoting a uniform distribution of the solution set in the target space and effectively maintaining the diversity and coverage of the solution set. Together, these two methods form a dual optimization mechanism of "sorting guiding convergence and crowding regulating distribution." While ensuring the convergence of the algorithm, it comprehensively explores the trade-offs between various objectives, achieving an effective balance between convergence performance and distribution uniformity, significantly improving the overall quality and decision value of the solution set.

[0058] During prediction, based on the high-dimensional nonlinear mapping relationship learned by the convergent deep neural network, potential new samples can be predicted through an inference process. The predicted result (objective function value, dependent variable) and its corresponding prediction input (process parameters, independent variables) are combined into a single data set.

[0059] For details on entering the multi-objective optimization phase, please refer to the following process: Figure 3 As shown, the trained DNN model is used as the fitness evaluator for the NSGA-II algorithm, with the optimization objectives set as minimizing the fracture objective function value and the wrinkling objective function value. The NSGA-II parameters are set as follows: population size 100, maximum number of iterations 2000, crossover probability 0.9, and mutation probability 0.1. During the algorithm's operation, the DNN model quickly predicts the multi-objective performance of each candidate solution, and NSGA-II performs non-dominated sorting and crowding calculation based on the prediction results, guiding the population to evolve towards the Pareto optimal front.

[0060] The initial population generation of the NSGA-II algorithm described above can be partially guided by the preliminary predictions of the DNN model to improve the optimization convergence efficiency.

[0061] In an embodiment, a model self-enhancing step is further included: the high-performance solution found in the NSGA-II optimization process is fed back to the DNN model for incremental learning, continuously improving the accuracy and generalization ability of the prediction model.

[0062] The high-performance solution specifically refers to the point formed by all the main and secondary strains that can be read out in the forming limit diagram, and the distance from the fracture line and the pure shear line in the coordinate axis of the forming limit diagram. The minimum value of the numerical addition of the distance values is the minimum risk of fracture and wrinkling, which is the high-performance solution.

[0063] In the present embodiment, regarding step three, the TOPSIS-based final solution, the Pareto optimal solution set obtained by NSGA-II optimization is subjected to multi-attribute decision analysis using the ideal solution approximation method. First, the objective function values are normalized, the weight coefficients of each objective are determined, then the distance between each Pareto solution and the positive ideal solution and the negative ideal solution is calculated, and finally all solutions are sorted according to the relative closeness, and the solution with the largest relative closeness is selected as the final variable pressure ring force implementation scheme.

[0064] It should be noted that the core idea of TOPSIS is to select a scheme that is closest to the ideal target and farthest from the negative ideal target. The specific content is: first, the objective function values are normalized, the weight coefficients of each objective are determined, then the distance between each Pareto solution and the positive ideal solution and the negative ideal solution is calculated, and finally all solutions are sorted according to the relative closeness, and the solution with the largest relative closeness is selected as the final variable pressure ring force implementation scheme. The calculation formula is as follows.

[0065] First, the objective function values are normalized, the weight coefficients of each objective are determined, then the distance between each Pareto solution and the positive ideal solution and the negative ideal solution is calculated, and finally all solutions are sorted according to the relative closeness.

[0066] The Pareto solution set has P candidate solutions, each with Q evaluation indicators:

[0067] Where P is the total number of candidate solutions, Q is the total number of evaluation indicators, d PQ is the original value of the pth candidate solution at the Qth indicator.

[0068] Data standardization is performed using vector normalization method:

[0069] Where z pq is the standardized indicator value, the standardized value of the pth candidate solution at the qth indicator.

[0070] According to the weight of each index w q Construct the weighted normalized matrix:

[0071] Where v pq is the weighted normalized value of the pth candidate solution on the qth index.

[0072] The expressions of the positive and negative ideal solutions are as follows:

[0073] Where V + is the positive ideal solution vector, V- is the negative ideal solution vector, v q + is the positive ideal value of the qth index, v q - is the negative ideal value of the qth index.

[0074] The expression of the Euclidean distance from the positive ideal solution to the negative ideal solution is as follows:

[0075] Where S p + is the Euclidean distance from the pth candidate solution to the positive ideal solution, S p - is the negative ideal value of the pth index.

[0076] The expression of the relative closeness degree calculation is as follows:

[0077] C p is the relative closeness degree of the pth candidate solution, the relative closeness degree takes the value range [0, 1], and the larger the value, the better the solution.

[0078] The final selection of the candidate solution with the largest relative closeness degree is:

[0079] is the index number of the optimal solution, and the finally selected solution is the candidate solution with the largest relative closeness degree Cp.

[0080] After obtaining the Pareto optimal solution set, the TOPSIS method is used for final decision-making, and the decision-making process is as follows: Figure 4The two objective function values are normalized first, and the weight of the cracking risk is set to 0.6 and the weight of the wrinkling risk is set to 0.4 according to the actual engineering requirements. The Euclidean distance of each Pareto solution to the positive ideal solution and the negative ideal solution is calculated, and finally the solution with the largest relative closeness is selected as the optimal variable blank holder force implementation scheme.

[0081] The weight coefficient of the above TOPSIS method can be adjusted according to actual engineering requirements to reflect the importance difference of different quality indicators.

[0082] To verify the implementation effect, the variable blank holder force scheme obtained by the optimization of the application is applied to the actual production line trial model, compared with the average value based on the database set by the traditional experience, the product cracking defect rate and wrinkling defect rate can be reduced at the same time. At the same time, by constructing a high-precision prediction model and using an intelligent optimization process, the method significantly reduces the dependence on traditional entity experiments which are time-consuming and high-cost, thereby greatly reducing the research and development cost.

[0083] The above calculation-intensive integrated DNN and NSGA-II method can be effectively implemented in engineering practice. Theoretically, the results obtained by the method are superior to or at least not inferior to the results obtained by the traditional experience setting method. The advantage is derived from the integration of the method: it combines the experience basis of traditional experimental design, the powerful mapping ability of deep neural network for complex nonlinear relationship, and the global search ability of NSGA-II algorithm in multi-objective space. Even in the ideal case, that is, the global optimal solution is accidentally found by the method based on traditional experimental design, the solution set obtained by the integrated DNN and NSGA-II must contain the solution. Therefore, under the guarantee of the systematic global search mechanism of NSGA-II, the final result is theoretically superior to or at least not inferior to the result obtained by the traditional experience setting method.

[0084] The method described in the application is not only suitable for the deep drawing process of automobile coverings, but also can be extended to the sheet forming process optimization in the fields of household appliances, aerospace, etc., and has a wide engineering application prospect.

[0085] Based on the problems of low efficiency, difficulty in considering multi-objective conflicts, and dependence on experience in the decision-making process in the prior art for deep drawing process parameter optimization, according to the characteristics and quality control requirements of the deep drawing process, the application realizes intelligent prediction and optimization of process parameters by integrating deep neural network and NSGA-II multi-objective optimization algorithm and combining with the TOPSIS decision-making method, which not only guarantees product quality but also significantly improves research and development efficiency, and provides a new technical route for deep drawing process optimization.

[0086] In the present embodiment, as shown in FIG. 1, the deep drawing process of an automobile covering is taken as an example. Figure 1As shown in the left, the input node number is 6, including 3 blank holder force size control point parameters and 3 blank holder force stroke control point parameters. The three segments are combined into an optimal variable blank holder force curve, and the variable blank holder force curve is parameterized as the force values FV1, FV2 and FV3 of 3 key control points; the stroke of the variable blank holder force is divided into S1, S2 and S3 of 3 key control points.

[0087] The present application can effectively predict the optimal variable blank holder force curve in the deep drawing process, and has excellent performance in suppressing defects such as cracking and wrinkling, thereby improving product quality and process efficiency, while greatly reducing research and development cost and cycle.

[0088] Example two The purpose of this embodiment is to provide a computer device, including a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0089] Example three The purpose of this embodiment is to provide a computer readable storage medium.

[0090] A computer readable storage medium has a computer program stored thereon, which is executed by a processor to perform the steps of the above method.

[0091] Example four The purpose of this embodiment is to provide a multi-objective optimization deep drawing variable blank holder force prediction system, which comprises: A process data acquisition module is configured to establish a deep drawing analysis model of a part to be predicted, and acquire process data through finite element simulation for the established model; A deep neural network model training module is configured to train a deep neural network model based on the acquired process data, and obtain a trained deep neural network model; A Pareto optimal solution set acquisition module is configured to use the trained deep neural network model as a fitness evaluator to perform multi-objective optimization solving, and obtain a Pareto optimal solution set of variable blank holder force parameters; A solving module is configured to perform decision analysis on the optimal solution set, and select the solution with the largest relative closeness as the final implementation scheme.

[0092] Example five The purpose of this embodiment is to provide a computer program product containing instructions, which when executed on a computer, causes the computer to perform the methods and functions involved in any of the above embodiments. The steps involved in the apparatus of the above embodiments correspond to the method embodiment one, and the specific implementation can refer to the relevant description part of embodiment one. The term "computer readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying a set of instructions for execution by a processor and causing the processor to perform any of the methods in the present application.

[0093] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively made into each integrated circuit module, or a plurality of modules or steps among them can be made into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0094] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for predicting the variable blank holder force of deep drawing with multi-objective optimization, characterized in that, The method comprises the following steps: Establishing a deep drawing analysis model of a part to be predicted, and obtaining process data through finite element simulation based on the established model; Training a deep neural network model based on the obtained process data to obtain a trained deep neural network model; Using the trained deep neural network model as a fitness evaluator to perform multi-objective optimization and solution to obtain a Pareto optimal solution set of variable blank holder force parameters; Performing decision analysis on the optimal solution set and selecting a solution with the largest relative closeness as a final implementation scheme.

2. A method of predicting the variable blank holder force of a multi-objective optimized deep drawing process as claimed in claim 1, wherein, The process data is obtained through finite element simulation based on the established model, and specifically comprises the following steps: During deep drawing, the sheet metal is placed on the concave die, and then the blank holder is pressed down to clamp the outer edge of the sheet metal, and the material flow is accurately controlled through the drawing force and the segmented variable blank holder force; Finally, the convex die of the rigid body moves downward to draw the sheet metal into the concave die to form a product; After each simulation, the principal strain data of all elements on the sheet metal are extracted, and the strain state distribution is determined on the forming limit diagram; By analyzing the position of the element strain point relative to the safety area, the areas in different strain states of uniaxial tension, biaxial tension and pure shear are identified; On this basis, the evaluation value reflecting the risk of rupture and the evaluation value representing the degree of wrinkling in each simulation case are calculated, so as to realize the calculation of the rupture objective function value and the wrinkling objective function value of each simulation case.

3. A method of predicting the variable blank holder force of a multi-objective optimized deep drawing process as claimed in claim 1, wherein, The variable blank holder force curve is parameterized as the force values FV1, FV2 and FV3 of the 3 key control points; The stroke of the variable blank holder force is divided into S1, S2 and S3 of the 3 key control points.

4. A method of predicting the variable blank holder force in deep drawing of multi-objective optimization as claimed in claim 1, wherein, The deep neural network model comprises 1 input layer, 3 hidden layers and 1 output layer; The obtained process data is divided into a training set and a test set, a composite loss function combining MSE and MAE is used, an Adam optimizer is used for training, and after the training is completed, the correlation between the predicted value and the simulation value is verified. The number of input nodes in the input layer is 6, including 3 blank holder force size control point parameters and 3 blank holder force stroke control point parameters; The number of output nodes in the output layer is 2, which are the rupture objective function value and the wrinkling objective function value respectively.

5. A method of predicting the variable blank holder force in deep drawing of multi-objective optimization as claimed in claim 1, wherein, The trained deep neural network model is used as the fitness evaluator of the NSGA-II algorithm, and the optimization objectives are set as minimizing the rupture objective function value and the wrinkling objective function value; The parameters of the NSGA-II algorithm are set; During the running process of the NSGA-II algorithm, the deep neural network model quickly predicts the multi-objective performance of each candidate solution, the NSGA-II algorithm performs non-dominated sorting and crowding calculation according to the prediction results, guides the population to evolve towards the Pareto optimal frontier, and obtains a Pareto optimal solution set of variable blank holder force parameters.

6. A method of predicting the variable blank holder force of a multi-objective optimized deep drawing process as claimed in claim 1, wherein After obtaining the Pareto optimal solution set, the TOPSIS method is used for final decision: firstly, the two objective function values are normalized, and the weights of the rupture risk and the wrinkling risk are set according to the engineering actual requirements; The Euclidean distances of each Pareto solution from the positive ideal solution and the negative ideal solution are calculated, and finally the solution with the largest relative closeness is selected as the optimal variable blank holder force implementation scheme.

7. A multi-objective optimization deep drawing variable blank holder force prediction system, characterized by, The method comprises the following steps: The process data acquisition module is configured to establish a deep drawing analysis model of the part to be predicted, and acquire process data through finite element simulation for the established model; The deep neural network model training module is configured to train a deep neural network model based on the acquired process data, and obtain a trained deep neural network model; The pareto optimal solution set acquisition module is configured to use the trained deep neural network model as a fitness evaluator to perform multi-objective optimization solving, and obtain a pareto optimal solution set of the variable blank holder force parameters; The solving module is configured to perform decision analysis on the optimal solution set, and select a solution with the largest relative closeness as a final implementation scheme.

8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method of any one of claims 1 to 6.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method of any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, performs the steps of the method of any one of claims 1 to 6.

Images