A stamping prediction method and automatic compensation device based on multiple machine learning models

Abstract

This invention relates to the field of computer science and technology, specifically to a stamping prediction method and automatic compensation device based on multiple machine learning models. First, the microstructure and material characteristics of the sheet metal are obtained using transmission electron microscopy and atomic probe tomography. Then, finite element simulation sampling is performed, and an adaptive algorithm is used to expand the dataset. Multiple models are constructed and trained to obtain the final springback prediction model. Finally, an automatic compensation device is embedded to achieve automatic compensation. The automatic compensation device includes modules for visual parameter input, embedded springback prediction, springback compensation calculation, and automatic control. It allows for intuitive parameter input, accurate springback prediction, calculation of compensation values, and adjustment of the process. The purpose of this invention is to solve the problem of insufficient accuracy and efficiency of traditional prediction and compensation technologies in addressing the deviation of part dimensions and shapes from design requirements caused by springback during the stamping process of high-strength sheet metal.

Description

Technical Field

[0001] This invention belongs to the field of computer science and technology, and relates to a stamping prediction method and automatic compensation device based on multiple machine learning models. Background Technology

[0002] Sheet metal stamping is a widely used metal forming method in industrial fields. It utilizes molds and stamping equipment to apply pressure to metal sheets, causing plastic deformation to obtain stamped parts with specific shapes, dimensions, and properties. This technology plays a crucial role in industries such as aerospace, automotive manufacturing, electrical machinery, and consumer goods production, and is an important means to achieve large-scale, efficient production of various metal parts. With the development of materials science, high-strength sheet metal is increasingly widely used in stamping due to its superior properties. However, in the high-strength sheet metal stamping process, springback has become a key factor restricting product quality and production efficiency. Springback refers to the elastic recovery deformation that occurs when a metal sheet undergoes plastic deformation after stamping, due to the release of internal stress when the external force is removed. This deformation can cause the dimensions and shape of the stamped parts to deviate from design requirements, seriously affecting product precision and quality.

[0003] In the production of high-strength sheet metal, springback presents numerous challenges. From a product quality perspective, springback makes it difficult to guarantee the dimensional accuracy of stamped parts, increasing the cost of subsequent processing and adjustments. For example, in automobile manufacturing, springback in body panels can lead to a decrease in assembly accuracy between parts, affecting the overall appearance and performance of the vehicle. In the aerospace field, where the precision requirements for parts are extremely high, springback can cause parts to fail to meet design standards and even affect flight safety. From a production efficiency perspective, to compensate for the errors caused by springback, multiple trial runs and mold adjustments are often required, which not only prolongs the product development cycle but also increases production costs. Moreover, the springback amount of high-strength sheet metal is relatively large, making it difficult to effectively control using traditional stamping processes, further exacerbating these problems. Different batches of high-strength sheet metal exhibit varying springback characteristics due to slight differences in material properties, making quality control during the production process even more difficult.

[0004] Currently, the main solutions to the springback problem of high-strength sheet metal during stamping are as follows:

[0005] Relying on the process experience of technicians, springback is compensated for by manually adjusting die design, such as changing the bending radius or die opening, or by stamping parameters such as pressure and speed. For example, in some small stamping plants, technicians rely on their long-accumulated experience to roughly estimate and adjust the springback of simple stamped parts. However, this method has obvious limitations. The experience of different technicians varies greatly, making it difficult to guarantee the accuracy and consistency of compensation. Moreover, for parts with complex shapes and new high-strength sheet materials, the experience-based method is often ineffective.

[0006] Physical models can be established to predict springback, such as simple mathematical models based on the principles of materials mechanics and plasticity to describe the deformation and springback behavior of sheet metal during stamping. However, due to the many complex factors involved in the stamping process, such as the anisotropy of materials, work hardening, and friction between the die and the sheet metal, physical models often cannot accurately simulate the actual situation, resulting in limited prediction accuracy.

[0007] Finite Element Analysis (FEA) technology is used to construct finite element models of sheet metal, dies, and the stamping process. These models, combined with material constitutive models, friction models, and boundary conditions, are then used with CAE software to predict springback and perform springback compensation calculations. Patent CN202410211644.9, entitled "A Method and Device for Springback Compensation of Stamped Parts Based on a Machine Learning Model," illustrates this method. Its principle involves obtaining springback data of the stamped part through finite element simulation and then performing reverse compensation on the die surface based on the springback amount. This method improves the accuracy of springback prediction to some extent, but it requires high precision in model parameter settings and calculations, and the calculation process is complex and time-consuming. Different CAE software, as well as varying software settings and user skill levels, can affect the reliability of the prediction results.

[0008] For example, patent number CN202411909498.3 describes a fast springback prediction method and device for complex components based on CNN-LSTM. By performing convolution operations on 3D model images using a CNN model, it is possible to effectively capture macroscopic structural features such as the overall size of the component, the distribution of ribs, and the radius of curvature. For example, in the case of high-ribbed panels, CNN can automatically identify the correlation between key geometric parameters such as rib height and thickness and springback. By using pixel matrix transformation and pooling layer dimensionality reduction, it significantly reduces data preprocessing time compared to traditional finite element mesh generation, making it suitable for quickly generating training datasets. The LSTM module can handle the sequential dependence of process parameters (such as aging temperature and time), for example, analyzing the stress release law at different aging stages, avoiding the limitation of traditional methods that treat process parameters as independent variables. However, it relies solely on the macroscopic geometric features of the three-dimensional model, completely ignoring the decisive influence of the material's microstructure, such as grain boundary distribution, dislocation density, and elemental segregation, on springback. Although macro- and micro-constitutive equations are introduced, they are not combined with microscopic image data, making it impossible to verify the accuracy of microscopic parameters through experimental data. Furthermore, it only processes "image-numerical" data through the CNN-LSTM cascade structure, without considering the different feature types, such as the differentiated weights of spatial features of images and numerical features of process parameters, which may weaken key information. It also does not propose an automatic compensation scheme that is linked with the stamping equipment. In practical applications, manual adjustment of mold parameters based on prediction results is still required, which inevitably leads to human error and extended trial molding cycles.

[0009] In summary, traditional solutions to the springback problem in high-strength sheet metal stamping generally suffer from drawbacks such as heavy reliance on process experience, low precision, and complex calculations. To effectively address practical production issues, a stamping prediction method based on multiple machine learning models and an automatic compensation device are employed. This method utilizes machine learning algorithms to process and analyze stamping data, uncovering potential patterns within the data, more accurately predicting springback, and achieving automatic compensation. This approach is expected to overcome the shortcomings of traditional methods and improve the quality and efficiency of stamping production. Summary of the Invention

[0010] This invention provides a stamping prediction method and an automatic compensation device based on multiple machine learning models, which solves the problem that traditional prediction and compensation technologies are not accurate and efficient enough because springback causes the size and shape of parts to deviate from the design requirements during the stamping process of high-strength sheet metal.

[0011] To solve the above problems, the technical solution adopted by the invention is as follows:

[0012] A stamping prediction method based on multiple machine learning models, the method comprising:

[0013] S01 Microstructure and Feature Acquisition: The microstructure of the sheet metal was obtained by using transmission electron microscopy combined with atomic probe tomography. The microstructure characteristics of elemental distribution, grain boundary features and dislocation density at the atomic scale were determined, and the characteristics of the material's elastic modulus and hardening index were obtained.

[0014] S02 Initial Simulation Sampling: The stamping parts are sampled using the finite element method, and the stamping springback simulation is performed using ABAQUS combined with the improved UMAT subroutine, and the springback vector data is recorded.

[0015] S03 Dataset Expansion: For various stamping materials, an adaptive parameter adjustment algorithm is used to traverse the initial process parameters, adjust the parameter step size according to material properties and previous simulation results, and obtain the stamping springback dataset through simulation sampling.

[0016] Model S04 Construction and Training: An improved ResNet-50 model was adopted, incorporating an attention mechanism to extract features from the microstructure map. The extracted features were integrated with material property parameters and process parameters. The dataset was divided into training, validation, and test sets. Transfer learning and ensemble learning strategies were used to train five models: Support Vector Regression, Kernel Ridge Regression, KNN, Random Forest Regression, and Gradient Boosting Tree. The root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination were used as the benchmarks. As evaluation indicators, the evaluation results are integrated into the final rebound prediction model using a dynamic weight allocation algorithm. This is used to measure how well a model fits the data; the closer the value is to 1, the better the model fit. (Calculation...) The formula is

[0017] in It is the actual value. It is a predicted value. It is the mean of the true values. n It is the sample size;

[0018] Mean Squared Error (MSE) measures the squared average of the differences between predicted and actual values. The formula is:

[0019]

[0020] in It is the actual value. It is a predicted value. n It is the sample size;

[0021] The Mean Absolute Percentage Error (MAPE) measures the relative error between the predicted and actual values, expressed as a percentage. The formula is:

[0022] According to each model The ratio is used to determine the weight of each model in the final prediction, and the models are integrated into a compensated prediction model. The input is material properties and process parameters, and the model outputs the rebound prediction value.

[0023] S05 Automatic Compensation: The final springback prediction model is embedded in the automatic compensation device. The device's visualization parameter input module receives process parameters and material property parameters input by the operator; the embedded springback prediction model predicts the springback vector; the springback compensation calculation module calculates the springback compensation value based on the prediction result and preset compensation rules; the automatic compensation control module generates a control signal based on the compensation value to adjust the process parameters of the stamping equipment and automatically compensate for the springback of the stamped parts.

[0024] An automatic compensation device for high-strength sheet metal stamping parts based on multiple machine learning models includes the following modules:

[0025] The visualization parameter input module is used to visually set process parameters and material property parameters for the automatic springback compensation device and input the parameters into the embedded springback prediction module.

[0026] The embedded springback prediction module generates a prediction model based on a method for predicting springback of high-strength sheet metal stamped parts using a multi-machine learning model. This model is then embedded in the device to predict the springback vector value during the stamping process, provided that process parameters and material property parameters are provided.

[0027] The springback compensation calculation module is used to calculate the springback compensation vector value;

[0028] The automatic control module is used to generate control signals and adjust the process.

[0029] The principle and advantages of this scheme are as follows:

[0030] Transmission electron microscopy and atomic probe tomography (APT) techniques were used to acquire the microstructural characteristics of sheet metal at the atomic scale, including elemental distribution, grain boundary features, and dislocation density, as well as material characteristics such as elastic modulus and hardening index. This revealed the influence of the material's essential properties on stamping performance at the microscopic level. Secondly, using the finite element method (FEM), ABAQUS combined with an improved UMAT subroutine was used to sample and simulate the springback of stamped parts, recording springback vector data to obtain basic simulation data. Next, an adaptive parameter adjustment algorithm was employed to adjust the parameter step size based on material properties and previous simulation results. Process parameters for various stamping materials were sampled extensively to expand the stamping springback dataset, providing rich data for model training. Then, the ResNet-50 model was improved and an attention mechanism was incorporated to extract deep features from the microstructure map. These features were integrated with material property parameters and process parameters. Transfer learning and ensemble learning strategies were used to train five models, including support vector regression and kernel ridge regression. A dynamic weight allocation algorithm was used, with root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination as evaluation metrics, to integrate the multiple models into a final springback prediction model. Finally, the prediction model is embedded into the automatic compensation device to predict springback based on the parameters input by the operator. Through the springback compensation calculation module and the automatic compensation control module, the process parameters of the stamping equipment are automatically adjusted to complete the automatic compensation of springback of the stamped parts.

[0031] Compared to existing technologies, traditional empirical methods and manual adjustments rely on personal experience, making it difficult to guarantee accuracy and resulting in significant dimensional deviations in parts. This proposed solution utilizes precise data obtained from microstructure and features, combined with complex model construction and training, to improve springback prediction accuracy. For example, in automotive parts manufacturing, the dimensional deviation of high-strength sheet metal stampings produced using traditional methods can reach ±1mm. With this solution, the dimensional deviation can be controlled within ±0.2mm, significantly improving product quality and reducing scrap rates. This method achieves data acquisition by combining finite element CAE simulation software with material microstructure analysis, and incorporates microstructure features to significantly enhance prediction performance. Five models are trained separately and then integrated into a unified model, which is then embedded into an automatic compensation device to achieve automated compensation based on material properties and process parameters.

[0032] Traditional finite element analysis, while utilizing numerical simulation, suffers from poor adaptability to variations in material properties and process parameters. This proposed solution employs an adaptive parameter adjustment algorithm and a multi-model integration strategy, enabling it to rapidly adapt to different stamping materials and process conditions. When producing high-strength sheet metal stampings for different automotive models, extensive readjustments are unnecessary; rapid switching and high-precision prediction and compensation are maintained, effectively shortening product development cycles and improving production flexibility.

[0033] Springback compensation often requires multiple manual interventions, resulting in low efficiency. This solution's automatic compensation device achieves a fully automated process from parameter input and springback prediction to process parameter adjustment. On a continuous stamping production line, it can monitor and compensate for springback in real time, reducing manual operation and downtime. Compared with traditional methods, production efficiency is improved, labor costs are reduced, and the stability and consistency of the production process are enhanced. Because this solution improves prediction accuracy and production efficiency, reduces scrap rate and manual intervention, it also reduces costs in mold development and production processes.

[0034] Compared with traditional numerical or physical models, this method and apparatus offer advantages in accuracy, adaptability, real-time performance, and cost control, particularly demonstrating superior performance when dealing with complex high-strength sheet metal and variable process conditions. Compared to single-model prediction methods, this approach is more flexible and robust, better able to handle the complexities and variability of the stamping process, reducing the risk of model errors and improving prediction accuracy.

[0035] By integrating the advantages of multiple models, this method can more comprehensively handle diverse data and feature relationships, demonstrating significant superiority in improving prediction accuracy, computational efficiency, and model optimization. Furthermore, by fully considering the impact of process parameters and material properties on springback, this method further enhances the accuracy and efficiency of the stamping process.

[0036] Furthermore, in step S01, a deep learning algorithm is used to perform secondary analysis on the data acquired by transmission electron microscopy and atomic probe tomography (APT). This automatically identifies and marks regions with unique microstructural features. Transmission electron microscopy and APT generate massive amounts of data, containing numerous microstructural information. Deep learning algorithms can quickly process this massive amount of data, automatically identifying regions with unique microstructural features, avoiding the tedious process of manually examining and analyzing data one by one, thus significantly saving time and labor costs. For example, when analyzing the microstructure of large sheet metal, manual screening might take days or even weeks, while a deep learning algorithm could complete the same task in a few hours. The real-time processing capability of deep learning algorithms can provide analysis results in a short time, allowing operators to adjust production parameters promptly based on the results. For instance, when the algorithm identifies a unique microstructural region in the sheet metal that affects springback, operators can immediately adjust the stamping process to avoid producing defective products. Simultaneously, in the subsequent construction and training of the springback prediction model, the accurately marked unique microstructural feature region data can serve as an important input parameter. These data can more accurately reflect the relationship between the microstructure of materials and their rebound performance, thereby improving the accuracy and reliability of rebound prediction models. For example, when training models such as support vector regression and kernel ridge regression, using data processed by deep learning algorithms can help the models better learn the mapping relationship between microstructural features and rebound, thus improving prediction accuracy.

[0037] Furthermore, in S02, the improved UMAT subroutine employs parallel computing optimization technology, utilizing the parallel computing capabilities of the graphics processing unit (GPU) to accelerate the stamping springback simulation process. By distributing the computational tasks in the UMAT subroutine to multiple cores of the GPU for parallel execution, the simulation time can be significantly shortened. For example, for a complex stamping part springback simulation, traditional CPU serial computation might take hours or even days, while GPU parallel computation might only take tens of minutes, greatly improving simulation efficiency. GPU-accelerated stamping springback simulation can provide accurate results in a shorter time, enabling technicians to quickly adjust the production process based on the simulation results and reduce production cycles. For large stamping parts or complex stamping processes, the amount of simulation computation increases dramatically. The parallel computing capabilities of the GPU can effectively handle these large-scale problems, avoiding model simplification or incomplete calculations due to insufficient computing resources. For example, in the stamping springback simulation of automotive body panels, the GPU can be used to perform accurate simulation analysis of the entire body panel without excessive simplification of the model.

[0038] Furthermore, in S04, the improved ResNet-50 model's attention mechanism employs a dynamic weight update strategy. This strategy adjusts the attention weights in real-time based on the characteristics of different microstructure images. Different microstructure images exhibit differences in element distribution, grain boundary features, and dislocation density, and these differences have varying degrees of impact on stencil springback prediction. The dynamic weight update strategy allows the attention mechanism to automatically focus on the regions and features most critical to springback prediction for each image. For example, in some images, the specific morphology of grain boundaries may be a key factor affecting springback. Dynamic weight adjustment allows the model to pay more attention to the features of grain boundary regions, thereby extracting this important information more accurately and providing strong support for subsequent springback prediction. Microstructure images may contain information that is not closely related to springback prediction. If the model gives the same attention to all regions and features, it will increase computational load and may introduce noise interference, reducing the model's accuracy. By dynamically adjusting the attention weights, the model can reduce its focus on this redundant information, allocating more computational resources to processing key features and improving the efficiency and quality of feature extraction.

[0039] Furthermore, the adaptive parameter adjustment algorithm in S03 dynamically adjusts the step size of process parameters based on historical simulation data and material properties. In the initial stage of stamping simulation, the range of process parameter values ​​is usually quite wide. If a fixed step size is used for parameter traversal, a significant amount of computational resources and time may be wasted in invalid parameter regions. The adaptive parameter adjustment algorithm, based on historical simulation data and material properties, can quickly identify parameter ranges that may produce effective results, and appropriately reduce the step size for fine searching within these ranges, while increasing the step size to quickly skip invalid regions, thereby greatly shortening the simulation time.

[0040] Furthermore, the improved ResNet-50 model in S04 incorporates an attention mechanism, which can automatically focus on regions in the microstructure map. This attention mechanism automatically identifies regions closely related to stencil springback, such as specific grain boundaries and dislocation-dense regions, and assigns higher attention weights to these key areas. In this way, the model can focus more on key information during feature extraction, avoiding interference from irrelevant or secondary information, thus more accurately capturing the key features affecting stencil springback. The attention mechanism can keenly capture these subtle features and give them sufficient attention. For example, differences in elemental distribution at certain atomic scales may be able to be extracted by the model through the focus of the attention mechanism, providing richer and more accurate information for subsequent predictions.

[0041] Furthermore, the transfer learning strategy in S04 initializes five machine learning models using model parameters pre-trained in the relevant domain, which accelerates model convergence and reduces training time.

[0042] Furthermore, the dynamic weight allocation algorithm dynamically adjusts the weights of each model in the integrated model based on the evaluation indicators of the models under different working conditions. The working conditions during the stamping process are complex and variable; different material properties, process parameters, and environmental conditions all affect springback. A single model can hardly maintain good performance under all working conditions. The dynamic weight allocation algorithm can dynamically adjust the weights based on real-time working conditions and the evaluation results of each model, enabling the integrated model to better adapt to various complex working conditions and provide more accurate springback predictions. The distribution of stamping data may change with time, production batches, and other factors. The dynamic weight allocation algorithm can monitor the performance of each model on new data in real time and dynamically adjust the weights according to the evaluation indicators, enabling the integrated model to quickly adapt to changes in data distribution and maintain stable predictive performance. At the same time, the performance of a single model may fluctuate due to factors such as data noise and outliers. Through the dynamic weight allocation algorithm, the integrated model can integrate the prediction results of multiple models and dynamically adjust the weights according to the evaluation indicators of each model, reducing the impact of fluctuations in a single model on the overall prediction results and improving the stability of the model. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models, according to an embodiment of the present invention.

[0044] Figure 2 This is a flowchart of the dataset acquisition, processing, and model training process for the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models, as described in this embodiment of the invention.

[0045] Figure 3 This is a flowchart of the KNN execution process for the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models, according to an embodiment of the present invention.

[0046] Figure 4 This is a flowchart illustrating the RFE (Recovery Execution) process of the high-strength sheet metal stamping part springback prediction method based on multiple machine learning models according to an embodiment of the present invention.

[0047] Figure 5 This is an SVM execution flowchart of the high-strength sheet metal stamping part springback prediction method based on multiple machine learning models according to an embodiment of the present invention.

[0048] Figure 6 This is a flowchart of the kernel ridge regression execution process for predicting the springback of high-strength sheet metal stamped parts based on multiple machine learning models, according to an embodiment of the present invention.

[0049] Figure 7 This is a flowchart of the boosting tree execution of the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models according to an embodiment of the present invention.

[0050] Figure 8 This is a diagram showing the final model structure of the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models according to an embodiment of the present invention.

[0051] Figure 9 This is a diagram of the restNet-50 network structure of the springback prediction method for high-strength sheet metal stamped parts based on multiple machine learning models according to an embodiment of the present invention.

[0052] Figure 10 This is a schematic diagram of the automatic compensation device for high-strength sheet metal stamping parts based on multiple machine learning models, according to an embodiment of the present invention. Detailed Implementation

[0053] Example 1, such as Figure 1-10 As shown, an automatic compensation device for high-strength sheet metal stamping parts based on multiple machine learning models includes the following modules:

[0054] The visualization parameter input module is used to visually set process parameters and material property parameters for the automatic springback compensation device and input the parameters into the embedded springback prediction module.

[0055] The embedded springback prediction module generates a prediction model based on a method for predicting springback of high-strength sheet metal stamped parts using a multi-machine learning model. This model is then embedded in the device to predict the springback vector value during the stamping process, provided that process parameters and material property parameters are provided.

[0056] The springback compensation calculation module is used to calculate the springback compensation vector value;

[0057] The automatic control module is used to generate control signals and adjust the process.

[0058] A stamping prediction method based on multiple machine learning models, the method comprising:

[0059] S01 Microstructure and Feature Acquisition: The microstructure of the sheet metal was obtained by using transmission electron microscopy combined with atomic probe tomography. The microstructure characteristics of elemental distribution, grain boundary features and dislocation density at the atomic scale were determined, and the characteristics of the material's elastic modulus and hardening index were obtained.

[0060] S02 Initial Simulation Sampling: The stamping parts are sampled using the finite element method, and the stamping springback simulation is performed using ABAQUS combined with the improved UMAT subroutine, and the springback vector data is recorded.

[0061] S03 Dataset Expansion: For various stamping materials, an adaptive parameter adjustment algorithm is used to traverse the initial process parameters, adjust the parameter step size according to material properties and previous simulation results, and obtain the stamping springback dataset through simulation sampling.

[0062] Model S04 Construction and Training: An improved ResNet-50 model was adopted, incorporating an attention mechanism for feature extraction from the microstructure map. The extracted features were integrated with material property parameters and process parameters. The dataset was divided into training, validation, and test sets. Transfer learning and ensemble learning strategies were used to train five models: Support Vector Regression, Kernel Ridge Regression, KNN, Random Forest Regression, and Gradient Boosting Tree. The root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination were used as the benchmarks. As evaluation indicators, the evaluation results are integrated into the final rebound prediction model using a dynamic weight allocation algorithm. This is used to measure how well a model fits the data; the closer the value is to 1, the better the model fit. (Calculation...) The formula is

[0063] in It is the actual value. It is a predicted value. It is the mean of the true values. n It is the sample size;

[0064] Mean Squared Error (MSE) measures the squared average of the differences between predicted and actual values. The formula is:

[0065]

[0066] in It is the actual value. It is a predicted value. n It is the sample size;

[0067] The Mean Absolute Percentage Error (MAPE) measures the relative error between the predicted and actual values, expressed as a percentage. The formula is:

[0068] According to each model The ratio is used to determine the weight of each model in the final prediction, and the models are integrated into a compensated prediction model. The input is material properties and process parameters, and the model outputs the rebound prediction value.

[0069] S05 Automatic Compensation: The final springback prediction model is embedded in the automatic compensation device. The device's visualization parameter input module receives process parameters and material property parameters input by the operator; the embedded springback prediction model predicts the springback vector; the springback compensation calculation module calculates the springback compensation value based on the prediction result and preset compensation rules; the automatic compensation control module generates a control signal based on the compensation value to adjust the process parameters of the stamping equipment and automatically compensate for the springback of the stamped parts.

[0070] The principle of this scheme is based on the deep integration of multidisciplinary technologies. In the stage of obtaining microstructure and features, transmission electron microscopy combined with atomic probe tomography is used to deeply analyze the microstructure of the sheet material at the atomic scale, accurately measure key properties such as elemental distribution, grain boundary features and dislocation density, and obtain the elastic modulus and hardening index of the material. This provides solid microscopic basic data for subsequent analysis and reveals the influence mechanism of the material's intrinsic properties on springback.

[0071] In the initial simulation sampling phase, the stamped parts were sampled using the finite element method, and stamping springback simulation was performed using ABAQUS software combined with an improved UMAT subroutine. This approach incorporates microstructural characteristics into the simulation process, more accurately simulating the actual deformation of the sheet metal during stamping, thereby recording reliable springback vector data and providing realistic and effective samples for subsequent model training.

[0072] In the dataset expansion phase, an adaptive parameter adjustment algorithm was used to iterate through the initial process parameters for various stamping materials. This algorithm dynamically adjusts the parameter step size based on material properties and previous simulation results, intelligently acquiring stamping springback data under different combinations of process parameters. This effectively expands the diversity and representativeness of the dataset, ensuring that subsequent models can learn rich stamping springback characteristics.

[0073] During model construction and training, an improved ResNet-50 model incorporating an attention mechanism was used to extract features from the microstructure map, highlighting key microstructural features. The extracted features were integrated with material property parameters and process parameters. After dividing the dataset, transfer learning and ensemble learning strategies were used to train five models, including support vector regression and kernel ridge regression. Using root mean square error, mean absolute percentage error, and coefficient of determination as evaluation metrics, a dynamic weight allocation algorithm was employed to integrate the five models into a final rebound prediction model, fully leveraging the strengths of each model to improve the accuracy and reliability of the prediction.

[0074] Finally, in the automatic compensation stage, the final springback prediction model is embedded in the automatic compensation device. The parameters input by the operator are obtained through the visual parameter input module. The embedded springback prediction model predicts the springback vector. The springback compensation calculation module calculates the compensation value according to the preset rules. The automatic compensation control module adjusts the process parameters of the stamping equipment according to the compensation value to achieve real-time and accurate compensation for the springback of the stamped parts.

[0075] Compared to existing technologies, traditional empirical methods and manual adjustments rely on personal experience, making it difficult to guarantee accuracy and resulting in significant dimensional deviations in parts. This solution, however, utilizes precise data obtained from microstructure and features, combined with complex model construction and training, to improve springback prediction accuracy. For example, in automotive parts manufacturing, the dimensional deviation of high-strength sheet metal stampings produced using traditional methods can reach ±1mm. With this solution, the dimensional deviation can be controlled within ±0.2mm, significantly improving product quality and reducing scrap rates.

[0076] Traditional finite element analysis, while utilizing numerical simulation, suffers from poor adaptability to variations in material properties and process parameters. This proposed solution employs an adaptive parameter adjustment algorithm and a multi-model integration strategy, enabling it to rapidly adapt to different stamping materials and process conditions. When producing high-strength sheet metal stampings for different automotive models, extensive readjustments are unnecessary; rapid switching and high-precision prediction and compensation are maintained, effectively shortening product development cycles and improving production flexibility.

[0077] Springback compensation often requires multiple manual interventions, resulting in low efficiency. This solution's automatic compensation device achieves a fully automated process from parameter input and springback prediction to process parameter adjustment. On a continuous stamping production line, it can monitor and compensate for springback in real time, reducing manual operation and downtime. Compared with traditional methods, production efficiency is improved, labor costs are reduced, and the stability and consistency of the production process are enhanced. Because this solution improves prediction accuracy and production efficiency, reduces scrap rate and manual intervention, it also reduces costs in mold development and production processes.

[0078] In step S01, a deep learning algorithm is used to perform secondary analysis on the data acquired by transmission electron microscopy and atomic probe tomography (APT). This automatically identifies and marks regions with unique microstructural features. Transmission electron microscopy and APT generate massive amounts of data, containing numerous microstructural information. The deep learning algorithm can quickly process this massive amount of data, automatically identifying regions with unique microstructural features, avoiding the tedious process of manually examining and analyzing data one by one, thus significantly saving time and labor costs. For example, when analyzing the microstructure of large sheet metal, manual screening might take days or even weeks, while a deep learning algorithm can complete the same task in a few hours. The real-time processing capability of the deep learning algorithm can provide analysis results in a short time, allowing operators to adjust production parameters promptly based on the results. For instance, when the algorithm identifies a unique microstructural region in the sheet metal that affects springback, operators can immediately adjust the stamping process to avoid producing defective products. Furthermore, in the subsequent construction and training of the springback prediction model, the accurately marked unique microstructural feature region data can serve as an important input parameter. These data can more accurately reflect the relationship between the microstructure of materials and their rebound performance, thereby improving the accuracy and reliability of rebound prediction models. For example, when training models such as support vector regression and kernel ridge regression, using data processed by deep learning algorithms can help the models better learn the mapping relationship between microstructural features and rebound, thus improving prediction accuracy.

[0079] In step S02, the improved UMAT subroutine employs parallel computing optimization technology, leveraging the parallel computing capabilities of the graphics processing unit (GPU) to accelerate the stamping springback simulation process. By distributing the computational tasks in the UMAT subroutine across multiple GPU cores for parallel execution, the simulation time can be significantly shortened. For example, for a complex stamping part springback simulation, traditional CPU serial computation might take hours or even days, while GPU parallel computation could reduce it to tens of minutes, greatly improving simulation efficiency. GPU-accelerated stamping springback simulation can provide accurate results in a shorter time, enabling technicians to quickly adjust the production process based on the simulation results and reduce production cycles. For large stamping parts or complex stamping processes, the computational load increases dramatically. The parallel computing capabilities of the GPU can effectively handle these large-scale problems, avoiding model simplification or incomplete calculations due to insufficient computing resources. For example, in the stamping springback simulation of automotive body panels, the GPU can be used to perform accurate simulation analysis of the entire body panel without excessive simplification of the model.

[0080] In section S04, the improved ResNet-50 model's attention mechanism employs a dynamic weight update strategy. This strategy adjusts the attention weights in real-time based on the characteristics of different microstructure images. These images differ in elemental distribution, grain boundary features, and dislocation density, and these differences have varying degrees of impact on springback prediction. The dynamic weight update strategy allows the attention mechanism to automatically focus on the regions and features most critical to springback prediction for each image. For example, in some images, the specific morphology of grain boundaries may be a key factor affecting springback. Dynamic weight adjustment allows the model to pay more attention to the features of grain boundary regions, thereby extracting this important information more accurately and providing strong support for subsequent springback prediction. Microstructure images may contain information that is not closely related to springback prediction. If the model gives the same attention to all regions and features, it will increase computational load and may introduce noise interference, reducing the model's accuracy. By dynamically adjusting the attention weights, the model can reduce its focus on this redundant information, allocating more computational resources to processing key features and improving the efficiency and quality of feature extraction.

[0081] The adaptive parameter adjustment algorithm in S03 dynamically adjusts the step size of process parameters based on historical simulation data and material properties. In the initial stage of stamping simulation, the range of process parameter values ​​is usually quite wide. If a fixed step size is used for parameter traversal, a significant amount of computational resources and time may be wasted in invalid parameter regions. The adaptive parameter adjustment algorithm, based on historical simulation data and material properties, can quickly identify parameter ranges that may produce effective results, and appropriately reduce the step size for fine searching within these ranges, while increasing the step size to quickly skip invalid regions, thereby greatly shortening the simulation time.

[0082] The improved ResNet-50 model in S04 incorporates an attention mechanism, which automatically focuses on regions in the microstructure map. This attention mechanism automatically identifies regions closely related to stencil springback, such as specific grain boundaries and dislocation-dense regions, and assigns higher attention weights to these key areas. This allows the model to focus more on critical information during feature extraction, avoiding interference from irrelevant or secondary information, thus more accurately capturing the key features affecting stencil springback. The attention mechanism can also keenly capture these subtle features and give them sufficient attention. For example, differences in elemental distribution at certain atomic scales may be able to be extracted by the model through the focus of the attention mechanism, providing richer and more accurate information for subsequent predictions.

[0083] The transfer learning strategy in S04 initializes five machine learning models using model parameters pre-trained in the relevant domain, in order to accelerate model convergence and reduce training time.

[0084] The dynamic weight allocation algorithm dynamically adjusts the weights of each model in the integrated model based on the evaluation metrics of the models under different working conditions. The working conditions during the stamping process are complex and variable; different material properties, process parameters, and environmental conditions all affect springback. A single model can hardly maintain good performance under all working conditions. The dynamic weight allocation algorithm can dynamically adjust the weights based on real-time working conditions and the evaluation results of each model, enabling the integrated model to better adapt to various complex working conditions and provide more accurate springback predictions. The distribution of stamping data may change with time, production batches, and other factors. The dynamic weight allocation algorithm can monitor the performance of each model on new data in real time and dynamically adjust the weights according to the evaluation metrics, allowing the integrated model to quickly adapt to changes in data distribution and maintain stable predictive performance. At the same time, the performance of a single model may fluctuate due to factors such as data noise and outliers. Through the dynamic weight allocation algorithm, the integrated model can combine the prediction results of multiple models and dynamically adjust the weights according to the evaluation metrics of each model, reducing the impact of fluctuations in a single model on the overall prediction results and improving the stability of the model.

[0085] In practical application

[0086] Experimental equipment preparation: A high-resolution transmission electron microscope (TEM) and an atomic probe tomography (APT) were selected to ensure that the resolution, accuracy, and other performance indicators of the equipment met the requirements for observing the microstructure of high-strength sheet metal. The equipment was calibrated and adjusted to ensure the accuracy of the acquired data.

[0087] Sample preparation: Representative samples are selected from different batches of high-strength sheet metal produced using different processes. The samples are prepared into appropriate sizes and shapes according to the requirements of TEM and APT. For TEM, the sample thickness needs to be controlled to around tens of nanometers; for APT, the samples are typically prepared into needle-like shapes.

[0088] Data Acquisition: TEM was used to observe the samples, obtaining microstructure images of the sheet metal and recording information such as elemental distribution, grain boundary characteristics, and dislocation density. Simultaneously, APT technology was used to accurately determine microstructural characteristics such as elemental distribution and grain boundary chemical composition at the atomic scale.

[0089] Material property testing: The elastic modulus of the material was determined using a Dynamic Mechanical Analyzer (DMA), and the hardening index was obtained through tensile testing combined with a microstructure evolution model. These material property data were recorded and stored in conjunction with the corresponding microstructure data.

[0090] Secondary analysis using deep learning: The acquired microstructure data is imported into an analysis system based on deep learning algorithms. This system employs a Convolutional Neural Network (CNN) architecture, trained on a large amount of labeled microstructure image data to learn patterns of special microstructure features. After training, secondary analysis is performed on the data acquired by TEM and APT, automatically identifying and labeling regions with special microstructure features, such as anomalous grain boundaries and high-density dislocation regions. The analysis results are then integrated with the original data to provide more accurate data support for subsequent simulations and model training.

[0091] Finite element model establishment: Using professional finite element analysis software such as ABAQUS, an accurate three-dimensional finite element model is established based on the design drawings and actual dimensions of the stamped parts. The model is then meshed appropriately, with finer meshes used in critical areas such as those prone to springback to improve calculation accuracy.

[0092] Improved UMAT Subroutine Writing: Based on the microstructure and mechanical properties of the material, an improved UMAT subroutine was written. The subroutine considers the influence of microstructure on the material's constitutive relations, such as grain boundary strengthening and dislocation slip mechanisms. The written UMAT subroutine was then imported into the ABAQUS software to ensure its compatibility and stability.

[0093] GPU parallel computing configuration: Build a computing platform with a high-performance graphics processing unit (GPU), and install the corresponding drivers and parallel computing software such as CUDA. Parallelize the improved UMAT subroutine, rationally distributing computing tasks across multiple GPU cores. Enable GPU-accelerated computing in the ABAQUS software settings.

[0094] Stamping springback simulation: Initial process parameters are set for the stamping material, such as forming blank holder force, lubrication conditions, and forming speed. The stamped part is sampled using the finite element method, and stamping springback simulation is performed in ABAQUS software using a modified UMAT subroutine. During the simulation, the calculation progress and results are monitored in real time to ensure the stability and accuracy of the simulation. Springback vector data, including displacement, stress, and strain information at each sampling point, are recorded as sample data under the initial process parameter conditions.

[0095] Adaptive parameter adjustment algorithm implementation: An adaptive parameter adjustment algorithm was developed. Based on the historical simulation data and characteristics of the material, the algorithm adopts the idea of ​​reinforcement learning and aims to maximize the prediction accuracy. It dynamically adjusts the step size of the process parameters. Based on the previous simulation results, the algorithm evaluates the impact of different combinations of process parameters on springback and intelligently selects the parameter step size for the next simulation.

[0096] Multi-material, multi-parameter traversal simulation: For various stamping materials, an adaptive parameter adjustment algorithm is used to traverse the initial process parameters. During the traversal, the parameter values ​​of each parameter are adjusted according to the algorithm to obtain different process parameter conditions. Under each corresponding process parameter condition, the initial simulation sampling steps are repeated to perform finite element sampling on the stamped parts. Stamping springback simulation is performed using ABAQUS combined with an improved UMAT subroutine, and springback vector data is recorded. The sample data under the initial process parameter conditions and the sample data under each corresponding process parameter condition are integrated to form a stamping springback dataset.

[0097] Improved ResNet-50 Model Setup: An improved ResNet-50 model is adopted, incorporating an attention mechanism module into its convolutional layers. This attention mechanism module employs a dynamic weight update strategy, adjusting the attention weights in real-time based on the characteristics of different microstructure images. The improved ResNet-50 model is initialized by loading model parameters pre-trained in relevant fields, such as those used in materials microstructure analysis.

[0098] Dataset Processing: Microstructure diagrams, material property parameters, and process parameters are integrated to construct a dataset. The dataset is then divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Data normalization is performed to ensure that data with different characteristics have the same scale, improving model training efficiency and accuracy.

[0099] Model Training: Transfer learning and ensemble learning strategies were employed to train five models: Support Vector Regression (SVM), Kernel Ridge Regression, K-Nearest Neighbors (KNN), Random Forest Regression, and Gradient Boosting Tree. During training, a grid search method was used to find the optimal parameters for each model. The root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) were used as evaluation metrics to assess model performance. Based on the evaluation results, a dynamic weight allocation algorithm was used to calculate the R² ratio of each model, determining the weight of each model in the final prediction, and ensembled the five models into a final rebound prediction model. Throughout the training process, the model's training progress and performance metrics were monitored in real time, and training parameters were adjusted based on the validation set results to prevent overfitting.

[0100] Automatic Compensation Device Construction: An automatic compensation device is constructed, comprising a visual parameter input module, an embedded springback prediction module, a springback compensation calculation module, and an automatic control module. The visual parameter input module uses a graphical interface design, providing operators with an intuitive and convenient operating interface. Operators can input process parameters, such as forming blank holder force, lubrication condition parameters, forming speed parameters, and material property parameters, such as material thickness, yield strength, and tensile strength. The embedded springback prediction module embeds the final springback prediction model, receives parameters from the visual parameter input module, and predicts the springback vector value during the stamping process. The springback compensation calculation module calculates the springback compensation vector value based on the prediction results and preset compensation rules, using rules formulated based on material mechanics principles and empirical formulas. The automatic control module generates control signals based on the compensation values, and adjusts the process parameters of the stamping equipment, such as die clearance and stamping pressure, through an interface with the stamping equipment, to achieve automatic compensation for the springback of the stamped parts.

[0101] System Integration and Debugging: Integrate the automatic compensation device with the stamping production equipment, establish a data transmission channel, and ensure smooth data interaction between modules. Debug the entire system by inputting different process parameters and material property parameters to verify the system's predictive accuracy and compensation effect. During debugging, optimize and adjust the system to improve its stability and reliability.

[0102] Production Application and Optimization: The stamping prediction method based on multiple machine learning models and the automatic compensation device are applied to actual stamping production. During production, new stamping data, including process parameters, material properties, and springback data, are continuously collected. The stamping springback dataset is updated, and the final springback prediction model is periodically retrained and optimized. Model parameters and weights are adjusted based on new data to improve the model's prediction accuracy and adaptability. Simultaneously, the preset compensation rules and control strategies of the automatic compensation device are adjusted and optimized according to actual production conditions to further improve the quality and efficiency of stamping production.

[0103] Example 2

[0104] As per the instruction manual Figure 1-10 As shown, the springback prediction method for high-strength sheet metal stamping parts based on multiple machine learning models in this embodiment of the invention includes four steps:

[0105] S101: Using experimental techniques such as scanning electron microscopy (SEM), transmission electron microscopy (TEM), or X-ray diffraction (XRD) to obtain microstructural images of high-strength sheet metal, the microstructural characteristics such as grain morphology, grain size, orientation, and texture of the high-strength sheet metal can be clearly observed. These structural characteristics directly affect the strength, toughness, resilience, and fatigue resistance of the high-strength sheet metal. Microstructure greatly influences the resilience of the material, and the inclusion of microstructural features can significantly enhance the model's resilience prediction capability.

[0106] S102: Process parameters include forming blank holder force, part dimensional parameters, lubrication conditions, and forming speed. The stamped part is sampled using the finite element method, combined with the material definition module of CAE software, such as the UMAT subroutine in ABAQUS, and microstructure analysis, to complete springback prediction and generate sample data. Specifically, sampling locations are specified in the finite element model of the stamped part, and CAE simulation is used to evaluate springback, obtaining the dimensional deviation vector after springback at each sampling point, which is then used as sample data for initial process parameters.

[0107] S103: The material properties of a single specification of material are fixed. By iteratively changing the process parameters while using the sampling method in step S102, data for a single material under different process parameters can be obtained. Similarly, by selecting raw materials of different specifications and iteratively changing the process parameters again, and using CAE forming springback prediction, data for multiple materials under different process parameters can be obtained, thus generating a dataset.

[0108] Step S104: The microstructure map is feature extracted using a resNet-50 pre-trained model. The dataset is then reorganized by combining the obtained material property parameters and process parameters. The test and training sets are then divided in a 2:8 ratio, and five models are trained. The trained models are then tested on the test set to calculate the performance of each model. Performance indicators such as MSE and MAPE, and then such as Figure 10 As shown The proportions of the models are combined to form the final model. Essentially, this ensemble model involves inputting the predicted parameters into each of the trained models before calling them, and finally, the outputs of the five models are determined according to the values ​​obtained from the test set. The proportions are combined to obtain the final prediction data. Compared with single-model prediction methods, this multi-model integration approach has greater flexibility and robustness, effectively addressing the variability and complexity of high-strength sheet metal stamping processes, reducing model risk, and improving prediction accuracy. By combining the advantages of multiple models, it can more comprehensively handle different types of data and feature relationships, offering greater advantages in prediction accuracy, computational efficiency, and model optimization. Furthermore, this method fully considers the impact of process parameters and material properties on springback.

[0109] In regression analysis, R² R² is a metric for measuring the goodness of a model fit, representing the proportion of the independent variable input that explains the dependent variable output. The value of R² ranges from 0 to 1; the closer to 1, the better the model fit.

[0110]

[0111] in It is the actual value. It is a predicted value. It is the mean of the true values. n It is the sample size;

[0112] Mean Squared Error (MSE) measures the squared average of the differences between predicted and actual values. The formula is:

[0113]

[0114] in It is the actual value. It is a predicted value. n It is the sample size;

[0115] The Mean Absolute Percentage Error (MAPE) measures the relative error between the predicted and actual values, expressed as a percentage. The formula is:

[0116] MSE is a commonly used loss function in regression problems, representing the squared mean of the differences between predicted and actual values. The formula is:

[0117] in It is the actual value. It is a predicted value. n It is the sample size;

[0118] MAPE measures the relative error between predicted and actual values, usually expressed as a percentage. It can be used to evaluate the accuracy of regression models, especially in prediction problems with different dimensions, where MAPE provides a normalized error measure.

[0119]

[0120] in It is the actual value. It is a predicted value.

[0121] As per the instruction manual Figure 10 As shown, the automatic rebound compensation device of this invention includes:

[0122] The visualization parameter input module is used to visually set process parameters and material property parameters for the automatic springback compensation device, and input the parameters into the embedded springback prediction module.

[0123] As an optional example, the visualization parameter input module in this embodiment of the invention is specifically used to: visualize the parameters required for prediction input to the embedded prediction module.

[0124] The embedded springback prediction module generates a prediction model based on a multi-machine learning model for predicting springback of high-strength sheet metal stamped parts. This model is then embedded in the device to predict the springback vector value during the stamping process, provided that process parameters and material property parameters are provided.

[0125] As an optional example, the embedded springback prediction module of this embodiment of the invention is specifically used to: predict springback in the stamping process under determined process parameters and material properties, obtain springback vector values, and the prediction of springback vector values ​​can be used to calculate springback compensation vectors to adjust the process to achieve ideal stamped parts.

[0126] The springback compensation calculation module is used to calculate the springback compensation vector value;

[0127] As an optional example, the springback compensation calculation module of this embodiment of the invention is specifically used to: calculate and convert the springback value predicted by the embedded springback prediction module into a springback compensation value.

[0128] The automatic control module is used to generate control signals and adjust the process.

[0129] As an optional example, the automatic control module of this embodiment of the invention is specifically used to: convert the springback compensation vector value calculated by the springback compensation calculation module into a control signal to adjust the process so as to obtain the ideal stamped part.

[0130] The above are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A stamping prediction method based on multiple machine learning models, characterized in that, The method includes: S01 Microstructure and Feature Acquisition: The microstructure of the sheet metal was obtained by using transmission electron microscopy combined with atomic probe tomography. The microstructure characteristics of elemental distribution, grain boundary features and dislocation density at the atomic scale were determined, and the characteristics of the material's elastic modulus and hardening index were obtained. S02 Initial Simulation Sampling: The stamping parts are sampled using the finite element method, and the stamping springback simulation is performed using ABAQUS combined with the improved UMAT subroutine, and the springback vector data is recorded. S03 Dataset Expansion: For various stamping materials, an adaptive parameter adjustment algorithm is used to traverse the initial process parameters, adjust the parameter step size according to the material properties and previous simulation results, and obtain the stamping springback dataset through simulation sampling. The adaptive parameter adjustment algorithm dynamically adjusts the change step size of the process parameters based on the historical simulation data and characteristics of the material. Model S04 Construction and Training: An improved ResNet-50 model was adopted, incorporating an attention mechanism to extract features from the microstructure map. The extracted features were integrated with material property parameters and process parameters. The dataset was divided into training, validation, and test sets. Transfer learning and ensemble learning strategies were used to train five models: Support Vector Regression, Kernel Ridge Regression, KNN, Random Forest Regression, and Gradient Boosting Tree. Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (CDE) were used as the benchmarks. As evaluation indicators, the evaluation results are integrated into the final rebound prediction model using a dynamic weight allocation algorithm. This is used to measure how well a model fits the data; the closer the value is to 1, the better the model fit. (Calculation...) The formula is in It is the actual value. It is a predicted value. It is the mean of the true values. n It is the sample size; Mean Squared Error (MSE) measures the squared average of the differences between predicted and actual values. The formula is: in It is the actual value. It is a predicted value. n It is the sample size; The Mean Absolute Percentage Error (MAPE) measures the relative error between the predicted and actual values, expressed as a percentage. The formula is: According to each model The ratio is used to determine the weight of each model in the final prediction, and the models are integrated into a compensated prediction model. The input is material properties and process parameters, and the model outputs the rebound prediction value. S05 Automatic Compensation: The final springback prediction model is embedded in the automatic compensation device. The device's visualization parameter input module receives process parameters and material property parameters input by the operator; the embedded springback prediction model predicts the springback vector; the springback compensation calculation module calculates the springback compensation value based on the prediction result and preset compensation rules; the automatic compensation control module generates a control signal based on the compensation value to adjust the process parameters of the stamping equipment and automatically compensate for the springback of the stamped parts.

2. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, In step S01, a deep learning algorithm is used to perform secondary analysis on the data obtained by transmission electron microscopy and atomic probe tomography, automatically identifying and marking regions with special microstructural features.

3. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, In S02, the improved UMAT subroutine adopts parallel computing optimization technology, which utilizes the parallel computing capabilities of the graphics processing unit (GPU) to accelerate the stamping and springback simulation process.

4. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, In S04, the improved ResNet-50 model's attention mechanism employs a dynamic weight update strategy, adjusting the attention weights in real time according to the characteristics of different microstructure images.

5. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, The improved ResNet-50 model in S04 incorporates an attention mechanism, which can automatically focus on key regions in the microstructure map.

6. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, The transfer learning strategy in S04 initializes five machine learning models using model parameters pre-trained in the relevant domain, in order to accelerate model convergence and reduce training time.

7. The stamping prediction method based on multiple machine learning models according to claim 1, characterized in that, The dynamic weight allocation algorithm dynamically adjusts the weights of each model in the integrated model based on the evaluation indicators of the model under different operating conditions.

8. An automatic compensation device for high-strength sheet metal stamping parts based on a multi-machine learning model, using the stamping prediction method based on a multi-machine learning model as described in any one of claims 1-7, characterized in that, include: The visualization parameter input module is used to visually set process parameters and material property parameters for the automatic springback compensation device and input the parameters into the embedded springback prediction module. The embedded springback prediction module generates a prediction model based on a method for predicting springback of high-strength sheet metal stamped parts using a multi-machine learning model. This model is then embedded in the device to predict the springback vector value during the stamping process, provided that process parameters and material property parameters are provided. The springback compensation calculation module is used to calculate the springback compensation vector value; The automatic control module is used to generate control signals and adjust the process.

Images