Digital twin construction method and computer system for metal stamping

Abstract

The present invention provides a method and computer system for constructing a digital twin for metal stamping. The method obtains a multi-source monitoring data set in the metal stamping process and performs collaborative feature extraction processing to generate a material structure feature set and a process dynamic feature set. A hybrid feature space is constructed based on the material structure feature set and the process dynamic feature set. Multi-dimensional feature matching processing is performed in the hybrid feature space to generate a feature mapping relationship network and construct a dynamic twin model based on the network. The stamping process is deduced in real time based on the dynamic twin model to generate forming quality prediction results and process defect location results. A stamping process parameter adjustment scheme is generated and fed back to the stamping control system to trigger parameter optimization operations. The present invention improves the early warning capability of process defects and the accuracy of process parameter adjustment under complex working conditions, while ensuring the real-time and engineering feasibility of constructing digital twins for large-scale stamping processes.

Description

Technical Field

[0001] The present invention relates to the fields of digital control and data processing and analysis, and in particular to a digital twin construction method and computer system for metal stamping forming. Background Art

[0002] With the deepening application of intelligent manufacturing technologies, digital modeling and process optimization of the metal stamping process have become core components of modern industrial manufacturing. Digital twin technology, by constructing a virtual mapping model of the physical process, provides technical support for forming quality prediction and process defect analysis. Current mainstream approaches rely primarily on monitoring data (such as material thickness changes or equipment pressure curves) to construct static simulation models and generate process adjustment recommendations through offline parameter matching. However, such approaches struggle to effectively integrate the coupled relationships between material deformation characteristics, equipment dynamic response, and energy transfer processes, resulting in insufficient sensitivity for identifying hidden defects such as transient stress concentration and abnormal energy dissipation in forming quality prediction models. Furthermore, traditional process optimization strategies often rely on empirical rules to locally adjust single parameters (such as stamping speed or die pressure), lacking a global collaborative optimization mechanism driven by multi-source data, which can easily lead to parameter conflicts and secondary process defects. Furthermore, existing digital twin construction processes lack the ability to process large-scale, heterogeneous data in time and space, causing real-time model updates to lag behind actual production cycles and failing to meet the dynamic optimization requirements under complex working conditions. The above defects seriously restrict the improvement of quality control accuracy and process iteration efficiency of the stamping process. There is an urgent need for a digital twin construction method that can realize multi-physical field coupling analysis, dynamic closed-loop optimization and real-time state deduction. Summary of the Invention

[0003] The present invention provides a digital twin construction method and computer system for metal stamping.

[0004] In the first aspect, an embodiment of the present invention provides a digital twin construction method for metal stamping forming, including: obtaining a multi-source monitoring data set in the metal stamping forming process, the multi-source monitoring data set including material physical parameter data, equipment operation status data and forming process dynamic response data; performing collaborative feature extraction processing on the multi-source monitoring data set to generate a material structure feature set and a process dynamic feature set, the material structure feature set characterizing the deformation characteristics and stress distribution characteristics of the stamping material, and the process dynamic feature set characterizing the action timing characteristics and energy transfer characteristics of the stamping equipment; constructing a hybrid feature space based on the material structure feature set and the process dynamic feature set, and performing multi-dimensional feature matching processing in the hybrid feature space to generate a feature mapping relationship network; constructing a dynamic twin model according to the feature mapping relationship network, and performing real-time state deduction on the stamping forming process based on the dynamic twin model to generate forming quality prediction results and process defect location results; generating a stamping process parameter adjustment plan based on the forming quality prediction results and the process defect location results, and feeding back the stamping process parameter adjustment plan to the stamping control system to trigger parameter optimization operation.

[0005] In a second aspect, an embodiment of the present invention provides a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the digital twin construction method for metal stamping forming as described above.

[0006] The digital twin construction method for metal stamping provided by the present invention constructs a hybrid feature space covering material deformation mechanism, equipment action logic and energy transfer process by integrating multi-source monitoring data of material physical parameters, equipment operating status and dynamic response of the forming process, and realizes full-dimensional state deduction of the stamping process based on a dynamic mapping relationship network. Through collaborative feature extraction and spatiotemporal alignment processing of multi-source data, this method integrates the thickness change, stress distribution and equipment action parameters of traditional isolated analysis into a set of associated features, significantly improving the coupled analysis capability of forming quality prediction and process defect location; utilizing the real-time parameter correction mechanism of the dynamic twin model, the material deformation propagation path simulation is dynamically matched with the equipment operating status monitoring data, effectively identifying transient stress concentration and energy transfer anomalies that are difficult to capture with traditional offline detection; establishing a direct correlation between deformation prediction results and process parameter adjustment through an optimization strategy network, generating a multi-parameter collaborative optimization scheme based on the dual constraints of material flow uniformity and equipment operation stability, overcoming the parameter conflict and secondary defect risks caused by single-dimensional process adjustment. Without relying on a large number of historical defect sample labels, this method significantly improves the early warning capability of process defects and the accuracy of process parameter adjustment under complex working conditions through the adaptive mapping and deduction mechanism of feature space, while ensuring the real-time and engineering feasibility of constructing digital twins of large-scale stamping processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flowchart of a method for constructing a digital twin for metal stamping provided by an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] See also Figure 1 , Figure 1 A flowchart of a method for constructing a digital twin for metal stamping provided by an embodiment of the present invention, the method being executed by a computer system, comprising the following steps:

[0010] Step S100: Acquire a multi-source monitoring data set during a metal stamping forming process, wherein the multi-source monitoring data set includes material physical parameter data, equipment operating status data, and forming process dynamic response data.

[0011] Material physical parameter data is data that describes the physical properties of the stamping material itself, such as the material's density, elastic modulus, Poisson's ratio, etc., reflecting the basic properties of the material when deformed by force. Equipment operating status data is data about the operation of the stamping equipment, covering information such as the equipment's spindle speed, hydraulic system pressure, mold position and motion status, and can reflect the equipment's working status and performance during the stamping process. Forming process dynamic response data is data on the dynamic changes in materials and equipment caused by the stamping action during the stamping process, such as the plastic deformation of the material, equipment vibration, and energy transfer data, reflecting the real-time changes in each link of the stamping process.

[0012] For material physical parameter data, thickness sensors can be used to collect real-time thickness measurement data of stamping materials during the forming process, and yield strength testing equipment can be used to obtain yield strength test data of stamping materials at different forming stages. For equipment operating status data, spindle speed data of the stamping equipment can be obtained through spindle speed sensors, and hydraulic system pressure data can be collected through hydraulic system pressure sensors. For dynamic response data during the forming process, power sensors can be used to calculate the energy consumption rate of the material's plastic deformation, and infrared thermal imagers can be used to obtain temperature distribution data at the contact surface between the mold and the material.

[0013] Step S200: Perform collaborative feature extraction processing on the multi-source monitoring data set to generate a material structure feature set and a process dynamic feature set. The material structure feature set characterizes the deformation characteristics and stress distribution characteristics of the stamping material, and the process dynamic feature set characterizes the action timing characteristics and energy transfer characteristics of the stamping equipment.

[0014] Collaborative feature extraction comprehensively considers the correlations and interactions between different types of data in a multi-source monitoring data set to extract key features that reflect material and process characteristics. The material structure feature set is a set of features used to describe the deformation and stress distribution of stamping materials during the forming process. Deformation characteristics reflect the shape change of the material under the action of the stamping pressure, such as stretching, compression, and bending. Stress distribution characteristics reflect the magnitude and distribution of internal stresses in the material, such as the residual stress distribution pattern. The process dynamic feature set describes the timing of the movements and energy transfer of the stamping equipment during operation. Timing characteristics describe the temporal sequence and speed changes of the movements of various equipment components, such as the stamping speed variation curve and the delay time of pressure transfer. Energy transfer characteristics reflect the energy conversion and transfer during the stamping process, such as the energy consumption rate of material plastic deformation and the kinetic energy loss rate of the equipment.

[0015] As an implementation manner, step S200 may specifically include the following steps S210 to S240:

[0016] Step S210: performing structural deformation feature extraction processing on the material physical parameter data to obtain structural deformation features in the material structural feature set; wherein the structural deformation features include the material thickness change rate, yield strength change gradient and residual stress distribution pattern.

[0017] The purpose of structural deformation feature extraction is to extract key features from this data that can reflect the structural deformation of the material. The material thickness change rate is the ratio of the change in material thickness over time or position during the stamping process, reflecting the thinning or thickening of the material under the action of the stamping pressure. The yield strength change gradient is the rate of change of the material's yield strength at different positions or different forming stages, reflecting the changes in the material's mechanical properties during the deformation process. The residual stress distribution pattern is the distribution pattern of the internal residual stress of the material after stamping, including information such as the magnitude, direction, and distribution area of ​​the residual stress.

[0018] As an implementation manner, step S210 may specifically include the following steps S211 to S214:

[0019] Step S211: collecting real-time thickness measurement data of the stamping material during the forming process, and performing spatial interpolation processing on the real-time thickness measurement data to generate a thickness change rate distribution cloud map.

[0020] Real-time thickness measurement data is collected in real time by thickness measurement equipment during the stamping process. This data reflects the material's thickness at different times and locations. Spatial interpolation converts discrete thickness measurement data into continuous thickness variation information. This process allows the distribution of the material's thickness variation rate across the entire forming area to be determined. A thickness variation rate distribution cloud chart is a visualization tool that graphically displays the spatial distribution of the material's thickness variation rate, providing a direct reflection of the material's thickness variation during the stamping process. Various thickness measurement devices, such as laser thickness sensors and ultrasonic thickness sensors, can be used to collect real-time thickness measurement data during the stamping process. For example, during the stamping process of a metal pipe, laser thickness sensors can be installed at different locations within the stamping die to collect thickness data at predetermined intervals. Spatial interpolation of real-time thickness measurement data can employ various interpolation algorithms, such as kriging and bilinear interpolation. Specifically, the collected thickness data is first sorted by spatial location. The distance and correlation between each data point and other data points are then calculated, and a covariance matrix is ​​constructed based on this information. Next, the covariance matrix is ​​solved using the least squares method to obtain the weight coefficient for each unknown location. Finally, the known thickness data is weighted and summed according to the weight coefficient to obtain the thickness prediction value at the unknown location, thereby generating a thickness change rate distribution cloud map.

[0021] Step S212: Obtaining yield strength test data of the stamping material at different forming stages, and performing gradient calculation processing on the yield strength test data to generate a yield strength change gradient curve.

[0022] Yield strength test data is obtained by mechanically testing stamping materials, yielding values ​​at different forming stages. This data reflects the stress at which plastic deformation begins at varying degrees of deformation. The yield strength gradient curve is a visualization tool that graphically displays the rate of change of yield strength as a function of position or forming stage, providing a direct reflection of the material's mechanical properties during the stamping process.

[0023] Obtaining the yield strength test data of stamping materials at different forming stages can be achieved through mechanical property testing methods such as tensile tests and compression tests. For example, in the stamping process of a metal sheet, samples can be cut at different stamping stages for tensile tests to measure their yield strength. Numerical differentiation methods, such as the finite difference method, can be used to perform gradient calculation processing on the yield strength test data. Specifically, it is assumed that a series of discrete yield strength test data points are known, and each data point corresponds to a forming stage or position. First, calculate the yield strength difference between two adjacent data points, and then divide the difference by the corresponding position or stage interval to obtain the average gradient value in the interval. By calculating all adjacent data points, a series of gradient values ​​can be obtained. Finally, these gradient values ​​are sorted according to the corresponding positions or stages, and a yield strength change gradient curve is drawn.

[0024] Step S213: collecting material surface stress distribution data based on a preset residual stress detection device, and performing pattern recognition processing on the stress distribution data to generate a residual stress distribution pattern feature vector.

[0025] Pre-configured residual stress detection devices are pre-configured devices used to detect residual stress distribution on a material surface, such as X-ray diffractometers and blind hole stress detectors. Pattern recognition processing analyzes and processes the collected stress distribution data to extract representative features and patterns. The residual stress distribution pattern feature vector contains key information about the residual stress distribution, such as stress magnitude, direction, and distribution area.

[0026] Various algorithms can be used to perform pattern recognition processing on stress distribution data, such as principal component analysis and cluster analysis. Taking the principal component analysis algorithm as an example, the collected stress distribution data is standardized to have the same scale and range. Then, the covariance matrix of the data is calculated, and the principal components of the data are obtained by solving the eigenvalues ​​and eigenvectors of the covariance matrix. The first N principal components with larger eigenvalues ​​are selected as the main features, and the original data are projected onto these principal components to obtain the eigenvectors after dimensionality reduction. Finally, these eigenvectors are combined to generate the residual stress distribution pattern eigenvector.

[0027] Step S214: normalizing and fusing the thickness change rate distribution cloud map, the yield strength change gradient curve, and the residual stress distribution pattern characteristic vector to obtain the structural deformation characteristics.

[0028] Normalized fusion processing unifies feature data of different types and scales to bring them into the same dimension and range, then fuses these features to create a comprehensive representation. The thickness change rate distribution cloud, yield strength gradient curve, and residual stress distribution pattern eigenvectors each reflect the structural deformation of the material during the stamping process from different perspectives, but their numerical ranges and data formats may differ. Through normalized fusion processing, these features can be effectively integrated to create a comprehensive representation that comprehensively and accurately reflects the material's structural deformation characteristics.

[0029] As an implementation manner, step S214 may specifically include the following steps S2141 to S2147:

[0030] Step S2141: performing data format conversion processing on the thickness change rate distribution cloud map, mapping the pixel grayscale value into a continuous change rate numerical sequence, and generating a thickness change rate distribution cloud map with a unified format.

[0031] For example, a thickness change rate distribution cloud map is presented as an image, where the grayscale value of each pixel represents the thickness change rate at that location. Data format conversion converts these pixel grayscale values ​​into a continuous sequence of change rate values ​​for subsequent processing and analysis. This continuous change rate sequence represents thickness change rate data in numerical form, making it easier to integrate and compare with other feature data.

[0032] A mapping function approach can be used to convert the data format of a thickness change rate distribution cloud map. First, determine the mapping relationship between grayscale values ​​and change rate values. This mapping relationship can be determined experimentally or empirically. For example, by selecting locations with known thickness change rates in the cloud map, measuring their corresponding grayscale values, and then establishing a linear or nonlinear mapping function between grayscale values ​​and change rate values. Next, traverse each pixel in the cloud map and convert its grayscale value into a corresponding change rate value according to the mapping function. Finally, arrange these converted values ​​in order of pixel position to generate a continuous sequence of change rate values. For example, suppose the linear mapping function between grayscale value and change rate value obtained experimentally is: change rate value = a * grayscale value + b, where a and b are constants determined experimentally. For each pixel in the cloud map, read its grayscale value, substitute it into the mapping function to calculate the corresponding change rate value, and then store these values ​​sequentially in an array to obtain a continuous sequence of change rate values.

[0033] Step S2142: Perform sampling point encryption processing on the yield strength change gradient curve, convert the discrete gradient value into a gradient continuous signal consistent with the time scale of the continuous change rate numerical sequence based on the stamping forming time axis, and generate a time-aligned yield strength change gradient curve.

[0034] The yield strength change gradient curve is obtained by performing gradient calculation on discrete yield strength test data, so the gradient value on the curve is discrete. The sampling point encryption process is to insert new points between discrete gradient values ​​to make the curve smoother and continuous, so as to better reflect the change trend of the yield strength. The stamping forming time axis is a coordinate axis based on the time of the stamping process, which is used to determine the time position corresponding to each data point. The discrete gradient value is converted into a gradient continuous signal that is consistent with the time scale of the continuous change rate numerical sequence in order to make the yield strength change gradient curve and the thickness change rate distribution cloud map consistent in time scale, so as to facilitate subsequent feature fusion processing.

[0035] Linear interpolation can be used to intensify the sampling points of the yield strength gradient curve. First, determine the sampling interval to be intensified, for example, inserting a new point at a predetermined time interval. Then, for each location where a new point is to be inserted, use the linear interpolation formula to calculate the gradient value of the new point based on the two adjacent discrete gradient values. The linear interpolation formula is: New point gradient value = (1-t) * previous discrete gradient value + t * next discrete gradient value, where t is the relative position of the new point between the two adjacent discrete gradient values. In this way, new points are inserted between discrete gradient values, making the curve smoother and more continuous.

[0036] To convert discrete gradient values ​​into a gradient continuous signal consistent with the time scale of the continuous rate of change numerical sequence based on the stamping time axis, the encrypted curve needs to be time-aligned. First, determine the time scale of the continuous rate of change numerical sequence, that is, the time interval corresponding to each data point. Then, based on the stamping time axis, adjust each data point in the encrypted yield strength change gradient curve to the same time position as the continuous rate of change numerical sequence. If there is no corresponding gradient value at a certain time position, linear interpolation or other interpolation methods can be used for estimation.

[0037] Step S2143: performing spatial coordinate system conversion processing on the residual stress distribution pattern eigenvector, aligning the dimensional direction of the eigenvector with the spatial grid distribution of the thickness change rate distribution cloud map, and generating a spatially unified residual stress distribution pattern eigenvector.

[0038] The spatial coordinate system conversion process transforms the spatial coordinate system of the eigenvector so that its dimensional orientation aligns with the spatial grid distribution of the thickness change rate distribution cloud map. The spatial grid distribution of the thickness change rate distribution cloud map is the spatial arrangement of pixels within the cloud map, defining a spatial coordinate system. By aligning the dimensional orientation of the residual stress distribution pattern eigenvector with the spatial grid distribution of the cloud map, the two features are spatially consistent, facilitating subsequent feature fusion processing.

[0039] The spatial coordinate system conversion processing of the residual stress distribution pattern eigenvector can be performed by using transformation operations such as rotation and translation. First, determine the original spatial coordinate system of the eigenvector and the spatial grid coordinate system of the cloud map. Then, calculate the rotation angle and translation vector between the two coordinate systems. The rotation angle can be determined by calculating the angle between the coordinate axes of the two coordinate systems, and the translation vector can be determined by calculating the displacement between the origins of the two coordinate systems. According to the calculated rotation angle and translation vector, the eigenvector is rotated and translated. Specifically, for each component in the eigenvector, a corresponding transformation is performed according to the rotation matrix and the translation vector to obtain the converted eigenvector. Through these transformation operations, the dimensional direction of the residual stress distribution pattern eigenvector is aligned with the spatial grid distribution of the thickness change rate distribution cloud map to generate a spatially unified residual stress distribution pattern eigenvector.

[0040] Step S2144: Perform spatial alignment processing on the thickness change rate distribution cloud map after format unification, the yield strength change gradient curve after time alignment, and the residual stress distribution pattern feature vector after spatial unification, and project each feature data to the same spatial coordinate grid node based on the surface grid division rule of the stamping material to generate a multi-feature space superposition data set.

[0041] Spatial alignment processing is to align different types of feature data in space so that they have corresponding feature values ​​at the same spatial position. The stamping material surface grid division rule is a rule for dividing the stamping material surface into a number of grid nodes, and these grid nodes constitute a spatial coordinate system. Projecting each feature data to the same spatial coordinate grid node is to assign the eigenvalues ​​in the thickness change rate distribution cloud map after format unification, the yield strength change gradient curve after time alignment, and the residual stress distribution pattern eigenvector after spatial unification to the corresponding grid nodes according to the grid division rule. The multi-feature space overlay dataset is a dataset containing multiple feature information obtained by spatially superimposing these projected feature data.

[0042] The following steps can be used to perform spatial alignment processing on the thickness change rate distribution cloud map after format unification, the yield strength change gradient curve after time alignment, and the residual stress distribution pattern eigenvector after spatial unification. First, determine the spatial coordinates of each grid node according to the surface grid division rules of the stamping material. Then, for the thickness change rate distribution cloud map after format unification, project it onto the corresponding grid node according to the position of each pixel point in the cloud map and the corresponding change rate value. If a pixel point is located between multiple grid nodes, the interpolation method can be used to distribute its change rate value to the adjacent grid nodes. For the yield strength change gradient curve after time alignment, project its gradient value onto the corresponding grid node according to the time and spatial position corresponding to each data point in the curve. Similarly, if a data point is located between multiple grid nodes, the interpolation method can be used to distribute it. For the residual stress distribution pattern eigenvector after spatial unification, project its eigenvalue onto the corresponding grid node according to the spatial coordinates of the eigenvector.

[0043] After completing the projection of each feature data, the different feature values ​​projected onto the same grid node are superimposed to generate a multi-feature space superposition data set.

[0044] Step S2145: The thickness change rate values, yield strength gradient values, and residual stress eigenvalues ​​in the multi-feature space superposition data set are dimensionlessly processed, and the dynamic range compression algorithm is used to adjust the numerical range of each eigenvalue to the same proportional interval to generate a standardized multi-feature data set.

[0045] Dimensionless processing removes the dimensions of different eigenvalues ​​to make them comparable. Dynamic range compression algorithms compress the numerical range of eigenvalues ​​to a smaller interval. This algorithm aligns the numerical ranges of all eigenvalues ​​to the same proportional range, preventing some eigenvalues ​​from dominating others due to their large numerical ranges. A standardized multi-feature data set is a dataset obtained after dimensionless processing and dynamic range compression, in which all eigenvalues ​​have the same dimensions and numerical range, facilitating subsequent analysis and processing. Dimensionless processing can be achieved using various methods, such as normalization and standardization. Dynamic range compression algorithms can be used to align the numerical ranges of all eigenvalues ​​to the same proportional range using logarithmic compression. Logarithmic compression algorithms perform a logarithmic transformation on the numerical range of eigenvalues, compressing a larger numerical range to a smaller interval. Specifically, for each normalized eigenvalue, the logarithm is taken to obtain the compressed eigenvalue. For example, if the normalized eigenvalue is x, the compressed eigenvalue is y = log(x + 1). Through the logarithmic compression algorithm, the numerical range of each eigenvalue can be adjusted to the same proportional interval to generate a standardized multi-feature data set.

[0046] Step S2146: Based on the material deformation sensitivity weight distribution strategy in the stamping stage, the thickness change rate data, yield strength gradient data and residual stress data in the standardized multi-feature data set are weighted superimposed to generate a fusion feature distribution map.

[0047] The strategy for allocating weights to material deformation sensitivity during the stamping stage is to assign different weights to each feature based on the material's sensitivity to thickness changes, yield strength changes, and residual stresses during different stamping stages. During the stamping process, the material has different mechanical properties and deformation characteristics at different stages, and therefore has different sensitivities to different features. By rationally allocating weights, the features that have a greater impact on material deformation at different stages can be highlighted, thereby improving the accuracy and effectiveness of feature fusion. Weighted superposition processing is to perform a weighted summation of the thickness change rate data, yield strength gradient data, and residual stress data in the standardized multi-feature data set according to the assigned weights to obtain a comprehensive eigenvalue. The fusion feature distribution map is a visualization tool that graphically displays the spatial distribution of the comprehensive eigenvalues ​​after weighted superposition. It can intuitively reflect the comprehensive deformation of the material during the stamping process.

[0048] Specifically, the weight of each feature can be determined based on the mechanical properties and deformation characteristics of the material at different stamping stages. For example, in the early stage of stamping, the thickness change of the material has a greater impact on the overall deformation, so a larger weight can be assigned to the thickness change rate data; in the later stage of stamping, the distribution of residual stress has a greater impact on the quality and performance of the material, so a larger weight can be assigned to the residual stress data. Then, for each grid node in the standardized multi-feature data set, its corresponding thickness change rate data, yield strength gradient data, and residual stress data are multiplied by the corresponding weights to obtain the weighted eigenvalues. Finally, the weighted eigenvalues ​​are summed to obtain the comprehensive eigenvalue of the grid node. The comprehensive eigenvalues ​​of each grid node are arranged according to their spatial position to generate a fusion feature distribution map.

[0049] Step S2147: Perform local feature enhancement processing on the fused feature distribution map, detect the fused feature value of the deformation mutation area and perform smooth transition correction to generate structural deformation features.

[0050] Local feature enhancement processing is to enhance local areas in the fused feature distribution map to highlight the characteristic information of these areas. Deformation mutation areas are areas where the material deformation changes suddenly during the stamping process. These areas may correspond to defects such as cracks and wrinkles in the material. The purpose of detecting the fused feature values ​​in the deformation mutation areas and performing smooth transition correction is to eliminate the abnormal feature values ​​in these mutation areas, making the fused feature distribution map smoother and more continuous, while retaining the true deformation information of the material. Structural deformation features are features obtained after local feature enhancement processing and smooth transition correction. They can more accurately reflect the structural deformation of the material during the stamping process.

[0051] Edge detection algorithms, such as the Sobel operator and the Canny operator, can be used to enhance local features in the fused feature distribution map. These algorithms can detect edges and sudden changes in the fused feature distribution map and enhance the feature values ​​in these areas to make them more distinct. For example, the Sobel operator can be used to calculate the horizontal and vertical gradients of the fused feature distribution map. Regions with larger gradient values ​​are then marked as edges and sudden changes, and the feature values ​​in these regions are amplified.

[0052] Smoothing filtering algorithms, such as Gaussian filtering and median filtering, can be used to detect the fusion eigenvalues ​​of the deformation mutation area and perform smooth transition correction. These algorithms can smooth the eigenvalues ​​of the mutation area, eliminate outliers, and make the eigenvalues ​​more continuous and stable.

[0053] Finally, the fused feature distribution map after local feature enhancement and smooth transition correction is used as the structural deformation feature.

[0054] Step S220: performing action timing decomposition processing on the equipment operation status data to obtain equipment action characteristics in the process dynamic feature set; wherein the equipment action characteristics include the stamping speed change curve, the pressure transmission delay time and the mold contact surface friction coefficient.

[0055] Motion sequence decomposition processes decompose equipment operating status data into chronological order, extracting the temporal characteristics and sequence information of each device action. Equipment operating status data contains various parameters and status information of the stamping equipment during operation, such as spindle speed and hydraulic system pressure. Through this process, key information reflecting the equipment's motion characteristics can be extracted from this data. Equipment motion characteristics, part of the process dynamic feature set, describe the operation of the stamping equipment during operation, including changes in stamping speed, pressure transmission delay time, and the friction coefficient of the die contact surface. The stamping speed variation curve graphically displays the stamping speed over time, reflecting the speed changes of the equipment during the stamping process. The pressure transmission delay time is the time it takes for pressure to be applied by the equipment and transmitted to the die and material, reflecting the efficiency and response speed of the equipment's pressure transmission. The die contact surface friction coefficient is the ratio of friction to normal pressure at the die-material interface, which affects the material's forming quality and the life of the die.

[0056] Time series analysis methods such as Fourier transform and wavelet transform can be used to decompose the action sequence of equipment operation status data.

[0057] As an implementation manner, step S220 may specifically include the following steps S221 to S224:

[0058] Step S221: Obtain the spindle speed sensor data of the stamping equipment, and perform time series segmentation processing on the spindle speed sensor data to generate a stamping speed change curve; wherein the time series segmentation processing is based on synchronous alignment of the equipment action cycle and the material feeding cycle.

[0059] Spindle speed sensor data is collected by a speed sensor installed on the main shaft of the stamping equipment. This data reflects the changes in the rotational speed of the main shaft of the stamping equipment. Time series segmentation processing is to segment the spindle speed sensor data in chronological order and extract the time periods related to the equipment operation cycle and the material feed cycle. The equipment operation cycle is the time required for the stamping equipment to complete a complete stamping action, and the material feed cycle is the time interval between each feeding of the material during the stamping process. By synchronizing the time series segmentation processing based on the equipment operation cycle and the material feed cycle, the spindle speed data corresponding to each stamping action can be accurately extracted, thereby generating a stamping speed change curve.

[0060] Spindle speed sensor data from stamping equipment can be obtained by installing a speed sensor on the equipment spindle, such as a photoelectric encoder or magnetoelectric encoder. To segment the spindle speed sensor data into time series, first determine the duration of the equipment's operating cycle and the material feed cycle. This can be determined by analyzing the equipment's operating principle and operating parameters, or through actual measurements. Then, the spindle speed sensor data is segmented chronologically based on the duration of the equipment's operating cycle and the material feed cycle, with the data within each cycle treated as a separate time series. Finally, each segmented time series is processed to remove noise and outliers, resulting in the spindle speed data corresponding to each stamping action. The spindle speed data for each stamping action is then arranged chronologically to plot a stamping speed curve. For example, with time as the horizontal axis and spindle speed as the vertical axis, the spindle speed value at each time point is plotted in a coordinate system. Connecting these points yields the stamping speed curve.

[0061] Step S222: Collect the hydraulic system pressure data of the stamping equipment, and perform delay time calculation processing on the hydraulic system pressure data to obtain a pressure transmission delay time series. The delay time calculation processing is determined based on the difference between the pressure peak arrival time and the mold closing time.

[0062] Hydraulic system pressure data is collected by pressure sensors installed in the hydraulic system of the stamping equipment. This data reflects the pressure changes in the hydraulic system during the stamping process. Delay time calculation is the process of calculating the pressure transmission delay time based on the hydraulic system pressure data. Pressure transmission delay time is the time it takes for pressure to be applied by the equipment and transmitted to the mold and material, reflecting the equipment's pressure transmission efficiency and response speed. Peak pressure arrival time is when the hydraulic system pressure reaches its maximum value, and mold closing time is when the mold is fully closed. The pressure transmission delay time is calculated by calculating the difference between the peak pressure arrival time and the mold closing time.

[0063] Collecting hydraulic system pressure data for stamping equipment can be achieved by installing pressure sensors, such as strain gauge pressure sensors and piezoelectric pressure sensors, at key locations within the hydraulic system. When calculating the delay time for the hydraulic system pressure data, first identify the peak pressure arrival time from the hydraulic system pressure data. The pressure data can be analyzed to determine the moment of maximum pressure, which is the peak pressure arrival time. Next, determine the mold closing time. Position sensors or other monitoring devices installed on the mold can be used to obtain mold closing time information. Finally, calculate the difference between the peak pressure arrival time and the mold closing time to determine the pressure transmission delay time. Arrange the pressure transmission delay times corresponding to each stamping action in chronological order to create a pressure transmission delay time series.

[0064] Step S223: obtaining real-time friction coefficient data through the friction force sensor on the mold contact surface, and performing sliding window statistical processing on the real-time friction coefficient data to generate dynamic change characteristics of the friction coefficient.

[0065] Sliding window statistical processing is to divide the real-time friction coefficient data according to the set time window, perform statistical analysis on the data in each window, and extract the dynamic change characteristics of the friction coefficient. The dynamic change characteristics of the friction coefficient are used to describe the characteristics of the friction coefficient changing over time during the stamping process, which can reflect the friction characteristics and wear of the contact surface between the mold and the material. When performing sliding window statistical processing on the real-time friction coefficient data, the size and step size of the sliding window are first determined. The size of the sliding window determines the number of data points contained in each statistical window, and the step size determines the distance the window moves each time. Then, the real-time friction coefficient data is divided according to the sliding window, and the data in each window is statistically analyzed, such as calculating statistical quantities such as the mean and standard deviation. Finally, the statistics of each window are arranged in chronological order to generate the dynamic change characteristics of the friction coefficient.

[0066] Step S224: performing time sequence alignment and feature splicing processing on the stamping speed change curve, the pressure transmission delay time series, and the dynamic change characteristics of the friction coefficient to obtain the equipment action characteristics.

[0067] Specifically, step S224 may include the following steps S2241 to S2247:

[0068] Step S2241: Perform time-axis unified processing on the stamping speed change curve, determine the main time axis based on the equipment action cycle, and synchronize and compensate the starting time of the pressure transfer delay time series with the equipment start time of the main time axis to generate a time-synchronized pressure transfer delay time series.

[0069] Time axis unification processing is to unify the stamping speed change curve and the pressure transfer delay time series onto the same time axis for subsequent timing alignment processing. The equipment action cycle is the time required for the stamping equipment to complete a complete stamping action. Determining the main time axis based on the equipment action cycle allows different characteristic data to be compared and analyzed on the same time scale. Synchronous compensation processing is to adjust the starting time of the pressure transfer delay time series with the equipment start time of the main time axis to make them consistent in time. The time-synchronized pressure transfer delay time series is the pressure transfer delay time series obtained after synchronous compensation processing, which is consistent with the stamping speed change curve on the time axis.

[0070] When performing time-axis unification processing on the stamping speed variation curve, first determine the time length of the equipment action cycle. This can be done by analyzing the working principle and operating parameters of the equipment, or by actual measurement. Then, starting from the equipment startup moment, divide the main time axis according to the time length of the equipment action cycle. For the stamping speed variation curve, remark each data point on the curve according to the time scale of the main time axis to align it with the main time axis. When performing synchronous compensation processing on the starting moment of the pressure transfer delay time series and the equipment startup moment of the main time axis, first determine the time difference between the starting moment of the pressure transfer delay time series and the equipment startup moment of the main time axis. Then, based on this time difference, shift the pressure transfer delay time series so that its starting moment is aligned with the equipment startup moment of the main time axis.

[0071] Step S2242: performing sampling rate matching processing on the dynamic change characteristics of the friction coefficient, performing linear interpolation processing on the discrete friction coefficient data according to the time resolution of the main time axis, and generating a continuous friction coefficient change curve.

[0072] Sampling rate matching is the process of matching the sampling rate of the dynamic change characteristics of the friction coefficient with the time resolution of the main time axis, so that the friction coefficient data is aligned with the main time axis in time. The time resolution of the main time axis is the time interval between two adjacent time points on the main time axis, which determines the time accuracy of the data. Discrete friction coefficient data is friction coefficient data collected by a friction force sensor. These data are discrete, that is, there are only measured values ​​at certain time points. Linear interpolation is the process of inserting new data points between discrete friction coefficient data to make the data more continuous in time. The continuous friction coefficient change curve is the friction coefficient change curve obtained after sampling rate matching and linear interpolation. It is continuous in time and can more accurately reflect the changes in the friction coefficient.

[0073] To perform sampling rate matching on the dynamic characteristics of the friction coefficient, first determine the time resolution of the primary time axis. Based on the equipment's operating cycle and data acquisition requirements, determine the time interval between two consecutive time points on the primary time axis. Next, analyze the sampling rate of the discrete friction coefficient data—that is, the time interval between two consecutive measurements. If the sampling rate of the discrete friction coefficient data does not match the time resolution of the primary time axis, sampling rate matching is necessary.

[0074] A linear interpolation formula can be used to linearly interpolate discrete friction coefficient data based on the time resolution of the main time axis. Assume that two adjacent discrete friction coefficient data points (t1, f1) and (t2, f2) are known, where t1 and t2 are time points and f1 and f2 are the corresponding friction coefficient values. For any time point t between t1 and t2 on the main time axis, the corresponding friction coefficient value f can be calculated using the linear interpolation formula: f = f1 + (f2 - f1) * (t - t1) / (t2 - t1). In this way, new data points are inserted between discrete friction coefficient data points, making the data more continuous in time. The friction coefficient values ​​at each time point are arranged in chronological order to plot a continuous friction coefficient change curve.

[0075] Step S2243: Input the punching speed change curve, the time-synchronized pressure transfer delay time series and the continuous friction coefficient change curve into the multi-channel time window alignment module, perform timestamp matching processing on the data points of the three based on the sliding window mechanism, and generate the time-aligned punching speed data sequence, pressure delay data sequence and friction coefficient data sequence.

[0076] The multi-channel time window alignment module is used to perform time alignment processing on multiple time series data. It can process data from multiple channels simultaneously and perform timestamp matching on data points based on a sliding window mechanism. The sliding window mechanism slides a fixed-size window on the time series data to process and analyze the data within the window. The timestamp matching processing is to match the data points in the stamping speed change curve, the time-synchronized pressure transfer delay time series, and the continuous friction coefficient change curve in time so that they have corresponding characteristic values ​​at the same time point. The time-aligned stamping speed data series, pressure delay data series, and friction coefficient data series are data sequences obtained after timestamp matching processing. They are aligned in time, which facilitates subsequent analysis and processing.

[0077] After the stamping speed change curve, the time-synchronized pressure transfer delay time series, and the continuous friction coefficient change curve are input into the multi-channel time window alignment module, the module first determines the size and step size of the sliding window. The size of the sliding window determines the number of data points contained in each window, and the step size determines the distance the window moves each time. Then, the sliding window is moved to each time point in turn, and the stamping speed data, pressure delay data, and friction coefficient data in the window are timestamp matched. Specifically, for each time point in the window, the data point closest to the time point is found from the stamping speed change curve, the time-synchronized pressure transfer delay time series, and the continuous friction coefficient change curve, and the characteristic values ​​of these data points are used as the corresponding characteristic values ​​of the time point.

[0078] The matching results for each time point are arranged in chronological order to generate time-aligned punch speed data series, pressure delay data series, and friction coefficient data series. For example, at each time point, the matched punch speed value, pressure transmission delay time value, and friction coefficient value are recorded in sequence to form three time series.

[0079] Step S2244: performing velocity-pressure coupling relationship analysis on the time-aligned punching velocity data sequence, extracting the pressure delay eigenvalues ​​corresponding to the punching velocity extreme points in each time window, and generating coupling eigenvectors.

[0080] Specifically, the time-aligned stamping speed data sequence and pressure delay data sequence are first divided according to time windows. The size of each time window can be adjusted according to actual conditions. Then, the extreme points of the stamping speed are found in each time window. By comparing the stamping speed values ​​of each data point in the window, the points corresponding to the maximum and minimum values ​​can be found as extreme points. For each stamping speed extreme point, the corresponding pressure delay eigenvalue is extracted from the pressure delay data sequence. The pressure delay eigenvalues ​​corresponding to the stamping speed extreme points in each time window are arranged in the order of the time windows to generate a coupling eigenvector. Through the coupling eigenvector, the coupling relationship between the stamping speed and the pressure transmission delay time can be analyzed, providing a basis for monitoring and optimizing the operating status of the equipment.

[0081] Step S2245: perform feature dimension expansion processing on the coupling feature vector and the friction coefficient data sequence in the corresponding time window, splice the friction coefficient mean, fluctuation amplitude and change trend parameters to the expanded dimension of the coupling feature vector, and generate a multidimensional joint feature matrix.

[0082] Feature dimension expansion processing is to add new feature dimensions on the basis of coupled feature vectors and integrate other relevant feature information into the coupled feature vectors.

[0083] Specifically, the mean, fluctuation amplitude, and trend parameter of the friction coefficient data within the corresponding time window can be calculated. To calculate the mean, all friction coefficient data within the time window can be added together and then divided by the number of data points to obtain the average value. To calculate the fluctuation amplitude, the maximum and minimum values ​​of the friction coefficient data within the time window can be found and their difference can be calculated. To calculate the trend parameter, the friction coefficient data within the time window can be fitted using methods such as linear regression, and the slope of the fitted line can be obtained as the trend parameter.

[0084] The friction coefficient mean, fluctuation amplitude, and trend parameters are concatenated into the extended dimensions of the coupled eigenvector. Specifically, three new dimensions are added to the end of the coupled eigenvector to store the friction coefficient mean, fluctuation amplitude, and trend parameters, respectively. These calculated parameters are sequentially inserted into the corresponding dimensions to generate a multidimensional joint feature matrix.

[0085] Step S2246: Fill in the null values ​​and smooth out the abnormal points of the multi-dimensional joint feature matrix, use the adjacent time window feature interpolation method to complete the missing data and perform local weighted average filtering to generate a standardized device action feature set.

[0086] Specifically, first check whether there are null values ​​in the multidimensional joint feature matrix. If there are null values, use the adjacent time window feature interpolation method to fill them. Specifically, for a time window with a null value, find its two adjacent time windows, perform linear interpolation based on the feature data of these two time windows, and obtain an estimated value of the null value. For example, suppose there is a null value in a feature dimension of a time window, and the values ​​of its two adjacent time windows on this feature dimension are a and b respectively. The time interval between this time window and the previous time window is t1, and the time interval between this time window and the next time window is t2. Then the estimated value of the null value is: Estimated value = a + (b a) * t1 / (t1 + t2).

[0087] Outlier smoothing is performed on the multidimensional joint feature matrix using a local weighted average filtering method. For each data point, its neighborhood is determined, for example, by selecting the k data points before and after the data point as the neighborhood. Each neighboring data point is then assigned a weight based on its distance from the data point, with closer points receiving a larger weight. Finally, a weighted average is performed on the data points in the neighborhood to obtain the smoothed data value. For example, assuming a data point has n data points in its neighborhood, with values ​​x1, x2, ..., xn, and corresponding weights w1, w2, ..., wn, respectively, the smoothed data value is: Smoothed value = (w1*x1+w2*x2+...+wn*xn) / (w1+w2+...+wn).

[0088] The multidimensional joint feature matrix after null value filling and outlier smoothing is used as the standardized device action feature set.

[0089] Step S2247: Based on the equipment action stage division rule, the standardized equipment action feature set is segmented and labeled, and the feature subsets corresponding to the acceleration stage, steady-state stage and deceleration stage are dynamically weighted and the connection relationship is optimized to generate equipment action features.

[0090] The equipment action stage division rule is a rule that divides the equipment's action process into different stages based on the working principle and operating characteristics of the stamping equipment. For example, the action process of the stamping equipment can be divided into the acceleration stage, the steady-state stage, and the deceleration stage. The segmented labeling process is to mark each data point in the standardized equipment action feature set according to the equipment action stage division rule to determine the action stage to which it belongs. The dynamic weight allocation is to assign different weights to the feature subsets corresponding to each stage according to the degree of influence of different action stages on equipment performance. The connection relationship optimization is to adjust and optimize the connection relationship between the feature subsets corresponding to different stages, so that the association between the feature subsets is more reasonable and effective. The equipment action feature is a feature obtained after segmented labeling processing, dynamic weight allocation and connection relationship optimization. It can more accurately reflect the working status of the stamping equipment in different action stages.

[0091] Specifically, first determine the basis for dividing the equipment's motion phases. The motion phases can be divided based on changes in parameters such as the stamping equipment's spindle speed and pressure. For example, when the spindle speed gradually increases from 0, this phase can be labeled the acceleration phase; when the spindle speed stabilizes near a fixed value, this phase can be labeled the steady-state phase; and when the spindle speed gradually decreases from a stable value, this phase can be labeled the deceleration phase. Then, for each data point in the standardized equipment motion feature set, determine the motion phase to which it belongs based on the changes in its corresponding parameters, such as spindle speed and pressure, and label them accordingly.

[0092] Dynamic weight assignment and connection optimization are performed on the feature subsets corresponding to the acceleration phase, steady-state phase, and deceleration phase. For dynamic weight assignment, different weights can be assigned to the feature subsets corresponding to each phase based on the degree of impact of different action phases on device performance. For example, in the acceleration phase, the acceleration and speed changes of the device have a greater impact on device performance, so a larger weight can be assigned to the feature subset of the acceleration phase; in the steady-state phase, the stability and accuracy of the device have a greater impact on product quality, so an appropriate weight can be assigned to the feature subset of the steady-state phase; in the deceleration phase, the braking performance and accuracy of the device have a greater impact on production safety and efficiency, so a corresponding weight can be assigned to the feature subset of the deceleration phase.

[0093] To optimize connectivity, methods such as graph neural networks can be used to model and optimize the connectivity between feature subsets corresponding to different phases. Specifically, feature subsets corresponding to the acceleration phase, steady-state phase, and deceleration phase are used as nodes in the graph, and edges between the nodes are constructed based on the temporal order and physical connections between the different phases. The graph-structured data is then fed into the graph neural network for training. By optimizing the network parameters, the connectivity between the nodes becomes more reasonable and effective. During the training process of the graph neural network, supervised or unsupervised learning methods can be used. If labeled training data is available, supervised learning methods can be used, using the labeled information as the training objective. The network parameters are optimized by minimizing the error between the predicted results and the labeled information. If labeled training data is not available, unsupervised learning methods can be used to optimize the network parameters by learning the intrinsic structure and feature representation of the data. For example, unsupervised learning models such as autoencoders can be used to encode and decode the graph-structured data, allowing the network to learn the underlying feature representation of the data and thus optimize the connectivity between nodes.

[0094] After dynamic weight assignment and connection optimization, the feature subsets corresponding to each stage are combined according to the optimized connection relationships to generate the device action signature. Specifically, for each feature dimension, the feature values ​​corresponding to different stages are weighted and summed according to the assigned weights to obtain the final feature value for that feature dimension. The final feature values ​​of all feature dimensions are combined to form the device action signature.

[0095] Step S230: Perform energy transfer analysis on the dynamic response data of the forming process to obtain energy transfer characteristics in the process dynamic feature set; wherein the energy transfer characteristics include the material plastic deformation energy consumption rate, the equipment kinetic energy loss rate and the heat energy loss distribution.

[0096] The dynamic response data of the forming process refers to various data that reflect the dynamic changes of materials and equipment during the metal stamping process, such as material deformation, equipment vibration, energy transfer, etc.

[0097] As an implementation manner, step S230 may specifically include the following steps S231 to S234:

[0098] Step S231: Calculating the material plastic deformation energy consumption rate based on the power sensor data of the stamping equipment, wherein the plastic deformation energy consumption rate is determined by the ratio of the material absorbed energy to the total input energy per unit time.

[0099] Specifically, the total input power Ptotal is first obtained from the power sensor data, and then the energy absorbed by the material per unit time, Pabsorb, is calculated. To calculate the energy absorbed by the material, the material deformation and mechanical properties and other parameters can be measured and calculated in combination with the material's mechanical properties model. For example, in the stamping process of a metal sheet, strain gauges and other equipment can be used to measure the strain change of the sheet per unit time. Based on the stress-strain curve and deformation volume of the sheet, the energy absorbed by the material per unit time can be calculated. Specifically, assuming that the stress-strain curve of the sheet can be represented by a function σ(ε), where σ is stress, ε is strain, and the deformation volume of the sheet is V, the energy absorbed by the material can be calculated by integration: Pabsorb = ∫σ(ε)dε*V, where the integration interval is the strain change range of the sheet per unit time.

[0100] Finally, divide the material absorption energy Pabs by the total input energy Ptotal per unit time to obtain the material plastic deformation energy consumption rate: Energy consumption rate = Pabs / Ptotal.

[0101] Step S232: collecting kinetic energy loss data of the stamping equipment transmission system, and performing frequency domain decomposition processing on the kinetic energy loss data to obtain the frequency spectrum distribution characteristics of the equipment kinetic energy loss rate.

[0102] Kinetic energy loss data for stamping equipment transmission systems is related to the kinetic energy loss caused by various factors such as friction and vibration during operation. Data such as vibration acceleration and torque changes are included. Frequency domain decomposition converts time-domain kinetic energy loss data into frequency domain data and analyzes the energy distribution of different frequency components. Frequency domain decomposition is used to determine the spectral distribution of the equipment's kinetic energy loss rate, reflecting the distribution of the equipment's kinetic energy loss rate at different frequencies.

[0103] Frequency domain analysis methods such as Fourier transform can be used to perform frequency domain decomposition processing on kinetic energy loss data. Based on the results of Fourier transform, the spectrum distribution characteristics of the kinetic energy loss rate of the equipment are calculated. Specifically, for each frequency component, the proportion of its corresponding energy to the total energy is calculated to obtain the kinetic energy loss rate at that frequency. The kinetic energy loss rates at different frequencies are arranged in order of frequency to obtain the spectrum distribution characteristics of the kinetic energy loss rate of the equipment. For example, the spectrum distribution characteristics are plotted into a spectrum diagram, where the horizontal axis represents the frequency and the vertical axis represents the kinetic energy loss rate. The kinetic energy loss of the equipment at different frequencies can be intuitively observed through the spectrum diagram. By analyzing the spectrum distribution characteristics of the kinetic energy loss rate of the equipment, the frequency components with large energy loss in the transmission system can be found, providing a basis for equipment maintenance and optimization.

[0104] Step S233: obtaining temperature distribution data of the contact surface between the mold and the material through an infrared thermal imager, and performing heat conduction simulation processing on the temperature distribution data to generate a heat energy loss distribution map.

[0105] Numerical simulation methods such as finite element analysis can be used to simulate heat conduction on temperature distribution data. Specifically, this method discretizes a continuous physical system into a finite number of elements. Heat conduction can be simulated by solving the heat conduction equation. Specifically, a heat conduction model of the mold and material is first established, including defining the material's thermal conductivity performance parameters (such as thermal conductivity and specific heat capacity) and boundary conditions (such as the contact surface temperature between the mold and the material and the ambient temperature). The temperature distribution data is then input into the heat conduction model as initial conditions, and the heat conduction equation is solved using finite element software to obtain the temperature distribution at different times and locations. Based on the results of the heat conduction simulation, a heat dissipation distribution map is generated. Color mapping can be used to assign different temperature values ​​to different colors. The temperature distribution image, thus the heat dissipation distribution map, can then be plotted in two or three-dimensional space.

[0106] Step S234: performing space-time joint coding processing on the spectrum distribution characteristics of the plastic deformation energy consumption rate, the equipment kinetic energy loss rate, and the heat energy loss distribution diagram to obtain energy transfer characteristics.

[0107] As an implementation manner, step S234 may specifically include the following steps S2341 to S2347:

[0108] Step S2341: performing time discretization processing on the plastic deformation energy consumption rate, dividing the continuous energy consumption rate data into equally spaced energy consumption rate time series based on the time nodes of the stamping forming stage, and generating a discretized plastic deformation energy consumption rate set.

[0109] The plastic deformation energy consumption rate (PER) is the ratio of the energy consumed by the material undergoing plastic deformation to the total energy input during the metal stamping process. It is a continuously varying quantity. Time discretization involves segmenting the continuous PER data into set time intervals, converting it into a discrete time series. The time nodes of the stamping process are divided into several stages based on the characteristics and requirements of the stamping process. The start and end times of each stage are the time nodes. The discretized PER set, obtained after time discretization, consists of a series of discrete PER values. During time discretization, the time nodes of the stamping stage are first determined. This can be based on the specific conditions of the stamping process. For example, the time nodes can be determined based on factors such as the stamping equipment's operating cycle and the material's deformation process. Continuous PER data is then acquired. The energy input and material deformation during the stamping process can be measured in real time using devices such as power sensors to calculate the continuous PER. For each time node ti, the corresponding PER value Pi is extracted from the continuous PER data. Arrange these energy consumption rate values ​​in chronological order to form a discretized plastic deformation energy consumption rate set {P0, P1, P2, ..., Pn-1} of length n.

[0110] Through time discretization processing, the continuous plastic deformation energy consumption rate data is converted into discrete time series, which is convenient for subsequent processing and analysis.

[0111] Step S2342: performing spatial mapping processing on the spectrum distribution characteristics of the equipment kinetic energy loss rate, mapping the main frequency band energy value of the spectrum characteristics to the three-dimensional spatial coordinate points of the stamping equipment transmission system, and generating a kinetic energy loss spatial distribution diagram.

[0112] The spectral distribution characteristics of the equipment's kinetic energy loss rate are derived by performing frequency domain decomposition on the kinetic energy loss data of the stamping equipment's transmission system. These characteristics reflect the distribution of the equipment's kinetic energy loss rate at different frequencies. Spatial mapping involves mapping the energy values ​​of the main frequency bands in the spectral distribution characteristics to the three-dimensional spatial coordinates of the stamping equipment's transmission system, assigning the energy values ​​to corresponding spatial locations.

[0113] Specifically, first determine the three-dimensional spatial coordinate system of the stamping equipment transmission system. You can use a fixed point of the transmission system as the origin to establish a three-dimensional rectangular coordinate system and determine the position coordinates of each component in the coordinate system. Then, analyze the spectrum distribution characteristics of the equipment's kinetic energy loss rate to find the main frequency band energy value. The main frequency band is the frequency range where energy is mainly concentrated in the spectrum, and the main frequency band energy value is the sum of the energy within this frequency range. For each main frequency band energy value, determine the transmission system component to which the energy value corresponds based on its corresponding frequency component and the working principle of the transmission system. For example, if the frequency component corresponding to a main frequency band energy value is related to the rotation frequency of a gear, then the component corresponding to the energy value is the gear.

[0114] Map the mainband energy values ​​to the corresponding 3D coordinate points. For each component, assign its corresponding mainband energy value to the coordinate point of the component in 3D space. You can use color mapping or numerical annotation to visualize the energy values.

[0115] The energy values ​​of all coordinate points are integrated to generate a spatial distribution map of kinetic energy loss. Three-dimensional drawing software or visualization tools can be used to plot the coordinate points and their corresponding energy values ​​to create a kinetic energy loss spatial distribution map. This kinetic energy loss spatial distribution map allows for intuitive visualization of kinetic energy loss at different locations in the transmission system, providing a basis for equipment maintenance and optimization. For example, if the map reveals a component with high kinetic energy loss, it may indicate wear or failure, requiring prompt inspection and replacement.

[0116] Step S2343: performing time axis resampling processing on the heat energy loss distribution map, adjusting the time resolution of the temperature data according to the time nodes of the discretized plastic deformation energy consumption rate set, and generating a heat energy loss distribution map after time unification.

[0117] The heat energy loss distribution diagram is obtained through equipment such as infrared thermal imagers. It reflects the distribution of heat energy loss at the contact surface between the mold and the material and the surrounding space, and contains temperature distribution information at different time points. The time axis resampling process resamples the temperature data in the heat energy loss distribution diagram on the time axis and adjusts the time resolution of the data to make it consistent with the time nodes of the discretized plastic deformation energy consumption rate set. The time nodes of the discretized plastic deformation energy consumption rate set are the time points determined when the plastic deformation energy consumption rate is time discretized. These time points divide the stamping process into several stages. The heat energy loss distribution diagram after time unification is obtained after the time axis resampling process. The time resolution of the temperature data is consistent with the time nodes of the discretized plastic deformation energy consumption rate set.

[0118] When resampling the heat dissipation distribution graph over time, first determine the time nodes of the discretized plastic deformation energy consumption rate set. Then, analyze the temporal resolution of the temperature data in the heat dissipation distribution graph, that is, the time interval between two adjacent time points. If the temporal resolution of the temperature data is inconsistent with the time nodes of the discretized plastic deformation energy consumption rate set, resampling is required. For each time node ti of the discretized plastic deformation energy consumption rate set, find the two closest time points tj and tj+1 from the heat dissipation distribution graph, where tj ≤ ti ≤ tj+1. Based on the temperature data at these two time points, use interpolation to calculate the temperature distribution corresponding to time node ti. For example, using linear interpolation, assuming that the temperature distributions corresponding to time points tj and tj+1 are Tj and Tj+1, respectively, the temperature distribution Ti corresponding to time node ti can be calculated using the following formula: Ti = Tj + (Ti - Tj) * (ti - tj) / (tj+1 - tj). The temperature distribution corresponding to the time node of each discretized plastic deformation energy consumption rate set is arranged in chronological order to generate a time-unified heat energy loss distribution diagram.

[0119] Step S2344: Input the discretized plastic deformation energy consumption rate set, the kinetic energy loss spatial distribution map, and the time-unified heat energy loss distribution map into the spatiotemporal coding structure, extract the spatial dimension features through the convolution layer, and capture the temporal dimension correlation through the recurrent layer to generate a spatiotemporal joint coding feature tensor.

[0120] In the spatiotemporal coding architecture, convolutional layers can be used to extract spatial dimensional features from the kinetic energy loss spatial distribution map and the time-normalized thermal energy dissipation distribution map. For example, a two-dimensional convolutional layer is used to convolve the kinetic energy loss spatial distribution map and the thermal energy dissipation distribution map. Different convolution kernels are used to extract local features at different scales and directions. It is understood that the convolution kernel size, step size, and padding can be set according to actual needs. Each convolution kernel slides over the input data, calculating the convolution result and generating a series of feature maps. These feature maps reflect the characteristic information of the input data at different locations and scales.

[0121] In the spatiotemporal coding structure, the recurrent layer can be used to process the discretized set of plastic deformation energy consumption rates. Through the iterative operation of the recurrent unit, the time series pattern of the data can be learned. Examples of recurrent layers include long short-term memory networks (LSTMs) and gated recurrent units (GRUs).

[0122] The spatial features extracted by the convolutional layer and the temporal correlations captured by the recurrent layer are fused to generate a spatiotemporal joint encoding feature tensor. Specifically, the feature maps output by the convolutional layer and the hidden states output by the recurrent layer are concatenated or weighted summed to produce a comprehensive feature representation. This feature representation incorporates both the spatial and temporal correlations of the input data, providing a more comprehensive picture of energy transfer during the stamping process.

[0123] Step S2345: Perform feature channel fusion processing on the spatiotemporal joint coding feature tensor, perform cross-channel weighted superposition on the feature maps of the plastic deformation energy consumption channel, the kinetic energy loss channel, and the heat energy dissipation channel to generate a preliminary fused energy feature map.

[0124] Specifically, the weight of each channel is first determined, and the distribution of weights can be adjusted according to the importance of different channels to energy transfer. For example, the weights of the plastic deformation energy consumption channel, the kinetic energy loss channel, and the heat energy dissipation channel can be determined through experiments or experience to be w1, w2, and w3, respectively, and w1+w2+w3=1. These weights represent the importance of each channel in the fusion process. Then, the feature map of each channel is weighted. For the feature map F1 of the plastic deformation energy consumption channel, the feature map F2 of the kinetic energy loss channel, and the feature map F3 of the heat energy dissipation channel, they are multiplied by the corresponding weights w1, w2, and w3, respectively, to obtain the weighted feature maps w1*F1, w2*F2, and w3*F3.

[0125] The weighted feature maps are superimposed across channels. Specifically, w1*F1, w2*F2, and w3*F3 are added across the channel dimension to obtain a preliminary fused energy feature map F. This means F = w1*F1 + w2*F2 + w3*F3. This cross-channel weighted superposition method fuses feature maps from different channels to produce a preliminary fused energy feature map that contains multiple energy transfer information.

[0126] The preliminary fused energy signature map can more comprehensively reflect the energy transfer during the stamping process, providing richer characteristic information for subsequent energy distribution analysis and optimization. For example, the preliminary fused energy signature map can more clearly observe the energy distribution in different regions and at different time points, as well as the interrelationships between different energy transfer channels. Further analysis of the preliminary fused energy signature map can identify areas and stages with significant energy loss, providing a basis for optimizing the stamping process and improving energy utilization efficiency.

[0127] Step S2346: Based on the energy absorption characteristics of the stamping material, the preliminary fusion energy characteristic map is subjected to regional feature enhancement processing, the fusion characteristics of the high energy density area are detected and edge smoothing correction is performed to generate an optimized energy distribution feature.

[0128] The energy absorption characteristics of stamping materials are the ability of materials to absorb and dissipate energy during the stamping process. Different stamping materials have different energy absorption characteristics. Regional feature enhancement processing is to carry out targeted feature enhancement of different areas in the preliminary fusion energy characteristic map based on the energy absorption characteristics of the stamping material, highlighting those important features related to the energy absorption characteristics of the material. High energy density areas are areas with relatively high energy values ​​in the preliminary fusion energy characteristic map. These areas may correspond to areas where deformation of the material is concentrated, high load areas of the equipment, etc. Edge smoothing correction processing is to smooth the edge parts of the high energy density area to avoid sudden changes in characteristic values ​​and make the energy distribution characteristics more continuous and natural. The optimized energy distribution characteristics are the energy distribution characteristics obtained after regional feature enhancement processing and edge smoothing correction. It can more accurately reflect the energy distribution related to the energy absorption characteristics of the material during the stamping process.

[0129] The regional feature enhancement process can specifically begin by analyzing the energy absorption characteristics of the stamping material. Experiments or theoretical calculations can be used to understand the material's energy absorption patterns under different deformation conditions, such as the material's stress-strain curve and energy absorption coefficient. Based on the material's energy absorption characteristics, the feature regions requiring enhancement are identified. For example, if the material has strong energy absorption in certain deformation regions, the corresponding features in these regions are enhanced. High energy density regions are detected within the preliminary fused energy feature map. Methods such as threshold segmentation can be used to mark regions with energy values ​​above a certain threshold as high energy density regions. For detected high energy density regions, their fused features are enhanced. For example, these features can be highlighted by increasing their feature values ​​or enhancing their contrast. The edges of high energy density regions are smoothed and corrected. Smoothing filtering algorithms, such as Gaussian filtering or median filtering, can be used to smooth the feature values ​​of edge regions. These filtering algorithms adjust the current feature values ​​based on the feature values ​​within the neighborhood, making the feature values ​​in edge regions change more gradually.

[0130] After regional feature enhancement and edge smoothing, an optimized energy distribution feature is generated. This optimized energy distribution feature can more accurately reflect the energy distribution related to the material's energy absorption characteristics during the stamping process, providing a more targeted basis for stamping process optimization and material selection.

[0131] Step S2347: Match and verify the optimized energy distribution characteristics with the energy conservation constraints of the equipment operation, eliminate the characteristic areas that violate the law of conservation of energy, and fill in the missing characteristic values ​​to generate energy transfer characteristics.

[0132] Specifically, the total input energy, Etotal, can be calculated based on the device's input power and operating time. Then, from the optimized energy distribution, energy information such as the material's plastic deformation energy consumption, Eplastic, the device's kinetic energy loss, Ekinetic, and thermal energy dissipation, Ether, can be extracted. These energy values ​​are summed to obtain the calculated total energy, Etotal = Eplastic + Ekinetic + Ether.

[0133] Compare Etotal and Emeter. If the difference between Emeter and Etotal exceeds a preset error range, it is considered that a characteristic region violates the law of conservation of energy. These violating regions require further analysis and processing. This can be done to check for errors in the data acquisition and processing process, or for unaccounted energy loss factors. If the characteristic region itself is identified as a problem, it is removed from the optimized energy distribution signature.

[0134] Missing eigenvalues ​​in the optimized energy distribution features are completed. Interpolation methods such as linear interpolation and spline interpolation can be used to estimate missing eigenvalues ​​based on the eigenvalues ​​of the surrounding areas. For example, for a region with missing eigenvalues, several adjacent regions with known eigenvalues ​​are found. Based on the eigenvalues ​​and spatial position relationships of these regions, the estimated value of the missing eigenvalue is calculated using the linear interpolation formula.

[0135] After matching verification, eliminating violation areas, and filling in missing values, an energy transfer signature is generated. This energy transfer signature accurately reflects the energy transfer during the stamping process, providing a reliable basis for optimizing the stamping process and improving equipment performance.

[0136] Step S240: performing feature dimension unification processing on the structural deformation features, the device motion features, and the energy transfer features, so that the dimensions of each feature are consistent in both the time scale and the space scale.

[0137] The unification of feature dimensions is to process these different types of features so that they have the same dimension and data format in terms of time scale and space scale, so as to facilitate subsequent comprehensive analysis and processing. Specifically, the time scale is first unified, and the time resolution and time range of the structural deformation features, equipment motion features, and energy transfer features are analyzed to determine a unified time scale. For example, the equipment motion cycle can be selected as a unified time scale, and the various feature data can be resampled and aligned according to the equipment motion cycle. For structural deformation features, their time series are interpolated or sampled according to the equipment motion cycle so that their time resolution is consistent with the equipment motion cycle. For equipment motion features and energy transfer features, the time axis is also adjusted so that they are aligned with the structural deformation features in time.

[0138] Secondly, the spatial scale is unified. The spatial resolution and spatial range of the structural deformation characteristics, equipment motion characteristics, and energy transfer characteristics are analyzed to determine a unified spatial scale. An appropriate spatial grid division rule can be determined based on the size and shape of the stamping material, as well as the structure and layout of the equipment. The individual feature data are projected and interpolated according to the unified spatial grid division rule, so that they have the same spatial resolution and coordinate system. For example, the thickness change rate distribution cloud map and residual stress distribution pattern feature vectors in the structural deformation characteristics are projected onto the unified spatial grid nodes; the stamping velocity change curve and pressure transfer delay time series in the equipment motion characteristics are divided and interpolated according to spatial position to align with the unified spatial grid; the kinetic energy loss spatial distribution map and heat dissipation distribution map in the energy transfer characteristics are similarly adjusted and interpolated to align with the unified spatial scale.

[0139] After unifying the temporal and spatial scales, the dimensions of each feature are processed. The dimensions of the structural deformation features, equipment motion features, and energy transfer features are examined, and features of different dimensions are non-dimensionalized to make them comparable. Dimensionless processing can be achieved through normalization and standardization. For example, for structural deformation features such as the thickness change rate and yield strength gradient, normalization can be used to adjust their numerical range to the [0, 1] range. For equipment motion features such as the stamping speed and pressure transfer delay time, standardization can be used to convert them to data with a mean of 0 and a standard deviation of 1.

[0140] Step S300: constructing a hybrid feature space based on the material structure feature set and the process dynamic feature set, and performing multi-dimensional feature matching processing in the hybrid feature space to generate a feature mapping relationship network.

[0141] A hybrid feature space integrates features from both the material structure feature set and the process dynamic feature set, creating a high-dimensional space containing diverse feature information. Multidimensional feature matching analyzes and matches the correlations and mapping relationships between different features within the hybrid feature space to identify the inherent connections between them. The feature mapping relationship network, derived through multidimensional feature matching, reflects the mapping relationships between different features and can be used to describe the interactions and influences between material structure features and process dynamic features during the stamping process.

[0142] Specifically, the features in the material structure feature set and the process dynamic feature set are first merged to form a feature vector containing all the features. For example, the structural deformation features, equipment motion features and energy transfer features are spliced ​​together to obtain a high-dimensional feature vector. Then, according to the nature and characteristics of the features, the feature vectors are dimensionally divided and organized to form the structure of a mixed feature space. There are many methods that can be used to perform multi-dimensional feature matching processing in a mixed feature space, such as correlation analysis, cluster analysis, neural networks, etc. Taking correlation analysis as an example, the correlation coefficient between different features in the mixed feature space can be calculated to find feature pairs with strong correlation. The correlation coefficient can be calculated using methods such as the Pearson correlation coefficient and the Spearman correlation coefficient. For feature pairs with strong correlation, it can be considered that there is a mapping relationship between them.

[0143] Based on the results of multi-dimensional feature matching, a feature mapping relationship network is generated. Each feature in the hybrid feature space can be treated as a node in the network, and the mapping relationships between features as edges. Edge weights can be set based on the strength of the correlation between features; the stronger the correlation, the greater the edge weight. This feature mapping relationship network can intuitively demonstrate the relationships between different features in the stamping process, providing a basis for subsequent dynamic twin model construction and stamping process optimization.

[0144] As an implementation manner, step S300 may specifically include the following steps S310 to S340:

[0145] Step S310: performing feature cross-correlation processing on the structural deformation features in the material structure feature set and the equipment motion features in the process dynamic feature set to generate a first-level feature correlation matrix.

[0146] Feature cross-correlation processing combines and analyzes structural deformation features and equipment motion features to identify their correlations and associations. The first-level feature correlation matrix is ​​obtained through feature cross-correlation processing and reflects the associations between structural deformation features and equipment motion features.

[0147] Specifically, first determine the characteristic vectors of the structural deformation characteristics and the equipment action characteristics. The structural deformation characteristic vector can be composed of characteristics such as the thickness change rate, the yield strength change gradient, and the residual stress distribution pattern. The equipment action characteristic vector can be composed of characteristics such as the stamping speed change curve, the pressure transmission delay time, and the mold contact surface friction coefficient. Then, calculate the correlation between the structural deformation characteristic vector and the equipment action characteristic vector. Correlation analysis methods such as the Pearson correlation coefficient and the Spearman correlation coefficient can be used to calculate the correlation coefficient between each structural deformation feature and each equipment action feature. The calculated correlation coefficients are combined into a matrix, namely the first-level feature correlation matrix. The rows of the matrix represent the structural deformation characteristics, the columns represent the equipment action characteristics, and each element in the matrix represents the correlation coefficient between the corresponding structural deformation characteristics and the equipment action characteristics.

[0148] Step S320: performing spatiotemporal alignment processing on the residual stress distribution pattern feature vector in the material structure feature set and the energy transfer feature in the process dynamic feature set to generate a second-level feature correlation matrix.

[0149] The residual stress distribution pattern eigenvectors within the material structure feature set reflect the distribution of residual stress within the stamped material after forming. The energy transfer features within the process dynamic feature set describe the energy conversion and transfer during the stamping process. Spatiotemporal alignment aligns the residual stress distribution pattern eigenvectors and energy transfer features in time and space, ensuring they have corresponding eigenvalues ​​at the same time point and spatial location. The second-level feature correlation matrix, derived through spatiotemporal alignment and correlation analysis, reflects the correlation between the residual stress distribution pattern eigenvectors and the energy transfer features.

[0150] Specifically, the time scales of the residual stress distribution pattern eigenvectors and the energy transfer features are first unified. The time series of the residual stress distribution pattern eigenvectors and the energy transfer features can be resampled and aligned according to the time nodes of the stamping process so that they have the same resolution and range in time. The spatial scales of the residual stress distribution pattern eigenvectors and the energy transfer features are unified. According to the size and shape of the stamping material, as well as the structure and layout of the equipment, a unified spatial grid division rule is determined. The residual stress distribution pattern eigenvectors and the energy transfer features are projected and interpolated according to the unified spatial grid division rule so that they have the same resolution and coordinate system in space. For example, the residual stress distribution pattern in the residual stress distribution pattern eigenvector is projected onto the unified spatial grid node, and the kinetic energy loss spatial distribution map and the heat energy dissipation distribution map in the energy transfer feature are also adjusted and interpolated in spatial coordinates so that they are aligned with the unified spatial grid.

[0151] After completing the spatiotemporal alignment, the correlation between the residual stress distribution pattern eigenvectors and the energy transfer features is calculated. For example, using a correlation analysis method, the calculated correlation coefficients are combined into a matrix, the second-level feature correlation matrix. The rows of the matrix represent the residual stress distribution pattern features, and the columns represent the energy transfer features. Each element in the matrix represents the correlation coefficient between the corresponding residual stress distribution pattern feature and the energy transfer feature.

[0152] The correlation between the residual stress distribution pattern eigenvectors and the energy transfer characteristics can be analyzed through the second-level characteristic correlation matrix.

[0153] Step S330: performing matrix fusion processing on the first-level feature correlation matrix and the second-level feature correlation matrix to obtain a hybrid feature correlation topological structure.

[0154] As an implementation manner, step S330 may specifically include the following steps S331 to S334:

[0155] Step S331: performing singular value decomposition on the first-level feature correlation matrix to obtain a first eigenvector set and a first singular value distribution.

[0156] Numerical methods such as the Jacobi method and QR decomposition can be used to perform singular value decomposition on the first-level eigencorrelation matrix. Taking the Jacobi method as an example, the matrix is ​​gradually converted to a diagonal matrix through an iterative process to obtain singular values ​​and eigenvectors. First, the matrices U and V are initialized to the identity matrix, and Σ is the first-level eigencorrelation matrix A. Then, in each iteration, the off-diagonal element with the largest absolute value in matrix Σ is selected and transformed using a rotation matrix, reducing this off-diagonal element to zero. Simultaneously, the U and V matrices are updated. This process is repeated until matrix Σ is approximately diagonal. The resulting column vectors of the U matrix are the set of first eigenvectors, and the diagonal elements of the Σ matrix are the first singular value distributions. The set of first eigenvectors reflects the primary eigendirections of the first-level eigencorrelation matrix, and the first singular value distribution reflects the importance of these eigendirections. The larger the singular value, the greater the role played by the corresponding eigenvector in the matrix.

[0157] Step S332: performing principal component analysis on the second-level feature correlation matrix to obtain a second eigenvector set and principal component contribution rates.

[0158] Specifically, the second-level feature association matrix is ​​first standardized so that its mean is 0 and its standard deviation is 1, so as to eliminate the influence of dimensions between different features and improve the accuracy of the analysis. Then, the covariance matrix of the standardized second-level feature association matrix is ​​calculated. The covariance matrix reflects the correlation between different features. Next, the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalue represents the size of the variance explained by each principal component, and the eigenvector represents the direction of the principal component. The eigenvalues ​​and eigenvectors can be solved using methods such as eigenvalue decomposition or singular value decomposition. The eigenvector with a larger eigenvalue is selected as the principal component to obtain the second eigenvector set. Through principal component analysis, the information of the second-level feature association matrix can be compressed and extracted to obtain a few important principal components.

[0159] Step S333: performing orthogonal projection processing on the first eigenvector set and the second eigenvector set to generate a joint eigenvector space.

[0160] Specifically, a common spatial dimension is first determined. Based on the dimensions of the first and second eigenvector sets, an appropriate spatial dimension can be selected so that both eigenvector sets can be projected into this space. Then, the projection matrix of the first and second eigenvector sets onto the common spatial dimension is calculated. The projection matrix can be obtained by calculating the inner product between the first and second eigenvector sets and the basis vectors of the common spatial dimension.

[0161] The projected first eigenvector set and the second eigenvector set are merged to generate a joint eigenvector space. In the joint eigenvector space, the vectors are orthogonal to each other, which makes the different feature directions independent of each other, facilitating the subsequent weighted fusion process.

[0162] Step S334: performing weighted fusion processing on the joint feature vector space based on the first singular value distribution and the principal component contribution rate to obtain a hybrid feature association topology structure.

[0163] Specifically, the first singular value distribution and the principal component contribution rate are normalized so that the sum of their weights is 1. , calculate the normalized singular value weights ; For the principal component contribution rate , calculate the normalized principal component weights .

[0164] Then, according to the normalized weights, weights are assigned to the projection vectors in the joint eigenvector space. The projection vectors in the joint eigenvector space include the projection vectors of the first eigenvector set and the projection vector of the second eigenvector set . For the projection vector of the first eigenvector set , whose weight is ; For the projection vector of the second eigenvector set , whose weight is .

[0165] The weighted projection vectors are combined. The mixed feature association topology can be obtained by concatenating the weighted projection vectors into a matrix by column.

[0166] This weighted fusion process comprehensively considers the importance of the relationships between structural deformation characteristics and equipment motion characteristics, as well as the residual stress distribution pattern eigenvectors and energy transfer characteristics. This allows the hybrid feature association topology to more accurately reflect the complex relationships between different features. This topological structure provides a key foundation for the subsequent construction of a feature mapping relationship network, facilitating in-depth analysis of the interactions and influences between material structural characteristics and process dynamics during the stamping process.

[0167] Step S340: constructing a dynamic weight distribution network based on the hybrid feature association topology structure, and optimizing the connection relationship of feature nodes in the hybrid feature space through the dynamic weight distribution network to generate a feature mapping relationship network.

[0168] The hybrid feature association topology describes the complex relationships between material structural features and process dynamics, providing the foundation for constructing a dynamic weight allocation network. This network dynamically adjusts the connection weights between feature nodes based on the degree of correlation between different features and the actual situation. Feature nodes represent individual features in the hybrid feature space. Connection optimization adjusts the connection weights between feature nodes to make the mapping relationships between features more reasonable and effective. The feature mapping network, derived from connection optimization, accurately reflects the mapping relationships between different features.

[0169] A neural network architecture can be used to construct a dynamic weight allocation network based on a hybrid feature association topology. First, the hybrid feature association topology serves as the network's input layer, with the nodes in the input layer corresponding to the feature nodes in the hybrid feature space. Next, a hidden layer is added. The nodes in the hidden layer can be configured based on the actual situation. In the hidden layer, each node is connected to the nodes in the input layer via connection weights, which can be initialized to random values.

[0170] To achieve dynamic weight distribution, a weight adjustment mechanism is introduced into the network. Adaptive learning algorithms, such as gradient descent and the Adam algorithm, can be used to continuously adjust connection weights based on the network's output and the objective function. The objective function can be designed based on specific application requirements. For example, the goal can be to minimize the mapping error between feature nodes, allowing the network to learn more accurate feature mapping relationships.

[0171] A dynamic weight allocation network optimizes the connections between feature nodes in a hybrid feature space. During network training, data on the hybrid feature association topology is continuously input, the network's output error is calculated based on the objective function, and an adaptive learning algorithm is used to adjust the connection weights. As training progresses, the connection weights gradually converge to an optimal value, resulting in more reasonable connections between feature nodes.

[0172] After connection optimization is complete, a feature mapping network is generated. This network can be represented as a graph structure, where nodes represent feature nodes in the hybrid feature space, edges represent the connections between feature nodes, and edge weights represent the strength of the connections. This network intuitively displays the mapping relationships between different features, providing an important basis for the subsequent construction of dynamic twin models and real-time state deduction of the stamping process.

[0173] Step S400: constructing a dynamic twin model according to the feature mapping relationship network, and performing real-time state deduction of the stamping process based on the dynamic twin model to generate forming quality prediction results and process defect location results.

[0174] The feature mapping relationship network reflects the accurate mapping relationship between the material structural characteristics and the process dynamic characteristics during the stamping process. Based on this network, a dynamic twin model can be constructed. The dynamic twin model is a virtual model that can simulate the stamping forming process in real time. By continuously updating the model parameters, it keeps in sync with the actual stamping process. Real-time state deduction uses the dynamic twin model to predict the state of the stamping process at future moments based on the current stamping state and feature mapping relationship. The forming quality prediction result is obtained through real-time state deduction, which predicts the forming quality of the stamped product, such as the thickness deviation of the material, the internal stress distribution, etc. The process defect location result is obtained by real-time state deduction to find the location and cause of possible process defects in the stamping process.

[0175] As an implementation manner, step S400 may specifically include the following steps S410 to S460:

[0176] Step S410: converting the node connection relationship in the feature mapping relationship network into a dynamic differential equation group, determining the coupling coefficient matrix of the equation group based on the feature transfer path between the nodes, and generating an initial state transfer model.

[0177] The nodes in the feature mapping network represent the various features of the stamping process, while the edges represent the connections between features, reflecting the interactions and influences between them. A dynamic differential equation system is a mathematical model used to describe the dynamic changes in a system. It uses a set of differential equations to represent the temporal changes in the variables in the system. The coupling coefficient matrix is ​​a matrix of coefficients between the equations in the dynamic differential equation system, reflecting the degree of coupling between the different variables. The initial state transition model is obtained by converting the feature mapping network into a dynamic differential equation system. It is used as an initial model to describe the state transitions in the stamping process.

[0178] Specifically, a variable is first defined for each node in the feature mapping network. These variables represent various characteristics of the stamping process, such as material thickness, stamping speed, and pressure. Then, based on the connections between the nodes, differential equations between these variables are established. For each node, the rate of change of its variable can be expressed as a function of the variables of the other nodes connected to it.

[0179] The coupling coefficient matrix of the system of equations is determined based on the characteristic transfer paths between nodes. The characteristic transfer path is the order and method by which characteristics are transferred between nodes, which determines the degree of coupling between different variables. For each equation in the dynamic differential equation system, the coefficients between it and other equations are determined based on the characteristic transfer path.

[0180] Combining the dynamic differential equations and the coupling coefficient matrix, we get the initial state transfer model. The initial state transfer model can be expressed as ,in is a variable vector, and A is a coupling coefficient matrix. The model describes the temporal relationship between various characteristics during the stamping process, providing a basis for subsequent real-time state deduction.

[0181] Step S420: collecting real-time pressure data and material deformation data of the stamping equipment in the current operation stage, inputting the real-time pressure data and material deformation data into the observation variable interface in the initial state transfer model, and generating a model input parameter set.

[0182] The observation variable interface is used in the initial state transition model to receive external observation data. By inputting real-time pressure data and material deformation data into this interface, actual stamping state information can be introduced into the model. The model input parameter set is obtained by inputting real-time pressure data and material deformation data into the initial state transition model and is used to drive the model's state deduction.

[0183] The real-time pressure data and material deformation data are input into the observation variable interface of the initial state transfer model. The observation variable interface in the initial state transfer model is, for example, a vector. The real-time pressure data and material deformation data are combined in sequence into a vector and input into the observation variable interface. For example, if the observation variable interface of the initial state transfer model requires a three-dimensional vector, where the first element represents pressure and the second and third elements represent the thickness change and strain of the material, the collected real-time pressure data, material thickness change data, and strain data are combined in this order into a three-dimensional vector and input into the observation variable interface.

[0184] Step S430: calling the dynamic parameter correction algorithm to perform online adjustment on the coupling coefficient matrix of the initial state transition model, calculating the parameter correction amount based on the residual between the real-time data stream and the model prediction value, and generating an updated state transition model.

[0185] When using the dynamic parameter correction algorithm to perform online adjustments to the coupling coefficient matrix of the initial state transfer model, you can first obtain the real-time data stream and the model prediction value. During the stamping process, real-time data is continuously collected, and the initial state transfer model is used to predict the current input parameters to obtain the model prediction value. The residual between the real-time data stream and the model prediction value is calculated. Various methods can be used to calculate the residual, such as mean square error (MSE) and absolute error.

[0186] As an implementation manner, step S430 may specifically include the following steps S431 to S436:

[0187] Step S431: Obtain the material thickness prediction value output by the initial state transfer model and the real-time thickness value measured by the actual sensor, and calculate the residual sequence between the two within the continuous time window.

[0188] The material thickness prediction output by the initial state transfer model is the value predicted by the model based on the current input parameters and coupling coefficient matrix for the material thickness at a future time. The real-time thickness value measured by the actual sensor is the actual material thickness value at the current moment, collected in real time by a thickness sensor installed on the stamping die, such as a laser thickness sensor or ultrasonic thickness sensor. The residual sequence is the difference between the predicted and measured material thickness values ​​within a continuous time window, reflecting the degree of discrepancy between the model's predicted and actual values. To calculate the residual sequence, the predicted and measured material thickness values ​​are acquired sequentially in chronological order within the continuous time window, and the difference between them is calculated. By calculating the residual sequence, the discrepancy between the model's predicted and actual values ​​can be visually observed. If the residual sequence is large and fluctuates frequently, the model's prediction accuracy is low and the model parameters need to be adjusted. If the residual sequence is small and stable, the model's prediction accuracy is high.

[0189] Step S432: Perform autocorrelation analysis on the residual sequence to identify the periodic characteristics of the residual fluctuation and the location of abnormal mutation points, and generate a residual characteristic analysis report.

[0190] Specifically, we first calculate the autocorrelation function of the residual sequence. The autocorrelation function represents the correlation of the residual sequence at different time delays and can be calculated using the following formula: , where r i is the i-th element of the residual sequence, is the mean of the residual series and k is the time lag.

[0191] Then, based on the results of the autocorrelation function, the periodic characteristics of the residual fluctuations are identified. By observing the autocorrelation function curve, the time delays k at which peaks appear in the curve are found. The time delays corresponding to these peaks are the periods of the residual fluctuations. At the same time, by observing and analyzing the residual sequence, the locations of abnormal mutation points can be identified. A threshold can be set; when the absolute value of an element in the residual sequence exceeds this threshold, the point is considered an abnormal mutation point.

[0192] The residual characteristic analysis report includes information such as the periodic characteristics of residual fluctuations and the location of abnormal mutation points, and provides analysis and interpretation of this information. For example, for periodic characteristics, the cause can be analyzed to determine whether it is related to factors such as the stamping equipment's duty cycle and the characteristics of the material. For the location of abnormal mutation points, the relevant data of the stamping process at that point in time can be further examined to identify possible causes of the failure.

[0193] Step S433: adjusting the update frequency of the coupling coefficient matrix based on the periodic characteristics in the residual characteristic analysis report, and triggering an emergency correction mechanism at the abnormal mutation point.

[0194] The periodic characteristics in the residual characteristic analysis report reflect the cyclical variation of the model's prediction error. By adjusting the update frequency of the coupling coefficient matrix, the model can better adapt to this cyclical variation and improve its prediction accuracy. Abnormal mutation points correspond to abnormal conditions in the stamping process. Triggering the emergency correction mechanism at these locations allows timely adjustment of model parameters to avoid excessive prediction errors caused by abnormal conditions.

[0195] Specifically, the update period of the coupling coefficient matrix is ​​first determined based on the periodic characteristics. If the residual sequence has obvious periodic fluctuations with a period of T, the update period of the coupling coefficient matrix can be set to an integer multiple of T. This ensures that the model can update the parameters within each period to adapt to the periodic changes in the prediction error.

[0196] Then, the execution frequency of the dynamic parameter correction algorithm is adjusted according to the update cycle. The dynamic parameter correction algorithm is used to calculate the correction amount of the coupling coefficient matrix. Its execution frequency is adjusted to be consistent with the update cycle, that is, the dynamic parameter correction algorithm is executed once in each update cycle.

[0197] An emergency correction mechanism is triggered at the location of an abnormal mutation point. When an abnormal mutation point is detected in the residual sequence, the current stamping process is immediately suspended and the emergency correction mechanism is triggered. The emergency correction mechanism may include the following steps: First, a detailed analysis of the real-time data before and after the abnormal mutation point is performed to identify the cause of the abnormality, such as equipment failure or material defects. Then, based on the analysis results, the parameters of the coupling coefficient matrix are manually adjusted, or a more aggressive dynamic parameter correction algorithm is used, such as increasing the correction step size and changing the correction strategy, to quickly adjust the model parameters and reduce the prediction error.

[0198] Step S434: using the recursive least squares method to iteratively optimize the dynamic weight parameters in the coupling coefficient matrix, and adjusting the step size in each iteration according to the current residual value and the historical correction amount calculation parameter.

[0199] The recursive least squares method is an algorithm used to estimate model parameters online. It continuously updates the estimated parameters to minimize the sum of squared residuals. The dynamic weight parameters in the coupling coefficient matrix are the values ​​of each element in the matrix. These parameters determine the degree of coupling between different variables in the model. Iterative optimization involves gradually adjusting the values ​​of the dynamic weight parameters over multiple iterations to continuously reduce the model's prediction error. The parameter adjustment step size is the adjustment amplitude of the dynamic weight parameters in each iteration, which determines the speed and stability of the parameter update.

[0200] Specifically, the dynamic weight parameters and related matrices are initialized first. The dynamic weight parameters in the coupling coefficient matrix are initialized to an initial value, and the matrices required for the recursive least squares method, such as the gain matrix k and the covariance matrix P, are also initialized.

[0201] In each iteration, the parameter adjustment step size is calculated based on the current residual value and the historical correction amount. The current residual value is the difference between the predicted material thickness and the actual measured value at the current time point. The historical correction amount is the adjustment amount of the dynamic weight parameter in the previous iteration. The parameter adjustment step size can be calculated using the following formula: ,in is the parameter adjustment step size, K is the gain matrix, and r is the current residual value. The gain matrix K can be updated according to the covariance matrix P and the input data. The specific update formula is: ,in is the input data vector, is the forgetting factor, which is used to control the weight of historical data.

[0202] Update the dynamic weight parameters according to the parameter adjustment step size. Add the current dynamic weight parameters to the parameter adjustment step size to obtain the updated dynamic weight parameters. The covariance matrix P reflects the uncertainty of the parameter estimation and needs to be updated in each iteration based on the input data and the parameter adjustment step size. The update formula is: , where I is the identity matrix.

[0203] Repeat the above steps until the iteration termination condition is met. The iteration termination condition can be set according to the actual situation, such as the sum of squares of the residuals is less than a certain threshold, the number of iterations reaches the maximum limit, etc.

[0204] Step S435: Substitute the optimized dynamic weight parameters into the dynamic differential equation group to resolve the material deformation propagation equation and generate a corrected state transition trajectory.

[0205] Specifically, the optimized dynamic weight parameters are first updated to the coupling coefficient matrix. The elements in the coupling coefficient matrix are the dynamic weight parameters. The optimized parameter values ​​replace the original parameter values ​​to obtain the updated coupling coefficient matrix A'.

[0206] Substitute the updated coupling coefficient matrix A' into the dynamic differential equation system , where x is a vector representing various state variables in the stamping process, such as material thickness, strain, pressure, etc. The material deformation propagation equation is part of the dynamic differential equation system and can be extracted from the equation system.

[0207] Select an appropriate numerical solution method to solve the material deformation propagation equation, such as the Euler method, Runge-Kutta method, etc. Taking the Euler method as an example, for the material deformation propagation equation , where x is the material deformation variable and t is time. The iterative formula of the Euler method is: , where x nis the material deformation value at the nth time step, h is the time step, t n is the time of the nth time step.

[0208] Starting from the initial state, the numerical solution is iterated according to the formula to calculate the material deformation value at each time step. These values ​​are arranged in chronological order to obtain the corrected state transition trajectory. The corrected state transition trajectory can more accurately reflect the actual state changes of the material during the stamping process.

[0209] Step S436: Verify the matching degree between the corrected state transition trajectory and the actual measurement data. When the residual decrease rate of S consecutive iterations is less than the preset convergence threshold, the correction process is terminated and an updated state transition model is generated, S≥3.

[0210] Verifying the degree of match between the corrected state transition trajectory and the actual measured data is to evaluate whether the model can more accurately reflect the actual state of the stamping process after parameter correction. The residual is the difference between the predicted value in the corrected state transition trajectory and the actual measured data, and the residual descent rate is the reduction ratio of the residual in two adjacent iterations. The preset convergence threshold is a pre-set value used to determine whether the model has converged. When the residual descent rate for 3 or more consecutive iterations is less than the preset convergence threshold, it means that the parameters of the model have basically stabilized and the correction process can be terminated. The updated state transition model is obtained after parameter correction and verification, and can more accurately reflect the state transition of the stamping process.

[0211] Specifically, the actual measurement data is obtained first. During the stamping process, the state data of the material, such as thickness, strain, pressure, etc., are collected in real time through various sensors, and these data are used as actual measurement data. The predicted value in the corrected state transition trajectory is compared with the actual measurement data to calculate the residual. Different methods can be used to calculate the residual, such as mean square error, absolute error, etc. After each iteration, the residual descent rate is calculated to determine whether the residual descent rate for three consecutive iterations is less than the preset convergence threshold. If the residual descent rate for three consecutive iterations is less than the preset convergence threshold, the model is considered to have converged and the correction process is terminated.

[0212] Step S440: Simulate the deformation propagation path of the stamping material in the next forming stage through the updated state transition model, extract the material thickness change prediction surface and internal stress diffusion trajectory, and generate a deformation process simulation data set.

[0213] To simulate the deformation propagation path of the stamping material in the next forming stage, the updated state transition model is used to predict the deformation of the material in the future forming stage based on the current material state and process parameters. Specifically, the current material state and process parameters are first determined. The current material state includes information such as the thickness, strain, and stress of the material, and the process parameters include information such as the stamping speed, pressure, and mold shape. This information is used as the input of the updated state transition model. The model is solved based on the structure and parameters of the updated state transition model. Numerical solution methods, such as the Runge-Kutta method and the finite difference method, can be used to solve the state changes of the model in the future forming stage. During the solution process, the thickness changes and internal stress distribution of the material at different positions and times are recorded.

[0214] Based on the recorded data, a material thickness change prediction surface and internal stress diffusion trajectory are generated. For the material thickness change prediction surface, the material thickness values ​​at different locations and times are used as the surface height to plot a 3D surface graph. For the internal stress diffusion trajectory, the magnitude and direction of the internal stress are used as vectors to plot the stress diffusion trajectory in 3D space. By integrating these information, a dataset for the deformation process simulation is generated.

[0215] Step S450: Compare the maximum thickness deviation value in the deformation process simulation data set with the preset forming tolerance threshold. If the deviation value exceeds the threshold, it is marked as a deformation exceeding standard area, and a forming quality prediction result is generated.

[0216] Specifically, material thickness data is first extracted from the deformation process simulation dataset. Material thickness data can be stored in a three-dimensional array or matrix, where each element represents the material thickness at a specific location and time. The deviation between the material thickness at each location and the target thickness is calculated. The target thickness is the thickness that the material should achieve based on the stamping process design requirements. The deviation values ​​at all locations are traversed to find the maximum value. The maximum thickness deviation value is compared with the preset forming tolerance threshold. If the maximum thickness deviation value exceeds the threshold, it is considered that there is an area of ​​excessive deformation. In the three-dimensional material space, locations where the thickness deviation value exceeds the threshold are identified and marked as areas of excessive deformation. Excessive deformation areas can be marked using different colors or symbols for intuitive display. The forming quality prediction results can be presented in the form of a report or image. The report should include information such as the presence of excessive deformation areas, their location, and their extent. The presence of excessive deformation areas indicates potential problems with the forming quality of the stamped product.

[0217] Step S460: Identify stress concentration areas based on the curvature variation characteristics of the internal stress diffusion trajectory, perform similarity matching processing on the stress concentration areas and crack initiation patterns in the historical defect database, and generate process defect location results.

[0218] Specifically, the curvature of the internal stress diffusion trajectory can be calculated first. Stress concentration areas can be identified based on the curvature value. Areas with larger curvature values ​​indicate higher stress concentration. A curvature threshold can be set. When the curvature value of a point on the trajectory exceeds the threshold, the area at that point is marked as a stress concentration area. Similarity matching is then performed between the stress concentration area and crack initiation patterns in the historical defect database. Feature extraction and matching algorithms, such as stress concentration area descriptors (e.g., maximum principal stress gradient amplitude, stress direction dispersion, curvature radius), and machine learning algorithms can be used to compare the characteristics of the stress concentration area with those of the crack initiation patterns in the historical defect database. A similarity score is calculated (e.g., using a dynamic time warping (DTW) algorithm to calculate trajectory similarity), and the pattern with the highest similarity score is identified. Based on the similarity matching results, a process defect localization result is generated. The process defect localization result should include information such as the possible process defect type, defect location, and range. For example, if the crack initiation pattern with the highest similarity is an edge crack, it can be determined that an edge crack is likely to occur in the stress concentration area, and the specific location and range of the crack can be determined. Through the results of process defect location, possible process defects in the stamping process can be discovered in time, and corresponding measures can be taken to prevent and repair them, thereby improving the quality of stamping products.

[0219] Step S500: generating a stamping process parameter adjustment plan based on the forming quality prediction result and the process defect location result, and feeding back the stamping process parameter adjustment plan to the stamping control system to trigger a parameter optimization operation.

[0220] The forming quality prediction results include information on the assessment of the forming quality of the stamped product, such as whether there are areas with excessive deformation, as well as the location and range of these areas. The process defect location results clearly identify the location and type of process defects that may occur during the stamping process.

[0221] As an implementation manner, step S500 may specifically include the following steps S510 to S540:

[0222] Step S510: determining a stamping speed adjustment coefficient and a pressure compensation amount according to the deformation exceeding standard area information in the forming quality prediction result.

[0223] The excessive deformation area information in the forming quality prediction results includes information such as the location, range, and thickness deviation of the excessive area. The stamping speed adjustment coefficient is used to adjust the stamping speed of the stamping equipment. By changing this coefficient, the stamping speed can be adjusted. The pressure compensation amount is the additional pressure applied to the stamping equipment to compensate for pressure changes caused by excessive deformation.

[0224] Specifically, first analyze the thickness deviation and distribution of areas where deformation exceeds the specified value. If the thickness deviation is large and concentrated, it indicates that the material is deforming unevenly during the stamping process, which may be caused by excessive stamping speed or unreasonable pressure distribution.

[0225] Based on the analysis results, determine the stamping speed adjustment factor. If the deformation exceeds the standard due to excessive stamping speed, the stamping speed can be appropriately reduced. The stamping speed adjustment factor can be determined based on the thickness deviation value and the range of the exceeding standard area.

[0226] The pressure compensation amount can be calculated based on the material properties and thickness deviation of the area with excessive deformation. The greater the thickness deviation, the greater the pressure required for compensation. Through experiments or simulations, a relationship model between thickness deviation and pressure compensation can be established, and the pressure compensation amount calculated based on the model. The determined stamping speed adjustment coefficient and pressure compensation amount are recorded as part of the stamping process parameter adjustment plan. By adjusting the stamping speed and pressure compensation amount, the material deformation uniformity can be improved, the occurrence of excessive deformation areas can be reduced, and the quality of the stamped product can be improved.

[0227] Step S520: Based on the coordinates of the stress concentration area in the process defect location result, a mold contact surface optimization solution and a material feed path correction vector are generated.

[0228] Specifically, first determine the location and characteristics of the stress concentration area and then determine how to optimize the mold contact surface. If the stress concentration area is located at an edge or corner of the mold, the mold surface in that area can be ground, polished, or rounded to improve stress distribution. If the stress concentration area is caused by an unreasonable mold shape or structure, the mold can be redesigned or modified.

[0229] The mold contact surface optimization plan includes information such as specific optimization measures, optimized locations, and expected effects after optimization. For example, if it is decided to grind a certain edge of the mold, the plan should state the grinding range, the grinding roughness requirements, and the expected improvement in stress distribution after grinding. Determine the material feed path correction vector based on the location of the stress concentration area. If the stress concentration area is located at a certain position of the material, the material can be adjusted to avoid the area by adjusting the feed direction and distance of the material. The material feed path correction vector can be expressed as a two-dimensional or three-dimensional vector, where the direction of the vector indicates the direction in which the material needs to be adjusted, and the length of the vector indicates the distance the material needs to be adjusted.

[0230] Step S530: performing multi-objective optimization processing on the stamping speed adjustment coefficient, the pressure compensation amount, the die contact surface optimization scheme, and the material feed path correction vector to generate a candidate parameter adjustment set.

[0231] Specifically, the optimization objectives are first determined. These objectives may include improving the quality of stamped products, reducing production costs, and increasing production efficiency. For example, improving the quality of stamped products can be achieved by reducing areas with excessive deformation and process defects; reducing production costs can be achieved by reducing mold wear and material waste; and improving production efficiency can be achieved by shortening the stamping cycle.

[0232] Then, assign a weight to each optimization goal. The weight represents the importance of each goal in the optimization process and can be adjusted according to the actual situation. For example, if you pay more attention to product quality, you can assign a larger weight to the goal of improving product quality.

[0233] Next, an optimization model is established, which can be based on mathematical programming models, genetic algorithms, simulated annealing algorithms, and other methods. Taking the mathematical programming model as an example, the stamping speed adjustment coefficient, pressure compensation amount, die contact surface optimization plan, and material feed path correction vector can be used as decision variables, and the optimization goal as the objective function, while considering various constraints such as equipment performance limitations and process feasibility.

[0234] An optimization algorithm is used to solve the optimization model. Based on the selected optimization algorithm, the optimization model is solved to obtain an optimal set of parameter combinations. During the solution process, the algorithm continuously searches and compares different parameter combinations to find the solution that strikes the best balance between multiple objectives. The resulting optimal parameter combination and some similar parameter combinations are considered as a set of candidate parameter adjustments. Each solution in the candidate parameter adjustment set strikes a balance between different objectives and can be selected and applied based on the actual situation.

[0235] Step S540: Perform feasibility screening on the candidate parameter adjustment set using a preset process stability evaluation model to obtain a stamping process parameter adjustment plan.

[0236] The process stability assessment model can be a model based on a machine learning algorithm, such as a neural network model, a support vector machine model, etc., or a model based on a physical model, such as a finite element model. The feasibility of each solution is judged based on the output of the process stability assessment model. The output of the model can be an evaluation indicator, such as a process stability score, a defect incidence rate, etc. A feasibility threshold can be set. When the evaluation indicator of the solution exceeds the threshold, the solution is considered feasible. The solutions whose evaluation indicators exceed the feasibility threshold in the candidate parameter adjustment set are screened out to form a set of feasible solutions. The optimal solution is selected from the set of feasible solutions as the stamping process parameter adjustment solution. Based on the actual situation, the solution with the best evaluation indicator can be selected, or the most appropriate solution can be selected by comprehensively considering other factors such as cost and efficiency.

[0237] It is understandable that the various commonly known algorithms involved in the above-mentioned introductions of the embodiments of the present invention, such as Euclidean distance, cosine distance, Kriging interpolation algorithm, genetic algorithm, simulated annealing algorithm, etc., can all be obtained from the relevant content in the prior art. Their calculation formulas or methods are known. In order to save space, they are not expanded too much in the embodiments of the present invention. In addition, when implementing the scheme of the present invention, those skilled in the art can supplement the details according to the common knowledge in this field. For example, according to the common knowledge in this field, normalization can be used to eliminate dimensional conflicts before feature fusion, interpolation can be used to eliminate dimensional differences, thresholds can be reasonably set in combination with historical data or actual needs, and models can be trained based on a general model training method, etc. The present invention no longer redundantly introduces the overly detailed implementation process.

[0238] See also Figure 2 , Figure 2This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 may be connected via a bus or other means. The processor 101 (also known as the Central Processing Unit (CPU)) is the computing and control core of the computer system, capable of parsing various instructions within the computer system and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, a mobile communication interface, etc.), which can be used to send and receive data under the control of the processor 101. The communication interface 102 may also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system for storing programs and data. It is understood that the memory 103 herein may include both the built-in memory of the computer system and, of course, the extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system, but this is not limited to this in the present invention.

[0239] In one embodiment, the processor 101 executes the digital twin construction method for metal stamping provided in the above embodiment of the present invention by running the computer program in the memory 103.

Claims

1. A digital twin construction method for metal stamping, characterized in that: include: Acquire a multi-source monitoring data set during a metal stamping forming process, wherein the multi-source monitoring data set includes material physical parameter data, equipment operating status data, and forming process dynamic response data; Performing collaborative feature extraction processing on the multi-source monitoring data set to generate a material structure feature set and a process dynamic feature set, wherein the material structure feature set characterizes the deformation characteristics and stress distribution characteristics of the stamping material, and the process dynamic feature set characterizes the action timing characteristics and energy transfer characteristics of the stamping equipment; Performing feature cross-correlation processing on the structural deformation features in the material structure feature set and the equipment motion features in the process dynamic feature set to generate a first-level feature correlation matrix; Performing spatiotemporal alignment processing on the residual stress distribution pattern feature vector in the material structure feature set and the energy transfer feature in the process dynamic feature set to generate a second-level feature correlation matrix; Performing matrix fusion processing on the first-level feature correlation matrix and the second-level feature correlation matrix to obtain a hybrid feature correlation topological structure; Constructing a dynamic weight distribution network based on the hybrid feature association topology structure, and optimizing the connection relationship of feature nodes in the hybrid feature space through the dynamic weight distribution network to generate a feature mapping relationship network; Converting the node connection relationship in the feature mapping relationship network into a dynamic differential equation group, determining the coupling coefficient matrix of the equation group based on the feature transfer path between the nodes, and generating an initial state transfer model; Collecting real-time pressure data and material deformation data of the stamping equipment during the current operation phase, inputting the real-time pressure data and the material deformation data into the observation variable interface of the initial state transfer model to generate a model input parameter set; Invoking a dynamic parameter correction algorithm to perform online adjustment on the coupling coefficient matrix of the initial state transition model, calculating parameter corrections based on residuals between real-time data streams and model prediction values, and generating an updated state transition model; The updated state transition model is used to simulate the deformation propagation path of the stamping material in the next forming stage, extract the material thickness change prediction surface and the internal stress diffusion trajectory, and generate a deformation process simulation data set; Comparing the maximum thickness deviation value in the deformation process simulation data set with a preset forming tolerance threshold, if the deviation value exceeds the threshold, marking it as a deformation exceeding tolerance area, and generating a forming quality prediction result; Identifying stress concentration areas based on curvature variation characteristics of the internal stress diffusion trajectory, performing similarity matching processing on the stress concentration areas and crack initiation patterns in a historical defect database to generate process defect location results; A stamping process parameter adjustment plan is generated according to the forming quality prediction result and the process defect positioning result, and the stamping process parameter adjustment plan is fed back to the stamping control system to trigger a parameter optimization operation.

2. The method according to claim 1, characterized in that The collaborative feature extraction process is performed on the multi-source monitoring data set to generate a material structure feature set and a process dynamic feature set, including: Performing structural deformation feature extraction processing on the material physical parameter data to obtain structural deformation features in the material structural feature set; wherein the structural deformation features include material thickness change rate, yield strength change gradient, and residual stress distribution pattern; Performing action time sequence decomposition processing on the equipment operation status data to obtain equipment action features in the process dynamic feature set; wherein the equipment action features include a stamping speed change curve, a pressure transmission delay time, and a mold contact surface friction coefficient; Performing energy transfer analysis on the dynamic response data of the forming process to obtain energy transfer features in the process dynamic feature set; wherein the energy transfer features include the material plastic deformation energy consumption rate, the equipment kinetic energy loss rate, and the heat energy dissipation distribution; The structural deformation characteristics, the device motion characteristics and the energy transfer characteristics are processed in a unified manner so that the characteristics maintain dimensional consistency in time and space scales.

3. The method according to claim 2, characterized in that The performing structural deformation feature extraction processing on the material physical parameter data to obtain the structural deformation features in the material structural feature set includes: Collecting real-time thickness measurement data of the stamping material during the forming process, and performing spatial interpolation processing on the real-time thickness measurement data to generate a thickness change rate distribution cloud map; Acquiring yield strength test data of the stamping material at different forming stages, and performing gradient calculation processing on the yield strength test data to generate a yield strength change gradient curve; Collecting material surface stress distribution data based on a preset residual stress detection device, and performing pattern recognition processing on the stress distribution data to generate a residual stress distribution pattern feature vector; The thickness change rate distribution cloud map, the yield strength change gradient curve and the residual stress distribution mode characteristic vector are normalized and fused to obtain the structural deformation characteristics.

4. The method according to claim 2, characterized in that The performing action time sequence decomposition processing on the equipment operation status data to obtain the equipment action features in the process dynamic feature set includes: Obtaining spindle speed sensor data of a stamping device and performing time series segmentation processing on the spindle speed sensor data to generate a stamping speed variation curve; wherein the time series segmentation processing is based on synchronous alignment of the device operation cycle and the material feeding cycle; Collecting hydraulic system pressure data of the stamping equipment and performing delay time calculation processing on the hydraulic system pressure data to obtain a pressure transmission delay time series; wherein the delay time calculation processing is determined based on the difference between the pressure peak arrival time and the mold closing time; Acquire real-time friction coefficient data through a friction force sensor on the mold contact surface, and perform sliding window statistical processing on the real-time friction coefficient data to generate a dynamic change feature of the friction coefficient; The stamping speed change curve, the pressure transmission delay time series and the dynamic change characteristics of the friction coefficient are subjected to time sequence alignment and feature splicing processing to obtain the equipment action characteristics.

5. The method according to claim 2, characterized in that The performing of energy transfer analysis on the dynamic response data of the forming process to obtain the energy transfer features in the process dynamic feature set includes: Calculating the material plastic deformation energy consumption rate based on the power sensor data of the stamping equipment, wherein the plastic deformation energy consumption rate is determined by the ratio of the material absorbed energy to the total input energy per unit time; Collecting kinetic energy loss data of the stamping equipment transmission system and performing frequency domain decomposition processing on the kinetic energy loss data to obtain the frequency spectrum distribution characteristics of the equipment kinetic energy loss rate; Acquire the temperature distribution data of the contact surface between the mold and the material through an infrared thermal imager, and perform heat conduction simulation processing on the temperature distribution data to generate a heat energy loss distribution map; The energy transfer characteristics are obtained by performing space-time joint coding processing on the plastic deformation energy consumption rate, the spectrum distribution characteristics of the equipment kinetic energy loss rate, and the heat energy loss distribution diagram.

6. The method according to claim 1, characterized in that The performing matrix fusion processing on the first-level feature association matrix and the second-level feature association matrix to obtain a hybrid feature association topological structure includes: Performing singular value decomposition on the first-level feature correlation matrix to obtain a first eigenvector set and a first singular value distribution; Performing principal component analysis on the second-level feature correlation matrix to obtain a second eigenvector set and a principal component contribution rate; Performing orthogonal projection processing on the first eigenvector set and the second eigenvector set to generate a joint eigenvector space; The joint feature vector space is weightedly fused based on the first singular value distribution and the principal component contribution rate to obtain the hybrid feature association topology structure.

7. The method according to claim 1, characterized in that The calling of the dynamic parameter correction algorithm to adjust the coupling coefficient matrix of the initial state transition model online, calculating the parameter correction amount based on the residual between the real-time data stream and the model prediction value, and generating an updated state transition model includes: Obtaining the material thickness prediction value output by the initial state transfer model and the real-time thickness value measured by the actual sensor, and calculating the residual sequence between the two within a continuous time window; Performing autocorrelation analysis on the residual sequence to identify the periodic characteristics of residual fluctuations and the location of abnormal mutation points, and generating a residual characteristic analysis report; Adjusting the update frequency of the coupling coefficient matrix based on the periodic characteristics in the residual characteristic analysis report, and triggering an emergency correction mechanism at the abnormal mutation point; The dynamic weight parameters in the coupling coefficient matrix are iteratively optimized using a recursive least squares method, and the step size is adjusted in each iteration based on the current residual value and the historical correction amount calculation parameter; Substituting the optimized dynamic weight parameters into the dynamic differential equation group to resolve the material deformation propagation equation and generate a revised state transition trajectory; Verify the matching degree between the corrected state transition trajectory and the actual measurement data, terminate the correction process when the residual decrease rate of S consecutive iterations is less than a preset convergence threshold, and generate the updated state transition model, S≥3.

8. A computer system, characterized in that: include: a memory, wherein the computer program is stored in the memory; A processor, configured to load the computer program to implement the method for constructing a digital twin for metal stamping as described in any one of claims 1 to 7.

Images