Method and device for correlating and predicting process parameters of cold roll forming and mechanical properties of products

Abstract

The application provides a cold bending forming process parameter and product mechanical property correlation prediction method and device, relates to the artificial intelligence technical field, and is based on material parameters to decouple forming process parameters, can separate the influence of material scale on geometric deformation. Based on the forming process parameters and the material parameters, the plastic history index before the start of each pass is determined, the previous cumulative plastic deformation intensity can be quantified, and the invisible history path is converted into a numerical feature. Further, the corresponding process parameters and material parameters are organized into a multi-pass cold bending forming sequence, which can retain the evolution law and time sequence dependence of plastic accumulation. Finally, the prediction model can have the physical decoupling and history perception ability based on the input features, distinguish samples with the same geometry but different histories, and give a physically self-consistent residual stress prediction, realizing accurate correlation prediction from process parameters to product mechanical properties.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for predicting the correlation between cold bending forming process parameters and product mechanical properties. Background Technology

[0002] Cold bending is a highly efficient manufacturing process that continuously bends metal strip into a specific cross-sectional shape using multiple rolling passes. It is widely used in the manufacture of automotive structural components, building profiles, and rail transit parts. In actual production, the core evaluation indicators for forming quality include the springback of the formed section, the distribution of residual stress, and changes in the material's yield strength and hardness. These mechanical responses not only determine the final dimensional accuracy and assembly performance of the product but also directly affect the fatigue life and service safety of the components.

[0003] Because cold bending involves a complex alternating process of multi-pass progressive bending and elastic-plastic loading and unloading of the material, the mechanical state after forming is highly dependent on the combination of process parameters for each pass, the constitutive properties of the material, and the plastic history accumulated from previous passes. How to quickly and accurately predict the mechanical properties after forming based solely on material parameters and pass process parameters during the process design stage is a technical problem that needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for predicting the correlation between cold bending forming process parameters and product mechanical properties, so as to predict the mechanical properties (such as yield strength, tensile strength, hardness, springback and cross-sectional residual stress distribution) of the formed product before actual production, based on its material parameters and designed multi-pass forming process parameters.

[0005] In a first aspect, embodiments of the present invention provide a method for predicting the correlation between cold bending forming process parameters and product mechanical properties. The method includes: obtaining material parameters of the sample to be predicted, and forming process parameters for multiple passes of forming the sample; scaling and decoupling the geometric processing parameters of the forming process parameters based on the material yield strength and material thickness parameters in the material parameters to obtain decoupled geometric processing parameters corresponding to each pass; determining the plasticity history index of the sample to be predicted before each pass based on the pass spacing, bending angle, and material thickness in the forming process parameters; constructing a multi-pass cold bending forming sequence for the sample to be predicted based on the plasticity history index, decoupled geometric processing parameters, and material parameters; inputting the multi-pass cold bending forming sequence into a pre-trained mechanical response prediction model, and using the mechanical response prediction model to predict the mechanical properties of the multi-pass cold bending forming sequence to determine the predicted mechanical properties of the sample after forming.

[0006] In conjunction with the first aspect, the present invention also provides a first implementation of the first aspect, wherein the step of constructing a multi-pass cold bending forming sequence of the sample to be predicted based on the plasticity history index, decoupled geometric processing parameters, and material parameters includes: constructing a historical position vector corresponding to each pass based on the plasticity history index, plasticity history increment, and the relative position of the corresponding pass in the entire forming line; splicing the decoupled geometric processing parameters, the historical position vector, and the material parameters to form a geometrically decoupled pass feature vector corresponding to each pass; and stacking the geometrically decoupled pass feature vectors of all passes in pass order to form a multi-pass cold bending forming sequence of the sample to be predicted.

[0007] In conjunction with the first aspect, the present invention also provides a second implementation of the first aspect, wherein the step of determining the plastic history index of the sample to be predicted before the start of each pass, based on the pass spacing and bending angle in the molding process parameters and the plate thickness in the material parameters, includes: determining the equivalent bending radius of the current pass according to the pass spacing and bending angle of the corresponding pass; determining the pass plastic increment of the pass according to the plate thickness and the equivalent bending radius; and weighted summing the pass plastic increments of all passes before the current pass to determine the plastic history index before the start of the current pass.

[0008] In conjunction with the first aspect, this invention also provides a third implementation of the first aspect, wherein the step of inputting a multi-pass cold bending sequence into a pre-trained mechanical response prediction model, and using the mechanical response prediction model to predict the mechanical properties of the multi-pass cold bending sequence to determine the post-forming mechanical property prediction result of the sample to be predicted includes: determining the gating coefficient vector of the mechanical response prediction model based on the plastic history strength corresponding to the multi-pass cold bending sequence; determining the features to be predicted corresponding to the multi-pass cold bending sequence based on the gating coefficient vector; inputting the features to be predicted into the mechanical response prediction model, and performing branch prediction on the features to be predicted through the scalar mechanical response output branch and the residual stress output branch of the mechanical response prediction model to determine the corresponding scalar mechanical response result and residual stress prediction result; and determining the post-forming mechanical property prediction result of the sample to be predicted based on the scalar mechanical response result and the residual stress prediction result.

[0009] In conjunction with the first aspect, this embodiment of the invention also provides a fourth implementation of the first aspect, wherein the step of determining the post-molding mechanical property prediction result of the sample to be predicted based on the scalar mechanical response result and the residual stress prediction result includes: determining a first predicted value corresponding to the scalar mechanical response result based on the training label statistics corresponding to the scalar mechanical response result; determining a second predicted value corresponding to the residual stress prediction result based on the training label statistics corresponding to the residual stress prediction result; and determining the first predicted value and the second predicted value as the post-molding mechanical property prediction result of the sample to be predicted.

[0010] In conjunction with the first aspect, the present invention also provides a fifth implementation of the first aspect, wherein the method further includes: acquiring pre-constructed training samples; wherein the training samples include sample labels, the sample labels including scalar mechanical response labels and residual stress labels; training the mechanical response prediction model based on the training samples and their sample labels, and constructing scalar mechanical response output branches and residual stress output branches corresponding to the training samples; wherein the total loss used for training the mechanical response prediction model is constructed based on scalar mechanical response loss, residual stress loss, and the cross-sectional equilibrium constraints and residual stress smoothing regularization terms corresponding to the residual stress.

[0011] In conjunction with the first aspect, the present invention also provides a sixth implementation of the first aspect, wherein the method further includes: determining a local nearest neighbor set corresponding to a multi-pass cold bending forming sequence from a preset pass reference sample; determining a center projection vector and a perturbation scale vector corresponding to the multi-pass cold bending forming sequence based on the local nearest neighbor set; constructing enhanced input features corresponding to the multi-pass cold bending forming sequence based on the center projection vector; or, adding random perturbation to the enhanced input features in combination with the perturbation scale vector to construct enhanced input features after feature perturbation; and constructing training samples based on the enhanced input features or the enhanced input features after feature perturbation to train the mechanical response prediction model.

[0012] Secondly, embodiments of the present invention also provide a device for predicting the correlation between cold bending forming process parameters and product mechanical properties. The device includes: a parameter acquisition module for acquiring material parameters of the sample to be predicted, and forming process parameters for multiple passes of forming the sample; a first processing module for scaling and decoupling the geometric processing parameters of the forming process parameters based on the material yield strength and material thickness parameters in the material parameters, to obtain decoupled geometric processing parameters corresponding to each pass; a second processing module for determining the plasticity history index of the sample to be predicted before each pass based on the pass spacing, bending angle, and plate thickness in the forming process parameters; a data construction module for constructing a multi-pass cold bending forming sequence of the sample to be predicted based on the plasticity history index, decoupled geometric processing parameters, and material parameters; and an execution module for inputting the multi-pass cold bending forming sequence into a pre-trained mechanical response prediction model, predicting the mechanical properties of the multi-pass cold bending forming sequence through the mechanical response prediction model, and determining the predicted mechanical properties of the sample after forming.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, which includes a processor and a memory. The memory stores machine-executable instructions that can be executed by the processor. The processor executes the machine-executable instructions to implement the method for predicting the correlation between cold bending forming process parameters and product mechanical properties in any of the above embodiments.

[0014] Fourthly, embodiments of the present invention also provide a machine-readable storage medium storing machine-executable instructions. When the machine-executable instructions are invoked and executed by a processor, the machine-executable instructions cause the processor to implement the method for predicting the correlation between cold bending forming process parameters and product mechanical properties in any of the above embodiments.

[0015] The embodiments of this invention bring the following beneficial effects: This invention provides a method and apparatus for predicting the correlation between cold bending forming process parameters and product mechanical properties. It utilizes the yield strength and plate thickness in material parameters to scale and decouple geometric processing parameters, making the characteristic values ​​of the same geometric path under different material scales tend to be consistent, thereby separating the influence of "geometric deformation" from "material scale." Simultaneously, based on the pass spacing, bending angle, and plate thickness, it calculates the plastic history index before the start of each pass, quantifying the accumulated plastic deformation intensity of previous passes, transforming the originally invisible historical path into an input numerical feature. Furthermore, the corresponding process parameters and material parameters are organized into a multi-pass cold bending forming sequence according to the pass order, completely preserving the evolution law of plastic accumulation and the temporal dependence between passes. Finally, when the prediction model predicts the mechanical properties of the sample to be predicted, because the input features themselves have physical decoupling and historical perception capabilities, the model can distinguish samples with the same geometry but different histories, and provide physically consistent residual stress predictions, ultimately achieving accurate correlation prediction from process parameters to product mechanical properties.

[0016] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for predicting the correlation between cold bending forming process parameters and product mechanical properties, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the impact of the feature enhancement vector on prediction accuracy in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the comparison of prediction effects provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a device for predicting the correlation between cold bending forming process parameters and product mechanical properties, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] To facilitate understanding, the method for predicting the correlation between cold bending forming process parameters and product mechanical properties provided in this embodiment of the invention will be described first, with reference to... Figure 1 The method includes the following steps: Step S102: Obtain the material parameters of the sample to be predicted, and the molding process parameters of the multiple passes for molding the sample to be predicted.

[0023] The sample to be predicted is a target product that requires cold bending forming and for which the material grade and all process parameters for each pass have been determined. Material parameters include the product's sheet thickness, initial yield strength, and hardening index; these parameters remain constant throughout the forming process and reflect the material's inherent mechanical properties. The forming process parameters for each pass include at least the pass spacing, bending angle, forming speed, and roll geometry description. These parameters characterize the geometric deformation path of the sheet material in each pass. Specifically, the pass spacing is the distance between the centers of two adjacent rolls along the production line direction, affecting the release of residual stress from previous passes and tension transfer; the bending angle is the additional bending angle (in degrees or radians) generated in the strip during that pass, used to control the amount of plastic deformation; the forming speed is the linear velocity of the strip through that pass (e.g., meters per minute), used to affect the strain rate and frictional heat effect; and the roll geometry description includes fillet radius, groove depth, groove width, and working surface curvature, which determines the contact pressure distribution and local forming accuracy.

[0024] In cold bending forming, metal strip typically undergoes several passes (e.g., 8-12 passes) of progressive bending, causing the material to gradually undergo plastic deformation until the target cross-sectional shape is achieved. Each pass can consist of a pair of rolls, and the roll profile determines the bending angle and local forming amount introduced in that pass. The bending angles of multiple passes can be distributed in a "first increase then decrease" or "step-increase" pattern to avoid excessive deformation in a single pass, which could lead to material cracking or uncontrolled springback.

[0025] Step S104: Based on the material yield strength and material plate thickness parameters in the material parameters, the geometric processing parameters of the forming process parameters are scaled and decoupled to obtain the decoupled geometric processing parameters corresponding to the corresponding pass.

[0026] In cold bending forming, even if two products use the exact same geometric deformation path (i.e., the pass spacing, bending angle, forming speed, and roller geometry are all identical), as long as their plate thickness or initial yield strength differs, the springback, residual stress amplitude, and yield strength changes after forming will exhibit significant differences. For example, the springback of thick plates or high-strength steel at the same bending angle is usually much greater than that of thin plates or ordinary strength steel.

[0027] If the original geometric processing parameters and material parameters are directly predicted, the model will struggle to distinguish between "response changes caused by the geometric path itself" and "response changes caused by material scale" during the learning process. More seriously, the model may incorrectly attribute response amplitude fluctuations caused by material differences to certain geometric parameters, thus learning spurious correlations. For example, when thick plate samples in the data are mostly accompanied by large bending angles, the model may incorrectly assume that "large bending angles necessarily lead to large springback," ignoring the actual cause being plate thickness. When applied to thin plate materials, even with the same bending angle, the model may underestimate the springback; when applied to thicker materials, it may overestimate it. This spurious correlation severely limits the model's generalization ability across different grades and specifications of products.

[0028] To address this, this invention also proposes decoupling the geometric path information of the forming process based on material parameters, ensuring that the geometric parameters input into the subsequent model no longer carry the influence of material properties. The initial yield strength and plate thickness can be considered as material-scale indicators; the larger these indicators, the stronger the "amplification effect" of the material on the mechanical response. Correspondingly, the original geometric processing parameters (pass spacing, bending angle, forming speed, roller geometry, etc.) can be divided by a power of these indicators, so that samples with identical geometric paths but different material conditions will have scaled geometric processing parameters falling into similar numerical ranges, becoming purely geometric path features. Based on this, the model can learn the contribution of geometric deformation to the mechanical response and the adjustment of the response amplitude by material properties, eliminating the false correlation where the model misjudges material differences as geometric regularities, making the prediction results more consistent with physical reality.

[0029] Step S106: Based on the pass spacing and bending angle in the molding process parameters and the plate thickness in the material parameters, determine the plastic history index of the sample to be predicted before the start of each pass.

[0030] Cold bending is a multi-pass progressive bending process. The plastic deformation applied in previous passes gradually accumulates in the material, leading to work hardening, increased yield point, and superposition of residual stress. These historical effects alter the material's mechanical response behavior in subsequent passes. For example, materials that have undergone significant plastic deformation exhibit less springback and a more complex residual stress distribution during subsequent bending.

[0031] To characterize the impact of the entire molding process on the final mechanical properties, this embodiment of the invention also calculates a plastic history index for each pass to quantify the cumulative deformation strength. For example, a larger index indicates more sufficient prior accumulation, a higher degree of material hardening, and a more significant impact on subsequent springback and residual stress. With this index, even if two samples have identical geometric parameters in the current pass, as long as their prior histories are different, the model can output different springback and residual stress prediction results through different plastic history index values, thereby achieving a precise correlation between process parameters and mechanical properties.

[0032] Step S108: Based on the plasticity history index, decoupled geometric processing parameters, and material parameters, construct a multi-pass cold bending forming sequence for the sample to be predicted.

[0033] The plastic history index can quantify the cumulative deformation strength of preceding processes, while decoupling geometric processing parameters can eliminate material-scale interference and retain pure geometric path information. Material parameters can characterize inherent mechanical properties. Furthermore, a multi-pass cold bending sequence can be constructed based on these three parameters in pass order, enabling subsequent models to identify the evolution of plastic accumulation along the pass direction. For example, how large deformations in early passes change the springback trend in later passes, or how parameter adjustments in intermediate passes affect the accumulation path of residual stress. This transforms the complex temporal dependency between the process parameter sequence and the final mechanical properties into a high-dimensional feature representation that the model can learn, achieving end-to-end performance prediction.

[0034] Step S110: Input the multi-pass cold bending forming sequence into the pre-trained mechanical response prediction model, and use the mechanical response prediction model to predict the mechanical properties of the multi-pass cold bending forming sequence to determine the predicted mechanical properties of the sample after forming.

[0035] The completed multi-pass cold bending forming sequence integrates the decoupled geometric processing parameters (characterizing the pure geometric deformation path), plastic history index (quantifying the cumulative deformation strength of the preceding passes), and inherent material properties (plate thickness, initial yield strength, hardening index) for each pass in sequence. This sequence fully records the evolution of the process scheme, material properties, and plastic accumulation from the first pass to the last pass.

[0036] The sequence is input into a pre-trained mechanical response prediction model. This model can identify the contribution of pure geometric deformation to performance using decoupled geometric features and perceive the influence of prior cumulative effects using the plastic history index, thereby outputting the expected mechanical properties of the product under this process scheme (such as strength, springback, and residual stress distribution in the cross section). Compared with traditional methods, the input features of this invention already carry historical path information and material-independent geometric paths, enabling the model to distinguish samples with the same geometry but different histories and predict the residual stress field that satisfies the self-equilibrium of the cross section, achieving accurate correlation prediction between cold bending forming process parameters and final mechanical properties.

[0037] In summary, in the cold bending forming process, the final mechanical properties of the product are influenced by the inherent properties of the material, the geometric deformation paths of each pass, and the cumulative effect of preceding plasticity. This invention utilizes the yield strength and plate thickness in the material parameters to scale and decouple the geometric processing parameters, making the characteristic values ​​of the same geometric path consistent across different material scales, thus separating the influence of "geometric deformation" from that of "material scale." Simultaneously, based on the pass spacing, bending angle, and plate thickness, the plastic history index before each pass is calculated, quantifying the accumulated plastic deformation intensity of preceding passes and transforming the previously invisible historical path into an input numerical feature. Furthermore, the corresponding process parameters and material parameters are organized into a multi-pass cold bending forming sequence according to the pass order, fully preserving the evolution law of plasticity accumulation and the temporal dependence between passes. Finally, when the prediction model predicts the mechanical properties of the sample after forming, because the input features themselves possess physical decoupling and historical perception capabilities, the model can distinguish samples with the same geometry but different histories and provide physically consistent residual stress predictions, ultimately achieving accurate correlation prediction from process parameters to product mechanical properties.

[0038] Furthermore, based on the above embodiments, this invention also provides another method for predicting the correlation between cold bending forming process parameters and product mechanical properties, to illustrate the steps of the above embodiments. Corresponding to the above embodiments, after completing model training, this invention can deploy the model to cold bending forming process design software or production line parameter evaluation system to quickly predict new process samples. In one embodiment, all pass parameters and material parameters of the new process sample can be read to form an original process parameter vector (i.e., forming process parameters) and material parameter vector (i.e., material parameters) in the same format as the training stage. Further, the plasticity history index of each pass of the new process sample is calculated, and all geometrically decoupled pass feature vectors (i.e., decoupled geometric processing parameters) are generated and stacked to obtain a cold bending forming sequence feature matrix (i.e., a multi-pass cold bending forming sequence).

[0039] Furthermore, the corresponding sequence can be input into the trained cold-bending mechanical response prediction model to obtain the scalar mechanical response prediction vector and the residual stress prediction vector. In one embodiment, the cold-bending sequence feature matrix can also be expanded into sample-level feature vectors, and then the nearest neighbor retrieval index saved during the training phase can be used to search for the nearest neighbor. For each training sample, calculate the local mean vector. and local projection matrix The center projection vector is obtained. In the online prediction phase, the enhanced feature vector can also be directly applied. Random perturbations can be added during the training phase to improve the robustness of the model.

[0040] Furthermore, the prediction results in the normalized space can be inversely transformed into actual engineering units, outputting the yield strength, tensile strength, hardness, springback, and residual stress distribution of the cross-section after molding. In one embodiment, as an example, assuming the model has been trained, the normalized scalar output is predicted for a new process sample. and normalized residual stress vector The actual values ​​are recovered based on the mean and standard deviation saved during the training phase, as follows: Yield strength prediction: Tensile strength prediction: Hardness prediction: Rebound amount prediction: Similarly, the residual stress distribution is inversely normalized to obtain the actual stress value at each discrete location. Based on this, it is possible to extract enhancing features from the original multi-pass cold bending forming process parameters that are sensitive to plastic history, stable to changes in material dimensions, and robust to local process disturbances. Under the condition of satisfying cross-sectional equilibrium constraints, a predictive model between the cold bending forming process parameters and the mechanical response after forming can be established. After the model is deployed, it can quickly predict the mechanical properties (yield strength, tensile strength, hardness, springback) and residual stress distribution of the formed material under given cold bending forming process parameters. These prediction results can be directly used for: process parameter optimization: adjusting parameters such as pass spacing and bending angle to minimize springback or make the residual stress distribution more uniform; virtual mold testing: replacing some physical mold testing to shorten the new product development cycle; quality prediction: assessing whether the current process scheme will lead to insufficient strength or excessive stress concentration after forming; and equipment parameter setting guidance: providing feedforward reference values ​​for the production line control system.

[0041] It is understandable that residual stress and springback in cold bending are not solely determined by the current pass, but are also closely related to the degree of plastic deformation and material dimensions already formed in previous passes. If we directly apply the original cold bending parameter matrix... Performing uniform normalization can easily mix up "geometric path differences" and "material scale differences," making it difficult to compare samples with different plate thicknesses and different initial yield strengths.

[0042] Based on the above embodiments, the embodiments of the present invention mainly describe how to calculate the plastic history index corresponding to each pass, and how to express the geometric parameters and material scale effects separately.

[0043] Regarding step S104 above, a material scaling factor can be constructed based on the initial yield strength and plate thickness in the material parameters. The geometric processing parameters are then scaled using this material scaling factor to obtain decoupled geometric processing parameters. It is important to note that the mechanical response amplitude corresponding to the same geometric path is not consistent under different plate thicknesses and different initial yield strengths. If geometric and material parameters are directly mixed and input, the model may easily mislearn the material scale differences as geometric regularities. This embodiment of the invention calculates the material scaling factor to scale and decouple the geometric parameters.

[0044] 1> From the original process parameter vector Extracting geometric parameter subvectors Geometric parameter subvectors Used to characterize the The geometric processing path for each pass. In one implementation, the geometric parameter subvector... Including track spacing Bending angle Molding speed and roller geometry description vector The components in the text. Furthermore, regarding the plate thickness... Initial yield strength and hardening index Perform normalization to obtain the normalized material parameter vector. Normalization can be achieved by normalizing the training set mean and standard deviation or by minimax normalization. In one implementation, the plate thickness, initial yield strength, and hardening index are subtracted from the training set mean and divided by the training set standard deviation, respectively.

[0045] 2> Construction material scaling factor Material scaling factor To express the combined effect of plate thickness and initial yield strength on the stress response scale, in one implementation, it can be calculated using the following formula:

[0046] in, Indicates the first Material scaling factor for each sample; This indicates the reference initial yield strength, for example, 235 MPa; This indicates the thickness of the reference plate, for example, 1 mm; This represents the thickness sensitivity index, which can be set to 0.15 for example. The larger the material scaling factor, the more significant the amplification effect of material strength and plate thickness on the mechanical response.

[0047] Furthermore, using a material scaling factor For geometric parameter subvectors Perform scaling decoupling to obtain the scaled geometric vector. In one implementation, the geometric parameter subvectors are... according to Scaling is performed, and the calculation method is as follows: ,in, This represents the scaling factor, for example, 0.8. Based on this, the geometric path features under different material conditions will fall into a closer numerical range, which is more conducive to unified modeling.

[0048] Regarding step S106 above, the equivalent bending radius of the current pass can be determined based on the pass spacing and bending angle of the corresponding pass; the pass plasticity increment of the pass can be determined based on the plate thickness and the equivalent bending radius; and the pass plasticity increments of all passes before the current pass can be weighted and accumulated to determine the plasticity history index before the start of the current pass.

[0049] During cold bending forming, the bending and correction of the sheet metal in previous passes alters the yield point and springback trend in subsequent passes. To represent this cumulative effect, this step calculates a plastic history index for each sample and each pass, thereby quantitatively expressing the degree of cumulative plasticity from previous passes. This index serves as a separate dimension for "history path sensitivity" in subsequent features. The specific steps are as follows: 1> Read the first Plate thickness of each sample All track spacing and all track curvature angles When the bending angle When recording in degrees, convert them to radians before using them in subsequent calculations.

[0050] Furthermore, the equivalent bending radius is calculated for each pass. Equivalent bending radius The curvature scale used to approximate the curve of this passage can be calculated in one implementation as follows:

[0051] in, Indicates the first The sample at the th The equivalent bending radius of each pass; Indicates the first The sample at the th Track spacing; Indicates the first The sample at the th The curvature angle of the track; This represents a minimal constant to prevent the denominator from being zero; for example, it can be taken as... It will automatically increase, indicating that the current track's curvature effect is relatively weak.

[0052] 2> Based on plate thickness and equivalent bending radius Calculate the plastic increment per pass .

[0053] incremental plasticity per pass Used to characterize the The sample at the th In one embodiment, the representative plastic tensile strength of the outer layer fiber in a pass can be calculated using the following formula:

[0054] in, Indicates the first The sample at the th The plastic increment per pass. As shown in the expression, the greater the plate thickness and the smaller the equivalent bending radius, the greater the plastic increment per pass. To avoid excessively large values ​​due to a few abnormal parameters, in practical implementation, the plastic increment per pass can be... The value is limited to between 0 and 0.4.

[0055] 3> Plastic increment of the preceding pass Perform a weighted summation to obtain the first... Plasticity history index before the start of the course , Indicates the first The sample at the th The accumulated plastic history intensity before the start of the next stage; the larger the value, the more obvious the preceding plastic effect.

[0056] To reflect the principle that "the closer to the current track, the stronger the historical influence," a positional attenuation weight is introduced for preceding tracks. In one implementation, the plastic history index... It can be calculated using the following formula:

[0057] in, Indicates the first The sample at the th Plasticity history index before the start of the course; This represents the historical attenuation coefficient, with a value ranging from 0 to 1, for example, 0.85. Indicates the previous track index. When... When there is no preceding path, directly set .

[0058] Furthermore, the aforementioned calculation can be repeated for all samples and all passes to obtain the plasticity history index matrix. Plasticity history index matrix This represents the distribution of plastic history intensity of all samples at all pass locations, with a size of [size missing]. In one embodiment, for example, the first... Plate thickness of each sample The interval between tracks is 1.2. The bending angle is 20. The value is 30°. First, convert 30° to radians and then substitute it into the equation. The calculation formula is used to calculate the equivalent bending radius for that pass. Then, based on the plate thickness... and equivalent bending radius Calculated plastic increment per pass If, before the fourth pass, the plastic increments of the first three passes were 0.030, 0.045, and 0.060 respectively, then the historical attenuation coefficient... Taking 0.85, the plastic history index for the 4th pass is: Therefore, the proportion of plastic effect is greater closer to the current pass, which is more in line with the historical transfer characteristics in cold bending forming.

[0059] It should be noted that, in order for subsequent models to distinguish between samples with the same current pass parameters but different previous forming paths, this step no longer categorically incorporates the history of previous passes into the current pass parameters, but instead uses the plasticity history index. Expressing "the degree of plastic accumulation that has been formed before the current pass" separately can more stably express the influence of the preceding plastic history on the springback and residual stress.

[0060] Furthermore, regarding step S108 above, a historical position vector corresponding to each pass can be constructed based on the plasticity history index, plasticity history increment, and the relative position of the corresponding pass in the entire forming line; the decoupled geometric processing parameters, historical position vector, and material parameters are spliced ​​together to form a geometric decoupled pass feature vector corresponding to each pass; the geometric decoupled pass feature vectors of all passes are stacked in pass order to form a multi-pass cold bending forming sequence of the sample to be predicted.

[0061] Correspondingly, based on the above steps, the aforementioned plasticity history index can be injected into each pass feature to obtain a geometrically decoupled pass feature vector, ultimately forming a multi-pass cold bending forming sequence for the sample to be predicted. The specific steps are as follows: 1> Constructing historical position vectors Historical position vector This is used to simultaneously inject the plasticity history index and the passage order into the current passage feature. In one implementation, the historical position vector... It consists of three quantities, namely the plasticity history index of the current track. , Plasticity historical increment and the position of the unified path .in, , Indicates the first The sample at the th Plasticity history index before the start of the course, when Direct order ; This is used to indicate the relative position of the current pass within the entire forming line.

[0062] 2> Scale the geometric vector Historical position vector and normalized material parameter vector splicing together to form the first The sample at the th Geometric decoupling of trace eigenvectors Among them, the geometrically decoupled trace eigenvectors It can simultaneously contain three types of information: geometric path, plastic history, and material state, serving as the basic unit for subsequent sequence feature reconstruction. Furthermore, it can stack the layers sequentially according to the pass order. All geometrically decoupled trace feature vectors of each sample , obtained the Cold bending forming sequence feature matrix of each sample .

[0063] Among them, the cold bending forming sequence feature matrix Indicates the first Each sample contains a complete multi-pass enhancement pre-feature sequence (i.e., a multi-pass cold bending forming sequence), with a size of [size missing]. .in, Represents a single geometrically decoupled trace eigenvector The dimension. In one embodiment, as an example, let the first... Initial yield strength of each sample 350 MPa, plate thickness It is 1.5 mm thick, with reference to the initial yield strength. Take 235 MPa as the reference plate thickness. Taking 1 mm as an example, the thickness sensitivity index is... If we set it to 0.15, we can first calculate the material scaling factor. If the first Geometric parameter subvectors of the track Composed of five dimensions: pass spacing, bending angle, forming speed, and the geometric parameters of the two rollers, the first is to use... Scale these 5-dimensional parameters and then stitch them together with the plasticity history index. , Plasticity historical increment Normalization of Dao positions and normalized material parameter vector This yields the geometrically decoupled trace eigenvectors. It should be noted that, in order to avoid the material strength and plate thickness variations masking the differences in the actual geometric path, this step expresses the "geometric path" and "material scale" separately, and directly injects the plastic history index into each pass feature. Based on this, the comparability and generalization ability of the model under different material grades and plate thicknesses can be improved.

[0064] Furthermore, this embodiment of the invention also determines the local nearest neighbor set corresponding to the multi-pass cold bending forming sequence from a preset pass reference sample; based on the local nearest neighbor set, it determines the center projection vector and the perturbation scale vector corresponding to the multi-pass cold bending forming sequence. Further, it can also construct enhanced input features corresponding to the multi-pass cold bending forming sequence based on the center projection vector; or, it can add random perturbations to the enhanced input features in combination with the perturbation scale vector to construct perturbed enhanced input features. Training samples are further constructed based on the enhanced input features or the perturbed enhanced input features to train the mechanical response prediction model.

[0065] Among them, the above-mentioned cold bending forming sequence feature matrix Although it already includes plastic history and decoupled geometric information, it still suffers from high dimensionality, significant local noise, and potential uneven sample distribution caused by abrupt changes in roller parameters. To obtain a more stable sample-level representation without introducing a complex pre-trained network, this embodiment of the invention also generates enhanced feature vectors through local neighborhood projection. .

[0066] In one implementation, the following steps are followed: 1> The first Cold bending forming sequence feature matrix of each sample Expanded into sample-level feature vectors in a fixed order Sample-level feature vectors Indicates the first The concatenation result of all trace features of each sample, with a length of [length missing]. In practical implementation, the order can be expanded line by line, starting with the number of tracks and then the dimensions, as long as the training and prediction phases remain completely consistent. In one embodiment, as an example, let the number of tracks for a given sample be... Each Dimension ,but for The matrix is ​​expanded in the order of "course number first, dimension second," meaning it first takes the 10-dimensional vector of the first course, then the 10-dimensional vector of the second course, and so on, resulting in a matrix of length [missing information]. Sample-level feature vectors .

[0067] 2> The preset track reference samples can be all samples in the model's training set (e.g., refer to the training sample section below). After feature engineering, a set of corresponding sample-level feature vectors can be formed. Further, these all samples are compared with the aforementioned sample-level feature vectors. A nearest neighbor index can be established to quickly find the nearest reference sample in a high-dimensional space, forming a neighborhood set. For example, materials with similar inherent properties can be used. In one implementation, a KD-tree, a ball tree, or an approximate nearest neighbor index can be employed. In one embodiment, as an example, all samples in the training set can be indexed... A KD-tree index is constructed on top of each [the existing index]. Using 80-dimensional coordinates as spatial points, the variance of each dimension is pre-calculated for weighted distance (optional), and subsequent new samples are then processed. It can quickly find the nearest Training samples.

[0068] 3> For the current sample-level feature vector Search for the nearest neighbor in the training set. Each set of neighbor samples forms a local neighborhood. (Number of neighbors) The distance can range from 8 to 20, and in one implementation, it can be 12. The distance metric can be Euclidean distance or weighted Euclidean distance weighted by the standard deviations of each dimension.

[0069] 4> Furthermore, the sample-level feature vectors in the local neighborhood set can be first subjected to mean-removal processing to obtain the local mean vector. Local mean vector This represents the average process path center in the vicinity of the current sample. The local covariance matrix is ​​calculated based on the mean-removed local neighborhood set, and eigenvalue decomposition is performed on the local covariance matrix. The top [values] can be selected... The principal directions form a local projection matrix. This maps high-dimensional sample-level feature vectors to a low-dimensional stable subspace. In one implementation, the dimensionality of the feature vectors is increased. The value can be 16, 32, or 64; if determined by the cumulative explained variance, the principal direction with 90% or 95% of the cumulative explained variance is retained. This local projection matrix is ​​used to dynamically construct a "personalized" low-dimensional feature space for each sample during prediction. By capturing the local true data manifold of the sample, high-dimensional features are projected onto a more stable principal direction, thereby filtering out local noise such as roller parameter jumps.

[0070] Sample-level feature vectors Projected onto the local projection matrix The corresponding low-dimensional subspace yields the central projection vector. Center projection vector Indicates the first The deterministic representation of a sample in a locally stable subspace can be calculated by first... Subtract the local mean vector Then, with the local projection matrix Matrix multiplication is represented as ,in, Represents the local projection matrix The transpose of .

[0071] 5> Furthermore, the degree of dispersion of the coefficients of the local neighborhood set in the low-dimensional subspace can be calculated to obtain the perturbation scale vector. This serves as a benchmark to measure the "normal fluctuation range" of reference samples most similar to the current sample across various feature dimensions in a local low-dimensional space. In practical implementation, we first... All neighbor samples are projected onto the local projection matrix. In the corresponding low-dimensional subspace, the standard deviation of the projection coefficients in each dimension is calculated to obtain the projection vector with respect to the center. Same-dimensional perturbation scale vector To prevent excessive perturbation in individual dimensions, the perturbation scale vector can be adjusted. Limited to 0 to Between, for example 0.1 is acceptable.

[0072] In practical implementation, the local neighborhood will be... The sample-level feature vectors of each sample are projected onto... In the corresponding low-dimensional space, we obtain indivual 3D projection vector. For each dimension Calculate the value in this dimension The standard deviation of each projection value yields the perturbation scale vector. The Each component Finally, each component is truncated to... ,For example .

[0073] 6> Furthermore, it can be based on the central projection vector and / or perturbation scale vector Generate enhanced feature vectors During the training phase, the enhanced feature vector can be generated using the following formula. :

[0074] in, Indicates the first Enhanced feature vectors of each sample; This represents element-wise multiplication; Represents a standard normally distributed random vector, with dimensions and enhanced feature vectors. Same. During the deployment phase, to ensure output stability, it can be directly taken as... Random disturbances will no longer be introduced.

[0075] In one embodiment, as an example, suppose a sample has a total of 8 channels, and the feature vector of each geometrically decoupled channel is... If the dimension is 10, then the cold bending forming sequence feature matrix The size is After expansion, a sample-level feature vector of length 80 is obtained. After finding the 12 nearest neighbors in the training set, the local covariance matrix is ​​calculated and 32 principal directions are retained to obtain the local projection matrix. Then the sample-level feature vectors The projection yields a 32-dimensional central projection vector. Then, based on the degree of dispersion of the coefficients of the 12 neighbors in the 32-dimensional space, a 32-dimensional perturbation scale vector is obtained. This ultimately forms a 32-dimensional enhanced feature vector. It should be noted that, to adapt to scenarios with limited sample size, mixed variable types, and significant differences in local structures in cold bending forming data, this step does not employ a complex encoder requiring a large number of pre-trained parameters. Instead, it directly generates sample-level enhanced feature vectors using local neighborhood projection. Based on this, the robustness of features to local noise and roller parameter jumps can be improved while maintaining the simplicity of implementation.

[0076] In summary, by constructing a nearest-neighbor retrieval index in a high-dimensional sample-level feature space, and performing covariance decomposition on the local neighborhood of each sample and projecting it onto a low-dimensional subspace, an enhanced feature vector with both stability and expressive power can be generated. In this process, the number of neighbors K determines the breadth of the local statistical scope and the enhancement dimension. The amount of information retained in the projected subspace is determined by these factors, and both can jointly affect the performance of subsequent prediction models. Therefore, this invention also analyzes the impact of local neighborhood projection parameters on prediction accuracy, and studies two key hyperparameters in the local neighborhood projection method (local neighborhood size K and enhanced feature vector dimension). The influence of parameters on the final prediction accuracy is investigated, and the rationality of the parameter combination selected in this invention is verified. (Refer to...) Figure 2 This diagram illustrates the impact of feature enhancement vectors on prediction accuracy. Figure 2 In the middle, the horizontal axis represents "enhanced feature vector dimension". The vertical axis represents the "local neighborhood size K". The figure uses a filled contour plot, with colors ranging from dark purple to bright yellow to represent the numerical variation of the "root mean square error of rebound prediction", in millimeters. The color bars on the right indicate the corresponding color mapping of the error values. Pentagrams are also used to mark the parameter combinations that can be used in this invention. The marking positions clearly show that this parameter combination is located precisely at the core of the low-error region, with the darkest surrounding color representing the lowest error level.

[0077] Furthermore, the model input (multi-pass cold bending sequence) already includes multi-pass plastic history, geometric path, and material scaling decoupling information. However, there is still a nonlinear relationship between this input and the scalar mechanical response and residual stress distribution after forming. Especially in cold bending, the residual stress distribution not only needs to fit the numerical magnitude but also needs to satisfy the approximate balance between the resultant force and resultant moment of the cross section. Correspondingly, this embodiment of the invention constructs a multi-task prediction model, which predicts scalar mechanical responses such as yield strength, tensile strength, hardness, and springback after forming, and predicts the residual stress distribution of the cross section, and improves the physical rationality of the results through balance constraints. Correspondingly, for the above step S110, the predicted results of the mechanical properties after forming can be determined through the following steps: 1) Determine the gating coefficient vector of the mechanical response prediction model based on the strength of the plastic history corresponding to the multi-pass cold bending forming sequence.

[0078] a) The plasticity history index matrix can be compared The first in The data is aggregated to form a plastic history summary vector. Plastic history summary vector This is used to represent the strength of plasticity history from a sample-level perspective. It can consist of three components: the average and maximum plasticity history index of all passes, and the plasticity history index of the last pass, thus forming a plasticity history summary vector. It can characterize the overall plasticity level, local peak values, and end-stage accumulation states, respectively. In one embodiment, as an example, a plasticity history index sequence is used. Taking a total of 8 passes as an example, the average value is 0.08, the maximum value is 0.14, and the value of the last pass (the 8th pass) is... If the value is 0.14, then The length is 3.

[0079] b. The dimension of the model's main input (such as a multi-pass cold bending sequence or an enhanced feature vector) can be denoted as... , Furthermore, the corresponding main input can be fed into the first fully connected layer to obtain the initial shared features. In one implementation, the input dimension of the first fully connected layer is... The output dimension is 128, followed by a linear rectification activation function. Based on this, a general nonlinear representation related to the mechanical response is initially extracted, which can provide a rich feature channel foundation for subsequent gating regulation.

[0080] c. The plastic history summary vector of step a above can be obtained. Input the gating mapping layer to obtain the gating coefficient vector. Gating coefficient vector Used to adjust the enhanced feature vector The responses of each channel after the first fully connected layer. In one implementation, the gated mapping layer consists of one fully connected layer and a sigmoid activation function, with the output dimension sharing the initial features. Consistent, meaning 128 dimensions; the gating coefficient of each dimension is between 0 and 1.

[0081] 2) Based on the gating coefficient vector, determine the features to be predicted corresponding to the multi-pass cold bending forming sequence.

[0082] In one implementation, the gating coefficient vector can be... Shared features with the initial Element-wise multiplication yields the plastic history-gated features. When the current sample has a strong plastic history, some channels related to plastic accumulation will be amplified; when the plastic history is weak, the corresponding channel responses will be suppressed.

[0083] Furthermore, plastic history gating features can be used. Inputting the second fully connected layer yields shared predicted features. Shared predictive features It can serve as a unified input (i.e., the aforementioned features to be predicted) for subsequent multi-task output layers. In one implementation, the second fully connected layer has an input dimension of 128 and an output dimension of 64, followed by a linear rectified activation function.

[0084] In one embodiment, for example, if the first The plasticity history indices of the eight channels of a sample are [0, 0.02, 0.05, 0.08, 0.10, 0.12, 0.13, 0.14], then the plasticity history summary vector is... The possible values ​​are [average 0.08, maximum 0.14, final pass value 0.14]. The plastic history summary vector... After inputting into the gating mapping layer, a 128-dimensional gating coefficient vector is obtained. Then, it shares features with the initial 128-dimensional output of the first fully connected layer. Element-wise multiplication yields the plastic history gating feature. .

[0085] It should be noted that, in order to enable the model to establish different responses to the degree of plasticity accumulation during the feature extraction stage, this step does not simply use the plasticity history index as an additional input, but rather utilizes the plasticity history summary vector. Generate gate coefficient vector This allows for direct adjustment of shared feature channels. Based on this, the model's ability to identify "historically sensitive parts" in the rebound amount and residual stress distribution can be enhanced.

[0086] 3) Input the feature to be predicted into the mechanical response prediction model, and perform branch prediction on the feature to be predicted through the scalar mechanical response output branch and the residual stress output branch of the mechanical response prediction model to determine the corresponding scalar mechanical response result and residual stress prediction result.

[0087] After extracting the above-mentioned features to be predicted (shared predictive features) Following this, embodiments of the present invention further map the outputs to two prediction targets: scalar mechanical response and residual stress distribution. Since the two types of outputs differ significantly in dimensions, numerical range, and physical meaning, embodiments of the present invention also design independent branch networks and employ normalization processing to stabilize training. For example, for the scalar mechanical response, the aforementioned shared prediction features can be used... The scalar mechanical response output branch is fed into the scalar mechanical response prediction vector. Scalar mechanical response prediction vector It is used to predict post-molding yield strength, post-molding tensile strength, post-molding hardness, and springback. In one embodiment, the scalar mechanical response output branch includes one hidden layer and one output layer. The hidden layer dimension can be 32, and the output layer dimension is... For the aforementioned residual stress distribution, shared prediction features can be used. The residual stress is fed into the residual stress output branch to obtain the residual stress prediction vector. Residual stress prediction vector Used for predicting cross-sections The residual stress values ​​at discrete locations. In one embodiment, the residual stress output branch includes one hidden layer and one output layer, where the hidden layer dimension can be 64 and the output layer dimension is... After training is complete, all trainable parameters in the scalar mechanical response output branch and the residual stress output branch can be saved for subsequent joint optimization.

[0088] 4) Based on the scalar mechanical response results and residual stress prediction results, determine the predicted mechanical properties of the sample after molding.

[0089] Specifically, the predicted mechanical properties of the sample after molding can be determined based on the training label statistics corresponding to the scalar mechanical response results and the residual stress prediction results, as well as the corresponding first and second predicted values. During training, the training set statistics of the scalar mechanical response labels and residual stress labels can be normalized before model training to avoid excessive differences in output dimensions affecting training stability. Upon completion of the prediction, the normalized prediction results can be restored to actual engineering units based on the corresponding statistics. In one embodiment, as an example, if the mean of the yield strength training set is... The standard deviation is The model predicts the normalized value as The actual predicted value is .

[0090] Furthermore, model training can be achieved through the following steps: 1) Obtain pre-constructed training samples. The training samples contain sample labels, including scalar mechanical response labels and residual stress labels.

[0091] This involves the unified collection and organization of training samples to obtain the original cold-bending forming parameter matrix, scalar mechanical response labels, and residual stress labels, which can be used for feature generation and model training. For example, multiple cold-bending forming samples can be collected for the same type of cross-section product or the same pass arrangement. Each cold-bending forming sample contains a fixed number of passes, denoted as the number of passes. For example, 8, 10, or 12 can be used. If different numbers of passes exist in actual production, models can be made separately according to the same number of passes. Alternatively, shorter sequences can be padded with zeros at the end and the valid pass markers can be saved simultaneously. It is preferable to use samples with the same number of passes to reduce implementation complexity.

[0092] Furthermore, the process parameters for each cold-bending sample at each pass can be read to form the first... The sample at the th Original process parameter vector for each pass Original process parameter vector Used to characterize the actual processing status of the current pass, in one implementation, the original process parameter vector... Including track spacing Bending angle Molding speed and roller geometry description vector .in, Indicates the first The sample at the th The spacing between adjacent forming positions in a pass, in millimeters; Indicates the first The sample at the th The bending angle introduced by each pass is calculated using radians. Indicates the first The sample at the th Linear velocity per pass, in meters per minute; roller geometric description vector. Used to characterize the The profile of the rollers for each pass can be composed of parameters such as the roller fillet radius, groove depth, groove width, and working surface curvature. In one embodiment, as an example, let a certain sample number be... , No. Dao Ci (Total) (Pass). The original process parameter vector for this pass. It can be constructed as follows: Track spacing Bending angle (correspond Molding speed Roller geometric description vector From the radius of the fillet trench depth , slot width Working surface curvature Composition, that is .but Dimension .

[0093] Furthermore, the material parameters and global parameters of each cold-bending sample are read to form the first... Material parameter vector of each sample Material parameter vector To characterize material properties that do not change with each pass or are shared throughout the entire forming line, in one implementation, a material parameter vector is used. Including plate thickness Initial yield strength and hardening index .in, Indicates the first The plate thickness of each sample is in millimeters; Indicates the first The initial yield strength of each sample, in megapascals; Indicates the first The material hardening index of each sample.

[0094] Furthermore, all samples can be organized into an original cold-bending forming parameter matrix according to sample index and pass index. Original cold bending forming parameter matrix This represents a dataset consisting of all samples, all tracks, and all track parameters, with a size of [size missing]. ,in, Indicates the number of samples. This represents the parameter dimension for a single track.

[0095] Furthermore, scalar mechanical response labels and residual stress labels can be collected for each cold-bent sample. The scalar mechanical response label is denoted as... This is used to characterize the key scalar results after the sample has been shaped. In one implementation, This includes post-molding yield strength, post-molding tensile strength, post-molding hardness, and springback. Residual stress is denoted as... This is used to characterize the residual stress distribution at multiple discrete locations on a cross section. If the cross section is co-discrete... If there are several positions, then The length is In practice, residual stress labels can be obtained through X-ray diffraction, blind hole method, or experimentally calibrated finite element results.

[0096] Furthermore, the scalar mechanical response labels of all samples are organized into a scalar mechanical response matrix. The size is ,in, This represents the number of scalar mechanical responses, for example, 4; the residual stress labels of all samples are organized into a residual stress matrix. The size is It should be noted that this step does not treat the cold-bending sample as a single static parameter point, but explicitly retains the three-layer structure of "sample-pass-parameter" to provide a unified input for subsequent calculation of the plasticity history index and generation of pass sequence features, which can avoid losing pass sequence information in subsequent feature processing.

[0097] 2) Based on the training samples and their labels, the mechanical response prediction model is trained to construct the scalar mechanical response output branch and the residual stress output branch corresponding to the training samples. The total loss used in training the mechanical response prediction model is constructed based on the scalar mechanical response loss, the residual stress loss, and the cross-sectional equilibrium constraints and residual stress smoothing regularization terms corresponding to the residual stress.

[0098] Training a multi-task network using only mean squared error can lead to a good numerical fit for residual stress prediction, but it disrupts the overall force balance of the cross-section, thus affecting the reliability of derived indicators such as springback. This embodiment of the invention also introduces a physical constraint term into the loss function and jointly optimizes all trainable parameters.

[0099] a. In one implementation, the first... True values ​​of scalar mechanical response for each sample and the true value of residual stress , respectively with the scalar mechanical response prediction vector and residual stress prediction vector By comparison, the scalar mechanical response loss and residual stress loss can be calculated, and both the scalar mechanical response loss and residual stress loss can be expressed in the form of mean square error.

[0100] b. In this context, the residual stress on the cross-section of the cold-bent part must satisfy the internal self-equilibrium condition of "zero resultant force and zero resultant moment" when no external force is applied; otherwise, the part will deform on its own. To ensure that the predicted residual stress distribution satisfies basic static equilibrium, this embodiment of the invention also constructs cross-sectional equilibrium constraints for model training. This constraint penalizes unbalanced stress distributions that are numerically accurate but physically invalid, thereby ensuring that the prediction results are physically true and reliable. Specifically, let the cross-section be co-discrete as... The position, the first The representative area corresponding to each location is The distance to the reference neutral axis of the cross section is Then, the resultant force and resultant moment of the cross section corresponding to the predicted residual stress distribution can be calculated separately. The resultant force and resultant moment of the cross section are approximately balanced. Axial resultant force Resultant bending moment .

[0101] c. Furthermore, the sum of the squares of the above two equilibrium residuals can be used as a regularization penalty, and the cross-sectional equilibrium regularization term can be applied. It can be constructed as follows:

[0102] in, Indicates the first The sample at the th Predicted residual stress at discrete locations on each cross section; Indicates the first The representative area corresponding to each discrete location; Indicates the first The distance from each discrete location to the reference neutral axis of the cross section. The term is used to penalize unbalanced resultant forces at cross sections. The term is used to penalize unbalanced resultant moments at cross sections.

[0103] d. To avoid unreasonable spikes in the predicted residual stress distribution between adjacent discrete locations, a residual stress smoothing regularization term can be further constructed. In one implementation, the residual stress smoothing regularization term... The sum of squares of the residual stress differences between adjacent discrete locations can be taken, i.e., for Squaring each point and summing the results, where, Indicates the first The sample at the th The predicted residual stress at discrete locations on each cross section can be used to suppress local oscillations without altering the overall stress level of the cross section.

[0104] e. Furthermore, the scalar mechanical response loss, residual stress loss, and cross-sectional equilibrium regularity term can be further considered. and residual stress smoothing regularization term The weighted sums are used to form the total loss. The weight coefficients in the total loss can be set according to the size of the training set. For example, the weight of the residual stress loss can be 1.0, the weight of the cross-sectional equilibrium regularization term can be 0.1, and the weight of the residual stress smoothing regularization term can be 0.01.

[0105] In one embodiment, as an example, if the scalar mechanical response loss of a training sample is 0.020, the residual stress loss is 0.150, and the cross-sectional equilibrium regularization term... The residual stress smoothing regularization term is 0.080. The weight of the residual stress loss is 0.030, the weight of the cross-sectional equilibrium regularization term is 1.0, the weight of the residual stress smoothing regularization term is 0.1, and the weight of the residual stress smoothing regularization term is 0.01. Therefore, the total loss of this sample can be calculated as follows: The calculation involves adjusting the scalar mechanical response prediction error, residual stress prediction error, and cross-sectional equilibrium deviation based on the total loss.

[0106] f. Furthermore, an adaptive moment estimation optimizer can be used to update all trainable parameters. The initial learning rate can be chosen as follows: The batch size can be 16 or 32, and the total number of training cycles can be 500 to 800. After each training cycle, the scalar mechanical response error and residual stress error are evaluated simultaneously using the validation set. The model parameters corresponding to the optimal comprehensive index are saved as the final output cold bending forming mechanical response prediction model.

[0107] It should be noted that simply pursuing a close fit of residual stress point values ​​can easily lead to results where local values ​​are close but the overall static equilibrium is not satisfied. This step places scalar mechanical response prediction and residual stress distribution prediction into the same training framework and adds cross-sectional equilibrium constraints. Based on this, the engineering reliability of the residual stress prediction results can be improved, and the prediction stability of highly correlated indicators such as springback can be improved simultaneously. Furthermore, the embodiments of the present invention also conduct a fine comparison of cross-sectional residual stress distribution prediction to verify the accuracy, smoothness, and physical rationality of the method of the present invention in the task of residual stress distribution prediction, especially the actual prediction effect after introducing the above-mentioned cross-sectional equilibrium constraints.

[0108] Because the residual stress distribution of the cross-section after cold bending requires not only accurate stress values ​​at each discrete point, but also the basic static equilibrium condition that the resultant force and resultant moment of the cross-section are approximately zero, otherwise the prediction results will lose their reference value in practical engineering. This invention selects a typical sample with a high degree of accumulated plastic history and compares its actual residual stress distribution with three prediction results: the traditional method, the variant method of this invention that removes equilibrium constraints, and the complete method of this invention, to verify the prediction effect. (Refer to...) Figure 3 The diagram showing the comparison of prediction results uses the following: The horizontal axis of the upper part of the figure represents the "position of the cross-section along the thickness direction," in millimeters. Zero values ​​represent the center of the plate thickness, negative values ​​represent the lower surface direction, and positive values ​​represent the upper surface direction. The vertical axis represents "residual stress," in megapascals (MPa). In the figure, the solid black dots represent the measured residual stress distribution, the dashed red squares represent predictions using traditional methods, the purple diamond-shaped dashed lines represent predictions from the present invention without equilibrium constraints, and the solid green triangles represent predictions from the complete method of the present invention.

[0109] As can be observed in the figure, the predicted curve of the traditional method deviates significantly from the measured values ​​at multiple locations, especially in the peak region of the compressive stress on the cross-section surface (to the left of the negative half-axis), where the predicted value is significantly lower, and the overall curve exhibits unreasonable and drastic fluctuations. The predicted curve of the unconstrained version of this invention can better match the overall trend of the measured values, but there are still slight deviations at some local locations, and the calculated resultant force of the cross-section deviates from zero. The predicted curve of the complete method of this invention not only has the highest degree of agreement with the measured values, but also fits particularly accurately at the peak of the surface compressive stress, and the curve trend is smooth, without the abnormal spikes found in the traditional method. The magnified area in the lower part of the figure focuses on the compressive stress region near the lower surface of the cross-section (the side of the plate subjected to compressive stress due to bending in the thickness direction, i.e., the inner surface of the bend), and the meaning of the horizontal and vertical axes is consistent with that of the upper part of the figure. The lower surface of the cross-section allows for a more detailed distinction of the differences between the four curves in this key region in the magnified view: the traditional method severely underestimates the stress peak, the unconstrained version of this invention improves slightly but still has deviations, while the complete method of this invention almost coincides with the measured values. Experimental results show that the cross-sectional equilibrium regularization term added to the loss function in this invention can penalize the degree to which the resultant force and resultant moment of the predicted residual stress distribution deviate from zero, thereby automatically guiding the prediction results toward a physically reasonable direction during the optimization process.

[0110] This invention proposes a method for predicting the correlation between cold bending forming process parameters and product mechanical properties, which is innovative compared to existing technologies in the following aspects: 1. A feature expression system based on the decoupling of plastic history index and material scaling is constructed. By defining the plastic history index with pass attenuation weight, the preceding cumulative plastic deformation is quantified into an independent feature. The material scaling factor is used to separate the geometric path and material scale effect, which solves the problem of poor comparability of process parameters under different plate thicknesses and strength specifications.

[0111] 2. An enhanced feature generation method based on local neighborhood projection is adopted. It does not rely on complex pre-trained networks. It maps high-dimensional sequence features to a low-dimensional stable subspace only through nearest neighbor retrieval and local covariance decomposition, which effectively suppresses the interference of roller parameter jumps and local noise on model stability.

[0112] 3. A plastic history gating mechanism is introduced, which transforms the plastic history summary vector into a channel response of shared features that dynamically adjusts the gating coefficient. This enables the model to adaptively amplify or suppress the expression of relevant features based on the cumulative degree of plasticity, thereby enhancing the ability to identify sensitive areas of springback and residual stress history.

[0113] 4. The loss function incorporates the balance constraints of the resultant force and resultant moment of the cross section, which improves the prediction of residual stress from a simple point value fitting to a physically consistent distribution that meets the static equilibrium condition. At the same time, the smoothing regularization term suppresses unreasonable spikes at adjacent discrete points, which significantly improves the engineering credibility of the prediction results.

[0114] Based on the above embodiments, this invention also provides a device for predicting the correlation between cold bending forming process parameters and product mechanical properties, referring to... Figure 4 The device includes: a parameter acquisition module 10, used to acquire material parameters of the sample to be predicted, and forming process parameters for multiple passes of forming the sample to be predicted; a first processing module 20, used to scale and decouple the geometric processing parameters of the forming process parameters based on the material yield strength and material thickness parameters in the material parameters, to obtain the decoupled geometric processing parameters corresponding to the corresponding passes; a second processing module 30, used to determine the plastic history index of the sample to be predicted before the start of each pass based on the pass spacing, bending angle and plate thickness in the forming process parameters; a data construction module 40, used to construct a multi-pass cold bending forming sequence of the sample to be predicted based on the plastic history index, decoupled geometric processing parameters and material parameters; and an execution module 50, used to input the multi-pass cold bending forming sequence into a pre-trained mechanical response prediction model, and use the mechanical response prediction model to predict the mechanical properties of the multi-pass cold bending forming sequence, to determine the predicted mechanical properties of the sample after forming.

[0115] The cold bending forming process parameter and product mechanical property correlation prediction device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0116] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figure 1 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figure 1 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 5 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figure 1 The method shown.

[0117] exist Figure 5 In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5The diagram uses only a single double-headed arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figure 1 The method shown.

[0118] The computer program product of the method and apparatus for predicting the correlation between cold bending forming process parameters and product mechanical properties provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, which will not be repeated here. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0119] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of this invention, used to illustrate the technical solutions of this invention, and not to limit it. The scope of protection of this invention is not limited thereto. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting the correlation between cold bending forming process parameters and product mechanical properties, characterized in that, The method includes: Obtain the material parameters of the sample to be predicted, and the molding process parameters of multiple passes for molding the sample to be predicted; Based on the material yield strength and material plate thickness parameters in the material parameters, the geometric processing parameters of the forming process parameters are scaled and decoupled to obtain the decoupled geometric processing parameters corresponding to the corresponding pass. Based on the pass spacing and bending angle in the molding process parameters and the plate thickness in the material parameters, the plasticity history index of the sample to be predicted is determined before the start of each pass. Based on the plasticity history index, the decoupled geometric processing parameters, and the material parameters, a multi-pass cold bending forming sequence of the sample to be predicted is constructed. The multi-pass cold bending sequence is input into a pre-trained mechanical response prediction model. The mechanical performance of the multi-pass cold bending sequence is predicted by the mechanical response prediction model to determine the predicted mechanical performance of the sample after forming.

2. The method according to claim 1, characterized in that, The steps for constructing a multi-pass cold bending sequence for the sample to be predicted based on the plastic history index, the decoupled geometric processing parameters, and the material parameters include: Based on the plasticity history index and plasticity history increment of each pass, as well as the relative position of the corresponding pass in the entire forming line, a historical position vector corresponding to the pass is constructed. The decoupled geometric processing parameters, the historical position vector, and the material parameters are concatenated to form the geometric decoupled pass feature vector corresponding to the pass. The geometrically decoupled feature vectors of all passes are stacked in pass order to form a multi-pass cold bending forming sequence of the sample to be predicted.

3. The method according to claim 1, characterized in that, Based on the pass spacing and bending angle in the molding process parameters and the plate thickness in the material parameters, the step of determining the plastic history index of the sample to be predicted before the start of each pass includes: Determine the equivalent bending radius of the current track based on the track spacing and bending angle of the corresponding track. The pass plastic increment is determined based on the plate thickness and the equivalent bending radius. The plasticity increments of all previous treads are weighted and accumulated to determine the historical plasticity index before the start of the current tread.

4. The method according to claim 1, characterized in that, The steps of inputting the multi-pass cold bending sequence into a pre-trained mechanical response prediction model, predicting the mechanical properties of the multi-pass cold bending sequence using the mechanical response prediction model, and determining the predicted mechanical properties of the sample to be predicted after forming include: Based on the strength of the plastic history corresponding to the multi-pass cold bending forming sequence, the gating coefficient vector of the mechanical response prediction model is determined. Based on the gate coefficient vector, the features to be predicted corresponding to the multi-pass cold bending forming sequence are determined; The feature to be predicted is input into the mechanical response prediction model. The scalar mechanical response output branch and the residual stress output branch of the mechanical response prediction model are used to perform branch prediction on the feature to be predicted, and the corresponding scalar mechanical response result and residual stress prediction result are determined. Based on the scalar mechanical response results and the residual stress prediction results, the post-molding mechanical property prediction results of the sample to be predicted are determined.

5. The method according to claim 4, characterized in that, The step of determining the post-molding mechanical property prediction results of the sample to be predicted based on the scalar mechanical response results and the residual stress prediction results includes: Based on the training label statistics corresponding to the scalar mechanical response result, determine the first predicted value corresponding to the scalar mechanical response result; Based on the training label statistics corresponding to the residual stress prediction result, a second predicted value corresponding to the residual stress prediction result is determined. The first predicted value and the second predicted value are determined as the predicted mechanical properties of the sample after molding.

6. The method according to claim 4, characterized in that, The method further includes: Obtain pre-constructed training samples; wherein the training samples contain sample labels, the sample labels including scalar mechanical response labels and residual stress labels; Based on the training samples and their labels, the mechanical response prediction model is trained to construct the scalar mechanical response output branch and the residual stress output branch corresponding to the training samples. The total loss used in the mechanical response prediction model training is constructed based on the scalar mechanical response loss, the residual stress loss, and the cross-sectional equilibrium constraint and the residual stress smoothing regularization term corresponding to the residual stress.

7. The method according to claim 6, characterized in that, The method further includes: Determine the local nearest neighbor set corresponding to the multi-pass cold bending forming sequence from the preset pass reference sample; Based on the local nearest neighbor set, the center projection vector and the perturbation scale vector corresponding to the multi-pass cold bending forming sequence are determined. Based on the center projection vector, construct the enhanced input features corresponding to the multi-pass cold bending forming sequence; or, combine the perturbation scale vector to add random perturbation to the enhanced input features to construct the enhanced input features after feature perturbation. Training samples are constructed based on the enhanced input features or the enhanced input features after feature perturbation in order to train the mechanical response prediction model.

8. A device for predicting the correlation between cold bending forming process parameters and product mechanical properties, characterized in that, The device includes: The parameter acquisition module is used to acquire the material parameters of the sample to be predicted, as well as the molding process parameters of multiple passes for molding the sample to be predicted. The first processing module is used to perform scaling and decoupling processing on the geometric processing parameters of the forming process parameters based on the material yield strength and material plate thickness parameters in the material parameters, so as to obtain the decoupled geometric processing parameters corresponding to the corresponding pass. The second processing module is used to determine the plasticity history index of the sample to be predicted before the start of each pass based on the pass spacing and bending angle in the molding process parameters and the plate thickness in the material parameters. The data construction module is used to construct a multi-pass cold bending forming sequence of the sample to be predicted based on the plasticity history index, the decoupled geometric processing parameters, and the material parameters. The execution module is used to input the multi-pass cold bending forming sequence into a pre-trained mechanical response prediction model, and use the mechanical response prediction model to predict the mechanical properties of the multi-pass cold bending forming sequence to determine the predicted mechanical properties of the sample after forming.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the method for predicting the correlation between cold bending forming process parameters and product mechanical properties as described in any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method for predicting the correlation between cold bending forming process parameters and product mechanical properties as described in any one of claims 1 to 7.

Images