A method and system for quality control of a photovoltaic connector metal core forming process

Abstract

The present application relates to the technical field of quality control, and more particularly to a quality control method and system for a photovoltaic connector metal core forming process. The method comprises the following steps: constructing a process space correlation model, generating a size deviation tensor through differential encoding, determining equivalent elastic-plastic state information in combination with material attribute parameters and thermal state data, performing residual propagation operation to predict a target size deviation value, decomposing into pressure, temperature and speed compensation components by using a decoupling control model, and writing into an execution unit and adaptively updating model parameters after being verified by a geometric boundary rule. The present application solves the problems of size deviation prediction and collaborative control in the continuous stamping forming of the multi-zone metal assembly of the photovoltaic connector, and realizes multi-dimensional decoupling adaptive control, thereby improving the prediction accuracy and control stability.

Description

Technical Field

[0001] This invention relates to the field of quality control technology, and in particular to a quality control method and system for the metal core forming process of photovoltaic connectors. Background Technology

[0002] Photovoltaic connectors are core components of the power transmission system in photovoltaic power plants. The forming precision of their metal cores directly determines the connector's electrical contact performance, mechanical mating life, and long-term operational reliability. Currently, photovoltaic connector metal cores are generally mass-produced using a continuous high-speed stamping process, which offers significant advantages in terms of high production efficiency and high material utilization.

[0003] Existing stamping quality control technologies mainly include two categories: offline die springback pre-compensation and online single-parameter feedback control. During the die design phase, technicians calculate material springback using finite element simulation and pre-correct the die cavity. During production, simple optical sensors detect a few key single-point dimensions of the finished product, and based on the detection results, adjust the die pressure or closing height individually for feedback compensation.

[0004] However, the metal core of a photovoltaic connector integrates multiple functional areas such as a crimping sleeve, spring, and transition fillet. These different areas have fundamentally different stress states, heat accumulation characteristics, and forming requirements. A single compensation strategy often results in dimensional discrepancies, where the crimping sleeve meets the specifications, but the spring is misaligned or has excessive springback. The plastic deformation heat and frictional heat generated by continuous high-speed stamping create localized heat accumulation fields within the mold cavity, leading to a nonlinear decrease in the material's dynamic yield strength. Traditional static material parameter models cannot accurately predict forming deviations. Furthermore, there is a strong coupling relationship between process parameters such as pressure, temperature, and feeding speed. Adjusting a single parameter can easily trigger system oscillations and deviation transfers. Additionally, mold wear and batch material variations can cause the control model to drift over time, making it difficult to maintain long-term stable production quality. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a quality control method and system for the forming process of photovoltaic connector metal cores. This method solves the problems of inaccurate springback prediction caused by local heat accumulation during continuous high-speed stamping, contradictory dimensions of multi-functional partitions, and quality fluctuations caused by strong coupling of process parameters. It improves the forming accuracy and production stability of photovoltaic connector metal cores and reduces the scrap rate.

[0006] In a first aspect, the present invention provides a quality control method for the forming process of a photovoltaic connector metal core, characterized in that the method comprises: S1. Obtain the thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and update the process spatiotemporal correlation model; perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; S2. Based on the material property parameters, the thermal data, and the dimensional deviation tensor, determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly; S3. Input the equivalent elastoplastic state information and the mechanical process parameters into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value and the mechanical process parameters, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component; S4. Convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional partition, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, write the pressure compensation component, the temperature compensation component, and the velocity compensation component into the corresponding execution unit respectively. S5. Update the model parameters of the process spatiotemporal correlation model based on the result of the geometric boundary rule verification.

[0007] Secondly, the present invention provides a quality control system for the forming process of a photovoltaic connector metal core, used to implement the above-mentioned method, the system comprising: Modeling construction units are used to acquire thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and to construct a spatiotemporal correlation model of the process. The deviation calculation unit is used to perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; The elastoplastic inversion unit is used to determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly based on the material property parameters, the thermal data, and the dimensional deviation tensor. The decoupled prediction unit is used to input the equivalent elastoplastic state information into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component; The verification execution unit is used to convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional zone, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, the pressure compensation component, the temperature compensation component, and the velocity compensation component are written into the corresponding execution unit respectively. The parameter update unit is used to update the model parameters of the process spatiotemporal correlation model based on the result of the geometric boundary rule verification.

[0008] Thirdly, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0009] The beneficial effects of this invention are as follows: This invention establishes a spatiotemporal correlation model to create a coupling relationship between the three dimensions of heat, force, and feeding in the spatiotemporal topology. The dimensional deviation tensor, generated by differential encoding of geometric feature data, is input into the elastoplastic inversion model along with the dynamic yield strength corrected based on thermal data. After stripping the elastic rebound, the equivalent elastoplastic state information of each functional zone is obtained. This equivalent elastoplastic state information is then used as node features in the spatiotemporal correlation model. The model obtains the deviation evolution characteristics of adjacent nodes in historical strokes through the three spatiotemporal edges of heat, force, and feeding. After attention-weighted aggregation, it applies corrections based on three physical consistency constraints: material continuity, volume invariance, and geometric topology, generating residual propagation characteristics. These characteristics are decoded and mapped into a deviation evolution trend vector for the next forming cycle, and then weighted and fused with the current dimensional deviation tensor to output the target value. The dimensional deviation value is marked; the predicted result is decomposed into three independent compensation components of pressure, temperature and speed through a decoupled control model. Before being written into the execution unit, the geometric boundary rule is checked on the predicted shape after superimposed compensation. The compensation command is only allowed to be executed if the closed interval of the pressing cylinder, the symmetry of the spring sheet and the overall assembly boundary all fall within the rigid threshold range. After the verification is passed, the residual between the actual measured value and the predicted target value is used to back-correct the connection weights and decoding parameters of the spatiotemporal correlation model of the process through gradient descent, so that the model can adaptively track the die wear and material batch changes during continuous stamping. If the verification fails, a conservative parameter rollback mechanism is immediately triggered, the compensation components are reset to safe initial values ​​and manual review is initiated. The model topology weights are corrected according to the feedback from the process engineer, and the model correction is completed under the premise of ensuring equipment safety. Through the cooperation of the above technical solutions, the accuracy of dimensional deviation prediction and long-term control stability during continuous high-speed stamping are significantly improved. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments 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 based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a quality control method for the metal core forming process of a photovoltaic connector, as shown in the embodiment. Figure 2 This is a curve showing the convergence of the predicted residuals during continuous stamping in the example. Figure 3 This is a structural diagram of a quality control system for the metal core forming process of a photovoltaic connector, as shown in the embodiment. Detailed Implementation

[0012] This invention provides a quality control method and system for the metal core forming process of a photovoltaic connector. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] For ease of understanding, the specific process of the embodiments of the present invention will be described below, such as... Figure 1 As shown in the figure, a quality control method for the forming process of a photovoltaic connector metal core in an embodiment of the present invention includes: S1. Obtain the thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and construct a process spatiotemporal correlation model; perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; In S1, the spatiotemporal correlation model of the construction process includes: The thermal data is obtained by collecting real-time temperature sequences from multiple sampling points inside the mold cavity using temperature sensing devices; the mechanical process parameters are obtained by simultaneously collecting forming force curves, mold closing height, and material feeding displacement during the stamping process using pressure and displacement sensing devices; and the geometric feature data is obtained by acquiring the outline image of the formed component using optical imaging devices and performing edge extraction processing on the outline image. Each functional zone of the multi-zone metal component is defined as a model node, and the cumulative relationships of heat conduction, mechanical traction, and feeding are defined as model edges. A spatiotemporal correlation model of the process is constructed based on the thermal data, the mechanical process parameters, and the geometric feature data.

[0014] Specifically, due to the dispersed and isolated nature of multi-source heterogeneous data during the metal core forming process of photovoltaic connectors, and the lack of spatiotemporal correlation, it is impossible to accurately reflect the inherent coupling relationship between various process parameters and forming quality.

[0015] To address the aforementioned technical challenges, the temperature sensing device employs a hybrid sensing network combining a high-response-speed thermocouple array and an infrared thermal imaging probe. The thermocouple array is arranged in a matrix topological configuration for each functional zone within the mold cavity, with at least one sampling point configured for each zone. The infrared thermal imaging probe captures the transient temperature of the exposed metal strip surface at the moment the mold opens. Each sampling point continuously records temperature values ​​at a millisecond-level sampling frequency. The temperature values ​​from each sampling point are then timestamped to generate a real-time temperature sequence for each sampling point. Furthermore, the real-time temperature sequences from all sampling points within the mold cavity are spatially aggregated to form a multi-dimensional data set containing spatial coordinates, time series, and temperature amplitude—i.e., thermal data.

[0016] The aforementioned pressure sensing equipment employs a piezoelectric force sensor mounted on the bottom of the punch press slide or the die base. The displacement sensing equipment includes a high-precision optical encoder mounted on the punch press guide post and a rotary encoder or laser displacement sensor mounted on the servo motor end of the feeder. Both the pressure and displacement sensing equipment are activated by a unified external trigger signal to collect data. Simultaneously, at the moment the slide descends and contacts the material, the forming force curve, the dynamic fluctuation of the die closing height, and the material feeding displacement are recorded. This data is then structured and packaged to obtain the mechanical process parameters.

[0017] The aforementioned optical imaging equipment is deployed on the exit side of the punch press and includes a high-resolution industrial area array camera, a telecentric lens, and a strobe ring light source or a back-facing parallel light source. During the extremely short window period when the metal strip steps into place and is stationary, the punch press controller sends a synchronous trigger pulse to the optical imaging equipment to obtain a clear, motion-free image of the component outline. After acquiring the image, a Gaussian filtering algorithm is first used to eliminate ambient light interference and random noise, and histogram equalization is used to enhance contrast. Subsequently, sub-pixel-level edge extraction processing is performed. Specifically, the gray-level gradient operator is first used to roughly locate the edges at the pixel level to form an initial edge band. Then, in the normal direction of the initial edge band, the spatial moment or gray-level centroid method is used to perform continuous interpolation calculation on the gray values ​​of discrete pixels. By finding the extreme points of the fitted curve, the edges are precisely located to the sub-pixel level, and the outer contour line of the component and the boundary line of the internal holes are extracted. These are then transformed into an ordered set of two-dimensional coordinate points, and geometric attribute labels for each functional area are attached to obtain geometric feature data. The aforementioned geometric attribute labels, such as "crimping cylinder - inner contour - arc segment", "spring sheet - left edge - straight segment", "transition fillet - outer contour - feature point" or "crimping cylinder - open end - corner point", enable subsequent dimensional deviation calculations to be oriented and restructured according to functional zones.

[0018] The above-mentioned spatiotemporal correlation model is a dynamic graph structure network model based on physical topology priors. It is used to characterize the multi-physics coupling effect and the evolution law of dimensional deviation during the stamping process. In the process of model construction, the multi-partition metal component is first divided into multiple functional partition nodes according to the geometric topology. Each node represents a specific spatial region and carries the multi-dimensional state characteristics of the region in the current forming cycle, namely the above-mentioned thermal data, mechanical process parameters and geometric feature data.

[0019] Next, the spatial edges of the model are defined. These spatial edges connect adjacent functional partition nodes, representing the physical interactions between different regions within the same molding cycle. For any two adjacent nodes, the weight of their spatial edges is determined by a weighted average of the equivalent thermal influence coefficient and the equivalent stiffness coupling coefficient after normalization. The equivalent thermal influence coefficient is calculated based on the steady-state heat conduction principle and is equal to the thermal conductivity of the mold material multiplied by the effective contact area between adjacent partitions, divided by the heat transfer path length. This coefficient quantifies the flux intensity of heat conduction from the high-temperature partition to the low-temperature partition. The equivalent stiffness coupling coefficient is calculated based on the deformation compatibility principle and is equal to the equivalent stiffness of adjacent partitions multiplied by the stress transfer ratio. This stress transfer ratio is determined based on the continuity ratio of the material cross-sectional area at the interface between the two partitions. This coefficient quantifies the mechanical tensile or compressive effects caused by the material continuity between adjacent partitions.

[0020] The maximum value among the equivalent thermal influence coefficients of all spatial edges in the current forming cycle is used as the denominator. The equivalent thermal influence coefficient of each edge is divided by this maximum value to obtain the normalized thermal influence relative factor. Similarly, the maximum value among the equivalent stiffness coupling coefficients of all spatial edges in the current forming cycle is used as the denominator. The equivalent stiffness coupling coefficient of each edge is divided by this maximum value to obtain the normalized stiffness coupling relative factor.

[0021] Subsequently, a fusion weight is assigned based on the degree of dominance of thermal accumulation and mechanical deformation within the current molding cycle. The degree of dominance of thermal accumulation is determined by the ratio of the overall temperature rise rate of the mold cavity to the historical average temperature rise rate, while the degree of dominance of mechanical deformation is determined by the deviation ratio between the current peak value of the forming force curve and the nominal peak value of the forming force. The relative factor of thermal influence is multiplied by the degree of dominance of thermal accumulation, and the relative factor of stiffness coupling is multiplied by the degree of dominance of mechanical deformation. The sum of these two products yields the comprehensive weight of this spatial edge.

[0022] Define the time edge of the model. The time edge connects nodes of the same functional zone between adjacent forming cycles, representing the cross-cycle inheritance effect of historical physical states. For any functional zone node, its time edge weight is determined by the cross-cycle inheritance coefficient, which is calculated based on the heat inheritance ratio and the wear accumulation factor. The heat inheritance ratio equals the heat remaining in the mold area of ​​the zone at the end of the previous forming cycle divided by the cumulative total heat generated in the area during the current cycle. The residual heat equals the product of the specific heat capacity of the mold material, the effective mass of the mold, and the difference between the mold temperature at the end of the cycle and the initial temperature rise. The cumulative total heat generated equals the product of the specific heat capacity of the mold material, the effective mass of the mold, and the difference between the peak temperature of the mold during the current cycle and the initial temperature rise. The wear accumulation factor equals the ratio of the current total number of stampings to the total number of stampings in the mold's design life. Multiplying the heat inheritance ratio and the wear accumulation factor yields a dimensionless cross-cycle inheritance coefficient, which is directly used as the time edge weight, thereby quantifying the intensity of the inheritance effect of heat accumulation and mold wear from the previous stamping on the dimensional deviation of the current stamping.

[0023] The above technical solution, by constructing a process spatiotemporal correlation model, realizes the organic integration of multi-source heterogeneous data, accurately reveals the spatiotemporal coupling relationship between various process parameters during the metal core forming process of photovoltaic connectors, provides a reliable basis for subsequent accurate location of quality anomalies and adaptive adjustment of process parameters, and effectively improves the comprehensiveness and accuracy of quality control.

[0024] Further, in S1, the generation of the size deviation tensor includes: The geometric feature data is subjected to subpixel-level coordinate transformation to determine the measured contour point set; the measured contour point set is registered with the nominal feature data to calculate the Euclidean distance vector between the corresponding reference points; the Euclidean distance vector is recombined in multiple dimensions according to the topology of the multi-partition metal component to obtain the size deviation tensor.

[0025] Specifically, since geometric feature data resides in the camera pixel coordinate system while nominal feature data resides in the physical world absolute coordinate system, there are coordinate system differences, optical distortion, and rigid body displacement errors between the two. Furthermore, simple numerical subtraction cannot preserve the spatial directionality and topological correlation of the deviation. Therefore, the camera intrinsic parameter matrix, distortion coefficient matrix, and extrinsic parameter matrix, pre-calibrated using a high-precision calibration board, are used. For each coordinate point in the geometric feature data, a nonlinear distortion inverse operation is first performed using the distortion coefficient matrix to restore the geometrically distorted coordinates caused by lens defects to an ideal, distortion-free projection surface. Then, the distortion-free pixel coordinates are multiplied by the pixel physical size using the intrinsic parameter matrix to convert them into camera physical coordinates in millimeters. Finally, the camera physical coordinates are uniformly transformed to a global world coordinate system with the mold reference pin as the origin using the extrinsic parameter matrix through rotation and translation transformations. After the above coordinate mapping, distortion correction, and spatial transformation, a set of measured contour points with unified physical dimensions is obtained.

[0026] The aforementioned nominal feature data originates from a computer-aided design model of a multi-zone metal component, containing a set of key control points discretized according to a preset sampling density. A nearest-neighbor iterative algorithm is employed, using the discrete key control point set in the nominal feature data as a fixed reference benchmark and the measured contour point set as a floating target. Iterative calculations are performed, specifically searching for the nearest neighbor point in the nominal point set for each point in the measured contour point set, establishing a matching relationship between corresponding points. Based on this matching relationship, the optimal rotation matrix and translation vector of the measured point set relative to the nominal point set are calculated to minimize the overall mean square error between corresponding points. This optimal rotation matrix and translation vector are then used to update the measured contour point set through rigid body transformation. This iterative calculation is repeated until the overall mean square error converges to a preset threshold, thereby eliminating the non-deformable rigid body displacement error generated during the component's feeding or stamping process, resulting in the registered measured contour point set.

[0027] The Euclidean distance vector between each measured point in the registered measured contour point set and its corresponding reference point in the nominal feature data is calculated. For each discrete reference point in the nominal feature data, the local tangent at that point is determined based on the tangent direction of the parameter curve of the CAD model where it is located. The direction perpendicular to this tangent is taken as the local normal, and a local normal coordinate system is established. In the local normal coordinate system, the perpendicular distance from the measured point to the tangent of the corresponding reference point is calculated as the normal distance. This normal distance reflects the dimensional deviation caused by excessive flow or insufficient springback of the material in the thickness or depth direction. The projected offset of the measured point in the tangent direction is calculated as the tangential distance. This tangential distance reflects the tensile deformation or positioning offset of the material in the length or circumference direction. The normal distance and the tangential distance are combined into a two-dimensional vector to obtain the Euclidean distance vector, which quantifies both the magnitude and spatial direction of the deviation.

[0028] Based on the topology of the multi-zone metal component, the aforementioned Euclidean distance vectors are recombined in multiple dimensions to obtain the dimensional deviation tensor. Specifically, the spatial topology map of the multi-zone metal component is extracted from its computer-aided design model, dividing the component into several interconnected functional zones, and defining the adjacency matrix between zones. Each Euclidean distance vector is mapped according to the spatial coordinates of its respective functional zone. For each functional zone, the normal and tangential distance components of all Euclidean distance vectors within it are extracted, and the maximum absolute value, arithmetic mean, and variance of each component are calculated to form the statistical feature set of that zone. A three-dimensional tensor structure is constructed. The first dimension arranges the functional zone nodes according to the connection order of the topology map; the second dimension arranges the normal and tangential distance components; and the third dimension arranges the three statistical feature types: maximum deviation, average deviation, and variance. The statistical feature sets of each functional zone are sequentially filled into this tensor to obtain the dimensional deviation tensor. This dimensional deviation tensor not only contains local deviation details but also implicitly includes the global deformation mode and the mutual restraint effect between zones.

[0029] The above technical solution eliminates coordinate system differences and rigid body displacement errors, transforming discrete contour deviations into high-dimensional structured data that retains spatial distribution gradients and partition constraints. This enables subsequent calculations based on the actual deformation directions and mutual constraints of each partition to deduce the spatial transmission path and temporal evolution trend of errors, thereby providing a physically interpretable structured and quantitative basis for predicting the target size deviation value of the next forming cycle.

[0030] S2. Based on the material property parameters, the thermal data, and the dimensional deviation tensor, determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly; specifically including: The local thermal accumulation field of the multi-zone metal component during continuous stamping is determined based on the cavity temperature distribution in the thermal data; the dynamic yield strength of each functional zone in the multi-zone metal component is calculated by combining the material property parameters and the local thermal accumulation field. The dimensional deviation tensor and the dynamic yield strength of each functional zone are input into a preset elastoplastic inversion model to obtain the plastic deformation of each functional zone; based on the plastic deformation and the dynamic yield strength, the elastic rebound of each functional zone after unloading is determined. The plastic deformation and elastic rebound are concatenated using multidimensional feature vectors to generate the equivalent elastoplastic state information.

[0031] Specifically, after generating the dimensional deviation tensor, relying solely on surface geometric deviations cannot reveal the intrinsic mechanical root cause of dimensional deviations, nor can it distinguish the contributions of plastic deformation and elastic rebound to the final dimensions. Therefore, this step uses multi-source data fusion to invert the equivalent elastic-plastic state inside the material, analyzes the physical causes of dimensional deviations, and provides a data basis for subsequent deviation prediction.

[0032] Based on the cavity temperature distribution in the aforementioned thermal data, the local thermal accumulation field of the multi-zone metal component during continuous stamping is determined. Specifically, the steady-state temperature values ​​at the end of the current forming cycle and the temperature peak envelopes of several past consecutive stamping cycles are extracted from the thermal data. A three-dimensional temperature distribution cloud map inside the cavity is constructed using radial basis function interpolation. Radial basis function interpolation is a general numerical method for generating a continuous spatial field based on discrete sampling points, which can accurately fit the isotropic heat conduction law inside the metal mold. The first derivative of the temperature peak envelopes from historical stamping cycles is calculated to obtain the transient temperature rise rate of each functional zone. The temperature values ​​of each functional zone in the three-dimensional temperature distribution cloud map are multiplied by the corresponding transient temperature rise rate to generate the aforementioned local thermal accumulation field.

[0033] Combining the aforementioned material property parameters and the local thermal accumulation field, the dynamic yield strength of each functional zone in the multi-zone metal assembly is calculated. Dynamic yield strength is the critical stress at which a material begins to undergo plastic deformation under specific temperature conditions, and its value decreases non-linearly with increasing temperature. Material property parameters are pre-stored in the system, including the initial yield strength, thermal softening coefficient, elastic modulus, and hardening index of the metal strip at room temperature. For each functional zone, the equivalent average temperature of that zone is extracted from the local thermal accumulation field. This equivalent average temperature is then substituted into the temperature term of the Johnson-Cook thermoviscoplastic constitutive equation for calculation. The Johnson-Cook thermoviscoplastic constitutive equation is a general mathematical model describing the mechanical behavior of materials under high temperature and dynamic loads, accurately reflecting the softening effect of temperature on the yield strength of materials. The formula for calculating the temperature term is as follows: in, For dynamic yield strength, The initial yield strength is given by , where T is the equivalent average temperature of the functional zones. For reference temperature, Let m be the melting point of the material and m be the thermal softening coefficient. Finally, the dynamic yield strength distribution vector of each functional zone is obtained.

[0034] The aforementioned dimensional deviation tensor is mapped to a preset elastoplastic inversion model, and the plastic deformation of each functional zone is determined by combining the aforementioned dynamic yield strength. The plastic deformation is a physical quantity that characterizes the degree of irreversible deformation of the material after forming. The elastoplastic inversion model is an inverse solution neural network model built on the large deformation theory of continuum mechanics. Its training data consists of finite element simulation results under different thermal states and process parameters, including the corresponding dimensional deviation tensor, dynamic yield strength, and true plastic strain field. The training objective is to minimize the mean square error between the model's predicted plastic strain field and the simulated true plastic strain field. The model inputs are the dimensional deviation tensor and the dynamic yield strength of each zone, and the output is the plastic strain field of each zone. In practice, the dimensional deviation tensor is input as the target boundary condition into the elastoplastic inversion model, and the dynamic yield strength of each zone is input as the yield criterion threshold into the model. The model starts an iterative solution process. First, an initial plastic strain field is assumed, and the theoretical geometry corresponding to the strain field is derived forward. The residual between the theoretical geometry and the dimensional deviation tensor is calculated. The gradient descent method is used to continuously adjust the plastic strain field until the residual converges to the preset threshold. At this point, the output plastic strain field is the amount of plastic deformation of each functional zone.

[0035] Based on the aforementioned plastic deformation and dynamic yield strength, the elastic rebound of each functional zone after unloading is determined. Elastic rebound is the geometric deformation caused by the release of elastic strain energy within the material after mold unloading, and it is positively correlated with dynamic yield strength and plastic deformation. Specifically, the plastic deformation and dynamic yield strength of each functional zone are extracted, and an unloading stiffness matrix is ​​constructed using the elastic modulus in the material properties. Considering that the apparent elastic modulus degrades after severe plastic deformation, the unloading stiffness matrix is ​​corrected for damage reduction based on the magnitude of the plastic deformation. The damage reduction coefficient is linearly negatively correlated with the plastic deformation; the larger the plastic deformation, the smaller the reduction coefficient. Using the dynamic yield strength as the upper limit of stress at the unloading starting point, the strain recovery corresponding to complete stress release is calculated using the corrected unloading stiffness matrix. The strain recovery is converted into geometric displacement to obtain the elastic rebound of each functional zone. The calculation formula is as follows: Where ΔR is the elastic rebound and α is the damage reduction factor. The plastic deformation amount is E, the elastic modulus of the material is E, and the nominal characteristic data of the functional partition are the characteristic geometric dimensions of each functional partition in the stamping deformation direction, such as the wall thickness of the press cylinder and the cantilever length of the spring sheet. The plastic deformation amount and the elastic rebound amount are spliced ​​together as multi-dimensional feature vectors to generate the equivalent elastoplastic state information. Specifically, the feature vectors of all functional partitions are spliced ​​together according to the topological order of the multi-partition metal component to obtain high-dimensional comprehensive state information containing the degree of irreversible deformation and elastic recovery capability of each partition, that is, the equivalent elastoplastic state information.

[0036] The above technical solution, by inverting the equivalent elastic-plastic state inside the material, eliminates the nonlinear influence of local heat accumulation on the material's mechanical properties during continuous stamping, accurately separates the contributions of plastic deformation and elastic springback to the final size, and provides a core physical basis for subsequent prediction of residual propagation deviation under physical constraints. It effectively solves the technical problem that traditional static models cannot handle the inaccurate springback prediction caused by heat accumulation effects, thereby improving the quality of the forming process.

[0037] S3. Input the equivalent elastoplastic state information into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component; In S3, the target dimensional deviation value for the next molding cycle is predicted, including: The equivalent elastoplastic state information is input into the process spatiotemporal correlation model. For each node in the process spatiotemporal correlation model, the deviation evolution characteristics of its adjacent nodes in historical impulses are obtained. The deviation evolution characteristics include the deviation amplitude and the deviation direction. In the spatiotemporal correlation model of the process, feature aggregation processing modified by physical consistency constraints is performed to obtain residual propagation features. The physical consistency constraints include material continuity constraints, volume invariance constraints, and geometric topology constraints. The deviation evolution trend vector for the next forming cycle is calculated based on the residual propagation characteristics; the deviation evolution trend vector is then weighted and fused with the size deviation tensor of the current cycle to determine the target size deviation value.

[0038] Specifically, the equivalent elastoplastic state information of each functional zone is input into the process spatiotemporal correlation model. This process spatiotemporal correlation model is a pre-built and continuously updated dynamic graph network structure. During the model training phase, the equivalent elastoplastic state information obtained from thermal data, mechanical process parameters, and geometric feature data through differential encoding and elastoplastic inversion processing in multiple consecutive molding cycles is used as the node input features. The historical deviation labels of adjacent nodes are constructed using the size deviation tensor sequence of the corresponding cycle. The actual measured size deviation tensor of the next molding cycle is used as the supervision label. The training task is to predict the size deviation of the next molding cycle based on the equivalent elastoplastic state information of each node and the historical deviation evolution characteristics of adjacent nodes. The loss function uses the mean square error between the predicted value and the actual measured value. The spatiotemporal adjacent edge weights, message passing layer parameters, and decoding layer parameters are jointly optimized using the backpropagation gradient descent method, enabling the model to learn the propagation law and dissipation characteristics of errors in the spatiotemporal dimension. For each target node in the model, the system traverses all its adjacent nodes in the three dimensions of heat, force, and feeding. It extracts the dimensional deviation tensor sequence of each adjacent node in the past several consecutive molding cycles from the historical database, and performs feature deconstruction on the sequence to obtain the deviation evolution features. The deviation evolution features include deviation amplitude and deviation direction. The deviation amplitude is obtained by calculating the Euclidean norm of each deviation vector in the historical dimensional deviation tensor, which is used to characterize the absolute degree of dimensional deviation from the nominal value. The deviation direction is obtained by extracting the spatial axial component of each deviation vector in the historical dimensional deviation tensor and normalizing it to a unit direction vector, which is used to characterize the specific direction of the dimensional offset in three-dimensional space.

[0039] After acquiring the equivalent elastoplastic state information of each node and the deviation evolution characteristics of all adjacent nodes, the model performs feature aggregation processing to fuse spatiotemporal neighborhood information. This processing first employs an attention mechanism, which is a computational mechanism that adaptively allocates information aggregation weights based on the strength of physical associations between nodes. The model calculates the attention coefficient of each adjacent node relative to the target node based on the spatiotemporal adjacency edge weights of the three dimensions of heat, force, and feeding. This coefficient quantifies the proportion of the influence of the historical deviations of adjacent nodes on the future state of the target node in different physical dimensions. The deviation evolution characteristics of each adjacent node are weighted and summed using this coefficient to generate a preliminary neighborhood influence vector. Subsequently, the model concatenates and maps the neighborhood influence vector with the equivalent elastoplastic state feature vector of the target node itself. During the mapping process, three major physical consistency constraints are forcibly applied through the model's built-in constraint correction mapping operator. Among them, the material continuity constraint requires that the displacement field and strain field of adjacent nodes at the boundary remain continuous. This constraint is achieved by applying a consistency penalty to the displacement gradient of adjacent nodes in the mapping operator, thereby eliminating non-physical predictions of material tearing or overlapping at the algorithm level. The volume invariance constraint is based on the physical law that the volume of metallic materials remains basically constant during the plastic deformation stage. It requires that the component representing the volume expansion rate in the mapped feature approaches zero. If the aggregation calculation causes a certain region to show a stretching trend in the in-plane direction, this constraint automatically corrects its thickness direction feature to maintain volume balance. The geometric topology constraint ensures that the relative spatial position relationship of each functional zone of the multi-zone metal component does not change fundamentally after deformation. This constraint is achieved by setting the degree of freedom components that violate the fixed connection or relative position topology relationship to zero in the mapped feature vector. After the above-mentioned constraint correction, the original purely mathematical aggregation characteristics are given real physical meaning. The final output residual propagation characteristics are high-dimensional node latent variable representations that contain the error propagation path and cumulative effect in the spatiotemporal network. This representation not only retains the physical essence of how local deformation triggers global chain reactions through the material continuum, but also satisfies the basic conservation laws of continuum mechanics.

[0040] After obtaining the residual propagation features, the model maps them back to three-dimensional physical space through a decoding network to calculate the deviation evolution trend vector for the next forming cycle. This decoding network is a pre-trained multi-layer fully connected perceptron structure in the process spatiotemporal correlation model. Its input is the high-dimensional residual propagation features of each node, and its output is the expected deviation increment in each geometric dimension of the corresponding functional partition within the next forming cycle. During the training phase, the decoding network is jointly optimized with the graph network. The training label is the difference sequence between the actual measured size deviation tensor of the next forming cycle and the size deviation tensor of the current cycle. The training objective is to ensure that the trend vector output by the network accurately reproduces the slope and direction of error evolution in historical data. During the inference phase, the decoding network receives the residual propagation features and performs forward computation, performing affine transformations and nonlinear activations layer by layer, ultimately outputting a deviation evolution trend vector. Each component of this vector corresponds to the error development rate and direction of each functional partition in a specific geometric dimension, revealing the evolution trend of the stamping system itself without external intervention.

[0041] To determine the target dimensional deviation value ultimately used to guide equipment operation, the deviation evolution trend vector and the dimensional deviation tensor of the current cycle are weighted and fused. Since the deviation evolution trend vector represents the expected incremental change in deviation for the next cycle, while the dimensional deviation tensor represents the absolute geometric deviation of the current forming cycle, adaptive weight allocation is performed based on the sensitivity matrix of each functional zone. This sensitivity matrix refers to the weight allocation basis obtained by calibrating the statistical distribution characteristics and evolution rate of dimensional deviations in each geometric dimension from historical data. It is used to quantify the contribution priority of the current deviation and trend increment in the fusion process. Based on this matrix, it is assessed whether the current relative deviation index is approaching the tolerance boundary and whether the rate of change of the relative trend index exceeds the stability threshold. Dimensions where the deviation is close to the boundary are given a higher weight to ensure immediate correction, while dimensions with rapidly evolving trends are given a higher weight to achieve proactive suppression. The weight coefficients of each dimension are dynamically calculated based on the above evaluation indicators using a normalized exponential function, ensuring that all are positive and sum to one. Furthermore, the fusion process incorporates a time phase compensation factor. This factor is a pre-calibrated advance correction coefficient based on the mechanical response hysteresis characteristics of the stamping equipment's actuator. It is used to perform phase advance correction on the deviation evolution trend vector, ensuring that the weighted fusion target size deviation value can be synchronously matched with the equipment's operating cycle. After adaptive weighting and phase correction, the final output target size deviation value comprehensively reflects the superposition effect of the current static error base and the future dynamic error increment, providing a precise quantitative basis for subsequent decoupling control.

[0042] The above technical solution enables accurate prediction of the evolution trend of dimensional deviation in the next forming cycle, effectively solving the technical problem that traditional single stamping independent control cannot capture historical cumulative effects and multi-zone coupling interference, thereby significantly improving the accuracy of dimensional deviation prediction and long-term control stability of complex multi-zone metal components during continuous high-speed stamping.

[0043] Further, in S3, the total compensation amount is decomposed into a pressure compensation component, a temperature compensation component, and a velocity compensation component, including: The target size deviation value is input into a preset decoupling control model; the target is optimized based on the Jacobian sensitivity matrix through the decoupling control model to determine the first compensation weight corresponding to the mold pressure, the second compensation weight corresponding to the mold temperature, and the third compensation weight corresponding to the feeding speed. The total compensation amount is linearly decomposed based on the first compensation weight, the second compensation weight, and the third compensation weight to obtain the pressure compensation component, the temperature compensation component, and the speed compensation component.

[0044] Specifically, after obtaining the target dimensional deviation value, this value is an error vector in the geometric dimension, while the stamping equipment actuator can only receive control commands in the mechanical, thermal, or kinematic dimensions. The stamping forming of multi-zone metal components is a multivariable, strongly coupled system; adjusting a single process parameter will trigger a chain reaction of changes in multiple geometric dimensions. Traditional single-loop independent control easily leads to system oscillations and deviation transfer. Therefore, the target dimensional deviation value needs to be inversely decomposed into three independent compensation components—pressure, temperature, and speed—through a decoupled control model to solve the technical problem of mutual interference between compensation strategies in a multivariable, strongly coupled system.

[0045] During implementation, the aforementioned target size deviation values ​​are input into a pre-defined decoupling control model. This decoupling control model is an inverse dynamics algorithm framework based on the Jacobian sensitivity matrix and constrained least squares optimization. It is pre-calibrated using offline stamping experimental data and finite element simulation data. The input data consists of the target size deviation values ​​for each functional zone, and the output data consists of pressure compensation components, temperature compensation components, and velocity compensation components. Internally, this model encapsulates the quantitative relationships between the unit increments of pressure, temperature, and velocity on the geometric dimensions of each functional zone under the current elastoplastic state and thermal accumulation field.

[0046] Target optimization is performed based on the Jacobian sensitivity matrix using a decoupled control model to determine the first, second, and third compensation weights. Since the target size deviation and process parameters belong to different dimensions (length, force, temperature, and velocity), directly combining them would lead to dimensional confusion. Therefore, before optimization, each column of the Jacobian sensitivity matrix is ​​divided by the nominal adjustment range of the corresponding process parameter, and the target size deviation is divided by the nominal tolerance upper limit to achieve dimensionless optimization. The nominal adjustment range is the maximum allowable adjustment range of the process parameter relative to the reference value within the current equipment safety range, and the nominal tolerance upper limit is the allowable deviation limit of the geometric dimensions of each functional zone. The dimensionless target size deviation is used as the objective function vector, and the dimensionless Jacobian sensitivity matrix is ​​used as the coefficient matrix. A least-squares problem is solved with physical boundary constraints of the process parameters, i.e., pressure, temperature, and velocity boundary constraints under normal operating conditions, to minimize the residual deviation norm. This yields the dimensionless optimal process parameter adjustment vector. This is the dimensionless Jacobian sensitivity matrix, where the rows correspond to the geometric dimension deviation of each functional zone, and the columns correspond to the pressure. ,temperature ,speed The matrix elements, representing the dimensionless geometric dimensional changes caused by a unit increment of the dimensionless process parameters under the current operating conditions, are defined by three process parameter dimensions. This represents the dimensionless vector of target size deviation values. The sum of the absolute values ​​of the three values ​​is taken as the total compensation amount Q, and the first compensation weight is calculated for each. =∣ | / Q、Second compensation weight w2=| | / Q、Third compensation weight w3=| | / Q, this weight represents the relative contribution of each process parameter to the total compensation.

[0047] The total compensation amount is linearly decomposed based on the first, second, and third compensation weights to obtain pressure, temperature, and velocity compensation components. The total compensation amount Q is then multiplied by each of the three compensation weights to obtain three dimensionless allocation quantities; each allocation quantity is then assigned a corresponding dimensionless adjustment intensity. , , The original symbols are then multiplied by the nominal adjustment range of the corresponding process parameters to restore the physical dimensions, yielding the pressure compensation component ΔP, the temperature compensation component ΔT, and the velocity compensation component Δv. This decomposition ensures that the sum of the dimensionless intensities of the three compensation components equals the total compensation, and each component retains the adjustment direction and amplitude obtained from the original optimization, achieving a complete distribution of the total compensation energy across the three physical dimensions.

[0048] The above technical solution, through the decoupling control model, reversely maps the target size deviation value in the geometric dimension into three independent compensation components: pressure, temperature, and speed. Based on the normalization optimization of the Jacobian sensitivity matrix, it eliminates harmful interference in a multivariable strongly coupled system, solves the technical problem of system oscillation and deviation transfer caused by the adjustment of a single parameter, and lays the foundation for improving the quality of metal core forming.

[0049] S4. Convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional partition, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, write the pressure compensation component, the temperature compensation component, and the velocity compensation component into the corresponding execution unit respectively. In S4, geometric boundary rule verification is performed, including: The pressure compensation component, temperature compensation component, and velocity compensation component are converted into expected geometric correction values ​​for each functional zone. The expected geometric correction values ​​are then algebraically superimposed with the current period size deviation tensor to obtain the compensated expected residual deviation field. Combined with the nominal feature data, the predicted geometric shape is obtained. Extract the geometric features from the predicted geometry, including the closed interval value corresponding to the crimping cylinder region, the symmetry value corresponding to the spring plate region, and the assembly boundary value corresponding to the overall structure; determine whether the closed interval value, the symmetry value, and the assembly boundary value are all within their respective preset rigidity threshold ranges; if the determination result is yes, then the geometric boundary rule verification is confirmed to be passed.

[0050] Specifically, before writing the pressure compensation component, temperature compensation component, and speed compensation component into the execution unit, although the decoupled control model and the process spatiotemporal correlation model have undergone rigorous algorithm design, there is still a risk of calculating incorrect compensation amounts due to sensor noise, material batch anomalies, or local overfitting of the model in complex industrial sites.

[0051] To address the aforementioned technical issues, the linearized mapping relationship between process parameters and geometric dimensions obtained from the decoupled control model calibration is utilized. The pressure, temperature, and velocity compensation components in the total compensation are then converted into expected geometric corrections for each functional zone. This conversion is based on the geometric response corresponding to a unit process parameter increment, as represented by the Jacobian sensitivity matrix. This expected geometric correction is then algebraically superimposed with the current cycle size deviation tensor to obtain the compensated expected residual deviation field. Based on the nominal feature data, the expected residual deviation field is decoded into predicted coordinates for each contour point, generating the predicted geometric shape after superimposing the total compensation.

[0052] Three key geometric values ​​are extracted from the predicted geometry. The first corresponds to the closed-section value of the crimping cylinder area, obtained by calculating the chord distance between two feature points at the open end of the crimping cylinder's inner contour. This value directly determines the contact resistance and pull-out force after the crimping cylinder is crimped with the external wire. If it is too large, the crimping will be loose; if it is too small, the wire may be easily cut. The second corresponds to the symmetry value of the spring contact area, obtained by establishing a symmetry center reference plane for the spring contact design, i.e., the vertical plane where the spring contact's geometric center axis is located. The maximum normal offset distance of the two sides of the spring contact's contour lines relative to this center reference plane is extracted, and the absolute value of the difference between the two sides' maximum normal offset distances is calculated. This value determines the uniformity of the connector's insertion and extraction force; if it exceeds the tolerance, it will lead to uneven insertion and extraction force and reduced mechanical life. The third corresponds to the assembly boundary value of the overall structure, obtained by calculating the envelope rectangle size of the component's maximum outer contour, the center distance of key positioning holes, and the relative position coordinates between each functional zone. This value determines whether the component can be assembled into the insulating plastic shell without interference.

[0053] The values ​​of the closed interval, symmetry, and assembly boundary are compared with their respective preset rigid threshold ranges. The rigid threshold range is an absolute safety baseline pre-set based on product failure mode analysis, with its upper and lower limits corresponding to the critical geometric limits of product functional failure, leaving an appropriate safety margin compared to the nominal tolerances on the drawings. Only when all three values ​​simultaneously fall within their respective rigid threshold ranges is the geometric boundary rule verification considered passed, allowing the compensation component to be written into the physical execution unit; if any value exceeds the corresponding rigid threshold range, the verification fails, triggering a subsequent conservative parameter rollback mechanism. Through this technical solution, a forward-looking virtual trial model is performed on the superimposed compensation prediction results in a digital twin environment. Using three core functional indicators—the closed interval of the press-fit cylinder, the symmetry of the spring sheet, and the overall assembly boundary—as the basis for safety verification, the technical problem of unpredictable safety of the compensation strategy is solved. This provides a reliable safety gate for physical execution, effectively intercepting extreme compensation commands caused by model prediction deviations or abnormal operating conditions, and significantly reducing the risk of mold collision damage and scrap.

[0054] S5. Update the model parameters of the process spatiotemporal correlation model based on the results of the geometric boundary rule verification; specifically including: If the result of the geometric boundary rule verification is not passed, the conservative parameter rollback mechanism is triggered to reset the pressure compensation component, the temperature compensation component and the velocity compensation component to the preset safe initial value. Initiate a manual review signal and adjust the connection weights of each node in the spatiotemporal correlation model of the process based on the correction data provided by the manual reviewer. If the result of the geometric boundary rule verification is passed, the actual geometric feature data of the next forming cycle is obtained, and the residual between the target size deviation value and the actual measured value is calculated; the model parameters of the process spatiotemporal correlation model are updated according to the residual using the gradient descent method.

[0055] Specifically, after the geometric boundary rule verification is completed, the process spatiotemporal correlation model needs to be updated based on feedback, regardless of whether the verification passes or fails. During continuous stamping, die wear, lubricant performance degradation, and batch-to-batch material differences can cause the model to gradually deviate from the actual physical laws. Without a closed-loop adaptive update mechanism, the prediction accuracy will continue to deteriorate. At the same time, directly continuing the error compensation strategy when the verification fails will lead to equipment safety risks. Therefore, it is necessary to perform conservative backtracking and manual correction according to the dual-branch path of the verification results, or to perform gradient descent updates based on residuals, to solve the technical problems of model drift and safety adaptation under abnormal operating conditions.

[0056] During implementation, if the geometric boundary rule verification fails, a conservative parameter rollback mechanism is triggered, replacing the currently calculated pressure compensation component, temperature compensation component, and speed compensation component with preset safety initial values. These safety initial values ​​are derived from the basic process parameters that have been repeatedly verified during the initial debugging phase to ensure stable equipment operation, or from the moving average of the corresponding process parameters over several stable stamping cycles. This forces the stamping process back to a known and controllable physical state, avoiding destructive impacts from extreme compensation commands on the mold and equipment.

[0057] Simultaneously, a manual review signal is activated, a high-priority alarm prompt pops up on the control interface, and the on-site audible and visual alarm device is activated. After manual intervention, the thermal data, mechanical process parameters, geometric feature data, and abnormal compensation amounts predicted by the model at the time of alarm triggering are retrieved for root cause analysis. Based on actual measurement results and experience judgment, correction data is input into the system. This correction data is a correction ratio coefficient for the physical coupling relationship between specific functional zones. According to this correction ratio coefficient, the connection weights of each node in the process spatiotemporal correlation model are adjusted. This connection weight includes the spatial edge weight representing the physical interaction between adjacent zones within the same molding cycle, and the temporal edge weight representing the historical cumulative effect across cycles. If it is determined manually that the heat conduction effect between two adjacent functional zones is overestimated by the model, the fusion ratio of the equivalent thermal influence coefficient of the spatial edge is multiplied by a correction ratio coefficient less than 1; if it is determined that the mechanical traction effect is underestimated, the fusion ratio of the equivalent stiffness coupling coefficient of the spatial edge is multiplied by a correction ratio coefficient greater than 1; if it is determined that the cross-cycle thermal genetic intensity or wear genetic intensity is misjudged, the temporal edge weight is multiplied by the corresponding correction ratio coefficient. By introducing prior knowledge from human experts, the topological weights at the bottom layer of the model are forcibly intervened, enabling the model to re-establish the correlation logic that conforms to the current actual physical state.

[0058] If the geometric boundary rule verification passes, after the next molding cycle, the component outline image after molding is acquired again using an optical imaging device. Subpixel-level edge extraction is then performed to obtain the actual geometric feature data, and the actual size deviation tensor is generated according to the aforementioned differential encoding process. This actual size deviation tensor is subtracted from the target size deviation value predicted in the previous cycle, corresponding one-to-one by functional partition and geometric direction, to obtain the residual vector. This residual vector precisely quantifies the magnitude and direction of the model prediction error. The model parameters of the process spatiotemporal correlation model are updated using gradient descent. Specifically, the loss function is constructed as half the sum of the squares of the components of the residual vector. The model parameters to be updated in the process spatiotemporal correlation model include the connection weight matrix of the spatial and temporal edges of each node, and the connection weights in the decoding network used to map the residual propagation features to the deviation evolution trend vector. The gradient of the loss function with respect to the above parameters is calculated, and the parameters are updated along the negative gradient direction with a preset learning rate step size to reduce the loss function value. This iteration continues until convergence or a preset number of updates is reached, thereby completing the online adaptive evolution of the model parameters. Figure 2 As shown, the variation curves of the model prediction residuals over 30 consecutive forming cycles are presented. The residuals of the traditional method show no obvious convergence trend; however, the method of this invention, through residual propagation calculation and online updating via gradient descent, causes the prediction residuals to decrease rapidly, stabilize after the 10th cycle, and remain in a convergent state. This demonstrates that the above-mentioned adaptive update strategy for model parameters can effectively suppress model drift and significantly improve the long-term accuracy and stability of dimensional deviation prediction during continuous stamping.

[0059] The above technical solution, through a dual-branch verification mechanism, reverts to a safe state in abnormal situations and introduces human expert knowledge to correct the model topology. In normal situations, it automatically optimizes model parameters based on measured residuals, thus solving the technical problems of model drift and safe self-adaptation under abnormal working conditions, thereby improving the quality of the metal forming process.

[0060] This invention also provides a quality control system for the forming process of photovoltaic connector metal cores, used to implement the above-mentioned method, such as... Figure 3 As shown, the system includes: Modeling construction units are used to acquire thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and to construct a spatiotemporal correlation model of the process. The deviation calculation unit is used to perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; The elastoplastic inversion unit is used to determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly based on the material property parameters, the thermal data, and the dimensional deviation tensor. The decoupled prediction unit is used to input the equivalent elastoplastic state information into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component; The verification execution unit is used to convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional zone, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, the pressure compensation component, the temperature compensation component, and the velocity compensation component are written into the corresponding execution unit respectively. The parameter update unit is used to update the model parameters of the process spatiotemporal correlation model based on the result of the geometric boundary rule verification.

[0061] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0062] In summary, this invention establishes a spatiotemporal coupling relationship between three dimensions—thermal, mechanical, and feeding—on the spatiotemporal topology by constructing a process spatiotemporal correlation model. The dimensional deviation tensor generated from geometric feature data through differential encoding, along with the dynamic yield strength corrected based on thermal data, is input into the elastoplastic inversion model. After stripping away elastic rebound, the equivalent elastoplastic state information of each functional zone is obtained. This equivalent elastoplastic state information is then used as node features in the process spatiotemporal correlation model. The model obtains the deviation evolution characteristics of adjacent nodes in historical strokes through the three spatiotemporal edges of thermal, mechanical, and feeding. After attention-weighted aggregation, it applies corrections based on three physical consistency constraints: material continuity, volume invariance, and geometric topology, generating residual propagation characteristics. These characteristics are decoded and mapped into a deviation evolution trend vector for the next forming cycle, and then weighted and fused with the current dimensional deviation tensor before being input into the model. The system calculates the target dimensional deviation value. The predicted result is decomposed into three independent compensation components—pressure, temperature, and speed—via a decoupled control model. Before being written into the execution unit, the geometric boundary rules of the superimposed compensation prediction are validated. The compensation command is only allowed to execute if the closed section of the press-fit cylinder, the symmetry of the spring sheet, and the overall assembly boundary all fall within the rigidity threshold range. After successful validation, the residual between the actual measured value and the predicted target value is used to reverse-correct the connection weights and decoding parameters of the spatiotemporal correlation model through gradient descent, enabling the model to adaptively track die wear and material batch changes during continuous stamping. If the validation fails, a conservative parameter rollback mechanism is immediately triggered, resetting each compensation component to a safe initial value and initiating manual review. Based on feedback from process engineers, the model topology weights are corrected, completing model correction while ensuring equipment safety. Through the synergy of these technical solutions, the accuracy of dimensional deviation prediction and long-term control stability during continuous high-speed stamping are significantly improved.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0064] If the integrated unit is implemented as 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 the present invention, in essence, or the part that contributes to the prior art, or all 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, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications 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 the present invention.

Claims

1. A quality control method for the forming process of a photovoltaic connector metal core, characterized in that, The method includes: S1. Obtain the thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and construct a process spatiotemporal correlation model; perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; S2. Based on the material property parameters, the thermal data, and the dimensional deviation tensor, determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly; S3. Input the equivalent elastoplastic state information into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value and the mechanical process parameters, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component. S4. Convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional partition, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, write the pressure compensation component, the temperature compensation component, and the velocity compensation component into the corresponding execution unit respectively. S5. Update the model parameters of the process spatiotemporal correlation model based on the result of the geometric boundary rule verification.

2. The method according to claim 1, characterized in that, In S1, a spatiotemporal correlation model of the process is constructed, including: The thermal data is obtained by collecting real-time temperature sequences from multiple sampling points inside the mold cavity using temperature sensing devices; the mechanical process parameters are obtained by simultaneously collecting forming force curves, mold closing height, and material feeding displacement during the stamping process using pressure and displacement sensing devices; and the geometric feature data is obtained by acquiring the outline image of the formed component using optical imaging devices and performing edge extraction processing on the outline image. Each functional zone of the multi-zone metal component is defined as a model node, and the cumulative relationships of heat conduction, mechanical traction, and feeding are defined as model edges. A spatiotemporal correlation model of the process is constructed based on the thermal data, the mechanical process parameters, and the geometric feature data.

3. The method according to claim 1, characterized in that, In S1, the size deviation tensor is generated, including: The geometric feature data is subjected to subpixel-level coordinate transformation to determine the measured contour point set; the measured contour point set is registered with the nominal feature data to calculate the Euclidean distance vector between the corresponding reference points; the Euclidean distance vector is recombined in multiple dimensions according to the topology of the multi-partition metal component to obtain the size deviation tensor.

4. The method according to claim 1, characterized in that, In S2, the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly is determined, including: The local thermal accumulation field of the multi-zone metal component during continuous stamping is determined based on the cavity temperature distribution in the thermal data; the dynamic yield strength of each functional zone in the multi-zone metal component is calculated by combining the material property parameters and the local thermal accumulation field. The dimensional deviation tensor and the dynamic yield strength of each functional zone are input into a preset elastoplastic inversion model to obtain the plastic deformation of each functional zone; based on the plastic deformation and the dynamic yield strength, the elastic rebound of each functional zone after unloading is determined. The plastic deformation and elastic rebound are concatenated using multidimensional feature vectors to generate the equivalent elastoplastic state information.

5. The method according to claim 1, characterized in that, In S3, the target dimensional deviation value for the next molding cycle is predicted, including: The equivalent elastoplastic state information of each functional partition is input into the process spatiotemporal correlation model. For each node in the process spatiotemporal correlation model, the deviation evolution characteristics of its adjacent nodes in historical impulses are obtained. The deviation evolution characteristics include deviation amplitude and deviation direction. In the spatiotemporal correlation model of the process, feature aggregation processing modified by physical consistency constraints is performed to obtain residual propagation features. The physical consistency constraints include material continuity constraints, volume invariance constraints, and geometric topology constraints. The deviation evolution trend vector for the next forming cycle is calculated based on the residual propagation characteristics; the deviation evolution trend vector is then weighted and fused with the size deviation tensor of the current cycle to determine the target size deviation value.

6. The method according to claim 5, characterized in that, In S3, the total compensation amount is decomposed into pressure compensation component, temperature compensation component, and velocity compensation component, including: The target size deviation value is input into a preset decoupling control model; the target is optimized based on the Jacobian sensitivity matrix through the decoupling control model to determine the first compensation weight corresponding to the mold pressure, the second compensation weight corresponding to the mold temperature, and the third compensation weight corresponding to the feeding speed. The total compensation amount is linearly decomposed based on the first compensation weight, the second compensation weight, and the third compensation weight to obtain the pressure compensation component, the temperature compensation component, and the speed compensation component.

7. The method according to claim 1, characterized in that, In S4, geometric boundary rule verification is performed, including: The pressure compensation component, temperature compensation component, and velocity compensation component are converted into expected geometric correction values ​​for each functional zone. The expected geometric correction values ​​are then algebraically superimposed with the current period size deviation tensor to obtain the compensated expected residual deviation field. Combined with the nominal feature data, the predicted geometric shape is obtained. Extract the geometric features from the predicted geometry, including the closed interval value corresponding to the crimping cylinder region, the symmetry value corresponding to the spring plate region, and the assembly boundary value corresponding to the overall structure; determine whether the closed interval value, the symmetry value, and the assembly boundary value are all within their respective preset rigidity threshold ranges; if the determination result is yes, then the geometric boundary rule verification is confirmed to be passed.

8. The method according to claim 1, characterized in that, In S5, the model parameters of the process spatiotemporal correlation model are updated, including: If the result of the geometric boundary rule verification is not passed, the conservative parameter rollback mechanism is triggered to reset the pressure compensation component, the temperature compensation component and the velocity compensation component to the preset safe initial value. Initiate a manual review signal and adjust the connection weights of each node in the spatiotemporal correlation model of the process based on the correction data provided by the manual reviewer. If the result of the geometric boundary rule verification is passed, the actual geometric feature data of the next forming cycle is obtained, and the residual between the target size deviation value and the actual measured value is calculated; the model parameters of the process spatiotemporal correlation model are updated according to the residual using the gradient descent method.

9. A quality control system for the forming process of a photovoltaic connector metal core, used to implement the method as described in any one of claims 1-8, characterized in that, The system includes: Modeling construction units are used to acquire thermal data, mechanical process parameters, and geometric feature data of the current molding cycle, and to construct a spatiotemporal correlation model of the process. The deviation calculation unit is used to perform differential encoding processing on the geometric feature data and the preset nominal feature data to generate a size deviation tensor; The elastoplastic inversion unit is used to determine the equivalent elastoplastic state information of each functional zone in the multi-zone metal assembly based on the material property parameters, the thermal data, and the dimensional deviation tensor. The decoupled prediction unit is used to input the equivalent elastoplastic state information into the process spatiotemporal correlation model, perform residual propagation calculation, and predict the target size deviation value of the next molding cycle; determine the total compensation amount based on the target size deviation value, and decompose the total compensation amount into pressure compensation component, temperature compensation component and speed compensation component; The verification execution unit is used to convert the pressure compensation component, temperature compensation component, and velocity compensation component into the expected geometric correction amount of each functional zone, combine the current cycle size deviation tensor and nominal feature data to obtain the predicted geometric shape, extract the geometric features, and perform geometric boundary rule verification. After the verification is passed, the pressure compensation component, the temperature compensation component, and the velocity compensation component are written into the corresponding execution unit respectively. The parameter update unit is used to update the model parameters of the process spatiotemporal correlation model based on the result of the geometric boundary rule verification.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the method as described in any one of claims 1-8.

Images