An adaptive control method and system for electromagnetic shield stamping process

Abstract

The present application belongs to the technical field of intelligent control of stamping forming, and particularly relates to a self-adaptive control method and system for a stamping process of an electromagnetic shield, the method comprising: obtaining die displacement, forming force and sheet strain data to form a state vector; updating a posterior probability to form a probability-weighted simulation model set and constructing a game search tree, the maximum layer generating a control action and the minimum layer representing uncertainty; using a Jacobian matrix to screen sensitive dominant directions to non-uniformly expand child nodes; deducing and calculating leaf node evaluation values, performing a minimax search to solve an optimal control strategy and driving equipment to execute. The present application can effectively represent stamping uncertainty, solve the problem of insufficient sensitive boundary identification, and improve the real-time performance of the solution and the robustness of the control strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for stamping forming. More specifically, this invention relates to an adaptive control method and system for the stamping process of electromagnetic shielding covers. Background Technology

[0002] Electromagnetic shielding covers are crucial structural components for suppressing electromagnetic interference and ensuring the stable operation of electronic equipment. With the increasing miniaturization and high reliability requirements of end products, the manufacturing precision and consistency requirements for electromagnetic shielding covers are constantly rising. Stamping, with its advantages of high production efficiency and suitability for mass production, has become the primary method for processing electromagnetic shielding covers. However, electromagnetic shielding covers typically have thin walls, complex contours, and small local feature dimensions. During the forming process, the sheet metal is affected by factors such as batch-to-batch material variations, die friction conditions, ambient temperature, and equipment disturbances, which can easily lead to phenomena such as localized strain concentration, uneven material flow, excessive thinning, or localized accumulation, resulting in quality defects such as cracking, wrinkling, and dimensional deviations. Therefore, how to perceive the die displacement, forming force, and sheet metal strain characteristics in real time during the stamping process, and dynamically adjust the blank holder force and slider motion trajectory parameters based on the forming state, is a key aspect of improving the yield and process stability of electromagnetic shielding covers, and also an important problem that precision stamping control needs to solve.

[0003] Currently, existing technologies address the uncertainty problem in stamping forming by establishing optimized design models and optimizing parameters based on offline simulations and approximate models. Chinese patent application CN105260532A discloses a design method for uncertainties in the variable blank holder force during thin-plate stretching based on sequential approximate optimization. This application utilizes interval descriptions to represent uncertain parameters affecting the quality of stretching forming, establishes a multi-objective optimization design model including the variable blank holder force and uncertain parameters, constructs a training sample point set through finite element analysis, trains a radial basis function neural network to establish an approximate model, and combines iterative optimization with a multi-objective genetic algorithm. Based on these techniques, uncertainties can be considered during the design phase, resulting in a variable blank holder force curve with a certain degree of stability.

[0004] However, the aforementioned technical solutions can improve the forming performance and accuracy of materials to some extent. However, existing technologies mainly rely on offline finite element simulation optimization or deterministic approximation models for feedforward adjustment. The actual stamping process exhibits strong nonlinearity, strong coupling, and transient evolution characteristics. Uncertainties arising from material property fluctuations, transient friction changes, and equipment perturbations are dynamically changing in actual processing, and a single offline model cannot fully characterize these real-time evolving uncertainties. When multiple operating conditions fluctuate simultaneously, offline-generated fixed control curves or conventional feedback adjustment methods often fail to identify the sensitive boundaries of cracking and wrinkling risks in a timely manner, resulting in insufficient stability of the generated control strategy under adverse conditions. Therefore, there is an urgent need for an adaptive control method that can integrate multi-dimensional real-time monitoring data, take into account model prediction uncertainties, and select robust temporal control strategies from multiple possible responses within a finite control period to improve the forming quality and production reliability of the electromagnetic shield stamping process. Summary of the Invention

[0005] To address the technical problems in existing technologies where deterministic models are insufficient to characterize the stamping uncertainties under material and friction fluctuations, resulting in inadequate identification of sensitive boundaries for fracture and wrinkling, and poor real-time performance and robustness of online control strategies, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an adaptive control method for the stamping process of an electromagnetic shield, comprising: S1, acquiring and dimensionlessly processing the die displacement, forming force, and sheet strain characterization data of the stamping process to form a state vector; updating the posterior probability of the confidence of each reduced-order surrogate model according to real-time data to form a probability-weighted simulation model set; constructing a game search tree with the probability-weighted simulation model set, wherein the maxima layer generates control actions, the minima layer represents the model prediction uncertainty, and the Jacobian matrix of the state vector with respect to the control parameters is calculated when the nodes are expanded, and the parameter combination with the highest sensitivity to the risk of cracking and wrinkling is identified as the dominant expansion direction, along the... S2. Non-uniformly generate child nodes in the dominant expansion direction and orthogonal direction; S3. For the generated control parameter sequence, use the probability weighted simulation model set to deduce and calculate the scalar evaluation value of the leaf node. The evaluation value is obtained by mapping the weighted expectation vector and weighted covariance matrix of the state vector on the model set; Perform minimax search. The minimization layer searches for the response boundary that makes the evaluation value the worst in the model set. The maximization layer optimizes the risk avoidance utility function to maximize the forming quality under the worst boundary and solves the optimal control strategy; S4. Generate the time-series control parameter sequence of blank holder force and slider motion trajectory parameters according to the solved optimal control strategy, and output it to the stamping equipment for execution.

[0007] This invention effectively characterizes the stamping uncertainty caused by material and friction fluctuations by updating and forming a probabilistically weighted simulation model set, and solves the problem of insufficient sensitive boundary identification by optimizing the control strategy under the worst boundary through minimax search; by identifying the dominant direction and expanding non-uniform nodes to reduce the number of nodes, the real-time performance of the solution and the robustness of the control strategy are improved.

[0008] Preferably, constructing the game search tree using the probability-weighted simulation model set includes: using the state vector as the root node of the game search tree, and alternately generating maximal and minimum level nodes downwards; setting a maximum depth threshold for the game search tree, and terminating node expansion when the search depth reaches the maximum depth threshold; extracting the complete control parameter sequence and state deduction path from the root node to the current leaf node, and storing the state deduction path in a cache.

[0009] This invention assesses the risk of hysteresis forming caused by material rheology in advance by setting a maximum depth threshold for the search tree, and limits the size of the search tree to meet the real-time requirements of online solving.

[0010] Preferably, the minimum layer searches for the worst response boundary of the model set, which includes: inputting the control action generated by the maximum layer into each reduced-order surrogate model in the model set for extrapolation calculation; extracting the maximum thinning rate and maximum thickening rate of the sheet metal output by each reduced-order surrogate model to calculate a comprehensive risk assessment index; adjusting the confidence weights to maximize the weighted expectation of the comprehensive risk assessment index when the posterior probability confidence interval constraints of each model are satisfied and the sum of the confidence weights is 1; and using the state vector distribution output by the model set after weight adjustment as the worst response boundary of the current node of the minimum layer.

[0011] This invention constructs a conservative worst-case response state by adjusting weights under interval constraints to maximize the risk-weighted expectation, thereby improving the system's decision-making safety when model predictions deviate.

[0012] Preferably, the calculation of the Jacobian matrix of the state vector with respect to the control parameters includes: applying a step disturbance to each control parameter variable, using the blank holder force and the slider motion trajectory parameters as control parameter variables; calculating the disturbance state vector output after applying the step disturbance using the probability-weighted simulation model set; subtracting the disturbance state vector from the state vector to obtain the state vector deviation value; dividing the state vector deviation value by the corresponding step disturbance to obtain the partial derivatives of each state variable with respect to each control parameter, and constructing the Jacobian matrix from all partial derivatives.

[0013] Preferably, the non-uniform generation of child nodes along the dominant expansion direction and the orthogonal direction includes: normalizing the dimensions of the Jacobian matrix, extracting the control parameter with the greatest influence on the state vector as the dominant expansion direction; extracting multiple continuous sampling points at equal intervals within the feasible parameter interval of the dominant expansion direction; using the remaining control parameters as orthogonal directions, extracting upper limit, lower limit, and center value sampling points within the feasible parameter interval of each orthogonal direction; and orthogonally combining the continuous sampling points of the dominant expansion direction with the sampling points of the orthogonal directions to generate a candidate set of control parameter combinations.

[0014] This invention significantly reduces the number of candidate nodes and lowers the online computational burden of the system by extracting the dominant expansion direction and performing coarse-grained dimensionality reduction sampling in the orthogonal direction, while ensuring the search accuracy of sensitive directions.

[0015] Preferably, the calculation of the leaf node scalar evaluation value includes: obtaining the multidimensional predicted state vectors of the leaf nodes under each reduced-order surrogate model in the model set; multiplying each multidimensional predicted state vector with its corresponding confidence level and summing the results to obtain a weighted expected vector of the state vectors; calculating the deviation vector between each multidimensional predicted state vector and the weighted expected vector; multiplying the product of the deviation vector and its own transpose matrix with the corresponding confidence level and summing the results to obtain a weighted covariance matrix; mapping the weighted expected vector to a forming quality expected scalar through the weight vector; calculating the trace of the weighted covariance matrix and mapping it to a forming quality fluctuation scalar; and calculating the scalar evaluation value based on both.

[0016] Preferably, the method of maximizing the risk aversion utility function to solve for the optimal control strategy includes: using a preset forming limit strain threshold and a device forming force limit threshold to constrain and verify the predicted state vectors derived from each sequence; pruning and removing the corresponding nodes when the predicted sheet strain value exceeds the forming limit strain threshold or the predicted forming force value exceeds the forming force limit threshold; constructing a risk aversion utility function in the retained sequences with the goal of maximizing the expected forming quality scalar and minimizing the forming quality fluctuation scalar; selecting the sequence that maximizes the value of the utility function as the optimal control strategy and converting it into a control signal format.

[0017] Preferably, the generation of the timing control parameter sequence for the blank holder force and slider motion trajectory parameters includes: parsing the discrete control action variables in the optimal control strategy, and using a cubic spline interpolation algorithm to smooth them into a continuous blank holder force reference curve and slider motion trajectory reference curve; calculating the slider running speed curve and acceleration rate of change based on the slider motion trajectory reference curve; and merging and arranging the blank holder force reference curve, slider motion trajectory reference curve, and slider running speed curve to form a structured timing control parameter sequence containing timestamp, blank holder force, slider displacement, and slider speed parameters.

[0018] This invention uses a cubic spline interpolation algorithm to smooth discrete variables into continuous running curves and introduces speed planning to prevent impact vibration, effectively preventing sudden changes in sheet material strain and improving processing quality.

[0019] Preferably, the acquisition of die displacement, forming force, and sheet metal strain characterization data during the stamping process includes: acquiring die displacement data through a displacement sensor, acquiring forming force data through a force sensor, and acquiring surface strain data through a visual image correlation system; based on the die displacement data, forming force data, surface strain data, and control parameters from the previous cycle, inverting and estimating the sheet metal strain distribution using a pre-trained reduced-order surrogate model; and using the acquired surface strain data or the inverted and estimated sheet metal strain distribution as sheet metal strain characterization data.

[0020] Secondly, the present invention provides an adaptive control system for the stamping process of an electromagnetic shielding cover, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned adaptive control method for the stamping process of an electromagnetic shielding cover is implemented.

[0021] By adopting the above technical solution, the adaptive control method for the stamping process of the electromagnetic shielding cover is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention acquires data on mold displacement, forming force, and sheet metal strain, dynamically updates the posterior confidence of each reduced-order surrogate model, and constructs a probabilistically weighted simulation model set to characterize forming uncertainties caused by material fluctuations, friction variations, and equipment disturbances. A game-theoretic search tree is constructed using a minimax search mechanism. The minimax layer evaluates the worst-case response boundary under model uncertainty, while the maxima layer selects robust control strategies based on a risk-avoidance utility function, thereby improving forming stability under adverse conditions.

[0024] Furthermore, during the node expansion stage, the Jacobian matrix is ​​used to identify the dominant parameter directions that significantly impact the risk of cracking and wrinkling. Sub-nodes are then generated non-uniformly along the dominant and orthogonal directions to reduce unnecessary traversal calculations and improve online solution efficiency. Finally, the timing control commands for the blank holder force and slider motion trajectory parameters are output, which helps reduce the risk of stamping defects and improve the forming quality of the electromagnetic shielding cover. Attached Figure Description

[0025] Figure 1 This is a flowchart of an adaptive control method for the stamping process of an electromagnetic shielding cover according to the present invention; Figure 2 This is a schematic diagram of the curve showing the change of the maximum principal strain characterization value of the sheet metal over time in this invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] This invention discloses an adaptive control method for the stamping process of an electromagnetic shielding cover, referring to... Figure 1 This includes steps S1-S3: S1. Obtain the stamping state and generate child nodes along the dominant direction.

[0028] In an optional embodiment, die displacement data, forming force data, and sheet strain characterization data of the stamping process are acquired. The acquired die displacement data, forming force data, and sheet strain characterization data are then processed to be dimensionless to form a state vector. The posterior probability of the confidence of each reduced-order surrogate model in the model library is updated according to the real-time data to form a probability-weighted simulation model set. A game search tree is constructed using the probability-weighted simulation model set. The maxima layer generates control actions, and the minima layer represents the uncertainty of model prediction. When expanding nodes, the Jacobian matrix of the state vector with respect to the control parameters is calculated. Then, the parameter combination with the highest sensitivity to the risk of breakage and wrinkling is identified as the dominant expansion direction. Finally, child nodes are generated non-uniformly along the dominant direction and the orthogonal direction.

[0029] Real-time displacement data of the stamping die is acquired using a wire displacement sensor, forming force data is acquired in real-time using a piezoelectric force sensor, and surface strain data of the die during the incomplete closing stage or the observable area of ​​the sheet metal is acquired using a binocular vision digital image correlation system. To address the issue of some areas being difficult to observe directly after the die closes, the overall strain distribution of the sheet metal is estimated by inversion based on die displacement data, forming force data, sheet metal surface strain data of the observable area, and the blank holder force parameters and slider motion trajectory parameters actually issued in the previous control cycle, combined with a pre-trained reduced-order surrogate model. This yields the sheet metal strain distribution estimation data. The sheet metal surface strain data of the observable area or the sheet metal strain distribution estimation data are used as the sheet metal strain characterization data.

[0030] The reduced-order surrogate model is a rapid prediction model pre-established based on historical finite element simulation data or actual stamping production data. It approximates the mapping relationship between blank holder force parameters, slider motion trajectory parameters, and sheet metal forming state. When establishing the reduced-order surrogate model, sheet metal strain data, thinning rate data, thickening rate data, and forming force data are first extracted from the complete finite element simulation results. Then, dimensionality reduction fitting is performed using intrinsic orthogonal decomposition, response surface regression, radial basis function, or neural network methods to obtain a prediction model with lower computational complexity. During online control, the current state vector and control actions are input into the reduced-order surrogate model, which quickly outputs the corresponding estimated values ​​of sheet metal strain distribution, maximum thinning rate, maximum thickening rate, and predicted forming force.

[0031] Subsequently, the minimum-maximum normalization algorithm is used to perform dimensionless processing on the die displacement data, forming force data, and sheet metal strain characterization data, mapping each type of data to the real number interval of 0 to 1, and splicing them in a preset order to form a one-dimensional state vector at the current moment. The current state vector is compared with the predicted state vector output by each reduced-order surrogate model based on the actual control input of the previous control cycle, and then the residual vector between the current state vector and the predicted state vector is calculated. Then, the residual distance is obtained based on the Euclidean distance or Mahalanobis distance of the residual vector. A Gaussian likelihood function is constructed based on the residual distance, so that the reduced-order surrogate model with a smaller residual distance obtains a higher likelihood value. Then, a recursive Bayesian weighted algorithm is used to multiply the model weights of the previous control cycle with the current likelihood value and perform normalization processing to obtain the posterior probability weights of each reduced-order surrogate model representing the actual stamping process behavior. In the initial control cycle, the initial weights of each reduced-order surrogate model are set to equal weights, or the initial weights are preset according to the historical verification error.

[0032] A multi-branch tree structure is constructed in system memory, with the current one-dimensional state vector as the root node of the game search tree. Maximal and minima are sequentially set according to the discrete control cycle. The maximal layer generates candidate control actions, including adjustments to the pressure force and slider trajectory parameters. The slider trajectory parameters include at least one of the following: slider displacement setpoint parameters, slider velocity setpoint parameters, acceleration constraint parameters, or piecewise motion curve control point parameters. The minima characterizes the uncertainty of the model prediction, selecting all reduced-order surrogate models from the probability-weighted simulation model set as the uncertainty response branch.

[0033] During node expansion, the finite difference method is used to calculate the Jacobian matrix of the predicted state vector derived from the current state vector through control actions, relative to each control parameter. This allows us to obtain the sensitivity of each control parameter to subsequent changes in the forming state. The Jacobian matrix is ​​then subjected to linear algebraic decomposition, and the right singular vector corresponding to the maximum singular value is extracted as the dominant expansion direction most sensitive to fracture and wrinkling risks. The remaining directions are designated as orthogonal expansion directions. Subsequently, denser action sampling points are set along the dominant expansion direction, and sparser action sampling points are set along the orthogonal expansion direction. This non-uniformly generates candidate child nodes for the maxima layer, thereby reducing the number of search nodes and improving the computational efficiency of online solution.

[0034] In an optional embodiment, a game search tree is constructed using a probabilistically weighted simulation model set. The maxima layer generates control actions, and the minima layer represents the uncertainty of model prediction. Specifically, the current state vector is used as the root node of the game search tree. Starting from the root node, maxima layer nodes representing control actions and minima layer nodes representing model responses are generated alternately downwards. The maximum depth threshold of the game search tree is set to 6 layers. When the search depth reaches the maximum depth threshold, the node expansion process of the current branch is terminated. The complete control parameter sequence and state deduction path from the root node to the current leaf node of the game search tree are extracted, and the state deduction path is stored in the device cache.

[0035] In the actual stamping process control of the electromagnetic shielding cover, the game-theoretic search tree is constructed starting from the dimensionless state vector obtained at the current moment as the root node. The inference mechanism is initiated, with the first layer being the maxima layer, where the control algorithm generates the currently selectable blank holder force control action and slider motion trajectory parameter control action. The second layer is the minima layer, where each control action is input into a probabilistic weighted simulation model set, calculating the worst-case response of the probabilistic weighted simulation model set under various uncertainties. This process proceeds alternately downwards until a set maximum depth threshold is reached. The maximum depth threshold is counted according to the control decision layers, set to 6 control decision layers; a corresponding model response layer is set after each control decision layer. When each control cycle is 0.05s, the prediction time domain is 0.3s.

[0036] When the search algorithm determines that the current branch has reached the set maximum depth threshold, it triggers a truncation condition and stops expanding the current subtree. At this point, a complete deduction path containing multiple continuous control parameters and corresponding state evolutions is formed from the root node of the game search tree to the current leaf node. The generated complete control parameter sequence data and the data matrix corresponding to the state deduction path are encapsulated and then stored in the PLC or in the SRAM device cache of the host industrial control computer, occupying a set cache block, for example, allocating 2MB of space for cyclic overwriting. The cached data is used for subsequent optimal path comparison evaluation and backtracking evaluation processes. By setting a maximum depth threshold, the controller can assess in advance the risk of wrinkling or cracking caused by the rheology of the metal sheet, and limit the size of the game search tree, so that the minimax search calculation process meets the solution requirements within the current control cycle.

[0037] In an optional embodiment, the Jacobian matrix of the state vector with respect to the control parameters is calculated. Specifically, using the blank holder force parameter and the slider motion trajectory parameter as control parameter variables, a preset step disturbance is applied to each control parameter variable; using the aforementioned probabilistic weighted simulation model set, the disturbance state vector output after applying the step disturbance is calculated; the disturbance state vector is subtracted from the base state vector when no step disturbance is applied to obtain the deviation value of the state vector; the deviation value of the state vector is divided by the corresponding step disturbance to obtain the partial derivative of each state variable with respect to each control parameter, and all partial derivatives constitute the Jacobian matrix.

[0038] When performing local linearization analysis on complex nonlinear stamping processes during node expansion, the blank holder force parameter and slide trajectory parameter in the stamping control system are selected to form the control vector. For example, the control vector includes the blank holder force parameter, the slide speed setpoint, and the slide displacement trajectory control point. Assume the initial rated control vector for the current cycle is [200kN, 50mm / s, 30mm], where 200kN is the blank holder force setpoint, 50mm / s is the slide speed setpoint, and 30mm is the slide displacement trajectory control point. The corresponding multidimensional state vector is calculated using a probabilistic weighted simulation model set. This multidimensional state vector includes the die displacement parameter, forming force parameter, key node strain parameter, and maximum thinning rate parameter or maximum thickening rate parameter. The finite difference method is used to perturb each of the aforementioned control parameters, setting the perturbation step size to 0.1% of their initial rated values. When the initial rated value of a control parameter is 0 or it is not suitable for proportional perturbation, 0.1% of the allowable variation range of the control parameter is used as the perturbation step size. For example, a positive step disturbance is applied to the blank holder force, adjusting the blank holder force to 200.2 kN, while keeping the other control parameters unchanged. The new control vector is then input into the probabilistic weighted simulation model set for calculation to obtain the new disturbance state vector after being disturbed by the blank holder force.

[0039] The response difference of the state vector before and after the application of a small perturbation is calculated by subtracting the perturbed state vector from the aforementioned multidimensional state vector term by term to obtain a multidimensional deviation vector. The difference in each dimension of the multidimensional deviation vector is divided by the actual applied step perturbation amount, i.e., 0.2 kN, to obtain the local partial derivatives of each state variable, such as the die displacement parameter, forming force parameter, and strain parameter, with respect to the blank holder force. The perturbation and difference operations are repeated to sequentially obtain the partial derivative vectors corresponding to the slider speed setpoint step size with a logarithmic value of 0.05 mm / s and the slider displacement trajectory control point step size with a logarithmic value of 0.03 mm. The obtained partial derivative vectors are arranged side-by-side to form a multidimensional Jacobian matrix. This multidimensional Jacobian matrix represents the direction and magnitude of the influence of a small change in the control action on the sheet metal thinning rate response, thickening rate response, strain distribution response, and forming force response in the subsequent predicted forming state under the current stamping transient conditions.

[0040] In an optional embodiment, the parameter combination with the highest sensitivity to fracture risk and wrinkling risk is identified as the dominant expansion direction, and child nodes are generated non-uniformly along the dominant direction and orthogonal directions. Specifically, the constructed multidimensional Jacobian matrix is ​​normalized; based on the normalized Jacobian matrix, the control parameter with the largest influence weight on the state vector is extracted as the dominant expansion direction; five sampling points are extracted at equal intervals within the parameter feasible interval of the dominant expansion direction; other control parameters besides the dominant expansion direction are taken as orthogonal directions, and three sampling points (lower limit, upper limit, and center value) are extracted within the parameter feasible interval of each orthogonal direction; the sampling points of the dominant expansion direction and the sampling points of the orthogonal directions are orthogonally combined to generate a candidate set of control parameter combinations, and the set of control parameter combinations is used as the expanded child nodes.

[0041] To reduce computational load, a non-uniform discretization strategy for the control parameter space is implemented based on a pre-calculated multidimensional Jacobian matrix. Since the dimensions of the control parameters corresponding to each column in the Jacobian matrix vary (e.g., kN, mm / s, mm), the dimensional Jacobian matrix is ​​transformed into a dimensionless normalized Jacobian matrix by multiplying each parameter by its maximum allowable variation range (e.g., blank holder force variation range of 40kN, slider speed variation range of 10mm / s, slider displacement trajectory control point variation range of 5mm). The L2 norm of each column vector in the normalized matrix is ​​then calculated to represent the comprehensive influence weight of each control parameter on the sheet metal forming state. For example, when the L2 norm corresponding to the blank holder force is the largest, it indicates that the blank holder force is most sensitive and has a significant impact on changes in breakage or wrinkling risk; in this case, the blank holder force is determined to be the dominant expansion direction of the current node. For the determined dominant expansion direction, within the current feasible adjustment range, such as the range of 180kN to 220kN, a fine-grained discretization scheme is adopted to extract five consecutive sampling points at equal intervals of 180kN, 190kN, 200kN, 210kN and 220kN.

[0042] For the remaining control parameters not selected as the dominant expansion direction, such as the slider speed setpoint and slider displacement trajectory control points, these control parameters are set as orthogonal directions, and coarse-grained spatial dimensionality reduction sampling is performed. Within the feasible range of each control parameter, for example, if the slider speed is between 40 mm / s and 60 mm / s, and the slider displacement trajectory control points are between 27.5 mm and 32.5 mm, only three sampling points are extracted: the lower limit, the center value, and the upper limit. That is, sampling points of 40 mm / s, 50 mm / s, and 60 mm / s for the slider speed are extracted, and sampling points of 27.5 mm, 30 mm, and 32.5 mm for the slider displacement trajectory control points are extracted. The five sampling points of the dominant direction are orthogonally combined with the three sampling points of each of the two orthogonal directions by performing a Cartesian product operation, generating a total of 45 representative control action combinations. The 45 sets of control action combinations are used as maximal sub-nodes of the game search tree for expansion. Compared with uniformly discrete nodes in each dimension, such as collecting all 5 points for a total of 125 nodes, this strategy reduces the number of search nodes while taking into account the search accuracy in the control-sensitive direction. This helps to improve node search efficiency and reduce the online computing burden of the system.

[0043] S2. Calculate the scalar evaluation value and optimize the solution for the optimal control strategy.

[0044] In an optional embodiment, the generated control parameter sequence is extrapolated using a probabilistic weighted simulation model set, and leaf node scalar evaluation values ​​are calculated. The leaf node scalar evaluation values ​​are calculated based on the weighted expectation vector and weighted covariance matrix of the state vector on the probabilistic weighted simulation model set. A minimax search is performed, with the minimization layer searching for the response boundary of the probabilistic weighted simulation model set that results in the worst scalar evaluation value, and the maximization layer optimizing the risk aversion utility function to maximize the forming quality under the worst boundary and solving for the optimal control strategy.

[0045] For a complete sequence of control parameters from the root node to any leaf node in the game search tree, forward state deduction is performed using various reduced-order surrogate models from the probabilistic weighted simulation model set, resulting in multiple predicted state vectors at the end of the prediction time domain. Based on the posterior probability weights of each reduced-order surrogate model, the multiple predicted state vectors are weighted and summed to obtain a weighted expectation vector; simultaneously, a weighted covariance matrix is ​​calculated based on the deviation of each predicted state vector from the weighted expectation vector.

[0046] A weighted covariance matrix is ​​used as the covariance matrix for the Mahalanobis distance. After superimposing a preset small regularization coefficient on the main diagonal, the Mahalanobis distance between the weighted expectation vector and the preset ideal forming state vector is calculated, and the Mahalanobis distance is used as the basic error penalty term. The trace of the weighted covariance matrix is ​​calculated and used as the model prediction uncertainty penalty term. The basic error penalty term and the model prediction uncertainty penalty term are weighted and added according to a preset ratio to obtain the leaf node risk scalar, where a larger leaf node risk scalar indicates a higher forming defect risk.

[0047] Based on the risk scalar of each leaf node, a minimax search process with αβ pruning is performed from the leaf node to the root node. In the minimax nodes, under the condition that the weights of each reduced-order surrogate model satisfy the confidence interval constraint and the sum of the weights is 1, the model weight combination that maximizes the risk scalar of the leaf node is found, and the maximum risk scalar is passed upwards as the worst response result. In the maxima nodes, the worst risk scalar is input into the risk aversion utility function, such that the higher the risk, the lower the utility, and the control action branch with the largest utility function value is selected. The search process continues to backtrack to the root node, and the current step action corresponding to the maximum utility value at the root node is taken as the optimal control strategy for the current control cycle.

[0048] In an optional embodiment, the minimum layer searches for the worst response boundary of the scalar evaluation value on the probabilistic weighted simulation model set. Specifically, the control actions generated by the maximum layer are input into each reduced-order surrogate model in the probabilistic weighted simulation model set for inference calculation; the maximum thinning rate parameter and the maximum thickening rate parameter of the sheet metal output by each reduced-order surrogate model are extracted, and a comprehensive risk assessment index is calculated based on the maximum thinning rate parameter and the maximum thickening rate parameter; under the premise of satisfying the posterior probability confidence interval constraint of each reduced-order surrogate model and the sum of the confidence weights of each reduced-order surrogate model being 1, the confidence weights of each reduced-order surrogate model are adjusted to maximize the weighted expectation of the comprehensive risk assessment index; the state vector distribution output by the probabilistic weighted simulation model set after adjusting the confidence weights is used as the worst response boundary of the minimum layer at the current node and passed to the maximum layer for subsequent evaluation calculation.

[0049] Assume the current probabilistic weighted simulation model set contains three reduced-order surrogate models based on orthogonal configuration, representing standard operating conditions, high friction coefficient operating conditions, and negative thickness tolerance operating conditions, respectively. The current control actions generated by the maxima layer are input into the three reduced-order surrogate models for parallel and rapid extrapolation. For example, parameters such as a 150kN blank holder force, a 45mm / s slider running speed, and a 30mm slider displacement trajectory control point are input into the reduced-order surrogate models. Then, the maximum thinning rate of the sheet metal in the electromagnetic shield corner area or flange area is extracted from the extrapolation results, for example, the maximum thinning rate is output as 15%, 22%, and 18%, respectively. Simultaneously, the maximum thickness increase rate is extracted, for example, the maximum thickness increase rate is output as 8%, 12%, and 10%, respectively. Based on the rule that the comprehensive risk assessment index is the product of the thinning rate weighting coefficient and the maximum thinning rate plus the product of the thickening rate weighting coefficient and the maximum thickening rate, the comprehensive risk assessment index of each reduced-order proxy model is calculated. The thinning rate weighting coefficient is set to 0.7 and corresponds to the risk of breakage, and the thickening rate weighting coefficient is set to 0.3 and corresponds to the risk of wrinkling.

[0050] After calculating the risk indices of each reduced-order surrogate model, within a pre-defined and Bayesian-updated posterior probability confidence interval, linear programming is used to tilt the weights towards the model with the highest output risk index. For example, the weight fluctuation range of the standard operating condition model is limited to 0.4 to 0.6, the weight fluctuation range of the high friction coefficient operating condition model is limited to 0.1 to 0.3, and the weight fluctuation range of the thickness negative tolerance operating condition model is limited to 0.2 to 0.4, while keeping the sum of the weights of the three reduced-order surrogate models at 1. For instance, when the high friction coefficient operating condition model outputs the highest comprehensive risk assessment index, the weight of the high friction coefficient operating condition model is increased from 0.2 to the upper confidence limit of 0.3, while the weight of the standard operating condition model is decreased from 0.5 to 0.4, and the weight of the thickness negative tolerance operating condition model is kept at 0.3. This results in the adjusted weights of the three models being 0.4, 0.3, and 0.3, respectively, with each model weight falling within its respective confidence interval and the sum of the weights still being 1. This maximizes the weighted expected value of the comprehensive risk assessment indicators, thereby forming a conservative worst-case response assessment state.

[0051] After the aforementioned weight reallocation operation, the worst-case expected vector and the corresponding fluctuation covariance matrix are recalculated based on the adjusted weights. These worst-case expected vector and corresponding fluctuation covariance matrix are then used as the distribution of the worst-case response boundary state vector caused by model prediction uncertainty at the current node, and subsequently passed to the maxima layer. The maxima layer then uses the worst-case response boundary as a benchmark for subsequent optimization decisions, thereby improving the stability of stamping parameter selection even when model predictions may exhibit unfavorable deviations.

[0052] In an optional embodiment, a probabilistic weighted simulation model set is used to deduce and calculate the scalar evaluation value of the leaf node. The scalar evaluation value of the leaf node is calculated based on the mapping of the weighted expectation vector and the weighted covariance matrix of the state vector on the probabilistic weighted simulation model set. Specifically, the multidimensional predicted state vector of the leaf node under each reduced-order surrogate model in the probabilistic weighted simulation model set is obtained; the multidimensional predicted state vector under each reduced-order surrogate model is multiplied with the corresponding posterior probability distribution confidence and summed to calculate the weighted expectation vector of the state vector; the deviation vector between the multidimensional predicted state vector and the weighted expectation vector under each reduced-order surrogate model is calculated; the product of the deviation vector and the transpose matrix corresponding to the deviation vector is multiplied with the corresponding confidence and summed to calculate the weighted covariance matrix of the state vector; the weighted expectation vector is mapped to the forming quality expectation scalar through a preset weight vector; the trace of the weighted covariance matrix is ​​calculated and mapped to the forming quality fluctuation scalar; finally, the scalar evaluation value of the leaf node is calculated based on the forming quality expectation scalar and the forming quality fluctuation scalar.

[0053] When the game search tree reaches a leaf node, the high-dimensional state vector is converted into a single, comparable scalar evaluation value to support the minimax decision process. Assuming the current leaf node is under three reduced-order surrogate models within the probabilistic weighted simulation model set, with corresponding posterior confidence levels of 0.5, 0.3, and 0.2, respectively, a 20-dimensional multidimensional predicted state vector containing sheet metal strain parameters and mold stress parameters is predicted. A weighted summation operation is performed, multiplying the three predicted state vectors by their corresponding confidence levels and summing them to obtain the weighted expectation vector of the state vectors. The deviation vector of the predicted state vector from the weighted expectation vector for each reduced-order surrogate model is calculated, and the autocovariance outer product matrix of the deviation vector is calculated. This is then weighted and summed using the confidence levels to obtain a 20×20-dimensional state weighted covariance matrix. This state weighted covariance matrix represents the degree of dispersion of the prediction results under model prediction uncertainty. After obtaining the distribution parameters, a mapping calculation process from the high-dimensional space to the one-dimensional evaluation space is performed. Input a set of weight vectors pre-tuned by process experts, for example, setting the weight of the penalty thinning rate to -100 and the weight of the reward strain uniformity to +50, and perform the inner product calculation of the expectation vector and the weight vector, thereby mapping the multidimensional expectation vector to a forming quality expectation scalar representing the forming quality level, wherein the value direction of the forming quality expectation scalar is set so that the larger the value, the better the forming quality.

[0054] The sum of the main diagonal elements of the weighted covariance matrix is ​​extracted, i.e., the trace of the weighted covariance matrix is ​​obtained. This trace is then multiplied by a variance penalty coefficient, preferably set to 0.8, to calculate the forming quality fluctuation scalar. A leaf node risk scalar function is constructed; for example, the expected forming quality scalar is negatively weighted and added to the forming quality fluctuation scalar according to the risk penalty coefficient. A larger leaf node risk scalar indicates a higher forming quality risk.

[0055] In an optional embodiment, the risk aversion utility function is optimized at the maximum level to maximize the forming quality under the worst-case boundary and solve for the optimal control strategy. Specifically, using preset sheet metal forming limit strain thresholds and equipment forming force limit thresholds, the predicted state vectors obtained by probabilistic weighted simulation model sets for each control parameter sequence are constrained and verified. When the predicted sheet metal strain value in the predicted state vector exceeds the sheet metal forming limit strain threshold, or when the predicted forming force value in the predicted state vector exceeds the equipment forming force limit threshold, the nodes corresponding to the current control parameter sequence are pruned and removed. In the retained control parameter sequences, a risk aversion utility function is constructed with the goal of maximizing the expected forming quality scalar and minimizing the forming quality fluctuation scalar. The control parameter sequence that maximizes the risk aversion utility function value is selected as the optimal control strategy. The obtained optimal control strategy is converted into the control signal format of the stamping equipment actuator and the control signal is sent to the execution unit of the stamping equipment for corresponding control operations.

[0056] After the worst-case response boundary assessment values ​​caused by all possible actions are propagated back from the minimum layer, the maximum layer initiates a reverse backtracking process of constraint verification and optimal decision-making. It reads the safe operating boundary conditions of the current process, for example, setting the principal strain limit forming curve threshold of the sheet metal to 0.25, and simultaneously setting the safe forming force limit threshold of the servo press to 1500kN. It scans all control action sequences and their corresponding extrapolated states in the search tree. If a path is found to have a future extrapolated principal strain exceeding 0.25, indicating a risk of fracture, or if the predicted forming force exceeds 1500kN, indicating a risk of equipment overload, a hard pruning and removal operation is immediately triggered for the problematic node and its preceding related control sequences. The risk avoidance utility value corresponding to the problematic node is set to negative infinity, and the problematic node is removed from the candidate pool.

[0057] For the remaining valid control parameter sequence after filtering, the maximization layer constructs a risk aversion utility function using the pre-calculated expected scalar and volatility scalar. This is achieved by subtracting the product of the risk aversion coefficient and volatility scalar from the expected scalar, where the risk aversion coefficient is set to 1.5 to reinforce the requirement of minimizing quality fluctuations. The utility function values ​​of all feasible paths are compared, and the action sequence that maximizes the utility function is selected from the feasible paths. The first step control vector of the action sequence at the current moment is extracted as the optimal control strategy. For example, the determined optimal parameters are a blank holder force of 190kN, a slider speed of 45mm / s, and a slider displacement trajectory control point of 30mm. Subsequently, the system packages these parameters into network packets and sends them via Ethernet. The underlying motion control card performs the command and analog-to-inductive conversion, thereby controlling the stamping execution process of the next cycle.

[0058] S3. Generate timing control parameters and drive the stamping equipment to execute.

[0059] In an optional embodiment, a sequence of timing control parameters for blank holder force and slider motion trajectory is generated based on the optimal control strategy obtained by solving, and the generated sequence of timing control parameters is output to the stamping equipment for execution.

[0060] The discrete control action variables extracted from the root node of the game search tree are analyzed. A cubic spline interpolation algorithm is used to smooth the discrete step length action variables according to a 1ms hardware control cycle, thus converting them into continuous blank holder force reference curves and slider motion trajectory reference curves. Based on the slider motion trajectory reference curves, the slider running speed curve and acceleration rate of change are calculated, and a position incremental proportional-integral-derivative (PID) controller algorithm is used to generate a slider running speed control curve to prevent impact vibration. The values ​​of the three curves are merged and arranged to form a structured timing control parameter sequence containing synchronization timestamps, blank holder force parameters, slider displacement trajectory parameters, and slider running speed parameters. The generated timing control parameter sequence is aligned to 16-bit bytes and packaged into network packets according to the industrial standard data format supported by the stamping field programmable logic controller, and an industrial Ethernet communication protocol stack based on Ethernet control technology is invoked. The packaged network packets are sent to the underlying motion control card of the stamping equipment via the Ethernet socket interface of the control host. The underlying motion control card converts the blank holder force parameters into analog commands or bus control commands that the hydraulic proportional valve can recognize. At the same time, it converts the slider displacement trajectory parameters and slider running speed parameters into position commands, speed commands or pulse control commands that the servo driver can recognize, thereby driving the servo hydraulic cushion and main servo motor of the electromagnetic shielded stamping machine to execute the stamping task of the current cycle according to the set timing.

[0061] Reference Figure 2 This reflects the trend of the maximum principal strain value of the sheet metal gradually increasing and accompanied by minor local fluctuations as the stamping time progresses from the start to approximately 0.3 seconds. In the initial stage of the stamping process, the principal strain value remains at a low level and changes relatively slowly; as the stamping process progresses, the increase in the principal strain value gradually increases, reaching a relatively high level of nearly 0.20 at the end of the stamping cycle. This trend indicates that by using a continuous and smooth control curve to drive the stamping equipment to perform the processing task, the principal strain state of the sheet metal during the forming process can be effectively monitored and controlled, preventing abrupt changes in the principal strain value or exceeding the safe forming limit threshold. This demonstrates that the generated time-series control parameter sequence has a good and stable effect in controlling forming quality.

[0062] Verification operations were conducted using an electromagnetic shielding cover stamping production line equipped with a servo press and a multi-point CNC hydraulic cushion. The sheet material used was 0.5mm thick stainless steel, and the control cycle was set to 0.05s. Three groups of experiments were set up for comparative analysis: the basic control group used a traditional full-dimensional uniform grid search predictive control algorithm; the ablation group introduced a minimax game search mechanism and a minimum layer adversarial evaluation mechanism on top of the basic control group; and the comprehensive experimental group used the non-uniform node generation strategy and risk-avoidance game search control strategy proposed in this application. Each of the three groups continuously stamped 1000 electromagnetic shielding covers, and a 5% negative tolerance in material thickness and high friction coefficient fluctuations were introduced during the production process to simulate unfavorable working conditions. During the experiments, the single-step solution time, maximum thinning rate, and finished product pass rate of each group were recorded.

[0063] Experimental results show that the average single-step solution time of the basic control group is 125ms, exceeding the control cycle requirement of 0.05s, which easily leads to lag in control commands. The maximum thinning rate of the stamped parts in the basic control group is 24.5%, and the finished product qualification rate is 82.3%. After adding the adversarial evaluation mechanism, the ablation group improved the adaptability of the control strategy to model prediction uncertainty and working condition disturbances, reducing the maximum thinning rate of the finished product to 19.2% and increasing the finished product qualification rate to 93.5%. However, since the ablation group still uses the full-dimensional grid search method, the average single-step solution time still reaches 112ms. Due to the introduction of the αβ pruning algorithm in the minimax backtracking process, the evaluation process of some invalid branches is reduced, so the average single-step solution time is slightly lower than that of the basic control group, but it still exceeds the control cycle requirement of 0.05s. After adopting a non-uniform node discretization strategy, the average time for single-step solution in the comprehensive experimental group was reduced to 42ms, which met the requirement of a 0.05s control cycle under the set experimental conditions. At the same time, the maximum thinning rate of the stamped parts was reduced to 17.8%, and the finished product qualification rate was increased to 99.1%.

[0064] Therefore, using the Jacobian matrix to extract the dominant expansion direction and performing non-uniform node sampling helps reduce the node size of the game search tree and the online computational burden, allowing the single-step solution time to be controlled within the set control cycle, thereby reducing the adverse effects of control command lag on forming quality. The mechanism combining minimax game search with worst-case boundary evaluation helps to assess potential cracking and wrinkling risks in advance under adverse working conditions such as material thickness fluctuations or changes in friction state, and generates more stable blank holder force parameter control strategies and slider motion trajectory parameter control strategies, thereby improving the forming stability and finished product yield of the electromagnetic shield stamping process.

[0065] This invention also discloses an adaptive control system for the stamping process of an electromagnetic shielding cover, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an adaptive control method for the stamping process of an electromagnetic shielding cover according to the present invention is implemented.

[0066] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0067] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An adaptive control method for the stamping process of an electromagnetic shielding cover, characterized in that, include: S1. Obtain the die displacement, forming force and sheet metal strain characterization data of the stamping process and perform dimensionless processing to form a state vector; The posterior probability of the confidence of each reduced-order surrogate model is updated according to real-time data to form a probability-weighted simulation model set; a game search tree is constructed with the probability-weighted simulation model set, the maximal layer generates control actions, the minimum layer represents the uncertainty of model prediction, the Jacobian matrix of the state vector with respect to the control parameters is calculated when the node is expanded, the parameter combination with the highest sensitivity to the risk of cracking and wrinkling is identified as the dominant expansion direction, and child nodes are generated non-uniformly along the dominant expansion direction and the orthogonal direction; S2. For the generated control parameter sequence, the leaf node scalar evaluation value is derived and calculated using the probability-weighted simulation model set. The evaluation value is obtained by mapping the weighted expectation vector and weighted covariance matrix of the state vector on the model set. A minimax search is performed. The minimization layer searches the model set for the response boundary that makes the evaluation value the worst. The maximization layer optimizes the risk aversion utility function to maximize the forming quality under the worst boundary and solves the optimal control strategy. S3. Generate a sequence of timing control parameters for blank holder force and slider motion trajectory parameters based on the solved optimal control strategy, and output it to the stamping equipment for execution.

2. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The step of constructing a game search tree using the probability-weighted simulation model set includes: using the state vector as the root node of the game search tree, and alternately generating maximal and minimum level nodes downwards; setting a maximum depth threshold for the game search tree, and terminating node expansion when the search depth reaches the maximum depth threshold; extracting the complete control parameter sequence and state deduction path from the root node to the current leaf node, and storing the state deduction path in a cache.

3. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The minimum layer searches for the worst response boundary of the model set, including: inputting the control actions generated by the maximum layer into each reduced-order surrogate model in the model set for extrapolation calculation; extracting the maximum thinning rate and maximum thickening rate of the sheet metal output by each reduced-order surrogate model to calculate a comprehensive risk assessment index; adjusting the confidence weights to maximize the weighted expectation of the comprehensive risk assessment index when the posterior probability confidence interval constraints of each model are satisfied and the sum of the confidence weights is 1; and using the state vector distribution output by the model set after weight adjustment as the worst response boundary of the current node of the minimum layer.

4. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The calculation of the Jacobian matrix of the state vector with respect to the control parameters includes: applying a step disturbance to each control parameter variable, using the blank holder force and the slider motion trajectory parameters as control parameter variables; calculating the disturbance state vector output after applying the step disturbance using the probability weighted simulation model set; subtracting the disturbance state vector from the state vector to obtain the state vector deviation value; dividing the state vector deviation value by the corresponding step disturbance to obtain the partial derivatives of each state variable with respect to each control parameter; and constructing the Jacobian matrix from all the partial derivatives.

5. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 4, characterized in that, The non-uniform generation of child nodes along the dominant expansion direction and orthogonal directions includes: normalizing the dimensions of the Jacobian matrix, extracting the control parameter with the greatest influence on the state vector as the dominant expansion direction; extracting multiple continuous sampling points at equal intervals within the parameter feasible interval of the dominant expansion direction; using the remaining control parameters as orthogonal directions, extracting upper limit, lower limit, and center value sampling points within the parameter feasible interval of each orthogonal direction; and orthogonally combining the continuous sampling points of the dominant expansion direction with the sampling points of the orthogonal directions to generate a candidate set of control parameter combinations.

6. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The calculation of the leaf node scalar evaluation value includes: obtaining the multidimensional predicted state vectors of the leaf nodes under each reduced-order surrogate model in the model set; multiplying each multidimensional predicted state vector with its corresponding confidence level and summing the results to obtain a weighted expected vector of the state vectors; calculating the deviation vector between each multidimensional predicted state vector and the weighted expected vector; multiplying the product of the deviation vector with its own transpose matrix and its corresponding confidence level and summing the results to obtain a weighted covariance matrix; mapping the weighted expected vector to a forming quality expected scalar through the weight vector; calculating the trace of the weighted covariance matrix and mapping it to a forming quality fluctuation scalar; and calculating the scalar evaluation value based on both.

7. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 6, characterized in that, The method of optimizing the risk aversion utility function at the maximum level to solve for the optimal control strategy includes: using a preset forming limit strain threshold and a device forming force limit threshold to constrain and verify the predicted state vectors derived from each sequence; pruning and removing the corresponding nodes when the predicted sheet strain value exceeds the forming limit strain threshold or the predicted forming force value exceeds the forming force limit threshold; constructing a risk aversion utility function in the retained sequences with the objective of maximizing the expected forming quality scalar and minimizing the forming quality fluctuation scalar; selecting the sequence that maximizes the value of this utility function as the optimal control strategy and converting it into a control signal format.

8. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The generation of the timing control parameter sequence for the blank holder force and slider motion trajectory parameters includes: parsing the discrete control action variables in the optimal control strategy, calling the cubic spline interpolation algorithm to smooth them into a continuous blank holder force reference curve and slider motion trajectory reference curve; calculating the slider running speed curve and acceleration change rate based on the slider motion trajectory reference curve; and merging and arranging the blank holder force reference curve, slider motion trajectory reference curve, and slider running speed curve to form a structured timing control parameter sequence containing timestamp, blank holder force, slider displacement, and slider speed parameters.

9. The adaptive control method for the stamping process of an electromagnetic shielding cover according to claim 1, characterized in that, The acquisition of die displacement, forming force, and sheet metal strain characterization data during the stamping process includes: acquiring die displacement data through a displacement sensor, acquiring forming force data through a force sensor, and acquiring surface strain data through a visual image correlation system; based on the die displacement data, forming force data, surface strain data, and control parameters from the previous cycle, inverting and estimating the sheet metal strain distribution using a pre-trained reduced-order surrogate model; and using the acquired surface strain data or the inverted and estimated sheet metal strain distribution as sheet metal strain characterization data.

10. An adaptive control system for the stamping process of an electromagnetic shielding cover, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement an adaptive control method for the stamping process of an electromagnetic shielding cover according to any one of claims 1-9.

Images