An automobile sheet metal part stamping forming precision control method and system

Abstract

The present application relates to sheet metal part processing technical field, specifically to a kind of automobile sheet metal part stamping forming precision control method and system, first, online acquisition sheet metal three-dimensional profile, thickness distribution and material performance parameter, input rebound prediction proxy model obtains initial rebound deviation field;Further, with minimum profile deviation as target, generate partition blank holder force dynamic curve and adjustable drawbead height distribution, complete feedforward setting, in stamping process, real-time acquisition flange strain, blank holder ring displacement and mould temperature, rebound trend is dynamically predicted by digital twin model and closed-loop regulation actuator, after demoulding, measured profile deviation and process data are input into process drift early warning model, to predict the precision risk level of future stamping cycle, when threshold is reached, automatic warning is realized, self-awareness, self-compensation, self-warning and self-learning of stamping forming precision are realized, to long-term maintain high-precision forming and actively prevent mould wear and process drift caused batch quality defects.

Description

Technical Field

[0001] This invention relates to the field of sheet metal processing technology, and in particular to a method and system for controlling the stamping precision of automotive sheet metal parts. Background Technology

[0002] Existing methods and systems for controlling the precision of stamping in automotive sheet metal parts generally rely on a combination of offline finite element simulation and die surface compensation. This approach involves pre-setting springback compensation during the die design phase and using segmented constant blank holder forces or fixed draw beads during production to constrain material flow, thereby reducing springback and improving part dimensional compliance to some extent. Some advanced solutions also incorporate online 3D measurement and statistical process control, enabling post-demolding detection and trend analysis of forming deviations. This allows for the identification and alarm of systematic deviations in single batches or continuous production, providing a preliminary capability for monitoring forming quality. Most existing control modes are open-loop or post-processing corrections, which cannot detect changes in the sheet metal state in real time during the stamping process and dynamically compensate for springback. This results in insufficient adaptability to fluctuations in the performance of incoming batches and disturbances in operating conditions. At the same time, existing technologies generally lack proactive early warning mechanisms for long-term degradation processes such as die wear and slow drift of process parameters. They often only respond passively after batch deviations occur. Moreover, once the prediction model is trained offline, it becomes fixed and cannot learn online or automatically track the evolution of equipment status. This makes it difficult to maintain high-precision forming in the long term and also makes it difficult to prevent batch quality defects caused by die failure or process drift. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for controlling the stamping precision of automotive sheet metal parts. Based on the combination of online sensing, surrogate model prediction, digital twin closed-loop control and process drift early warning, it can achieve self-sensing, self-compensation, self-early warning and self-learning of stamping precision, so as to maintain high-precision forming in the long term and actively prevent batch quality defects caused by mold wear and process drift.

[0004] To achieve the above objectives, the present invention provides a method for controlling the stamping precision of automotive sheet metal parts, comprising the following steps: The initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed can be obtained online. The acquired data is input into a pre-trained springback prediction proxy model to predict the initial springback deviation field after the sheet metal is formed. Based on the initial springback deviation field, with the goal of minimizing the surface deviation after forming, a dynamic curve of partitioned blank holder force and an initial height distribution of adjustable draw beads are generated, and the stamping actuator is set accordingly. During continuous stamping, strain data, pressure ring displacement data, and mold temperature data of the sheet metal flange area inside the mold are collected in real time, and input into the digital twin model to dynamically predict the real-time springback trend, and to perform closed-loop adjustment of the actuator. After each stamping and demolding, the measured surface deviation of the formed part is obtained. The time series containing the measured surface deviation and process data is input into the process drift early warning model to predict the forming accuracy risk level of future stamping cycles. When the predicted risk level reaches the preset warning threshold, a warning message is generated.

[0005] The online acquisition of the initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed includes: The initial three-dimensional contour data is obtained by a structured light three-dimensional contour sensor installed on a feeding robot, and the thickness distribution data is obtained by an electromagnetic ultrasonic thickness measuring probe. By scanning the information code on the sheet metal, the material property fluctuation parameters, including the statistical distribution parameters of yield strength, tensile strength, and thickness anisotropy coefficient, are retrieved from the manufacturing execution system.

[0006] The springback prediction proxy model is a deep learning model that integrates convolutional neural networks and long short-term memory networks. Its inputs also include mold geometric feature encoding, preset stamping speed and lubrication condition parameters.

[0007] The generation of the dynamic curve of the partition blank holder force and the initial height distribution of the adjustable draw bead include: A decomposition-based multi-objective evolutionary algorithm is used for optimization. The dynamic curve of the partitioned blank holder force is parameterized using a B-spline curve. The initial height distribution of the adjustable drawing bead is set by the independent extension and retraction of multiple points of the array-type micro-electrically controlled lifting pins on the mold.

[0008] The real-time acquisition of strain data, blank holder displacement data, and mold temperature data in the flange area of ​​the sheet metal inside the mold, and inputting them into the digital twin model to dynamically predict the real-time springback trend, includes: The strain data is acquired by an array of fiber optic strain sensors embedded in the surface of the die and the pressure ring; The displacement data of the pressing ring is collected by a magnetostrictive displacement sensor integrated in the partition pressing cylinder; The digital twin model is based on the finite element reduced-order model and uses the collected strain and displacement data as boundary conditions for simulation, outputting the real-time rebound trend.

[0009] The closed-loop adjustment of the actuator includes: Based on the real-time rebound trend, the blanking force of each zone of the partition blanking force servo control unit is independently adjusted, and the heating or cooling power of the local temperature regulating element array embedded in the punch or die is controlled. The local temperature regulating element array is composed of a mixture of thermoelectric coolers and resistance heating wires, which can regulate the material flow characteristics by changing the local temperature field of the sheet.

[0010] The process drift early warning model is a deep learning model based on temporal convolutional networks or long short-term memory networks. Its inputs include the partition blank holder force curves in the most recent N stamping cycles, the mold temperature change sequence, the stress distribution feature sequence output by the digital twin model, and the corresponding measured surface deviation sequence. Its output, the forming accuracy risk level, is the probability level of the formed part dimensions exceeding tolerance in the next few stamping cycles.

[0011] The warning information includes the remaining number of punches for mold replacement or the compensation value for the height of the drawing bead, and is displayed through a human-machine interface.

[0012] The springback prediction proxy model uses the comparison results between the measured surface deviation and the initial springback deviation field to perform online reinforcement learning updates. This update adopts a near-end strategy optimization algorithm to reduce the prediction deviation as a reward signal. The process drift early warning model is periodically retrained after accumulating a new batch of stamping data.

[0013] One of them is an automotive sheet metal stamping forming precision control system, used to implement the automotive sheet metal stamping forming precision control method, including a sheet metal parameter online sensing module, a springback prediction module, a process parameter generation module, a stamping forming actuator, a real-time simulation and control module, an online detection module, a process drift early warning module, an early warning output module, and a closed-loop learning module. The output of the online sheet metal parameter sensing module is connected to the input of the springback prediction module; the output of the springback prediction module is connected to the input of the process parameter generation module; the output of the process parameter generation module is connected to the control of the stamping forming actuator; the sensor output of the stamping forming actuator is connected to the input of the real-time simulation and control module; the control output of the real-time simulation and control module is connected to the actuator of the stamping forming actuator; the output of the online detection module is connected to the input of the process drift early warning module; the output of the process drift early warning module is connected to the early warning output module; the input of the closed-loop learning module is connected to both the online detection module and the real-time simulation and control module, and its output is connected to both the springback prediction module and the process drift early warning module. The online sensing module for sheet material parameters is used to acquire the initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed. The rebound prediction module is equipped with an online-updable rebound prediction proxy model for predicting the initial rebound deviation field. The process parameter generation module is used to generate a dynamic curve of partitioned blank holder force and an adjustable initial height distribution of draw beads based on the initial springback deviation field. The stamping forming actuator includes a partitioned blank holder ring, an adjustable drawing bead, an electro-hydraulic servo blank holder cylinder, a local temperature regulating element array composed of a thermoelectric cooler and a resistance heating wire, and fiber optic strain sensors, magnetostrictive displacement sensors and temperature sensors distributed in the mold. The real-time simulation and control module runs a digital twin model to receive sensor data and output control commands to adjust the actuator in a closed loop. The online detection module is used to acquire the measured surface data of the molded part and calculate the measured surface deviation. The process drift early warning module has a built-in process drift early warning model, which is used to predict the molding accuracy risk level based on the time series of process data and measured surface deviation. The early warning output module is used to output early warning information including the remaining number of punches for mold replacement or the height compensation value of the drawing bead when the risk level reaches the early warning threshold. The closed-loop learning module is used to feed back the measured surface deviation to the springback prediction module and to feed back the process and deviation data to the process drift early warning module.

[0014] This invention discloses a method and system for controlling the precision of stamping forming of automotive sheet metal parts. It constructs a closed-loop control system covering the entire process from sheet metal loading and forming to demolding inspection. First, it senses the true three-dimensional contour, thickness distribution, and batch material properties of each sheet metal online, and uses a deep learning surrogate model to predict the springback deviation field. Based on this, a multi-objective evolutionary algorithm is used to optimize and generate a dynamic curve of the partitioned blank holder force and the height distribution of the draw bead formed by an array of electrically controlled lifting pins, achieving feedforward setting. During the stamping process, fiber optic strain sensors and magnetostrictive displacement sensors embedded in the mold collect data in real time, driving a finite element reduced-order digital twin model. Millisecond-level simulation and dynamic adjustment of blank holder force in each zone and local thermoelectric hybrid temperature control array actively compensate for springback; after demolding, online measurement is used to obtain the measured surface deviation. On the one hand, the proxy model is updated by reinforcement learning to optimize the near-end strategy to adapt to mold wear. On the other hand, multi-source process data and deviation sequence are input into the process drift early warning model based on time-series deep network to predict the out-of-tolerance risk level of future stamping cycles in advance. When the threshold is exceeded, early warning information containing the remaining stamping cycles or compensation value suggestions is output. Thus, the stamping forming accuracy is self-sensing, self-compensating, self-early warning and self-learning, maintaining high accuracy in the long term and actively preventing batch quality defects. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is a schematic diagram of the overall structure of the method for controlling the stamping precision of automotive sheet metal parts according to the present invention.

[0017] Figure 2 This is a flowchart of the method for controlling the stamping precision of automotive sheet metal parts according to the present invention.

[0018] Figure 3 This is a flowchart of the process drift warning and self-learning closed-loop process of the present invention.

[0019] Figure 4 This is a schematic diagram of the automotive sheet metal stamping precision control system of the present invention.

[0020] In the diagram: 1-Online sensing module for sheet metal parameters, 2-Springback prediction module, 3-Process parameter generation module, 4-Stamping forming actuator, 5-Real-time simulation and control module, 6-Online detection module, 7-Process drift early warning module, 8-Early warning output module, 9-Closed-loop learning module. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] In the description of this invention, it should be understood that "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] Please see Figures 1 to 3 This invention provides a method for controlling the stamping precision of automotive sheet metal parts, comprising the following steps: S1: Online acquisition of initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed; S2: Input the acquired data into the pre-trained springback prediction proxy model to predict the initial springback deviation field after the sheet metal is formed; S3: Based on the initial springback deviation field, with the goal of minimizing the surface deviation after forming, generate the dynamic curve of the partition blank holder force and the initial height distribution of the adjustable draw bead, and set the stamping actuator accordingly. S4: During the continuous stamping process, the strain data, pressure ring displacement data and mold temperature data of the sheet metal flange area inside the mold are collected in real time, and input into the digital twin model to dynamically predict the real-time springback trend, and the actuator is adjusted in a closed loop. S5: After each stamping and demolding, the measured surface deviation of the formed part is obtained, and the time series containing the measured surface deviation and process data is input into the process drift early warning model to predict the forming accuracy risk level of future stamping cycles. S6: When the predicted risk level reaches the preset warning threshold, a warning message is generated.

[0024] In this embodiment, during actual stamping production line operation, when a sheet of material to be formed is conveyed to the press by a feeding robot, the system automatically initiates online sensing of sheet parameters, simultaneously acquiring the sheet's three-dimensional morphology, thickness distribution, and performance fluctuation data of the batch of material. Subsequently, the springback prediction proxy model quickly predicts the possible springback deviation field after forming based on these actual incoming material conditions. Based on this prediction, the process parameter generation module, aiming to minimize surface deviation, calculates the optimal dynamic curve of the zoned blank holder force and the adjustable drawbead height distribution through optimization algorithms, and completes the feedforward setting of the stamping actuator accordingly. During the die-closing stamping process, the sensor array embedded in the die collects the strain, blank holder displacement, and die temperature of the sheet flange area in real time. The digital twin model running on the edge computing device uses this data to perform simulation at millisecond speeds, dynamically predicting the current real-time springback trend, and immediately adjusts the blank holder force and local temperature of each zone through closed-loop control, actively compensating for springback during the forming process.

[0025] After the parts are demolded, an online 3D measurement system located at the end of the production line scans the parts and calculates the measured surface deviation. This deviation, along with process data from the stamping process, is fed into a process drift early warning model. This model predicts the accuracy risk level for multiple future stamping cycles based on time-series characteristics. Once the risk level exceeds a preset threshold, the system issues an early warning message containing specific maintenance suggestions through the human-machine interface, guiding operators to perform timely mold maintenance or process adjustments. This eliminates batch deviations that may be caused by mold wear and process drift at the outset, achieving a fully intelligent closed loop from perception, decision-making, control to early warning self-learning.

[0026] Furthermore, the online acquisition of the initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed includes: The initial three-dimensional contour data is obtained by a structured light three-dimensional contour sensor installed on a feeding robot, and the thickness distribution data is obtained by an electromagnetic ultrasonic thickness measuring probe. By scanning the information code on the sheet metal, the material property fluctuation parameters, including the statistical distribution parameters of yield strength, tensile strength, and thickness anisotropy coefficient, are retrieved from the manufacturing execution system.

[0027] In this embodiment, after the feeding robot grabs the sheet material, during the process of moving towards the mold, the structured light three-dimensional contour sensor installed on the end effector of the robot quickly scans the entire surface of the sheet material in a non-contact manner to reconstruct the initial three-dimensional contour with micron-level precision; at the same time, the electromagnetic ultrasonic thickness probe measures the thickness of the sheet material at different preset positions to obtain the thickness distribution.

[0028] A QR code or RFID tag attached to one corner of the sheet metal is scanned by a reader. Based on the batch information read, the system automatically accesses the manufacturing execution system and retrieves the statistical distribution parameters of the yield strength, tensile strength, and thickness anisotropy coefficient for the corresponding heat number of that batch of material. In this way, the true initial geometry and material properties of each sheet metal are precisely digitized, providing a more accurate input for subsequent personalized process optimization than the traditional method of using only nominal dimensions and material grades. This fundamentally adapts to the interference of incoming material fluctuations on forming accuracy.

[0029] Furthermore, the springback prediction proxy model is a deep learning model that integrates convolutional neural networks and long short-term memory networks, and its inputs also include mold geometric feature encoding, preset stamping speed and lubrication condition parameters.

[0030] In this embodiment, the springback prediction proxy model is trained in advance using massive historical stamping data and corresponding springback test results. It adopts a deep learning architecture that integrates convolutional neural networks and long short-term memory networks. The convolutional layers automatically extract spatial correlation patterns from the three-dimensional contour, thickness distribution, and die geometry of the sheet metal, while the long short-term memory layers process the temporal dependencies of these features as they evolve with the stamping process.

[0031] In addition to the incoming material data, the model's input also includes the geometric feature encoding of the current mold, preset stamping speed, and lubrication condition parameters, enabling the model to comprehensively consider the coupling effects of multiple factors. When new sheet material data is input, the model can output a distribution map of the springback amount in each region after the entire sheet material is formed within seconds, i.e., the initial springback deviation field, providing accurate feedforward basis for subsequent process parameter optimization.

[0032] Furthermore, the generation of the dynamic curve of the partition blank holder force and the initial height distribution of the adjustable draw bead include: A decomposition-based multi-objective evolutionary algorithm is used for optimization. The dynamic curve of the partitioned blank holder force is parameterized using a B-spline curve. The initial height distribution of the adjustable drawing bead is set by the independent extension and retraction of multiple points of the array-type micro-electrically controlled lifting pins on the mold.

[0033] In this embodiment, to simultaneously minimize post-forming surface deviation and avoid cracking due to excessive thinning, the process parameter generation module employs a decomposition-based multi-objective evolutionary algorithm. This algorithm decomposes the multi-objective problem into multiple single-objective sub-problems and performs co-evolution to search for a set of Pareto optimal solutions. The partitioned blank holder force is no longer a constant value, but a continuously changing curve parameterized by a B-spline curve. By adjusting the curve control points, the rise, holding, and fall of the blank holder force during the stamping stroke can be precisely designed.

[0034] For adjustable draw beads, a series of closely spaced micro-electrically controlled lifting pins are arranged on the die along the draw bead trajectory. Each pin is controlled by an independent stepper motor or piezoelectric driver and can extend and retract independently, thereby locally changing the equivalent height of the draw bead.

[0035] During the solution process, the optimization algorithm continuously calls the surrogate model to predict the molding results under different parameter combinations, and finally gives the optimal dynamic curve control parameters of the partition blank holder force and the target height value of each lifting pin. This enables precise and differentiated constraints on the material flow along the path of the sheet metal flowing into the die, thereby reducing the springback driving force from the source.

[0036] Furthermore, the real-time acquisition of strain data, blank holder displacement data, and mold temperature data in the flange area of ​​the sheet metal within the mold, and inputting them into the digital twin model to dynamically predict the real-time springback trend, includes: The strain data is acquired by an array of fiber optic strain sensors embedded in the surface of the die and the pressure ring; The displacement data of the pressing ring is collected by a magnetostrictive displacement sensor integrated in the partition pressing cylinder; The digital twin model is based on the finite element reduced-order model and uses the collected strain and displacement data as boundary conditions for simulation, outputting the real-time rebound trend.

[0037] In this embodiment, a fiber optic strain sensor array is embedded in the contact surface of the die and the blank holder according to an optimized layout. These sensors can measure the minute deformation of the sheet flange under the blank holder force in real time and transmit it as an optical wavelength signal. They have strong anti-electromagnetic interference capability and fast response speed.

[0038] Meanwhile, each independent zoned blank holder cylinder integrates a magnetostrictive displacement sensor to accurately measure the real-time displacement of each zone's blank holder ring during the stamping process, reflecting the dynamic changes in sheet thickness. These strain and displacement data are synchronously acquired at kilohertz frequencies via a high-speed acquisition card and input as boundary conditions into a digital twin model deployed on a real-time controller. This digital twin model is not a complete finite element model, but a reduced-order model constructed based on techniques such as intrinsic orthogonal decomposition. It can complete a rapid simulation from the current state to the end of stamping within milliseconds, predicting the real-time springback trend that may occur after demolding if stamping continues under the current conditions, i.e., the springback direction and amplitude of each region.

[0039] Furthermore, the closed-loop adjustment of the actuator includes: Based on the real-time rebound trend, the blanking force of each zone of the partition blanking force servo control unit is independently adjusted, and the heating or cooling power of the local temperature regulating element array embedded in the punch or die is controlled. The local temperature regulating element array is composed of a mixture of thermoelectric coolers and resistance heating wires, which can regulate the material flow characteristics by changing the local temperature field of the sheet.

[0040] In this embodiment, the real-time simulation and control module generates dynamic compensation commands based on the real-time rebound trend predicted by the digital twin model. For the zoned blank holder force, it independently sends correction signals to the control valve of each electro-hydraulic servo blank holder cylinder, reducing the blank holder force in areas that require more material flow and increasing the blank holder force in areas that need to suppress wrinkling, thus forming a dynamically differentiated blank holder force distribution.

[0041] Simultaneously, an array of localized temperature-regulating elements embedded within the punch or die begins operation. This array consists of alternating micro-thermoelectric coolers and resistance heating wires, with each unit independently controllable. When positive springback is predicted in a certain area, the sheet metal in that area is locally heated to reduce its yield strength, making the material flow more easily and generating greater stretch during subsequent molding, thus offsetting the springback after demolding. Conversely, areas predicted to experience negative springback are locally cooled.

[0042] This method of actively regulating the temperature field to change the material flow characteristics and internal stress distribution provides a new control dimension for springback compensation in addition to mechanical force, realizing online active compensation during the stamping process.

[0043] Furthermore, the process drift early warning model is a deep learning model based on a temporal convolutional network or a long short-term memory network. Its input includes the partition blank holder force curves in the most recent N stamping cycles, the die temperature change sequence, the stress distribution feature sequence output by the digital twin model, and the corresponding measured surface deviation sequence; its output, the forming accuracy risk level, is the probability level of the formed part dimensions exceeding tolerance in the next few stamping cycles.

[0044] In this embodiment, the model is a dedicated deep learning prediction model for the time-series drift characteristics of stamping processes. It adopts a multi-level modular architecture, consisting of five core modules: a data preprocessing unit, a time-series feature extraction unit, a feature fusion mapping unit, a risk quantification output unit, and a model adaptive update unit. Each module works in tandem, fulfilling its specific function, and is adapted to the nonlinear, weakly drifting, and long-dependency characteristics of continuous time-series data in stamping production. The specific structure is as follows: Data preprocessing unit: As the model input preprocessor, it receives multi-source heterogeneous time-series data, including the partitioned blank holder force curve sequence, mold temperature change sequence, and sheet metal stress distribution characteristic sequence output by the real-time simulation and control module, as well as the measured surface deviation sequence output by the online detection module. This unit has data cleaning, alignment, normalization, and time-series slicing functions. It can remove abnormal noise data from sensors, unify the time axis scale of data from different acquisition frequencies, and trim continuous production data into fixed-length N-period time-series samples to eliminate dimensional differences and provide standardized input for subsequent feature extraction.

[0045] Temporal Feature Extraction Unit: This is the core perception unit of the model, employing a dual-architecture selectable mode. It can switch between a Temporal Convolutional Network (TCN) and a Long Short-Term Memory Network (LSTM) depending on the production conditions. The TCN, through multi-layer causal convolutions and dilated convolution structures, efficiently captures local short-term process fluctuations within the stamping cycle, adapting to short-term changes such as instantaneous fluctuations in blank holder force and local temperature shifts. The LSTM, through a dedicated gating structure, solves the gradient vanishing problem of traditional recurrent neural networks, accurately capturing long-term, slowly accumulating temporal change features caused by mold wear and process drift, covering implicit drift patterns across cycles.

[0046] Feature Fusion Mapping Unit: Responsible for deeply fusing the extracted blank holder force time-series features, temperature field drift features, stress distribution evolution features, and surface deviation accumulation features. Through a fully connected layer, it completes the adaptive weight allocation of multi-dimensional features, and explores the implicit correlations under multi-parameter coupling, such as the correspondence between blank holder force response delay caused by mold wear and the increase in springback deviation, the coupling law between temperature field offset and local forming deviation, and constructs the mapping relationship between process state and forming accuracy.

[0047] Risk quantification output unit: Based on the fused deep time series features, it completes the risk probability calculation through the classification activation function, transforms the abstract process drift features into quantifiable molding accuracy deviation probability, and matches the preset level classification rules to output standardized risk level results.

[0048] Model Adaptive Update Unit: Connects to the closed-loop learning module, supports periodic incremental retraining of the model, and can update model weights based on the latest production time series data to adapt to long-term working condition changes such as mold aging, environmental changes, and material batch fluctuations, thus avoiding model performance degradation.

[0049] This model runs continuously online during the stamping continuous production process. It uses a single stamping cycle as the data acquisition unit and N historical cycles as the time series analysis window to achieve fully automated operation of the entire process of "data acquisition - preprocessing - feature extraction - trend analysis - risk prediction - result output". The specific steps are as follows: Real-time acquisition and accumulation of multi-source time-series data: The model continuously connects with various functional modules of the system, collecting production data cycle by cycle, and caching the complete time-series datasets of the most recent N stamping cycles in real time. Among them, the process status data includes the actual following curve of blank holder force in each zone, the time-series change sequence of temperature across the entire die domain, and the stress distribution evolution sequence of key forming areas of the sheet metal; the forming result data is the measured three-dimensional deviation sequence of the corresponding part, forming a time-series sample library with a one-to-one correspondence between "process parameters and forming results".

[0050] Standardized preprocessing and temporal reconstruction: The data preprocessing unit performs noise reduction on the cached N-cycle data, eliminating outliers caused by stamping start-up and shutdown and instantaneous sensor interference; it completes multi-source data alignment based on a unified timestamp to ensure the temporal synchronization of blank holder force, temperature, stress, and deviation data; it unifies the data scale through a normalization algorithm and simultaneously completes temporal sample reconstruction to form a model input matrix with unified dimensions and continuous temporal sequence.

[0051] Multi-level temporal feature deep extraction: The corresponding feature extraction architecture is automatically called according to the production conditions. Short-term fluctuation features are extracted by temporal convolutional networks to capture abnormal disturbances in process parameters within a single cycle; long-term drift features are extracted by long short-term memory networks to uncover the trend shift of parameters within continuous cycles and accurately identify tiny hidden drifts that traditional threshold detection cannot detect, including core drift phenomena such as the gradual increase of blank holder force response hysteresis, the slow shift of the steady-state value of the mold temperature, and the gradual accumulation of springback deviation over cycle.

[0052] Multi-feature fusion and drift pattern analysis: The feature fusion unit integrates all temporal features, quantifies the influence weight of each process parameter drift on molding accuracy, constructs a process state evolution trend curve, determines whether the current process is in a stable state, a slight drift state, or a severe drift state, and identifies the core causes of drift (mold wear, temperature deviation, process parameter decay, etc.).

[0053] Future cycle risk probability prediction: Based on historical drift patterns and current process status, the model extrapolates the process evolution trend of several stamping cycles in the future, calculates the probability value of part surface dimensions exceeding tolerance, and outputs the corresponding forming accuracy risk level in combination with preset level rules.

[0054] Risk result output and closed-loop storage: After completing a single prediction, the model outputs a quantitative risk level and stores the time series data, feature parameters, and prediction results into the database to provide data support for subsequent model retraining and process traceability.

[0055] This model uses multi-parameter time-series drift trend and out-of-tolerance probability quantification as its core judgment criteria. Combined with the characteristics of sheet metal stamping forming process, it presets classification thresholds to achieve accurate and early risk assessment. The core judgment criteria and classification rules are as follows: The core judgment criteria are: first, the time-series drift characteristics, including four core drift indicators: blank holder force response delay, die temperature steady-state offset, cumulative deviation of sheet metal stress distribution, and periodic growth rate of part surface springback deviation; second, the probability quantification criteria, using the predicted probability of part forming dimensions exceeding tolerance within the next 3-10 stamping cycles as the direct judgment standard, comprehensively reflecting the severity of process drift and batch deviation risk.

[0056] Risk Level Classification and Judgment Rules: The model classifies molding accuracy risk into three levels, corresponding to different early warning and handling strategies: Level 1 (Low Risk): Predicted out-of-tolerance probability ≤ 10%, the time-series fluctuations of various process parameters are within the normal tolerance range, there is no continuous drift trend, molding accuracy is stable, and no intervention is required; Level 2 (Medium Risk): 10% < predicted out-of-tolerance probability ≤ 30%, there is slight continuous process drift, springback deviation, blank holder force response, and temperature field show slow shifts. Although there are currently no out-of-tolerance parts, there is potential batch risk, triggering an early warning prompt, and regular monitoring is recommended; Level 3 (High Risk): Predicted out-of-tolerance probability > 30%, the process drift trend is significant, multiple core parameters continuously deviate from the standard range, and there is a high probability of batch out-of-tolerance in the future, triggering a mandatory early warning and pushing specific maintenance and process adjustment plans.

[0057] Through the above-mentioned structured model architecture, full-time workflow and quantitative judgment criteria, this model can realize the early perception and accurate judgment of stamping process drift, transforming the traditional quality control mode of post-event quality inspection and error correction into a predictive maintenance mode based on data trends, and solving the problem of batch forming accuracy deviation caused by progressive mold wear and slow process drift.

[0058] Furthermore, the warning information includes the remaining number of punches for mold replacement or the draw bead height compensation value, and is displayed through a human-machine interface.

[0059] In this embodiment, when the risk level calculated by the process drift early warning model exceeds the preset early warning threshold, the early warning output module generates corresponding early warning information and pushes it to the human-machine interface on site. The early warning information not only informs the operator of the current risk level but, more importantly, provides specific action suggestions. For example, by analyzing the cumulative rate of springback deviation and the margin of the current compensation amount, the system estimates the remaining number of stamping cycles the die can safely continue to produce and suggests replacing the punch insert near the specified number of stamping cycles; or, if the risk can be compensated by adjusting the drawbead height, it will provide a specific numerical value for the drawbead height compensation. Operators can accurately perform maintenance operations without relying on personal experience or guesswork, maximizing die life while ensuring product quality.

[0060] Furthermore, the springback prediction proxy model uses the comparison results of the measured surface deviation and the initial springback deviation field to perform online reinforcement learning updates. This update adopts a near-end strategy optimization algorithm to reduce the prediction deviation as a reward signal. The process drift early warning model is periodically retrained after accumulating a new batch of stamping data.

[0061] In this embodiment, the closed-loop learning module is responsible for the adaptive update of the two models. For the springback prediction proxy model, after each part is demolded and the measured surface deviation is obtained through online 3D measurement, the system compares the measured deviation field with the initial springback deviation field predicted by the model before the current stamping and calculates the prediction error.

[0062] This error serves as a reward signal for reinforcement learning. The weights of the surrogate model are adjusted incrementally online through a proximal policy optimization algorithm, enabling the model to automatically adapt to the slow changes in springback characteristics caused by mold wear and the seasonal fluctuations in ambient temperature.

[0063] For the process drift early warning model, after accumulating a certain number of new stamping cycle data, the system periodically retrains the model using the latest dataset, including the new data, to update its internal trend representation. This ensures that the early warning model can always track the latest equipment degradation trajectory and process drift patterns. These two complementary learning mechanisms together constitute the system's self-evolution capability, maintaining control accuracy over the long term and preventing model degradation.

[0064] Please see Figure 4 A precision control system for stamping and forming of automotive sheet metal parts, used in the precision control method for stamping and forming of automotive sheet metal parts, includes a sheet metal parameter online sensing module 1, a springback prediction module 2, a process parameter generation module 3, a stamping and forming actuator 4, a real-time simulation and control module 5, an online detection module 6, a process drift early warning module 7, an early warning output module 8, and a closed-loop learning module 9. The output of the online sheet metal parameter sensing module 1 is connected to the input of the springback prediction module 2. The output of the springback prediction module 2 is connected to the input of the process parameter generation module 3. The output of the process parameter generation module 3 is connected to the control of the stamping forming actuator 4. The sensor output of the stamping forming actuator 4 is connected to the input of the real-time simulation and control module 5. The control output of the real-time simulation and control module 5 is connected to the actuator of the stamping forming actuator 4. The output of the online detection module 6 is connected to the input of the process drift warning module 7. The output of the process drift warning module 7 is connected to the warning output module 8. The input of the closed-loop learning module 9 is connected to both the online detection module 6 and the real-time simulation and control module 5. Its output is connected to both the springback prediction module 2 and the process drift warning module 7. The online sensing module 1 for sheet material parameters is used to acquire the initial three-dimensional contour data, thickness distribution data and material property fluctuation parameters of the sheet material to be formed. The rebound prediction module 2 is equipped with an online-updable rebound prediction proxy model for predicting the initial rebound deviation field. The process parameter generation module 3 is used to generate a dynamic curve of partitioned blank holder force and an adjustable initial height distribution of draw beads based on the initial springback deviation field. The stamping forming actuator 4 includes a partitioned blank holder ring, an adjustable drawing bead, an electro-hydraulic servo blank holder cylinder, a local temperature regulating element array composed of a thermoelectric cooler and a resistance heating wire, and fiber optic strain sensor, magnetostrictive displacement sensor and temperature sensor distributed in the mold. The real-time simulation and control module 5 runs a digital twin model to receive sensor data and output control commands to adjust the actuator in a closed loop. The online detection module 6 is used to acquire the measured surface data of the molded part and calculate the measured surface deviation; The process drift early warning module 7 has a built-in process drift early warning model, which is used to predict the molding accuracy risk level based on the time series of process data and measured surface deviation. The early warning output module 8 is used to output early warning information including the remaining number of punches for mold replacement or the height compensation value of the drawing bead when the risk level reaches the early warning threshold. The closed-loop learning module 9 is used to feed back the measured surface deviation to the springback prediction module 2, and to feed back the process and deviation data to the process drift early warning module 7.

[0065] In this embodiment, after receiving the online sheet material signal, the online sheet material parameter sensing module 1 triggers the sensor to collect data and sends the data to the springback prediction module 2.

[0066] The springback prediction module 2 uses a springback prediction proxy model to quickly calculate the initial springback deviation field and outputs it to the process parameter generation module 3.

[0067] The process parameter generation module 3 generates a dynamic curve of the partitioned blank holder force and a drawing bead height distribution scheme through an optimization algorithm, and converts them into control commands to set the initial state of the stamping forming actuator 4. During the stamping process, fiber optic strain sensors, magnetostrictive displacement sensors, and temperature sensors distributed within the mold collect data in real time and transmit it to the real-time simulation and control module 5. The digital twin model running in this module simulates and predicts the springback trend in real time based on the sensor data, and outputs control commands to the electro-hydraulic servo blank holder cylinder and the local temperature regulating element array to achieve dynamic closed-loop compensation.

[0068] After the stamping is completed and the part is demolded, the online detection module 6 scans the part to obtain the measured surface deviation. This deviation is sent to the process drift early warning module 7 to predict the future risk level in combination with the process data, and is also obtained by the closed-loop learning module 9.

[0069] The closed-loop learning module 9 receives measured deviations and process data, performs online reinforcement learning updates on the springback prediction proxy model, and periodically triggers the retraining of the process drift early warning model.

[0070] When the warning output module 8 receives a risk level exceeding the threshold, it immediately displays information including maintenance suggestions on the operator's human-machine interface. All modules of the entire system work together to form a complete technical closed loop from real-time perception, feedforward prediction, online closed-loop control to trend warning and continuous self-learning.

[0071] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that implementing all or part of the above embodiments and making equivalent changes in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for controlling the stamping precision of automotive sheet metal parts, characterized in that, Includes the following steps: The initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed can be obtained online. The acquired data is input into a pre-trained springback prediction proxy model to predict the initial springback deviation field after the sheet metal is formed. Based on the initial springback deviation field, with the goal of minimizing the surface deviation after forming, a dynamic curve of partitioned blank holder force and an initial height distribution of adjustable draw beads are generated, and the stamping actuator is set accordingly. During continuous stamping, strain data, pressure ring displacement data, and mold temperature data of the sheet metal flange area inside the mold are collected in real time, and input into the digital twin model to dynamically predict the real-time springback trend, and to perform closed-loop adjustment of the actuator. After each stamping and demolding, the measured surface deviation of the formed part is obtained. The time series containing the measured surface deviation and process data is input into the process drift early warning model to predict the forming accuracy risk level of future stamping cycles. When the predicted risk level reaches the preset warning threshold, a warning message is generated.

2. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The online acquisition of the initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed includes: The initial three-dimensional contour data is obtained by a structured light three-dimensional contour sensor installed on a feeding robot, and the thickness distribution data is obtained by an electromagnetic ultrasonic thickness measuring probe. By scanning the information code on the sheet metal, the material property fluctuation parameters, including the statistical distribution parameters of yield strength, tensile strength, and thickness anisotropy coefficient, are retrieved from the manufacturing execution system.

3. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The springback prediction proxy model is a deep learning model that integrates convolutional neural networks and long short-term memory networks. Its inputs also include mold geometric feature encoding, preset stamping speed and lubrication condition parameters.

4. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The generation of the dynamic curve of the partition blank holder force and the initial height distribution of the adjustable draw bead include: A decomposition-based multi-objective evolutionary algorithm is used for optimization. The dynamic curve of the partitioned blank holder force is parameterized using a B-spline curve. The initial height distribution of the adjustable drawing bead is set by the independent extension and retraction of multiple points of the array-type micro-electrically controlled lifting pins on the mold.

5. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The real-time acquisition of strain data, blank holder displacement data, and mold temperature data in the flange area of ​​the sheet metal inside the mold, and inputting them into the digital twin model to dynamically predict the real-time springback trend, includes: The strain data is acquired by an array of fiber optic strain sensors embedded in the surface of the die and the pressure ring; The displacement data of the pressing ring is collected by a magnetostrictive displacement sensor integrated in the partition pressing cylinder; The digital twin model is based on the finite element reduced-order model and uses the collected strain and displacement data as boundary conditions for simulation, outputting the real-time rebound trend.

6. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The closed-loop adjustment of the actuator includes: Based on the real-time rebound trend, the blanking force of each zone of the partition blanking force servo control unit is independently adjusted, and the heating or cooling power of the local temperature regulating element array embedded in the punch or die is controlled. The local temperature regulating element array is composed of a mixture of thermoelectric coolers and resistance heating wires, which can regulate the material flow characteristics by changing the local temperature field of the sheet.

7. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 3, characterized in that, The process drift early warning model is a deep learning model based on temporal convolutional networks or long short-term memory networks. Its inputs include the partition blank holder force curves in the most recent N stamping cycles, the die temperature change sequence, the stress distribution feature sequence output by the digital twin model, and the corresponding measured surface deviation sequence. Its output, the forming accuracy risk level, is the probability level of the formed part dimensions exceeding tolerance in the next few stamping cycles.

8. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 1, characterized in that, The warning information includes the remaining number of punches for mold replacement or the compensation value for the height of the drawing bead, and is displayed through a human-machine interface.

9. The method for controlling the stamping precision of automotive sheet metal parts as described in claim 7, characterized in that, The rebound prediction proxy model is updated online using the comparison results between the measured surface deviation and the initial rebound deviation field. This update adopts a near-end policy optimization algorithm to reduce the prediction deviation as a reward signal. The process drift early warning model is periodically retrained after accumulating a new batch of stamping data.

10. A precision control system for stamping automotive sheet metal parts, used to implement the precision control method for stamping automotive sheet metal parts as described in any one of claims 1 to 9, characterized in that, It includes a sheet metal parameter online sensing module, a springback prediction module, a process parameter generation module, a stamping forming actuator, a real-time simulation and control module, an online detection module, a process drift early warning module, an early warning output module, and a closed-loop learning module; The output of the online sheet metal parameter sensing module is connected to the input of the springback prediction module; the output of the springback prediction module is connected to the input of the process parameter generation module; the output of the process parameter generation module is connected to the control of the stamping forming actuator; the sensor output of the stamping forming actuator is connected to the input of the real-time simulation and control module; the control output of the real-time simulation and control module is connected to the actuator of the stamping forming actuator; the output of the online detection module is connected to the input of the process drift early warning module; the output of the process drift early warning module is connected to the early warning output module; the input of the closed-loop learning module is connected to both the online detection module and the real-time simulation and control module, and its output is connected to both the springback prediction module and the process drift early warning module. The online sensing module for sheet material parameters is used to acquire the initial three-dimensional contour data, thickness distribution data, and material property fluctuation parameters of the sheet material to be formed. The rebound prediction module is equipped with an online-updable rebound prediction proxy model for predicting the initial rebound deviation field. The process parameter generation module is used to generate a dynamic curve of partitioned blank holder force and an adjustable initial height distribution of draw beads based on the initial springback deviation field. The stamping forming actuator includes a partitioned blank holder ring, an adjustable drawing bead, an electro-hydraulic servo blank holder cylinder, a local temperature regulating element array composed of a thermoelectric cooler and a resistance heating wire, and fiber optic strain sensors, magnetostrictive displacement sensors and temperature sensors distributed in the mold. The real-time simulation and control module runs a digital twin model to receive sensor data and output control commands to adjust the actuator in a closed loop. The online detection module is used to acquire the measured surface data of the molded part and calculate the measured surface deviation. The process drift early warning module has a built-in process drift early warning model, which is used to predict the molding accuracy risk level based on the time series of process data and measured surface deviation. The early warning output module is used to output early warning information including the remaining number of punches for mold replacement or the height compensation value of the drawing bead when the risk level reaches the early warning threshold. The closed-loop learning module is used to feed back the measured surface deviation to the springback prediction module and to feed back the process and deviation data to the process drift early warning module.

Images