A cold heading forming control method and system for a vehicle body flange sleeve

Abstract

The application relates to the technical field of automatic control, and discloses a cold upsetting forming control method and system for a flange sleeve of an automobile body, the method comprising the following steps: intelligently judging the deviation trend of a material flow demarcation point by monitoring deformation data, then extracting and iteratively correcting a root stress model to obtain a stress evolution sequence; positioning and identifying a mismatch peak value and coordinates of the mismatch peak value which indicate a cracking risk by time sequence analysis on the sequence; calculating a strain gradient of a transition zone of a cylinder wall according to the mismatch peak value and the coordinates, determining an alternating state of tensile stress and compressive stress, and then calculating an accurate stroke compensation increment; fine-tuning simulation is carried out based on the increment, the stable position of the demarcation point can be predicted, and the uniform distribution of the root stress can be verified; intelligently matching the uniform state with a material performance limit, and generating a final stroke optimization vector after the requirement is met; and finally realizing real-time monitoring and dynamic stroke compensation of a production process. The method can improve the stability of production forming quality.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and in particular to a method and system for controlling the cold heading of automotive body flange sleeves. Background Technology

[0002] Currently, the automotive industry increasingly demands lightweight and high reliability. As a critical connecting component, the cold heading quality of automotive body flange sleeves directly impacts overall vehicle performance. With the widespread application of high-strength steel, traditional cold heading processes face challenges in controlling material flow and stress distribution, easily leading to defects such as root cracking. Against this backdrop, this invention proposes an intelligent control method integrating finite element simulation and neural network models. This method aims to achieve real-time optimization and compensation of process parameters through an industrial automatic control system, thereby improving forming accuracy and stability.

[0003] In a current technology, the cold heading control of automotive body flange sleeves primarily relies on manual experience to set process parameters and repeatedly adjust the mold through trial molding to approximate the target shape. This method is typically based on static mold design and a fixed stroke path, lacking real-time monitoring and feedback of the dynamic stress state during material deformation. Specifically, operators need to estimate the possible location of the material flow boundary point based on experience and manually adjust the equipment stroke, but cannot accurately capture the real-time offset trend of the boundary point during the deformation stage, and it is even more difficult to quantify the correlation between root stress evolution and cracking risk. Due to the lack of support from an industrial automatic control system, the current technology struggles to achieve closed-loop control of data acquisition, real-time analysis, and dynamic compensation, resulting in lagging and coarse process adjustments that cannot effectively suppress defects caused by stress mismatch in high-strength steel during cold heading. Therefore, the core deficiency of the current technology lies in its reliance on offline experience-based trial and error, lacking intelligent decision-making capabilities driven by real-time data, and failing to instantly identify the offset of the material flow boundary point and fine-tune the stroke during the forming process, thus restricting the stability and reliability of the flange sleeve forming quality.

[0004] In summary, existing technologies suffer from unstable molding quality. Summary of the Invention

[0005] This invention provides a method and system for controlling the cold heading of automotive body flange sleeves, in order to solve the problem of unstable forming quality in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for controlling the cold heading of an automotive body flange sleeve, comprising:

[0007] Acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset condition. Based on the offset trend, the stress distribution vector at the root of the flange sleeve is extracted. A stress model is constructed based on the stress distribution vector and iteratively corrected. The corrected stress model is then used for calculation to obtain the stress evolution sequence. A time-series analysis was performed on the stress evolution sequence to obtain the mismatch peak value, and the real-time location coordinates corresponding to the mismatch peak value were located. Based on the real-time position coordinates, the strain gradient of the transition zone of the flange sleeve wall is calculated, and the alternation state of tensile and compressive stress is determined based on the strain gradient. Based on the alternation state of tensile and compressive stress, the adjustment increment required for stroke compensation is calculated. Based on the adjustment increment, the cold heading process stroke is fine-tuned and simulated. By analyzing the simulation results, the stable position of the material flow boundary point is predicted. Based on the stable position, the uniform distribution state of the root stress is determined. The uniform distribution state and the preset high-strength steel material limit are matched to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector. Based on the travel optimization vector, the preset deformation control model is updated. Through the updated deformation control model, the dynamic stress of the production process is monitored in real time, and compensation adjustment instructions are generated.

[0008] Secondly, the present invention provides a cold heading forming control system for an automotive body flange sleeve, comprising: The data acquisition module is used to acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset conditions. The model iteration module is used to extract the stress distribution vector at the root of the flange sleeve according to the offset trend, construct a stress model based on the stress distribution vector and perform iterative correction, and apply the corrected stress model to calculate and obtain the stress evolution sequence. The coordinate positioning module is used to perform time-series analysis on the stress evolution sequence, obtain the mismatch peak value, and locate the real-time position coordinates corresponding to the mismatch peak value. The compensation calculation module is used to calculate the strain gradient of the transition zone of the flange sleeve wall based on the real-time position coordinates, determine the alternating state of tensile and compressive stress based on the strain gradient, and calculate the adjustment increment required for stroke compensation based on the alternating state of tensile and compressive stress. The process simulation module is used to fine-tune the cold heading process stroke according to the adjustment increment, predict the stable position of the material flow boundary point by analyzing the simulation results, and determine the uniform distribution state of the root stress according to the stable position. The stroke optimization module is used to match the uniform distribution state with the preset high-strength steel material limit to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector. The control output module is used to update the preset deformation control model according to the stroke optimization vector, and to monitor the dynamic force of the production process in real time through the updated deformation control model, and generate compensation adjustment instructions.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention acquires and calculates cold heading deformation data in real time, obtains the initial offset of the material, and classifies it according to preset rules. Then, when the classification result meets the offset condition, it accurately determines the offset trend of the material flow boundary point. This step realizes the accurate identification and trend prediction of the material flow behavior in the early stage of forming, providing a key initial decision basis for subsequent precise control, avoiding the lag and deviation caused by the reliance on experience judgment in traditional methods, and improving the directionality and predictability of process control from the source.

[0010] (2) Based on the determined offset trend, this invention specifically extracts the stress distribution vector at the root of the flange sleeve and constructs an iteratively correctable stress model. By applying the corrected model for calculation, a stress evolution sequence that reflects the dynamic change law of stress is obtained. This process realizes the quantitative modeling and dynamic deduction of the stress state at the root, overcomes the limitation of traditional static analysis in reflecting process changes, and provides an accurate data foundation for in-depth analysis of the process and the causes of defects.

[0011] (3) This invention performs time-series analysis on the stress evolution sequence, identifies key peaks characterizing stress mismatch, and obtains their corresponding real-time location coordinates, thereby establishing a quantitative relationship between these location coordinates and the risk of cracking defects. This step realizes the association between abstract stress data and specific, locatable physical defect risks, completing the transformation from process monitoring to risk warning, enabling the location and severity of potential defects to be assessed in advance and quantitatively, significantly improving the pertinence and effectiveness of defect prevention.

[0012] (4) Based on the risk coordinates, this invention calculates the strain gradient of the transition zone of the cylinder wall and analyzes its stress characteristics, thereby calculating the required stroke compensation increment; then, using this increment, it performs fine-tuning simulation of the process stroke, predicting the stable position of the boundary point and the corresponding uniform distribution of root stress. This series of operations, through the closed-loop path of "analyzing characteristics - calculating increment - simulation verification", realizes the refined and scientific adjustment and pre-simulation of process parameters, ensuring that the proposed compensation amount can effectively guide the material flow to stabilize and optimize the stress distribution, thereby improving the accuracy of process adjustment and the success rate.

[0013] (5) This invention determines the matching degree between the simulated uniform distribution state and the material performance limit, and extracts key process data to generate the final stroke optimization vector when the requirements are met; then, this vector is used to update the deformation control model, and finally, the production process is monitored in real time and the stroke is dynamically adjusted through this model. This step completes the transformation from the optimization theoretical vector to the control model, and then to the on-site execution command, establishing an industrial automatic control system that can self-adjust according to the real-time state, realizing the leap from static setting to dynamic optimization of the process, and continuously ensuring the stability and reliability of the molding quality. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of a cold heading control method for an automobile body flange sleeve provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a cold heading control system for an automobile body flange sleeve provided in the second embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 The first embodiment of the present invention provides a method for controlling the cold heading of an automotive body flange sleeve, comprising the following steps: S11, acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset condition. S12, Based on the offset trend, extract the stress distribution vector at the root of the flange sleeve, construct a stress model based on the stress distribution vector and perform iterative correction, apply the corrected stress model to perform calculations, and obtain the stress evolution sequence. S13, perform time series analysis on the stress evolution sequence to obtain the mismatch peak value, and locate the real-time position coordinates corresponding to the mismatch peak value; S14. Calculate the strain gradient of the transition zone of the flange sleeve wall based on the real-time position coordinates, determine the alternating state of tensile and compressive stress based on the strain gradient, and calculate the adjustment increment required for stroke compensation based on the alternating state of tensile and compressive stress. S15, according to the adjustment increment, the cold heading process stroke is fine-tuned and simulated. By analyzing the simulation results, the stable position of the material flow boundary point is predicted. Based on the stable position, the uniform distribution state of the root stress is determined. S16, Match the uniform distribution state with the preset high-strength steel material limit to obtain the matching degree. If the matching degree meets the preset forming quality requirements, extract process data from the uniform distribution state to generate the final stroke optimization vector. S17. Based on the stroke optimization vector, update the preset deformation control model. Through the updated deformation control model, monitor the dynamic stress of the production process in real time and generate compensation adjustment instructions.

[0017] In step S11, cold heading deformation data is acquired, and the initial offset of the material is calculated based on the cold heading deformation data. The initial offset is then classified, and when the classification result meets a preset offset condition, the offset trend of the boundary point is determined, including: Acquire cold heading deformation data, and calculate the initial offset of the material flow boundary point based on the cold heading deformation data; The initial offset is input into a preset classification neural network model for preliminary classification to obtain classification labels and confidence levels; If the confidence level exceeds a preset offset threshold, the offset trend of the boundary point is determined.

[0018] It should be noted that the cold heading deformation data is obtained through finite element simulation of the high-strength steel cold heading process. This simulation is based on the material constitutive model, mold geometry parameters, and a preset process path. It calculates and outputs a full-field dataset containing the three-dimensional coordinate changes, displacement, stress, and strain of each node of the material. The focus is on the material flow trajectory data to identify the deformation mode. Based on this dataset, the material flow boundary point is located by analyzing the spatial gradient abrupt change characteristics of the displacement field or velocity field. This boundary point represents the key interface for the material to switch from one dominant flow mode to another. When calculating the initial offset, the system symmetrically selects several sampling points on both sides of the boundary point and calculates the average displacement of these sampling points along a specific direction (such as radial or axial). Then, the difference between the average displacement values ​​on both sides is taken as the initial offset value. This value quantifies the degree of asymmetry in the material flow.

[0019] Then, the offset is input into the neural network model for preliminary classification. The specific process of obtaining the classification label and confidence score is as follows: The model is a standard multilayer perceptron classification model. Its network structure consists of an input layer (1 node, corresponding to the offset scalar), three fully connected hidden layers (64, 32, and 16 neurons respectively), and an output layer (the number of nodes is consistent with the preset number of flow pattern classifications, such as 5 classes). The hidden layers use the ReLU activation function to introduce nonlinearity, and the output layer uses the Softmax function to convert the original score into a probability distribution. The data used for model training is constructed as follows: First, thousands of cold heading process cases are collected from historical production databases and high-fidelity finite element simulation case libraries. Each case includes an initial material offset value calculated by an image correlation algorithm, and a flow pattern classification (such as "no offset" or "slight positive offset") jointly determined by at least two process experts based on the final metallographic flow line morphology of the corresponding workpiece as the "true label".

[0020] Subsequently, all collected offset values ​​were cleaned to remove outliers that clearly exceeded physical meaning, and then randomly divided into training, validation, and test sets in an 8:1:1 ratio. Finally, based on the statistical characteristics (mean and standard deviation) of the training set, Z-score standardization was performed on all data (including training, validation, and test sets) to transform them into a distribution with a mean of 0 and a standard deviation of 1, thus eliminating dimensions and accelerating model training convergence. The core training parameters of the model include: using the cross-entropy loss function to measure the difference between the predicted distribution and the true label; using the Adam optimizer for weight updates with an initial learning rate set to 0.001; and employing an early stopping strategy to prevent overfitting. Before training, all historical offset data underwent standardization preprocessing, and the training period was typically set to 200 epochs with a batch size of 32. When applied, the model receives the currently calculated offset value as input. After nonlinear transformation by the hidden layer of the network, it outputs a classification label and a confidence score between 0 and 1 at the output layer. This confidence score is the maximum value in the probability distribution output by Softmax, reflecting the model's confidence in the classification judgment.

[0021] If the confidence level exceeds a preset offset threshold, the threshold setting process is as follows: After model training is completed, an independent validation dataset is used for evaluation. By analyzing the model's prediction results for different confidence level intervals on the validation set, a "confidence level - classification accuracy" relationship curve is plotted. Balancing the requirements of high accuracy with sufficient sample coverage, a critical value that stably meets the minimum reliability standard (e.g., classification accuracy for the corresponding confidence level interval is not less than 95%) is selected from the curve as the threshold. For example, if set to 0.8, the system adopts the classification label and, based on the offset direction (e.g., "right offset") and degree (e.g., "severe") indicated by the label, combined with the typical material flow vector field characteristics under this classification mode, determines a specific boundary point offset trend vector. This vector not only includes the offset direction but may also include the expected offset rate or range, providing a clear and quantifiable directional input basis for subsequent root stress calculation.

[0022] In step S12, based on the offset trend, the stress distribution vector at the root of the flange sleeve is extracted. A stress model is constructed based on the stress distribution vector and iteratively corrected. The corrected stress model is then applied for calculation to obtain the stress evolution sequence, including: Based on the offset trend, locate the root region of the flange sleeve and extract the stress data of the node; Feature extraction is performed on the stress data to construct a stress distribution feature vector, and a mapping matrix between the stress distribution feature vector and the dynamic force relationship is established. The theoretical stress value is predicted based on the mapping matrix, and a finite element simulation is performed on the mapping matrix. If the stress deviation between the theoretical stress value and the finite element simulation result exceeds a preset stress deviation threshold, the mapping matrix is ​​updated based on the stress deviation to obtain an updated mapping matrix. The stress state of the flange sleeve root region is analyzed using the updated mapping matrix to obtain the stress evolution sequence of the root stress with the deformation stage.

[0023] It should be noted that when locating the flange sleeve root region and extracting nodal stress data based on the aforementioned offset trend, the system first analyzes the direction and intensity information contained in the offset trend vector of the dividing point determined in step S101. This vector not only indicates the offset direction (e.g., pointing to the right of the mold's central axis), but its length also quantifies the severity of the offset. Based on this information, the system defines a local coordinate system at the flange sleeve root of the finite element model and automatically delineates a fan-shaped monitoring area using the direction angle of this vector as a reference. For example, with the vector direction as the center line, it extends 15 degrees to both sides to form a fan with a total angle of 30 degrees. A preset search depth (e.g., 3-5 mesh elements) is extended along the vector direction from the dividing point to the inner side of the root fillet. All finite element mesh nodes within this three-dimensional fan-shaped area are included in the filtering range. Simultaneously, to establish a comparison benchmark, the system symmetrically selects the same number of nodes in the opposite direction of the vector direction as a reference group, thereby accurately locking down the area with the most active mechanical behavior and requiring key monitoring. Subsequently, complete stress tensor data, including three normal stress components and three shear stress components, are extracted from the output results of each selected node within this area.

[0024] Then, when extracting features from the stress data, constructing a stress distribution feature vector, and establishing a mapping matrix between the stress distribution feature vector and the dynamic force relationship, the system performs principal stress analysis on the collected nodal stress tensor, calculates the first, second, and third principal stress values ​​and principal directions for each node, and simultaneously calculates the von Mises equivalent stress. Subsequently, the system splices the above key indicators (each node contains 3 principal stress values, 3 principal direction cosines, and 1 equivalent stress, for a total of 7 scalars) of all nodes (e.g., N nodes) in the monitoring area according to a predetermined node order to form a row vector with a dimension of 7N. This vector is the stress distribution feature vector, which comprehensively characterizes the intensity, direction, and distribution heterogeneity of the stress field in the monitoring area. To establish the mapping relationship, the system calls a historical training database, which stores data pairs consisting of a large number of finite element simulation cases. Each case contains a stress distribution feature vector (as input features) and its corresponding dynamic stress state parameter vector (as target values). The latter is output by the solver and typically includes the resultant reaction force components, total strain energy, and average maximum principal stress of all nodes in the monitoring area. The system uses a multiple linear regression algorithm to train this dataset. The specific training process is as follows: First, all stress distribution feature vectors in the training set are arranged in rows to form a feature matrix, and the corresponding dynamic stress state parameter vectors are arranged in rows to form a target matrix. Then, the weight coefficient matrix is ​​calculated by solving the least squares solution of the linear equation system.

[0025] Next, the mapping matrix is ​​iteratively verified through finite element simulation. When the theoretical stress value is predicted based on the mapping matrix, the system inputs the stress distribution feature vector calculated in real time into the initial mapping matrix, and directly outputs a set of predicted dynamic stress parameters through matrix multiplication. Then, based on the material constitutive relation, these parameters are converted into the equivalent stress theoretical value at the key root position.

[0026] In the specific process of iterative verification, the system first calls the finite element solver to perform a one-step incremental calculation based on the geometric and boundary conditions of the current cold heading deformation stage, directly obtaining the true stress tensor data of all nodes in the monitoring area at the root of the flange sleeve as a benchmark; simultaneously, the system inputs the stress distribution feature vector extracted under the same deformation stage into the mapping matrix to be verified, and obtains the predicted dynamic stress parameter vector through matrix multiplication. Then, according to the preset material constitutive model (such as the Johnson-Cook dynamic constitutive model), these dynamic stress parameters are mapped to the equivalent theoretical stress values ​​of key locations at the root (such as stress concentration points); the Johnson-Cook model is an empirical constitutive relation, it... A set of material constants characterizes the coupling relationship between flow stress, equivalent plastic strain, strain rate, and temperature in metallic materials during plastic deformation. Core parameters include yield strength at a reference strain rate, strain hardening coefficient, strain rate sensitivity coefficient, and thermal softening coefficient. These parameters are calibrated through uniaxial tensile, compressive, and Hopkinson bar dynamic mechanical experiments on the high-strength steel. In this application, the system combines the dynamic stress parameters output from the mapping matrix (such as the loading force at a specific strain rate) with the equivalent plastic strain and estimated temperature rise of the current calculation step, substituting them into the calculation framework of the constitutive model to solve for the corresponding flow stress. This flow stress serves as the theoretical equivalent stress value at that location under the deformation conditions. The system then calculates the relative deviation between this theoretical equivalent stress value and the corresponding true equivalent stress value output by the finite element simulation to quantify the prediction error.

[0027] If the stress deviation between the theoretical stress value and the finite element simulation result exceeds a preset stress deviation threshold, the mapping matrix is ​​updated according to the stress deviation. The specific update process is as follows: First, the system defines a loss function, typically using the sum of squares (mean square error) of the differences between the predicted theoretical stress value and the actual finite element stress value as the objective function. Next, the system uses automatic differentiation to calculate the partial derivative of the loss function with respect to each weight element in the mapping matrix, i.e., the gradient vector. The gradient direction indicates the direction of the fastest increase in the loss function. Then, the system sets an initial learning rate parameter (e.g., 0.01) and updates all current weight elements of the mapping matrix along the opposite direction of the gradient vector. The value is the product of the learning rate and the corresponding gradient component. This step aims to reduce the value of the loss function. When the deviation exceeds the limit, the system uses gradient descent to update. Specifically, the error between the predicted theoretical stress value and the finite element true stress value is calculated, and the gradient of this error relative to each weight element in the mapping matrix is ​​obtained. Then, the weight elements are fine-tuned along the gradient in the opposite direction with a preset learning rate. The above calculation and update steps are performed in an iterative loop. After each iteration, the updated mapping matrix is ​​used to recalculate the theoretical stress value and evaluate the loss. If the loss value is still higher than the allowable error threshold, a new round of gradient calculation and weight update is continued until the prediction error drops below the threshold or the preset maximum number of iterations is reached.

[0028] The determination of the preset stress deviation threshold is first based on the final requirements for the cold heading quality of the flange sleeve, especially the maximum allowable stress prediction error to prevent micro-cracking in the root region. Engineers define an acceptable range of relative stress prediction error (e.g., ±5%) based on the material's mechanical properties and historical process data. Subsequently, high-precision finite element simulations are performed under a large number of typical process parameter combinations. The prediction results of the simplified model (i.e., using a mapping matrix) are compared with the simulation results of the full model, and the deviation distribution is statistically analyzed. Finally, to ensure the reliability of the model prediction, the critical error value that meets the prediction accuracy requirements of the vast majority (e.g., 99%) of cases in the statistical distribution is formally set as the fixed threshold for system operation.

[0029] Finally, when using the updated mapping matrix to analyze the stress state of the flange sleeve root region and obtain the stress evolution sequence of the root stress with deformation stages, the specific analysis and generation process is as follows: The system first extracts the original three-dimensional stress signals of all monitoring points in the root region at each deformation time (at a fixed time interval, such as 0.5 milliseconds) from the online monitoring data stream, after coordinate transformation, and calculates its stress invariants to form a feature vector representing the current overall stress distribution state. Subsequently, this feature vector is multiplied by a time-varying, staged mapping matrix (this matrix is ​​updated after each process optimization iteration, and its element values ​​are calibrated by least-squares fitting of a large amount of offline simulation and measured data), directly analyzing and outputting the representative equivalent stress value of the root region at the current time and its first and third principal stress components. The system repeats this operation for each time step to obtain a discrete stress-time point sequence. Next, the system employs a timestamp-based strict alignment algorithm to sort the analytical results of all time steps according to the deformation process (from initial upsetting, intermediate upsetting to final upsetting). A cubic spline interpolation algorithm is then used to fit the discrete points, filling any potential data acquisition gaps and ensuring the curve's second-order continuity and differentiability. Finally, a moving average filter is applied to the interpolated continuous curve to suppress high-frequency noise, generating a smooth, continuous stress evolution sequence that accurately reflects the dynamic evolution of stress throughout the entire process.

[0030] In step S13, time-series analysis is performed on the stress evolution sequence to obtain the mismatch peak value, and the real-time location coordinates corresponding to the mismatch peak value are located, including: The stress evolution sequence is input into a preset predictive neural network model to generate a time-series prediction curve of stress. By calculating the time-series prediction curve and comparing it with a preset benchmark stress threshold, the mismatch peak value and the time index corresponding to the mismatch peak value are identified. Based on the time index, retrieve the node displacement data of the preset finite element mesh model, locate the mesh node that generated the mismatch peak, and obtain its real-time position coordinates.

[0031] It should be noted that when the stress evolution sequence is input into the preset neural network model to generate the time-series prediction curve of stress, the preset neural network model used by the system is a Long Short-Term Memory (LSTM) network. The specific configuration of this model is as follows: the network structure includes an input layer, two stacked LSTM hidden layers, and an output layer. The input layer receives a normalized stress sequence of length 60 (corresponding to the critical time steps in the cold heading process). The first LSTM hidden layer contains 128 neurons, and the second contains 64 neurons. Each LSTM layer is followed by a Dropout layer with a dropout rate of 0.2 to prevent overfitting. The output layer is a fully connected layer with a linear activation function, used to predict the stress values ​​for the next 8 consecutive time steps. The model uses mean squared error as the loss function and is trained using the Adam optimizer, with a learning rate set to 0.001, a first-moment estimation decay factor beta1 of 0.9, and a second-moment estimation decay factor beta2 of 0.999. The training batch size was set to 32, the maximum number of training rounds was 200, and a training strategy was adopted to stop the training early if the validation set loss did not decrease for 10 consecutive rounds.

[0032] The training data construction process is as follows: Thousands of complete stress evolution sequences are extracted from historical production databases and high-fidelity simulation libraries. For each sequence, a sliding window of fixed length 60 and a sliding step size of 1 is used to segment it, generating multiple continuous subsequences of length 60 as model input samples. For each input sample, its corresponding training label is a continuous segment of stress values ​​of length 8 immediately following the 60-step subsequence. Subsequently, all generated "input-label" sample pairs are randomly divided into training, validation, and test sets in a 7:2:1 ratio. Finally, Z-score standardization is performed on the entire dataset based on the statistical values ​​(mean and standard deviation) of all samples in the training set to eliminate dimensions and accelerate model convergence. The network captures the long-term dependencies and short-term fluctuation patterns of stress data in the time dimension through its internal gating mechanism. Finally, by performing a forward propagation calculation, the model outputs a continuous curve predicting future stress changes based on the input stress evolution sequence, thus obtaining the time-series stress prediction curve.

[0033] Then, by calculating the generated time-series prediction curve and comparing it with the preset benchmark stress threshold to identify the mismatch peak and the corresponding time index, the system uses the five-point central difference formula to perform discrete point numerical differentiation on the time-series prediction curve to calculate its second derivative sequence.

[0034] The specific process for setting the benchmark stress threshold is as follows: All molding process cases marked as "qualified" are selected from the historical database, and the second derivative sequence of the prediction curve corresponding to each case is extracted. Subsequently, the system uses a kernel density estimation algorithm to fit the overall probability distribution of the absolute values ​​of the second derivatives of all qualified cases, and calculates the 99.5th percentile value of this distribution as the initial candidate value for the benchmark threshold. Finally, combining the strength dispersion coefficient of the current batch of materials and the mold wear state, linear compensation is applied to this candidate value within the range of 0.9 to 1.1 to generate the final dynamic benchmark stress threshold for real-time comparison. During the identification process, the system establishes a sliding detection window with a length of 5 time steps, calculating the absolute value of the second derivative of each point within the window in real time. If the absolute values ​​of more than 80% of the data points (i.e., at least 4 points) within a window exceed the dynamic benchmark stress threshold, a mismatch is determined to exist near the center point of the window, and the point with the largest absolute value of the second derivative within the window is identified as the mismatch peak point. Simultaneously, the integer position number of this point in the original stress input sequence is output. This output integer position number is the aforementioned time index.

[0035] Next, when retrieving the node displacement data of the preset finite element mesh model based on the obtained time index to locate the mesh node that caused the mismatch peak and obtain its real-time position coordinates, the system uses the time index as a key query condition. The specific operation process is as follows: The system first converts the time index into a specific time step file path in the simulation result database, and directly reads the displacement field data of all nodes in the predefined "root monitoring region" under that time step through an efficient data I / O interface (such as the HDF5 library). Subsequently, the system uses a node displacement gradient algorithm based on spatial adjacency to locate the node: for each node in the region, the Euclidean distance between its displacement vector and the average displacement vector of all its directly adjacent nodes is calculated. This distance is the local displacement mutation intensity value. The system traverses all nodes and finds the node with the largest mutation intensity value, determining it as the key location where material flow is most significantly hindered and most likely to cause stress mismatch. By traversing the displacement field data of all nodes in the monitoring area at that time step, the system finds the node where the displacement vector changes abruptly or differs the most from the adjacent nodes. This node usually corresponds to the key location where material flow is obstructed or stress is released. Once the node is locked, the system directly reads its three-dimensional coordinates in the global coordinate system from its attribute data, that is, its real-time position coordinates.

[0036] In step S14, the strain gradient of the transition zone of the flange sleeve wall is calculated based on the real-time position coordinates, and the alternating state of tensile and compressive stress is determined based on the strain gradient. The adjustment increment required for stroke compensation is calculated based on the alternating state of tensile and compressive stress, including: Based on the real-time position coordinates, locate the finite element mesh nodes in the transition zone of the flange sleeve wall, and calculate the strain gradient tensor field of the finite element mesh nodes. The strain gradient field is input into a preset material constitutive model to calculate the dynamic force distribution and generate a scalar sequence coupled with the dynamic force distribution and the strain gradient tensor field. Analyzing the scalar sequence, if there is an alternating state of tensile and compressive stress at the node of the transition zone of the cylinder wall, the adjustment increment required for stroke compensation is calculated based on the amplitude of the alternating state.

[0037] It should be noted that when locating the finite element mesh nodes of the flange sleeve wall transition zone based on the real-time position coordinates and calculating the strain gradient tensor field, the system first performs a spherical range query based on three-dimensional spatial distance in the preset finite element model database, using these coordinates as the center. Specifically, the system compares the real-time position coordinates with the coordinates of all nodes in the database according to the model's global coordinate system. Using a preset R-tree spatial index, it quickly filters out all nodes whose Euclidean distance from the center point is within 5 mm, forming a local node set (i.e., a node cluster). Then, based on the topological connection relationship of the finite element mesh, the system constructs a "neighborhood node set" (usually composed of nodes directly connected by element edges) for each node in the node cluster. Next, the system retrieves the displacement vector data of all nodes in the node cluster from the deformation process database at each time step of the entire cold heading process. For each node, the system uses a central difference scheme to calculate its displacement gradient tensor: using the coordinate differences and displacement differences between the node and its neighboring nodes along the three coordinate axes, a least-squares problem is solved to obtain the rate of change of the node's displacement field in the three spatial directions, i.e., the displacement gradient tensor. Finally, based on the small deformation geometry equations, the linearized strain tensor is calculated from the displacement gradient tensor. The specific calculation process is as follows: the system reads the obtained displacement gradient tensor F. According to the definition of the Cauchy strain tensor, the strain tensor ε is equal to half the sum of the displacement gradient tensor and its transpose tensor, i.e. The system decomposes this tensor operation into component calculations, and for each node, calculates the six independent components of its strain tensor. , , , , , The normal strain component is obtained by directly substituting the corresponding diagonal elements of the displacement gradient into the formula, while the shear strain component is calculated by half the sum of the corresponding displacement gradient cross terms. Then, the spatial gradient of the strain tensor is calculated to obtain the strain gradient tensor field describing the non-uniformity and directionality of deformation in this local region.

[0038] When the strain gradient tensor field is input into a preset material constitutive model to calculate the dynamic force distribution and generate a coupled scalar sequence, the preset constitutive model of the system is the Johnson-Cook dynamic constitutive model. The material constitutive model parameters used by the system are obtained by systematically calibrating the high-strength steel material through mechanical experiments: including obtaining the strain hardening coefficient and reference yield stress through quasi-static uniaxial tensile tests, obtaining the strain rate sensitivity coefficient through Hopkinson bar impact tests at different strain rates, and obtaining the thermal softening coefficient through compression tests combined with temperature rise measurements.

[0039] In actual calculations, the system executes an independent calculation process for each node in the node cluster: First, it extracts the strain gradient tensor components of the node from the strain gradient tensor field, and reads the equivalent plastic strain value of the current deformation step from the process data. Based on the accumulated plastic deformation work of the node, it estimates its local temperature rise in combination with the material's specific heat capacity and thermal conductivity. Then, it substitutes the strain gradient components, equivalent plastic strain, equivalent strain rate derived from the strain gradient, and estimated temperature rise into the Johnson-Cook constitutive equation for iterative solution until the convergence condition is met. Finally, it outputs the Cauchy stress tensor of the node, completing the calculation of the dynamic stress distribution.

[0040] After obtaining the Cauchy stress tensor (i.e., the dynamic stress distribution vector) for all nodes, the system performs a tensor double dot product operation on each node: the Cauchy stress tensor of the node is condensed with its corresponding strain gradient tensor to obtain a scalar value, which physically characterizes the coupling strength between the stress work and the deformation gradient at that point; finally, the system arranges the scalar values ​​calculated by all nodes in the node cluster at the current deformation time step in ascending order of node ID, forming an ordered array, which is the scalar sequence for the current time step; as the incremental steps of the finite element simulation advance, the above process is repeated for each time step, thereby generating a set of scalar sequences arranged in chronological order.

[0041] Finally, when analyzing the scalar sequence and calculating the adjustment increment required for stroke compensation based on the amplitude of the alternating tensile and compressive stress states, the system employs a sliding time window method to perform time series analysis on the continuous scalar sequence. The specific setup process is as follows: The system first divides the entire scalar sequence into a series of sliding windows with a length of 10 time steps and an overlap of 5 time steps between adjacent windows. For the data within each window, the system performs zero-mean preprocessing, then calculates its autocorrelation function, and preliminarily assesses periodicity by identifying local minima of the autocorrelation function at lags of 1 to 3 steps. Simultaneously, the system uses a first-order difference combined with a sign detection algorithm to accurately locate all points in the sequence that cross the zero line from positive to negative or from negative to positive. If three or more consecutive complete zero-crossing oscillations are identified within a window, the system calculates the absolute difference between all adjacent peak and trough pairs within that oscillation period and takes the arithmetic mean of these differences as the average amplitude. When the average amplitude exceeds a threshold (e.g., 0.70) jointly determined by high-cycle fatigue testing of materials and historical process statistics, a significant alternation of tensile and compressive stresses is identified at the corresponding node in the cylinder wall transition zone. The system then records this state and extracts the maximum single amplitude within the oscillation cycle (i.e., the maximum value among all peak-trough pairs). When calculating the adjustment increment, the system inputs the maximum amplitude into a pre-calibrated empirical relationship, which is established based on regression analysis of historical successful compensation case data, mapping the amplitude to a dimensionless compensation coefficient. This empirical relationship adopts an exponential growth form, and its specific parameters are obtained by fitting historical data; for example, the compensation coefficient increases exponentially with the portion of the amplitude exceeding the threshold. Finally, the calculated compensation coefficient is multiplied by the current process-set baseline stroke value (e.g., the current value of the compression stroke), and the product is the axial stroke adjustment increment in millimeters.

[0042] The specific method for determining the threshold is as follows: First, a series of high-cycle fatigue tests are conducted on the standard specimens of the high-strength steel material, and its stress amplitude-life curve is plotted. The stress amplitude that causes the specimen to fail at 10^7 cycles is converted into an equivalent scalar sequence amplitude through a constitutive model, which is used as the fatigue strength threshold of the material itself. Simultaneously, the average amplitude of the scalar sequence of all qualified cases in the stable molding stage was extracted from the historical high-quality process database, its distribution was statistically analyzed, and the 99th percentile was taken as the process statistical threshold. The final system judgment threshold is taken as follows: and The smaller of the two values ​​is chosen to ensure that both the material fatigue limit and the historically stable process constraints are met simultaneously.

[0043] In step S15, the cold heading process stroke is fine-tuned and simulated according to the adjustment increment. By analyzing the simulation results, the stable position of the material flow boundary point is predicted. Based on the stable position, the uniform distribution state of the root stress is determined, including: Based on the adjustment increment, a corrected stroke sequence is generated, and the corrected stroke sequence is input into a finite element simulation to calculate the velocity vector field of the material flow. The velocity vector field is subjected to feature extraction and identification, and the coordinates of the material flow boundary point are output. Calculate the variance of the coordinates. If the variance is less than a preset variance threshold, mark the coordinates as having reached a stable position and output the stable coordinates. Based on the stable coordinates, the equivalent stress dispersion of the root region is calculated. If the equivalent stress dispersion is lower than a preset uniformity threshold, the root stress is marked as reaching a uniform distribution state.

[0044] It should be noted that when generating the corrected stroke sequence based on the adjustment increment and inputting this sequence into the finite element simulation to calculate the material flow velocity vector field, the system first performs time series alignment and interpolation processing on the adjustment increment. The adjustment increment values ​​at discrete time steps are constructed into a continuous time-increment curve using a cubic spline interpolation algorithm. Subsequently, this curve is vector-superimposed with the baseline axial stroke-time curve stored in the process parameter database to generate a new, smooth stroke control command sequence. The system first interpolates and superimposes the axial stroke adjustment increments calculated in millimeters from the previous steps onto the preset standard cold heading process stroke path according to the corresponding time steps, generating a smooth corrected stroke control curve as the new loading boundary condition. The system then submits a complete input file containing the updated stroke commands, the 3D mesh model of the mold assembly, material constitutive model parameters (such as Johnson-Cook model parameters), and contact friction coefficients to an explicit dynamic finite element solver (such as LS-DYNA) for recalculation. During the simulation, at the end of each computational increment step, in addition to outputting the standard stress-strain results, the solver also records and outputs the velocity vectors of all nodes in the flange sleeve wall transition region (a predefined set of nodes based on geometric features) at the current increment step, according to user-defined output requests. These velocity vectors contain components in the X, Y, and Z directions. The system aggregates these velocity vector data, organized by node number and increment step time index, to form a complete multidimensional dataset describing the instantaneous material flow state in the region from the start to the end of deformation, thus obtaining the required material flow velocity vector field.

[0045] Next, when calculating the variance of the coordinates and marking them as stable coordinates when the variance is less than a preset threshold, the system first uses a sliding window with a fixed length of 8 time steps to continuously collect the latest predicted 3D coordinates of the boundary point obtained by the convolutional neural network model, forming a coordinate sequence arranged in chronological order. For the coordinate data within this window, the system calculates the sample variance for each of the X, Y, and Z coordinate axis components using the variance formula of unbiased estimation. The system repeats the above boundary point coordinate prediction at multiple consecutive deformation time steps (e.g., the most recent 8 time steps) to obtain a set of 3D coordinate sequences arranged in chronological order. Then, the system calculates the numerical variance of this sequence in the X, Y, and Z coordinate axis directions. The system then makes a judgment: if the variance values ​​calculated in the X, Y, and Z directions within the current sliding window are all strictly less than the preset variance threshold, then the spatial position of the material flow boundary point is determined to have reached a statistically stable state. Subsequently, the system calculates the arithmetic mean of all coordinate points within the sliding window (i.e., the average of the coordinate values ​​in the X, Y, and Z directions), marks the coordinates of this average value, and outputs it as the final "stable coordinates." In the process of setting the preset variance threshold, firstly, "high-quality and qualified" molding cases are selected from historical data based on strict dimensional tolerance and non-destructive testing standards. Next, the coordinate sequence of the boundary points of these cases in the molding stabilization stage is extracted, and the variance in each direction is calculated. Then, the distribution of the variance values ​​of all cases is statistically analyzed, and the maximum value among the 99th percentile values ​​in the X, Y, and Z directions is taken as the unified threshold. This threshold is also periodically recalculated and updated using newly added qualified case data.

[0046] Finally, when calculating the equivalent stress dispersion in the root region based on the stable coordinates and assessing whether it has reached a uniform distribution state, the system first uses the stable coordinates as the center and leverages the topological connectivity and spatial index of the finite element mesh to quickly retrieve and filter all mesh nodes located within a spherical space with a radius of 2 mm, forming a local node set for stress uniformity assessment. Next, the system reads the complete Cauchy stress tensor of each node in this node set from the result file of the current deformation step, and solves for the equivalent stress value of each node by calculating its stress tensor invariants according to the von Mises yield criterion. Subsequently, the system calculates the arithmetic mean and standard deviation of these node equivalent stress values, and divides the standard deviation by the mean to obtain its coefficient of variation, which serves as a dimensionless index for measuring the dispersion of local stress distribution.

[0047] The process for setting the preset uniformity threshold (e.g., 0.08) is as follows: Collect all historically tested and qualified molding cases, extract the stress variation coefficient data of the same region at the root during the stable stage, statistically analyze its distribution, and take the 95th percentile as a reference benchmark. Simultaneously, calibrate this benchmark value by combining it with the allowable range of mechanical property fluctuations regarding microstructure uniformity as specified in the high-strength steel material handbook, and finally determine the threshold. If the calculated current variation coefficient is lower than this preset threshold, the stress distribution in that region is considered highly uniform.

[0048] In step S16, the uniform distribution state and the preset high-strength steel material limit are matched to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector, including: When the root stress reaches the uniform distribution state, the stress data at each location is extracted and input into the preset high-strength steel constitutive relation model to calculate and generate the stress safety margin field. The stress safety margin field is input into a preset matching neural network model to calculate a quantitative index that characterizes the degree of matching between the uniform distribution state and the preset high-strength steel ultimate state. When the quantification index meets the preset molding quality requirements, the calculated stroke adjustment sequence is obtained; The travel adjustment sequence is fitted to generate a compensation control matrix, and a gradient projection transformation is performed on the compensation control matrix to output the final travel optimization vector.

[0049] It should be noted that when the root stress reaches the aforementioned uniform distribution state, the system extracts the von Mises equivalent stress data of all nodes at the flange sleeve root and transition zone under this state from the finite element simulation results. This stress data is then input into a preset high-strength steel constitutive model, specifically the Johnson-Cook dynamic constitutive model. This model describes the flow stress behavior of the material under high-speed deformation through an equation consisting of a product of strain hardening, strain rate strengthening, and thermal softening terms. Specifically, the flow stress is equal to a value obtained by multiplying the material's static yield stress, strain hardening term, strain rate-related term, and temperature-related term. The strain hardening term reflects the material's strengthening characteristic as plastic strain increases; the strain rate-related term characterizes the dynamic effect of the material's yield stress increasing with the deformation rate; and the thermal softening term describes the softening phenomenon caused by heat generated during plastic deformation.

[0050] The five key material constants required for the model, including yield stress at the reference strain rate, strain hardening coefficient, strain hardening exponent, strain rate sensitivity coefficient, and thermal softening coefficient, have all been calibrated using data from quasi-static tensile tests, a series of Hopkinson bar impact tests, and compression tests at different temperatures for the high-strength steel. In application, the system first calculates the instantaneous dynamic yield limit of the material at each node under the current specific deformation conditions using the constitutive equations, based on the current equivalent plastic strain, equivalent strain rate, and local temperature rise estimated through energy dissipation. Subsequently, the system divides this dynamic yield limit value by the current von Mises equivalent stress value extracted from the finite element results for that node, obtaining a ratio greater than or less than 1. This ratio represents the safety margin of the node under the current processing conditions. The system performs the above calculations on all nodes within the evaluation area and arranges the safety margin values ​​of each node according to their spatial location, ultimately forming a complete stress safety margin field.

[0051] Then, when the stress safety margin field is input into a preset neural network model to calculate the matching degree quantification index, the preset model is a feedforward neural network containing an input layer, three fully connected hidden layers, and an output layer. The number of nodes in its input layer corresponds to the length of a one-dimensional vector after the two-dimensional safety margin field is meshed and flattened (e.g., 2500 nodes for a 50x50 mesh). The first hidden layer has 512 neurons, the second has 256, and the third has 128, all using the ReLU activation function to introduce nonlinearity. The output layer is a neuron using the Sigmoid activation function, ensuring that the output value falls within the range of 0 to 1. The training set of this network consists of thousands of pairs of finite element simulation and physical inspection data of historical cold heading processes. The input of each sample is the normalized safety margin field data, and the label is a normalized quality score (0-1 points) based on a comprehensive evaluation of multiple indicators such as the metallographic structure rating, fatigue life test results, and dimensional accuracy of the corresponding part. During network training, mean squared error is used as the loss function, and the Adam optimizer is employed for weight updates. The initial learning rate is set to 0.001, and a Dropout layer with a dropout rate of 0.3 and an early stopping mechanism are introduced to prevent overfitting. In the application phase, the network performs forward propagation calculations on the input safety margin field data, ultimately producing a scalar value in the output layer.

[0052] When the quantification index meets the preset molding quality requirements, the system calculates a stroke adjustment sequence. The index threshold corresponding to the molding quality requirements (e.g., 0.88) is then determined to be met. After determining that the requirements are met, the system retrieves all axial stroke adjustment increments dynamically calculated and applied at each deformation time step since the start of the current process to achieve the current uniform state from the real-time process database. These increments are arranged in chronological order to form the stroke adjustment sequence. The specific setting process of the preset index threshold (e.g., 0.88) is as follows: The system first constructs a database of historical "high-quality cases." These cases not only meet dimensional tolerance requirements, but the fatigue life test results of their corresponding parts must also be higher than the minimum value specified by the industry standard. For each case in the database, the system calls a pre-trained feedforward neural network model to calculate the corresponding quantification index based on its safety margin field data, thereby forming a high-quality case index sample set. The calibrated value is then set as a fixed index threshold for online quality judgment. This threshold will be periodically (e.g., quarterly) refitted and re-evaluated as the number of cases in the database increases.

[0053] Finally, when fitting the travel adjustment sequence to generate the compensation control matrix and output the final travel optimization vector, the system first uses polynomial least squares to fit the travel adjustment increment data in time series form. Specifically, the system attempts different polynomial models from order 3 to 7, selecting the optimal fitting order (usually order 4 or 5) by calculating the sum of squared residuals and combining it with the AIC criterion, thus obtaining a continuous and smooth function describing the increment's change over time. Subsequently, the system re-discretely samples this function at fixed time intervals synchronized with the control cycle (e.g., every 0.01 seconds) to obtain an equally spaced increment sequence. To construct the compensation control matrix, the system uses this sequence as the main diagonal element of the matrix and adds the first and second order time lag terms (i.e., the increment values ​​of the previous and two time steps) as additional columns to its right, forming a matrix where rows correspond to time steps and columns contain current and historical values. This matrix characterizes the autocorrelation and dynamic dependence of the adjustment at different time points. Finally, the system performs a gradient projection transformation on the matrix: First, it calculates the numerical gradient field along the time dimension of the matrix to identify the intensity and direction of the adjustment trend; then, it projects the gradient field onto the feasible solution space defined by the process constraints (the next stroke adjustment amount must not exceed the maximum speed and acceleration limit of the servo motor). This projection operation is achieved by solving a quadratic programming problem with linear inequality constraints, the goal of which is to minimize the change in the adjustment vector before and after the projection; the new vector obtained after the solution, each element of which is the final stroke compensation amount for each time step that satisfies all physical constraints and transitions smoothly, is the final stroke optimization vector.

[0054] In step S17, the preset deformation control model is updated according to the stroke optimization vector. The updated deformation control model is used to monitor the dynamic stress during the production process in real time and generate compensation adjustment instructions. The preset deformation control model is updated according to the stroke optimization vector, the updated deformation control model is obtained and driven by the real-time load data stream, and the dynamic force data of the flange sleeve root is output. A real-time monitoring signal is generated based on the dynamic force data, and the correction value of the servo motor pulse is calculated. The correction value is superimposed on the original control data stream of the flange sleeve production stroke to obtain a compensation adjustment command, which is then executed to achieve compensation adjustment of the flange sleeve production stroke.

[0055] It should be noted that when updating the preset deformation control model based on the stroke optimization vector, the system inputs this vector as a set of time-varying parameters into a real-time mechanical prediction model based on simplified finite element theory. This model is essentially a state-space model whose parameters can be updated in real time. Its core is the reduced-order stiffness matrix and mass matrix of the key region (including the root and transition zone) of the flange sleeve. The update process is as follows: the system updates parameters through a specially designed mapping network. This network is a multilayer perceptron with two fully connected hidden layers. The number of nodes in its input layer corresponds to the dimension of the stroke optimization vector, and the number of nodes in its output layer is consistent with the number of independent elements in the stiffness matrix that need to be adjusted. The training data for this network is constructed as follows: a large number of offline finite element simulations covering different process conditions are performed. In each simulation, a known stroke disturbance is artificially introduced and its corresponding stiffness matrix change is recorded, thus forming a large number of "stroke disturbance vector - stiffness matrix increment" sample pairs. The network uses mean squared error as the loss function and is trained using the Adam optimizer to learn the nonlinear mapping relationship from the stroke compensation amount to the scaling factor of the stiffness matrix elements. In application, the compensation amount corresponding to each discrete time point in the stroke optimization vector is mapped to a scaling factor for a specific element of the time-varying stiffness matrix in the model. These scaling factors act on the stiffness matrix in real time, which is equivalent to simulating the local hardening or softening effect of the material caused by stroke fine-tuning, thereby completing the online refresh of the model parameters.

[0056] When acquiring and utilizing the real-time load data stream to drive the updated deformation control model, the system uses high-frequency dynamic force sensors integrated on the cold heading machine punch and die holder to collect triaxial load signals in real time at a sampling rate of 2000Hz. These signals are then processed by a Kalman filter-based data fusion algorithm for noise reduction and synchronization, forming a load data stream. This data stream serves as a time-varying boundary condition and is input into the updated state-space model. The model employs an explicit Newmark integral algorithm, combined with the updated system matrix, to solve the dynamic equations in real time. It calculates and outputs the three-dimensional dynamic stress tensor, equivalent stress, and their rate of change over time at the preset monitoring node at the root of the flange sleeve under the current load and the adjusted process. This set of real-time calculated mechanical parameters constitutes the required dynamic force data.

[0057] When generating real-time monitoring signals and calculating correction values ​​for servo motor pulses based on the dynamic force data, the system first extracts a key monitoring indicator—the minimum safety margin of the root monitoring area—from the dynamic force data. The preset multi-level early warning thresholds for this indicator are determined through statistical analysis of its distribution in historical high-quality production cases: the first-level yellow warning threshold is set at the 5th percentile of historical data, and the second-level red intervention threshold is set at the 1st percentile of historical data. The system continuously checks the current minimum safety margin value in 10-millisecond cycles. If the value is lower than the first-level threshold for three consecutive cycles, a yellow warning signal is generated; if the value is lower than the second-level threshold in any cycle, a red intervention signal is immediately generated.

[0058] Simultaneously with signal generation, the system initiates the calculation process for servo motor pulse correction values. The core algorithm is a discretized incremental digital PI (proportional-integral) controller, whose input is the deviation e(t) between the target safety margin (usually set to 1.2) and the current measured minimum safety margin. The values ​​of the proportional coefficient Kp and integral coefficient Ki are obtained through extensive offline simulation and bench testing calibration, aiming to balance response speed and system stability. The integral term accumulates historical deviations, while the proportional term responds to immediate deviations. In addition, the system calculates the linear regression slope of the deviation sequence within the last 50 milliseconds, which is introduced into the calculation as a feedforward compensation term. Finally, the basic correction value and feedforward compensation value output by the PI controller are vector-synthesized with the current moment's reference compensation value interpolated from the stroke optimization vector. The synthesized total correction value is then smoothed by a ramp function based on the servo motor's maximum acceleration limit, ultimately converting it into an equivalent number of pulses (the conversion is based on the servo system's pulse equivalent, for example, 1000 pulses per millimeter displacement). This value is the correction value of the servo motor pulses within the current control cycle.

[0059] When the correction value is superimposed onto the original control data stream of the flange sleeve production stroke to achieve compensation adjustment, the specific implementation process is as follows: The system packages the calculated pulse correction value as a process data object and sends it to the motion controller of the servo motor via a real-time industrial Ethernet protocol (such as EtherCAT). The motion controller maintains a circular buffer internally, storing the pre-compiled stroke instruction sequence for several future control cycles. After receiving a new pulse correction value, the controller's core state machine is immediately triggered: First, the correction value is timestamped according to the precise system clock to ensure that it is strictly synchronized in the time domain with the currently executing original instruction sequence; then, a vector superposition algorithm is used to incrementally superimpose the correction value onto the original position instruction at the corresponding time point in the buffer. To ensure that the superimposed trajectory is smooth and meets kinematic constraints (such as continuous velocity, acceleration, and jerk), the controller calls the built-in cubic spline interpolator to perform real-time interpolation on the superimposed discrete instruction points, generating a new, high-order continuous position-time reference curve. Finally, the reference curve is input into the three-loop control loop of the servo driver, consisting of the position loop, speed loop, and current loop. Through space vector pulse width modulation technology, the servo motor is driven to perform precise angular displacement, thereby achieving minute and continuous online compensation and adjustment of the cold heading machine slider stroke trajectory.

[0060] Reference Figure 2 This invention provides a cold heading forming control system for automotive body flange sleeves, comprising: The data acquisition module is used to acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset conditions. The model iteration module is used to extract the stress distribution vector at the root of the flange sleeve according to the offset trend, construct a stress model based on the stress distribution vector and perform iterative correction, and apply the corrected stress model to calculate and obtain the stress evolution sequence. The coordinate positioning module is used to perform time-series analysis on the stress evolution sequence, obtain the mismatch peak value, and locate the real-time position coordinates corresponding to the mismatch peak value. The compensation calculation module is used to calculate the strain gradient of the transition zone of the flange sleeve wall based on the real-time position coordinates, determine the alternating state of tensile and compressive stress based on the strain gradient, and calculate the adjustment increment required for stroke compensation based on the alternating state of tensile and compressive stress. The process simulation module is used to fine-tune the cold heading process stroke according to the adjustment increment, predict the stable position of the material flow boundary point by analyzing the simulation results, and determine the uniform distribution state of the root stress according to the stable position. The stroke optimization module is used to match the uniform distribution state with the preset high-strength steel material limit to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector. The control output module is used to update the preset deformation control model according to the stroke optimization vector, and to monitor the dynamic force of the production process in real time through the updated deformation control model, and generate compensation adjustment instructions.

[0061] It should be noted that the cold heading control system for automotive body flange sleeves provided in this embodiment of the invention is used to execute all the process steps of the cold heading control method for automotive body flange sleeves described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0062] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0063] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for controlling the cold heading of an automobile body flange sleeve, characterized in that, include: Acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset condition. Based on the offset trend, the stress distribution vector at the root of the flange sleeve is extracted. A stress model is constructed based on the stress distribution vector and iteratively corrected. The corrected stress model is then used for calculation to obtain the stress evolution sequence. A time-series analysis was performed on the stress evolution sequence to obtain the mismatch peak value, and the real-time location coordinates corresponding to the mismatch peak value were located. Based on the real-time position coordinates, the strain gradient of the transition zone of the flange sleeve wall is calculated, and the alternation state of tensile and compressive stress is determined based on the strain gradient. Based on the alternation state of tensile and compressive stress, the adjustment increment required for stroke compensation is calculated. Based on the adjustment increment, the cold heading process stroke is fine-tuned and simulated. By analyzing the simulation results, the stable position of the material flow boundary point is predicted. Based on the stable position, the uniform distribution state of the root stress is determined. The uniform distribution state and the preset high-strength steel material limit are matched to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector. Based on the travel optimization vector, the preset deformation control model is updated. Through the updated deformation control model, the dynamic stress of the production process is monitored in real time, and compensation adjustment instructions are generated.

2. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The process of acquiring cold heading deformation data, calculating the initial offset of the material based on the cold heading deformation data, classifying the initial offset, and determining the offset trend of the boundary point when the classification result meets the preset offset conditions includes: Acquire cold heading deformation data, and calculate the initial offset of the material flow boundary point based on the cold heading deformation data; The initial offset is input into a preset classification neural network model for preliminary classification to obtain classification labels and confidence levels; If the confidence level exceeds a preset offset threshold, the offset trend of the boundary point is determined.

3. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, Based on the offset trend, the stress distribution vector at the root of the flange sleeve is extracted. A stress model is constructed based on the stress distribution vector and iteratively corrected. The corrected stress model is then applied to calculate the stress evolution sequence, including: Based on the offset trend, locate the root region of the flange sleeve and extract the stress data of the node; Feature extraction is performed on the stress data to construct a stress distribution feature vector, and a mapping matrix between the stress distribution feature vector and the dynamic force relationship is established. The theoretical stress value is predicted based on the mapping matrix, and a finite element simulation is performed on the mapping matrix. If the stress deviation between the theoretical stress value and the finite element simulation result exceeds a preset stress deviation threshold, the mapping matrix is ​​updated based on the stress deviation to obtain an updated mapping matrix. The stress state of the flange sleeve root region is analyzed using the updated mapping matrix to obtain the stress evolution sequence of the root stress with the deformation stage.

4. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The step of performing time-series analysis on the stress evolution sequence to obtain the mismatch peak value and locating the real-time position coordinates corresponding to the mismatch peak value includes: The stress evolution sequence is input into a preset predictive neural network model to generate a time-series prediction curve of stress. By calculating the time-series prediction curve and comparing it with a preset benchmark stress threshold, the mismatch peak value and the time index corresponding to the mismatch peak value are identified. Based on the time index, retrieve the node displacement data of the preset finite element mesh model, locate the mesh node that generated the mismatch peak, and obtain its real-time position coordinates.

5. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The process of calculating the strain gradient of the flange sleeve wall transition zone based on the real-time position coordinates, determining the alternating state of tensile and compressive stress based on the strain gradient, and calculating the adjustment increment required for stroke compensation based on the alternating state of tensile and compressive stress includes: Based on the real-time position coordinates, locate the finite element mesh nodes in the transition zone of the flange sleeve wall, and calculate the strain gradient tensor field of the finite element mesh nodes. The strain gradient field is input into a preset material constitutive model to calculate the dynamic force distribution and generate a scalar sequence coupled with the dynamic force distribution and the strain gradient tensor field. Analyzing the scalar sequence, if there is an alternating state of tensile and compressive stress at the node of the transition zone of the cylinder wall, the adjustment increment required for stroke compensation is calculated based on the amplitude of the alternating state.

6. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The process of fine-tuning the cold heading process stroke according to the adjustment increment, predicting the stable position of the material flow boundary point by analyzing the simulation results, and determining the uniform distribution state of the root stress based on the stable position, includes: Based on the adjustment increment, a corrected stroke sequence is generated, and the corrected stroke sequence is input into a finite element simulation to calculate the velocity vector field of the material flow. The velocity vector field is subjected to feature extraction and identification, and the coordinates of the material flow boundary point are output. Calculate the variance of the coordinates. If the variance is less than a preset variance threshold, mark the coordinates as having reached a stable position and output the stable coordinates. Based on the stable coordinates, the equivalent stress dispersion of the root region is calculated. If the equivalent stress dispersion is lower than a preset uniformity threshold, the root stress is marked as reaching a uniform distribution state.

7. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The process involves matching the uniform distribution state with a preset high-strength steel material limit to obtain a matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate a final stroke optimization vector, including: When the root stress reaches the uniform distribution state, the stress data at each location is extracted and input into the preset high-strength steel constitutive relation model to calculate and generate the stress safety margin field. The stress safety margin field is input into a preset matching neural network model to calculate a quantitative index that characterizes the degree of matching between the uniform distribution state and the preset high-strength steel ultimate state. When the quantification index meets the preset molding quality requirements, the calculated stroke adjustment sequence is obtained; The travel adjustment sequence is fitted to generate a compensation control matrix, and a gradient projection transformation is performed on the compensation control matrix to output the final travel optimization vector.

8. The cold heading control method for automobile body flange sleeves according to claim 1, characterized in that, The preset deformation control model is updated based on the stroke optimization vector. The updated deformation control model is used to monitor the dynamic stress during the production process in real time and generate compensation adjustment instructions. The preset deformation control model is updated according to the stroke optimization vector, the updated deformation control model is obtained and driven by the real-time load data stream, and the dynamic force data of the flange sleeve root is output. A real-time monitoring signal is generated based on the dynamic force data, and the correction value of the servo motor pulse is calculated. The correction value is superimposed on the original control data stream of the flange sleeve production stroke to obtain a compensation adjustment command, which is then executed to achieve compensation adjustment of the flange sleeve production stroke.

9. A cold heading forming control system for an automobile body flange sleeve, characterized in that, include: The data acquisition module is used to acquire cold heading deformation data, calculate the initial offset of the material based on the cold heading deformation data, classify the initial offset, and determine the offset trend of the boundary point when the classification result meets the preset offset conditions. The model iteration module is used to extract the stress distribution vector at the root of the flange sleeve according to the offset trend, construct a stress model based on the stress distribution vector and perform iterative correction, and apply the corrected stress model to calculate and obtain the stress evolution sequence. The coordinate positioning module is used to perform time-series analysis on the stress evolution sequence, obtain the mismatch peak value, and locate the real-time position coordinates corresponding to the mismatch peak value. The compensation calculation module is used to calculate the strain gradient of the transition zone of the flange sleeve wall based on the real-time position coordinates, determine the alternating state of tensile and compressive stress based on the strain gradient, and calculate the adjustment increment required for stroke compensation based on the alternating state of tensile and compressive stress. The process simulation module is used to fine-tune the cold heading process stroke according to the adjustment increment, predict the stable position of the material flow boundary point by analyzing the simulation results, and determine the uniform distribution state of the root stress according to the stable position. The stroke optimization module is used to match the uniform distribution state with the preset high-strength steel material limit to obtain the matching degree. If the matching degree meets the preset forming quality requirements, process data is extracted from the uniform distribution state to generate the final stroke optimization vector. The control output module is used to update the preset deformation control model according to the stroke optimization vector, and to monitor the dynamic force of the production process in real time through the updated deformation control model, and generate compensation adjustment instructions.

Images