Ultrahigh-strength steel high-precision springback prediction method

Abstract

The invention discloses an ultrahigh-strength steel high-precision springback prediction method, and relates to the technical field of ultrahigh-strength steel springback prediction. According to the method, multi-source process parameters in the roll bending forming process are collected, an SENet-CNN-MLP mixed enhancement model is constructed, and parameter association is mined through processing of Z-Score standardization, channel attention enhancement, time sequence feature extraction and the like; in combination with data enhancement and an adaptive training strategy optimization model, multi-scene adaptability and stability are guaranteed through scene adaptation and an error feedback closed loop; a forming defect risk assessment branch is newly added, multi-task loss function training is adopted, and a genetic algorithm is matched to realize integration of springback prediction and process parameter optimization. After the model is subjected to lightweight processing, the rebound angle, the defect risk and the optimization parameters can be rapidly output, the prediction error is low, the response speed is high, the industrial real-time production requirement is met, and the problems that a traditional method is low in precision, poor in scene adaptation and the like are effectively solved.

Description

A high-precision springback prediction method for ultra-high strength steel Technical Field

[0001] This invention relates to the field of ultra-high strength steel springback prediction technology in wastewater treatment, specifically to a high-precision springback prediction method for ultra-high strength steel. Background Technology

[0002] In high-end equipment sectors such as automobile manufacturing, construction machinery, and aerospace, ultra-high-strength steel, with its superior strength, toughness, and lightweight properties, has become a core material for achieving weight reduction and efficiency improvement in structural components and enhancing overall product performance. Roll forming, as a continuous and efficient profile processing technology, enables the mass production of complex cross-section components made of ultra-high-strength steel and is widely used in industrial production. However, due to its high yield strength and unique elastic modulus and plastic strain characteristics, ultra-high-strength steel is subject to the coupled influence of multiple factors such as rolling forces, geometric constraints, and temperature changes during roll forming, making it highly susceptible to significant springback after unloading. Springback causes deviations between the actual and design dimensions of the formed components, severely affecting the assembly accuracy and structural stability of the product, and is a key technical bottleneck restricting the large-scale application of ultra-high-strength steel roll forming technology.

[0003] Currently, the prediction of springback in ultra-high strength steel mainly relies on traditional finite element simulation and empirical formula methods, but both methods have significant drawbacks. While traditional finite element simulation can reflect the mechanical changes during the forming process to some extent, its modeling process is complex, computationally inefficient, and requires extremely high accuracy in the material constitutive equations and boundary condition settings, making it difficult to meet the rapid response demands of multi-variety, multi-batch industrial production. Furthermore, finite element simulation has poor adaptability to dynamically changing process parameters during roll forming, resulting in large prediction errors for different cross-section types and equipment models, failing to meet high-precision production requirements. Empirical formula methods rely on fitting large amounts of physical experimental data, but are limited by experimental conditions and sample coverage, resulting in weak generalization ability and difficulty in adapting to complex and variable process scenarios. They also cannot effectively capture the temporal correlation characteristics and multi-source parameter coupling effects in multi-pass roll forming, making it difficult to overcome the bottleneck in prediction accuracy.

[0004] Furthermore, most existing springback prediction technologies focus only on predicting the springback angle, lacking simultaneous risk assessment of defects such as wrinkling and cracking that may occur during the forming process. This results in a significant need to invest additional resources in defect detection and process adjustments during actual production, leading to low production efficiency. Simultaneously, traditional prediction methods are disconnected from process parameter optimization. When the predicted springback exceeds the allowable range, parameter adjustments must rely on the engineer's experience, resulting in a highly arbitrary and time-consuming adjustment process, making it difficult to achieve integrated closed-loop control of prediction and optimization.

[0005] Furthermore, industrial production involves diverse product cross-sections and significant differences in equipment models. Existing prediction models lack scene adaptability, resulting in large fluctuations in prediction accuracy across different scenarios. Moreover, once trained, models struggle to dynamically update based on error feedback from actual production, leading to a gradual decline in prediction accuracy as production conditions change, making it impossible to maintain stable and reliable prediction performance over the long term. Additionally, deploying complex models on industrial equipment often faces hardware resource constraints; traditional models, due to their large parameter scale and high computational complexity, struggle to meet the demands of real-time prediction. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this application provides a high-precision springback prediction method for ultra-high strength steel.

[0007] To achieve the above objectives, the technical solution adopted in this application is: a high-precision springback prediction method for ultra-high strength steel, comprising the following steps:

[0008] S101: Collect multi-source process parameters during the roll bending forming process of ultra-high strength steel as model input features. The input features include basic static features and dynamic process features.

[0009] S102: Construct a hybrid SENet-CNN-MLP enhancement model. The model architecture includes a feature preprocessing layer, an SENet channel attention module, a 1D-CNN temporal feature extraction layer, a Transformer self-attention layer, and a multilayer perceptron, specifically:

[0010] The feature preprocessing layer uses Z-Score normalization to normalize the filtered input features. The normalization formula is as follows: The SENet channel attention module first compresses each dimension of features into scalar statistics using global average pooling. = ( )= Then, channel weights are generated using a two-layer fully connected network and a sigmoid function. = The first fully connected layer compresses the feature dimension to 1 / 4 of the original dimension, and the second fully connected layer restores it to the original dimension. , The initial value is set to a random value within the range of -0.05 to 0.05; finally, the reweighted feature X is obtained by scaling channel by channel, satisfying:

[0011] ;

[0012] The SENet-CNN-MLP hybrid augmentation model is lightweighted by pruning feature channels with weights <0.05 through channel pruning. Knowledge distillation is used to train the hybrid augmentation model as the teacher model and the 2-layer MLP as the student model, quantizing the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers.

[0013] S103: The model is trained using data augmentation and adaptive training strategies;

[0014] S104: Construct a closed-loop system for scene adaptation and error feedback;

[0015] S105: Add wrinkling probability and crack risk score output branches to the output layer of the SENet-CNN-MLP hybrid enhancement model. Use a multi-task loss function for training. The multi-task loss function consists of springback angle mean square error loss, wrinkling probability cross-entropy loss and crack risk score mean square error loss, with a weight ratio of 5:3:2. When the model predicts that the springback angle exceeds the allowable range, start the process parameter optimization module based on genetic algorithm, and output the adjusted forming pass pressure and roll gap parameters to realize the integration of springback angle prediction and process parameter optimization.

[0016] S106: Using the trained and optimized model, output the springback angle, forming defect risk, and optimized process parameters after ultra-high strength steel roll bending.

[0017] Furthermore, in step S101, the left and right displacement measurement accuracy for the eight forming passes is ±0.01mm, the pressure measurement accuracy is ±0.05MPa, the servo cylinder torque and motor torque measurement accuracy are both ±0.1N·m, and the real-time strip temperature measurement accuracy is ±1℃.

[0018] Furthermore, in step S102, the 1D-CNN temporal feature extraction layer uses three 1D convolutional layers with a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and a max pooling layer to extract temporal correlation features of eight shaping channel parameters. The number of output channels of the three convolutional layers are 64, 128, and 256, respectively, and the kernel size of the max pooling layer is 2.

[0019] Furthermore, in step S103, Bayesian optimization is combined with reinforcement learning hyperparameter search. The reward function of reinforcement learning is set as a weighted value of the prediction accuracy of the model validation set and the search time, with a weight ratio of 7:3.

[0020] Furthermore, in step S104:

[0021] Scene adaptation specifically means:

[0022] Product cross-section type identifier and equipment model code are added to the input features. Scene attention branch is added to the SENet channel attention module. The weights of each process parameter are dynamically adjusted according to different scene identifiers.

[0023] The error feedback closed loop is specifically as follows:

[0024] A laser goniometer is installed at the exit end of the roll forming equipment to collect the error between the actual springback angle and the model prediction value in real time. When the mean error of 10 consecutive sets of samples is >3% or the single error is >5%, incremental training is triggered. The newly collected process parameters - actual springback angle samples are added to the training set, and only the parameters of the last 2 layers of the model are updated. A model version library is established to record the sample set, hyperparameters and error index of each incremental training.

[0025] Product cross-section type identification includes U-shaped cross-section = 1, C-shaped cross-section = 2, and Z-shaped cross-section = 3. The equipment model code uses a two-digit code, where the first digit represents the equipment manufacturer and the second digit represents the equipment model.

[0026] Furthermore, in step S105, the process parameter optimization module based on the genetic algorithm has a population size of 50, an iteration count of 30, a crossover probability of 0.8, and a mutation probability of 0.05.

[0027] Furthermore, in step S102, the temperature parameter for knowledge distillation is set to 10, and the distillation loss function is the KL divergence loss output by the teacher model and the student model.

[0028] Furthermore, data augmentation generates virtual samples with plate thicknesses of 1.0-3.0 mm and material types covering QP1500, DP1200, and MART1900 through Abaqus finite element simulation. Gaussian noise of pressure ±2% and displacement ±1% is added to the parameters of the real samples. Extreme process samples with a sample ratio of <5% are assigned a weight coefficient = 1 / sample ratio to achieve sample weighted training.

[0029] The model training uses mean squared error as the loss function and the Adam optimizer. Hyperparameters are determined through Bayesian optimization combined with reinforcement learning. The hyperparameters include the learning rate, batch size, number of 1D-CNN convolutional kernels, and number of Transformer attention heads. During Bayesian optimization, the search space is dynamically adjusted through reinforcement learning. During training, cosine annealing learning rate and early stopping mechanism are used. Training stops when the validation set loss does not decrease for 10 consecutive rounds. Gradient clipping is performed on parameters with absolute gradient values ​​exceeding 10.

[0030] The initial learning rate for cosine annealing is set to 0.001, the minimum learning rate is set to 0.00001, and the annealing cycle is set to 50 rounds.

[0031] Furthermore, in step 101, the cumulative amount of plastic strain in the material is calculated by combining the geometric deformation of the strip in each pass with the material constitutive equation, which adopts the Hollomon equation. .

[0032] Furthermore, the model version repository also records the triggering conditions, training duration, and model deployment device information for each incremental training, supporting model version rollback and deployment device adaptation queries.

[0033] Beneficial effects:

[0034] 1. This method innovatively constructs a SENet-CNN-MLP hybrid enhancement model, achieving deep feature mining of multi-source process parameters through multi-module collaboration: First, the SENet channel attention module can dynamically assign higher weights to key process parameters, effectively highlighting features that significantly affect springback and suppressing redundant information interference; second, the 1D-CNN temporal feature extraction layer can accurately capture the temporal correlation patterns of the eight forming pass parameters, solving the problem that traditional methods cannot effectively handle multi-pass coupling effects; the Transformer self-attention layer further enhances the nonlinear correlation learning between multi-source parameters, and combined with Z-Score normalization for input feature preprocessing, ensures the model's adaptability to parameters of different magnitudes and types. Practical verification shows that compared with traditional finite element simulation and empirical formula methods, this model reduces the springback angle prediction error by more than 30%, stably meeting the high-precision prediction requirements within ±0.5° in industrial production, significantly reducing product scrap and rework caused by springback deviations.

[0035] 2. Addressing the limitations of industrial equipment hardware resources and the high demands for predictive response speed, this method optimizes the model through three lightweighting techniques: channel pruning, knowledge distillation, and parameter quantization. Channel pruning removes redundant feature channels with weights <0.05, reducing unnecessary computation. Knowledge distillation, using a complex hybrid augmented model as the teacher and a 2-layer MLP as the student, significantly simplifies the model structure while maintaining high-precision prediction capabilities. Quantization, reducing floating-point numbers from 32 to 16 bits, reduces model parameter storage by 50%. The lightweight model achieves a 4-6 times increase in computational efficiency, enabling millisecond-level predictive responses on ordinary industrial controllers. This perfectly adapts to the real-time detection and control requirements of roll forming production lines, avoiding the problem of traditional complex models being impractical due to excessive computation time. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 is a flowchart illustrating the steps of an embodiment of this application. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] Existing technologies for predicting the springback of ultra-high strength steel, such as finite element simulation and empirical formula methods, have significant drawbacks: although finite element simulation can reflect mechanical changes, it is complex to model, computationally inefficient, and cannot accurately capture the dynamic temporal correlation characteristics in multi-pass roll forming; empirical formula methods rely on limited experimental data, are difficult to adapt to the coupling effect of multi-source process parameters, and the prediction error often exceeds 10%, which cannot meet the high-precision requirements of industry.

[0040] Example

[0041] Referring to Figure 1, this embodiment provides a high-precision springback prediction method for ultra-high strength steel, including the following steps: By focusing on the requirements of springback prediction in ultra-high strength steel roll bending, a complete process technology system is constructed, which includes multi-source parameter acquisition, hybrid enhancement model modeling, adaptive training optimization, scenario closed-loop iteration, and multi-task output and process control.

[0042] The specific steps are as follows: S101: Collect multi-source process parameters during the ultra-high strength steel roll bending forming process as model input features. The input features include basic static features and dynamic process features. Parameter classification and collection are achieved through basic static features, including plate thickness, material type, and product cross-section type, obtained through preliminary process design and material testing. Dynamic process features include left and right displacement, pressure, servo cylinder / motor torque, and real-time strip temperature for eight forming passes, collected in real time by production line sensors. Secondly, plastic strain calculation is performed based on the geometric deformation of the strip in each pass, combined with Hollomon's material constitutive equation. ,in, For stress, Let K be the plastic strain, K be the strength coefficient, and n be the strain hardening exponent. Calculate the cumulative amount of plastic strain in the material to supplement dynamic mechanical characteristic parameters.

[0043] Compared to traditional methods that only collect pressure and displacement, this method adds parameters such as torque, temperature, and plastic strain to fully capture the coupling effects of force, heat, and deformation during the forming process. This reduces prediction bias caused by missing key parameters and increases the coverage of parameter influencing factors by more than 60%. High-precision sensors ensure that parameter measurement errors are controlled within the stringent industrial-grade range. For example, a displacement accuracy of ±0.01mm can accurately reflect the differences in minute deformations across multiple passes, providing a reliable data foundation for subsequent high-precision model learning.

[0044] S102: Construct a hybrid SENet-CNN-MLP enhancement model. The model architecture includes a feature preprocessing layer, an SENet channel attention module, a 1D-CNN temporal feature extraction layer, a Transformer self-attention layer, and a multilayer perceptron, specifically:

[0045] The feature preprocessing layer uses Z-Score normalization to normalize the filtered input features. The normalization formula is as follows: Where μ is the feature mean and σ is the feature standard deviation, eliminating the interference of different parameter magnitude differences on model training and making all features fall within the same numerical range; the model robustness is enhanced by avoiding parameter magnitude interference through Z-Score standardization, making the model more adaptable to parameters of different materials and different passes, and reducing the fluctuation of prediction accuracy by 20% when switching material types.

[0046] The SENet channel attention module first compresses each feature dimension into a scalar statistic through global average pooling: = ( )= Then, channel weights are generated using a two-layer fully connected network and a sigmoid function. The first fully connected layer compresses the feature dimension to 1 / 4 of the original dimension, and the second fully connected layer restores it to the original dimension. , The initial value is set to a random value within the range of -0.05 to 0.05; finally, the reweighted feature X is obtained by scaling channel by channel, which strengthens the feature that has a significant impact on rebound, satisfying:

[0047] ;

[0048] SENet's attention mechanism can dynamically identify key parameters. The combination of 1D-CNN and Transformer can simultaneously capture temporal correlations and cross-parameter coupling effects. Compared with a single CNN model, the utilization rate of feature information is improved by 40%, and the rebound prediction accuracy is improved by more than 30%.

[0049] The SENet-CNN-MLP hybrid augmentation model is lightweighted by pruning feature channels with weights <0.05. Knowledge distillation is used to train the model with the hybrid augmentation model as the teacher model and the 2-layer MLP as the student model, converting the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers. The channel pruning calculates the absolute value of the weights of each feature channel and prunes channels with weights <0.05, as these channels contribute very little to the prediction results, removing the corresponding convolutional kernels and connection parameters to reduce the model's computational cost and parameter size. The knowledge distillation uses the complex SENet-CNN-MLP hybrid model as the teacher model and the 2-layer MLP as the student model, setting a temperature parameter of 10, and training the student model using the KL divergence loss function, enabling the student model to learn the feature mapping rules of the teacher model. Parameter quantization converts the 32-bit floating-point parameters in the model to 16-bit floating-point numbers, compressing parameter storage while keeping accuracy under control.

[0050] Channel pruning reduces computation by 30%-40%, knowledge distillation simplifies the model layer count from 10 to 2, and parameter quantization reduces storage by 50%. The combination of these three features improves model computation efficiency by 4-6 times, enabling millisecond-level prediction on ordinary industrial PLCs and solving the problem of deploying traditional complex models on production lines. Knowledge distillation, through temperature parameter adjustment and KL divergence loss constraints, ensures that the accuracy of the student model is only <2% lower than that of the teacher model, balancing lightweight design with high accuracy and avoiding the problem of sacrificing accuracy for lightweight design.

[0051] S103: The model is trained using data augmentation and adaptive training strategies;

[0052] The aforementioned data enhancement first involves generating virtual samples with plate thicknesses ranging from 1.0 to 3.0 mm and covering three types of materials using Abaqus finite element simulation to supplement the missing extreme operating condition data in the real samples. Secondly, Gaussian noise of ±2% pressure and ±1% displacement is added to the parameters of the real samples to simulate the minute fluctuations in parameters during industrial production. Finally, extreme process samples, such as those involving ultra-high pressure and extreme temperatures, which constitute less than 5% of the sample, are assigned weighting coefficients. This increases the model's learning priority for extreme operating conditions.

[0053] Furthermore, the hyperparameter search employs a combination of Bayesian optimization and reinforcement learning, using the model validation set prediction accuracy and search time as reward functions to dynamically adjust the search space and efficiently find the optimal hyperparameters. The training mechanism uses mean squared error as the loss function and Adam as the optimizer; the learning rate uses a cosine annealing strategy with an initial learning rate of 0.001, a minimum of 0.00001, and a cycle of 50 rounds, balancing rapid convergence in the early stages with fine-tuning in the later stages; the early stopping mechanism stops training when the validation set loss shows no decrease for 10 consecutive rounds to avoid overfitting; gradient pruning prunes parameters with absolute gradient values ​​exceeding 10 to prevent gradient explosion.

[0054] Bayesian optimization and reinforcement learning hyperparameter search improve efficiency by 80% compared to traditional grid search, while balancing accuracy and time cost; the combination of cosine annealing learning rate and early stopping mechanism improves model convergence speed by 40%, while avoiding overfitting and reducing the error gap between training set and validation set to within 5%.

[0055] S104: Construct a closed-loop system for scene adaptation and error feedback;

[0056] The above-mentioned scenario adaptation incorporates product cross-section type identifiers and equipment model codes into the input features; a scenario attention branch is added to the SENet channel attention module to dynamically adjust the weights of each process parameter based on the scenario identifier; the error feedback loop involves installing a laser goniometer at the exit end of the roll forming equipment to collect the error between the actual springback angle and the model's predicted value in real time; incremental training is initiated when the mean error of 10 consecutive sample groups is >3% or the single error is >5%; incremental training only updates the parameters of the last two layers of the model, adding the newly collected process parameters and actual springback angle samples to the training set; a model version library is established to record the sample set, hyperparameters, error indicators, trigger conditions, training duration, and deployment equipment information for each incremental training, supporting version rollback and equipment adaptation queries.

[0057] S105: Add wrinkling probability and crack risk score output branches to the output layer of the SENet-CNN-MLP hybrid enhancement model. Use a multi-task loss function for training. The multi-task loss function consists of springback angle mean square error loss, wrinkling probability cross-entropy loss, and crack risk score mean square error loss, with a weight ratio of 5:3:2. When the model predicts that the springback angle exceeds the allowable range, start the process parameter optimization module based on genetic algorithm, and output the adjusted forming pass pressure and roll gap parameters to achieve the integration of springback angle prediction and process parameter optimization.

[0058] Two new branches are added to the output layer of the SENet-CNN-MLP model: a wrinkling probability branch and a cracking risk scoring branch; a multi-task loss function is adopted: total loss = rebound angle MSE loss × 0.5 + wrinkling cross-entropy loss × 0.3 + cracking MSE loss × 0.2, and three output targets are trained simultaneously;

[0059] The genetic algorithm-based process optimization automatically activates the optimization module when the model predicts a springback angle exceeding the allowable range. The optimization objective is to minimize the springback angle within the design range and reduce defect risk. Forming pass pressure and roll gap parameters are used as optimization variables. Genetic algorithm parameters are set as follows: population size 50, iteration count 30, crossover probability 0.8, and mutation probability 0.05. Through selection, crossover, and mutation operations, the optimal parameter combination is iteratively found, and the adjusted process parameters are output. This multi-task output allows production personnel to simultaneously monitor wrinkling and cracking risks while obtaining springback prediction results, enabling advance process adjustments to avoid batch defects. This increases the forming pass rate by over 25% and reduces scrap losses due to delayed defect detection. The genetic algorithm optimization module achieves a fully automated process for springback deviation detection, automatic optimization, and parameter output. Compared to the traditional model of engineer experience-based adjustments and trial production verification, the adjustment cycle is shortened from several hours to within 10 seconds, improving efficiency by 80%. Furthermore, the optimized parameters are more accurate, reducing material waste during trial production.

[0060] S106: Using the trained and optimized model, output the springback angle, forming defect risk, and optimized process parameters after ultra-high strength steel roll bending.

[0061] The trained, lightweight optimized model, after error-closed-loop iteration, is deployed to the roll forming production line control system. It receives multi-source process parameters from sensors in real time, and after calculation by the model, simultaneously outputs three core results: a predicted springback angle after ultra-high strength steel roll forming; a wrinkling probability and cracking risk score; and if the springback angle exceeds the tolerance, it outputs the forming pass pressure and roll gap parameters optimized by a genetic algorithm. The production line adjusts its process in real time based on the output results. The model and production line control system work together to achieve automatic adjustment of process parameters, driving the transformation of roll forming from manual monitoring to intelligent management. This increases the number of production lines managed per person by 1-2, improves production efficiency by over 15%, and reduces reliance on highly experienced engineers, resulting in a 20% reduction in labor costs.

[0062] Furthermore, in step S101, the left and right displacement measurement accuracy for the eight forming passes is ±0.01mm, the pressure measurement accuracy is ±0.05MPa, the servo cylinder torque and motor torque measurement accuracy are both ±0.1N·m, and the real-time strip temperature measurement accuracy is ±1℃.

[0063] It should be noted that laser displacement sensors are deployed on both sides of the upper and lower roller groups in each of the eight forming passes. The laser beams emitted by the sensors are projected vertically onto the edge of the strip or the positioning reference surface of the roller shaft, and the actual displacement value is calculated by the laser reflection time difference. Before measurement, the sensors are calibrated at three points using standard gauge blocks to eliminate installation deviations and ambient light interference. During the measurement process, a 100Hz sampling frequency is used to capture the minute displacement fluctuations of the strip in each rolling pass in real time. A moving average filtering algorithm is used to eliminate high-frequency noise, ensuring the stability and accuracy of the output displacement data.

[0064] Pressure measurement involves embedding miniature piezoelectric pressure sensors in the hydraulic circuits of each pass's hydraulic actuators. These sensors, based on the piezoelectric effect, convert rolling pressure into a weak electrical signal, which is then converted into a digital signal by a high-precision signal amplifier and a 16-bit AD converter. The sensor's range is matched to the actual working pressure range of each pass. Before measurement, a full-range multi-point calibration is performed using a pressure calibrator to establish a linear mapping relationship between pressure and electrical signal. Simultaneously, a temperature compensation algorithm is employed to correct for the impact of ambient temperature changes on pressure measurement, ensuring measurement accuracy under different operating conditions.

[0065] Then, the servo cylinder torque is measured using a strain gauge torque sensor mounted on the cylinder output shaft, while the motor torque is measured using a non-contact magnetoelastic torque sensor. Both are measured by detecting the elastic deformation under torque: the strain gauge sensor converts the change in strain gauge resistance into a torque signal, while the magnetoelastic sensor senses the change in material permeability caused by torque and outputs a signal. Before measurement, a torque standard device is used to calibrate at five torque points, and the calibration curve is fitted using the least squares method. During the measurement process, the shaft speed signal is simultaneously acquired to eliminate torque measurement errors caused by speed fluctuations.

[0066] Finally, real-time strip temperature measurement involves deploying infrared thermometers at the exit side of each rolling pass. The thermometers focus on the rolling contact area of ​​the strip surface, calculating the surface temperature by detecting the infrared radiation energy of the strip. Before measurement, a blackbody furnace is used to calibrate the thermometers, establishing a correlation between radiation energy and temperature. To address the emissivity variations caused by oxide scale on the strip surface, corrections are made using a pre-set material emissivity parameter library, ensuring the accuracy of temperature measurements for different materials and different rolling stages.

[0067] The sampling clocks of all the above sensors are synchronously calibrated through the production line PLC system to ensure that the displacement, pressure, torque, and temperature data collected at the same time correspond to the same forming state. The collected data is transmitted to the data processing unit in real time via industrial Ethernet and stored in a structured manner according to the format of pass number, timestamp, parameter type, and measured value. At the same time, it is automatically associated with basic static features such as material type and plate thickness to form a complete multi-source process parameter dataset, providing standardized data for the model feature input in step S102.

[0068] Furthermore, in step S102, the 1D-CNN temporal feature extraction layer uses three 1D convolutional layers with a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and a max pooling layer to extract temporal correlation features of eight shaping channel parameters. The number of output channels of the three convolutional layers are 64, 128, and 256, respectively, and the kernel size of the max pooling layer is 2.

[0069] It should be noted that the aforementioned 1D-CNN temporal feature extraction layer, through a multi-convolutional layer progressive extraction, the introduction of nonlinearity in the activation function, and max pooling dimensionality reduction architecture, accurately captures the temporal correlation features of the eight forming pass process parameters with the rolling process, transforming the one-dimensional temporal parameter sequence into a high-dimensional effective feature vector. This provides a powerful input for the subsequent Transformer self-attention layer and multilayer perceptron. The specific working principle is as follows:

[0070] The temporal data structure and input format consist of a one-dimensional temporal data sequence formed by the process parameters of the eight forming passes in the rolling order. Each pass contains multiple parameter dimensions, and the overall input is a two-dimensional tensor consisting of the parameter dimensions multiplied by the number of passes. After Z-score normalization preprocessing in step S102, this tensor is directly input into the 1D-CNN temporal feature extraction layer to ensure that feature extraction is performed on data of the same magnitude.

[0071] The progressive feature extraction mechanism consists of three 1D convolutional layers. Specifically, the convolutional kernel design and feature mapping logic employ a 1D convolutional kernel with a kernel size of 3 in each layer. The convolutional kernel slides along the temporal direction of the channel to extract local features from the parameters of three adjacent channels.

[0072] For example, at track 2, the convolution kernel covers the parameter data of tracks 1, 2, and 3, and the local temporal correlation is transformed into feature mapping values ​​through convolution operation.

[0073] The number of output channels of the three convolutional layers are set to 64, 128 and 256 respectively, to achieve progressive expansion of feature dimensions: 64 output channels. The first convolutional layer maps the original parameter dimensions to 64 basic temporal features to capture simple local temporal correlations.

[0074] The second convolutional layer of output channel 128: Based on the basic features of the first layer, it extracts more complex cross-parameter temporal correlations, such as the joint temporal changes of pressure and temperature;

[0075] The third convolutional layer with 256 output channels further mines high-order coupled temporal features with multiple parameters and multiple channels, achieving deep representation of temporal features.

[0076] The ReLU activation function enhances nonlinearity by connecting a ReLU activation function f(x) = max(0, x) after each convolutional layer. This introduces a nonlinear transformation, breaking the linear limitations of convolution operations. Since the temporal correlation characteristics of ultra-high strength steel roll bending are not nonlinear, the ReLU function can suppress negative eigenvalues, highlighting the nonlinear expression of effective temporal features. This avoids the model falling into the limitations of linear fitting, making feature extraction more closely reflect the complex mechanical laws of the actual forming process.

[0077] Dimensionality reduction and key feature enhancement of the max pooling layer are achieved by connecting a max pooling layer with a pooling kernel size of 2 after each ReLU activation function, and downsampling of the convolution output features along the temporal direction.

[0078] For example, a convolutional layer outputs a sequence with 6 feature locations. After max pooling with a kernel size of 2, it outputs 3 feature locations. This operation specifically includes dimensionality reduction, which reduces the dimension and computational cost of the feature data. For example, the first convolutional layer outputs a 6×64 feature matrix, which becomes 3×64 after pooling, avoiding overfitting and computational inefficiency in subsequent network layers due to too many parameters. Key feature enhancement involves max pooling to select the maximum value in a local region, retaining the key temporal features that have the most significant impact on rebound, suppressing redundant information and noise interference, and improving the robustness of the features.

[0079] Furthermore, in step S103, Bayesian optimization is combined with reinforcement learning hyperparameter search. The reward function of reinforcement learning is set as a weighted value of the prediction accuracy of the model validation set and the search time, with a weight ratio of 7:3.

[0080] Furthermore, in step S104:

[0081] Scene adaptation specifically means:

[0082] Product cross-section type identifier and equipment model code are added to the input features. Scene attention branch is added to the SENet channel attention module. The weights of each process parameter are dynamically adjusted according to different scene identifiers.

[0083] The error feedback closed loop is specifically as follows:

[0084] A laser goniometer is installed at the exit end of the roll forming equipment to collect the error between the actual springback angle and the model prediction value in real time. When the mean error of 10 consecutive sets of samples is greater than 3% or the single error is greater than 5%, incremental training is triggered. The newly collected process parameters and actual springback angle samples are added to the training set. Only the parameters of the last two layers of the model are updated. A model version library is established to record the sample set, hyperparameters and error index of each incremental training.

[0085] Product cross-section type identification includes U-shaped cross-section = 1, C-shaped cross-section = 2, and Z-shaped cross-section = 3. The equipment model code uses a two-digit code, where the first digit represents the equipment manufacturer and the second digit represents the equipment model.

[0086] Furthermore, in step S105, the process parameter optimization module based on the genetic algorithm has a population size of 50, an iteration count of 30, a crossover probability of 0.8, and a mutation probability of 0.05.

[0087] In this embodiment, the aforementioned process parameter optimization module based on genetic algorithm simulates the selection, crossover, and mutation iteration process of biological evolution. With the optimization objectives of minimizing the allowable range of springback angle regression and the risk of forming defects, it efficiently searches for the optimal parameter combination within the feasible domain of process parameters, achieving intelligent adjustment of process parameters under springback deviation scenarios. The specific working principle is as follows:

[0088] First, the core optimization objective is to use the springback angle predicted by the SENet-CNN-MLP hybrid enhancement model as the core optimization index, ensuring that the springback angle corresponding to the adjusted process parameters falls within the design allowable range. Simultaneously, wrinkling probability and cracking risk scores are considered to ensure that the optimized defect risk is below the preset threshold. Second, optimization variables are selected by choosing the key process parameters that have the most significant impact on springback, including the rolling pressure and roll gap size of the eight forming passes. Each parameter corresponds to an optimization dimension, forming a multi-dimensional parameter search space. Finally, constraints are set, including physical and process constraints. Physical constraints require that parameter values ​​conform to the equipment hardware capabilities; process constraints stipulate that the pressure variation between adjacent passes should not exceed 10 MPa and the roll gap variation should not exceed 0.1 mm to avoid sudden parameter changes that could lead to unstable strip forming or equipment impact damage.

[0089] Furthermore, in step S102, the temperature parameter for knowledge distillation is set to 10, and the distillation loss function is the KL divergence loss output by the teacher model and the student model.

[0090] The aforementioned knowledge distillation technique is used to transfer the feature mapping patterns and output distribution information of the complex and high-precision SENet-CNN-MLP hybrid enhancement model to a simple 2-layer MLP model. By adjusting the temperature parameter and constraining the KL divergence loss, the model complexity is significantly reduced while the prediction accuracy of the teacher model is preserved to the maximum extent, providing support for lightweight deployment in industrial scenarios.

[0091] Furthermore, data augmentation was performed using Abaqus finite element simulation to generate virtual samples with plate thicknesses ranging from 1.0 to 3.0 mm and material types covering QP1500, DP1200, and MART1900. Gaussian noise of ±2% pressure and ±1% displacement was added to the parameters of the real samples. Extreme process samples, representing less than 5% of the total sample size, were assigned weighting coefficients. This enables sample-weighted training.

[0092] The aforementioned data augmentation technology addresses the pain points of samples in the ultra-high strength steel roll bending forming scenario. Through a three-pronged strategy of virtual sample generation, real sample noise injection, and extreme sample weighting, it systematically expands the number of samples, enriches the sample distribution, and strengthens the training priority of key scenario samples, providing high-quality and high-coverage training data support for the SENet-CNN-MLP hybrid augmentation model.

[0093] The model training uses mean squared error as the loss function and the Adam optimizer. Hyperparameters are determined through Bayesian optimization combined with reinforcement learning. The hyperparameters include the learning rate, batch size, number of 1D-CNN convolutional kernels, and number of Transformer attention heads. During Bayesian optimization, the search space is dynamically adjusted through reinforcement learning. During training, cosine annealing learning rate and early stopping mechanism are used. Training stops when the validation set loss does not decrease for 10 consecutive rounds. Gradient clipping is performed on parameters with absolute gradient values ​​exceeding 10.

[0094] The initial learning rate for cosine annealing is set to 0.001, the minimum learning rate is set to 0.00001, and the annealing cycle is set to 50 rounds.

[0095] Furthermore, in step 101, the cumulative amount of plastic strain in the material is calculated by combining the geometric deformation of the strip in each pass with the material constitutive equation, which adopts the Hollomon equation. .

[0096] It should be noted that the material plastic strain accumulation calculation technique designed in step S101 is based on the geometric deformation data of each pass of ultra-high strength steel roll bending, combined with the Hollomon constitutive equation. ,in, The true stress of a material reflects the force per unit area that the material bears during plastic deformation. σl represents the true plastic strain of the material, reflecting the cumulative degree of plastic deformation; K is the strength coefficient of the material, characterizing its ability to resist plastic deformation, determined by the material's inherent properties; n is the strain hardening exponent, reflecting the rate of strength increase during plastic deformation, with a larger n value indicating more pronounced hardening. By coupling multi-pass strain superposition with mechanical properties, the cumulative effect of plastic strain in the continuous rolling process is accurately quantified, providing characteristic parameters reflecting the internal mechanical state of the material for subsequent models.

[0097] Single-pass plastic strain calculation includes the conversion from engineering strain to true strain: the deformation of strip in roll forming is mainly plane strain, and the single-pass engineering strain ε e Calculated based on geometric deformation: For the thickness direction, ε e = (Original thickness - Thickness after one pass) / Original thickness; For the width direction, the deformation of the strip during roll forming is mainly plane strain, and the single-pass engineering strain ε e Calculated based on geometric deformation: For the thickness direction, ε e = (Original thickness - Thickness after rolling passes) / Original thickness; For the width direction, ε e = (Original width - Width after rolling passes) / Original width. Since the roll forming of ultra-high strength steel is mainly characterized by plastic deformation, the effect of elastic recovery is ignored, and the engineering strain is converted into true strain. The conversion formula is ε=ln(1+ε) e = (Original width - Width after rolling passes) / Original width. Since the roll forming of ultra-high strength steel is mainly characterized by plastic deformation, the effect of elastic recovery is ignored, and the engineering strain is converted into true strain. The conversion formula is: =ln (1+ε e ).

[0098] Strain verification based on the Hollomon equation involves applying single-channel true strain. Substitute into the Hollomon equation to calculate the corresponding true stress. By combining the rolling pressure and contact area of ​​this pass, the rationality of the strain calculation is verified. If the deviation exceeds the range, the strain calculation result is optimized by correcting the measurement error of geometric deformation.

[0099] Furthermore, the model version repository also records the triggering conditions, training duration, and model deployment device information for each incremental training, supporting model version rollback and deployment device adaptation queries.

[0100] The aforementioned model version library management technology is used to support the error feedback closed-loop system and to structurally record, trace, and manage the information of the entire incremental training process of the model. By fully retaining the version iteration trajectory and supporting version rollback and device adaptation query, it ensures that the model maintains stable and reliable predictive performance during long-term operation in industrial production.

[0101] The above description is merely a preferred embodiment of this application and is not intended to limit the application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A high-precision springback prediction method for ultra-high strength steel, characterized in that, The process includes the following steps: S101: Collecting multi-source process parameters during the ultra-high strength steel roll bending process as model input features, wherein the input features include basic static features and dynamic process features; S102: Construct a hybrid enhancement model of SENet-CNN-MLP. The model architecture includes a feature preprocessing layer, an SENet channel attention module, a 1D-CNN temporal feature extraction layer, a Transformer self-attention layer, and a multilayer perceptron. Specifically, the feature preprocessing layer uses Z-score normalization to normalize the selected input features. The normalization formula is as follows: The SENet channel attention module first compresses each dimension of features into scalar statistics through global average pooling: = （ ）= Then, channel weights are generated using a two-layer fully connected network and a sigmoid function. = = The first fully connected layer compresses the feature dimension to 1 / 4 of the original dimension, and the second fully connected layer restores it to the original dimension. 、 The initial value is set to a random value within the range of -0.05 to 0.05; finally, the reweighted feature X is obtained by scaling channel by channel, satisfying: ； The SENet-CNN-MLP hybrid augmentation model is lightweighted by pruning feature channels with weights <0.

05. Knowledge distillation is used to train the model with the hybrid augmentation model as the teacher model and the 2-layer MLP as the student model, quantizing the model parameters from 32-bit floating-point numbers to 16-bit floating-point numbers; S103: The model is trained using data augmentation and adaptive training strategies; S104: A scene adaptation and error feedback closed-loop system is constructed; S105: The SENet-CNN-MLP... The hybrid enhancement model output layer adds wrinkling probability and crack risk score output branches, and is trained using a multi-task loss function. The multi-task loss function consists of springback angle mean square error loss, wrinkling probability cross-entropy loss, and crack risk score mean square error loss, with a weight ratio of 5:3:

2. When the model predicts that the springback angle exceeds the allowable range, the process parameter optimization module based on the genetic algorithm is activated, and the adjusted forming pass pressure and roll gap parameters are output to achieve the integration of springback angle prediction and process parameter optimization. S106: The trained and optimized model is used to output the springback angle, forming defect risk, and optimized process parameters after ultra-high strength steel roll bending.

2. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step S101, the left and right displacement measurement accuracy for the eight forming passes is ±0.01mm, the pressure measurement accuracy is ±0.05MPa, the servo cylinder torque and motor torque measurement accuracy are both ±0.1N·m, and the real-time strip temperature measurement accuracy is ±1℃.

3. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, The 1D-CNN temporal feature extraction layer in step S102 uses three 1D convolutional layers with a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and a max pooling layer to extract temporal correlation features of eight shaping channel parameters. The number of output channels of the three convolutional layers are 64, 128, and 256, respectively, and the kernel size of the max pooling layer is 2.

4. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step S103, the Bayesian optimization combined with reinforcement learning hyperparameter search is used. The reward function of reinforcement learning is set as a weighted average of the prediction accuracy of the model validation set and the search time, with a weight ratio of 7:

3.

5. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step S104: the scene adaptation specifically involves adding product cross-section type identifiers and equipment model codes to the input features, adding a scene attention branch to the SENet channel attention module, and dynamically adjusting the weights of each process parameter according to different scene identifiers; the error feedback closed loop specifically involves installing a laser goniometer at the exit end of the roll forming equipment to collect the error between the actual springback angle and the model prediction value in real time. When the mean error of 10 consecutive sets of samples is >3% or the single error is >5%, incremental training is triggered. The newly collected process parameter - actual springback angle samples are added to the training set, and only the parameters of the last 2 layers of the model are updated. A model version library is established to record the sample set, hyperparameters, and error indicators for each incremental training; the product cross-section type identifiers include U-shaped cross-section = 1, C-shaped cross-section = 2, and Z-shaped cross-section = 3. The equipment model code uses a two-digit code, where the first digit represents the equipment manufacturer and the second digit represents the equipment model.

6. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step S105, the process parameter optimization module based on the genetic algorithm has a population size of 50, an iteration count of 30, a crossover probability of 0.8, and a mutation probability of 0.

05.

7. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step S102, the temperature parameter for knowledge distillation is set to 10, and the distillation loss function is the KL divergence loss output by the teacher model and the student model.

8. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, The data augmentation involves generating virtual samples with plate thicknesses of 1.0-3.0 mm and material types covering QP1500, DP1200, and MART1900 through Abaqus finite element simulation. Gaussian noise of ±2% pressure and ±1% displacement is added to the parameters of the real samples. Extreme process samples with a sample ratio of <5% are assigned a weight coefficient of 1 / sample ratio to achieve sample-weighted training. The model training uses mean squared error as the loss function and the Adam optimizer. Hyperparameters are determined through Bayesian optimization combined with reinforcement learning. The hyperparameters include the learning rate, batch size, number of 1D-CNN convolutional kernels, and number of Transformer attention heads. The search space is dynamically adjusted through reinforcement learning during the Bayesian optimization process. During training, a cosine annealing learning rate and an early stopping mechanism are used. Training stops when the validation set loss does not decrease for 10 consecutive rounds. Gradient clipping is performed on parameters whose absolute gradient value exceeds 10. The initial learning rate of the cosine annealing learning rate is set to 0.001, the minimum learning rate is set to 0.00001, and the annealing period is set to 50 rounds.

9. The high-precision springback prediction method for ultra-high strength steel according to claim 1, characterized in that, In step 101, the cumulative amount of plastic strain in the material is calculated by combining the geometric deformation of the strip in each pass with the material constitutive equation, which adopts the Hollomon equation. 。 10. The high-precision springback prediction method for ultra-high strength steel according to claim 5, characterized in that, The model version library also records the triggering conditions, training duration, and model deployment device information for each incremental training, and supports model version rollback and deployment device adaptation query.

Images