Data-driven battery 1060 aluminum foil rolling plate shape prediction and optimization method and system

Abstract

This invention relates to the field of intelligent manufacturing technology, and in particular to a data-driven method and system for predicting and optimizing the shape of 1060 aluminum foil rolled for batteries. The method includes the following steps: S101: Real-time acquisition of multi-source rolling feature data and mill status data; S102: Data cleaning and temporal feature extraction; S103: Accurate shape prediction using a deep learning model; S104: Calculation of the residual matrix between the predicted shape and the target shape; S105: Determination of the process threshold for the residual matrix; S106: Construction of a multi-objective optimization function for shape flatness and energy efficiency; S107: Improved genetic algorithm for multi-parameter global optimization; S108: Feedforward compensation control and online adaptive model update. This invention can accurately capture nonlinear features that are difficult to quantify, such as the dynamic changes in roll thermal crown, nonlinear deformation of metal, and micron-level fluctuations in oil film thickness during the rolling process of ultra-thin aluminum foil. Simultaneously, by combining the fusion analysis of multi-source sensor data and mill status data, it achieves temporal and accurate prediction of the shape, thereby improving the accuracy of shape prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a data-driven method and system for predicting and optimizing the shape of 1060 aluminum foil rolled into battery plates. Background Technology

[0002] With the rapid iteration of the new energy vehicle industry and the continuous upgrading of power battery technology, the market's performance requirements for aluminum foil used in power battery current collectors are becoming increasingly stringent. In particular, 1060 grade aluminum foil needs to be rolled multiple times to an extremely thin specification of 10-15μm, with thickness variations controlled within the micrometer level and plate flatness within ±2I-Unit. As a core component for internal current conduction in power batteries, 1060 aluminum foil with plate defects is prone to wrinkles and breaks during battery packaging and use, directly leading to poor battery contact, increased internal resistance, and seriously affecting battery performance and safety stability.

[0003] Currently, the shape control in the rolling of 1060 aluminum foil mainly relies on classical mechanistic models such as shape equations for description and regulation. However, in the rolling process of ultra-thin strip, there are many complex and difficult-to-quantify physical factors. Specifically, the thermal expansion of the rolls causes dynamic changes in the roll thermal crown, the thickness of the rolling oil film fluctuates at the micrometer level, and the metal material exhibits nonlinear deformation characteristics at extremely thin specifications. These factors are coupled with each other, making it difficult for traditional physical mechanism models to accurately quantify and describe the actual rolling process. The model prediction accuracy is significantly reduced, failing to meet the shape control requirements of ultra-thin aluminum foil.

[0004] Furthermore, in actual production control, traditional PID feedback control relies on the detection results of the outlet side shape gauge for lag adjustment, resulting in a time and space lag of approximately 0.5-2 seconds. This control method is difficult to effectively suppress periodic edge waviness, center waviness, and other shape defects caused by factors such as changes in thermal crown and oil film fluctuations, leading to a high proportion of out-of-tolerance shape defects. This necessitates frequent intervention and adjustments based on operator experience, affecting the continuity and stability of rolling production.

[0005] Therefore, there is an urgent need to design a method that can accurately capture the nonlinear characteristics during the rolling of ultra-thin aluminum foil and achieve accurate prediction and optimized control of the sheet shape in order to solve the above-mentioned problems in the existing technology. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a data-driven method and system for predicting and optimizing the shape of 1060 aluminum foil rolled into battery plates.

[0007] The technical solution adopted in this invention is as follows: a data-driven method for predicting and optimizing the rolling profile of 1060 aluminum foil for batteries, applied to the process of rolling 1060 aluminum foil to extremely thin specifications. Currently, in the field of power batteries, 10-20μm is a widely accepted conventional extremely thin specification. This invention is applicable to this range, and is particularly applicable to rolling 1060 aluminum foil to an extremely thin specification of 10-15μm. Of course, it is also applicable to thinner specifications, such as 4.5-9μm. The method includes the following steps: S101: Real-time acquisition of multi-source rolling characteristic data and mill status data: real-time acquisition of multi-source rolling characteristic data during the rolling process of 1060 aluminum foil for batteries at a millisecond-level sampling frequency, and simultaneous acquisition of real-time mill status data; the multi-source rolling characteristic data includes temperature cloud map of the roll system, and the real-time mill status data includes rolling oil film thickness monitoring data; S102: Data cleaning and temporal feature extraction: The multi-source rolling feature data is cleaned by noise reduction and normalization, and then temporal features are extracted based on the dynamic continuity of the rolling process to obtain a standardized feature vector. S103: Deep Learning Model for Accurate Plate Shape Prediction: The standardized feature vector is fused with the real-time status data of the rolling mill and input into the deep learning model to predict the plate shape values ​​of each transverse segment of the 1060 aluminum foil within a future time period, and outputs a plate shape deviation cloud map; the deep learning model is an LSTM network with an integrated attention mechanism, which assigns high weights to the key influencing factor data in the rolling process to capture the nonlinear characteristics of the dynamic changes in roll thermal crown and the micron-level fluctuations in oil film thickness during the rolling process; the key influencing factor data in the rolling process includes rolling force, roll temperature, and tension; S104: Calculate the residual matrix between the predicted plate shape and the target plate shape: Compare the predicted plate shape with the target zero plate shape curve set by the process channel by channel, calculate the plate shape residual value of each measurement channel in the transverse direction of the aluminum foil, and form a residual matrix that reflects the overall plate shape deviation. S105: Residual matrix process threshold determination: Determine whether the maximum value of the residual matrix exceeds the preset process threshold. If it does not exceed the threshold, return to step S101 to continue collecting data. If it exceeds the threshold, execute step S106. S106: Constructing a multi-objective optimization function for plate shape flatness and energy efficiency: Combining the need to improve plate shape with the goal of controlling production energy consumption, construct a multi-objective optimization function based on plate shape flatness and energy efficiency; S107: Improved Genetic Algorithm for Multi-Parameter Global Optimization: Using the control parameters of the rolling mill actuator as optimization variables, an improved genetic algorithm is used to perform global iterative optimization on the multi-objective optimization function to obtain the optimal combination of control parameters that minimizes the objective function; the control parameters of the rolling mill actuator include bending roll force, roll tilt amount, and segmented cooling spray valve opening degree; S108: Feedforward Compensation Control and Online Adaptive Model Update: The optimal control parameter combination is sent to each actuator of the rolling mill in real time, and feedforward compensation control is performed to correct the plate shape deviation in advance. The actual adjusted plate shape detection results are fed back to the deep learning model, and the model weight parameters are updated through online learning algorithm. After completing one optimization, the process returns to step S101 to form a continuous closed-loop optimization control.

[0008] As a further description of the above technical solution: In step S101, the multi-source rolling characteristic data is collected by a sensor array deployed at key locations of the rolling mill. The sensor array includes a tension sensor, a pressure sensor, a roll temperature infrared sensor, and a shape meter. The collected data includes rolling force, rolling speed, aluminum foil inlet and outlet tension, roll system temperature cloud map, and residual stress distribution on the aluminum foil surface. The real-time status data of the rolling mill includes the rolling mill operating load, current action parameters of the actuator, rolling oil film thickness monitoring data, and roll wear status data.

[0009] As a further description of the above technical solution: the noise reduction process in step S102 adopts wavelet transform to filter out environmental noise and equipment vibration noise collected by the sensor; the normalization process adopts minimum-maximum normalization to map the data to the [0,1] interval; the time series feature extraction includes extracting the 5-second moving average of the time series, the rate of change of adjacent sampling times, and the first-order difference features of the data.

[0010] As a further description of the above technical solution: In step S103, the deep learning model is an LSTM network with an integrated attention mechanism. The LSTM network includes an input layer, three LSTM hidden layers, an attention feature layer, two fully connected layers, and an output layer. Specifically: each LSTM hidden layer has 128 nodes; the attention feature layer uses an additive attention mechanism; the first fully connected layer has 64 nodes, and the second layer has 32 nodes, using the ReLU activation function; the number of nodes in the output layer is consistent with the number of transverse measurement channels of the aluminum foil; the deep learning model completes the plate shape prediction based on the fused data of the current and past 5 seconds, predicts the plate shape values ​​of each transverse segment of the 1060 aluminum foil within the next second, and outputs a plate shape deviation cloud map. The plate shape values ​​are measured in the industry-standard I-Unit.

[0011] As a further description of the above technical solution: the residual matrix in step S104 is calculated as follows: the difference between the predicted plate shape value and the target plate shape value of each measurement channel is calculated, and the differences of all channels are arranged in order according to the horizontal position of the aluminum foil to obtain the residual matrix. The target plate shape is the zero plate shape curve that meets the process requirements of 1060 aluminum foil for power batteries.

[0012] As a further description of the above technical solution: the expression of the multi-objective optimization function in step S106 is: ,in, To predict the improvement in plate shape, its value is the difference between the current maximum value of the residual matrix and the predicted maximum value of the residual matrix after parameter adjustment; The energy consumption variation of the rolling mill actuator is the sum of the energy consumption fluctuations of the bending roll mechanism, the tilting roll mechanism, and the segmented cooling spray mechanism, calculated according to their respective power characteristics and operating times. The weighting coefficients are dynamically adjusted based on actual rolling process requirements using the analytic hierarchy process (AHP), and satisfy the following conditions: .

[0013] As a further description of the above technical solution: the control parameters of the mill actuator in step S107 include the bending force, the roll inclination, and the opening of the segmented cooling spray valve. Each parameter is set with a process-allowed value range as an optimization constraint. The improved genetic algorithm introduces an elite retention strategy and an adaptive crossover and mutation probability. The crossover probability and mutation probability are dynamically adjusted according to the population fitness to avoid premature convergence of the algorithm.

[0014] Further, in step S107, the control parameters of the mill actuator are in the following ranges: bending force 0-500kN, roll tilt 0-5mm, and segmented cooling spray valve opening 0-100%. The improved genetic algorithm introduces an elite retention strategy and adaptive crossover and mutation probabilities, wherein: the elite retention ratio is 10%, and the top 10% of individuals with the best fitness in each generation are directly retained to the next generation; the adaptive crossover probability and mutation probability are dynamically adjusted according to the population fitness: individuals with high fitness values ​​use a crossover probability of 0.9 and a mutation probability of 0.05; individuals with low fitness values ​​use a crossover probability of 0.6 and a mutation probability of 0.01; the lower limit of the crossover probability is 0.5, and the upper limit of the mutation probability is 0.06.

[0015] As a further description of the above technical solution: the response time of the feedforward compensation control in step S108 is less than 500ms, and the mill actuator completes the action adjustment before the actual plate shape deviation occurs; the online learning algorithm is a mini-batch gradient descent algorithm, which incrementally updates the weights of the deep learning model based on the actual plate shape feedback data to achieve adaptive evolution of the model.

[0016] Another technical solution provided by this invention is: a data-driven prediction and optimization system for the rolled shape of 1060 aluminum foil for batteries, employing the method described above, comprising: The data acquisition module is used to acquire multi-source rolling characteristic data and mill real-time status data during the rolling process of 1060 aluminum foil for batteries at a millisecond sampling frequency; the multi-source rolling characteristic data includes temperature cloud map of the roll system, and the mill real-time status data includes rolling oil film thickness monitoring data. The data preprocessing module is used to perform noise reduction, normalization and cleaning of the multi-source rolling feature data, and extract time-series features to obtain a standardized feature vector. The shape prediction module is used to fuse the standardized feature vector with the real-time status data of the rolling mill and input it into a deep learning model. The deep learning model predicts the shape values ​​of each transverse segment of the 1060 aluminum foil in the future time period based on the fused data from the current period and a past period, and outputs a shape deviation cloud map. The deep learning model is an LSTM network with an integrated attention mechanism. The attention mechanism assigns high weights to the key influencing factor data in the rolling process to capture the nonlinear characteristics of the dynamic changes in roll thermal crown and the micron-level fluctuations in oil film thickness during the rolling process. The key influencing factor data in the rolling process includes rolling force, roll temperature, and tension. The residual calculation module is used to compare the predicted plate shape with the target zero plate shape curve channel by channel, calculate the plate shape residual value of each measurement channel, and form a residual matrix. The threshold determination module is used to determine whether the maximum value of the residual matrix exceeds a preset process threshold. The multi-objective optimization module is used to construct a multi-objective optimization function based on plate flatness and energy efficiency when the maximum value of the residual matrix exceeds a threshold. It uses the control parameters of the rolling mill actuator as optimization variables and employs an improved genetic algorithm for global iterative optimization to find the optimal combination of control parameters. The control parameters of the rolling mill actuator include bending roll force, roll tilt, and segmented cooling spray valve opening. The feedforward control module is used to send the optimal control parameter combination to each actuator of the rolling mill in real time with a response time of less than 500ms, and to perform feedforward compensation control to correct the plate shape deviation in advance. The online update module is used to feed back the actual adjusted plate shape detection results to the plate shape prediction module and update the weight parameters of the deep learning model through an online learning algorithm.

[0017] The present invention has the following beneficial effects: 1. This invention employs an LSTM deep learning network with an integrated attention mechanism, which can accurately capture difficult-to-quantify nonlinear characteristics during the rolling of ultra-thin aluminum foil, such as the dynamic changes in roll thermal crown, nonlinear deformation of the metal, and micron-level fluctuations in oil film thickness. Simultaneously, by combining multi-source sensor data with mill status data fusion analysis, it achieves precise temporal prediction of sheet shape, thus improving prediction accuracy. The sheet shape deviation cloud map output by the model can intuitively and comprehensively reflect the sheet shape change trend across the entire transverse region of the aluminum foil, providing accurate and reliable decision-making basis for subsequent sheet shape control. This solves the core technical problem of insufficient description and inaccurate prediction by physical models in the rolling of ultra-thin aluminum foil.

[0018] 2. This invention innovatively employs a feedforward compensation control mechanism, using a deep learning model to predict the trend of plate shape changes in advance. The response time of the control command is less than 500ms, allowing the mill actuator to complete the coordinated adjustment of optimal parameters before the actual plate shape deviation occurs. This completely overcomes the drawbacks of traditional PID feedback control, which relies on the outlet-side plate shape gauge for detection and suffers from a 0.5-2 second time lag. This method effectively avoids periodic edge waviness, mid-wave, and warping defects caused by lag, reducing the incidence of plate shape defects in battery aluminum foil. The plate flatness is stably controlled within ±2I-Unit, fully meeting the stringent process requirements of power battery current collectors.

[0019] 3. In constructing the multi-objective optimization function, this invention incorporates plate flatness and actuator energy efficiency into a unified optimization system. By dynamically adjusting the weight coefficients using the analytic hierarchy process (AHP), it effectively avoids excessive movement of actuators such as bending rolls and spray pumps while ensuring plate quality, thereby optimizing production energy consumption and reducing the overall energy consumption of rolling production. Simultaneously, the smooth and coordinated movements of the actuators significantly reduce mechanical wear on the equipment, extend the service life of key rolling mill components, and lower equipment maintenance and replacement costs, achieving a dual improvement in product quality and production economy.

[0020] 4. This invention establishes a closed-loop control architecture encompassing multi-source data acquisition, deep feature mining, precise time-series prediction, multi-objective global optimization, feedforward compensation control, and online model evolution. After completing feedforward control, actual strip shape feedback data is used for online incremental updates of the deep learning model. This enables the model to continuously adapt to complex and changing working conditions such as roll wear, raw material composition fluctuations, ambient temperature changes, and rolling speed adjustments, achieving dynamic self-learning and adaptive evolution. This closed-loop system provides strip shape control with autonomous adjustment capabilities without human intervention, significantly improving the continuity and stability of rolling production, reducing production interruptions caused by manual intervention, and increasing production efficiency. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the structure of the deep learning model of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention

[0023] Reference Figure 1-2 The present invention provides a data-driven method for predicting and optimizing the shape of 1060 aluminum foil rolled into sheet, which is applied to the process of rolling 1060 aluminum foil to extremely thin specifications, and includes the following steps: S101: Real-time acquisition of multi-source rolling characteristic data and mill status data: A multi-dimensional sensor array is built and deployed at key locations of the mill, and with a millisecond-level sampling frequency of 10-50ms, multi-source rolling characteristic data such as rolling force, rolling speed, aluminum foil inlet and outlet tension, roll system temperature cloud map, and residual stress distribution on the aluminum foil surface are collected in real time to comprehensively reflect the instantaneous mechanical and thermodynamic dynamic changes in the rolling zone; at the same time, the mill control system synchronously acquires real-time mill status data such as mill operating load, current action parameters of actuators, rolling oil film thickness monitoring data, and roll wear status data, providing a comprehensive and accurate data source for subsequent strip shape prediction.

[0024] S102: Data Cleaning and Temporal Feature Extraction: The professional data preprocessing module performs layered processing on the collected raw data. First, wavelet transform is used for noise reduction to filter out invalid interference signals such as environmental noise and equipment vibration noise during sensor acquisition, ensuring the validity of the data. Then, min-max normalization is used to normalize the denoised data, mapping data with different dimensions and numerical ranges to the [0,1] interval, eliminating the impact of dimensional differences on model training and prediction. Targeting the dynamic continuity characteristics of the rolling process, the moving average of the time series, the rate of change of adjacent sampling times, and the first-order difference features of the data are extracted. The extracted effective temporal features are integrated into a standardized feature vector, which serves as the input data for the deep learning model.

[0025] S103: Deep Learning Model for Accurate Slab Shape Prediction: Standardized feature vectors are fused with real-time mill status data and input into a deep learning prediction engine. The core of this engine is an LSTM network with an integrated attention mechanism. The network consists of an input layer, hidden layers, attention feature layers, fully connected layers, and an output layer. The attention mechanism assigns higher weights to key factors influencing slab shape, such as rolling force and roll temperature, enhancing the model's ability to capture key features. The LSTM network effectively solves the long-term dependency problem of traditional time-series models, accurately mining temporal correlation features in the data. Based on the fused data from the current and past 5 seconds, the model predicts the slab shape value (I-Unit) of each transverse segment of 1060 aluminum foil within the next second and outputs a visualized slab shape deviation cloud map, intuitively displaying the lateral slab shape deviation distribution of the aluminum foil.

[0026] Traditional PID feedback control relies on an outlet side shape gauge for detection. From the generation of shape deviation to sensor capture, signal transmission, controller calculation, and actuator action, the entire process has a time lag of 0.5-2 seconds. This invention selects 1 second as the prediction window, which not only covers the traditional lag time range (the median of 0.5-2 seconds) but also reserves sufficient adjustment time for the actuator (response time <500ms). The aluminum foil rolling process is characterized by continuous changes; within a 1-second window, the changes are small, resulting in high prediction accuracy. However, beyond 2 seconds, the uncertainty of the changes increases, and the prediction accuracy decreases significantly.

[0027] S104: Calculate the residual matrix between the predicted plate shape and the target plate shape: The target zero plate shape curve that meets the stringent process requirements of 1060 aluminum foil for power batteries is pre-calibrated and stored in the system. The plate shape deviation cloud map predicted by the deep learning model is compared with the target zero plate shape curve channel by channel. The difference between the predicted plate shape value and the target plate shape value of each measurement channel is calculated. The differences of all channels are arranged in order according to the horizontal position of the aluminum foil to obtain a two-dimensional residual matrix that can comprehensively and quantitatively reflect the overall plate shape deviation of the aluminum foil.

[0028] S105: Residual Matrix Process Threshold Determination: The system presets a plate shape process threshold of ±2I-Unit to meet the power battery production standards. This threshold can be flexibly adjusted according to the actual product specifications and process requirements. Numerical analysis is performed on the residual matrix to determine whether its maximum value exceeds the preset process threshold: If the maximum residual value does not exceed ±2I-Unit, it indicates that the current aluminum foil plate shape meets the process requirements, and the system maintains the existing rolling control parameters and returns to step S101 to continue real-time acquisition of multi-source data; if the maximum residual value exceeds ±2I-Unit, it indicates that the aluminum foil is about to have plate shape defects, and the system immediately triggers the subsequent intelligent parameter optimization process.

[0029] S106: Constructing a Multi-Objective Optimization Function Based on Plate Flatness and Energy Efficiency: To balance the dual objectives of plate flatness optimization and production energy consumption control, a multi-objective optimization function based on plate flatness and energy efficiency is constructed. ,in To predict the improvement in plate shape, we need to characterize the degree of reduction in plate shape deviation after parameter adjustment. This represents the change in energy consumption of the rolling mill actuators, characterizing the energy consumption fluctuations after adjusting the parameters of actuators such as bending rolls, tilting, and spraying. The weighting coefficients are dynamically adjusted based on actual rolling process requirements using the analytic hierarchy process (AHP), and satisfy the following conditions: When pursuing ultra-high plate quality, the size can be increased. In pursuing energy-saving production, it can increase This ensures that the objective function can accurately match the needs of different production conditions.

[0030] S107: Improved Genetic Algorithm for Multi-Parameter Global Optimization: Using bending roll force, roll tilt, and segmented cooling spray valve opening as core optimization variables, and based on the performance of the rolling mill equipment and rolling process requirements, allowable value ranges are set for each optimization variable as optimization constraints. The constructed multi-objective optimization function is used as the optimization objective, and an improved genetic algorithm is employed for global iterative optimization through a multi-objective optimization scheduler. This improved genetic algorithm introduces an elite retention strategy and an adaptive crossover mutation probability. The elite retention strategy directly preserves the best individuals from each generation to the next, ensuring optimization accuracy. The adaptive crossover mutation probability is dynamically adjusted according to the population fitness to avoid premature convergence. The number of iterations is set to 100-200 to balance optimization accuracy and efficiency, ultimately solving for the objective function. Minimize the optimal combination of control parameters.

[0031] S108: Feedforward Compensation Control and Online Adaptive Model Update: The optimal control parameter combination obtained from the solution is sent to each actuator of the rolling mill in real time with a response time of less than 500ms, including the bending roll mechanism, the roll tilting mechanism, and the segmented cooling spray mechanism. Each actuator quickly and collaboratively completes parameter adjustment according to the instructions. Since this adjustment is completed before the actual occurrence of the shape deviation, feedforward compensation control of the shape is realized, avoiding the generation of shape defects from the root. At the same time, the aluminum foil shape result actually detected by the shape meter after the actuator adjustment is used as feedback data and input into the deep learning prediction engine. Through the online learning algorithm of small batch gradient descent, the model weights are incrementally updated based on the feedback data to realize the adaptive evolution of the model, enabling the model to continuously adapt to complex and variable rolling conditions such as roll wear, raw material performance fluctuations, and ambient temperature changes. After the model parameter update is completed, the system returns to step S101 to continue collecting multi-source data, forming a full-process, continuous closed-loop optimization control of acquisition, prediction, judgment, optimization, control, and update.

[0032] Specifically: Step S101: Real-time acquisition of multi-source rolling characteristic data and mill status data: Sensor array arrangement: Tension sensors are installed at the aluminum foil inlet and outlet rolls to collect the inlet and outlet tension of the aluminum foil; pressure sensors are installed at the roll bearing housing to collect real-time changes in rolling force; roll temperature infrared sensors are installed on the side of the rolls without contact with the roll surface to collect temperature cloud maps of the roll system; and a shape meter is installed on the mill outlet side, with the distance from the rolling deformation zone controlled within 1.5m, to collect the residual stress distribution and current shape value of the aluminum foil surface.

[0033] Data acquisition requirements: The sampling frequency of all sensors is uniformly set to 10-50ms, and can be flexibly adjusted according to the rolling speed. The higher the rolling speed, the higher the sampling frequency, to ensure that the acquired data can accurately reflect the instantaneous mechanical and thermodynamic dynamic changes in the rolling zone. The acquired multi-source rolling characteristic data include rolling force (0-2000kN), rolling speed (0-1500m / min), aluminum foil inlet and outlet tension (0-50kN), roll system temperature cloud map (20-150℃), and residual stress distribution on the aluminum foil surface (0-100MPa).

[0034] Mill status data acquisition: The mill's PLC control system synchronously acquires mill operating load (0-100%), current operating parameters of the bending roll mechanism, tilting mechanism, and spraying mechanism, as well as rolling oil film thickness monitoring data (0.1-1μm) and roll wear status data. All data is timestamped and synchronized with multi-source rolling characteristic data to ensure temporal consistency. Rolling oil film thickness monitoring data is acquired in real-time by eddy current displacement sensors installed at the support roll bearings. The sensors have an accuracy of 0.1μm, a measurement range of 0-2μm, and a sampling frequency synchronized with the mill's PLC control system (10-50ms), enabling real-time capture of micron-level dynamic fluctuations in oil film thickness. The sensor output signal is converted into a 4-20mA standard signal by a preamplifier, then acquired by the PLC analog input module and converted into an oil film thickness value.

[0035] Step S102: Data cleaning and temporal feature extraction: Noise reduction: Wavelet transform is used to reduce noise in the original acquired data. The db4 wavelet is selected as the base wavelet, and the original signal is decomposed into 5 layers. The high-frequency coefficients are threshold quantized and then the signal is reconstructed, effectively filtering out invalid interference signals such as environmental noise and equipment vibration noise.

[0036] Normalization: The min-max normalization method is used to map the denoised data to the [0,1] interval. The calculation formula is as follows: ,in This is the original data. The minimum value of the data. For the maximum value of the data, The normalized data eliminates the impact of dimensional differences on model training and prediction.

[0037] Temporal Feature Extraction: Considering the dynamic continuity of the rolling process, three main categories of temporal features are extracted from the normalized data: First, a 5-second moving average, calculated using a 5-second sliding window to smooth local fluctuations in the data; second, the rate of change between adjacent sampling times, calculated as the ratio of the difference between two adjacent sampling points to time, reflecting the data's trend; and third, the first-order difference feature, eliminating the influence of data trends and highlighting the data's fluctuation characteristics. These three extracted features are then concatenated to form a standardized feature vector. The vector dimension is determined based on the number of sensor acquisition channels to ensure a comprehensive reflection of the temporal changes in the rolling process.

[0038] Step S103: Accurate prediction of plate shape using deep learning model: 1. Model Structure Design: The deep learning model is an LSTM network integrating additive attention mechanism. It adopts an end-to-end network architecture, consisting of five layers: input layer, LSTM hidden layer, attention feature layer, fully connected layer, and output layer. The functions and parameters of each layer are designed as follows: Input layer: The number of nodes is exactly the same as the dimension of the standardized feature vector. It is connected to the next layer in a fully connected manner, and the activation function is a linear activation function to achieve distortion-free transmission of input data. LSTM Hidden Layers: A three-layer cascaded structure is used, with 128 nodes in each layer. Each layer includes a forget gate, an input gate, an output gate, and a cell state. The forget gate uses the Sigmoid activation function, the input gate uses a dual activation function of Sigmoid and Tanh, and the output gate uses the Sigmoid activation function. The cell state uses element-wise addition and multiplication operations to achieve long-term dependency feature mining of time-series data. A Dropout layer is set between each layer with a Dropout rate of 0.2 to prevent model overfitting. The optimal number of nodes for each of the three LSTM hidden layers is 128. This design is customized for the feature dimensions and industrial computing scenarios of multi-source data from aluminum foil rolling. The 128-node dimension can fully cover multi-source rolling feature data and the feature dimensions after the fusion of rolling mill status data, including the time-series features of more than 20 types of monitoring data such as rolling force, roll temperature, tension, rolling speed, and oil film thickness. It can fully explore the coupling relationship between various data, such as the linkage change between rolling force and roll thermal crown, and avoid feature loss due to insufficient number of nodes. Compared with larger node numbers such as 256 and 512, the 128-node dimension significantly reduces the computing load of industrial servers while ensuring feature capture capabilities, keeping the computation time of single-frame shape prediction of the model within 10ms, meeting the real-time requirement of continuous time-series prediction within the next second. Combined with the Dropout layer (Dropout rate 0.2) setting, the 128-node dimension can effectively avoid model overfitting, enabling the model to maintain stable prediction accuracy under different rolling conditions (such as rolling speed adjustment and raw material composition fluctuation), adapting to the variability of industrial production conditions.

[0039] This embodiment compares the plate shape prediction performance of different network structures under the same working conditions:

[0040] The results show that a 3-layer, 128-node system achieves optimal prediction accuracy while maintaining real-time performance. Increasing the number of layers or nodes can slightly improve accuracy by 0.08 to 0.07, but it leads to response time exceeding the limit (>500ms), failing to meet the real-time requirements of feedforward control. Reducing the number of layers or nodes significantly decreases prediction accuracy. Therefore, the 3-layer, 128-node system represents a unique technical balance point in this invention, which cannot be directly determined by those skilled in the art through conventional optimization.

[0041] Attention Feature Layer: An additive attention mechanism is preferred, and the design is customized to address the differences in dimensionality and feature importance among multi-source data related to aluminum foil rolling. The additive attention mechanism maps the query vector (Q) and key vector (K) to the same 64-dimensional space through a fully connected layer, which can effectively solve the problem of inconsistent feature dimensions in rolling data (such as rolling force in the kN range, roll temperature in the °C range, and oil film thickness in the μm range), and avoid the weight calculation deviation caused by the difference in dimensions of dot product attention. The additive attention mechanism calculates the attention score through Tanh activation and single-node fully connected layer, which has a higher degree of differentiation in the weight of core features such as rolling force and roll temperature. It can increase the weight ratio of core features to more than 80%, accurately focus on the factors that play a decisive role in the shape, and improve the model's ability to capture nonlinear features. The computational complexity of the additive attention mechanism is moderate and has a high matching degree with the 128-node LSTM hidden layer. It will not cause the overall model computation delay due to the introduction of the attention mechanism, and ensure the real-time performance of shape prediction.

[0042] The output of the LSTM hidden layer is used as the query vector, key vector, and value vector. A fully connected layer maps the query vector and key vector to the same dimension (64-dimensional). After Tanh activation, attention weights are calculated through a single-node fully connected layer, then normalized using the Softmax function. Finally, the attention weights are weighted and summed with the value vector to enhance the weights of key influencing factors on sheet shape, such as rolling force, roll temperature, and tension, thereby improving the model's ability to capture core features. Specifically: 1. Roll temperature: Assigns the highest initial attention weight, as it is the primary cause of changes in sheet shape.

[0043] 1060 aluminum foil is a soft aluminum alloy. During ultra-thin rolling with a thickness of 10-15μm, the frictional heat generated between the rolls and the foil, as well as the heat generated by plastic deformation in the rolling zone, lead to dynamic changes in the roll system temperature, which in turn causes real-time fluctuations in the roll thermal crown. Every 0.01mm change in roll thermal crown directly results in a 0.1-0.3μm fluctuation in the transverse thickness difference of the aluminum foil, ultimately causing typical shape defects such as center waviness and edge waviness. Furthermore, the spatial distribution of the roll system temperature contour map directly determines the distribution characteristics of thermal crown, which is the root physical factor causing shape deviations. Based on this mechanism, the attention mechanism assigns the highest initial weight to the roll temperature data, ensuring that the model prioritizes capturing this core feature.

[0044] 2. Rolling force: Assigns the second highest initial attention weight, which is the core cause of local deviations in plate shape.

[0045] During the rolling of ultra-thin aluminum foil, the metal material exhibits nonlinear plastic deformation characteristics due to excessive deformation. Instantaneous fluctuations in rolling force (range 0-2000kN) directly alter the degree of metal deformation in the rolling zone. Uneven lateral distribution of rolling force leads to lateral stress imbalance in the aluminum foil, resulting in shape defects such as local warping and single-sided waviness. Simultaneously, there is a strong coupling effect between rolling force and roll thermal crown (increased rolling force exacerbates roll thermal expansion, further amplifying the shape impact of thermal crown). Based on this mechanism, the attention mechanism classifies rolling force as the second core influencing factor, assigning it the second-highest initial weight.

[0046] 3. Tension: Assigns the third highest initial attention weight, which is a key parameter for plate stress control.

[0047] The inlet and outlet tension of aluminum foil (range 0-50kN) is a core process parameter for controlling the lateral stress distribution of aluminum foil. Insufficient tension can easily lead to wrinkles during aluminum foil rolling, and uneven lateral tension distribution can exacerbate edge waviness. At the same time, tension can indirectly offset some of the shape deviations caused by roll thermal crown and rolling force fluctuations by changing the rolling deformation resistance of aluminum foil (e.g., appropriately increasing the outlet tension can suppress mid-wave defects). Based on its direct control effect on shape and its coupling mechanism with core factors, the attention mechanism assigns tension the third highest initial weight.

[0048] 4. Secondary factors: assigned low initial weight, have an indirect influence, and play no core leading role.

[0049] The effects of rolling speed, rolling oil film thickness (0.1-1μm), and roll wear on sheet shape are indirectly reflected through core factors (e.g., oil film thickness fluctuations only affect the transmission efficiency of rolling force, roll wear has a long-term, slow effect, and instantaneous sheet shape changes can be ignored). Therefore, the attention mechanism assigns low initial weights to these factors to avoid secondary factors interfering with the capture of core features and to ensure that the weight allocation is highly consistent with the physical logic of sheet shape evolution.

[0050] Fully connected layer: A two-layer cascaded structure is used, with 64 nodes in the first layer and 32 nodes in the second layer. Both layers employ the ReLU activation function to achieve non-linear mapping and dimensionality reduction of attention features, uncovering deep correlations between features. This node count configuration and the ReLU activation function are an optimal combination, ensuring accurate mapping of the core features output by the attention layer to the plate shape value. Gradient Node Count Reduction Design: From the high-dimensional features of the attention layer to the plate shape value dimension of the output layer, the number of nodes is reduced from 64 to 32 to achieve progressive feature dimensionality reduction. During the dimensionality reduction process, the effective information of the core features is preserved to the greatest extent, avoiding feature fragmentation caused by direct dimensionality reduction and ensuring the accuracy of plate shape prediction.

[0051] Adaptability of ReLU activation function: ReLU activation function can effectively enhance the nonlinear features in rolling data (such as nonlinear deformation of metal and micron-level fluctuations in oil film thickness), while solving the gradient vanishing problem in deep network training, making the feature mapping of fully connected layers more efficient, and improving the model training convergence speed by more than 30%.

[0052] The combination of 64+32 nodes has a small computational load and a high degree of matching with the computing power of industrial servers, further shortening the computation time of model prediction and meeting the response requirement of less than 500ms for mill feedforward compensation control.

[0053] Output layer: The number of nodes is consistent with the number of transverse measurement channels for the aluminum foil. The activation function is a linear activation function, and the output is the shape value (I-Unit) of each channel, realizing the quantitative output of the shape value. These input and output parameters are optimized settings to fit the industrial rolling scenario. Input time window: 5 seconds: To capture the optimal time window for dynamic changes in core factors such as roll thermal crown and rolling force, experiments have verified that the change in aluminum foil shape has a strong temporal correlation with the rolling data in the first 5 seconds. The 5-second time window can not only extract sufficient temporal dependency features, but also avoid data redundancy and increased computation caused by an excessively long window, keeping the mean square error of the model prediction within 0.01. Unit of measurement: I-Unit: This is the industry-standard unit of measurement for plate shape detection. The output plate shape value of the model is unified to I-Unit, eliminating the need for additional unit conversion. It directly matches the detection results of the plate shape instrument in the industrial field with the process threshold (±2I-Unit), achieving seamless integration between the model prediction results and the requirements of industrial production processes, and improving the model's engineering implementation capability.

[0054] The attention mechanism and network structure parameters of this invention are not selected based on general experience, but are specifically designed for the physical characteristics of the dynamic changes in the thermal crown of the rolls and the micron-level fluctuations in the oil film thickness during the ultra-thin rolling of 1060 aluminum foil.

[0055] Time constant of thermal crown: Based on heat conduction theory, the time constant of the effect of roll thermal crown on sheet shape is approximately 4-6 seconds. Therefore, selecting 5 seconds as the input time window can fully cover the main dynamic process of thermal crown change. If the window is too short (e.g., 2 seconds), the complete evolution of thermal crown cannot be captured; if the window is too long (e.g., 10 seconds), noise will be introduced and redundant calculations will be increased.

[0056] Physical allocation of attention weights: Based on the mechanical and thermodynamic mechanisms of aluminum foil rolling, this invention pre-determines the initial distribution of attention weights: roll temperature has the highest weight, followed by rolling force, and then tension. This physical mechanism-based weight initialization allows the model to focus on the correct physical quantities in the early stages of training, accelerating convergence and improving prediction accuracy. General attention mechanisms do not possess this physical prior.

[0057] 2. Model Training and Validation: Dataset Construction: Collect nearly 6 months of 1060 aluminum foil rolling production data, remove abnormal operating conditions such as machine shutdown, roll change, and raw material change, and construct a historical rolling dataset of no less than 1 million records. Each data record contains a standardized feature vector, mill status data, and the actual sheet shape value at the corresponding time. Dataset partitioning: The dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for iterative updating of model parameters, the validation set is used for hyperparameter tuning and overfitting monitoring, and the test set is used for final performance verification of the model. Training parameter settings: The Adam optimizer is used to train the model, and a learning rate decay strategy is adopted, with the learning rate decaying to 0.5 of the original value every 20 training rounds. The mean squared error (MSE) loss function is used to accurately quantify the deviation between the predicted value and the actual value. Model tuning and validation: During training, the model performance is monitored in real time using the validation set. Training is stopped when the validation set loss no longer decreases after 10 consecutive rounds to prevent overfitting. The model's hyperparameters (such as the number of LSTM nodes, Dropout rate, and learning rate) are tuned using a grid search to ensure that the mean squared error of the model's prediction on the test set is controlled within 0.01 and the overall prediction error is controlled within 5%, meeting the high-precision prediction requirements of industrial production.

[0058] 3. Data fusion processing: The standardized feature vector obtained in step S102 is combined with the real-time status data of the rolling mill for dimensional splicing and data fusion. During the fusion process, the real-time status data of the rolling mill is synchronously normalized to ensure that the data dimensions are consistent. The fused data is used as the input data of the deep learning model. The data dimension is the sum of the dimension of the standardized feature vector and the dimension of the rolling mill status data, realizing the feature fusion and complementarity of multi-source data.

[0059] 4. Plate Shape Prediction Implementation: The fused input data is fed into the trained and deployed deep learning model on the industrial server. Based on the current and past 5 seconds of continuous fused data, the model calculates the plate shape prediction results for the next second frame by frame through forward propagation. The prediction time step is consistent with the sensor sampling frequency (10ms, 20ms, 50ms), realizing continuous temporal prediction of plate shape, and obtaining the plate shape value (I-Unit) of each horizontal segment of 1060 aluminum foil at each time step in the next second.

[0060] 5. Prediction Result Output: The model output is divided into two categories. The first is the numerical plate shape prediction result, which is the plate shape value data of each measurement channel of the aluminum foil at each time step. This data is stored in the industrial database for subsequent residual matrix calculation. The second is the visualized plate shape deviation cloud map. The numerical prediction result is converted into a color cloud map through industrial visualization software. Different colors represent different plate shape deviations, which intuitively show the distribution and trend of plate shape deviation in the lateral direction of the aluminum foil, making it convenient for operators to monitor in real time.

[0061] Step S104: Calculate the residual matrix between the predicted plate shape and the target plate shape: Target plate shape calibration: Based on the process requirements of 1060 aluminum foil for power batteries, the target zero plate shape curve is calibrated in the system. This curve is the ideal curve in which the plate shape value of each measurement channel in the transverse direction of the aluminum foil is 0I-Unit. It can be flexibly adjusted according to the actual product specifications (such as width and thickness).

[0062] Residual matrix calculation: The predicted plate shape value of the model is compared with the target zero plate shape curve channel by channel. The difference between the predicted plate shape value and the target plate shape value of each measurement channel is calculated to obtain the plate shape residual value of a single channel. The plate shape residual values ​​of all measurement channels are arranged in order from left to right according to the horizontal position of the aluminum foil to form a two-dimensional residual matrix. The rows of the matrix represent the sampling time, the columns represent the horizontal measurement channels of the aluminum foil, and the matrix elements are the plate shape residual values ​​of the corresponding channels, which accurately quantifies and reflects the overall plate shape deviation state of the aluminum foil.

[0063] Step S105: Determining the process threshold for the residual matrix: Process threshold setting: The system presets the plate shape process threshold to ±2I-Unit. This threshold is determined based on the industry standard and actual production of 1060 aluminum foil for power battery current collectors. If customers have higher plate shape requirements, they can flexibly adjust the threshold range in the system.

[0064] Threshold determination implementation: Perform numerical analysis on the residual matrix, extract the maximum value in the matrix, and determine whether the maximum value exceeds the preset ±2I-Unit process threshold: If it does not exceed the threshold, it means that the current aluminum foil shape meets the process requirements, the system maintains the existing rolling control parameters, and returns to step S101 to continue real-time acquisition of multi-source data; if it exceeds the threshold, it means that the aluminum foil is about to have a shape defect, the system reminds the operator through audible and visual alarms, and immediately triggers the subsequent intelligent parameter optimization process.

[0065] Step S106: Construct a multi-objective optimization function for plate flatness and energy efficiency: Objective function design: Constructing a multi-objective optimization function ,in, To predict the improvement in plate shape, the calculation formula is the difference between the maximum value of the current residual matrix and the maximum value of the predicted residual matrix after parameter adjustment. The change in energy consumption of the rolling mill actuators is calculated by the power changes and running time of each actuator. For the weighting coefficients, satisfying . To quantify the reduction in plate shape deviation after parameter adjustment, the maximum value of the plate shape residual matrix is ​​used as the core evaluation index. This is quantitatively calculated by the difference between the maximum values ​​of the predicted residual matrix before and after optimization. A positive value indicates that the plate shape deviation decreases as the parameter is adjusted; a larger value indicates a more significant improvement in plate shape. If the candidate solution is a suboptimal solution... It could be negative, indicating an aggravated deviation in plate shape; This is a quantitative value for the real-time energy consumption fluctuation of the three main actuators—bending roll, roll tilting, and segmented cooling spray—after the control parameters are adjusted. The calculation is based on the real-time energy consumption under the current control parameters as a benchmark, and the difference between the predicted energy consumption under the candidate solution control parameters and the benchmark energy consumption. It also considers the action response time of each mechanism and the real-time nature of the rolling process, achieving quantitative calculation by summing the energy consumption changes of each individual mechanism. The value is non-negative, representing the energy consumption fluctuation range of the actuator after parameter adjustment. The smaller the value, the better the energy consumption control effect, which can reduce energy waste while optimizing the plate shape.

[0066] Weighting coefficients were determined using the analytic hierarchy process (AHP). First, a judgment matrix was constructed, using flatness and energy efficiency as two evaluation indicators. Pairwise comparisons were performed based on actual rolling process requirements to determine the relative importance of each indicator. Then, a consistency check was used to verify the rationality of the judgment matrix. Finally, the weighting coefficients were calculated. For example, in the production of high-end power battery aluminum foil, where flatness requirements are higher, a weighting coefficient of [value missing] could be used. , In conventional aluminum foil production, it is possible to take... , .

[0067] Step S107: Improve the genetic algorithm for multi-parameter global optimization: 1. Optimize variable and constraint settings: The three core optimization variables are: bending force, roll tilt, and segmented cooling spray valve opening. These are all continuous numerical variables that directly correspond to the action parameters of each actuator in the mill. Constraint Setting: Based on the performance parameters of the rolling mill equipment, rolling process requirements, and equipment safety operation specifications, strict allowable value ranges are set for each optimization variable as hard constraints for optimization. Specifically, the ranges are: roll bending force (0-500kN, step size 0.1kN), roll tilt (0-5mm, step size 0.01mm), and segmented cooling spray valve opening (0-100%, step size 0.1%). All candidate solutions during the optimization process must satisfy these value ranges; candidate solutions exceeding these ranges are directly judged as invalid solutions and eliminated. Specifically: (a) Bending roller force: 0-500kN (preferred range) This value range is determined based on three major engineering criteria: the plastic deformation resistance of 1060 aluminum foil, the bending mechanical properties of the rolls, and the effective force range for sheet shape correction. The details are as follows: Material deformation characteristics are based on the following: 1060 aluminum foil is soft pure aluminum. When rolled to an ultra-thin specification of 10-15μm, its transverse plastic deformation resistance is approximately 80-120MPa. The corresponding effective force range for roll bending force correction of the sheet shape is 50-450kN. 0-50kN is the micro-adjustment range (suitable for slight sheet shape deviations), and 450-500kN is the extreme correction range (suitable for more severe medium or edge wave defects). This full range covers all the process requirements for sheet shape correction of ultra-thin 1060 aluminum foil, and the force adjustment is linearly related to the sheet shape deviation correction, demonstrating clear process effectiveness.

[0068] The mechanical performance of the equipment is based on the following: The power battery aluminum foil special rolling mill adapted to this invention has a rated working load of 600kN for its bending roll mechanism. After reserving a safety margin of 100kN, 500kN is the maximum safe operating value of the bending roll force. This avoids equipment failures such as overload of the roll bearing seat and bending deformation of the roll caused by the bending roll force exceeding 500kN. At the same time, 0kN is the initial reset value of the bending roll mechanism, which conforms to the equipment's conventional operating specifications.

[0069] Industrial experimental verification basis: In actual 1060 aluminum foil rolling experiments, this invention found that when the bending roll force exceeds 500kN, it not only cannot further correct the plate shape deviation, but also increases the transverse thickness difference of the aluminum foil due to excessive bending of the roll; when the bending roll force is less than 0kN (reverse bending roll), it will destroy the rolling stability of the extremely thin aluminum foil, causing the aluminum foil to deviate and crack. Therefore, 0-500kN is the only optimal range of bending roll force.

[0070] (ii) Roll tilt: 0-5mm (preferred range)

[0071] This value range is determined based on the requirements for lateral thickness difference correction of ultra-thin aluminum foil, the tilting mechanical stroke of the rolling mill roll system, and the stability requirements of the contact area in the rolling zone. It represents the optimal range that balances the shape correction effect with the stability of equipment operation. Specific basis: The basis for the plate shape process correction is as follows: When 1060 aluminum foil is rolled to 10-15μm, the transverse thickness difference corresponding to the plate shape deviation (within ±2I-Unit) is 0.1-0.3μm. According to process calculation, every 0.1mm adjustment of the roll tilt can correct the transverse thickness difference of the aluminum foil by about 0.02μm. A tilt of 0-4mm can cover the correction requirements of all compliant plate shape deviations. 4-5mm is the process redundancy range (to accommodate non-conventional slight deviations caused by fluctuations in raw material composition and adjustments in rolling speed). The tilt adjustment in this range can achieve precise and linear correction of plate shape deviations without correction blind spots. The mechanical stroke of the equipment is based on the roll tilting mechanism of the special rolling mill for aluminum foil for power batteries. The maximum mechanical tilting stroke is designed to be 6mm, with 0mm as the initial horizontal position of the roll system and 5mm as the maximum safe tilting stroke. A 1mm mechanical clearance margin is reserved to avoid equipment problems such as loose bolts at the roll system connection and roll system misalignment caused by tilting amount exceeding 5mm. At the same time, it ensures the action response accuracy of the tilting mechanism (0.01mm step). Rolling stability is based on the following: When the roll tilt exceeds 5mm, the contact area between the roll and the aluminum foil in the rolling zone will be greatly reduced, resulting in uneven distribution of rolling force and causing new plate shape defects (such as single-sided waviness and local warping); when the tilt is less than 0mm (reverse tilt), it will cause abnormal force on the side of the aluminum foil entering the roll, causing cracks at the edge of the aluminum foil. Therefore, 0-5mm is the optimal range for roll tilt in terms of both process and equipment.

[0072] (iii) Segmented cooling spray valve opening: 0-100% (preferred range)

[0073] This value range is a generally accepted and optimal range in the industry, determined by considering the temperature adjustment requirements of the roll thermal crown, the fluid control characteristics of the spray system, and the temperature sensitivity of 1060 aluminum foil. Specific engineering examples are as follows: The basis for roll thermal crown correction: In the ultra-thin rolling of 1060 aluminum foil, the change in roll thermal crown is determined by the temperature distribution of the roll system. The segmented cooling spray system controls the local temperature of the roll by adjusting the spray volume of the rolling oil. The valve opening degree of 0-100% corresponds to the spray volume of 0-rated spray volume, which can achieve precise adjustment of the temperature of each section of the roll body from 0-20℃, completely covering the temperature adjustment range required for roll thermal crown correction (a roll temperature adjustment of 1-5℃ can achieve a thermal crown correction of 0.01-0.05mm, which is suitable for the thermal crown adjustment requirements corresponding to all plate shape deviations). The characteristics of the spray system equipment are based on the following: The segmented cooling spray valve for industrial rolling is a proportional regulating valve. Its effective adjustment range is 0-100%, where 0% means the valve is completely closed (no spray) and 100% means the valve is completely open (rated spray volume). Within this range, the valve opening and the spray volume have a linear proportional relationship, and the adjustment accuracy can reach 0.1%. This can meet the process requirements of fine adjustment of roll temperature and is the standard safe and effective operating range of the spray valve. Material temperature sensitivity is based on the following: The suitable rolling temperature for 1060 aluminum foil is 40-60℃. When the spray valve opening exceeds 100% (overspraying), the roll temperature will be too low, the plastic deformation capacity of the aluminum foil will decrease, and rolling cracks will occur. An opening of less than 0% has no practical engineering significance. Therefore, 0-100% is the only optimal range that balances roll thermal crown correction and aluminum foil rolling stability.

[0074] 2. Core Algorithm Improvements: Based on the traditional genetic algorithm, three core improvements are made to address the problems of premature convergence, low optimization accuracy, and poor constraint adaptability in the traditional algorithm. The specific improvements are as follows: Introducing an elite retention strategy: The top 10% of the best individuals (the individuals with the lowest fitness values) in each generation of the population are directly retained to the next generation of the population without participating in crossover or mutation operations, ensuring that the high-quality genes in the population are not destroyed and greatly improving the accuracy of optimization. The design incorporates adaptive crossover and mutation probabilities: Crossover and mutation probabilities are not fixed values, but are dynamically adjusted based on the fitness values ​​of individuals in the population. Individuals with high fitness values ​​(poor performance) are given a high crossover probability (0.9) and a high mutation probability (0.05) to accelerate the population's evolution; individuals with low fitness values ​​(good performance) are given a low crossover probability (0.6) and a low mutation probability (0.01) to reduce gene damage to high-performing individuals; at the same time, a lower limit of 0.5 for crossover probability and an upper limit of 0.06 for mutation probability are set to ensure population diversity. Add adaptive constraint verification: After population initialization, crossover, and mutation operations, add an adaptive constraint verification step to verify the variable value range of the generated individuals. Individuals that exceed the constraints are corrected by boundary truncation to adjust the variable values ​​to the boundary values ​​of the constraint range, ensuring that all individuals are valid solutions and improving the algorithm's adaptability to constraints.

[0075] 3. Algorithm Implementation Steps: The improved genetic algorithm is implemented using Python and the Scikit-learn library, and deployed on an industrial optimization scheduling server. The specific implementation steps are as follows: Population initialization: The optimization variables are encoded using real number encoding. Each individual is a set of candidate solutions for control parameters consisting of bending force, roll inclination, and spray valve opening. The population size is set to 50. An initial population that meets the constraints is randomly generated to ensure the diversity of the initial population. Fitness calculation: The value of the multi-objective optimization function constructed in step S106 is used as the fitness value of an individual. The smaller the fitness value, the better the plate shape improvement effect and energy consumption control level of the candidate solution, and the stronger the competitiveness of the individual in the population. The fitness value is calculated for each individual in the initial population to form a fitness value list. Selection operation: The selection is carried out by combining the roulette wheel selection method with the elite retention strategy. First, the top 10% of the best individuals are directly retained to the next generation. The remaining individuals are selected by the roulette wheel selection method. The reciprocal of the individual fitness value is used as the selection probability. The lower the fitness value, the higher the probability of the individual being selected. A sufficient number of individuals are selected from the parent population to form a crossover population. Crossover operation: The real number crossover method is used to cross individuals in the crossover population pairwise. Crossover points are generated according to the adaptive crossover probability. The variable values ​​after the crossover points are exchanged to generate offspring crossover individuals. The crossover process is carried out according to the pairing principle to ensure that each individual participates in the crossover operation. Mutation operation: The Gaussian mutation method is used to perform mutation operation on the population after the fusion of parent and offspring generations. The variable values ​​of individuals are perturbed by Gaussian according to the adaptive mutation probability to generate mutated offspring individuals; the perturbation amplitude is set according to the value range of the variables to ensure that the mutated individuals are as close as possible to the constraints. Constraint verification: Adaptive verification of constraints is performed on the new generation population generated after crossover and mutation, and invalid solutions are corrected by boundary truncation to obtain a valid new generation population that satisfies the constraints; Iteration Termination: Repeat the fitness calculation, selection, crossover, mutation, and constraint verification steps above to perform iterative optimization; stop iterating when the number of iterations reaches the set value or the population fitness value tends to stabilize (the average change rate of fitness value over 20 consecutive generations is <0.5%).

[0076] 4. Determination of Optimization Results: After the iteration terminates, the individual corresponding to the minimum fitness value in the final population is extracted as the optimal control parameter combination of the rolling mill actuator. This combination is the global optimal solution that can achieve the dual objectives of improving plate shape and controlling energy consumption under the current rolling conditions. The optimal control parameter combination is converted into action instructions that can be recognized by each actuator of the rolling mill and stored in the control instruction library for subsequent feedforward compensation control.

[0077] Step S108: Feedforward Compensation Control and Online Adaptive Model Update: 1. Control command issuance: The optimal control parameter combination obtained in step S107 is converted into standardized action commands for each actuator of the rolling mill. The commands are then issued in real time via industrial Ethernet to the PLC control system of the rolling mill and the frequency conversion, servo, and hydraulic control systems of each actuator with a response time of less than 500ms. Data encryption and verification mechanisms are used during command transmission to ensure the accuracy and reliability of command transmission. The command transmission error is controlled within ±0.1%.

[0078] 2. Implementation of feedforward compensation control: Actuator action adjustment: After receiving the action command, the bending roll mechanism, the roll tilting mechanism, and the segmented cooling spray mechanism quickly and collaboratively complete the parameter adjustment according to the command parameters. The action response time of each actuator is controlled within 200ms, and the overall action adjustment completion time is less than 400ms. This ensures that all adjustment actions are completed before the actual plate shape deviation occurs, realizing feedforward compensation control of plate shape and avoiding the generation of plate shape defects from the root. Action Coordination Control: The action adjustment of each actuator adopts a coordinated linkage control strategy. Based on the influence weight of each parameter on the plate shape and the action response speed, different action start times are set. The bending roller mechanism and the tilting mechanism start adjustment first, and the spraying mechanism starts 50ms later to ensure that the adjustment effects of each mechanism match each other and avoid secondary fluctuations in plate shape caused by asynchronous actions. Operational status monitoring: During the adjustment of the actuator's actions, the sensor array and the rolling mill control system monitor the operational status of each mechanism, the parameter adjustment value and the actual operating parameter in real time to ensure that the adjustment value is consistent with the command value and the parameter adjustment error is controlled within ±0.5%. If an adjustment deviation occurs, the system will immediately issue an alarm and make compensation adjustments.

[0079] 3. Multi-dimensional data feedback: After the actuators complete parameter adjustments, they enter the data feedback stage. Multi-dimensional feedback data is collected in real time at the original sampling frequency through sensor arrays, shape gauges, and the rolling mill control system, and stored in the industrial database. This provides complete and accurate data support for online model updates and process traceability. The feedback data includes: the actual shape value of the aluminum foil, the actual adjustment parameters of each actuator, the actual operating energy consumption, the real-time operating status data of the rolling mill, and multi-source characteristic data of the rolling process. All feedback data is timestamped to ensure time sequence consistency.

[0080] 4. Online Adaptive Model Update: Employing the Mini-BatchSGD online learning algorithm, the deep learning strip shape prediction model's weights are incrementally updated based on multi-dimensional feedback data. This ensures the model can continuously adapt to complex and changing rolling conditions. The specific implementation steps are as follows: Feedback dataset construction: Extract feature data (standardized feature vectors and mill status data) and actual plate shape values ​​corresponding to the model input from the feedback data to construct a small batch feedback dataset with a size of 128-256 records. The size of the dataset can be flexibly adjusted according to the actual production data volume. Prediction bias calculation: Input the feature data in the feedback dataset into the current deep learning model to obtain the model's predicted plate shape value, compare it with the actual plate shape value, calculate the prediction bias, and use it as a loss signal for model update; Incremental weight update: The mini-batch gradient descent algorithm is used, with the prediction bias as the loss function, to incrementally update the model's weights and biases. After each update, the learning rate is slightly reduced to prevent model update oscillations. Model performance validation: After the model weights are updated, the model performance is quickly validated using the validation set to ensure that the model's prediction accuracy has not decreased. If the validation results show that the prediction accuracy has decreased, the update is stopped immediately and the weight parameters are restored to those before the update. Model evolution: After each feedforward compensation control is completed, the model is updated online. Through continuous incremental updates, the model achieves dynamic self-learning and adaptive evolution, enabling the model to continuously adapt to complex and ever-changing rolling conditions such as roll wear, raw material performance fluctuations, ambient temperature changes, and rolling speed adjustments, thus ensuring the long-term prediction accuracy of the model.

[0081] 5. Closed-loop cycle: After completing the online adaptive update of the model, the system automatically returns to step S101 to continue the real-time acquisition of multi-source rolling feature data and mill status data, and enters the next closed-loop cycle of acquisition, prediction, judgment, optimization, control and update, so as to realize the full-process, continuous and adaptive closed-loop optimization control of 1060 aluminum foil rolling shape without frequent manual intervention.

[0082] Example 1

[0083] To accurately verify the technical effect of the method of the present invention, an industrial experimental platform for ultra-thin rolling of 1060 aluminum foil was built. The experimental object was 1060 aluminum foil for power batteries (target rolling thickness 10-15μm, width 1200mm). The experimental equipment was a 20-roll Sendzimir rolling mill (suitable for ultra-thin rolling of aluminum foil for power batteries). The experimental period was 30 days, divided into two stages: a control group experiment and an experimental group experiment. The same raw materials, rolling process benchmark parameters, and production conditions were used in each stage to ensure that the experimental variables were unique.

[0084] Control group: Employs the industry-standard classic plate shape mechanism model (plate shape equation) combined with traditional PID feedback control. The mechanism model is used for plate shape prediction, and PID control is used for plate shape deviation correction, representing a standard technical solution in the current aluminum foil rolling industry. Specifically, the PID controller parameters are tuned according to industry standards, with proportional, integral, and derivative coefficients all set within conventional ranges. The plate shape mechanism model uses a plate shape equation based on elastic flattening theory and lateral flow theory. The actuator adjustment cycle is in the millisecond range, typically around 100ms. There is no automatic optimization; adjustments rely on operator experience, and there is no online update mechanism.

[0085] Experimental Group: Employs the LSTM deep learning model with integrated attention mechanism of this invention, improved genetic algorithm for multi-objective optimization, and feedforward compensation control. All parameters are set according to the preferred values ​​of this invention. LSTM Model Structure: 3 LSTM hidden layers, 128 nodes per layer, additive attention mechanism, 2 fully connected layers (64 / 32 nodes), training uses the Adam optimizer with an initial learning rate of 0.001, 100 training epochs, batch size of 32, and an early stopping strategy; if the validation set loss does not decrease after 10 epochs, the training stops. Genetic Algorithm Parameters: Population size 50, number of iterations 100-200, elite retention ratio 10%, adaptive crossover probability 0.6-0.9, mutation probability 0.01-0.05. Online updates use mini-batch gradient descent with a batch size of 16, a learning rate of 0.0001, and are updated after each control action. The board shape for the next second is predicted based on the fused data of the current and past 5 seconds, with a feedforward compensation control response time <500ms.

[0086] Testing Indicators: Four core testing indicators are set: shape prediction accuracy, shape control accuracy, shape defect incidence rate, and rolling comprehensive energy consumption. Among them, shape prediction accuracy is quantified by root mean square error (RMSE) and mean relative error (MAPE); shape control accuracy is quantified by the maximum absolute value of shape deviation (I-Unit); shape defect incidence rate is quantified by the proportion of out-of-tolerance rolls to the total number of production rolls; and rolling comprehensive energy consumption is quantified by the comprehensive power consumption per ton of qualified aluminum foil (kWh / t).

[0087] Data Acquisition: An industrial-grade data acquisition system was used to collect and statistically analyze various indicators of the two experimental groups in real time. The control group and the experimental group each produced 500 tons of qualified 1060 aluminum foil. The plate shape, energy consumption and other indicators were recorded every hour. The effective data collection volume was more than 1.2 million records, ensuring the statistical significance of the experimental data.

[0088] After a 30-day industrial-scale comparative experiment, the quantitative comparison data of the four core technologies between the control group and the experimental group are shown in the table below, fully demonstrating the significant and superior effects of the invention. Statistical analysis was performed on all valid data, and the average values ​​of each indicator were calculated, resulting in the comparison results shown in Table 1. All data in Table 1 are experimental statistical averages. Table 1. Comparison of the effects of the method of the present invention and the prior art.

[0089] The experimental data in the table above show that the plate shape prediction RMSE was reduced by 90.59%, the defect rate by 97.6%, and the energy consumption by 27.06%. This order-of-magnitude synergistic improvement far exceeds the expectations of linear superposition.

[0090] Plate shape defect incidence rate: Traditional methods mainly result in edge waviness, center waviness, and local warping, with out-of-tolerance rolls accounting for 12.5% ​​of the total production rolls, requiring manual intervention or rework. In contrast, the plate shape defect incidence rate of the method of this invention is only 0.3%, and all of them are minor deviations that do not require manual intervention, improving the continuity of rolling production by more than 99%.

[0091] Overall rolling energy consumption: The method of this invention avoids excessive movement of the rolling mill actuator by using a multi-objective optimization function and an improved genetic algorithm, which significantly reduces the energy consumption of core actuators such as the bending roll mechanism and the segmented cooling spray mechanism. The overall power consumption per ton of qualified aluminum foil is reduced from 850 kWh / t to 620 kWh / t, and the economic efficiency of rolling production is significantly improved.

[0092] Example 2

[0093] This embodiment is a data-driven prediction and optimization system for the rolled shape of 1060 aluminum foil for batteries, using the method described in Embodiment 1, including: The data acquisition module is used to acquire multi-source rolling characteristic data and mill real-time status data during the rolling process of 1060 aluminum foil for batteries at a millisecond sampling frequency; the multi-source rolling characteristic data includes temperature cloud map of the roll system, and the mill real-time status data includes rolling oil film thickness monitoring data. The data preprocessing module is used to perform noise reduction, normalization and cleaning processes on the multi-source rolling feature data in sequence, and extract time-series features to obtain standardized feature vectors. The shape prediction module is used to fuse the standardized feature vector with the real-time status data of the rolling mill and input it into a deep learning model. The deep learning model predicts the shape values ​​of each transverse segment of the 1060 aluminum foil within the next second based on the fused data from the current period and a past period, and outputs a shape deviation cloud map. The deep learning model is an LSTM network with an integrated attention mechanism. The attention mechanism assigns high weights to the key influencing factor data in the rolling process to capture the nonlinear characteristics of the dynamic changes in roll thermal crown and the micron-level fluctuations in oil film thickness during the rolling process. The key influencing factor data in the rolling process includes rolling force, roll temperature, and tension. The residual calculation module is used to compare the predicted plate shape with the target zero plate shape curve channel by channel, calculate the plate shape residual value of each measurement channel, and form a residual matrix. The threshold determination module is used to determine whether the maximum value of the residual matrix exceeds a preset ±2I-Unit process threshold. The multi-objective optimization module is used to construct a multi-objective optimization function based on plate flatness and energy efficiency when the maximum value of the residual matrix exceeds a threshold. It uses the control parameters of the rolling mill actuator as optimization variables and employs an improved genetic algorithm for global iterative optimization to find the optimal combination of control parameters. The control parameters of the rolling mill actuator include bending roll force, roll tilt, and segmented cooling spray valve opening. The feedforward control module is used to send the optimal control parameter combination to each actuator of the rolling mill in real time with a response time of less than 500ms, and to perform feedforward compensation control to correct the plate shape deviation in advance. The online update module is used to feed back the actual adjusted plate shape detection results to the plate shape prediction module and update the weight parameters of the deep learning model through an online learning algorithm.

[0094] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven method for predicting and optimizing the rolled shape of 1060 aluminum foil for batteries, characterized in that, The process for rolling 1060 aluminum foil to extremely thin specifications includes the following steps: S101: Real-time acquisition of multi-source rolling characteristic data and mill status data: real-time acquisition of multi-source rolling characteristic data during the rolling process of 1060 aluminum foil for batteries at a millisecond-level sampling frequency, and simultaneous acquisition of real-time mill status data; the multi-source rolling characteristic data includes temperature cloud map of the roll system, and the real-time mill status data includes rolling oil film thickness monitoring data; S102: Data cleaning and temporal feature extraction: The multi-source rolling feature data is cleaned by noise reduction and normalization, and then temporal features are extracted based on the dynamic continuity of the rolling process to obtain a standardized feature vector. S103: Deep Learning Model for Accurate Plate Shape Prediction: The standardized feature vector is fused with the real-time status data of the rolling mill and input into the deep learning model to predict the plate shape values ​​of each transverse segment of the 1060 aluminum foil within a future time period, and outputs a plate shape deviation cloud map; the deep learning model is an LSTM network with an integrated attention mechanism, which assigns high weights to the key influencing factor data in the rolling process to capture the nonlinear characteristics of the dynamic changes in roll thermal crown and the micron-level fluctuations in oil film thickness during the rolling process; the key influencing factor data in the rolling process includes rolling force, roll temperature, and tension; S104: Calculate the residual matrix between the predicted plate shape and the target plate shape: Compare the predicted plate shape with the target zero plate shape curve set by the process channel by channel, calculate the plate shape residual value of each measurement channel in the transverse direction of the aluminum foil, and form a residual matrix that reflects the overall plate shape deviation. S105: Residual matrix process threshold determination: Determine whether the maximum value of the residual matrix exceeds the preset process threshold. If it does not exceed the threshold, return to step S101 to continue collecting data. If it exceeds the threshold, execute step S106. S106: Constructing a multi-objective optimization function for plate shape flatness and energy efficiency: Combining the need to improve plate shape with the goal of controlling production energy consumption, constructing a multi-objective optimization function based on plate shape flatness and energy efficiency; S107: Improved Genetic Algorithm for Multi-Parameter Global Optimization: Using the control parameters of the rolling mill actuator as optimization variables, an improved genetic algorithm is used to perform global iterative optimization on the multi-objective optimization function to obtain the optimal combination of control parameters that minimizes the objective function; the control parameters of the rolling mill actuator include bending roll force, roll tilt amount, and segmented cooling spray valve opening degree; S108: Feedforward Compensation Control and Online Adaptive Model Update: The optimal control parameter combination is sent to each actuator of the rolling mill in real time, and feedforward compensation control is performed to correct the plate shape deviation in advance. The actual adjusted plate shape detection results are fed back to the deep learning model, and the model weight parameters are updated through online learning algorithm. After completing one optimization, the process returns to step S101 to form a continuous closed-loop optimization control.

2. The data-driven method for predicting and optimizing the rolling shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, In step S101, the multi-source rolling characteristic data is collected by a sensor array deployed at key locations on the rolling mill. The sensor array includes tension sensors, pressure sensors, roll temperature infrared sensors, and a shape meter. The collected data includes rolling force, rolling speed, aluminum foil inlet and outlet tension, roll system temperature cloud map, and residual stress distribution on the aluminum foil surface. The real-time status data of the rolling mill includes rolling mill operating load, current action parameters of the actuators, rolling oil film thickness monitoring data, and roll wear status data.

3. The data-driven method for predicting and optimizing the rolled shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, In step S102, the noise reduction process uses wavelet transform to filter out environmental noise and equipment vibration noise collected by the sensor; the normalization process uses min-max normalization to map the data to the [0,1] interval; the time series feature extraction includes extracting the moving average of the time series, the rate of change of adjacent sampling times, and the first-order difference features of the data.

4. The data-driven method for predicting and optimizing the rolled shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, In step S103, the deep learning model is an LSTM network with an integrated attention mechanism. The LSTM network includes an input layer, three LSTM hidden layers, an attention feature layer, two fully connected layers, and an output layer. Specifically: each LSTM hidden layer has 128 nodes; the attention feature layer uses an additive attention mechanism; the first fully connected layer has 64 nodes, and the second layer has 32 nodes, using the ReLU activation function; the number of nodes in the output layer is consistent with the number of transverse measurement channels on the aluminum foil. The deep learning model performs shape prediction based on the fused data from the current and past 5 seconds, predicting the shape values ​​of each transverse segment of the 1060 aluminum foil within the next second, and outputs a shape deviation cloud map. The shape values ​​are measured in the industry-standard I-Unit.

5. The data-driven method for predicting and optimizing the rolling shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, The residual matrix in step S104 is calculated as follows: the difference between the predicted plate shape value and the target plate shape value of each measurement channel is calculated, and the differences of all channels are arranged in order according to the horizontal position of the aluminum foil to obtain the residual matrix. The target plate shape is the zero plate shape curve that meets the process requirements of 1060 aluminum foil for power batteries.

6. The data-driven method for predicting and optimizing the rolling shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, The expression for the multi-objective optimization function in step S106 is: ,in, To predict the improvement in plate shape, its value is the difference between the current maximum value of the residual matrix and the predicted maximum value of the residual matrix after parameter adjustment; The energy consumption variation of the rolling mill actuator is the sum of the energy consumption fluctuations of the bending roll mechanism, the tilting roll mechanism, and the segmented cooling spray mechanism, calculated according to their respective power characteristics and operating times. The weighting coefficients are dynamically adjusted based on actual rolling process requirements using the analytic hierarchy process (AHP), and satisfy the following conditions: .

7. The data-driven method for predicting and optimizing the shape of 1060 aluminum foil rolled for batteries according to claim 1, characterized in that, In step S107, the control parameters of the mill actuator include the bending force, the roll inclination, and the opening of the segmented cooling spray valve. Each parameter is set with a process-allowed value range as an optimization constraint. The improved genetic algorithm introduces an elite retention strategy and adaptive crossover and mutation probabilities. The crossover and mutation probabilities are dynamically adjusted according to the population fitness to avoid premature convergence of the algorithm.

8. The data-driven method for predicting and optimizing the rolling shape of 1060 aluminum foil for batteries according to claim 7, characterized in that, In step S107, the control parameters of the mill actuator are in the following ranges: bending force 0-500kN, roll tilt 0-5mm, and segmented cooling spray valve opening 0-100%. The improved genetic algorithm introduces an elite retention strategy and adaptive crossover and mutation probabilities. The elite retention ratio is 10%, and the top 10% of individuals with the best fitness in each generation are directly retained to the next generation. The adaptive crossover and mutation probabilities are dynamically adjusted according to the population fitness: individuals with high fitness values ​​have a crossover probability of 0.9 and a mutation probability of 0.05; individuals with low fitness values ​​have a crossover probability of 0.6 and a mutation probability of 0.

01. The lower limit of the crossover probability is 0.5, and the upper limit of the mutation probability is 0.

06.

9. The data-driven method for predicting and optimizing the rolling shape of 1060 aluminum foil for batteries according to claim 1, characterized in that, In step S108, the response time of the feedforward compensation control is less than 500ms, and the mill actuator completes the action adjustment before the actual plate shape deviation occurs; the online learning algorithm is a mini-batch gradient descent algorithm, which incrementally updates the weights of the deep learning model based on the actual plate shape feedback data to achieve adaptive evolution of the model.

10. A data-driven prediction and optimization system for the rolled shape of 1060 aluminum foil for batteries, employing the method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-source rolling characteristic data and mill real-time status data during the rolling process of 1060 aluminum foil for batteries at a millisecond sampling frequency; the multi-source rolling characteristic data includes temperature cloud map of the roll system, and the mill real-time status data includes rolling oil film thickness monitoring data. The data preprocessing module is used to perform noise reduction, normalization and cleaning of the multi-source rolling feature data, and extract time-series features to obtain a standardized feature vector. The plate shape prediction module is used to fuse the standardized feature vector with the real-time status data of the rolling mill and input it into the deep learning model. The deep learning model predicts the plate shape value of each transverse section of 1060 aluminum foil in the future time period based on the fused data of the current and past periods, and outputs a plate shape deviation cloud map. The residual calculation module is used to compare the predicted plate shape with the target zero plate shape curve channel by channel, calculate the plate shape residual value of each measurement channel, and form a residual matrix. The threshold determination module is used to determine whether the maximum value of the residual matrix exceeds a preset process threshold. The multi-objective optimization module is used to construct a multi-objective optimization function based on plate flatness and energy efficiency when the maximum value of the residual matrix exceeds a threshold. It uses the control parameters of the rolling mill actuator as optimization variables and employs an improved genetic algorithm for global iterative optimization to find the optimal combination of control parameters. The control parameters of the rolling mill actuator include bending roll force, roll tilt, and segmented cooling spray valve opening. The feedforward control module is used to send the optimal control parameter combination to each actuator of the rolling mill in real time with a response time of less than 500ms, and to perform feedforward compensation control to correct the plate shape deviation in advance. The online update module is used to feed back the actual adjusted plate shape detection results to the plate shape prediction module and update the weight parameters of the deep learning model through an online learning algorithm.

Images