A continuous plating control method for power battery wide-width pole piece

Abstract

The application provides a continuous plating control method for a wide width pole piece of a power battery, and relates to the field of electrochemical plating. The method comprises the following steps: inputting a current density distribution sequence into a time series prediction model to predict a predicted current density distribution sequence of a plurality of future continuous regulation periods; based on a target plating layer thickness and the predicted current density distribution sequence, a multi-objective optimization problem is constructed, which comprises a thickness deviation minimization target, a current change smoothness target and process constraint conditions, and an optimal current setting value sequence in the plurality of future continuous regulation periods is solved to control the independent power supply to perform partition power supply. The application aims to solve the technical problem that the plating layer thickness is not uniformly distributed in the transverse direction due to the edge effect in the existing continuous plating process of the wide width pole piece of the power battery, effectively suppresses the edge effect, improves the surface density uniformity of the plating layer thickness, and realizes high-precision and high-robustness intelligent control of the continuous plating process of the wide width pole piece.

Description

Technical Field

[0001] This invention relates to the field of electrochemical plating, and more particularly to a method for continuous plating control of wide-width electrode sheets for power batteries. Background Technology

[0002] During the continuous plating process of wide-width electrodes, due to the edge effect, the current spontaneously concentrates towards the edge of the electrode, resulting in a significantly thicker plating layer at the edge than in the middle, creating a laterally uneven distribution of "overplating at the edges and underplating in the middle." This problem directly affects the rate performance and consistency of the battery, and in severe cases, can even lead to safety hazards such as lithium plating and short circuits.

[0003] To improve the lateral uniformity of the plating coating, existing technologies have proposed a zoned power supply scheme, which divides the anode of the plating tank into multiple independently powered zones along the width of the electrode sheet, and compensates for edge effects by applying different currents to different zones. However, this scheme faces the following problems in practical applications: First, the conductivity of the plating solution causes current crosstalk between adjacent zones, making it difficult to accurately obtain the actual effective current density of each zone; second, the electroplating process has a large delay characteristic, and traditional proportional-integral-derivative (PID) control is prone to overshoot or oscillation; third, multiple process constraints are coupled with each other, making it difficult for traditional independent control strategies to achieve globally optimal current distribution.

[0004] To address the aforementioned issues, existing improvement solutions either only solve a single problem or rely on static models that cannot adapt to long-term drift in process parameters. Therefore, achieving a laterally uniform distribution of coating thickness during continuous plating of wide-width electrodes is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a continuous plating control method for wide-width electrode sheets of power batteries, to solve the technical problem of uneven lateral distribution of plating thickness caused by edge effects during the continuous plating process of wide-width electrode sheets of power batteries. The method includes: Step S1: Setting multiple independently controlled power supply zones along the width direction of the wide-width electrode sheet, acquiring the effective current density of each power supply zone in real time, and constructing a current current density distribution sequence; Step S2: Inputting the current current density distribution sequence into a pre-constructed time series prediction model to predict and obtain a predicted current density distribution sequence for multiple consecutive control cycles in the future; Step S3: Based on the target plating thickness and the predicted current density distribution sequence... The current density distribution sequence is measured, and a multi-objective optimization problem is constructed, which includes the objectives of minimizing thickness deviation, smoothing current change, and process constraints. The optimal current setpoint sequence for each power supply zone in the future multiple consecutive control cycles is obtained by solving the problem. Step S4: Extract the current setpoint corresponding to the first control cycle from the optimal current setpoint sequence and send it to the independent power supply corresponding to each power supply zone. Control the independent power supply to supply power to the corresponding power supply zone according to the current setpoint. Step S5: After each control cycle, Steps S1 to S4 are repeated based on the latest measured data to form a rolling time-domain closed-loop control.

[0006] The embodiments of the present invention have the following advantages:

[0007] This invention, through the establishment of multiple independently controlled power supply zones along the width of the electrode sheet, combined with an insulating shielding plate, multi-point flexible conductive contacts, and a crosstalk matrix decoupling algorithm based on singular value decomposition, can accurately obtain the effective current density of each zone, thus providing a reliable sensing basis for precise compensation of edge effects. By constructing a long short-term memory network prediction model with an attention mechanism, coupled with incremental fine-tuning and elastic weight consolidation regularization, it can accurately predict the current density distribution trend of multiple future control cycles, effectively overcoming the control lag problem caused by the large delay characteristics of the electroplating process. By constructing a multi-objective optimization problem including thickness deviation minimization, current change smoothness, and multiple process constraints, and solving it using convex quadratic programming, coupled with adaptive adjustment of weight coefficients and constraint relaxation mechanisms, it can achieve globally coordinated optimal allocation of current setpoints for each zone while satisfying all process constraints. Simultaneously, by employing a model predictive control framework that solves for the globally optimal sequence of future multiple cycles, executes only the first cycle instruction, and updates each cycle based on the latest measured data, it combines the advantages of predictive control and real-time feedback correction. In summary, this invention can effectively suppress edge effects and achieve lateral consistency and longitudinal stability of the coating thickness of wide-width electrode sheets. Attached Figure Description

[0008] Figure 1This is a flowchart illustrating the steps of a continuous plating control method for wide-width electrode sheets in a power battery according to the present invention. Detailed Implementation

[0009] This invention provides a continuous plating control method for wide-width electrode sheets in power batteries, solving the technical problem of uneven lateral distribution of plating thickness caused by edge effects during the continuous plating process of wide-width electrode sheets in existing power batteries. It effectively suppresses edge effects, improves the areal density uniformity of the plating thickness, and achieves high-precision, high-robust intelligent control of the continuous plating process for wide-width electrode sheets.

[0010] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0011] For examples, please refer to the appendix. Figure 1 This invention provides a method for continuous coating control of wide-width electrode sheets in power batteries, specifically including the following steps: Step S1: Set up multiple independently controlled power supply zones along the width direction of the battery's wide electrode, obtain the effective current density of each power supply zone in real time, and construct the current current density distribution sequence.

[0012] Furthermore, step S1 of the present invention also includes: An insulating shield is installed between adjacent power supply zones. The lower end of the insulating shield maintains a preset gap distance with the surface of the wide battery electrode. This preset gap distance is calculated based on the plating solution conductivity and the electrode's conveying speed according to a preset adjustment rule. Before the wide battery electrode enters the plating area, multiple flexible conductive contacts are installed along the width of the wide battery electrode. Each flexible conductive contact is independently configured with an elastic support element and a pressure sensor. The pressure sensor monitors the contact pressure between each contact and the electrode surface in real time. When the contact pressure of any contact deviates from the preset pressure range, a pressure adjustment mechanism is triggered to adjust the preload of the elastic support element of the contact. An offline calibration experiment is performed to construct an initial crosstalk matrix. The offline calibration experiment includes: sequentially applying a test current to each power supply zone individually, and using downstream thickness measurement... The thickness distribution curve is measured, and the thickness distribution curve is integrated and averaged over the width interval corresponding to each power supply zone to obtain the average thickness value of each zone. After normalizing all average thickness values, the current contribution ratio of the current energized zone to each zone is obtained. The current contribution ratios of all zones are combined to form the initial crosstalk matrix. During online operation, the total measured current of each power supply zone is collected in real time to form a measured current vector. Singular value decomposition is performed on the initial crosstalk matrix. After discarding the components with singular values ​​less than a preset threshold, the approximate matrix of the crosstalk matrix is ​​reconstructed. Then, the pseudo-inverse matrix of the approximate matrix is ​​calculated. The product of the measured current vector and the pseudo-inverse matrix is ​​calculated to obtain the effective current vector. Then, each effective current value in the effective current vector is divided by the electrode area of ​​the corresponding power supply zone to obtain the effective current density of each power supply zone.

[0013] Specifically, firstly, an insulating shield is installed between adjacent power supply zones. The lower end of this shield maintains a preset gap distance with the surface of the wide electrode sheet of the battery. The purpose of this gap distance is twofold: too small a gap increases the risk of scratching between the shield and the electrode sheet, while too large a gap weakens the suppression of current crosstalk. The preset gap distance is dynamically determined based on the conductivity of the plating solution and the electrode sheet's travel speed. The physical principle is that the higher the conductivity of the plating solution, the easier it is for the current to bypass the shield, requiring a smaller gap distance; the faster the electrode sheet travel speed, the faster the fluid renewal on the electrode sheet surface, requiring corresponding compensation for the gap distance. In actual operation, the optimal gap distance under different combinations of plating solution conductivity and electrode sheet travel speed can be pre-calibrated through experiments, forming a two-dimensional lookup table. During online operation, the control system collects the current plating solution conductivity and electrode sheet travel speed in real time, obtains the preset gap distance by looking up the table, and adjusts the shield position accordingly. This adjustment is triggered when operating conditions change, but remains constant during normal production.

[0014] Secondly, before the wide electrode sheet enters the plating area, multiple flexible conductive contacts are arranged along the width of the electrode sheet. These contacts enable simultaneous multi-point power supply to the electrode sheet, thereby shortening the lateral current flow distance and reducing the lateral potential drop caused by the electrode sheet's own resistance. If power is supplied only from a single point at the edge of the electrode sheet, the current needs to flow laterally across the entire width of the electrode sheet to reach the middle area. The electrode sheet resistance causes the cathode potential in the middle area to be significantly lower than that in the edge area, thus affecting the lateral uniformity of the plating thickness. By setting multiple contacts, each contact is responsible for injecting current into the electrode sheet area near it, significantly shortening the lateral current flow distance and effectively suppressing the potential drop problem. Each flexible conductive contact is independently equipped with an elastic support element and a pressure sensor. The elastic support element can be a spring, a spring sheet, or a pneumatic component, used to press the contact against the electrode sheet surface, ensuring a stable electrical contact between the contact and the electrode sheet. The pressure sensor monitors the contact pressure between the contact and electrode surfaces in real time and feeds the pressure signal back to the control system. The purpose of setting up pressure monitoring is that excessive contact pressure between the contact and electrode may cause scratches on the electrode surface or excessive wear of the contact, while insufficient contact pressure may lead to increased contact resistance or even power-off arcing. Therefore, it is necessary to control the contact pressure within a reasonable preset pressure range.

[0015] When the contact pressure of any contact deviates from the preset pressure range, the control system triggers the pressure regulating mechanism to adjust the preload of the corresponding elastic support element. The pressure regulating mechanism can be a stepper motor-driven screw mechanism, a proportional pneumatic regulating valve, or an electromagnetic regulating device. For example, when the pressure sensor detects that the contact pressure is below the preset lower limit, the control system controls the stepper motor to rotate forward, compressing the spring to increase the preload; when the contact pressure is above the preset upper limit, the control system controls the stepper motor to rotate in the reverse direction, releasing the spring to decrease the preload. Through this closed-loop regulation, the contact pressure of each contact can be independently controlled within the preset pressure range, ensuring a stable and consistent electrical contact state between all contacts and electrodes, providing a reliable current injection basis for accurately measuring the effective current density of each power supply zone.

[0016] Furthermore, the crosstalk matrix is ​​used to describe the current distribution of each power supply zone after the current drifts laterally in the plating bath, and its actual effect on each electrode area. To construct this matrix, an offline calibration experiment needs to be performed before the system is put into operation. The specific steps of the calibration experiment are as follows: First, a known test current is applied to each power supply zone individually, while other zones are not energized. For example, a test current is applied to the first zone first, and the current of the second to Nth zones is set to zero. During the energization process, the electrode runs continuously at the normal conveyor speed. A thickness gauge (such as a beta-ray thickness gauge or a laser thickness gauge) located downstream of the plating bath is used to measure the thickness distribution curve in the width direction of the electrode. Since the plating thickness is proportional to the current density, this thickness distribution curve actually reflects the current distribution of the currently energized zone on the electrode width. Theoretically, if there is no crosstalk, the thickness distribution curve should show a rectangular distribution in the width interval corresponding to the zone, with zeros on both sides. However, due to the existence of current crosstalk, the curve will form a main peak in the interval corresponding to the zone, while smaller secondary peaks appear in adjacent intervals. Then, the thickness distribution curve is integrated over the width interval corresponding to each power supply zone, and the average value is calculated to obtain the average thickness value of the region corresponding to each zone. For example, the thickness curve within the width interval corresponding to the first zone is integrated and divided by the interval width to obtain the average thickness value of that region. Similarly, the same integration and averaging operation is performed on the width intervals corresponding to the second, third, and so on up to the Nth zone to obtain N average thickness values. These N average thickness values ​​reflect the contribution of the current of the current energized zone to each electrode region. Finally, these N average thickness values ​​are normalized by dividing each average thickness value by the sum of all average thickness values ​​to obtain a set of proportionality coefficients. This set of proportionality coefficients indicates what proportion of the total current of the current energized zone flows to each electrode region, that is, the current contribution ratio of the current energized zone to the current of each zone. For example, the current contribution ratio of the first energized zone is 0.85, 0.08, 0.02, 0.00, 0.00, indicating that 85% of its current acts on its corresponding region, 8% flows to the adjacent region, and 2% flows to the next adjacent region. Repeat the above experiment with each zone as an energized zone to obtain N sets of current contribution ratios. Arrange these N sets of ratios in rows to form an N×N initial crosstalk matrix. Each column of this matrix corresponds to an energized zone, and each row corresponds to an electrode region. The matrix elements represent the current contribution ratio of the energized zone in the column to the electrode region in the row.

[0017] After offline calibration is completed and the initial crosstalk matrix is ​​obtained, the system enters the online operation phase. During online operation, each power supply zone's independent power source is equipped with a current sensor to collect the total output current of that zone in real time. Due to the existence of current crosstalk, this total measured current is not equal to the effective current actually acting on the corresponding electrode region of that zone, but includes components drifting to adjacent regions and components drifting in from adjacent regions. The total measured currents of all power supply zones are arranged in sequence to form a measured current vector, the dimension of which is equal to the number of power supply zones N. In order to calculate the effective current vector of each zone from the measured current vector, it is necessary to invert the crosstalk matrix. However, in actual engineering, due to measurement noise, calibration errors, and the possibility that the crosstalk matrix itself may have a near-linear correlation, directly inverting the crosstalk matrix can lead to numerical instability, or even make inversion impossible due to matrix singularity. To address this, this invention employs singular value decomposition (SVD) to preprocess the crosstalk matrix. SVD decomposes the crosstalk matrix into the product of three matrices: U, Σ, and the transpose of V. Σ is a diagonal matrix, and its diagonal elements are called singular values, arranged in descending order. The magnitude of the singular values ​​reflects the energy distribution of the crosstalk matrix in various directions; larger singular values ​​correspond to the main information components, while smaller singular values ​​correspond to noise or near-linearly correlated components. The specific processing procedure is as follows: First, SVD is performed on the crosstalk matrix to obtain all singular values. Then, a preset threshold is set, for example, multiplying the maximum value among all singular values ​​by 0.01 as the threshold. Singular values ​​smaller than this threshold are considered noise or redundant components and discarded. After discarding, only singular values ​​greater than or equal to the threshold and their corresponding U and V components are retained. These retained components are used to reconstruct an approximate matrix of the crosstalk matrix. This approximate matrix, after discarding noise and ill-conditioned components, has better numerical stability. Next, the pseudo-inverse matrix of the approximation matrix is ​​calculated. For non-square or ill-conditioned matrices, the pseudo-inverse matrix is ​​a generalized form of the inverse matrix, providing the optimal solution in the least squares sense. Finally, the measured current vector is multiplied by the pseudo-inverse matrix to obtain the effective current vector. Each element in the effective current vector represents the effective current value actually acting on the corresponding electrode region. Dividing this effective current value by the electrode area of ​​the corresponding power supply zone yields the effective current density of that zone.

[0018] Through the above singular value decomposition and pseudo-inverse solution, even if the crosstalk matrix has a certain degree of ill-conditioning or the measurement signal contains noise, the system can still stably calculate the effective current density of each partition, providing accurate state input for subsequent predictive control and optimization decisions.

[0019] Furthermore, the steps of the present invention also include: After a preset number of electrode rolls are produced, an online crosstalk matrix update process is triggered. In the online update process, the interruption closed-loop control is continuously interrupted for a preset duration. During the interruption duration, a disturbance current of preset amplitude is sequentially superimposed on each power supply zone, and the change in the total current of each zone and the change in the downstream thickness distribution curve are measured. The incremental change of the crosstalk matrix is ​​estimated online using a recursive least squares algorithm, and the incremental change is superimposed on the current crosstalk matrix to obtain the updated crosstalk matrix.

[0020] Specifically, during continuous production, process parameters such as plating solution composition, additive concentration, and electrode-anode spacing will slowly drift over time, causing the initial crosstalk matrix obtained from offline calibration to gradually become inaccurate. In order to maintain the accuracy of the solution, this invention designs an online crosstalk matrix update mechanism to periodically correct the crosstalk matrix without stopping the machine.

[0021] The online update process is triggered by completing a preset number of electrode rolls. For example, it is automatically triggered every 100 rolls of electrode produced. After triggering, the system temporarily interrupts the current closed-loop control mode and enters an update process with a preset interruption duration. During this interruption duration, the system sequentially adds a disturbance current of a preset amplitude to each power supply zone. The amplitude of this disturbance current is usually set to 1% to 5% of the normal production current to elicit a measurable response while ensuring safety. When a disturbance current is applied to a zone, due to the crosstalk effect, the disturbance current will not only change the coating thickness of the corresponding electrode area in this zone, but also cause slight changes in the thickness of adjacent areas through conduction in the plating solution. At this time, the system simultaneously collects two types of data: one is the change in the total current of each zone, i.e., the actual increment of the power supply output current; the other is the change in the thickness distribution curve measured by the downstream thickness gauge, i.e., the thickness increment of each electrode area. These two types of data constitute the input and output samples for estimating the change in the crosstalk matrix increment. After collecting a sufficient amount of disturbance response data, the system uses a recursive least squares algorithm to estimate the incremental change of the crosstalk matrix online. This algorithm can successively correct model parameters using newly collected data without reprocessing all historical data, resulting in high computational efficiency and suitability for online real-time applications. The algorithm takes the current change in each zone as input and the thickness change in each electrode region as output. Through recursive calculation, it obtains the deviation between the current crosstalk matrix and the true crosstalk matrix, i.e., the incremental change matrix. Finally, this incremental change is superimposed on the current crosstalk matrix to obtain the updated crosstalk matrix. After the update is complete, the system exits the interrupt state, resumes normal closed-loop control, and uses the updated crosstalk matrix for subsequent effective current density calculations.

[0022] Through the aforementioned online update mechanism, the crosstalk matrix can adaptively adjust itself to follow the slow drift of process conditions, maintaining long-term solution accuracy without frequent manual intervention and offline calibration.

[0023] Step S2: Input the current current density distribution sequence into the pre-built time series prediction model to predict and obtain the predicted current density distribution sequence for multiple consecutive control cycles in the future.

[0024] Furthermore, step S2 of the present invention also includes: Historical operational data is collected as a training dataset, including the effective current density, plating solution temperature, electrode conveying speed, total voltage, and current setpoints for each power supply zone in multiple past control cycles. The training dataset is normalized, scaling each feature variable to a preset numerical range to obtain a sample training set. A long short-term memory network structure with an attention mechanism is constructed. This attention mechanism automatically calculates the weight coefficients of each time step in the input sequence at each prediction time step, enabling the model to adaptively focus on historical moments where the weight coefficients are greater than a preset attention threshold. Mean squared error is used as the loss function, and offline training is performed using the sample training set until the validation set loss function value no longer decreases after multiple consecutive training rounds, thus generating a time series prediction model.

[0025] Specifically, to construct a time-series prediction model capable of accurately predicting future current density distribution, a sufficient amount of historical operating data needs to be collected as a training dataset. This data originates from the actual operating parameters recorded by the system during normal production. Specifically, within each control cycle, the system records a set of data, including: the effective current density of each power supply zone (calculated using the aforementioned crosstalk matrix decoupling method), plating solution temperature, electrode conveyor speed, total plating tank voltage, and current setpoints for each zone. This data covers the input, state, and output variables of the electroplating process, comprehensively describing the system's dynamic behavior. The time span of the data collection typically needs to cover multiple different operating conditions, such as different conveyor speeds, different plating solution temperatures, and different combinations of current settings, to ensure sufficient diversity and representativeness of the training data. In the collected raw data, the dimensions and numerical ranges of different characteristic variables vary significantly. For example, the plating solution temperature is typically between 20 and 60 degrees Celsius, the electrode conveyor speed may be between 5 and 30 meters per minute, and the current density value may be between 0.5 amperes per square decimeter and 4 amperes per square decimeter. If raw data with different dimensions is directly input into a neural network for training, feature variables with larger numerical ranges will dominate the loss function, causing the model to ignore feature variables with smaller numerical ranges but important physical meaning. To solve this problem, the training dataset needs to be normalized. The specific method of normalization is to scale each feature variable to a preset numerical range, such as 0 to 1 or -1 to 1. A commonly used normalization method is min-max normalization, calculated as follows: the normalized value equals the original value minus the minimum value of the feature, divided by the difference between the maximum and minimum values ​​of the feature. After normalization, all feature variables are mapped to the same numerical range, eliminating the influence of differences in dimensions and numerical ranges on model training. Arranging the normalized data in chronological order constitutes the sample training set. Each sample corresponds to complete data for one control cycle, and the chronological order of the samples reflects the dynamic evolution of the electroplating process. This sample training set will be used for subsequent offline training of the Long Short-Term Memory network, enabling the model to learn the nonlinear mapping relationship between input features and output predictions.

[0026] Next, during the continuous plating process of wide-width electrodes, the current density distribution at the current moment is not only related to the state of the most recent moments, but may also be affected by certain key events from earlier moments; for example, a fluctuation in the plating solution temperature may not have an observable impact on the plating thickness until tens of seconds later. Although traditional Long Short-Term Memory (LSTM) networks can remember long-term information, when processing extremely long time sequences, information from earlier moments is easily forgotten or diluted. To address this issue, this invention introduces an attention mechanism into the LSM network, enabling the model to automatically identify and focus on the moments in the historical sequence that are most important to the current prediction task. The core idea of ​​the attention mechanism is: at each prediction time step, the model calculates a weight coefficient for each historical time step in the input sequence. This weight coefficient reflects the importance of that time step to the current prediction. The larger the weight coefficient, the more important the information at that time step. The weight coefficient is calculated by: calculating the similarity between the hidden states output by the first LSM layer at each time step and the query vector of the current prediction time step, and then converting the similarity into a probability distribution form of weight coefficients using a normalized exponential function. After training, the model can automatically learn which types of events or which time ranges of data are most valuable for prediction. The system can pre-set an attention threshold, retaining only information from historical moments with weight coefficients greater than the threshold and discarding redundant information with weight coefficients that are too low, thereby improving prediction accuracy while reducing computational overhead.

[0027] In terms of network structure, the Long Short-Term Memory (LSTM) network with an attention mechanism designed in this invention consists of seven layers connected sequentially. The input layer receives a normalized sequence of feature vectors, which includes data such as effective current density, plating solution temperature, electrode travel speed, total voltage, and current setpoints from multiple past control cycles. The input layer passes the data to the first LSM layer, which has a certain number of memory units and outputs corresponding hidden states at each time step. These hidden states encode historical information from the beginning of the sequence to the current moment. The hidden states of all time steps output by the first LSM layer are simultaneously passed to the attention layer. The attention layer is the core innovative part of the network. This layer receives the complete hidden state sequence output by the first LSM layer at each time step, calculates the attention weight coefficient for each time step for the current prediction target, and then sums the hidden states of each time step according to the corresponding weight coefficients to obtain a fixed-dimensional context vector. This context vector gathers the most critical information for the current prediction from the entire historical sequence, and its dimension is independent of the length of the input sequence, thus solving the problem of information dilution when the LSM network processes long sequences. The attention layer outputs the weighted context vector to the second long short-term memory (LSTM) layer. The LSTM layer receives the context vector from the attention layer as input and has a certain number of memory units, but it only outputs the hidden state of the last time step, not the states of all time steps. The output of the LSTM layer is sequentially connected to the dropout layer, the fully connected layer, and the output layer. The dropout layer randomly sets the output of some neurons to zero with a preset probability to prevent overfitting. The fully connected layer uses a linear rectified function as the activation function to perform a non-linear transformation on the features. The number of nodes in the output layer is equal to the product of the number of power supply partitions and the prediction step size, and it uses a linear activation function to directly output the predicted current density distribution sequence for multiple consecutive control cycles. Through this seven-layer network structure, the model can adaptively focus on important moments in the historical sequence, achieving accurate prediction of the current density distribution trend.

[0028] During training, mean squared error (MSE) is used as the loss function. The MSE is calculated by comparing the predicted current density distribution sequence output by the model with the corresponding true current density distribution sequence, calculating the square of the difference between each predicted value and the true value, and then summing the squares of all differences and taking the average. The smaller the MSE value, the closer the model's prediction is to the true value. The advantage of using MSE as the loss function is that it amplifies the penalty effect of larger errors, prompting the model to pay more attention to sample points with larger prediction deviations. When training the network offline using a sample training set, the collected historical data is usually divided into a training set and a validation set. The training set is used to update the network's weight parameters, while the validation set is used to monitor the model's generalization ability and does not participate in parameter updates. In each training epoch, the optimizer calculates the gradient based on the loss function value and backpropagates, adjusting the network's weight parameters layer by layer to gradually decrease the loss function value. This invention employs an adaptive moment estimation optimizer, which can adaptively adjust the learning rate and has the advantages of fast convergence speed and insensitivity to hyperparameters. The training is stopped when the validation set loss function value no longer decreases for several consecutive training epochs. Specifically, after each training epoch, the current model's loss function value on the validation set is calculated and compared with the values ​​from previous epochs. If the validation set loss function value does not decrease for a predetermined number of epochs (e.g., 10 consecutive epochs), the model is considered to have converged, and continuing training may lead to overfitting. Therefore, training is stopped. Overfitting refers to the model overlearning noise and details in the training set, resulting in poor performance on the unseen validation set. This early stopping strategy achieves a balance between model performance and generalization ability.

[0029] After training, all weight parameters in the network are fixed, forming a deployable time-series prediction model. This model can receive input feature vectors from the current control cycle and several previous consecutive control cycles, and output a predicted current density distribution sequence for several future consecutive control cycles, providing predictive state information for subsequent rolling optimization control. This offline training process is completed before the system is put into operation, and the trained model is deployed to the online control system to perform real-time prediction tasks.

[0030] Furthermore, the steps of the present invention also include: After each preset number of electrode rolls are produced, the newly accumulated running data is added to the training dataset. A fine-tuning learning rate lower than the offline training learning rate is used to perform incremental fine-tuning training on the time series prediction model for a preset number of small batch training rounds. During the incremental fine-tuning training process, an elastic weight consolidation regularization term is used to suppress catastrophic forgetting. The elastic weight consolidation regularization term is assigned different constraint strengths according to the importance of network parameters in historical training.

[0031] Specifically, the time series prediction model obtained through offline training may experience slow changes in the dynamic characteristics of the electroplating system during long-term operation, such as adjustments to the plating solution formula, equipment aging, or seasonal fluctuations in ambient temperature. If the original model is used continuously without updates, the prediction accuracy will gradually decrease. To address this, this invention designs an incremental fine-tuning training mechanism to continuously update the model using newly accumulated operational data without shutting down the system.

[0032] After producing a preset number of electrode rolls, the system adds the newly accumulated operational data to the existing training dataset. This new data contains the latest dynamic characteristics of the system under current operating conditions. Then, the system uses a fine-tuning learning rate lower than the offline training learning rate to perform incremental fine-tuning training on the model for a preset number of mini-batch training epochs. The fine-tuning learning rate is typically set to one-tenth to one-hundredth of the offline training learning rate; for example, if the offline learning rate is 0.001, the fine-tuning learning rate is set to 0.0001. The mini-batch training epochs are usually set to several to dozens of epochs, rather than hundreds or thousands of epochs as in offline training. The purpose of using a smaller learning rate and fewer training epochs is to gently adjust the model within a limited number of update steps, allowing it to adapt to new data while maintaining the stability of existing knowledge.

[0033] A core problem faced by incremental fine-tuning training is catastrophic forgetting. Catastrophic forgetting refers to the phenomenon where neural networks drastically change their original weight parameters when learning new tasks or data, causing previously learned knowledge to be overwritten and forgotten. For example, after adapting to the characteristics of a new batch of plating solution, a model may forget the basic patterns reflected in the earlier training data. To suppress catastrophic forgetting, this invention introduces an elastic weight consolidation regularization term into the loss function of incremental fine-tuning training. The core idea of ​​the elastic weight consolidation regularization term is that each weight parameter in the network has a different degree of importance to the historical training task. Important parameters are subject to stronger constraints to minimize their changes, while less important parameters are allowed to change more significantly. The importance is quantified by calculating the second derivative of the loss function with respect to each weight parameter (i.e., the diagonal elements of the Fisher information matrix) after offline training. The larger the value, the greater the impact of the parameter on the output result, and therefore the more important it is. During incremental fine-tuning training, the regularization term penalizes large changes in important parameters. Specifically, it adds an additional term to the loss function, which is equal to the square of the change in each weight parameter multiplied by its corresponding preset constraint strength coefficient. In this way, the model can retain key knowledge learned previously while learning new data, thus achieving a balance between adapting to new working conditions and maintaining historical knowledge.

[0034] Step S3: Based on the target coating thickness and the predicted current density distribution sequence, construct a multi-objective optimization problem that includes the objectives of minimizing thickness deviation, smoothing current change, and process constraints, and solve it to obtain the optimal current setpoint sequence for each power supply zone in multiple consecutive control cycles in the future.

[0035] Furthermore, step S3 of the present invention also includes: The following objectives are established: A thickness deviation minimization objective function is established: the sum of squares of the differences between the predicted coating thickness (obtained by multiplying the predicted current density of each power supply zone by the Faraday constant) and the target coating thickness over multiple consecutive control cycles. A current change smoothness objective function is established: the sum of squares of the differences between the current setpoints of each power supply zone in two adjacent control cycles over multiple consecutive control cycles. A safety boundary penalty function is established: using a soft constraint, when the current setpoint of any power supply zone is greater than or equal to the sum of the preset lower current safety limit and the first safety boundary, or less than or equal to the difference between the preset upper current safety limit and the second safety boundary, the penalty is calculated based on the square of the distance between the current setpoint and the lower or upper current safety limit. The relation outputs a positive penalty term, where the first safety boundary and the second safety boundary are both preset positive values; the objective function for minimizing the thickness deviation is multiplied by a first weighting coefficient, the objective function for smoothing the current change is multiplied by a second weighting coefficient, and the objective function for penalizing the safety boundary is multiplied by a third weighting coefficient, and then summed to form the overall objective function; constraints are set as follows: the current setting value of each power supply zone is greater than or equal to a preset lower current safety limit and less than or equal to a preset upper current safety limit; the rate of change of current in each power supply zone between two adjacent control cycles is less than or equal to a preset maximum rate of change; the absolute value of the difference between the current setting values ​​of two adjacent power supply zones within the same control cycle is less than or equal to a preset maximum difference threshold between adjacent zones.

[0036] Specifically, the basic structure of the thickness deviation minimization objective function is as follows: First, for each power supply zone, the predicted current density over several consecutive control cycles is multiplied by the Faraday constant to obtain the predicted coating thickness for that zone in each prediction cycle. The Faraday constant is a fundamental physical constant in electrochemistry, establishing a linear relationship between current density and deposition rate; that is, the higher the current density, the thicker the metal layer deposited per unit time. Multiplying the predicted current density by the Faraday constant and the corresponding deposition time coefficient yields the predicted coating thickness. Then, each predicted coating thickness is subtracted from the pre-set target coating thickness to obtain the thickness deviation value. If the predicted thickness is greater than the target thickness, the deviation is positive; if the predicted thickness is less than the target thickness, the deviation is negative. To eliminate the mutual cancellation of positive and negative deviations, the deviation value is squared to convert all deviations to non-negative values. Larger deviations are further amplified after squaring, thus incurring a greater penalty during optimization. Finally, the squared deviations of all power supply zones over several consecutive control cycles are summed to obtain the value of the thickness deviation minimization objective function. The smaller the value of this objective function, the closer the overall predicted coating thickness is to the target thickness. In the subsequent optimization process, the control system searches for a sequence of current setpoints that minimizes the objective function value, thereby driving the actual coating thickness to converge towards the target thickness. Through this objective function, the system can quantitatively evaluate and optimize the thickness deviation, providing direct optimization guidance for suppressing edge effects and achieving a laterally uniform distribution of coating thickness.

[0037] In the continuous plating process of wide-width electrodes, in addition to striving to make the plating thickness as close as possible to the target value, it is also necessary to consider the smoothness of the current setpoint changes. If the current setpoint changes drastically between adjacent control cycles, it will cause large fluctuations in the power supply output current. On the one hand, this may impact the power supply equipment and shorten its service life; on the other hand, drastic current changes will cause abrupt changes in the plating deposition rate, which may lead to stress concentration, coarse grains, or quality defects such as burning inside the plating. To address this, this invention establishes a current change smoothness objective function. The basic structure of this objective function is as follows: For each power supply zone, the difference in current setpoint between two adjacent control cycles is calculated. Specifically, in multiple consecutive control cycles, the current setpoint of the later cycle is subtracted from the current setpoint of the previous cycle to obtain a set of difference sequences. Each difference reflects the magnitude and direction of the current adjustment of that zone between different cycles. If the adjustment magnitude is too large, the absolute value of the difference will be large; if the adjustment magnitude is gradual, the absolute value of the difference will be small. To penalize large current jumps during the optimization process, the difference is squared. The purpose of squaring is to convert all differences into non-negative values ​​after squaring, and to further amplify larger differences, thus incurring stronger penalties. Then, the squared differences of all power supply zones over all adjacent cycle pairs are summed to obtain the objective function for current change smoothness. The smaller the value of this objective function, the smoother the change in the current setpoint between cycles, and the smaller the fluctuations. In subsequent optimization, the control system needs to simultaneously minimize both the thickness deviation objective function and the current change smoothness objective function. There is a certain trade-off between these two objectives: excessively pursuing the minimization of thickness deviation may lead to drastic current adjustments, while excessively pursuing current smoothness may sacrifice thickness control accuracy. By setting reasonable weighting coefficients, the system can achieve a balance between the two, that is, ensuring the coating thickness accuracy while making the current setpoint change as smoothly as possible, thereby achieving a control effect that meets quality requirements while protecting equipment and maintaining process stability.

[0038] During the continuous deposition of wide-width electrodes, the current setting value cannot be arbitrarily selected; it must be limited to a safe range. Too low a current leads to low deposition efficiency or even side reactions, while too high a current may cause quality problems such as dendrite growth, coating burn-out, or hydrogen evolution. Therefore, this invention sets a lower and upper safe current limit as hard constraints. However, if the constraints are directly set as hard boundaries, the system will operate in an extreme state for a long time when the optimal solution falls exactly on the boundary, posing a certain safety risk. To avoid this situation, this invention introduces a safe boundary penalty function, using soft constraints to penalize current setting values ​​approaching the safe boundary.

[0039] The safety boundary penalty function is constructed as follows: a first safety boundary is added to the lower current safety limit, and a second safety boundary is subtracted from the upper current safety limit. Both the first and second safety boundaries are preset positive values, for example, 0.2 amperes per square decimeter. This divides the entire current range into three regions: a normal region, a warning region, and a prohibited region. The normal region refers to the interval where the current setpoint is greater than the sum of the lower current safety limit and the first safety boundary, but less than the difference between the upper current safety limit and the second safety boundary; the penalty function output is zero in this region. The warning region refers to the interval where the current setpoint is between the lower current safety limit and the sum of the lower current safety limit and the first safety boundary, or between the upper current safety limit minus the second safety boundary and the upper current safety limit. The prohibited region refers to the interval where the current setpoint is less than the lower current safety limit or greater than the upper current safety limit; this region is directly prohibited by hard constraints and will not appear in feasible solutions. When the current setting value of any power supply zone enters the warning area, the safety boundary penalty function outputs a positive penalty term according to the squared relationship between the current setting value and the lower or upper safety limit of the current. Specifically, if the current setting value is close to the lower limit, the difference between the current setting value and the lower safety limit of the current is calculated, the absolute value is squared, and then multiplied by a preset penalty coefficient. If the current setting value is close to the upper limit, the difference between the current setting value and the upper safety limit of the current is calculated, the absolute value is squared, and then multiplied by a preset penalty coefficient. The closer to the safety boundary, the smaller the squared value and the smaller the penalty term. The farther away from the safety boundary (i.e., the closer to the normal area), the larger the penalty term. This squared relationship makes the penalty term increase smoothly from the boundary of the normal area towards the safety boundary.

[0040] Through the above soft constraint design, the optimization solver tends to keep the current setpoint within the normal region when searching for the optimal solution. If it has to enter the warning region due to process limitations, the system will also prioritize the solution closer to the boundary of the normal region rather than the solution directly close to the safety boundary. This design provides a smooth objective function gradient for optimization while ensuring safety, avoids the problem of abrupt changes in the feasible region caused by hard constraints, and improves the stability and quality of optimization solutions.

[0041] The three objective functions are interdependent: minimizing thickness deviation may lead to drastic current changes, smoothing current may sacrifice thickness accuracy, and the safety boundary penalty function imposes constraints on current values ​​approaching the limit. To achieve a balance among these three, this invention assigns a weighting coefficient to each objective function. The thickness deviation minimization objective function is multiplied by a first weighting coefficient, the current smoothness objective function by a second weighting coefficient, and the safety boundary penalty function by a third weighting coefficient. The three weighted results are then summed to form the overall objective function. By adjusting the relative magnitudes of the three weighting coefficients, the control system's emphasis on different objectives can be altered. For example, increasing the first weighting coefficient makes the system focus more on thickness accuracy, increasing the second weighting coefficient makes the system focus more on current stability, and increasing the third weighting coefficient makes the system more inclined to move away from the safety boundary. The optimization solver aims to minimize the overall objective function and seeks the current setpoint sequence that optimally combines the three weighted objectives. Subsequent dependent claims will further explain the adaptive adjustment method of the weighting coefficients.

[0042] When constructing a multi-objective optimization problem, in addition to defining the objective function, it is also necessary to set explicit constraints to ensure that the solved current setpoint sequence is physically feasible and technologically acceptable. This invention sets the following three types of constraints: The first type is the current safety amplitude constraint. The current setpoint for each power supply zone must be greater than or equal to a preset lower current safety limit and less than or equal to a preset upper current safety limit. The lower current safety limit is usually set based on the minimum effective current density of the electroplating process; below this value, deposition efficiency may be too low or side reactions such as hydrogen evolution may occur. The upper current safety limit is set based on the limiting current density of the electroplating process; above this value, dendrite growth, coating burning, or excessive additive consumption may occur. This constraint directly limits the range of current setpoint values ​​for each zone in the form of a hard constraint. The second type is the current change rate constraint. The current change rate of each power supply zone between two adjacent control cycles must be less than or equal to a preset maximum change rate. This constraint limits the maximum adjustment range of the current setpoint from one cycle to the next. The purpose of this constraint is twofold: firstly, the power supply equipment itself has limited response capability, and the current cannot instantly jump to any value; secondly, a rapid change in current can lead to abrupt changes in the coating deposition rate, potentially causing uneven stress or defects within the coating. By limiting the rate of change, the current adjustment process can be ensured to be smooth and controllable. The third type is the current difference constraint between adjacent zones. Within the same control cycle, the absolute value of the difference between the current setpoints of two adjacent power supply zones must be less than or equal to the preset maximum difference threshold between adjacent zones. This constraint limits the degree of difference in current values ​​between adjacent zones. The reason for setting this constraint is that if the current difference between adjacent zones is too large, a sudden current change will occur at the boundary between the two zones, causing a sharp change in the coating thickness in that area, forming a local non-uniform band. By limiting the current difference between adjacent zones, the coating thickness can be smoothly transitioned in the width direction, avoiding steep thickness gradients. The above three types of constraints together constitute the feasible region of the optimization problem. Any current setpoint sequence that does not meet these constraints will not be adopted as the optimal solution, thus ensuring that the optimization results meet both process safety requirements and equipment physical limitations, while also ensuring the smoothness of the coating thickness in both the width and time directions.

[0043] Furthermore, step S3 of the present invention also includes: The system monitors the thickness deviation statistical characteristics fed back by the downstream thickness gauge in real time. The thickness deviation statistical characteristics include the deviation mean and the deviation variance. When the deviation mean is greater than a preset mean threshold, the first weight coefficient is increased by a preset first step size until the deviation mean is less than or equal to the preset mean threshold or the first weight coefficient reaches a preset upper weight value. When the deviation variance is greater than a preset variance threshold, the second weight coefficient is increased by a preset second step size until the deviation variance is less than or equal to the preset variance threshold or the second weight coefficient reaches the preset upper weight value. When the deviation mean is less than or equal to the preset mean threshold and the deviation variance is less than or equal to the preset variance threshold, the first weight coefficient and the second weight coefficient are decreased by a preset third step size until the first weight coefficient and the second weight coefficient return to their respective preset initial values. The sum of the first weight coefficient, the second weight coefficient, and the third weight coefficient is equal to 1, and each is restricted within a closed interval formed by a preset lower weight value and a preset upper weight value.

[0044] Specifically, the thickness deviation data fed back by the downstream thickness gauge is monitored in real time, and two statistical characteristics are calculated online: the mean deviation and the variance deviation. The mean deviation reflects the systematic deviation between the actual coating thickness and the target thickness, i.e., the degree of overall thickness deviation. The variance deviation reflects the degree of fluctuation in coating thickness along the electrode width direction or along the time direction, i.e., the uniformity. These two statistical characteristics correspond to two different aspects of thickness control: the mean represents accuracy, and the variance represents consistency.

[0045] When the mean deviation exceeds the preset mean threshold, it indicates that the actual coating thickness deviates significantly from the target value, indicating a systematic bias in the system. In this case, the control system increments the first weighting coefficient by a preset step size. The first weighting coefficient controls the proportion of the thickness deviation minimization objective function in the overall objective function. Increasing this coefficient makes the optimizer focus more on reducing the thickness deviation. The incrementing process continues, increasing the step size each time the mean deviation is detected to still be greater than the threshold, until the mean deviation decreases to less than or equal to the preset mean threshold, or the first weighting coefficient reaches the preset upper weight limit. Once the upper limit is reached, even if the mean deviation still does not meet the target, the increment stops to avoid weighting coefficient imbalance. When the variance exceeds the preset variance threshold, it indicates significant coating thickness fluctuations and insufficient lateral uniformity or longitudinal stability. In this case, the control system increments the second weighting coefficient by a preset second step size. The second weighting coefficient controls the proportion of the current change smoothness objective function in the overall objective function. Increasing this coefficient suppresses drastic fluctuations in the current setpoint, thereby indirectly stabilizing the coating thickness. Similarly, the incremental process continues until the deviation variance decreases to less than or equal to the preset variance threshold, or the second weight coefficient reaches the preset weight upper limit value.

[0046] When the mean thickness deviation is less than or equal to the preset mean threshold and the variance of the thickness deviation is less than or equal to the preset variance threshold, it indicates that the current coating thickness control effect has reached the expected target. There is no systematic deviation of overall thickness or thinness, nor is there a problem of large thickness fluctuations. In this case, the first and second weighting coefficients, which were previously increased due to excessive deviation, need to be restored to normal levels to avoid over-control. The specific restoration mechanism is as follows: the control system decreases the first and second weighting coefficients by a preset third step. The third step is a preset positive value, usually set to the same or smaller value as the first and second step to ensure the smoothness of the restoration process. The decreasing operation continues. After each control cycle or at each preset number of steps, the first and second weighting coefficients are each reduced by one third step until they are restored to their respective preset initial values. For example, assuming the initial preset value of the first weight coefficient is 0.5, and it is currently increased to 0.7, with a third step size of 0.01, the system will decrease by 0.01 each cycle, returning to 0.5 after 20 cycles. It should be noted that the decrease operation is performed simultaneously on both the first and second weight coefficients, as both were increased when the deviation exceeded the limit and should be restored synchronously after the deviation returns to normal. The sum of the first, second, and third weight coefficients equals 1. This constraint ensures that the relative weights of each component in the overall objective function have a normalized physical meaning, avoiding unexpected changes in the optimization results due to overall scaling of the weight coefficients. When the first and second weight coefficients decrease, the third weight coefficient automatically increases to maintain the sum of the three at 1. For example, if the first and second weight coefficients each decrease by 0.01, the third weight coefficient automatically increases by 0.02. Each of the three weighting coefficients is further constrained within a closed interval formed by a preset lower weight limit and a preset upper weight limit. The lower weight limit prevents any weighting coefficient from becoming too small or even negative, ensuring that each objective function retains its most basic contribution to the overall objective function. The upper weight limit prevents any weighting coefficient from becoming too large and dominating the entire optimization process, causing other objectives to be ignored. For example, the lower weight limit can be set to 0.1 and the upper weight limit to 0.8. The first weighting coefficient is always constrained within this interval during the adaptive adjustment process. When the decreasing operation attempts to lower it below the lower limit, the decreasing stops and it remains at the lower limit.

[0047] Through the aforementioned decreasing recovery mechanism and interval restrictions, the weighting coefficient can automatically return to the normal level after the deviation is eliminated, and maintain a reasonable numerical range throughout the entire adjustment process, thus achieving a smooth transition of the control system between different production stages.

[0048] Furthermore, step S3 of the present invention also includes: The overall objective function is transformed into a standard quadratic form expression, and the constraints are transformed into a system of linear inequalities, thus formalizing the multi-objective optimization problem into a convex quadratic programming problem. The interior-point method is used for solving this problem. A quadratic programming solver is invoked, and the quadratic coefficient matrix of the overall objective function, the linear coefficient vector, the coefficient matrix of the constraints, and the constraint boundary values ​​are input into the quadratic programming solver. The quadratic programming solver iteratively calculates and outputs the optimal solution vector that minimizes the overall objective function while satisfying all constraints. The optimal solution vector contains the optimal current setting value for each power supply zone in multiple consecutive control cycles. The current setting value of the first control cycle is extracted from the optimal solution vector as the execution instruction for the current control cycle. If the convex quadratic programming problem fails to converge within a preset maximum number of iterations, the first value in the optimal current setting value sequence obtained in the previous control cycle is used as the execution instruction for the current control cycle, and a solution failure event is recorded.

[0049] Specifically, after constructing the overall objective function and setting the constraints, a numerical optimization algorithm is needed to solve the problem to obtain the optimal current setpoint sequence. To ensure the solution process is efficient, stable, and guarantees finding the global optimum, this invention formalizes the aforementioned multi-objective optimization problem as a convex quadratic programming problem. Convex quadratic programming is a special type of mathematical optimization problem where the objective function is a quadratic and convex function, and the constraints are linear inequalities or equality equations. Convex quadratic programming has an important property: if a feasible solution exists, then any local optimum is also a global optimum. This property guarantees the global optimality of the solution. To achieve this formalization, the overall objective function needs to be transformed into a standard quadratic form expression. The general form of the standard quadratic form is: half multiplied by the transpose of the variable vector, multiplied by the quadratic coefficient matrix, multiplied by the variable vector, plus the transpose of the linear coefficient vector multiplied by the variable vector, plus a constant term. The quadratic coefficient matrix is ​​derived from the squared terms in the overall objective function (such as the square of the thickness deviation, the square of the current change, etc.), the linear coefficient vector is derived from the linear terms in the overall objective function (if any), and the constant term does not affect the position of the optimal solution in the optimization solution and can be ignored. Simultaneously, the three types of constraints (current safety amplitude constraint, current change rate constraint, and adjacent partition current difference constraint) are transformed into a system of linear inequalities, that is, expressed in matrix form as the constraint coefficient matrix multiplied by the variable vector being less than or equal to the constraint boundary value.

[0050] After formalizing the problem as described above, the interior-point method is used to solve the convex quadratic programming problem. The interior-point method is a classic algorithm for solving convex optimization problems. Its basic idea is to start from within the feasible region and iteratively search along the direction that reduces the objective function. Simultaneously, it incorporates constraints into the objective function through barrier functions or logarithmic barrier functions to ensure that the constraints are always satisfied during the iteration process. Compared to algorithms such as the simplex method, the interior-point method has the advantages of fast convergence and good numerical stability when dealing with large-scale, high-dimensional optimization problems. In specific implementation, the system calls a mature quadratic programming solver library, such as OSQP, and passes the quadratic and linear coefficient matrices of the formalized overall objective function, the coefficient matrices of the constraints, and the constraint boundary values ​​as input parameters to the solver. The solver internally implements the interior-point method algorithm and iteratively calculates and outputs the optimal solution vector that minimizes the overall objective function while satisfying all constraints. This optimal solution vector is the sequence of optimal current setpoints for each power supply zone over several consecutive control cycles. Through the formalization and solution process described above, the originally complex multi-objective and multi-constraint optimization problem is transformed into a standard mathematical form, which can be solved efficiently and reliably using mature numerical optimization tools.

[0051] After the overall objective function and constraints are input into the quadratic programming solver, the solver initiates an internal iterative calculation process. The basic principle of the iterative calculation is as follows: The solver starts with an initial feasible solution that satisfies all constraints. Then, in each iteration, the solver calculates the gradient of the objective function at the current solution, determines the search direction that causes the objective function value to decrease the fastest, and moves along that direction by a certain step size to obtain a new feasible solution. This process is repeated until a preset convergence condition is met, such as the difference between the objective function values ​​of two adjacent iterations being less than a preset tolerance, or the iteration step size being less than a preset threshold. When the iterative calculation converges, the solver outputs the optimal solution vector, which is a one-dimensional array whose dimension is equal to the number of power supply zones N multiplied by the prediction step size H. For example, with N equal to 10 power supply zones and H equal to 5 prediction steps, the length of the optimal solution vector is 50. The elements of this vector are organized according to a preset order, such as first arranging the current setpoints of all power supply zones in the first control cycle, then arranging the current setpoints of all power supply zones in the second control cycle, and so on, or arranging them according to the priority of the power supply zones. Regardless of the arrangement, this vector completely contains the optimal current setpoint for each power supply zone over multiple consecutive control cycles. Specifically, for the j-th power supply zone, the optimal solution vector contains H current setpoints for that zone over the 1st to Hth control cycles. These setpoints are the solution that minimizes the overall objective function value while satisfying all constraints (current safety amplitude constraint, current change rate constraint, and current difference constraint between adjacent zones). Minimizing the overall objective function value means achieving an optimal balance among the three weighted objectives of thickness deviation, current change smoothness, and safety boundary penalty. For example, if a zone is predicted to be thinner in the future, the solver will appropriately increase the current setpoint for that zone within the allowable current change rate range, while ensuring that this adjustment does not cause the difference between adjacent zones to exceed the limit. Through the above iterative calculation and output process, the control system obtains the globally optimal control sequence for multiple control cycles, providing a decision-making basis for subsequent rolling time-domain control. The quality of the solution vector directly determines the control accuracy of the coating thickness and the stability of the system.

[0052] After the quadratic programming solver outputs the optimal solution vector, this vector contains the optimal current setpoint for each power supply zone over several consecutive future control cycles. However, in actual control, the system does not need to, and should not, directly execute the setpoints for all future cycles. This is because there is an unavoidable deviation between the prediction model and the actual system, and random disturbances exist during production. If the setpoints for multiple future cycles are executed directly, the model error will accumulate over time, causing the control effect to gradually deviate from expectations. Therefore, this invention extracts only the current setpoint corresponding to the first control cycle from the optimal solution vector as the execution instruction for the current control cycle. That is, for each power supply zone, only its optimal current setpoint for the first future control cycle is taken and sent to the corresponding independent power source for execution. Although the setpoints for the remaining future cycles are solved, they are not executed directly but are overwritten after recalculation in the next control cycle. In actual solving, due to the complexity of the problem, the accuracy limitations of numerical calculation, or improper solver parameter settings, the convex quadratic programming problem may fail to reach the convergence condition within the preset maximum number of iterations. In this case, the solver cannot output the optimal solution that meets the accuracy requirements. To ensure that the control system can continue to operate even when the solution fails, this invention designs a convergence failure handling mechanism.

[0053] When a solution fails, the system does not use any result from the current solution. Instead, it calls the optimal current setpoint sequence obtained from the previous control cycle and extracts the first value as the execution command for the current control cycle. The reason for using the solution from the previous cycle is that it was calculated under the optimal conditions at that time and has been verified by measured data after actual execution. It is usually a feasible and near-optimal control command. In the absence of a better option, using the solution from the previous cycle ensures the continuity and stability of the control system, avoiding control interruptions or abnormal output values ​​due to solution failure. Simultaneously, the system records a solution failure event. The recorded information may include the time of failure, current operating parameters, and the solver's return status. These records are used for subsequent analysis of the cause of the failure, such as whether the feasible region is too small due to overly tight constraints or whether the problem is ill-conditioned due to model parameter drift. When the cumulative number of solution failure events reaches a preset number, the system can trigger an alarm or automatically adjust the solver parameters to facilitate appropriate maintenance and optimization by operators or the system. Through the aforementioned convergence failure handling mechanism, the robustness of the control system is significantly enhanced, and normal control output can still be maintained even in the event of occasional solver failure.

[0054] Furthermore, the steps of the present invention also include: When the number of consecutive failures in solving the convex quadratic programming problem reaches a preset threshold for consecutive failures, a constraint relaxation mechanism is triggered. In the constraint relaxation mechanism, a relaxation factor is introduced to relax the current rate of change constraint and the maximum difference threshold constraint between adjacent partitions, respectively. The initial value of the relaxation factor is set to one, and it is increased by a preset step size after each failure until the solution is successful or the preset upper limit of the relaxation factor is reached. When the solution is successful, the relaxation factor is gradually reduced to restore it to the initial value of one.

[0055] Specifically, in the process of optimization using a quadratic programming solver, there are sometimes situations where multiple consecutive solution failures occur. Common reasons for these consecutive failures include: overly strict constraints leading to a small feasible region or even the absence of a feasible solution; significant fluctuations in process parameters causing the optimal solution under the current operating conditions to fall near the boundary of the feasible region, making it difficult for the solver to converge stably; or deviations between the model parameters and the actual system, resulting in a mismatch between the constraints and the actual situation. To address these issues, this invention designs a constraint relaxation mechanism. The constraint relaxation mechanism is triggered when the number of consecutive solution failures of the convex quadratic programming problem reaches a preset consecutive failure threshold. For example, setting the consecutive failure threshold to 3 times, the constraint relaxation mechanism is triggered when the system fails to solve the problem for three consecutive control cycles. This threshold should not be set too low to avoid frequent relaxation triggers due to occasional solution failures; nor should it be set too high to prevent the system from remaining in a solutionless state for an extended period when relaxation is truly needed. After triggering the constraint relaxation mechanism, the system introduces a relaxation factor to relax the current rate of change constraint and the maximum difference threshold constraint between adjacent partitions, respectively. Specifically, the relaxation is achieved by multiplying the original upper limit of the current rate of change by the relaxation factor; similarly, the original maximum difference threshold between adjacent partitions is multiplied by the relaxation factor to obtain the relaxed maximum difference threshold. The initial value of the relaxation factor is set to one, and no relaxation is performed at this stage. After each solution failure, the relaxation factor increases by a preset step size, for example, by 0.1 each time. The increased relaxation factor expands the boundaries of the two constraints, thereby expanding the feasible region and making it easier for the solver to find a feasible solution. The incrementing process of the relaxation factor continues until a solution is successfully found or the relaxation factor reaches a preset upper limit, for example, an upper limit of 1.5, indicating that the constraints are relaxed by a maximum of 50%. The purpose of setting the upper limit is to prevent the constraints from being excessively relaxed, leading to excessively drastic current changes or excessively large current differences between adjacent partitions, thus affecting the coating quality.

[0056] After a successful solution, the system needs to gradually restore the relaxation factor to its initial value of one to bring the constraints back to normal levels. The restoration is achieved by gradually decreasing the relaxation factor in preset steps; for example, after each successful solution for a control cycle, the relaxation factor is reduced by 0.05 until it returns to one. If consecutive solution failures occur again during the restoration process, the constraint relaxation mechanism is retried, and the relaxation factor is increased again. Through this dynamic relaxation and restoration mechanism, the system can automatically address solution difficulties caused by changes in operating conditions or parameter drift while maintaining normal constraint strength. This ensures the robustness of the control system and avoids the negative impact of long-term use of relaxed constraints on coating quality.

[0057] Step S4: Extract the current setting value corresponding to the first control cycle from the optimal current setting value sequence and send it to the independent power supply corresponding to each power supply zone, and control the independent power supply to supply power to the corresponding power supply zone according to the current setting value.

[0058] Specifically, the current setpoint corresponding to the first control cycle is extracted from the optimal current setpoint sequence. That is, each power supply zone only takes its optimal current value within the first future cycle as the execution command for the current cycle. The control system sends these current setpoints to their respective independent power supplies via an industrial communication protocol. Upon receiving the command, each power supply adjusts its output current through its internal proportional-integral-derivative (PID) control loop, ensuring the actual output current follows the setpoint in real time. The power supply then supplies power to the corresponding power supply zone according to the adjusted current. The current is transferred to the electrode surface through the anode and plating solution, completing the electroplating deposition process. Because only the command for the first cycle is executed, and the system recalculates based on the latest measured data after each control cycle, it can continuously correct model deviations during rolling optimization, gradually converging the plating thickness to the target value.

[0059] Step S5: After each control cycle ends, steps S1 to S4 are repeated based on the latest measured data to form a rolling time-domain closed-loop control.

[0060] Specifically, after each control cycle ends, the system re-executes steps S1 to S4. First, the total measured current of each power supply zone in the current cycle is re-acquired, and the latest effective current density distribution sequence is obtained by combining it with the updated crosstalk matrix. Then, the latest current density distribution sequence is input into the prediction model, and the current density distribution trend for the next few cycles is re-predicted based on the latest measured data. Next, the multi-objective optimization problem is reconstructed and solved based on the new prediction sequence to obtain a new optimal current setpoint sequence. Finally, the current setpoint of the first cycle in the new sequence is extracted and sent to each independent power supply for execution. The above process is repeated in each control cycle, forming a rolling time-domain closed-loop control. Here, "rolling" means that the optimization time domain moves forward over time, and each cycle is re-predicted, re-optimized, and re-executed based on the latest measured data. "Closed-loop" means that the control decision not only depends on the prediction model, but also on the measured data fed back by downstream sensors such as thickness gauges. The system continuously corrects prediction deviations and model errors through feedback. The advantages of this rolling time-domain closed-loop control are twofold: firstly, each cycle utilizes the latest measured data for correction, effectively suppressing the impact of model bias and random disturbances on control accuracy; secondly, since only the first instruction in the optimized sequence is executed in each cycle, the system always maintains its ability to predict future trends, avoiding the lag problem inherent in pure feedback control. Through this mechanism, the system can achieve long-term stable high-precision control during the continuous plating process of wide-width electrode sheets.

[0061] In summary, the continuous plating control method for wide-width electrode sheets of power batteries provided by this invention has the following technical effects: First, at the sensing level, through the hardware design of insulating shielding plate and multi-point flexible conductive contact, combined with crosstalk matrix decoupling algorithm based on singular value decomposition, the effective current density of each zone is accurately obtained, eliminating the influence of current crosstalk and transverse potential drop on measurement accuracy.

[0062] Second, at the prediction level, a time series prediction model is constructed using a long short-term memory network with an attention mechanism. Combined with incremental fine-tuning and elastic weight consolidation regularization, it achieves accurate prediction of the current density distribution trend in multiple future control cycles, overcoming the control lag problem caused by the large delay characteristics of the electroplating process.

[0063] Third, at the optimization level, based on the model predictive control framework, a multi-objective optimization problem is constructed, which includes minimizing thickness deviation, smoothing current change, and various process constraints. By solving the problem through convex quadratic programming and with the adaptive adjustment of weight coefficients and constraint relaxation mechanism, the global collaborative optimal allocation of current setting values ​​for each power supply zone is achieved, which significantly improves the lateral consistency and longitudinal stability of the coating thickness.

[0064] In summary, this invention can effectively suppress edge effects, improve the areal density uniformity of coating thickness, thereby enhancing the rate performance and batch consistency of power batteries, while reducing the scrap rate caused by uneven coating, and realizing high-precision and robust intelligent control of the wide-area electrode continuous coating process.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0066] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for continuous plating control of wide-width electrode sheets for power batteries, characterized in that, The methods include: Step S1: Set up multiple independently controlled power supply zones along the width direction of the wide electrode sheet of the battery, obtain the effective current density of each power supply zone in real time, and construct the current current density distribution sequence. Step S2: Input the current current density distribution sequence into the pre-built time series prediction model to predict and obtain the predicted current density distribution sequence for multiple consecutive control cycles in the future; Step S3: Based on the target coating thickness and the predicted current density distribution sequence, construct a multi-objective optimization problem that includes the objectives of minimizing thickness deviation, smoothing current change, and process constraints, and solve it to obtain the optimal current setpoint sequence for each power supply zone in multiple consecutive control cycles in the future; Step S4: Extract the current setting value corresponding to the first control cycle from the optimal current setting value sequence and send it to the independent power supply corresponding to each power supply zone, and control the independent power supply to supply power to the corresponding power supply zone according to the current setting value; Step S5: After each control cycle ends, steps S1 to S4 are repeated based on the latest measured data to form a rolling time-domain closed-loop control.

2. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 1, characterized in that, Real-time acquisition of the effective current density of each of the power supply zones, including: An insulating shielding plate is provided between adjacent power supply zones. The lower end of the insulating shielding plate maintains a preset gap distance with the surface of the wide electrode sheet of the battery. The preset gap distance is calculated based on the conductivity of the plating solution and the electrode sheet conveying speed according to a preset adjustment rule. Before the wide electrode sheet of the battery enters the plating area, multiple flexible conductive contacts are arranged along the width direction of the wide electrode sheet of the battery. Each flexible conductive contact is independently configured with an elastic support element and a pressure sensor. The pressure sensor monitors the contact pressure between each contact and the surface of the electrode sheet in real time. When the contact pressure of any contact deviates from the preset pressure range, the pressure adjustment mechanism is triggered to adjust the preload of the elastic support element of the contact. An offline calibration experiment is performed to construct an initial crosstalk matrix. The offline calibration experiment includes: applying a test current to each of the power supply zones individually in sequence, measuring the thickness distribution curve using a downstream thickness gauge, integrating and averaging the thickness distribution curve according to the width interval corresponding to each of the power supply zones to obtain the average thickness value corresponding to each zone, normalizing all the average thickness values ​​to obtain the current contribution ratio of the current energized zone to each zone, and combining the current contribution ratios of all zones to form the initial crosstalk matrix. During online operation, the total measured current of each power supply zone is collected in real time to form a measured current vector. Singular value decomposition is performed on the initial crosstalk matrix. Components with singular values ​​less than a preset threshold are discarded, and an approximate matrix of the crosstalk matrix is ​​reconstructed. The pseudo-inverse matrix of the approximate matrix is ​​then calculated. The product of the measured current vector and the pseudo-inverse matrix is ​​calculated to obtain the effective current vector. Each effective current value in the effective current vector is then divided by the electrode area of ​​the corresponding power supply zone to obtain the effective current density of each power supply zone.

3. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 2, characterized in that, The method further includes: After a preset number of electrode rolls are produced, the crosstalk matrix online update process is triggered. In the online update process, the interruption closed-loop control is continuously interrupted for a preset duration. During the interruption duration, a disturbance current of preset amplitude is sequentially superimposed on each power supply zone, and the change in the total current of each zone and the change in the downstream thickness distribution curve are measured. The incremental change of the crosstalk matrix is ​​estimated online using the recursive least squares algorithm, and the incremental change is superimposed on the current crosstalk matrix to obtain the updated crosstalk matrix.

4. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 1, characterized in that, The steps involved in building a time series forecasting model include: Historical operating data is collected as a training dataset, which includes the effective current density, plating solution temperature, electrode tape speed, total voltage and current setpoints for each power supply zone in multiple past control cycles. The training dataset is normalized by scaling each feature variable to a preset numerical range to obtain the sample training set. A long short-term memory network structure with an attention mechanism is constructed. The attention mechanism is used to automatically calculate the weight coefficients of each time step in the input sequence at each prediction time step, so that the model adaptively focuses on historical moments with weight coefficients greater than a preset attention threshold. The mean squared error is used as the loss function, and offline training is performed using the sample training set until the loss function value on the validation set no longer decreases after several consecutive training rounds, thus generating a time series prediction model.

5. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 1, characterized in that, The method further includes: After each preset number of electrode rolls are produced, the newly accumulated running data will be added to the training dataset. The time series prediction model is subjected to incremental fine-tuning training with a preset small batch training round using a fine-tuning learning rate that is lower than the offline training learning rate. During the incremental fine-tuning training, an elastic weight consolidation regularization term is used to suppress catastrophic forgetting. The elastic weight consolidation regularization term is assigned different constraint strengths according to the importance of network parameters in historical training.

6. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 1, characterized in that, Construct a multi-objective optimization problem that includes the objectives of minimizing thickness deviation, smoothing current variation, and process constraints, including: The objective function for minimizing thickness deviation is established as the sum of squares of the differences between the predicted coating thickness and the target coating thickness obtained by multiplying the predicted current density of each power supply zone by the Faraday constant over multiple consecutive control cycles in the future. Establish the objective function for smoothness of current change: the sum of squares of the differences in current setpoints between two adjacent control cycles for each power supply zone in a series of future control cycles; Establish a safety boundary penalty function: adopting a soft constraint form, when the current setting value of any of the power supply zones is greater than or equal to the sum of the preset current safety lower limit and the first safety boundary, or less than or equal to the difference between the preset current safety upper limit and the second safety boundary, a positive penalty term is output according to the square relationship between the current setting value and the current safety lower limit or the current safety upper limit, wherein the first safety boundary and the second safety boundary are both preset positive values; The total objective function is formed by multiplying the thickness deviation minimization objective function by a first weighting coefficient, the current change smoothness objective function by a second weighting coefficient, and the safety boundary penalty function by a third weighting coefficient. The following constraints are set: the current setting value of each power supply zone is greater than or equal to the preset lower current safety limit and less than or equal to the preset upper current safety limit; the current change rate of each power supply zone between two adjacent control cycles is less than or equal to the preset maximum change rate; the absolute value of the difference between the current setting values ​​of two adjacent power supply zones within the same control cycle is less than or equal to the preset maximum difference threshold between adjacent zones.

7. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 6, characterized in that, The method further includes: Real-time monitoring of the thickness deviation statistical characteristics fed back by the downstream thickness gauge, wherein the thickness deviation statistical characteristics include the mean deviation and the variance of the deviation; When the mean deviation is greater than the preset mean threshold, the first weight coefficient is increased by a preset first step length until the mean deviation is less than or equal to the preset mean threshold or the first weight coefficient reaches the preset weight upper limit value. When the deviation variance is greater than the preset variance threshold, the second weight coefficient is increased by a preset second step size until the deviation variance is less than or equal to the preset variance threshold or the second weight coefficient reaches the preset weight upper limit value. When the mean deviation is less than or equal to the preset mean threshold and the variance deviation is less than or equal to the preset variance threshold, the first weight coefficient and the second weight coefficient are decreased by a preset third step size until the first weight coefficient and the second weight coefficient are restored to their respective preset initial values. The sum of the first weight coefficient, the second weight coefficient and the third weight coefficient is equal to 1, and each is restricted to a closed interval formed by a preset lower weight limit and a preset upper weight limit.

8. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 6, characterized in that, The optimal current setpoint sequence for each power supply zone over multiple consecutive control cycles is obtained by solving the following: The overall objective function is transformed into a standard quadratic form expression, the constraints are transformed into a system of linear inequalities, and the multi-objective optimization problem is formalized into a convex quadratic programming problem. The interior point method is used to solve the problem. The quadratic programming solver is called, and the quadratic coefficient matrix and linear coefficient vector of the total objective function, the coefficient matrix of the constraint conditions, and the constraint boundary values ​​are input into the quadratic programming solver. The quadratic programming solver outputs the optimal solution vector that minimizes the total objective function under all constraints through iterative calculation. The optimal solution vector contains the optimal current setting value for each power supply zone in multiple consecutive control cycles in the future. The current setpoint of the first control cycle is extracted from the optimal solution vector as the execution instruction of the current control cycle. When the convex quadratic programming problem fails to reach the convergence condition within the preset maximum number of iterations, the first value in the optimal current setpoint sequence obtained in the previous control cycle is used as the execution instruction of the current control cycle, and a solution failure event is recorded.

9. The continuous plating control method for wide-width electrode sheets of a power battery according to claim 8, characterized in that, The method further includes: When the number of consecutive failures in solving the convex quadratic programming problem reaches a preset threshold for consecutive failures, the constraint relaxation mechanism is triggered. In the constraint relaxation mechanism, a relaxation factor is introduced to relax the current change rate constraint and the maximum difference threshold constraint between adjacent partitions respectively. The initial value of the relaxation factor is set to one, and it is increased by a preset step size after each failure until the solution is successful or the preset upper limit of the relaxation factor is reached. Once the solution is successful, gradually decrease the relaxation factor until it returns to its initial value.

Images