A method for feedforward predictive control of a coating die and slurry pump based on ML

Abstract

The present application relates to the technical field of coating die control methods, in particular to a feedforward prediction control method for a coating die and slurry pump based on ML, which comprises the following steps: collecting coating process parameters in real time to construct an input feature vector; calculating a basic adjustment amount according to the partition area density deviation and the average area density deviation by using a PID control algorithm; inputting the basic adjustment amount and the input feature vector into a machine learning prediction model to obtain an area density change prediction value; calculating a prediction trust coefficient in combination with historical prediction accuracy, generating a feedforward correction amount based on the deviation between the area density change prediction value and a target value and the prediction trust coefficient; and outputting a final control instruction after superimposing the basic adjustment amount calculated by the PID control algorithm and the feedforward correction amount. Through the feedforward fusion of machine learning prediction and PID control, the present application effectively overcomes the hysteresis of traditional PID control, reduces waste caused by area density abnormalities, and significantly improves the stability and consistency of the coating process (higher CPK value and lower COV value).

Description

Technical Field

[0001] This invention relates to the technical field of coating die head control methods, and in particular to a feedforward predictive control method for coating dies and slurry pumps based on machine learning (ML). Background Technology

[0002] In lithium battery manufacturing, the coating process is a crucial step determining battery performance. The coating die system needs to precisely control the coating thickness and areal density of the slurry on the substrate to ensure battery consistency and performance. Traditional lithium battery electrode coating control mainly employs a PID closed-loop control scheme, adjusting the T-die gap (T-block stroke) and slurry pump speed based on feedback from an areal density detector. Specifically:

[0003] 1. Calculate the adjustment amount of T-block and the pump speed of slurry pump based on the feedback data from the dry film surface density meter.

[0004] 1.1 Feedback data is obtained through a dry film surface density detector. Because the wet film surface density differs greatly from the dry film surface density due to uneven water evaporation after baking in the oven, the dry film surface density data is generally used for closed-loop control.

[0005] 1.2 After obtaining the feedback data from the dry film surface density meter, input it into the PID controller to calculate the adjustment amount of the T block and the issued value of the slurry pump speed. Adjust the T block and the pump speed according to the adjustment amount of the T block and the issued value of the slurry pump speed respectively.

[0006] 2. The control methods for the above-mentioned T-block adjustment amount and the slurry pump speed have the following problems:

[0007] 2.1 Lag Control Problem: In the control methods described above, traditional feedback control systems can only adjust after detecting a deviation in areal density. The dry film areal density control cycle is limited by the physical location of the areal density meter and the coating speed, typically ranging from 2 to 3 minutes. This adjustment lag can lead to significant waste when the areal density is abnormal. An example illustrating the lag control problem is provided below. Figure 1 This is a production data point. The vertical axis represents the average areal density of the film area detected by the dry film thickness gauge, and the horizontal axis represents the measurement pass number of the thickness gauge. In actual production, the PID algorithm determines that the pump speed needs to be reduced and issues a command to the slurry pump at the 23rd pass. The foil travels at a speed of 50 m / min. At the 40th pass of the dry film thickness gauge, the average areal density begins to decrease. The scanning speed of the dry film thickness gauge is approximately 5 s / pass. Therefore, the hysteresis distance of the traditional PID algorithm control on the foil is at least 17. 5 50 / 60=70.83 meters: This results in waste material with an abnormal surface density of about 70.83 meters.

[0008] 2.2 Difficulty in Multivariable Coupled Control: There is a complex nonlinear coupling relationship between the position of the T-block and the speed of the slurry pump. Traditional control methods are difficult to establish an accurate mathematical model, resulting in limited control effects.

[0009] 2.3 Poor adaptability to changes in process parameters: When process conditions such as slurry viscosity, temperature, and humidity change, the parameters of traditional controllers need to be readjusted, which is not adaptable enough.

[0010] 2.4 Insufficient utilization of historical data: Traditional control systems cannot effectively utilize large amounts of historical production data to optimize control strategies, resulting in a waste of data resources.

[0011] Therefore, in the slurry coating process, how to solve the lag problem of PID control to avoid the generation of waste is an urgent problem to be solved. Summary of the Invention

[0012] The purpose of this invention is to provide a coating control method to solve the hysteresis problem of PID control.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: a feedforward predictive control method for a coating die head and slurry pump based on machine learning, comprising the following steps:

[0014] S1: Data acquisition is performed based on the coating production line. Data acquisition includes historical data acquisition, real-time data acquisition, and real-time areal density data acquisition. Real-time data acquisition includes process parameters such as T-block position offset, slurry pump speed change, actual physical position of T-block, backflow pressure, and cavity pressure, to obtain the input feature vector at time t+1. t+1 represents the current time; real-time areal density data acquisition includes the areal density value p(t+1) measured by the areal density meter at time t+1; historical data acquisition obtains the model training data set. ;

[0015] S2: Basic adjustment calculation, based on the zonal surface density deviation measured by the surface density meter at the current moment. and mean surface density deviation The basic adjustment amount of each partition T block is calculated by the PID controller. Adjustment amount of slurry pump foundation ;

[0016] S3: Feedforward prediction and correction, adjusting the basic adjustment amount of the T block. Adjustment amount of slurry pump foundation The current physical location of block T, the backflow pressure, and the cavity pressure are combined to form the prediction input vector. Predict the input vector Input into a machine learning prediction model to determine the predicted value of areal density change. Subsequently, based on the predicted value of areal density change... With the target value of areal density change The bias and prediction confidence coefficient γ are used to generate the feedforward correction amount of the T block. and feedforward correction amount of slurry pump ;

[0017] S4: Final control output, adjusting the T block's base adjustment amount. Adjustment amount of slurry pump foundation With T-block feedforward correction and feedforward correction amount of slurry pump Add them together to obtain and issue the final T-block adjustment amount. and the final slurry pump adjustment ;

[0018] S5: Model update: Based on the set conditions, the machine learning prediction model is updated on a rolling basis using new production data.

[0019] Preferred model training dataset include and Y Where t+1 represents the current time, Y is the model training input feature vector obtained from real-time data acquisition at time t in S1; In actual production, based on The change in areal density produced under the indicated working conditions; the change in areal density Y The input feature vector for model training at time t Historical data collection forms the model training dataset. Model training dataset Used for training machine learning prediction models. Built using a rolling time window approach, Represented as:

[0020] ,

[0021] in, ; Let be the surface density value collected by the surface density meter at time t. Let t be the areal density value collected when the foil at the die head moves to the areal density meter, N be the total number of training samples, and Y be the areal density value collected at time t. t+1 represents the change in surface density; t+1 represents the current time.

[0022] Preferably, the machine learning prediction model is a Long Short-Term Memory (LSTM) network model; the computation process of the LSTM model at time t includes forgetting gate, input gate, cell state update, and output gate operations, and its loss function... Defined as mean square error:

[0023] ;

[0024] Where N is the total number of training samples, and These are the actual and predicted values ​​of the areal density change for the i-th sample, respectively.

[0025] Then, at time t+1, the prediction formula of LSTM is:

[0026] ;

[0027] parameter and parameters loss function Parameters under the minimization condition.

[0028] Preferably, the machine learning prediction model is a support vector regression (SVR) model; the SVR model uses radial basis functions as kernel functions in the prediction phase. Its prediction function is defined as:

[0029] ;in The number of support vectors, for One of the support vectors, k=1 , and Let be the Lagrange multiplier and b be the bias term. These variables can be obtained through training.

[0030] Preferably, the prediction confidence coefficient γ is calculated based on historical prediction errors, and its calculation formula is as follows: Where h is the historical prediction window, The predicted value of the areal density change in the i-th prediction. The absolute value of the error between the actual value of the areal density change Y(i+1) and the actual value of the areal density change. The standard deviation of the error. It is an exponential function, providing nonlinear decay characteristics.

[0031] Preferably, the T-block feedforward correction amount and feedforward correction amount of slurry pump Based on the predicted value of areal density change With the target value of areal density change The bias and prediction confidence coefficient γ are generated, and the calculation formulas are as follows:

[0032] ;

[0033] ;

[0034] in, and These are the correction coefficients for T-block adjustment and pump speed adjustment, respectively, which can be adjusted according to actual working conditions. The target value for the change in areal density. γ is the predicted value of the areal density change, and γ is the prediction confidence coefficient.

[0035] Preferably, the machine learning prediction model further includes a multi-model adaptation step: in the model training step, the model training dataset is used... Parallel training includes at least several types of machine learning models, such as LSTM and SVR, and is based on the coefficients of determination of each model on the validation set. To select the optimal model under the current operating conditions for generating the T-block feedforward correction amount and feedforward correction amount of slurry pump .

[0036] Preferably, the determination coefficients of each model on the validation set are used. To select the optimal model under the current working conditions, specifically: calculate and compare the determination coefficients of each model. Select the coefficient of determination The largest model is used as the feedforward correction for generating T-blocks. and feedforward correction amount of slurry pump The prediction model; where the coefficient of determination The calculation formula is:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] in It is the sum of squared residuals; The total sum of squares; Y(i+1) is the predicted value of the areal density change for the i-th time, and Y(i+1) is the actual value of the areal density change for the i-th time. is the mean of the actual values ​​of the areal density change; n is the total number of samples in the validation set.

[0042] Preferably, in S2, the PID controller calculates the basic adjustment amount of each partition T block. Adjustment amount of slurry pump foundation The process is as follows: Surface density deviation of the m-th partition and mean surface density deviation The calculation is as follows:

[0043] ;

[0044] in, Let m be the areal density value of the m-th partition. The target value of areal density determined for the process;

[0045] This represents the total number of T blocks that need to be controlled. and The relationship between them is:

[0046] ;

[0047] PID controller calculates T-block basic adjustment Adjustment amount of slurry pump foundation The formula is:

[0048] ;

[0049] ;

[0050] Where kp is the proportionality coefficient, ki is the integral coefficient, and kd is the differential coefficient. , Representative proportion segment, , Represents the points system. , This represents the differential element.

[0051] Preferably, the trigger condition for the model update step is: the amount of sampled data reaches the set total number of N training samples.

[0052] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0053] This invention constructs a feedforward prediction and correction method by integrating machine learning prediction models. This method can predict the impact of adjustment commands (such as T-block movement and pump speed changes) on future areal density before the coating action is executed, and generate a T-block feedforward correction amount. and feedforward correction amount of slurry pump This allows the system to intervene before detecting areal density deviations, fundamentally solving the 2-3 minute control cycle lag problem caused by the physical location of the areal density meter and the coating speed in traditional PID feedback control. Figure 1 As shown in the background technology analysis, traditional methods result in approximately 70 meters of waste, while the present invention, through early intervention, can significantly shorten or even eliminate this abnormal range, effectively reducing the scrap rate.

[0054] By integrating PID basic regulation with prediction-based feedforward correction, the control method can more accurately and smoothly approximate the target surface density value. For example... Figure 5 and Figure 6 The experimental data comparison shows that after adopting the method of this invention (white area), the average process capability index (CPK) is significantly higher than that of traditional PID control (yellow area), while the average areal density coefficient of variation (COV) is significantly reduced. This indicates that the present invention not only makes the average areal density closer to the process center value, but also significantly reduces the lateral and longitudinal areal density fluctuations, thereby significantly improving the consistency and overall quality of coating.

[0055] This invention utilizes historical data from a rolling time window to update machine learning models (such as LSTM or SVR) online, enabling the prediction model to continuously adapt to changes in process parameters such as slurry viscosity, temperature, and ambient humidity. This self-learning capability overcomes the shortcomings of traditional PID control systems, which have fixed parameters and require manual retuning to adapt to different operating conditions, thus improving the control system's ability to adapt to parameter changes.

[0056] There is a complex nonlinear coupling relationship between the T-block position and the slurry pump speed. The machine learning model of this invention, through learning from historical data, can intrinsically capture and characterize this coupling relationship. During feedforward prediction, the model can comprehensively evaluate the impact of T-block adjustment and pump speed adjustment on surface density, thereby outputting a coordinated correction, which to some extent achieves multivariable control instead of traditional independent PID loop control.

[0057] This invention transforms historical production data into resources for training models, solving the problem of low data utilization in traditional control systems. The method supports the use of various machine learning models (such as LSTM and SVR) and can dynamically select the current optimal model through a multi-model adaptive selection mechanism (based on the R² metric). The framework is flexible and scalable, reserving space for the subsequent integration of more advanced algorithms. Attached Figure Description

[0058] Figure 1 A trend diagram showing the impact of changes in slurry pump speed on the mean surface density in existing control methods;

[0059] Figure 2 The overall system architecture flow of the control method of this invention;

[0060] Figure 3 This is a schematic diagram of the feature vector rolling acquisition of the control method of the present invention;

[0061] Figure 4 This is a framework diagram of the fusion control strategy of the control method of the present invention;

[0062] Figure 5The CPK trend graphs for the feedforward prediction algorithm and the traditional PID algorithm in the control method of this invention are shown.

[0063] Figure 6 This is a COV trend diagram of the feedforward prediction algorithm and the traditional PID algorithm in the control method of this invention. Detailed Implementation

[0064] As described in the background section, existing traditional PID control algorithms for calculating the T-block adjustment and the slurry pump speed have a lag control problem, which cannot properly address changes in process conditions such as slurry viscosity, temperature, and humidity. After numerous experiments and tests, the researchers have developed the method of this invention. In proposing this method, we considered solving the lag problem of existing PID control. At the same time, each output value is based on the current process environment, and historical process parameters and data are combined to further improve the reliability of T-block adjustment and pump speed adjustment control.

[0065] This invention proposes a feedforward predictive control method for coating die head and slurry pump based on machine learning algorithm. Machine learning algorithm is abbreviated as ML in the industry. It refers to the ability of machine learning model to predict the change value of areal density through basic adjustment by learning historical data. On this basis, the adjustment is intervened in advance to make the areal density adjustment more stable and accurate. The specific technical solution is as follows.

[0066] A feedforward predictive control method for a coating die and slurry pump based on machine learning includes the following steps:

[0067] S1: Data acquisition is performed based on the coating production line. Data acquisition includes historical data acquisition, real-time data acquisition, and real-time areal density data acquisition. Real-time data acquisition includes process parameters such as T-block position offset, slurry pump speed change, actual physical position of T-block, backflow pressure, and cavity pressure, to obtain the input feature vector at time t+1. t+1 represents the current time; real-time areal density data acquisition includes the areal density value p(t+1) measured by the areal density meter at time t+1; historical data acquisition obtains the model training data set. .

[0068] The model training data set include and Y Where t+1 represents the current time, Y is the model training input feature vector obtained from real-time data acquisition at time t in S1; In actual production, based on The change in areal density produced under the indicated working conditions; the change in areal density Y The input feature vector for model training at time t Historical data collection forms the model training dataset. Model training dataset Used for training machine learning prediction models. Built using a rolling time window approach, Represented as:

[0069] ;

[0070] in, ; Let be the surface density value collected by the surface density meter at time t. Let t be the areal density value collected when the foil at the die head moves to the areal density meter, N be the total number of training samples, and Y be the areal density value collected at time t. t+1 represents the change in surface density; t+1 represents the current time.

[0071] S2: Basic adjustment calculation, based on the zonal surface density deviation measured by the surface density meter at the current moment. and mean surface density deviation The basic adjustment amount of each partition T block is calculated by the PID controller. Adjustment amount of slurry pump foundation ;

[0072] S3: Feedforward prediction and correction, adjusting the basic adjustment amount of the T block. Adjustment amount of slurry pump foundation The current physical location of block T, the backflow pressure, and the cavity pressure are combined to form the prediction input vector. Predict the input vector Input into a machine learning prediction model to determine the predicted value of areal density change. Subsequently, based on the predicted value of areal density change... With the target value of areal density change The bias and prediction confidence coefficient γ are used to generate the feedforward correction amount of the T block. and feedforward correction amount of slurry pump .

[0073] The machine learning prediction model is either a Long Short-Term Memory (LSTM) network model or a Support Vector Regression (SVR) model. The LSTM model's computation process at time t includes forgetting gate, input gate, cell state update, and output gate operations. Its loss function... Defined as mean square error:

[0074] ;

[0075] N is the total number of training samples. and These are the actual and predicted values ​​of the areal density change for the i-th sample, respectively.

[0076] Then, at time t+1, the prediction formula of LSTM is:

[0077] ;

[0078] parameter and parameters loss function Parameters under the minimization condition.

[0079] The machine learning prediction model is a support vector regression (SVR) model; the SVR model uses radial basis functions as kernel functions in the prediction phase. Its prediction function is defined as:

[0080] ;in The number of support vectors, for One of the support vectors, k=1 , and Let be the Lagrange multiplier and b be the bias term. These variables can be obtained through training.

[0081] The prediction confidence coefficient γ is calculated based on historical prediction errors, and its calculation formula is as follows: Where h is the historical prediction window, The predicted value of the areal density change in the i-th prediction. The absolute value of the error between the actual value of the areal density change Y(i+1) and the actual value of the areal density change. The standard deviation of the error. It is an exponential function, providing nonlinear decay characteristics.

[0082] In the model training step, the model training dataset is used. Parallel training includes at least several types of machine learning models, such as LSTM and SVR, and is based on the coefficients of determination of each model on the validation set. To select the optimal model under the current operating conditions for generating the T-block feedforward correction amount and feedforward correction amount of slurry pump .

[0083] The T-block feedforward correction amount and feedforward correction amount of slurry pump Based on the predicted value of areal density change With the target value of areal density change The bias and prediction confidence coefficient γ are generated, and the calculation formulas are as follows:

[0084] ;

[0085] ;

[0086] in, and These are the correction coefficients for T-block adjustment and pump speed adjustment, respectively, which can be adjusted according to actual working conditions. The target value for the change in areal density. γ is the predicted value of the areal density change, and γ is the prediction confidence coefficient.

[0087] S4: Final control output, adjusting the T block's base adjustment amount. Adjustment amount of slurry pump foundation With T-block feedforward correction and feedforward correction amount of slurry pump Add them together to obtain and issue the final T-block adjustment amount. and the final slurry pump adjustment ;

[0088] S5: Model update. Based on set conditions, the machine learning prediction model is updated continuously using new production data. The amount of sampled data reaches the set total of N training samples.

[0089] The specific system architecture during implementation is as follows: Figure 2 As shown, it includes a data processing layer, a model training layer, a control decision layer, and an execution layer. A detailed description of each layer is as follows:

[0090] The data processing layer collects data based on the coating production line. Data collection includes historical data acquisition, real-time data acquisition, and real-time areal density data. Real-time data acquisition refers to the real-time acquisition of process parameters such as the T-block position offset, slurry pump speed change, actual physical position of the T-block, backflow pressure, and cavity pressure of the coating production line, in order to construct the input feature vector at the current time t+1. Real-time areal density data includes the areal density value p(t+1) measured by an areal density meter at time t+1; historical data collection is used to obtain the model training data set. The input feature vectors collected in previous data processing Combined with the input feature vector The obtained surface density change Y(t+1) is used as historical data collection, and the historical data collection yields the training data set. (t).

[0091] The model training layer includes the input training data set. The training database of (t), the machine learning prediction model trained online based on the training database, model selection, and the training data set obtained from historical data collection. Online training of machine learning prediction models. The best selected model is used to calculate the predicted output areal density change. Predicted value of areal density change The T-block basic adjustment is derived from the PID controller. 1. Slurry pump foundation adjustment amount and input feature vector Calculations yielded the predicted value based on the change in areal density. Target value for surface density change And the predicted confidence coefficient γ output T-block feedforward correction amount and feedforward correction amount of slurry pump .

[0092] The control decision layer includes inputting the areal density value p(t+1) measured from real-time areal density data into the PID control algorithm to calculate the basic adjustment amount. The calculation result of the basic adjustment amount is compared with the T-block feedforward correction amount output by the prediction model. and feedforward correction amount of slurry pump The control inputs are fused, and the T-block basic adjustment is obtained from the PID control algorithm. Adjustment amount of slurry pump foundation The feedforward correction amount of the T-block is generated by the feedforward predictive control of the predictive model. and feedforward correction amount of slurry pump The control quantities are fused to obtain the final T-block adjustment quantity. and the final slurry pump adjustment .

[0093] The execution layer includes T-block regulation and pump speed regulation, based on the final T-block regulation amount issued by the control decision layer. and the final slurry pump adjustment T-block adjustment and pump speed adjustment are performed by the T-block adjustment mechanism and the slurry pump adjustment mechanism, respectively.

[0094] Input feature vector in real-time data acquisition The establishment of a multi-dimensional feature space includes the relationship between the T-block position offset, slurry pump speed change, T-block position, slurry pump speed, process parameters (such as backflow pressure, cavity pressure, etc.) and areal density change. The input feature vector... The vector at time t+1 is represented as :

[0095] ;

[0096] Input feature vector It includes the following elements:

[0097] : T-block position offset;

[0098] Pump speed change;

[0099] : The actual physical location of block T;

[0100] : Backflow pressure;

[0101] : Cavity pressure;

[0102] ...

[0103] Other process parameters, such as slurry temperature, and a series of other characteristics.

[0104] Figure 3 A schematic diagram of feature vector acquisition for a single sampling period, and the calculation of the surface density change Y:

[0105] ;

[0106] in The surface density value collected by the surface density meter at time t; The isometric density value collected by the isometric density meter when the foil at the die head moves to the isometric density meter at time t.

[0107] Training dataset collected from historical data A rolling training strategy with a time window of N training samples is adopted. The training database and training dataset are updated when the number of samples reaches a set threshold N. as follows:

[0108] Training dataset This database is used for online training of machine learning prediction models. The machine learning prediction models utilize machine learning algorithms such as Long Short-Term Memory (LSTM) networks and Support Vector Regression (SVR) to build predictive models.

[0109] The LSTM model in the entire control method includes a training phase and a prediction phase. First, the training phase of the LSTM model includes the following four parts:

[0110] 1. Construct training samples from the training dataset W(t). The input feature vector X of each sample is usually a historical data sequence within a certain time window (e.g., from tN to t). The corresponding output surface density change Y is the target value to be predicted. The system usually provides feedback after one period, so the historical data sequence of Y is one period later than X (e.g., from t-N+1 to t+1).

[0111] 2. Input the input feature vector X(t) at time t from the training dataset W(t) into the LSTM network. The network processes the sequence information according to its gating mechanism (forget gate, input gate, output gate), such as the input feature vector X(t) at time t generating the predicted value of the areal density change at time t. ;

[0112] 3. Calculate the predicted value of the areal density change at time t. Change in actual surface density The differences between them, such as the predicted value of X(t). The actual surface density change corresponding to X(t) in W(t) loss function Defined as:

[0113] ;

[0114] 4. Calculate the loss function L using the backpropagation algorithm (BPTT) for all model parameters (including...). and The gradient of ) is calculated, and these parameters are updated using optimization algorithms such as gradient descent to minimize the loss function L.

[0115] The LSTM model performs the following during the prediction phase: After the training phase, the loss function is obtained. Parameters under minimization conditions and Then, at time t+1, the prediction formula of LSTM is:

[0116] 。

[0117] The Support Vector Regression (SVR) model in the entire control method includes a training phase and a prediction phase. The training phase of the SVR model consists of four parts:

[0118] 1. Construct training samples from the training dataset W(t), where each sample has an input feature vector. It is the feature vector at time t (which can be derived from a historical data window), and the input feature vector corresponds to the output. It is the actual change in surface density, reflecting the system feedback after one cycle.

[0119] 2. Input the training samples into the SVR model. Based on the training sample set W(t), SVR learns a set of support vectors and their corresponding Lagrange multipliers by solving a convex quadratic programming problem. and , and bias term b. These parameters together define the nonlinear mapping relationship between the input feature vector X and the surface density change value Y.

[0120] 3. The training process of SVR aims to minimize the following objective function:

[0121] ;

[0122] Where N is the total number of training samples; C is the penalty coefficient, the larger the coefficient, the more complex the model and the greater the risk of overfitting, the smaller the coefficient, the simpler the model and the greater the risk of underfitting; The smaller the insensitive loss coefficient, the more details the model considers, resulting in a more refined model. During training, i and j iterate through all feature vectors; after training, a selection will be made. *n* support vectors, where *i* and *j* represent traversing all feature vectors. Kernel function. The calculation formula is:

[0123] ;

[0124] The constraints are:

[0125] ;

[0126] 4. By solving the above optimization problem, the Lagrange multipliers are obtained. and And the bias term b. The support vectors are those that satisfy... sample points , This represents the number of support vectors.

[0127] The SVR model's prediction phase includes the following: for One of the support vectors, k=1 ; and For support vectors The corresponding Lagrange multipliers; at time t+1, for the new input feature vector SVR model predictions Depend on The weighted combination of support vectors generates the following:

[0128] ;

[0129] Kernel function Measuring the input feature vector With support vectors The similarity (the closer the value is to 1, the more similar they are).

[0130] The integrated control strategy framework of the control decision layer, such as Figure 4 As shown, this specifically includes the T-block basic adjustment of the PID controller. Adjustment amount of slurry pump foundation The calculation, feedforward prediction in the prediction model, feedforward correction calculation, and final output of the final T-block adjustment amount are performed. With the final slurry pump adjustment First, the PID controller calculates the basic adjustment of block T. Adjustment amount of slurry pump foundation Based on the surface density deviation of the m-th partition and mean surface density deviation To calculate the base adjustment amount ΔTm_base for block T and the base adjustment amount ΔP_base for the slurry pump, the surface density deviation of the m-th block at time t is then calculated. The formula is:

[0131] ;

[0132] in For the target surface density value, This represents the measured areal density value for the m-th partition.

[0133] Average surface density deviation calculate: and The relationship between them is,

[0134] ;

[0135] This represents the total number of T blocks that need to be controlled.

[0136] The PID controller calculates the basic adjustment T through proportional, integral, and derivative components. Adjustment amount of slurry pump foundation The specific formula is as follows:

[0137] ;

[0138] ;

[0139] Where kp is the proportionality coefficient, ki is the integral coefficient, and kd is the differential coefficient. Representative proportion segment, Represents the points system. This represents the differential element.

[0140] Feedforward prediction refers to the basic adjustment of block T. Adjustment amount of slurry pump foundation and the input feature vector of real-time data acquisition Combined into a prediction input vector Predict the input vector The input model predicts the impact on surface density. The prediction model can be a Long Short-Term Memory (LSTM) network or a Support Vector Regression (SVR) network. The predicted input vector of the prediction model... for:

[0141] ;

[0142] in , The inputs are the T-block and slurry pump foundation adjustments calculated at time t but not yet executed. Other inputs include the actual physical location of T-block at time t, backflow pressure, cavity pressure, and other process parameters. The prediction model output is the predicted value of the areal density change. :

[0143] .

[0144] If the LSTM model is selected, then = ;

[0145] If the SVR model is selected, then .

[0146] Feedforward correction calculation refers to predicting values ​​based on the areal density changes from the previous h iterations. The deviation from the actual value Y of the areal density change is used to set the prediction confidence coefficient. Predicting Trust Ratio calculate:

[0147] ;

[0148] Where h is the historical prediction window, such as the most recent 20 predictions; Predicted value of areal density change Actual value of areal density change The standard deviation of the error between them; It is an exponential function, providing nonlinear decay characteristics; It is a value between 0 and 1; The predicted value of the areal density change in the i-th prediction. Actual value of areal density change The absolute value of the error between them is calculated using the following formula:

[0149] ;

[0150] when When the time is right, it means that the prediction error is very small. Therefore The closer a value is to 1, the more reliable the prediction result. ∈(0,1).

[0151] Feedforward correction calculation refers to calculation based on the predicted value of surface density change. With the target value of areal density change Bias and prediction confidence coefficient Calculate the feedforward correction amount of the T-block and feedforward correction amount of slurry pump :

[0152] ;

[0153] ;

[0154] in and This is the correction coefficient for T-block and pump speed adjustment, which can be adjusted according to actual working conditions. This represents the target value for the change in areal density.

[0155] The final output refers to the T-block base adjustment calculated by the PID controller. Adjustment amount of slurry pump foundation With corresponding feedforward correction amount and Add them together to obtain and issue the final T-block adjustment amount. and the final slurry pump adjustment Provided to the execution layer; final T-block adjustment amount With the final slurry pump adjustment for:

[0156] ;

[0157] .

[0158] The machine learning prediction model selection process involves choosing between LSTM and SVR through a multi-model adaptive selection mechanism: multiple models (such as LSTM and SVR) are trained each time, and the optimal model is selected based on the determination coefficient of each model. Select the optimal model suitable for the current working conditions; coefficient of determination The calculation formula is as follows:

[0159] ;

[0160] in For the sum of squared residuals, Let i be the predicted value of the areal density change for the i-th time. Let be the actual value of the areal density change at the i-th time. and The sum of squares of the differences between the two reflects the unexplained variation in the model, while the sum of squares of the residuals reflects the variation. The calculation formula is:

[0161] ;

[0162] For the total sum of squares, This represents the average of the actual values ​​of the areal density variation. This represents the actual value of the change in areal density. and The sum of squares of the differences between the two sets of data reflects the total variation of the actual observations, where n is the total number of samples in the validation set. The calculation formula is:

[0163] ;

[0164] .

[0165] Coefficient of determination The coefficient of determination represents the ratio between the variance explained by the model and the total variance in the data. The closer it is to 1, the more variance the model can explain, and the better the model's predictive performance. The coefficient of determination should be the preferred choice when selecting a model. A model closer to 1.

[0166] The control flow of this invention mainly includes real-time data acquisition, PID control calculation, predictive model inference, feedforward correction calculation, actuator control command issuance, effect feedback, and model rolling update; in lithium battery coating scenarios, it is generally used... and To measure the quality of coating, The calculation formula is:

[0167] ;

[0168] in, This represents the upper limit of the areal density specified in the manufacturing process. This is the lower limit of the areal density specified in the process specifications. This represents the average of the actual areal density data produced. The standard deviation of the actual areal density data represents the magnitude of fluctuations in the process. The larger the value, the more it represents The larger or Smaller: The larger the value, the closer the areal density value is to the center of the process specification (areal density target value). The smaller the value, the smaller the standard deviation, and the smaller the fluctuation in the lateral areal density data.

[0169] The calculation formula is:

[0170] ;

[0171] in, This represents the average of the actual areal density data produced. The standard deviation of the actual areal density data produced. The smaller the value, the lower the dispersion of the data relative to its average value, the more stable the process output is, and the smaller the fluctuation.

[0172] Figure 5 , Figure 6 for and The trend graphs under traditional PID algorithm and feedforward predictive algorithm control show that the horizontal axis represents the number of passes of the areal density scanner. Figure 5 , Figure 6 The white area on the left side of the middle section represents the execution period of the feedforward prediction algorithm of this invention. Figure 5 , Figure 6 The yellow area on the right side represents the execution period of the traditional PID algorithm.

[0173] from Figure 5 , Figure 6 The data leads to the following conclusions: Figure 5 The average CPK value in the white area is higher than that in the yellow area, indicating that under the control of the feedforward prediction algorithm of this invention, the production process has stronger regulation capabilities, the output areal density data is closer to the target value of areal density required by the process, and the fluctuation is smaller, resulting in better overall quality performance. Figure 6 The mean COV value in the white area is lower than that in the yellow area. Under the control of the feedforward prediction algorithm of this invention, the relative volatility and dispersion of the areal density data are lower, and the production process is more stable and consistent.

[0174] Figure 5 , Figure 6 Data sourced from the following table:

[0175] Table 1: CPK / COV data table corresponding to enabling or disabling the algorithm of this invention

[0176]

[0177] In summary, this invention provides a machine learning-based feedforward predictive control method for coating dies and slurry pumps. This method operates in actual industrial coating equipment, directly controlling physical actuators (T-block adjustment, slurry pump adjustment) and altering the physical state of material coating, thus belonging to industrial process control methods. This method combines data-driven predictive capabilities with classical PID control by constructing a control process that integrates data acquisition, PID basic adjustment, and ML feedforward prediction with a new coefficient. Utilizing a continuously updated machine learning model, it predicts the impact trend of control commands on areal density in real time, and by introducing a prediction confidence coefficient γ, it ultimately generates feedforward correction commands deeply integrated with the PID basic adjustment. This enables the system to compensate for lag and smooth fluctuations, establishing a feedforward prediction mechanism for predicting and preventing deviations, in addition to the traditional feedback control loop of detecting and correcting deviations. Experimental data shows that this method solves the long-standing control lag problem in coating processes, improving the consistency and process stability of areal density control (manifested as higher CPK values ​​and lower COV values).

Claims

1. A method for feedforward predictive control of an ML-based coating die and slurry pump, the method comprising: The method comprises the following steps: ​ S1: data collection based on the coating production line, including historical data collection, real-time data collection and real-time area density data collection; The real-time data acquisition includes the process parameters of the T-block position offset, the slurry pump speed change, the T-block actual physical position, the backflow pressure and the cavity pressure, to obtain the input feature vector at t+1 moment , t+1 represents the current moment; the real-time area density data acquisition includes the area density value p(t+1) measured by the area density instrument at t+1 moment; the historical data acquisition obtains the model training data set ; S2: basic adjustment amount calculation, based on the current time zone density deviation measured by the densimeter and average density deviation , the basic adjustment amount of each partition T block is calculated by the PID controller and the basic adjustment amount of the slurry pump ; S3: Feedforward prediction and correction, combining the T-block base adjustment amount and the slurry pump base adjustment amount with the T-block actual physical position, backflow pressure and cavity pressure at the current time into a prediction input vector , the prediction input vector is input into a machine learning prediction model to determine a surface density change prediction value ; then based on the deviation of the surface density change prediction value from the surface density change target value and the prediction trust coefficient γ, the T-block feedforward correction amount and the slurry pump feedforward correction amount are generated; S4: final control output, add T-block base adjustment amount and slurry pump base adjustment amount and T-block feedforward correction amount and slurry pump feedforward correction amount to get the final T-block adjustment amount and the final slurry pump adjustment amount ; S5: model updating, according to the set condition, using new production data to update the machine learning prediction model; The prediction trust coefficient γ is calculated based on historical prediction errors, and the calculation formula is: ; wherein h is a historical prediction window, is the predicted value of the surface density change in the i-th prediction is the absolute value of the error between the predicted value and the actual value of the surface density change Y(i+1), is the standard deviation of the error, is an exponential function, providing a nonlinear decay characteristic; The T-block feedforward correction amount And the slurry pump feedforward correction amount Based on the surface density change prediction value The deviation of the surface density change target value And the prediction trust coefficient γ generation, the calculation formula is respectively: ； ； Wherein, And Respectively, T-block adjustment correction coefficient and pump speed adjustment correction coefficient, can be adjusted according to the actual working condition, The area density change target value, The area density change prediction value, and gamma is the prediction trust coefficient. The machine learning prediction model further comprises a multi-model adaptive step: in the model training step, a model training data set is used to train multiple types of machine learning models Parallel training of at least LSTM and SVR, and selection of the optimal model under the current working condition based on the determination coefficient of each model on the validation set to generate the T block feedforward correction amount and the slurry pump feedforward correction amount ; the determination coefficients of the models on the validation set to select the optimal model under the current working condition, specifically: calculating and comparing the determination coefficients of each model , and selecting the model with the largest determination coefficient as the prediction model of the T-block feedforward correction amount and the feedforward correction amount of the slurry pump ; wherein the calculation formula of the determination coefficient is: ； ； ； ； wherein, is the residual sum of squares; is the total sum of squares; is the i-th predicted value of the surface density variation, Y(i+1) is the i-th actual value of the surface density variation; is the mean value of the actual values of the surface density variation; n is the total number of samples in the validation set.

2. The method of claim 1, wherein the ML-based coater die and slurry pump feedforward predictive control method is characterized by: Model training data set Comprising and Y , wherein t+1 represents the current time, is the model training input feature vector obtained by real-time data collection at time t in S1; Y is the areal density change amount output based on the working condition represented by in actual production; the areal density change amount Y , the model training input feature vector at time t The model training data set is formed as a data source of historical data collection , the model training data set is used for machine learning prediction model training, is constructed in a rolling time window manner, is represented as: ， wherein, ; is the surface density value collected by the surface density instrument at time t, is the surface density value collected by the surface density instrument at time t, N is the total number of training samples, Y is the surface density change amount; t+1 represents the current time.

3. The method of claim 2, wherein the ML-based coater die and slurry pump feedforward predictive control method is characterized by: The machine learning prediction model is a long short-term memory network (LSTM) model. The calculation process of the LSTM model at time t includes a forgetting gate, an input gate, cell state updating, and an output gate operation, and a loss function is defined in the form of mean square error: ； wherein Ntotal is the total number of training samples, and respectively the actual value of the areal density variation and the predicted value of the areal density variation for the i-th sample. The prediction formula of the LSTM at t+1 is: ； parameters and parameters loss function parameters under the minimization condition.

4. The method of claim 2, wherein the ML-based coater die and slurry pump feedforward predictive control method is characterized by: The machine learning prediction model is a support vector regression (SVR) model; the SVR model uses a radial basis function as a kernel function in a prediction phase and a prediction function thereof is defined as: ； wherein, is the number of support vectors, is one of the K support vectors, k = 1... K , and is a Lagrange multiplier, and b is a bias term, which can be obtained after training.

5. The method of claim 1, wherein the ML-based coater die and slurry pump feedforward predictive control method is characterized by: The PID controller in S2 calculates the basic adjustment amount of each partition T block and the basic adjustment amount of the slurry pump The process is as follows: the mth partition surface density deviation and the average surface density deviation The calculation is as follows: ； wherein, is the areal density value for the mth zone, is the areal density target value determined by the process; the total number of T blocks to be controlled, and the relationship between them is: ； PID controller calculates the T block base adjustment amount and slurry pump base adjustment amount The formula is: ； ； Wherein, kp is proportional coefficient, ki integral coefficient, kd differential coefficient, , represents proportional link, , represents integral link, , represents differential link.

6. The method of claim 2, wherein the ML-based coater die and slurry pump feedforward predictive control method is characterized by: The triggering condition of the model updating step is that the sampling data amount reaches the total number of the set N training samples.

Images