Water-based adhesive coating control method and system based on artificial intelligence optimization

Abstract

The invention provides a water-based adhesive coating control method and system based on artificial intelligence optimization, and the method comprises the steps: obtaining a historical data set composed of process parameters of a coating process and coating quality parameters, carrying out the clustering, and obtaining a global induction point set according to a clustering result; determining a stage induction point subset according to the target parameter value of the current coating stage and the boundary of the target parameter, and updating the sparse Gaussian process regression model by using the global induction point set and the stage induction point subset; calculating a quality fluctuation index, determining the length of a prediction time domain based on the quality fluctuation index, constructing an optimization problem in the prediction time domain by adopting a multi-level opportunity constraint mode, and when the deviation value between the actual value of any key process parameter and the prediction trajectory based on the sparse Gaussian process regression model exceeds a deviation threshold value, determining that the prediction trajectory does not exceed the deviation threshold value. And solving the optimization problem to obtain an optimal control action sequence in the prediction time domain, and determining a final control action from the optimal control action sequence and sending the final control action to an execution mechanism.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a water-based adhesive coating control method and system based on artificial intelligence optimization. Background Art

[0002] Water-based adhesive coating plays a vital role in numerous industrial sectors, including packaging, labels, tapes, medical, electronics, automotive, and building materials. The process involves uniformly and precisely applying a liquid water-based adhesive to a continuously moving substrate. Through subsequent drying and curing steps, a coating with specific functionalities is formed. Coating quality directly impacts the performance and reliability of the final product, including the accuracy and uniformity of coating thickness, coating weight per unit area, surface flatness, and the presence of bubbles. Coating quality requires precise control of multiple interrelated process parameters, such as substrate speed, adhesive supply rate, and pressure. Traditional coating process control relies primarily on manual adjustments. While control methods exist for coating control, they are insufficient for complex, dynamic processes like water-based adhesive coating, which exhibit significant nonlinearity and time-varying characteristics. Model predictive control (MPC) can handle multivariable constraints and optimize control performance. However, existing MPC methods cannot adapt to the real-time fluctuations in coating quality. In addition, the quality requirements of different horizontal areas of the coated product often vary, making it impossible to ensure good adaptability at different stages. Summary of the Invention

[0003] To address the problems of insufficient accuracy and adaptability in existing water-based adhesive coating control methods, this application proposes a water-based adhesive coating control method based on artificial intelligence optimization, including: Obtain a historical data set consisting of process parameters and coating quality parameters of the water-based adhesive coating process within a preset time period before the current moment, perform density clustering on the historical data set, select the data point with the largest density in each cluster and add it to the global induction point set, calculate the local variance of the coating quality parameter, and add the data points with local variance greater than a threshold to the global induction point set; determine the target parameter value and the boundary of the target parameter for the current coating stage, determine the stage induction point subset based on the target parameter value and the boundary of the target parameter, and use the global induction point set and the stage induction point subset to update the sparse Gaussian process regression model; The quality fluctuation index is calculated according to the coating quality parameters collected in real time, and the length of the prediction time domain is determined based on the quality fluctuation index. The optimization problem is constructed in the prediction time domain by adopting a multi-level chance constraint method. When the deviation between the actual value of any key process parameter and the prediction trajectory based on the sparse Gaussian process regression model exceeds the deviation threshold, the optimization problem is solved to obtain the optimal control action sequence in the prediction time domain, the target control action is determined from the optimal control action sequence, and the target control action is sent to the actuator.

[0004] Optionally, determining the stage induction point subset according to the target parameter value and the boundary of the target parameter includes: Obtain a historical data subset corresponding to the current coating stage. If the absolute difference between a data point in the historical data subset and the target parameter value is less than a first preset threshold or the absolute difference with the boundary is less than a second preset threshold, add the data point to the stage induction point subset.

[0005] Optionally, the updating of the sparse Gaussian process regression model using the global induction point set and the stage induction point subset includes: The global induction point set and the stage induction point subset are merged and then duplicates are removed to obtain an induction point set; if the total number of the induction point set exceeds the maximum number of induction points, the average value of the kernel function similarity between each induction point in the induction point set and all other induction points in the set is calculated, and the induction point with the largest average kernel function similarity value is removed from the induction point set, and the process is repeated until the total number of induction points is equal to the maximum number of induction points; Obtaining, from the historical data set, input features and target outputs associated with each induction point in the induction point set to form a training data set for model updating; The inducing point set and the training data set are used to iteratively optimize the kernel function parameters, the likelihood function noise variance and / or the positions of the inducing points of the model by maximizing the marginal likelihood function of the sparse Gaussian process regression model.

[0006] Optionally, determining the length of the prediction time domain based on the quality fluctuation index includes: Selecting upper and lower limits of the quality fluctuation index, and setting the maximum and minimum allowable values ​​of the prediction time domain length; When the calculated quality fluctuation index is greater than an upper limit, setting the length of the predicted time domain to a maximum allowable value of the predicted time domain length; When the calculated quality fluctuation index is less than a lower limit, setting the length of the prediction time domain to a minimum allowable value of the prediction time domain length; When the calculated quality fluctuation index is between the upper and lower limits, the length of the current prediction time domain is kept unchanged.

[0007] Optionally, constructing the optimization problem in the prediction time domain by adopting a multi-level opportunity constraint approach includes: The coating is divided into three coating quality control areas along the coating width, including the critical area in the center and the non-critical areas on both sides of the edge; Setting a different coating quality parameter constraint satisfaction probability threshold for each of the coating quality control areas, wherein the probability threshold for the critical area is higher than the probability threshold for the non-critical area; When constructing the optimization problem, for each time step in the prediction time domain and each quality control area, the sparse Gaussian process regression model is used to predict the mean and variance of the corresponding coating quality parameter; For each quality control area and each time step, the upper confidence limit and lower confidence limit of the coating quality parameter are calculated based on the mean, the variance and the constraint satisfaction probability threshold corresponding to the quality control area, and the upper confidence limit and lower confidence limit are used as constraints of the optimization problem. The objective function of the optimization problem is to minimize the cumulative sum of deviations between the predicted values ​​of the coating quality parameters in each spatial area in the prediction time domain and the target parameter values, and / or minimize the cumulative sum of changes in the control actions.

[0008] Optionally, the calculating an upper confidence limit and a lower confidence limit of the coating quality parameter based on the mean, the variance, and a constraint satisfaction probability threshold corresponding to the quality control area includes: For each quality control area and each time step, a probability adjustment value is calculated using the inverse cumulative distribution function of the standard normal distribution and a constraint satisfaction probability threshold corresponding to the quality control area; The square root of the variance is calculated to obtain a prediction standard deviation, the product of the prediction standard deviation and the probability adjustment value is calculated, an upper confidence limit is obtained by adding the mean and the product, and a lower confidence limit is obtained by subtracting the mean and the product.

[0009] Optionally, determining a target control action from the optimal control action sequence includes: Calculate the confidence level of the sparse Gaussian process regression model for the prediction result after applying the first control action in the optimal control action sequence; If the confidence level is lower than a low confidence threshold, the amplitude of the first control action is reduced; if the confidence level is higher than a high confidence threshold, the amplitude of the first control action is increased.

[0010] This application also proposes a water-based adhesive coating control system based on artificial intelligence optimization, including: a processing unit configured to obtain a historical data set consisting of process parameters and coating quality parameters of a water-based adhesive coating process within a preset time period before a current moment, perform density clustering on the historical data set, select the data point with the largest density in each cluster and add it to a global induction point set, calculate the local variance of the coating quality parameter, and add the data points with a local variance greater than a threshold value to the global induction point set; determine a target parameter value and a boundary of the target parameter for a current coating stage, determine a stage induction point subset based on the target parameter value and the boundary of the target parameter, and update a sparse Gaussian process regression model using the global induction point set and the stage induction point subset; A control unit is used to calculate a quality fluctuation index based on coating quality parameters collected in real time, and determine the length of a prediction time domain based on the quality fluctuation index, and construct an optimization problem in the prediction time domain using a multi-level chance constraint method. When the deviation between the actual value of any key process parameter and the predicted trajectory based on the sparse Gaussian process regression model exceeds a deviation threshold, the optimization problem is solved to obtain an optimal control action sequence in the prediction time domain, a target control action is determined from the optimal control action sequence, and the target control action is sent to an actuator.

[0011] Optionally, determining the stage induction point subset according to the target parameter value and the boundary of the target parameter includes: Obtain a historical data subset corresponding to the current coating stage. If the absolute difference between a data point in the historical data subset and the target parameter value is less than a first preset threshold or the absolute difference with the boundary is less than a second preset threshold, add the data point to the stage induction point subset.

[0012] Optionally, the updating of the sparse Gaussian process regression model using the global induction point set and the stage induction point subset includes: The global induction point set and the stage induction point subset are merged and then duplicates are removed to obtain an induction point set; if the total number of the induction point set exceeds the maximum number of induction points, the average value of the kernel function similarity between each induction point in the induction point set and all other induction points in the set is calculated, and the induction point with the largest average kernel function similarity value is removed from the induction point set, and the process is repeated until the total number of induction points is equal to the maximum number of induction points; Obtaining, from the historical data set, input features and target outputs associated with each induction point in the induction point set to form a training data set for model updating; The inducing point set and the training data set are used to iteratively optimize the kernel function parameters, the likelihood function noise variance and / or the positions of the inducing points of the model by maximizing the marginal likelihood function of the sparse Gaussian process regression model.

[0013] Optionally, determining the length of the prediction time domain based on the quality fluctuation index includes: Selecting upper and lower limits of the quality fluctuation index, and setting the maximum and minimum allowable values ​​of the prediction time domain length; When the calculated quality fluctuation index is greater than an upper limit, setting the length of the predicted time domain to a maximum allowable value of the predicted time domain length; When the calculated quality fluctuation index is less than a lower limit, setting the length of the prediction time domain to a minimum allowable value of the prediction time domain length; When the calculated quality fluctuation index is between the upper and lower limits, the length of the current prediction time domain is kept unchanged.

[0014] Optionally, constructing the optimization problem in the prediction time domain by adopting a multi-level opportunity constraint approach includes: The coating is divided into three coating quality control areas along the coating width, including the critical area in the center and the non-critical areas on both sides of the edge; Setting a different coating quality parameter constraint satisfaction probability threshold for each of the coating quality control areas, wherein the probability threshold for the critical area is higher than the probability threshold for the non-critical area; When constructing the optimization problem, for each time step in the prediction time domain and each quality control area, the sparse Gaussian process regression model is used to predict the mean and variance of the corresponding coating quality parameter; For each quality control area and each time step, the upper confidence limit and lower confidence limit of the coating quality parameter are calculated based on the mean, the variance and the constraint satisfaction probability threshold corresponding to the quality control area, and the upper confidence limit and lower confidence limit are used as constraints of the optimization problem. The objective function of the optimization problem is to minimize the cumulative sum of deviations between the predicted values ​​of the coating quality parameters in each spatial area in the prediction time domain and the target parameter values, and / or minimize the cumulative sum of changes in the control actions.

[0015] Optionally, the calculating an upper confidence limit and a lower confidence limit of the coating quality parameter based on the mean, the variance, and a constraint satisfaction probability threshold corresponding to the quality control area includes: For each quality control area and each time step, a probability adjustment value is calculated using the inverse cumulative distribution function of the standard normal distribution and a constraint satisfaction probability threshold corresponding to the quality control area; The square root of the variance is calculated to obtain a prediction standard deviation, the product of the prediction standard deviation and the probability adjustment value is calculated, an upper confidence limit is obtained by adding the mean and the product, and a lower confidence limit is obtained by subtracting the mean and the product.

[0016] Optionally, determining a target control action from the optimal control action sequence includes: Calculate the confidence level of the sparse Gaussian process regression model for the prediction result after applying the first control action in the optimal control action sequence; If the confidence level is lower than a low confidence threshold, the amplitude of the first control action is reduced; if the confidence level is higher than a high confidence threshold, the amplitude of the first control action is increased.

[0017] This application utilizes global induction points and a subset of stage induction points determined according to the characteristics of the current coating stage to improve the prediction accuracy of the model and its adaptability to process changes; and adjusts the prediction time domain length according to the coating quality fluctuation index, so that the controller can have stronger disturbance suppression ability when the quality fluctuates violently, and reduce the amount of calculation when the quality is stable. Different quality constraint satisfaction probability thresholds are set for different areas in the coating width direction, realizing refined management of quality risks in each area, ensuring high quality in key areas, and avoiding excessive constraints on non-key areas. This application can reduce scrap rate, improve raw material utilization and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of Example 1; Figure 2 Schematic diagram of selecting the data point with the largest density in the cluster as the global induction point; Figure 3 Schematic diagram of guiding model update for induction points; Figure 4 Schematic diagram of thickness standard deviation and prediction time domain; Figure 5 Schematic diagram of the coating quality control area. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0020] In a specific embodiment, the present application proposes a water-based adhesive coating control method based on artificial intelligence optimization, such as Figure 1 Shown, including: S1, obtaining a historical data set consisting of process parameters and coating quality parameters of the water-based adhesive coating process within a preset time period before the current moment, performing density clustering on the historical data set, selecting the data point with the largest density in each cluster and adding it to the global induction point set, calculating the local variance of the coating quality parameter, and adding the data points with a local variance greater than a threshold to the global induction point set; determining the target parameter value and the boundary of the target parameter for the current coating stage, determining a stage induction point subset based on the target parameter value and the boundary of the target parameter, and updating the sparse Gaussian process regression model using the global induction point set and the stage induction point subset; The process parameter data and coating quality parameter data of the water-based glue coating process within the preset historical time window before the current moment are acquired from the data acquisition module, and the process parameter data and coating quality parameter data together constitute a historical data set. In one embodiment, the process parameters include but are not limited to the substrate line speed, glue pump speed, scraper pressure, drying temperature, etc.; the coating quality parameters include but are not limited to coating thickness, uniformity, surface defect count, etc., wherein the quality parameters are acquired through online sensors, such as beta-ray sensors, optical scanners, visual systems, etc. DBSCAN or OPTICS are used to cluster the historical data set to identify dense areas of data distribution, and each cluster in the cluster represents a common operating mode of the process. From each cluster, data points that can represent the core characteristics of the cluster are selected and added to the global induction point set, wherein the selection method includes but is not limited to the highest density point in the cluster, the cluster centroid, etc., Figure 2 A schematic diagram shows the use of the data point with the highest density within a cluster as a global induction point. Simultaneously, to obtain regional information, local statistical characteristics of the coating quality parameters in the historical dataset are calculated, such as local variance or local information entropy. The local variance is achieved by defining a neighborhood around each data point and calculating the variance of the coating quality parameters within that neighborhood. When the calculated local variance exceeds the variance sensitivity threshold, it indicates that the regional process fluctuations at that data point are significant, and the data point is added to the global induction point set. Each data point includes both process and quality parameters. For example, data point A is {substrate linear velocity: 50, glue pump speed: 1000, center coating thickness: 23, left edge coating thickness: 25, right edge coating thickness: 22}.

[0021] Coating involves multiple stages, each with different parameters. The current coating stage is determined, along with the target parameter values ​​and operating boundaries preset for that stage, such as the maximum / minimum temperature and speed limits. In one embodiment, data points adjacent to these target parameter values ​​or operating boundaries, or historical data points corresponding to the current stage, are obtained from the historical data corresponding to the current stage. The Euclidean distance between the historical data point and the target parameter point, or the minimum distance from the data point to each constraint boundary, is calculated. Data points with distances less than the proximity threshold are selected and added to the stage induction point subset.

[0022] The global induction point set is combined with the current stage induction point subset and duplicates are removed to obtain a valid induction point set. Using the induction point set and the historical data set, a sparse Gaussian process regression training algorithm is used to iteratively optimize the model's hyperparameters, such as the kernel function's length scale, signal variance, and noise variance, by maximizing the lower bound of evidence or approximate marginal likelihood. In one embodiment, the position of the induction point itself is also optimized to achieve online updating of the model. The induction point guides the model update, such as Figure 3 The input of the sparse Gaussian process regression model is the data points, i.e., the process parameters and the quality parameters. In another embodiment, the input also includes the control action at the previous moment. The output is the prediction of the future coating quality parameters, such as the predicted values ​​of the coating quality parameters at the next one or more time steps.

[0023] S2, calculates the quality fluctuation index according to the coating quality parameters collected in real time, and determines the length of the prediction time domain based on the quality fluctuation index, constructs the optimization problem in the prediction time domain by adopting the multi-level chance constraint method, and when the deviation between the actual value of any key process parameter and the prediction trajectory based on the sparse Gaussian process regression model exceeds the deviation threshold, solves the optimization problem to obtain the optimal control action sequence in the prediction time domain, determines the target control action from the optimal control action sequence, and sends the target control action to the actuator.

[0024] Based on one or more coating quality parameters collected in real time, such as the real-time mean and standard deviation of coating thickness, a quality fluctuation index is calculated. In a more specific embodiment, the fluctuation index is a weighted average variance or coefficient of variation. Based on the current value of the quality fluctuation index, the prediction time domain for subsequent model predictive control is determined. In one embodiment, a piecewise function or fuzzy logic inference system is used. Figure 4 The relationship between the thickness standard deviation and the prediction time domain length when the segmented method is used is shown. When the quality fluctuation index is high, a longer prediction time domain is selected to enhance the ability to suppress disturbances; when the quality fluctuation index is low, a shorter prediction time domain is selected to reduce the computational complexity of online optimization. Finally, an optimization problem is constructed, which uses multi-level chance constraints to deal with coating quality. Specifically, the coating is divided into multiple control areas along the coating width, and different coating quality parameter target values ​​and constraint limits are set for each area, as well as different constraint satisfaction probability thresholds. For example, the center area requires a 99% probability of satisfying the constraint, and the edge area requires a 95% probability of satisfying the constraint. At each time step of the optimization problem, For each control region j, the updated sparse Gaussian process regression model is used to predict the mean value of the coating quality parameter in the region under the action of the candidate control sequence. and variance Based on the predicted mean, variance and constraint satisfaction probability threshold corresponding to each region, the upper and lower confidence limits of the quality parameter are calculated. In one embodiment, Calculated, where is the inverse cumulative distribution function of the standard normal distribution, is the constraint satisfaction probability threshold for the jth coating quality control region. The upper and lower confidence limits are added as hard constraints to the optimization problem. The objective function of the optimization problem is to minimize the weighted sum of squares of the tracking errors between the predicted and target values ​​of the quality parameters for each region within the prediction horizon, and / or to minimize the weighted sum of squares of the rates of change or amplitudes of the control inputs.

[0025] The actual measured values ​​of one or more predefined key process parameters are obtained and compared with the corresponding parameter trajectory predicted by the model based on the current control strategy. When the cumulative deviation or instantaneous deviation between the actual value of any key process parameter and the model predicted trajectory exceeds the preset deviation alarm threshold, it means that the actual process has significantly deviated from the model expectation, and the solution of the above optimization problem is immediately triggered. The optimization solver calculates the optimal control action sequence within the current prediction time domain. ,in is the control time domain. Calculate if the first control action in the optimal control action sequence is applied After that, the prediction variance value corresponding to the next or more state predictions is used to determine the control action finally applied to the actuator. For example, when the prediction variance value is high, Attenuation is performed to avoid excessive control; when the prediction variance value is low, it is directly adopted The final control action is sent to the corresponding actuator of the coating machine, such as a valve controller, a motor controller, etc., to achieve control of the water-based adhesive coating. In an alternative embodiment, the first control action in the optimal control action sequence is directly sent to the actuator.

[0026] The water-based adhesive coating process involves multiple stages, such as equipment startup, steady-state operation, substrate or adhesive type switching, and shutdown. Not only do the behaviors of these stages differ significantly, but the core control focus also differs. While traditional global induction point selection strategies can focus on the overall process, they cannot adequately address the local dynamics of the current stage. In an optional embodiment, determining a subset of stage induction points based on target parameter values ​​and target parameter boundaries includes: Obtain a historical data subset corresponding to the current coating stage. If the absolute difference between a data point in the historical data subset and the target parameter value is less than a first preset threshold or the absolute difference with the boundary is less than a second preset threshold, add the data point to the stage induction point subset.

[0027] Identify the specific operational stage of the current water-based adhesive coating process and obtain the target parameter values ​​and operational constraint boundaries associated with that stage from a pre-configured parameter database. For each preset process target parameter value, calculate the absolute difference between the actual coating quality parameter value corresponding to that data point and the current stage target value. If the absolute difference is less than a first preset threshold, the system will use the data point as the stage induction point. Alternatively, for each constraint boundary of a key operational parameter, calculate the absolute difference between the actual process parameter value corresponding to the historical data point and these boundaries. If any absolute difference is less than a second preset threshold, the historical data point is used as the stage induction point.

[0028] In an optional embodiment, the updating of the sparse Gaussian process regression model using the global induction point set and the stage induction point subset includes: The global induction point set and the stage induction point subset are merged and then duplicates are removed to obtain an induction point set; if the total number of the induction point set exceeds the maximum number of induction points, the average value of the kernel function similarity between each induction point in the induction point set and all other induction points in the set is calculated, and the induction point with the largest average kernel function similarity value is removed from the induction point set, and the process is repeated until the total number of induction points is equal to the maximum number of induction points; Obtaining, from the historical data set, input features and target outputs associated with each induction point in the induction point set to form a training data set for model updating; The inducing point set and the training data set are used to iteratively optimize the kernel function parameters, the likelihood function noise variance and / or the positions of the inducing points of the model by maximizing the marginal likelihood function of the sparse Gaussian process regression model.

[0029] Merge the global induction point set with the current stage induction point subset to obtain a merged induction point set, perform deduplication on the merged set, and remove identical induction points. If the total number of induction points does not exceed the upper limit , then the set after deduplication is the final set of induction points. If it exceeds the upper limit, for each induction point, calculate the kernel function similarity value between it and all other induction points in the set, and then calculate the average of these similarity values. The induction point with the largest average kernel function similarity value is removed from the set because it has the highest average similarity with other points. Repeat until the total number of induction point sets is reduced to or slightly less than .

[0030] From the complete historical dataset or the portion most relevant to the current process characteristics, input feature vectors and corresponding target output values ​​corresponding to the inducing points are extracted, which together form the training dataset for model updating. The inducing point set is a set of points located in the input space, while the training dataset contains a large number of actual historical observations. Model updating involves approximating the complete Gaussian process based on the entire training data using these inducing points. Using the inducing point set and the training dataset, gradient ascent is employed to maximize the variational lower bound of the log-marginal likelihood function of the sparse Gaussian process regression model. During the optimization process, multiple model parameters are iteratively adjusted and optimized, including but not limited to the hyperparameters of the kernel function describing the correlation between data points, the noise variance of the likelihood function, and / or the location of the effective inducing point in the input feature space, until the optimization objective function converges or the preset maximum number of iterations is reached, thereby completing the update of the sparse Gaussian process regression model.

[0031] In an optional embodiment, determining the length of the prediction time domain based on the quality fluctuation index includes: Selecting upper and lower limits of the quality fluctuation index, and setting the maximum and minimum allowable values ​​of the prediction time domain length; When the calculated quality fluctuation index is greater than an upper limit, setting the length of the predicted time domain to a maximum allowable value of the predicted time domain length; When the calculated quality fluctuation index is less than a lower limit, setting the length of the prediction time domain to a minimum allowable value of the prediction time domain length; When the calculated quality fluctuation index is between the upper and lower limits, the length of the current prediction time domain is kept unchanged.

[0032] Suppose that during a water-based adhesive coating process, the quality fluctuation index is calculated by real-time monitoring of the standard deviation of the coating thickness. The upper and lower limits of the standard deviation are obtained, for example, an upper limit of 0.05mm and a lower limit of 0.01mm. At the same time, the maximum allowable value for the prediction time domain length is 10 time steps, and the minimum allowable value is 3 time steps. In each control cycle, the current coating thickness standard deviation is calculated as the quality fluctuation index. If this index is greater than 0.05mm, indicating that the coating quality fluctuates significantly, the prediction time domain length is set to 10 time steps. If this index is less than 0.01mm, indicating that the coating quality is stable, the prediction time domain length is set to 3 time steps to improve computational efficiency. If the coating thickness standard deviation is 0.03mm, which is between the upper and lower limits, the current prediction time domain length will remain unchanged. For example, if the current prediction time domain length is 5 time steps, this length will continue to be used for subsequent predictions and optimizations.

[0033] In an optional embodiment, constructing the optimization problem in the prediction time domain by adopting a multi-level opportunity constraint approach includes: The coating is divided into three coating quality control areas along the coating width, including the critical area in the center and the non-critical areas on both sides of the edge; Setting a different coating quality parameter constraint satisfaction probability threshold for each of the coating quality control areas, wherein the probability threshold for the critical area is higher than the probability threshold for the non-critical area; When constructing the optimization problem, for each time step in the prediction time domain and each quality control area, the sparse Gaussian process regression model is used to predict the mean and variance of the corresponding coating quality parameter; For each quality control area and each time step, based on the mean, the variance and the constraint satisfaction probability threshold corresponding to the quality control area, the probability adjustment value is calculated using the inverse cumulative distribution function of the standard normal distribution and the constraint satisfaction probability threshold corresponding to the quality control area, the square root of the variance is calculated to obtain the predicted standard deviation, the product of the predicted standard deviation and the probability adjustment value is calculated, the upper confidence limit is obtained by adding the mean and the product, and the lower confidence limit is obtained by subtracting the mean from the product, the upper confidence limit and the lower confidence limit are used as constraints of the optimization problem, and the objective function of the optimization problem is to minimize the cumulative sum of deviations between the predicted values ​​of the coating quality parameters of each spatial area in the prediction time domain and the target parameter values, and / or to minimize the cumulative sum of changes in the control actions.

[0034] Specifically, the width of the substrate is divided into three coating quality control areas. In one embodiment, the width of the central critical area occupies the main part of the effective width of the product, such as the middle 60%-80%. The coating quality of this area plays a decisive role in the performance of the final product; the two edge non-critical areas on both sides of the central area, where slight quality deviations have little impact on the overall function of the product, such as Figure 5 As shown in the figure, each quality control area is pre-set with a target coating quality parameter set value, as well as the upper and lower process constraints that the quality parameter must meet. Each quality control area also has a constraint satisfaction probability threshold. The probability threshold for the central critical area is high, such as 0.95, to ensure reliable quality. The probability threshold for the peripheral non-critical areas is relatively low, such as 0.90.

[0035] In each optimization cycle of the model predictive control, for the current dynamically determined prediction horizon For each future time step k in , and for each quality control area j, the updated sparse Gaussian process regression model is called, which predicts the key coating quality parameters, or the mean and variance of the parameters, for area j at time step k based on the current system state and the candidate control actions in the prediction time domain. Then, the inverse cumulative distribution function of the standard normal distribution is used , combined with the constraint satisfaction probability threshold of region j , calculate a probability adjustment factor , upper confidence limit of quality parameter and the lower confidence limit for: and ,in is the square root of the prediction variance, i.e., the prediction standard deviation. The confidence bound is used as a hard constraint in the optimization problem, which requires that for all j and k: and The objective function of the optimization problem is to minimize the predicted mean value of coating quality parameters in all spatial regions within the entire prediction time domain. and their respective target values The weighted cumulative sum of the deviations between them, and / or, minimizing the change in control action or total energy consumption, thereby achieving smooth and economical control while ensuring the differentiated quality requirements of each area.

[0036] In an optional embodiment, determining the target control action from the optimal control action sequence includes: Calculate the confidence level of the sparse Gaussian process regression model for the prediction result after applying the first control action in the optimal control action sequence; If the confidence level is lower than a low confidence threshold, the amplitude of the first control action is reduced; if the confidence level is higher than a high confidence threshold, the amplitude of the first control action is increased.

[0037] After simulating the first control action, the coating quality parameters at a certain key time point in the future are predicted to obtain the prediction variance In one embodiment, the prediction variance is further mapped to a normalized confidence score C, preferably, .

[0038] After calculating the confidence level C of the prediction result after the first control action is applied, if the calculated confidence level C is lower than the preset low confidence threshold, it indicates that the prediction uncertainty is large. In order to avoid overshoot, oscillation or constraint violation, the amplitude of the first control action is reduced. In one embodiment, the originally planned control action change u is adjusted to ,in is a decay factor. If the calculated confidence level C is higher than the preset high confidence threshold, it indicates that the prediction uncertainty is small. In order to achieve the control target faster, the amplitude of the first control action is increased. In one embodiment, u is adjusted to ,in is a gain factor; if the confidence level C is between the low confidence threshold and the high confidence threshold, the amplitude of the originally planned first control action is not changed.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. In addition, the various different implementations of the embodiments of the present invention can also be arbitrarily combined, as long as they do not violate the ideas of the embodiments of the present invention, and they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. A water-based adhesive coating control method based on artificial intelligence optimization, characterized in that: include: Obtain a historical data set consisting of process parameters and coating quality parameters of the water-based adhesive coating process within a preset time period before the current moment, perform density clustering on the historical data set, select the data point with the largest density in each cluster and add it to the global induction point set, calculate the local variance of the coating quality parameter, and add the data points with local variance greater than a threshold to the global induction point set; determine the target parameter value and the boundary of the target parameter for the current coating stage, determine the stage induction point subset based on the target parameter value and the boundary of the target parameter, and use the global induction point set and the stage induction point subset to update the sparse Gaussian process regression model; The quality fluctuation index is calculated according to the coating quality parameters collected in real time, and the length of the prediction time domain is determined based on the quality fluctuation index. The optimization problem is constructed in the prediction time domain by adopting a multi-level chance constraint method. When the deviation between the actual value of any key process parameter and the prediction trajectory based on the sparse Gaussian process regression model exceeds the deviation threshold, the optimization problem is solved to obtain the optimal control action sequence in the prediction time domain, the target control action is determined from the optimal control action sequence, and the target control action is sent to the actuator.

2. The method according to claim 1, characterized in that The step of determining the phase induction point subset according to the target parameter value and the target parameter boundary includes: Obtain a historical data subset corresponding to the current coating stage. If the absolute difference between a data point in the historical data subset and the target parameter value is less than a first preset threshold or the absolute difference with the boundary is less than a second preset threshold, add the data point to the stage induction point subset.

3. The method according to claim 1, characterized in that The method of updating the sparse Gaussian process regression model using the global induction point set and the stage induction point subset includes: The global induction point set and the stage induction point subset are merged and then duplicates are removed to obtain an induction point set; if the total number of the induction point set exceeds the maximum number of induction points, the average value of the kernel function similarity between each induction point in the induction point set and all other induction points in the set is calculated, and the induction point with the largest average kernel function similarity value is removed from the induction point set, and the process is repeated until the total number of induction points is equal to the maximum number of induction points; Obtaining, from the historical data set, input features and target outputs associated with each induction point in the induction point set to form a training data set for model updating; The inducing point set and the training data set are used to iteratively optimize the kernel function parameters, the likelihood function noise variance and / or the positions of the inducing points of the model by maximizing the marginal likelihood function of the sparse Gaussian process regression model.

4. The method according to claim 1, wherein The determining of the length of the prediction time domain based on the quality fluctuation index includes: Selecting upper and lower limits of the quality fluctuation index, and setting the maximum and minimum allowable values ​​of the prediction time domain length; When the calculated quality fluctuation index is greater than an upper limit, setting the length of the predicted time domain to a maximum allowable value of the predicted time domain length; When the calculated quality fluctuation index is less than a lower limit, setting the length of the prediction time domain to a minimum allowable value of the prediction time domain length; When the calculated quality fluctuation index is between the upper and lower limits, the length of the current prediction time domain is kept unchanged.

5. The method according to claim 1, wherein The optimization problem is constructed in the prediction time domain by adopting a multi-level opportunity constraint method, including: The coating is divided into three coating quality control areas along the coating width, including the critical area in the center and the non-critical areas on both sides of the edge; Setting a different coating quality parameter constraint satisfaction probability threshold for each of the coating quality control areas, wherein the probability threshold for the critical area is higher than the probability threshold for the non-critical area; When constructing the optimization problem, for each time step in the prediction time domain and each quality control area, the sparse Gaussian process regression model is used to predict the mean and variance of the corresponding coating quality parameter; For each quality control area and each time step, the upper confidence limit and lower confidence limit of the coating quality parameter are calculated based on the mean, the variance and the constraint satisfaction probability threshold corresponding to the quality control area, and the upper confidence limit and lower confidence limit are used as constraints of the optimization problem. The objective function of the optimization problem is to minimize the cumulative sum of deviations between the predicted values ​​of the coating quality parameters in each spatial area in the prediction time domain and the target parameter values, and / or minimize the cumulative sum of changes in the control actions.

6. The method according to claim 5, characterized in that The calculating the upper confidence limit and the lower confidence limit of the coating quality parameter based on the mean, the variance, and the constraint satisfaction probability threshold corresponding to the quality control area includes: For each quality control area and each time step, a probability adjustment value is calculated using the inverse cumulative distribution function of the standard normal distribution and a constraint satisfaction probability threshold corresponding to the quality control area; The square root of the variance is calculated to obtain a prediction standard deviation, the product of the prediction standard deviation and the probability adjustment value is calculated, an upper confidence limit is obtained by adding the mean and the product, and a lower confidence limit is obtained by subtracting the mean and the product.

7. The method according to claim 1, characterized in that Determine the target control action from the optimal control action sequence, including: Calculate the confidence level of the sparse Gaussian process regression model for the prediction result after applying the first control action in the optimal control action sequence; If the confidence level is lower than a low confidence threshold, the amplitude of the first control action is reduced; if the confidence level is higher than a high confidence threshold, the amplitude of the first control action is increased.

8. A water-based adhesive coating control system based on artificial intelligence optimization, characterized in that: include: a processing unit configured to obtain a historical data set consisting of process parameters and coating quality parameters of a water-based adhesive coating process within a preset time period before a current moment, perform density clustering on the historical data set, select the data point with the largest density in each cluster and add it to a global induction point set, calculate the local variance of the coating quality parameter, and add the data points with a local variance greater than a threshold value to the global induction point set; determine a target parameter value and a boundary of the target parameter for a current coating stage, determine a stage induction point subset based on the target parameter value and the boundary of the target parameter, and update a sparse Gaussian process regression model using the global induction point set and the stage induction point subset; A control unit is used to calculate a quality fluctuation index based on coating quality parameters collected in real time, and determine the length of a prediction time domain based on the quality fluctuation index, and construct an optimization problem in the prediction time domain using a multi-level chance constraint method. When the deviation between the actual value of any key process parameter and the predicted trajectory based on the sparse Gaussian process regression model exceeds a deviation threshold, the optimization problem is solved to obtain an optimal control action sequence in the prediction time domain, a target control action is determined from the optimal control action sequence, and the target control action is sent to an actuator.

9. The system according to claim 8, characterized in that The step of determining the phase induction point subset according to the target parameter value and the target parameter boundary includes: Obtain a historical data subset corresponding to the current coating stage. If the absolute difference between a data point in the historical data subset and the target parameter value is less than a first preset threshold or the absolute difference with the boundary is less than a second preset threshold, add the data point to the stage induction point subset.

10. The system according to claim 8, wherein: The method of updating the sparse Gaussian process regression model using the global induction point set and the stage induction point subset includes: The global induction point set and the stage induction point subset are merged and then duplicates are removed to obtain an induction point set; if the total number of the induction point set exceeds the maximum number of induction points, the average value of the kernel function similarity between each induction point in the induction point set and all other induction points in the set is calculated, and the induction point with the largest average kernel function similarity value is removed from the induction point set, and the process is repeated until the total number of induction points is equal to the maximum number of induction points; Obtaining, from the historical data set, input features and target outputs associated with each induction point in the induction point set to form a training data set for model updating; The inducing point set and the training data set are used to iteratively optimize the kernel function parameters, the likelihood function noise variance and / or the positions of the inducing points of the model by maximizing the marginal likelihood function of the sparse Gaussian process regression model.

Images